Professional Documents
Culture Documents
,. FOURTH EDITION
3. THEORY OF ERRORS 72
34 The normal or Gaussian law of error; 35 Applicability of
the normal law of error; 36 Normal error distributions;
37 Standard error of a sum or difference; 38 Standard error
of a product; 39 Standard error of a compound quantity;
40 Method of least squares; 41 Weighted mean; 42 Stan-
dard error of weighted mean; 43 Internal and external
consistency; 44 Other applications of the method of least
squares, solution of linear equations; 45 Solution of linear
equations involving observed quantities; 46 Curve fitting;
47 Line of regression; 48 Accuracy of coefficients; 49 Other
curves.
REFERENCES 115
BIBLIOGRAPHY 116
INDEX 117
6 7
CHAPTER 1
ERRORS OF OBSERVATION
"And so we see that the poetry fades out of the problem, and
by the time the serious application of exact science begins we
are left with only pointer readings."
EDDINGTON
On the other hand, systematic errors may arise from the observe 30sec
or the instrument; they are usually the more troublesome, for
repeated observations do not necessarily reveal them and even when
their existence or nature has been established they are sometimes
difficult to eliminate or determine; they may be constant or may
vary in some regular way. For instance, if a dial gauge is used in
which the pivot is not exactly centred, readings which are accurately
observed will be subject to a periodic systematic error. Again,
measurements of the rise of a liquid in a tube using a scale fixed to
the tube will be consistently too high if the tube is not accuratel Fig. 1. Record of the deflexions of a supersensitive torsion balance
vertical; in this case the systematic errors are positive and pro- showing irregular fluctuations in time due to the Brownian motion of the
instrument. (From an investigation by E. Kappler.)
portional to the height of liquid. Further, measuring devices may
be faulty in various ways; even the best possible instruments r are
limited in precision and it is important that the observer shoul •
2. Errors and fractional errors
appreciate their imperfections.
If a quantity Xo units is measured and recorded as x units we
Errors peculiar to a particular observer are often termed persona
shall call x - Xo the error in xo, and denote it usually bye. It
errors; we sometimes speak of the "personal equation." Errors 0
might be positive or negative but we shall assume throughout that
this kind are well authenticated in astronomical work. Bessel, fo
its numerical value is small compared with that of xo; this is usually
instance, examined tile estimates of time passages in astronomica
observations and fouad that systematic differences existed amongs written iel « IXol·
We can write
the leading astronomers of his time. That similar differences exis
amongst students and observers today will be familiar to teacher x=xo+e
and scientific workers alike. = xo(1 + I)
More fundamentally there is the "error" introduced by the ver
process of observation itself which influences in some measure th where 1= e/xo and is known as the fractional error in Xo. Also
phenomenon observed. In atomic physics this is specially im lOOe/xo is called the percentage error in Xo. Of course e/xo can be
portant and is enshrined in the uncertainty principle due t written as e]x approximately when lel<lxol, that is, when 1/1<1.
Heisenberg, but in macroscopic phenomena with which we shall b
mainly concerned it can be neglected. Further, there are man We note that ~= 1 +1
phenomena in physics, often included under the general term Xo
"noise," where fluctuations arise due to atomic or sub-atomic
particles which set a natural limit to the accuracy of measurements. and Xo
x=l+I':::.1 1- I if III « I
These fluctuations are, however, usually very much smaller tha
the errors which arise from other causes. For instance, molecula
e e
bombardments of the suspension of a suitably sensitive galvan Also, ~=~(l+f)':::.:'
meter produce irregular deflexions which can be recorded. Xo x Xo x x
10 11
CHAP. ERRORS OF OBSERVATION ERRORS OF OBSERVATION CHAP. 1
to be less than 0'1, and indeed it is possible on certain reasonable
3. Estimate of error
assumptions to calculate the probability that the error is not greater
If only a single measurement is made any estimate of the error than some assigned magnitude. We shall discuss this more fully
may be widely wrong. There may be a large personal or accidental later.
error, as well as that due to the lack of precision of the particular We note here that if a quantity Xo units is measured n times and
instrument used. recorded as x" X2, - - - xn units, we can write x, = Xo + e, where
To take a simple example, a student timing the oscillations of a e, is the error in the measurement x.. The arithmetic mean x of the
simple pendulum may count 49 swings as 50, and the stop-watch n measurements is
he uses may be accurate to perhaps O· 1 second. Also his "reaction
time," in using the stop-watch, may be different from that of another XI + X2 + - - - + xn =xo + _e!...1
~+_e=-2_+_-_-_-_+---"en
student, but perhaps in a timing such as this the effect of the "reaction n n
time" may be neglected as it is reasonable to assume it is the same at and as some of the errors el> e2, - - - en may be positive and some
the beginning as at the end of the time interval. In this case, assum- negative the value of (el + e2 + - - - + en)/n may be very small. In any
ing he counts the swings correctly the error in the measurement is case it must be smaller numerically than the greatest value of the
dictated by the accuracy of the stop-watch. The fractional error separate errors.
in the timing may be reduced by increasing the number of swings Thus if e is the largest numerical error in any of the measurements
timed but this tends to increase the error in counting the number wa have
of swings, unless some counting device is used. The following
times were obtained by ten different students using the same lei + e2 + -n- - + enl .;;;;e
pendulum and the same watch: 37'2, 37'0, 36'9, 36'7, 36'8, 36·2,
35'4,37'2,36'7,36,8 seconds for 20 swings and 73·8,74'3,74'0. and consequently [e - xol .;;;;e.
74'2,74'4,74'0,73'0,74'1,73'6,74'7 seconds for 40 swings. Hence in general x will be near to Xo and may be taken as the "best"
To obviate or reveal accidental errors, repeated measurements of value of the measured quantity which the measurements provide.
the same quantity are made by the same observer, whenever possible. In general the larger the value of n the nearer x approaches xo.
(If the phenomenon is unique the measurements cannot be repeated; It will be noticed that it is not possible to find e" e2' - - - en or e
for example, we cannot repeat the measurements of an eclipse. Of since Xo is not known. It is usual therefore to examine the scatter
course in a fundamental sense all measurements are unique; they or dispersion of the measurements not about Xo but about x. We
all refer to a particular instant.) A set of repeated measurements hall discuss this more fully later, but if we write x, = X + d, then
might be 10'1, 10'0, 10'0, 10·2, 12'3,10'1,10'1,10'0,10'1,10,2. d, denotes the deviation of xr from x: it is sometimes called the
It seems quite possible that some mistake was made in recording residual of x..
12·3 and it is reasonable to reject it on these grounds. (Results
should however not be rejected indiscriminately or without due We have x, = Xo + e, = x + d,
thought. Abnormal or unusual results, followed up rather than .0 that e; - dr = x - Xo
discarded, have in the hands of a great scientist often led to im- and el + e2 + - - - + en = n(x - xo)
portant discoveries.) The above measurements indicate that the
measured quantity lies between 10·0 and 10'2, and as the arithmetic whereas d + d + - - - + d = O.
l 2 n
mean of the nine measurements is 10 + (1/9)(0'8) ~ 10'1, we can We note then that by repeated measurements of the same quantity
say that the measured quantity is 10'1 ± 0·1 to indicate the scatter uccidental errors of the observer may be corrected in some degree,
of the measurements about the mean. In fact, in stating that the but systematic errors peculiar to him cannot thus be obviated nor
value of the measured quantity is 10'1 the numerical error is likely an any lack of accuracy in the instrument itself. To take an
12 13
CHAP. 1 ERRORS OF OBSERVATION ERRORS OF OBSERVATION CHAP. 1
obvious example, with a scale wrongly marked as 9 instead of 10 instrument. Of these the accidental error is usually assessed by
an observer might record measurements of 9 '0, 9 '1, 9 '0, - - - applying certain statistical concepts and techniques as we shall
instead of 10· 0, 10'1, 10· 0, - - - and the arithmetic mean of the explain in Chapter 2. The other two errors, (b) and (c), are some-
former readings, however many there be, will have a "systematic" times merely neglected or are assumed to be smaller than (a) which
error of 1'0. This example is not so absurd as at first sight it may is not always true. Indeed systematic errors are often the most
seem, for every instrument records m "instead of M where M - m serious source of inaccuracy in experimental or observational work
or (M - m)/M is a measure of the accuracy of the instrument. It and scientists have to devote much time and care to their elimination
is important that the observer should know what degree of accuracy and estimation. Various devices are used depending upon the
he can achieve with the instrument he is using. Indeed this nature of the measurements and of the apparatus used. There is
information helps him to decide whether the instrument is the no common or infallible rule. Certain systematic errors may be
right one to use, whether it is appropriate to the particular end eliminated by making the measurements in special ways or in
he has in view. particular combinations; others may be estimated by comparing
How accurately can a length of about 50 em be measured using the measurements with those obtained when a quite different
a metre rule? Is this sufficiently accurate for the purpose? If not, method or different instrument is employed. On the other hand
should a cathetometer be used? These are the sort of questions the personal error of an observer is usually treated as constant
which any scientist must ask and answer. The answers to these and is determined by direct comparison with an automatic recording
questions dictate the experimental apparatus he uses. It is of listle device or with the results of other observers. Once determined it
avail measuring one quantity with an accuracy of 1 in 1000, say, if is used to correct all the readings made by the observer. Various
another quantity which affects the result equally can only be other corrections may be applied to the readings to take account
measured with an accuracy of 1 in 100. of estimated systematic errors. In all cases the aim is so to correct
Here the distinction that is drawn between the terms accuracy the readings as to ensure that all the errors remaining are accidental.
and precision might be noted. * Accuracy refers to the closeness Sometimes the experiments are so designed and the measurements
of the measurements to the "actual" value or the "real" value of o randomized that any remaining systematic errors acquire the
the physical quantity, whereas the term precision is used to indicate nature of accidental errors. This device is used particularly in many
the closeness with which the measurements agree with one another biological experiments.
quite independently of any systematic error involved. Using the Of course, even when all this has been done systematic errors
notation introduced earlier in this section we say that a set of sometimes remain. Birge has pointed out, for instance, that
measurements Xl, X2 - - -, Xn are of high precision if the residuals d, "atomic weight determinations prior to 1920 were afflicted with
are small whatever the value of x - Xo, whereas the accuracy of unknown and unsuspected systematic errors that averaged fully
the measurements is high if the errors e r are small in which case' ten times the assumed experimental errors." Also, the history of
x - Xo is small too. Accuracy therefore includes precision but the measurement of fundamental physical constants such as the
the converse is not necessarily true. Sometimes more is known velocity of light, the electronic charge and Planck's constant records
of the precision of an instrument than of its accuracy. how various systematic errors have been revealed and corrected.
It will be clear that the assessment of the possible error in any For example, in 1929 the accepted value of the electronic charge was
measured quantity is of fundamental importance in science. It 4.7700 X 10-10 e.s.u.; later it was revised to 4·8025 X 10-10 e.s.u.,
involves estimating (a) the accidental error, (b) the systematic or the difference between these two values being much greater than the
personal error of the observer, and (c) the systematic error of the uccepted accidental error of the earlier value.
Students are recommended to read some of the accounts, par-
• I am grateful to Dr. H. S. Peiser who kindly brought to my notice ticularly the original papers, of experimental work of high accuracy.
Dr. Churchill Eisenhart's article in Photogrammetric Engineering, Vol.
XVIII, No.3, June 1952. If one instructive example may be selected, E. C. Bullard'!' has
14 15
CHAP. 1 ERRORS OF OBSERVATION ERRORS OF OBSERVATION CHAP. 1
discussed some gravity measurements made in East Africa. After The above result can easily be extended. First we note that by
examining very carefully both the accidental and systematic errors putting a = b it follows that the fractional err~r in a2 is t",:ice t~e
involved he concludes that the measurements form a consistent set fractional error in a. Alternatively, the fractional error In a IS
with a probable error of about 0·00001 ms-2• However, he cau- one-half the fractional error in a2, that is, the fractional error in
tiously adds: "While it is believed that the discussion of the errors Nt is one-half the fractional error in N.
includes all those large enough to be of any practical importance it
must be remembered that many apparently irreproachable gravity Thus if we write N= NoO +I)
measurements have in the past been found to be subject to unex we must have Ni ~ NJO + tf)
pected and unexplained errors, and until the source of these dis
crepancies has been found it would be unwise to be dogmatic about This follows of course from the result
the errors in the present work."
