Bartlett General Principles of The Method of Least Squares, With Applications (1915) PDF

ta
H. M.
Stewart, Esq
GENERAL PRINCIPLES
OF THK
METHOD OF LEAST SQUARES,

WITH APPLICATIONS,
BY
DANA
P.
BARTLETT,
S.B.,
PROFESSOR or MATHEMATICS. MASSACHUSETTS INBTTTUTB

or TECBNOLOOT.
THIRD EDITION.
BOSTON
THE AUTHOR
1915.
COPYRIGHT,
BY DANA
P.
1915.
BARTLETT.
Ol
TECHNOLOGY BRANCH
HARVARD COOPERATIVE SOCIETY
76 MAHBAQHUSETTH
AVENUE, CAMBRIDGE,
LfSt
PREFACE.
The preparation
of this
volume was undertaken with the
view of presenting in as simple and concise a manner an

possible the fundamental principles of the Method of Least
Squares.
While
the solution of
all
attempt has been
it is
believed that everything essential to
ordinary problems has been included, no
made
to develop at length those special
methods and forms that are so useful and almost necessary in

case large numbers of observations of certain kinds, such,
for instance, as those
met with
and astronomical
in geodetic
measurements, are to be adjusted.
Frequent references throughout the text, and more particularly the list 01 works given on page v of the Appendix, will,
however, enable the student to extend his studies in whatever special direction his profession
expected that this
book
may
require
will in such cases
it
being
be looked upon
;
merely as an introductory treatise. All of the works mentioned have been freely consulted in the preparation of these
pages, and the author desires in particular to acknowledge
his indebtedness for
many
of the examples.
DANA
P. BARTLETT.
CONTENTS.
CHAPTER
I.
GENERAL PRINCIPLES.
PAGES
3. Errors.
Object of the Method of Least Squares.
4. Constant Errors
Theoretical, Instrumental, Personal.
5. Mistakes.
6. Accidental Errors.
7. Direct Observa8. Real Errors.
tions The Arithmetical Mean.
9. Resid11. Weighted Observations.
13. The General Mean.
uals.
16. The Curve of Error.
17. Laws of Errors of Observa20. Derivation of the Equation of the Curve of Error.
tion.
22. The Method of Least Squares
1-16
1.
CHAPTER
II.
THE ADJUSTMENT OF OBSERVATIONS.

25, 27. Rules for Forming the
28. Reduction of Equations to Weight
23. Indirect Observations.
Normal Equations.
Unity.
and
its
tions.
tions.
29. Relation Between the Weight of an Observation

Measure of Precision.
30. Computation of Correc32. Significant Figures.
33. Conditioned Observa35. Special Cases.
36-43. Empirical Formulas and
Constants.
mic Solution.
42. Periodic Phenomena.

43. The Logarith44. Reduction of Equations to the Linear
Form
17-36
CHAPTER
III.
THE PRECISION OF OBSERVATIONS.

The Constant k.
49. The Value of k in Terms of h.
61. The Mean of the Errors, or Average Deviation.
52. The
Mean Error.
53. The Probable Error.
55, 56. The Rela-
48.
tions between
//,
a.d. t
58. Representation of
h, p, and p.
fi, r, a.d.
and r on the Curve of Error.
37-46
t
CONTENTS.
CHAPTER IV.
COMPUTATION OF THE PRECISION MEASURES.

PAGES
62. Direct
59-61. Direct Observations all of the Same Weight.
64-71. FuncObservations, the Weights Not Being All Alike.
72. The Precitions of Independent Observed Quantities.
74. Functions of the Same Varision of Measurements.
76. First Method of
76-85. Indirect Observations.
78. Second Method
77. Rule I.
Computing the Weights.
80. Third Method
79. Rule II.
of Computing the Weights.
82. The Mean
of Computing the Weights.
81. Rule III.
Error of an Observation.
85. Observations of Unequal
ables.
C6-89. Conditioned Observations
Weights.
47-82
CHAPTER V.
MISCELLANEOUS THEOREMS.
90.
The Distribution of Errors.
92.
The Rejection of Obser-
93, Criterion for the Rejection of a Single Doubtful

95. The Huge Error.
96. Constant Errors.
Observation.
vations.
98. Combination of Determinations having Different Con100. The Weighting of Observations.

Errors.
stant
101-103. Special Laws of Error.
104. Contradictory Obser-
vations
83-96
CHAPTER VI.
GAUSS'S
METHOD OF SUBSTITUTION.
107. Checks on the Formation of the Normal Equations.

108. The Reduced Normal Equations and the Elimination
109. Checks oil the Solution of the Normal
Equations.
110. Most Convenient Arrangement of the ComEquations.
113. Solution
111. Application of the Checks.
putations.
of the Elimination Equations.
115. The Weights of the
Unknown Quantities
METHOD OF COHHKLATIVKS
97-111
GAUSS'S
111-116
EXAMPLES
117-142
AlM'KNI'l.X.
TIIK
THEOUY OK PROBABILITY.
200. Definition; Simple Events.

204. DriH-mlfiJt Event*
Blin.MM.HAlMIV
TABLES.
202.
Compound
Events.
i-v
V-Vl
vli-xi
THE METHOD OF LEAST SQUARES.

CHAPTER
I.
GENERAL PRINCIPLES.
In scientific investigations of all kinds it is frequently
necessary to determine the values of certain quantities by
1.
means
of actual
of instruments.
measurements either with or without the aid

The observations may be made directly upon
unknown quantities or upon certain functions

unknowns. In the latter case the values of the required
quantities must be obtained by computation from the observed
In order to obtain more accurate
values of the functions.
values of the unknowns than would be given by a single
the values of the

of the
measurement, or set of measurements, the observations are

usually repeated either in the same way and under the same
conditions or in a variety of different
ing conditions.
ways and under vary-
Under these circumstances it will invariably be found that

the different measurements give discordant results, the amount
of the discrepancies varying with the character of the observations
and the question that now presents
itself
is
how
to
determine from these discordant observations the true values
From the nature of the case,

we
can
not
to
obtain our values with absolute
however,
expect
all
can
that
we
hope for is to obtain those values
accuracy;
of the required quantities.
which are rendered most probable
after all the observations
are taken into account, and, further, to determine the degree

of confidence that can be placed in those values.
MI-.TIIOD
2
2.
The attainment
OF LEAST SQUARES.
of
the above results constitutes the
primary object of the Method of Least Squares. The method

is also employed in comparing the relative worth of different
measurements of