O+f)i~l+tf
4. Estimate of the error in compound quantities which can be established independently.
Once the error in a measured quantity has been estimated it is a Again, if Q = abc - - - the fractional error in Q is approximately
fairly simple matter to calculate the value of the consequentia the sum of the fractional errors in a, b, c, - - - .
error in some other quantity on which it depends. " IIIn particular, the fractional error in an equals approximately n
If y is some function of a measured quantity x, the error in times the fractional error in a. This is true for all values of n,
due to some error in x can be found by using some simple positive or negative.
mathematical techniques which we shall now explain.
6. Error in a quotient
5. Error in a product If a quantity Q is expressed as the quotient aib where again a
If a quantity Q is expressed as the product of ab, where a and and b have fractional errors 11 and h respectively, we have
are measured quantities having fractional errorsfJ and/2 respectively,
we can write Q = alb
Q = ab aoO + 11)
= ao(l + fJ) x bo(l + I~ bo(l + I~
41T2/0 1 +
fi . . I b dy..,
<!lIvenapproxunate Y y dx ox.
~ TJ x 1 + 2/2
As an example, suppose y is the area of a circle of radius x, so that
41T2/0 (1 + fi - 2/~
~ T:2
o
x 1 y = 1Tx2
dy
~ go(1 + fi - 2/~. therefore dx = 21TX
Thus the fractional error in g is fi - 2/2, Since fi and 12 may be oy. ..,
positive or negative, it follows that the numerical value of the Hence 8x ~ 21TX, if oX is small
fractional error in g may be as large as lfil + 21121 or as small as nd oy ~ 21TXOX
lfil '" 21121·
If, however, as is often the case in practice, fi and 12 are not Thus the error in the calculated value of the area due to a small
known exactly but may have any value between certain limits, the rror ox in the measured value of the radius is 21TXOX, and this
greatest value of the fractional error can be estimated. For instance, may be positive or negative depending on the sign of ox. It is,
suppose fi lies between - FI and +FI, whilst 12 lies between - F2 of course, the area of the annulus having radii x and x + 8x.
and +F2; then fi - 2/2 can be as large numerically as FI + 2F2• The fractional error in the area is
Indeed we can write
lfi - 2/21 <: FI + 2F2 oy _ 21TXOX = 20X
y - 1TX2 x
. which gives the greatest fractional error in g.
Instead of this greatest value of the fractional error, a smaller that is, approximately twice the fractional error in the radius, in
quantity, sometimes called the "most probable" value, is often ordance with the result proved in Section 5.
quoted. It is given by the square root of the sum of the squares Of course, most simple examples including that above hardly
18 19
..
CHAP. 1 ERRORS OF OBSERVATION ERRORS o·p OBSER V A TlON CHAP. 1
need the use of the calculus; a little algebra is all that is necessary that is Ii - 2h, as we found in Section 6. This result is obtained
as we have shown earlier, but the calculus does facilitate the solution more simply by taking logarithms first, so that
of more complicated problems.
If a quantity Q is a function of several measured quantities logg = log 47T2 + log 1 - 210g T
x, y, z, - - - the error in Q due to errors ox, oy, oz, - - - in
x, y, z, - - - respectively is given by and hence ~ ~ ~/ _ 20T
g - I T
oQ
oQ ~ -
oX
ox + oQ
-oy oy + oQ
-oz OZ + ---
EXAMPLE 1
A quantity y is expressed in terms of a measured quantity x by
The first term ~~ ox is the error in Q due to an error ox in x only the relation y = 4x - (2/x). What is the percentage error in y
(that is, corresponding to oy, OZ, - - - all being zero), and similarly corresponding to an error of 1 % in x?
the second term °o~ oy is the error in Q due to an error oy in y only. We have dy[dx = 4 + (2/x2)
This result is often referred to as the principle of superposition of so that oy ~ [4 + (2/x2)]ox
errors. •
ercentage error in y = (oy/y)l00
Also, if we suppose that ox, oy, OZ, - - - can have any value
between -el and +e" -e2 and +e2, -e3 and +e3, - - - respec-
tively, then the "most probable" value of oQ is given by ~ 1~2/x)[4 + (2/x2)]ox
(OQ)2 =
oQ
(ox X e,
)2 + (OyoQ x e2) +
2 (OZsc )2
x e3 + ---
~ 100(4x2 + 2)ox
- x(4x2 - 2)
that is, 0 Q is the square root of the sum of the squares of the Thus if ox/x = 1/100, the percentage error in y is (4x2 + 2)/(4x2 - 2).
greatest errors due to an error in each variable separately. It might be noted that this percentage error varies with x and is
Taking as an example the simple pendulum formula used earlier approximately 1 % when x is large. Further, it is obviously large
we have when 4x2 - 2 is small, that is, when x lies near to ± 2-t; this is
not surprising since y = 0 when x = ±2-t. Actually using
g = 47T21/T2
oy ~ [4 + (2/x2)]ox, we find that when x = ± 2-t, oy ~ 80x.
therefore og ~ og 0/ + og sr
()/ oT
EXAMPLE 2
2
~ 47T 0/ _ 87T 2
[ sr A quantity y is calculated from the formula y = Cx/(1 + x2).
- T2 T3 If the error in measuring x is e units, find the corresponding error
in y and show that for all values of x it does not exceed Ce units.
og 01 sr
and - ~- -2-
g I T Now
dy 1 + x2 - x(2x) 1 - x2
dx =C (1 + X2)2 = C ,- . -'"
20 21
CHAP. I ERRORS OF OBSERVATION ERRORS OF OBSERVATION CHAP. ,'\
therefore oy ~ C/o
1 -
.
X2
,,,OX .\ EXAMPLE 3
The viscosity of water, 1), is calculated using Poiseuille's formula
giving the quantity of water Q flowing through a cylindrical tube
If the error ox in x is e unit, the error in y is f length I and radius a in time t under a pressure difference p.
Find the error in 1) due to errors in the measured quantities Q, I, a
1 -x2
oy ~ G__ . ".e andp.
'T1'pa4t
We have 1) = 81Q
This can be written oy ~ GeE
therefore ,
1 - x2 log 1) = log ('T1'18) + logp + 410g a + log t - log 1 - log Q
where
E = (1 + X2)2
~il
01) op
-=-+-+-----
40a ot 01 sa
How does E vary with x? When x =0, E =1; when x =I,
E = 0, and when x > 1, E < 0. Also E -+ as x -+ co. Further,
E does not change in value when x is replaced by x and hence
° +
1) pat I Q
Thus the fractional error in 1) equals a combination of the fractional
its graph is symmetrical about the axis of E and must have the rrors in p, a, t, I and Q. The term 40ala is usually the most
form shown in Fig. 2. important since a is very small and hence for an accurate deter-
I' mination of 1) special attention must be paid to the accuracy of
E measurement of the radius a. Indeed to ensure an accuracy of 1 %
in 1), the radius must be measured with an accuracy of at least
1 in 400.
A\
8. Error in a sum or difference
as If a quantity Q is expressed as the sum of two quantities a and b,
having errors el and e2 we have
-3 -2 -I 1 2 3 x Q =a +b
../
01 '--.....---- = ao + bo + el + e2
So writing the error in Q as e and the fractional error in Q as I
-o.s we have
e = el + e2
-I und 1= el + e2 = aofi + bo/2
Fig. 2. Graph of E = (1 - x2)/(1 + x2)2 ao +bo ao + bo
where II and 12 are the fractional errors in a and b respectively.
x
E has a maximum value when x = ° and minimum values when
= ± 3t. Thus E lies between 1 (when x = 0) and -t (when
If el and e2 are known, e and I can be calculated, but as we have
noted earlier all that is usually known is that el may have any
x = ± 3t). Hence the error in y lies between Ce and -GeI8. value between -EI and +EI say, whilst e2 may have any value
22 23
\
J~
between - E2 and + E2. What then are the limits between which EXAMPLE 2
e and j' may lie?
The viscosity TJ of a liquid is measured by a rotation viscometer.
Clearly e can have any value between -(EI + E2) and
The cylinders are of radii a and b, and a torque G is applied to the
+(EI + E2J. or lei lies between 0 and E[ + E2. It is usual to take
rotating cylinder so that
the "most probable" value of lei as (Er + EDt (see Section 6).
which is less than E[ + E2 but bigger than either E[ or E2•
On the other hand the fractional error I depends on the values G (1
TJ = 47T,Q Q2 -
1)
lJi.
of 00 and bo as well as on fi and 12. and varies between wide limits.
where ,Q is the angular velocity of rotation. Calculate the fractional
Writing 1= fi + _b+o b-(f2 - fi) = 12 + 0+0b (fi - 12) error in '1 given that a = 0·04 m, b = 0·05 m, that the greatest error
°0 0 00 0 in measuring both a and b is 0·000] m and that the error in G/D.may
it follows that if 00 is numerically much bigger than bo then I equals be neglected.
Again the "most probable" value of III may be taken to be OTJ = (- ~ fi + ~2 12)/(~ 2 - .!.)
Tj a2 b a b2
Veer + E~)
= 2(a2 h - b2fi)/(b2 - a2)
100 + bol
Hence in this case oTJ/TJ = 2(1612 - 25fi)/9
EXAMPLE 1
where the greatest values of I[ and 12 are 0'01/4 and 0'01/5
If two lengths of 11 and 12 are measured as 10'0 and 9·0 with
respectively.
possi ble errors of 0·1 in each case find (i) the greatest error and Hence 10Tj/TJI may be large as
(ii) the greatest fractional error in the values of 11 + 12 and 11 -/2.
We can write 2(16 x 0,01 + 25 x 0'01)
9 5 4
I[ + 12 = (10'0 ± 0·1) + (9'0 ± 0·1)
that is, 0·021 or about 2%.
= 19·0 ± 0'2
The "most probable" value of oTJ/TJ may be taken as
The greatest error in 11 + 12 is 0·2 and the greatest fractional error is
0'2/19'0 = 0·01. Also, 11 -/2 = 1·0 ± 0·2 so that the greatest error 29v[(16F2J2 + (25FI)2] = ~(0'07) = 0'016
9
in 11 -12 is 0·2 and the greatest fractional error is as high as 0'2/
1·0 = 0·2. We note that the "most probable" fractional error in If there were also errors in G and ,Q the fractional error in G/,Q
11 -/2 is v'[(0·1)2 + (0·])2]/1'0 = 0'14. would have to be added to the value of 101)/TJI calculated above.
24 25
t
CHAP. 1 ERRORS OF OBSERVATION ERRORS OF OBSER VATION CHAP. 1
EXAMPLE 3
In Heyl's method the gravitational constant G is calculated from EXERCISES 1
the formula
1. If y = x2/(l + x2) find oy/y when (a) x = 3, Ox = 0·1 and
(b) x = 2, Ox = 0·05.
G_ 4172/ (1 1)
- Al - A2 Tl - Tl
2. If y = sin (2wt + IX) find the fractional error in y due to an
error of 0·1 % in t when (i) t = 17/2w, (ii) t = 17/W. Find also
the values of t for which the fractional error in y is least.
where T, and T2 are times of oscillation, I is the moment of inertia
of the system about the axis of suspension and AI> A2 are constants. 3. Find the fractional error in eX corresponding to an error Ox
Estimate the fractional error in G due to errors in TI and T2 of in x. If x = 0·012 is correct to two significant figures show
0·1 s when TI = 1750 sand T2 = 2000 s. that eX may be calculated for this value of x correct to four
significant figures.
4. The mass m grammes of an electron moving with velocity u m s " '
Now since G= 4172/ X (T2 - TI)(T2 + T1) is mo/vlp -- (v2/e2)] where em S-1 is the velocity of light. Show
Al - A2 T?Tl that the fractional error in m is approximately v2/e2 times the
• fractional error in v, if v/e is small compared with unity.
using the method of Section 7
S. If i = k tan 8 find the value of 8 for which (i) the error in i is
least, and (ii) the fractional error in i is least, for a given error e
so
-=
O(T2 - T1) O(T2 + T ) 20T
+ ~+~ I ----~ I
20T2 in measuring 8.
G ~-~ ~ 6. Given T2 = h + (lOO/h) find the value of h for which the error
in T is least if the error in h is a constant e. Find also the value
of h for which the fractional error in T is least.
The first term on the right-hand side may be as large as ~(2~0)
7. The diameter of a capillary tube is found by weighing and
and is the most important term. Putting OT2 = - OTI = 1/10 measuring the length of a thread of mercury inserted into the
we get the largest value of OG/G is tube. Estimate the error in the calculated diameter of the tube
in terms of the errors in the length, mass and density of the
mercury thread.
2(1 1 1) 1
10 250 + 1750 - 2000 == 1200 8. Using Kater's pendulum g is given by
L. HOGBEN
Estimate the fractional error in y if h = 1'06 ern, rl = 0·07 ern,
r2 = 0·14 em, and the error in each of these measurements is
9. Frequency distributions
not greater than 0 :005 em.
Numerical data, including scientific measurements as well as
10. The deflexion dot a beam under certain conditions is given by industrial and social statistics, are often represented graphically to
d = 4We3/317Ea4• Find (i) the maximum fractional error, aid their appreciation.
(ii) the "most probable" fractional error in the value of Young's The first step in dealing with such data, if they are sufficiently
modulus E calculated from this formula if the error in d is numerous, is to arrange them in some convenient order; this is
±0·1 %, the error in e is ±0·05 % and the error in a is ±0'1 %. often done by grouping them into classes according to their
11. A coil of n turns of radius r carries a current J; the magnetic magnitude or according to suitable intervals of a variable on which
field at a point on its axis at a distance x from the centre is they depend. For instance, the percentage marks obtained in an
H = 27Tnr2J(r2 + x2)-3J2. If the error in measuring x is e, examination by a number of students could be grouped by counting
find the corresponding error in the value of the field. If e is the number of students who had marks between 0 and 9, 10 and
a constant find for what value of x the error in H is greatest. 19, - - - , 90 and 99, thus dividing them into 10 classes. The data
12. A large resistance R is measured by discharging a condenser of could then be tabulated as shown in Table 1, in which the marks
capacitance C charged to potential Vo; the time t taken for the of a sample of 120 students have been used.
potential to fall to V is noted. R is given by tl C log (Vo/V). The number of data in each class is usually called the frequency
Find the fractional error in R due to errors in t, C, Vo and V. for that class. Table 1 shows what is called the frequency distri-
If Vo and V are measured by a ballistic galvanometer, then bution. The pairs of numbers written in the columns headed
••• Vol V = dold where do and d are the corresponding deflexions "class," for example, 0 and 9, 10 and 19 and so on, are usually
"~·.of the instrument. Assuming the errors in do and d are equal called the lower and upper class limits. The width of any class 15
• in magnitude, show that the greatest value of the corresponding the difference between the first number specifying that class .and
. :.fractional error in R is fo( 1 + ~o)flOg -? where fo is the
the first number specifying the next, that is, 10 for each of .the
classes shown in Table 1. For some groupings, however, .the
widths of the classes may be unequal. ..
. fractional error in Vo Find the value of Vol V for which this
". is least.
Table 1
..•. IISC Lib 8'lore Class Frequency Class Frequency." •
511.43N72 0-9 2 50-59 32
10-19 5 60-69 25
1111111111111111111111111 20-29 6 70-79 10
79638
J 30-39
40-49
14
22
80-89
90-99
2
2
28 29
7963.8
CHAP. 2 SOME STATISTICAL IDEAS SOME STATISTICAL IDEAS CHAP. 2
The classification shown in Table 1 obviously helps us t Ire constructed of width equal to the class width and of area equal
appreciate the distribution of marks amongst the students; we ca to the frequency of the corresponding class, that is, the rectangles
see at a glance, for instance, how many students have fewer than In Fig. 4 have areas equal to 2, 5, 6, 14, 22, 32, 25, 10, 2, 2 units
40 marks and how many have 70 or more. But a graphical repre- respectively. The figure obtained is called a histogram, and the
sentation can make it possibly even clearer. The data are plotted total area of the histogram in this case is 120 units equal to the
in Fig. 3, where the marks are represented along the horizontal total number of students.
Since the area of each rectangle in a histogram represents the
40 frequency in the corresponding class, the heights of the rectangles
are proportional to the frequencies when the classes have equal
30 widths. In this case the mean height of the rectangles is pro-
>- portional to the mean frequency.
~ 20
(])
::::J
0- to. The mean
J 10 The mean frequency is usually not of particular significance, but
what is often important is the mean of the data. This is defined as
follows: if fi, 12, - - - In are the frequencies in the various classes
24·5 44·5 64·5 84·5 of which XI, X2, - - -, Xn are the mid-values of the variable, the
Marks mean value of the variable is given by
Fig. 3. Frequency polygon of examination marks.
(fix! + 12X2 + - - - + /',<)/(/1 + 12 + - - - + IJ (1)
axis and the frequencies along the vertical axis. The points obtained
by plotting the frequency against the mid-value of the corresponding This is the weighted mean of XI, x2, - - - ,xn' the weights being
class, namely, 4·5, 14·5, - - -, 94·5 are joined by straight lines the frequencies in the corresponding classes.
and the resulting figure is known as a frequency polygon. It can be written as
A different method is used in Fig. 4 where a series of rectangles n n
~ Isxs/~ Is or [Ix]/[/]
.=! .~I
unit area
30~ = and is often denoted by s,
To evaluate x using the expression (1) directly can sometimes be
20 laborious, but the arithmetic can be minimized by using the following
simple device.
10 Let us write x. = x~ + m where m is some constant. Then
histogram is shown in Fig. 5 which should be compared with t with recorded heights of 162, 162'5, 163, 163'5 em; strictly in that
original histogram shown in Fig. 4. ease it would include all those with heights between 161·75 em and
The area of ABCDEF in Fig. 5 is (15 + 17) units, the same 163·75em, whieh would be the class-boundaries, and the mid-value
the area of the corresponding rectangle ABCD in Fig. 4. This of the class would be 162'75 cm. However, in the class 162-164 there
true for all corresponding portions of the two histograms, so th might be included the men with recorded heights of 162'5, 163,
although they have different shapes they have equal areas. 163·5em and half of those with recorded heights of 162 and 164 cm ;
34 35
CHAP. 2 SOME STATISTICAL IDEAS SOME STATISTICAL IDEAS CHAP. 2
the mid-value of the 'Classwould then be 163 em. It is important tn regarded as a sample taken from an infinite number distributed in
there should be no doubt where the boundaries of the classes an , manner indicated by the frequency curve. The histogram of the
and that there should be no overlap or gaps between successi trnple is an approximation to the frequency curve, the degree of
classes. ipproximation depending in general on the number of observations
It will be noted that the mean of any number of observatio and on the class width. The histogram tends more and more
depends in general on the width of the classes into which t closely to the frequency curve as the number of observations
observations are grouped, since the mean is defined (see Section 1 Increases and the class width decreases, provided the unit of area is
in terms of the mid-values of the classes. However, provided t NOchosen that the total area under the histogram is always the
class width is not too large the effect of grouping is small and ame as that under the frequency curve, which as stated above is
usually neglected. For example, it is found that the mean of t usually taken as unity.
marks given in Table 1 is 51' 25; if these marks are grouped If the equation of the frequency curve is known, that is, f is
shown in Table 3 the mean is found to be 51'17. Again, us' expressed as some function of the variable x, the mean value of
only one class covering the whole range, 0--99, the mean is 49·5. the variable can usually be found by expressing it in terms of
With a very large number of data grouped into classes of sm Integrals, namely,
width it is clear that the outline of the histogram approximates
a continuous curve, as is illustrated in Fig. 6. Such a curve is kno
as a frequency curve.
X = Ja
b
fxdx/ J a
b
fdx
2[ -1
Fig. 6.
Variable. x
Frequency curve. J1
-1
2J1
f dx = -
71'
-1
-- dx
x2 + 1 = - tan 71'
x J1
-1
= 1
area being finite does not give the total frequency which is infinit
but the area under any part of the curve gives the relative frequen
for the corresponding range of the variable. The value of the variab which is zero. This is obvious from the symmetry of the curve
for which the ordinate divides the area under the curve into t ibout the axis x = O. The median is also x = 0 and so is the mode.
equal parts is therefore the median. On the other hand the mo
14. Measures of dispersion
corresponds to the maximum value of the frequency.
We shall take some examples of frequency curves later. A very important characteristic of a set of data is the dispersion
merely note here that any finite number of observations might or scatter of the data about some value such as the. mean.
36 37
CHAP. 2 SOME STATISTICAL IDEAS SOME STATISTICAL IDEAS CHAP. 2
The numbers 40, 50, 60, 70, 80 of which the mean is 60 have 16. The mean deviation
dispersion or scatter of 20 an each side of the mean; they are If Xl, X2, - - - Xn are a set of data of which .e is the mean, the
within the limits 60 ± 20. Similarly, the numbers 5, 6, 7, 8, 10, deviations of the data from the mean* are (x, - x), (X2 - x),
have a mean of 8 and are within the limits 8 - 3 and 8 + • - - (xn - x) respectively, which we will write as d., d2, - - - dn•
Different sets of data have, in general, different means and differe Some of these deviations are positive and same negative; in fact
dispersions. The data corresponding to' the frequency curve A i: their sum is zero, that is,
Fig. 7 have a smaller dispersion than those represented by t
curve D, whilst the mean of the data A is greater than that of t
data D. Various parameters are used to' measure the dispersio dl + d2 + - - - + d; = Xl + x2 + - - - + xn - nx = O.
We shall mention only three, namely, the range, the mean deviatl
and the standard deviation. The mean deviation (sometimes called the mean absolute deviation)
of a set of data is defined as the mean of the numerical values of
the deviations,
e
c Idd + Id21 + - - - + Idnl = ~ f Idsl·
V
::J n ns~l
f
Far example, the numbers 7, 5,8,10, 12,6 have a mean of 8 and
a mean deviation of (1/6)(1 + 3 + 0+ 2 + 4 + 2) = 2.
Variable If the data Xl' X2, - - - Xn have frequencies ii, 12, - - - In respec-
tively, the mean deviation is given by
Fig. 7. Frequency distributions with different means and dispersions
iildll + 121d21 + - - - + Inldnl }:f.fxs - xl
(3)
15. The range ii+/2+---+ln }:fs
The range of a frequency distribution is defined as the differen
EXAMPLE
between the least and greatest values of the variable. This is
very simple measure of dispersion and has, of course, limitations Consider the data given in Table 2. The mean was found to' be
because of its simplicity. It gives the complete interval of th
variable over which the data are distributed and So' includes tho 3 :~ ~ 3·9. In Table 4 below values of Idsl are given in column
data, usually at the ends of the interval, which occur very .
frequently. It may be much mare useful in certain cases to' kno three and values of f.ldsl in column four, from which it follows that
the portion of the range af the variable within which a giv .. . 58'0 I 32
I mean deviation
uie IS 44 = . .
fraction, say 50 %, of the data lie. .
Far example the range of the marks given in Table 1 is 0-9:
• deviations are sometimes taken from the median.
but 93/120 or 78 % of the marks lie between 30 and 69 inclusive.
38 39
.,~
il,r.~~Y --..-,I,('~
:f~{· tI'O'(Y
~\ ~\£>"I'\" )
~..Iir
CHAP. 2 SOME STATISTICAL IDEAS
SOME STATISTICAL IDEAS CHAP. 2
Table '4 runge (a, b), then the summations can be replaced by integrals and
XJ
0·5
1· 5
X = 3'9,
is
1
5
d. = x, - 3'9
Idsl
3'4
2·4
IMsl
3'4
12·0
wo have
a2 = ra
I(x - x)2dx / ra
fdx
The most imt>ortant measure of dispersion is the standa It might be noted that a coefficient of variation is sometimes used.
deviation, usually denoted by a. It is defined in terms of the squan It is defined as 100 a/x and is expressed as a percentage. In the
of the deviations from the mean as follows: ibove example the coefficient of variation is therefore
If d., d., - - -', t::ln are the deviations of the data xl, X2, - - -, X, 100(2'38)/8::::: 30%
from the mean x, then *
• The reader is ~ked to note the modification introduced in Section 3, which reduces to a = a.
40 41
CHAP. 2 SOME STATISTICAL IDEAS SOMB STATISTICAL IDEAS CHAP
18. Evaluation of standard deviation, a Using a working mean of 3 the necessary working is tabulated
In cases more complicated than the simple example abo below, where d now denotes the deviation from the working mean,
(Example 1), considerable economy in the arithmetic can be mad that is, x - m with m = 3.
by using a suitable working mean m, instead of x, The last two columns are ca.lculated for checking purposes as
As shown earlier explained later in Section 20.
"£fs(xs - m) = "£I.xs - m"£fs Table 5
"£fs(xs - m) "£I.xs - x f d fd fd2 fed + 1) fed + 1)2
so t hat = -- - m =x - m -3,
"£1. "£ I. 0 10 -30 90 -20 40
The summation on the left-hand side, involving deviations fro 1 40 -2 -80 160 -40 40
2 72 -1 -72 72 0 0
the working mean m, can thus be used to find x (see Section 10)
Further, 3 85 0 0 0 85 85
4 78 1 78 78 156 312
'E.fs(xs - x)2 = "£1. [(xs - m) + (m - x»)2 5 55 2 110 220 165 495
= "£fs(xs - m)2 + 2(m - x)"£fs(x. - m) + (m - x)2,,£ 6 32 3 96 288 128 512
Hence, dividing by "£1. and using equation (5) above - 7 18 4 72 288 90 450
8 7 5 35 175 42 252
a
2
= "£1S<xs- x)2
= "£fs(xs - m)2 + 2( m-x -)( -
x-m
) + (m-x -)2 9 2 6 12 72 14 98
"£1. "£1. • 10 J 7 7 49 8 64
= "£fs(xs - m)2 :..- (m _ x)2 sum 400 228 1492 628 2348
"£1.
Hence, from equation (5)
= fL2 - (m - x)2 (6)
We note that "£fs(xs - m)2 is the sum of the squares of th
x- 3= 228/400
deviations from the working mean m. It is usually easier, i therefore • x= 3·57
finding a2, to choose a suitable working mean m and calcula Also fL2 = 1492/400 and hence from equation (6)
fL2 = ,,£fs(x.-m)2/,,£fs and then use
1492
a2 = fL2 - (m - x)2 a2 =- - (0-57)2 = 3'73 - 0·32 = 3'41
400
The special case when m = 0 is of some importance. Also:
one further interpretation of equation (6) might well be noted here therefore a = 1'85
the least value 01 fL2 is a2 and this arises when m = x, that is, tho
We note that in this case the coefficient of variation equals
sum of the squares of the deviations of any set of data, measured
from any value m, is least when m equals the mean of the data. (1'85/3-57) x 100% = 52%
that to allow for the class width c the value of ,.,,2 as calculated abov be found, by using equation (6) of Section 18, in terms of
should under certain conditions be reduced by c2/12. /I,m, x, X', a and a'.
Since c = 1 in this case, we have Another important example of a sum, however, is where a
,.,,2 = 3'73 - 0'08 = 3·65 variable z is expressed as the sum of two or more independent
variables x, y, - - -, say z = x + y or more generally z = ax + by
and hence a2 = 3'65 - 0·32 = 3'33
where a and b are constants.
therefore a = 1·82 Suppose x has the values Xl, X2, - - - xn and y has the values
y" h, - - -Ym' Then if z = ax + by it follows that z has the
20. Charlier's checks values, mn in number, given by ax; + by} with i = 1, 2, 3, - - - n
As a check on the arithmetic involved in the calculation of x and j = 1, 2, 3, - - - m. If the means of z, x, yare denoted by
and a we can use the results , X, Y and the standard deviations by az, ax, ay respectively, it can
}:.f(d + 1) = }:.fd +}:.f be shown that
sr« + 1)2 = }:.fd + 2}:.fd2 +}:.f i = ax + by
\
CHAP. 2 SOME STATISTICAL IDEAS
1
SOME STATISTICAL IDEAS CHAP. 2
,
More generally James Bernoulli (1654-1705) showed that if the
show that the aggregate of (n! + n2) quantities has a varianc
probability of a certain event happening is p and the probability
given by
nlO'I + n20'~ + nln2(x - y)2 that it will not happen is q (so that p +
q = 1), the probabilities
that it will happen on 0, 1,2, 3, - - - n out of n occasions are given
nl + n2 (n! + nz}2
by the successive terms of the binomial expansion of (q p)n, +
15. If the variable x has the values 0, 1, 2, 3,4 and the variable namely,
has the values - 3, -1, 1, 3, find from first principles the meat n(n - 1)
and standard deviations of the function given by (i) x y, + qn + nqn-lp +1.2
qn-2p2 +
-I- pn
(ii) 2x - y, (iii) 3x +4y, and (iv) 3y - 2x. Verify the genera
Thus the probabilities of getting heads 0, 1, 2, - - - 10 times on
results quoted in Section 21.
tossing a coin 10 times are given by the successive terms of the
22. Certain special frequency distributions
The distributions of industrial and social statistics, of scientifi
expansion of (t + w o, that is,
48
49
\
CHAP. 2 SOME STATISTICAL IDEAS SOME STATISTICAL IDEAS CHAP. 2
values of p (and q) and of n. We note that in such distributio numbers of samples containing 0, 1, 2, 3, 4, 5 defective articles are
the variable n is not continuous; it can only assume positive integra 58,32, 7, 2, 1,0 respectively. Show that this distribution is approxi-
values. Further it can be proved that the mean of a binomial mately binomial and deduce the percentage of articles in the batch
distribution is np and the standard deviation is (npq)t. that are defective.
If a large batch of the products contains a fraction p that are
EXAMPLE 1
defective, the probability of choosing at random a defective article
Verify that the data given below, where f denotes the frequenc may be taken to be p. Consequently if we choose at random a
with which an event occurs n times, form a binomial distribution sample of n articles, the. probability that this sample will contain
and find the mean and standard deviation. s articles which are defective is given by the (s + l)th term of the
n 012345 expansion of (q + p)",
f 1 10 40 80 80 32 The mean of the above distribution is
Since the first and last terms of the expansion of (q + p)rI an (0 + 32 + 14 + 6 + 4 + 0)/100 = 0'56
qn and pn, if the above are distributed binomially then (Plq)5 mus If we equate this to np with n = 5 we get p = 0 ·112. Taking
equal 32 or plq = 2, that is, q = t and p = ,. for simplicity p = i the expansion of (q + p)n in this case is
Expanding (t + 1)5 we get
(~+ W = 9-5 (85 + 5 X 84 + 10 X 83 + 10 X 82 + 5 x 8 + 1)
;p + 10 + 40 + 80 + 80 + 32) 32768 + 20480 + 5120 + 640 + 40 + 1
59049
which shows that the above distribution is binomial. = 0·5549 + 0·3468 + 0'0867 + 0·0108 + 0·0007 + 0·0000.
Hence the mean equals np = 5(t) = 3t and the standard
deviation equals (npq)t = (5 x t x W = toO)t. Hence we should expect the 100samples to be distributed as follows:
To find the mean and standard deviation independently we can 55,35,9, 1,0,0, which agree well with the numbers actually found.
use an assumed mean of 3; the working is given below. The distribution is therefore approximately binomial and the
percentage of defectives in the batch is likely to be about 11 %.
n f d fd fd2
o 1 -3 - 3 9
1 10 -2 -20 40 24. The Poisson distribution
2 40 -1 -40 40 The form of the binomial distribution varies considerably
3 80 0 0 0 depending upon the values of p and n. One important practical
4 80 1 80 80 case is when p is very small, that is, the probability of the event
5 32 2 64 128 happening is very smail, but n is large, so large that np is not
sum 243 81 297 insignificant.
Itcan be shown that if p is small whilst n is large and such that
Hence the mean = 3 + 2~~ = 3t and the standard deviation np = m, the binomial expansion of (q + p)n approximates closely
to the series
= [297 _
243
(!)2J!
3
= (270)t
243
= !(1O)t
3 . m m2 m3 mn)
e-m ( 1 + T + fX2 + 1 x 2 x 3 + - - - + n!
EXAMPLE 2
A large batch of articles produced by a certain machine a where e = 2·71828 correct to 5 decimal places. This is known as
examined by taking samples of 5 articles. It is found that tb the Poisson series, and any distribution which corresponds to the
50 51
CHAP. 2 SOME STATISTICAL IDEAS SOME STATISTICAL IDEAS CHAP. 2
successive terms of this series is called a Poisson distritution. It i' hould happen rarely (p small) and that the number of trials should
so named after the French mathematician Poisson (1781-1840 he large, or strictly so large that the mean number of occurrences
who in 1837 developed the theory. appreciable.
What characterizes a distribution of the Poisson type is that t
relative frequency r with which an event happens n times vari EXAMPLE 1
according to the following table:
The following table gives the frequencies f of the number n of
n 0 1 2 3 --_ s
uccesses in a set of 500 trials. Find the mean of the distribution
emr 1 m m2/1 X 2 m3/1 X 2 x 3 - - _ nrls; md verify that the distribution is roughly of the Poisson type.
For a true probability distribution the sum of the probabilit» Verify also that the variance equals the mean approximately.
must be unity. For a Poisson distribution the sum of the relati
frequencies is n 0 1 2 3 4 5 6 7 8 9 sum
f 24 77 110 112 84 50 24 12 5 2 500
e-m 1 ( m + __m +
+_ 2
m')
+_ nf 0 77 220 336 336 250 144 84 40 18 1505
1 1x 2 s!
= e-m x e+m (approximately if s is large enough) The mean of the number of successes may be found by calculating
= 1 (approximately). the values of nf as shown in the table, whence the mean equals
To the same degree of approximation it can be shown that the mea 1505/500 = 3'01.
of the distribution is m and the standard deviation is mi, The terms of a Poisson series with this value of mare
In Fig. 9 are drawn histograms of Poisson distributions com
sponding to (a) m = 1 and (b) m = 4. e-3-0I(1 + 3·01 + 3'012/2 + 3'013/6 + - --)
On multiplying by 500 the successive terms become 24· 6, 74' 2,
emr emr 111'6, 112'0, 84'3, 50'8, 25'4, 11'0, 4'0, 1'4 which agree very
well with the values of f given in the table above.
12+ m=4 The variance is found to be 3 ·05. The working, using an assumed
m=1
o mean of 3, is given below, where d = n - 3.
8 n d [d [d2
05 6 0 -3 -72 216
4 1 -2 -154 308
2 2 -1 -110 110
0'1,
O , ' - - '=T'-
'I!
. - 3 0 0 0
012345 4 1 84 84
n 5 2 100 200
(0) (b) 6 3 72 216
Fig. 9. Poisson distributions having (a) mean = 1, (b) mean = 4. 7 4 48 192
Data conforming approximately to Poisson distributions occur 8 5 25 125
widely in science, particularly in biology. They also arise in man: 9 6 12 72
industrial problems. The indispensable conditions are that the evem sum 5 1523
52 53
SOME STATISTICAL IDEAS CHAP. 2
CHAP. 2 SOME STATISTICAL IDEAS
-co
A e-iJ2(x-m)2 dx
e
-m(1 5 ~ 1~ 6~
+9+162+2187+26244+---
) * "The role of the normal distribution in statistics is not unlike that of
the straight line in geometry. "(4)
54 55
SOME STATISTICAL IDEAS
CHAP. 2
CHAP. 2 SOME STATISTICAL IDEAS
..1
and this equals (Ay'7T)/h. If A is taken equal to h/y'7T the an
under the curve is unity and the equation of the curve then is 2(2rr1
y = _h e-h'{x-ftV
y'7T y
This is the frequency curve of what is called a normal or Gaussi
distribution. Strictly, such a distribution is one for which thi
_1
relative frequency of the observations having a value between (2rri
and x + ox is yox where y is given by equation (7). Another wa
of expressing this is to say that for a normal distribution tb
probability that an observation will lie between x and x + ox .
h
-- e-k'{x-m)2 ox
y'7T
o· x ~-I-~ x
The value of y above, equation (7), is therefore sometimes called 2
Fig.11. Normal error curves y = ~
av'(27T)
e_(x-.x')2/2a
the probability density or the relative frequency density of the
distribution.
It will be noted that any normal distribution is determined by 26. Relation between a normal and a binomial distribution
two parameters hand m. We can show that m is the mean of the It can be verified that the histogram of a binomial distribution
distribution (this is obvious from the symmetry of the curve in corresponding to (q + p)n approximates very closely, when n is
Fig. 10) and that h, sometimes called the precision constant, is large, to a normal error curve. Indeed the normal distribution can
related to the standard deviation a of the distribution by the be derived mathematically from the binomial. If the terms of the
relation 2a2h2 = 1. expansion of (q + p)" are plotted as ordinates against integral values
Writing m = x and 172 = 1/2a2 we have of x from 0 to n the points are found to lie approximately, if n is
1 large enough, on the normal curve
y = e-(x-x)2/2a2
ay'(27T) y = 1 e-(x-m)2/2a2
ay'(27T)
This then represents a normal distribution of which the mean is x
2
and the standard deviation is a. The shape of the curve for the with m equal to the mean np of the binomial distribution, and a
same value of x and for different values of a is shown in Fig. 11. equal to the variance npq.
When x = x ± a we have In Fig. 12 are drawn the histogram corresponding to (t + t)10,
and the normal curve with m = 10/3 and a2 = 20/9.
y = _1_ e-t _ 0·607 Of course there is one very important difference between the two
ay'(27T) - ay'(27T) distributions. The binomial is discontinuous consisting of a set
of discrete values corresponding to integral values of n, whereas
whilst when x = x ± 2a we have the normal distribution is continuous extending over all values of
1 -2 0·135 the variable. A distribution of discrete values may, of course,
y=--e ---- approximate closely to a normal distribution.
ay'(27T) - ay'(27T)
57
56
r
SOME STATISTICAL IDEAS CHAP. 2
CHAP. 2 SOME STATISTICAL IDEAS
y'
20 y'e59·05y
\ ___
ay(2TT)
1 e-(x-x)2/2a2 ox
Jr~ y = 1 e-(x-x)2/2a2
ay(2TT)
10 between the ordinates x and x + ox.
, \ Further, the probability that an observation will lie between two
\
\
r-j- values XI and x2 is represented by the area under the curve between
5 I
the ordinates XI and X2, and equals
bL
o
~
I 2 3 4 5678910 --- f
1
0"v'(217)
X2
Xl
e-(X-»2,2n2dx
x
Fig. 12. Histogram of binomial distribution given by G + ~rO, andi Writing (x - x)/a = 1 this expression becomes
(dotted curve)
with m = 10/3 and
the
(12
normal
= 20/9.
error curve y =
(1
1
v(2
17
) e-(x-m)2/2a2 __1
y(2TT)
J'2e-t,2dl
'1
27. The mean deviation of a normal distribution where tl = (XI - x)/a and 12 = (X2 - x)/a. It can therefore be
In Section 16 we defined the mean deviation of any set of data. expressed as the difference of two integrals of the type
If we apply this definition to a normal distribution we find that the
mean deviation, denoted by T), is given by __1 ITe-t,2dl
y(2TT)
T) = a/Y(;;TT) ~ 0·80a o
Thus for a normal distribution T) equals 4a/5 approximately. This integral has been evaluated for different values of T and is
The extent to which any other distribution departs from normality given in mathematical tables.
is sometimes estimated by calculating the ratio T)/a and comparing It might be noted that the function given by
it with 4/5.
For example, if we use the distribution shown in Table S 1
YTT JT
e-,2 dt = yTT
2 IT e-,2 dt
(Section 18) we find that :E/ldl = 592, whence
-T 0
= 592/400 = 1·48
T)
is known as the error function, because of its importance in the
But a = 1·85 and hence "J/a = 1·48/1·85 = 0·80. theory of errors, and is denoted by erf(T).
28. Area under the normal error curve T 1 Tv'2
We have stated earlier that for a normal distribution th~ t
Hence Y(~TT)l e- ,2dt = YTTI e-,2dt = terf(Ty2)
probability that an observation will lie between x and x + Ox is o 0
58 59
\
CHAP. 2 SOME STATISTICAL IDEAS
SOME STATISTICAL IDEAS CHAP. 2
This quantity is represented by the shaded area
Table 6
1 T erf (Tv'2)
Y = -_e-!12
V(217) o o
0·6745 0-5
shown in Fig. 13. The total area under this curve is unity. 1 0'6827
2 0'9543
y 3 0·9973
co 1
_1 g ± 1·96a is 0'95, so this is sometimes called the' 95% zone.
(211)2 Similarly the range x ± 3'09a is the 99'8% zone; the probability
that an observation will lie outside this zone is 1 in 500.
tions. It is clear that a perfect fit is unlikely, but it can be sho would be likely to lie between 2·341 - 0·018 and 2·341 + 0·018
that the most probable value of m is the mean of the observatio with a probability of 0·68. Further, the probability of the mean
and the most probable value of h2 is 1/2s2 where s is the standar lying between 2·341 - 0·036 and 2·341 + 0·036 is 0·95. These
deviation of the observations. We have pointed out earlier that fa probabilities vary with the value of n and hence it is important to
an infinite population of the normal type m = x and 112 = 1/2 quote the number of observations, especially if it is small, namely,
where X and a refer to that infinite population. x = 2·341 ± 0·018 (100 observations)
Of course, if we select at random" a sample of n data from Alternatively, the uncertainty in the value of a can be indicated
infinite normal population it is obvious that, in general, the mea. by using expression (8) and writing the standard error of the
of the sample will not be the same as the mean of the who] mean as 0'018(1 ± 0·07).
population; if n is large the two means may not differ very muc
EXAMPLE 1
but if n is small they may. It is possible to express mathernaticall
the manner in which the means of different samples of given si The mean height of 200 men is 1·705 ± 0·002 m and the mean
are distributed. In fact it can be shown that this distribution i height of another 300 men is 1·752 ± 0·001 m. Find the mean height
itself normal and such that its mean equals the true mean of tb of the 500 men and its standard error.
whole population, whilst its standard deviation equals a/n! when The mean height of the 500 men is
n is the number of data in each sample and a is the standac (1·705 x 2 + 1·752 x 3)/5
deviation of the whole population. Fig. 11 illustrates how th
0·01 + 0·156
distribution of the means of samples retains its normal form bu 1'70 + 5
= 1·733 m
decreases in dispersion as the size of the samples increases. I
dealing with any set of data it is therefore usual to find the me Also, the variance of their 500 heights about this new mean may
x and to write it in the form x ± ex where ex = oln» and is called be found by using relation (6) as follows.
the standard error of the mean. It should be noted that the quantity Variance of the 200 heights about the new mean
O· 6745a/v' n (cf. Section 28) which is called the probable error 01 = 200(0'002)2 + (0'028)2
the mean is sometimes quoted but the use of the standard error is
generally recommended and is now widely adopted.
= 0·0008 + 0·000 584 = 0·001 384
Similarly the standard error of the standard deviation is, Variance of the. 300 heights about the new mean
a/v'(2n), so that the standard deviation may be written as = 300(0,001)2 + (0'019)2
= 0·0003 + 0·000 361 = 0·000 661
a[ 1 ± v'~2nJ (8) Variance of the 500 heights about the mean
i: J;.(x
r-I
r - x)2/(N - 1) = £, I,. d;/(N
r-I
- 1) where N =~
n
.=1
f.. Hence the standard deviation a of the distribution
I
y = N e-(x-:I?)'/2a' 50
(Ty(217) ,
I
which' best fits a given frequency distribution by finding the m I
11. Assuming that the following distribution is approxima 18. The results of examining 213 discharge tubes and noting the
normal, find the mean and its standard error. Use (a) Besse voltage associated with a particular operating characteristic
formula, (b) Peters' forrnule. are given below. Verify that the distribution is approximately
normal. Sketch the normal curve and the histogram of the
x 25 26 27 28 29 30 31 32 33 34
1 1 5 17 49 85 52 25 11 4 1 data.
oltages 97 98 99 100 101 102 103 104 105 106 107 108 109
12. If the value of a quantity x and its standard error are given
10'23 ± 0'12, what is the probability that the accurate val Number 01
of x will lie between (i) 10· 11 and 10' 35, (ii) 10'11 and 10,4' tubes 1 8 10 25 32 42 42 27 15 4 3 3
(iii) 10'11 and 10'31, and (iv) 10'0 and 10·5. 19. Results of a count of warp breakages during the weaving of
particular lengths of cloth are given below. Show that the
13. One thousand electric lamps are tested, and it is found th
distribution is approximately Poissonian.
they have an average life of 950 burning hours with a standa
Warp breaks per length of
deviation of 150 hours. What number of lamps may be expect, cloth 0 1 2 3 4 5 6
to have a life of (a) less than 650 hours, (b) between 800 an Frequency 15 26 21 19 8 3 0
1100 hours, and (c) between 1100 and 1250 hours.
20. Tests made on electrical contacts to examine how often the
14. A sample of 20 data was found to have a mean of 80'1 wi contacts failed to interrupt the circuit due to their welding
standard deviation 2· 2. Another sample of 40 had a mean together yielded the following results.
81'4 with standard deviation 3 ,4. Assuming the two sampl
were taken from the same normal population, use the 60 da Number of welds per test 0 1 2 3 4 5
to estimate the mean of the population and its standard err, Frequency 10 13 13 8 4 2·
Show that the distribution is roughly Poissonian, and find its
15. The diameters of the spores of lycopodium can be found by aa
interference method. The results of an experiment are give variance.
below, where k = 5880 when the diameters are measured . 21. According to the kinetic theory the velocities of the molecules
centimetres. of a gas are such that the component velocities in anyone
k X diameter 14 15 16 17 18 19 20 21 22 direction are distributed normally (Maxwellian distribution).
Number 01 spores 1 1 8 24 48 58 35 16 8 Thus the probability that a molecule has a velocity-component,
measured parallel to the axis of x, lying between u and u + ou
Find the mean diameter of the spores and the standard erron is A e-hmu2ou where A, hand m are constants. Show that
Represent the results by a histogram and draw the correspon A = (hm/1T)t, and find the mean value of the velocity-component
ing normal error curve.
u.
16. Plot the frequency curve for the distribution given by 1
22. Using the result quoted in question 21, show that the probability
2/[7T(X2+ 1)] from x = - 1 to x = 1. Find approximately ~ that a molecule has a velocity lying between c and c oc +
standard deviation of this distribution by using the values of where c has components u, v, w parallel to the co-ordinate axes is
from 0 to 1 at intervals of 0·2.
(hm/1T)3/2 e-hmc2 ou ov ow
17. The median law is defined by 1= (l/2a)e-1xlla for all valu
of x. Sketch the frequency curve and show that the area und Find the mean value of c and of c2•
it is unity. Find the standard deviation. I
70 71
THEORY OF ERRORS CHAP. 3
Velocity d d2
CHAPTER 3
299782 2 4
299798 18 324
THEORY OF ERRORS 299786 6 36
299774 - 6 36
"Everybody believes in the exponential law of errors; the 299771 - 9 81
experimenters because they think it can be proved by 299776 - 4 16
mathematics; and the mathematicians because they believe
sum 7 497
it has been established by observation."(7)
A working mean of 299780 may be used. In the second column
IIIthe above table are given the values of the residuals d, relative
34. The normal or Gaussian law of error hI the working mean, and in the third column the squares, d2.
The function y = 1
Mean value = 299780 + 7/6 = 299781
aV(27T) e -(x-x)2/202
lurther a2 = 497 - 6(7/6)2 = 98
5
which defines a normal frequency distribution is often called tlu
Gaussian law of error. This "law" states that measurements of therefore a = 9·9
given quantity which are subject to accidental errors are distribute,
tnndard error of the mean = 9'9/V6 = 4
normally about the mean of the observations. More precisely, t
law infers that any set of measurements of a given quantity may velocity of light = 299 781 ± 4 km S-1 (6 observations)
regarded as a sample taken from a very large population-t
As there are only six values the uncertainty in the value of the
aggregate of all the observations that could be made if the inst
(andard error is considerable, namely, 1/10+, so that the standard
ments and time allowed-and that this population is normal.
rror might be written as 4(1 ± O' 32).
The two parameters x and a characterizing this normal distributio:
can be estimated by calculating the mean and standard deviatio:
of the given set of measurements, as explained in Section 30. Indeed 15. Applicability of the normal law of error
the mean of these measurements is the best estimate of the value 0: Some groups of observational data in science satisfy closely the
the measured quantity, whilst the standard deviation provides tb normal error law, but it is by no means universally true. Perhaps
best estimate of the accuracy so obtained. The value of the measure, this is not surprising since any theoretical derivation of the law is
quantity is usually written as x ± cc where cc = a/VI1 is the standard' bused on special assumptions which mayor may not correspond to
error of the mean and n is the number of measurements. (As we observational conditions.
have mentioned earlier some writers use not the standard error but This so-called Gaussian law of error was first deduced theoretically
the probable error of the mean which is 0'6745a/VI1, that is, hy Laplace in 1783; he started from the assumptions that the
approximately.) ,leviation of anyone of a set of observations from the mean of the
It is the resultant of a large number of very small deviations due to
EXAMPLE Independent causes, and that positive and negative deviations of
the same size are equally probable. (Cases can be quoted, however,
Certain values of the velocity of light obtained using differe
where a large number of independent causes, each giving rise to a
methods are given below in kilometres per second. Find the mea
and its standard error. mall deviation, do not result in deviations from the mean conform-
73
72
CHAP. 3 THEORY OF ERRORS THEORY OF ERRORS CHAP. 3
ing to the normal law, (8» Subsequently, Gauss gave a proof b olumn the frequency of its occurrence in a total of 300 determina-
on the postulate that the most probable value of any number lions. It will be noted that positive and negative errors are grouped
equally good observations is their arithmetic mean. together, and so we can only assume they are equally divided.
Both Laplace and Gauss discussed examples of measureme onsequently the mean of the distribution is zero.
which suggested a wide applicability of the normal law to Calculated
Limits of error Frequency f fx2 frequency
distribution of accidental errors and its truth was generally accep
But more recently its validity has been challenged on vari 0-0-0·1 114 0·285 103
grounds. In 1901,Karl Pearson showed that some series of measu 0-1- 84 1·890 85
ments which had been deemed to support the normal law sho 0-2- 53 3·312 57
significant departures from it. Later, with collaborators, he m 0·3- 24 2·940 32
six series of observations resembling those of the determination 0·4- 14 2-835 15
right ascension and declination using a transit circle, and he fou a-5- 6 1·815 6
again substantial departures from the normal law. Jeffreys'?' 0·6- 3 1·268 2
re-examined Pearson's data and confirmed the main result; 0-7- 1 0-562 o
addition he has shown that the data satisfy a law of Pearso 0-8- 1 0·722 o
type VII. 0·9- 0 0 o
.It is now generally agreed that the normal law of error is n 300
sum 300 15-629
universally valid. Nevertheless it is used widely unless there
evidence to show that some quite different probability distributi The third column of the table gives the product of the frequency
applies. It is worthy of note that even when the parent populati and the square of the deviation x. We measure x from the mean
is not normally distributed, the distribution of the means of samp to the mid-point of each class, that is, O· OS, 0 ·15,- - -,0·95.
of a given size is usually closer to a normal distribution than Hence we have
population itself. Jeffreys(6) concludes: "The normal law of e 0-2 = 15-629/299 = 0-05227
cannot be theoretically proved. Its justification is that in repres
ing many types of observations it is apparently not far wrong, Applying Sheppard's correction we get
is much more convenient to handle than others that might or 0-2 = 0-05227 - (0·1)2/12 = 0-05144
represent them better."
Of course the applicability of the normal or any other law or 0- = 0-227
be examined using the X2-test but the physicist, unlike the biolo Hence the normal error function corresponding to the data is
seldom uses this, perhaps because of the small number of observatio
with which he usually has to deal. However, certain large grou y =
300 e-x2/0-1029
of physical measurements do exist and some have been caref 0·227Y(27T)
examined. = 527 e-9·72x2
36. Normal error distributions In column four of the table above are given the calculated values
A much discussed example of observational data satisfying of f = 0 -2 y for the values of x corresponding to the mid-points
normal error law was given originally by Bessel.(lO) He provided of each class. The factor 2 is necessary to include negative as well
following data concerning the errors involved in measuring the ri as positive values of x and the factor O· 1 because of the class width,
ascension of stars. In the first column are given the magnitude The normal error curve and the histogram of the data are drawn in
the error of the observation in seconds of time, and in the seco: ig. 15.
74 75
CHAP. 3 TH EOR Y OF ERRORS THEORY OF ERRORS CHAP. 3
f !lnd J am more and more convinced that the deviations from such
70 IIcurve found so commonly in large groups of physical measure-
ments are due usually just to this cause." Several attempts to
provide suitable measurements have therefore been made. Besides
those of Karl Pearson mentioned earlier, Birge(13) made a series of
00 cross-hair settings "on a very wide but symmetrical solar
310 pectrum line, under conditions as favourable as possible to equal
reliability for all observations." His results are shown in Fig. 16
110
60
-08 ':(Yo-:04---=0-2 0 02 04 06 0 - x
Fig. 15. Histogram of Bessel's data and the corresponding normal cur so
It is clear that the agreement is very good. Hansmannt'D h
however, examined the data and shown that there are two ty .40
of Pearson curve that give even a better fit than the normal cur s'
Jeffreys(9) has also discussed the data and in particular examin
JO
how closely they follow the normal error law; he commen
"This agreement is therefore surprisingly good, indeed a little t
good, because it suggests that this series has been selected beca
it gives an unusually good agreement with the law and that otb
that may have disagreed violently with it may have been suppress
This danger of selection makes it undesirable to make use 0:
published series to test the law, unless there is some definite reaso -6
to believe that no suppression has taken place." Fig. 16. The distribution of residuals for 500 measurementsof a spectral
There are other difficulties in attempting to test the law, n line, compared with the Gaussian error curve evaluated by second
least for the physicist that observational data in physics are seldo moments. Abscissa represents residual Ii, in 0'001 mm units. Ordinate
sufficiently numerous. But even when the number of observatio representsthe number of residuals of magnitude Ii.
(Reproduced from Physical Review.)
is large they are sometimes invalidated in various ways. F,
example, Michelson, Pease and Pearson obtained 2885 values where the abscissa unit is 0·00 1 mm and the ordinate v' represents
the velocity of light from observations made on about 165 differenl the number of residuals of magnitude v. The smooth curve drawn
occasions extending over about two years. The data are sym- In Fig. 16 corresponds to the normal error curve
metrically distributed about the mean value of 299774 km S-1 bu y = _h e-h2x2
the distribution of the residuals is not normal. Birge(12)points ou VTr
that a very good fit can be obtained if one takes the sum of t
normal error curves, with standard deviations of 5 and 15 km s with h, as calculated from the observations, equal to 0 ·1965 and
respectively, which suggests, as one possibility, two groups l = 500 y. It is clear that over the whole range the observations
observations of unequal degrees of accuracy. He adds: "The poi follow the Gaussian curve very closely.
A similar series of observations, 1026 in number, have been
is that if a large assortment of observations are not of equal reliabili
made by Bond who viewed an illuminated slit using a travelling
their residuals cannot be expected to follow a normal error cur
76 77
CHAP. 3
CHAP. 3 THEOR Y OF ERRORS THEORY OF ERRORS
microscope slightly out of focus. The conditions resembled th In other words, D satisfies a normal law determined by the standard
in the measurement of a spectrum line. Bond's data were original deviation c,
given as an example of the normal error law, which they seemed Certain particular cases are specially important.
satisfy very well, but Jeffreys'P! subsequently examined them ani
found systematic departures from the normal law. (i) If D = dl + d2 we get
of each sample and the standard error of the difference of the It might be added here that in trying to assess the significance
means. of the difference of the means of two samples, especially small
samples, use may be made of what is usually called the t-distribution.
The standard error of the mean of the first sample is We have stated earlier (Section 29) that the mean x of a sample
3/y'250 = 0·19 of n taken from a normal population, of mean 111 and standard
deviation a, is such that the deviation x - m is distributed normally
The standard error of the mean of the second sample is with mean zero and standard deviation aly'n. This implies that
5/y'400 = 0·25 the ratio (x - m) : aly'n is distributed normally about zero with
unit standard deviation. However, when the standard deviation a
The standard error of the difference of the two means is estimated from a given sample of n as explained in Section 30,
the distribution of the ratio t = (x - m)y'n/a is not normal, but
it has been evaluated and is given in many books(lS) on statistics.
= ~[(y'~50)2 + (y'~)2J = ~~ = 0'31 From the table of t can be obtained, for any value of n, the
probability that a value of t will be exceeded in random sampling.
Thus if the two means have values ml and m2 the difference
the means can be written as
38. Standard error of a product
ml - m2 ± 0'31 Suppose z = x y where x and yare measured quantities of which
the means are }(and y respectively. Suppose the standard error of
The ratio 1m, - m21/0'31 is obviously of some practical impo
tance. From Table 6 (Section 2e) we note that if 1m, - m21 >3(0'31i x is ex and of y is {3.
If we write x = x + x' and y = y + y' the mean of x' is zero
it may be concluded that the two samples are not likely to belo and its standard error is IX, whilst the mean of y' is zero and its
to the same population; the probability that they so belong is 1<
than 0,2%. standard error is {3.
We have z = (x + x')(ji + y')
EXAMPLE 2 = xy + xy' + yx' + x'y'
Determinations of elm for the electron by two different grou
of methods give the following results: 1·75880 ± 0·000 42 a The mean of z, denoted by z, is therefore xy. Further, the
1·75959 ± 0·000 36, the unit being 10" C kg ' '. Do the two grou deviation of z from xy is xy' + yx' approximately, and hence
of measurement differ systematically'[ using equation (11), Section 37, the standard error of z is y where
The difference between the two mean values is
y2 = (x{3)2 + (yex)2
10-5[79 ± y'(422 + 362)]
We can write z = xy ± y
that is, 0·00079 ± 0·00055.
Thus the difference of the means is about 1·4 times its stand a and we note that (ylxy)2 = (ex/x? + ({3IYP
error, and so the existence of a systematic error is not ruled out.
Birge comments: "In view, however, of the many serious sourc ex/x is what might be called the fractional standard error of x, so
of systematic error that have been revealed, from time to time, that the above result indicates that the square of the fractional
almost every method of determining elm, I think that the sm standard error of the product equals the sum of the squares of the
remaining discrepancy should not be taken too seriously." fractional standard errors of the factors forming the product.
80 81
CHAP. 3 THEORY OF ERRORS THEOR Y OF ERRORS CHAP. 3
For example, suppose the sides of a rectangle are given as, The volume V cm3 of the cylinder is given by
20·00 ± 0· 04 cm and 10·00 ± 0· 03 ern. The area of the rectangl
can be written as 200'00 ± y ern? where V= 71(2'1)26'4 ± a
(L)2
200
= (0'04)2
20 +
(0'03)2
10 where (;)2 = (~1)2+ (1/
therefore y2 = 0·16 + 0·36 = 0'52
and al and a2 are the standard errors of r and I respectively
that is Y = 0·72
EXAMPLE 1
on t sin tA cos teA + D) - t sin teA + D) cos tA
oA = sin? tA
The radius r of a cylinder is given as 2·1 ± 0·1 em and tb
length I as 6· 4 ± 0·2 cm. Find the volume of the cylinder and itsl sintD sintD sin 23° 18·3'
standard error. 2 sin2 tA 1- cos A 1 - cos 60° 5' 2' = - 0' 79
82 83
THEORY OF ERRORS CHAP. 3
CHAP. 3 THEORY OF ERRORS
Also
on _ 1: cos -!(A + D) cos 53° 20·9' 3. Fifteen measurements of the surface tension of water have a
oD - sin-!A mean value of 0·072 52 N rn" with a standard deviation of
2 sin 30° 2·6'
0.00064. Find the standard error of the mean and its uncertainty.
= 0·597 = 0.60 Given that the accepted value of the surface tension at the tem-
1·00
perature of the laboratory is 0·073 05 N m -I show that there is
But O:A = ± 0·2' = ± 0·000058 radian likely to be a systematic error in the measurements.
4. The results of determinations of a physical quantity by two
and O:B = ± 0·4' = ± 0·000 116 radian
different methods were quoted as 15 + x units, where 100 x had
Hence the standard error IX of n is given by the following values:
0:2=(0'79 x 0·58 x 10-4)2+(0'60 x 1·16 x 10-4)2 (a) 37, 39, 38, 37, 39,40,41, 37, 36,39
(b) 40, 39, 41, 42, 40, 43, 38, 41, 40, 38.
= 0·69 x 10-8
Is there evidence that the two sets of values differ systematically
therefore 0: = ± 0·83 X 10-4
from one another?
Hence the refractive index of the glass of the prism 5. If m = 0·850 ± 0'012 find m2, m3 and 11m.
= 1·60246 ± 0·00008
6. If a = 1·16 ± 0'08 and b = 2·54 ± 0'12 find a + b, ab, alb,
(b - a)la and log. a.
EXERCISES 4 7. If v = 0·4107 ± 0·0003 m, u = 0·3513 ± 0·0002 m and
I
40. Method of least squares If all the weights are multiplied by the same constant factor
The principle of least squares, which was first formulated b the value of the weighted mean is unchanged. In the above
Legendre, may be expressed as follows: derivation the weight equals the frequency of occurrence of the
observation, but other significances are given to the weight often
the most probable value of any observed quantity is such tha
depending upon the personal assessment of the observer. Some-
the sum of the squares of the deviations of the observations fro times the weight attributed to any observation is determined, as
this value is least.
explained later, by the accuracy of the observation.
If Xl' Xl, - - -, x; are observed values of any given quantity, the:
according to the principle of least squares the most probable valu 41. Weighted mean
of this quantity is X chosen so that
Both Laplace and Gauss using different techniques established the
(xl - X)2 + (x2 - X)2 + - - - + (xn - X)2 principle of least squares mathematically. We have mentioned
is least. earlier that Gauss, assuming that the mean of the observations is
Now if x is the mean of Xl' Xl, - - -, xn so that L:xs = nx 0 the most probable value of a measured quantity, deduced the
L:(xs - x) = 0 we have normal error law. Conversely, we can use the normal error law to
L:(x, - X)2 = L:[(xs - x) + (x - X))2 deduce that the most probable value is the mean. The "proof"
= L:(xs - x)2 + n(x - X)2 proceeds as follows:
which is clearly least when X = x. The probability of the occurrence of an observation x, is given by
Thus applying the principle of least squares we derive the resul (h/V7T) e-h'(Xl-x)2
that the most probable value of a measured quantity is the arithmeti
where X is the magnitude of the measured quantity.
mean of the· observations.
Hence the probability of the occurrence of the observations
More generally, if the observations Xl' Xl, - - -, xn occur wit
Xl' X2, - - -, xm being the product of the probabilities of the
frequencies ft, h, - - -'!n respectively, then applying the princip
occurrence of each separately, equals
of least squares the most probable value of the measured quanti
is X, such that (h/V7T) n e-h'l:(x.-x)2
ft(x, - X)2 + h(xl - X)l + - - - + fn(xn - X)l if we assume that all the observations belong to the same Gaussian
is least. population specified by the precision constant h.
If now we write x = L:fsx,/L:fs, that is, the mean of the observa Thus the probability is a maximum if L:(xs - X)2 is a minimum,
tions each "weighted" according to the frequency of its occurrence, in keeping with the principle of least squares. As we have shown
we have above this occurs when x is the mean of Xl>X2, - - -, xn'
L:fs(x, - x) = 0 It may be, of course, that the measurements Xl' X2, - - -, XII
and hence L:fs(xs - X)2 = L:fs[(x, - x) + (x - X»)2 belong to different Gaussian populations; the accuracy of measure-
= L:fs(xs - x)2 + (x - X)2L:! ment may be different in each case.
We can then write the probability of the observation x. as
This is clearly least when X = s.
Thus the most probable value of the measured quantity is tb (hs/V7T) e-hHx.-x)'
weighted mean of the observations, that is, and hence the probability of the occurrence of the observations
(fix, + !2x2 + - - - + !nx,)IUi + i2 + - - - + j,) Xl' x2, - - -, Xn as
In such an expressionj', is known as the weight of the observation X h(h2 - - - hn7T-~n e-Eh~(x._x)2
86 87
CHAP. 3 THEORY OF ERRORS THEORY OF ERRORS CHAP. 3
This is greatest when L.h~(x. - x? is least, which occurs when distributed about zero. It can be shown that the standard error
is given by the weighted mean L.h~x$/L.h~. The proof is exactly of the weighted mean is given by a/(L.ws)!' that is,
given above.
If we write h~ = 1/2a; as quoted in Section 25, the most probablel
value of a measured quantity deducible from the observations [
L.w.(Xs - X?Ji
(n - 1)L.w.
XI' x2, - - -, xn is given by
If all the weights are equal this reduces to
L.h;x, _ L.xs/a;
L.h; - L.1/a~
L.(Xs -
[ (n - l)n
x?]!
This reduces, of course, to the arithmetic mean if 0'1 = a2 = in keeping with the usual formula. (Section 30.)
- - - = an'
Combining this result with that obtained above, it is clear that EXAMPLE 1
if the observations Xl X2, - - -, X" occur with frequencies J;, f2' - - -, In
or if Xs is the mean of Is observations, then the most probable value Find the most probable value of a quantity of which the following
of the measured quantity is Ire observed values: 9'4, 9'3, 9'4, 9'5, 9'6, 9'3, 9'7, 9'5,9·2,9'4.
Treating the observations as of equal weight, the most probable
L.hU.xs L.f.xs/a; L.Xs/IX; value is given by the arithmetic mean, which equals 9 + -fo (4'3) =
L.h~1s = L.1s/a; = L.1/IX; 9'43.
To find the standard error of the mean we calculate 0'2; using
where IX. = a/v Is and represents the standard error of the mean o .4 as the mean we get
(cf. Section 29) corresponding to the observations xs' Thus eac
0'2 = 0·28/9 = 0·031
observation x, can be given a weight proportional to the reciprocal 0
the square of its standard error. Alternatively, since the probabl If we used 9 ·43 as the mean we would get
error is a constant multiple of the standard error a weight pro-
portional to the reciprocal of the square of the probable error may a2 = 0·031 - (0'03)2 = 0·030
be used. The standard error of the mean = y'0'030/y'10 = 0·055.
The mean is therefore 9·43 ± O' 06, but in view of the accuracy
42. Standard error of weighted mean of the given values one might be inclined to write the mean as
9·4 ± 0·1. This would be misleading, however, especially if the
Quite generally, if the observations XI' x2, - - -, xn are given the number of observations were not quoted. It has to be remembered
weights WI> W2, - - -, w" respectively the weighted mean is that the accuracy 01 the standard error is not very high when the
x = L.w.x./L.Ws. As proved earlier, the sum L.w.(xs - X)2 is number of observations is small. For as we have stated earlier
least when X equals this weighted mean, X. (Section 30) the standard error of the standard deviation is
To estimate the accuracy of the weighted mean we find the /y'[2(n - 1)], so that the standard error of the mean, a/y'n, is
quantity a2 given by sometimes written as
0'2 =
n
2: w.(xs
s-I
- x)2/(n - 1) ~{1 +
y'n
1 }
- y'[2(n - 1)]
This is the variance with which the quantities w; (x, - x) are which in this case equals O'055(1 ± 0·24).
88 89
CHAP. 3 THEORY OF ERRORS THEORY OF ERRORS CHAP. 3
5. Two groups of determinations of elm for the electron give 10. Values of the velocity of light in kilometres per second as
following values: 1·75880 ± 0·00042 and 1,75959 ± 0·0003 corrected by Birge(12)are given below:
the unit being 1011C kg " '. Find the weighted mean of the Author Epoch Corrected result Probable error
values and its standard error. Rosa-Dorsey 1906'0 299784 10
6. Two determinations of the Faraday constant are: 96497·6 ± 6, Mercier 1923.0 299782 30
and 96506·6 ± 7·7 C mol " '. Find the difference of these tWI Michelson 1926' 5 299 798 15
Mittelstadt 1928' 0 299786 10
44. Other applications of the method of least squares; solution In determinantal notation we can write
linear equations x _ y
Legendre applied the method of least squares to the follow'
[ak][Qb]I-I[aa] [ak]1 [aa] [ab]1
problem.
Suppose we have a number of linear equations in the t l [bk ] [bb ] [ab ] [bk ] I[ab] [bb]
variables x and y of the form If different weights can be assigned to the given equations and
is the weight of the equation a.x + bsY = ks• then the normal
a.x + bsY = k,
IVs
equations become
where as. b., k, are constants. Suppose there are n equations whe
n > 2.
[waa]x + [wab]y - [wak] =0
Values of x and y can be found that satisfy any two of t [wab]x + [wbb]y - [wbk] = 0
equations. These values may not satisfy all the equations. that It is clear from these equations that if each of the original equations
the equations may not be consistent. In this case the questi is multiplied throughout by the square root of its weight the equations
arises: what are the values of x and y that satisfy all the equatio can be regarded as of equal weight.
as closely as possible?
The method of least squares can be applied to the solution
this problem. For writing EXAMPLE 1
Find the most probable values of x and y from the equations
a.x + bsY - k, = e.
2x + y = 5· I. x - y = 1'1. 4x - y = 7' 2 and x + 4 y = 5' 9
we can choose x and y such that the sum of the squares of t
We first construct the following table from the coefficients
"errors" es is least.
as. b•• k, of the given equations.
n
Thus ~ (a.x + bsY - ks)2 as b, k, a2, a.bs b~ ask, bsks
s=1 2 1 5·1 4 2 1 10·2 5 '1
must be a minimum. 1 -1 1'1 I -1 I I ·1 - 1'1
Differentiating partially with respect to x and y we get as necess 4 -1 7·2 16 -4 1 28·8 - 7·2
conditions for a minimum: 1 4 5·9 1 4 16 5'9 23·6
~a.(a.x + b.y
- k,) = 0 , sum 8 3 19·3 22 1 19 46'0 20'4
and 'i:.bs(a.x + bsY
- ks) = 0 •
Thus the normal equations are
From these two equations the values of x and y can be found. 22x + y = 46'0
These are the.i'rnost probable" values. x + 19y = 20'4
The above equations are usually written in the form . 874·0 - 20'4 853·6
[aa]x + [ably - [ak] =0 (1 Solving we get x = 418 _ 1 = 417- = 2·047
[ab]x + [bb]y - [bk] = 0 (1 448·8 - 46·0 402·8
n
and y = "n • = 417 = 0·966
where [ab] denotes the summation ~ a.b..
s=1 The most probable values are x = 2·05 and y = 0·97. With more
They are known as the normal equations. complicated examples certain arithmetical checks are introduced.
96 97
CHAP. 3 TH EOR Y OF ERRORS THEORY OF ERRORS CHAP. 3
Find the most probable values of x and y from the equatio del + de2 + de) = 0 (19)
Ws as b, ks wa3 wa.b, wb2 wa.k, Aldel + (AI + Avde2 + (AI + A2)de3 + A2de4 = 0
1 2 1 5 ·1 4 2 1 10·2 Comparing this with equation (18) we have
3 1 -1 1. 1 3 -3 3 3'3
2 4 -1 7·2 32 -8 2 57'6 el _ e2 _ e3 __ e4
sum 39 -9 6 71 '1 AI - Al + A2 - AI + A2 - A~
Thus the normal equations are so that e3 = e2 and e4 = e2 - el
- --
~·,lk:·,lAt - 1Cl -:
r~~'i;·
CHAP_ 3 THEORY OF ERRORS THEOR Y OF ERRORS CHAP_ 3
Then, as before, the 1)10stprobable values of x and yare tb We can write x = 2-05 ± 0-01 and y = 0-97 ± 0-01.
solutions of the normal equations As a further example, let us suppose that by direct observation
it has been found independently that
[aa]x + [ablY = [ak]
[ab]x + [bb]y = [bk] x = 1-00 ± 0-10 and y = 0-90 ± 0-07
Gauss(17) and others have discussed the problem of estimating the Further, suppose that a third independent observation leads to
weights that may be assigned to the values of x and y so obtained. x + 2y = 3-00 ± 0'07
Denoting these values by Xo and Yo and writing a"xo + b,yo - k., ""
d., that is, d1 d2, - - -, d; are the residuals when the most probabl What are the most probable values of x and y?
values Xo and Yo are substituted in the given equations, it has been. We will give the equations weights inversely proportional to the
shown that the standard error a to be expected in any expressio squares of their standard errors, namely, 1, 2, 2_ So multiplying
asxo + bsyo - k, is given by" the equations by I, y2, y2 respectively we get
a2 = [dd]/(n - 2) x = 1-00, yy2 = 0-90y2 and xy2 + 2yy2 = 3 -OOY2
Also if ax and ay denote the standard errors in each of which can now be regarded as having unit weight. We
respectively, then it can be shown that therefore have
a 2 a 2 a2 a. bs ks a2 b2s ashs ask,. bsks
.:»: =~=_,
[bb ] [aa ] t" 1 o 1-00 1 o o 1-00 o
o y2 0-9OV2 o 2 o o 1-80
where t" is the determinant y2 2y2 3-00y2 2 8 4 6-00 12-00
[aa] [ab]1 sum 3 10 4 7-00 13-80
1 [ab] [bb]
The normal equations are
If, as an example, we consider the equations solved in Example I, 3x + 4y = 7-00
Section 44, namely,
4x + 10y = 13-80
2x +y = 5 -1, x - y = I-I, 4x - y = 7 -2 and x + 4y = 5 -9
leading to x = 1,057 and y = 0-957_
the normal equations are Using x = 1-06 and y = 0-96 we get
22x +y = 46-0 [dd] = 0-0036 + 0-0072 + 0-0008 = 0-011
x + 19y = 20-4
Hence a~/l0 = a;/3 = 0-0116/14
leading to x = 2-05 and y = 0-97_
The residuals are therefore -0-03, -0-02, 0-03 and 0-03 sa leading to ax = 0-09 and ay = 0-05
that [Jd] = 0-0031 and a2 = 0-00155_ Hence x = 1-06 ± 0-09 and y = 0-96 ± 0-05
Hence
a~ _ a;_ 0-00155
19 - 22 - ----c 46. Curve fitting
4 17=---
C"7C
given by ax. + b - y.. Then the mean square error <x2 in An alternative method has been given by Bond.(lS) If we write
expressions ax. + b - Ys is given by the relation y = ax + b in the form Y = aX where Y = y - yand
<x2 = [dd]/(n - 2)
X = x - x we have shown above that a = [XY]/[XX].
Hence if the values of Xs are accurate and the standard error of the
Also, if <Xa and <Xb denote the standard errors in the values values of y, (and therefore of the values of Ys) is IX we can write
a and b we have, using the normal equations (21) and (22). formally
a = [X(Y ± <x)]/[XX]
a.~/n = <xU[xx] = <x2/tJ.
which, using standard error (i) of Section 39, leads to
where tJ. is the determinant
[XYJ ± <xy[XX] [XY] IX
[xx] [x]1 = n[xx] - [x)2 a= [XX] =[XX]±y'[XX]
I [x] n
Therefore the standard error of a is lXa = IX/Y[XX]. This result
In the example considered above the normal equations are
is in keeping with that given above since
285a +45b = 869·4
[XX] [xx] -2
45b +
lOb = 150'7 --=-
n n
-x
leading to a = 2·32 and b = 4'64.
that is, n[XX] = n[xx] - [xJ2
Hence we have = <x~/285=
<x~/1O 1X2/825
IX~ _ 1X2
where 1X2 = [dd]/8 and [dd] = 0,3500 as shown below.
and hence n - 'n[:-xx·]---r[x"F
x y d dd Further, since b = Y -- a,r and the standard error of y is <x/v'n
4·6
0 0'04 0'0016 we can write
1 7 ·1 -0,14 0·0196 b = y ± <x/yn - (a ± IXJX
2 9·5 -0,22 0·0484
3 11'5 0·10 0'0100 or b =y - ax ± ,J (:2 + <x~2)
4 13,7 0,22 0,0484
5 15·9 0·34 0·1156 <x2
6 18·6 -0'04 0·0016 so that <xl = - + 1X~2
n
7 20·9 -0,02 0·0004
8 23'5 -0'30 0'0900 =~ + IXUX)2
9 25'4 0'12 0·0144 n n2
sum 45 150'7 0,3500
=<x2 [xx]
Hence n[xx] - [x)2
lXa = 0·023 and <Xb = 0,123
as given above.
We can therefore write It is perhaps more convenient to write the equation in the form
y = (2'32 ±0'02)x + 4·64 ± 0'12 Y ± IX/yn = (a ± lXa)X
106 107
CHAP. 3 THEORY OF ERRORS THEORY OF ERRORS CHAP. 3
where therefore Y = (2'32 ± 0·02)x + 4·66 ± v[(0'07)2 + (0'09)2]
y = y - y, X = x - X, eeo = ee/v[XX] and ee2 = [DD]/(n - 2), = (2'32 ± 0'02)x + 4'66 ± 0'12
D. being the value of Y. - aX s: in keeping with the result found earlier.
Thus in the example considered above we have found tha It might be added that if the observed values x., Ys are given the
a = [XY]/ [XX] = 191.25/82'50 =2· 32, and hence as shown belo
weight w. the equations (21) and (22) above become
[DD] = 0'3580.
a[wxx] + b[wx] = [wxy]
X y D DD a [w,x] + blw] = [wy]
-4'5 -10'5 -0'06 0'0036 a = [w](wxy] - [wx](wy]
-3'5 - 8·0 0'12 0·0144 so that
[w][wxx] - [wxF
-2'5 - 5·6 0'20 0'0400
-1'5 - 3·6 -0'12 0·0144 b = [wy][wxx] - [wxH.wxy]
and
-0,5 - 1·4 -0'24 0·0576 [w] [wxx] - l wx F
0'5 0·8 -0'36 O'lf96
1'5 0·02 0·0004 Also the standard errors of a and b are given by
3'5
2'5 5·8 0 0 ee2 ee2
......!!...=_b_= ee2 ._
3'5 8·4 0·28 0'0784 [w] [wxx] [w][wxx] - [wxP
4'5 10·3 -0,14 0·0196
where ee2 = [wdd]/(n - 2).
sum 0·3580
Hence 49. Other curves
It often happens that the relation between the two variables x
ee2 = O'3580/8 and y is not linear. More generally, the relation may be of the form
therefore y = ao + alx + a2x2 I ••• + Gmx'"
ee~ = O'3580/(8 x 82' 50) involving m + I constants.
= 0'00054 If n corresponding values of x and yare known, say, (xs' yJ with
therefore s = I, 2, 3, •• " nand n> m + I, the values of the constants
eeo = 0·023 as before. ao, at> a2, - - -, am may be chosen such that the sum of the squares
Also, given by
n
....::..= /0' 3580 ~ (y. - ao - alx. - a2x; - • - - amx,;,)2
Hence the relati . Vn 'Y 80""" = 0·067 is least.
s ~ I
won IS
A particular example will illustrate the method.
Y ± 0'07 = (2'32 ± 0'02)X If y = ao + alx + a2x2 we choose ao, ai' a2 so that
Differentiating partially with respect to ao, ai' a2 respectively we We assume that y = ao + a, X + a2X2, where to simplify the
obtain the necessary conditions: arithmetic we take X = 5x - 3 and tabulate the working as follows.
0
I -
1
0
I
I
16
1
0
1
-4·8
-6,6
0
25·7
9·6
6·6
0
2·22 0'18
6·69 -0,09
14'46 -0'26
25'7 25·54 0·16
0'0324
0'()()81
0·0676
0·0256
2 4{)·1 4 8 16 80·2 160·4 39·93 0·17 0'0289
[y] - nao - [x]a, - [xx]a2 = 0
3 57·5 9 27 81 172·5 517'5 57·62 -0'12 0·0144
[xy] - [x]ao - [xx]a, - [xxx]a2 = 0
i sum o 147'5 28 0 196 264·0 728·8 0'1795
[xxy] - [xx]ao - [xxx]al - [xxxx]a2 =0
Hence the normal equations are
the solutions of which give the most probable values of ao, a, and a2'
Again, if we assume that the values of Xs are accurate and the 7ao + 28a2 = 147' 5
values of Ys are subject to experimental error, the standard errors 28al = 264·0
of the values of ao, aJ> a2 can be estimated by the method outlined 28ao + 196a2 = 728·8
in Section 45 above.
When the number of constants is large, the solution of the norma leading to ao = 14'463, al 9'428 and a2 = 1·652.
equations may be laborious. Special methods of solution have bee Hence we have
devised.U''- 20)
y = 14'4631 9'428X + l'652X2
The normal equations might more conveniently be written = 14·463 I 9'428(5x - 3) + I .652(5x - 3)2
soao + sJal + S2a2 = to . = 1'05 - 2'42x + 41'3x2
Siao + S2a, + S3a2 = tl
The uncertainties in the values of 00, ai' a2 can be estimated as
S2aO + S3a, + S4a2 = t2 follows. We first find the residuals d given by y - Go - a, X - a2X2
n n for the various values of X; these are tabulated above. We then
where Sk = ~ x~ and tk = ~ x~ Y, calculate d2 and find
,-I r=1
a2 = [dd]/(n - 3) = 0'1795/4
The general form of these equations is obvious.
If the standard errors of ao, ai' a2 are denoted by ao, aJ, a2
respectively, we have
EXAMPLE
ao2 aI _ a~ _ a2
Fit a parabolic curve to the following data: 7 ~1-1701-1-=-7 ---'0~28"
28 01
x o 0'2 0'4 0·6 0·8 1 ·0 1·2 1 o 196 1 28 196 0 28 0 28 0
y 1-0 2'4 6·6 14·2 25·7 40·1 57·5 28 0 196
110 11 I
CHAP. 3 THEORY OF ERRORS THEORY OF ERRORS CHAP. 3
(X~ _ (Xi _ (X~ _ 0'1795/4 5. Find the most probable position of the point of which the
therefore
112 - I2 - 4 - 336 measured distances from the points (1, 0), (3, 1) and (- I, 2)
therefore (xo = ± 0'122, (Xl = ± 0'040, (X2 = ± 0'023 are respectively 3 '1, 2·2 and 3' 2 units. Estimate the uncer-
tainties in the co-ordinates of the point.
Hence
y = 14·463 ± 0·122 + (9'428 ± 0'040)X + (1'652 ± 0'023)X2 6. The three angles of a plane triangle are measured and found
to be
The simplification of this expression is complex, since (Xo, (XI' (X2
are not independent. If, however, we treat them as independent we A = 48° 5' 10", B = 60° 25' 24" and C = 70° 42' 7".
get
y = 1·05 ± 0·27 - (2·42 ± O'72)x (41·3 ± 0·6)x2 + Find the most probable values of A, Band C assuming that
the measurements have (i) equal weights, and (ii) weights 1, 2, 3
The standard error 0·27 is overestimated, but the others are respectively.
satisfactory.
7. Plot the values of x and y given in the example discussed in
Section 46 and draw the best straight line through them. Find
EXERCISES 6 the equation of the line drawn.
1. From the equations Find the values of a and b using only the values (i) x = 0, 1,
3x+y=2'9, x-2y=0'9 and 2x-3y=i'9 2, 3, 4, 5 and the corresponding values of y, and (ii) x = 5, 6,
7, 8, 9 and the corresponding values of y.
find the most probable values of x and y.
8. Two quantities D and d are measured as follows:
2. The equations
x-3y=-5'6 d (in.) ! i I Ii 11 Ii 2
4x + y = 8'1 D (in.) 1'19 1·31 1'42 1'52 1·64 1'76 1·87
and 2x - y = 0·5
Show that they satisfy approximately a relation of the form
have weights 1, 2, 3 respectively. Find the most probable D = ad +
b, find the most probable values of a and b, and
values of x and y, and their standard errors. estimate the errors in those values.
3. Independent sets of observations led to the results 9. Fit a linear relation to the following data and estimate the errors
u = 1·23
0'06,± v = 2·17 ± 0·08 in the values of the constants obtained.
and u + v = 3· 50 ± O' 12.
x 10 12 13 17 19 20
Find the most probable values of u and v and their standard
errors.
v 11'0 7·6 6·2 -0,1 -3'2 -5'0
4. Find the most probable values of x, y and z that satisfy the 10. Values of the surface tension of water, y N m -1, at different
equations temperatures, to C, are given below. If y = a - be, where e is
x+y+z=4'01, 2x-y+z=l'04, the temperature on the Kelvin scale, find the most probable values
x +
3 y - 2z = 5' 02 and 3x y = 4· 97 + of a and b.
assuming the equations have equal weight. Find also the t 10 20 30 40 50 60
standard errors of x, y and z. Y X 103 74·22 72-75 71-18 69'56 67-91 66·18
112 113
CHAP. 3 THEORY OF ERRORS
c = 299777·27 ± 2·03 - (0'381 ± 0'234)(D - 1930) (5) RUTHERFORD, E., and GEIGER, H. Phil. Mag., 20, pp. 698-707
(1910).
May it be concluded that there is experimental evidence of
linear variation of c with time? (6) JEFFREYS,H. Theory 0/ Probability (Oxford: Clarendon Press,
1948).
13. Fit a parabolic curve to the data given in the example discussed
in Section 49, but excluding the value x = 1· 2. Use the (7) WmITAKER, E. T., and ROBINSON,G. Calculus a/Observations,
calculated expression for' y to find y when x = 1·2. p. 179 (London: Blackie and Son Ltd., 1929).
14. Values of the viscosity of water, 1'/ Ns m ->, at different. temper (8) See ref. (7), p. 177.
atures, to C, are given below. Fit a law of the form (9) JEfFREYS, H. Phil. Trons A, 237, pp. 23/-271 (1938).
1-1 = a bt +
ct2• + (10) BESSEL,F. W. Astron. Nachr .• IS, Nr. 358-359.
t 10 20 30 40 50 60 70
1'/ X 103
1·308 1·005 0·801 0·656 0·549 0·469 0'406 (1 I) HANSMANN,G. H. Biometrika. 26, pp. 128-195 (1934).
15. Assuming that TJ = Aek/T where T is the temperature on the (12) BeRGE, R. T. Rep. Progr. Phys., 8, p. 95 (1941).
Kelvin scale, use the data given in question J 4 above to find (13) BIRGE, R. T. Phys. Rev., 40. pp. 207--227 (1932).
the most probable values of A and k.
(14) See ref. (7), p. 175.
Mean 31
assumed or working 32 Random 62
deviation 38. 58 Range 38
of a sum 44 Ratio ?)/a 58
standard error of 62, 64, 72 Regression, line of 104
weighted 87 Relative frequency 33
Median 33 Residual 13
Median distribution 67, 70 Root mean square deviation 40
Method of least squares 86,96,101 Rutherford, E. 69, 115
Milne, W. E. 110,115
Multinomial law 67
Sample 61
"Noise" 10 Sampling, random 62
Normal distribution 55, 56, 57 Scatter, 13, 37
fitting of 66 Scrase, F. J. 54, 1J 5
Normal equations 96 Sheppard's corrections 43
Normal error Skew 67
curve 55 Solution of linear equations 96
distribution 74 Standard deviation 38, 40
law 72 evaluation of 42
applicability of 73 of a sum 44
Standard error
of compound quantities 82
Pearson, K. 74 of mean 62, 64, 72
Pearson types of curve 67, 74 of product 81
Personal of standard deviation 64
equation 10 of sum or difference 78
error 10, 14 of weighted mean 88
118 119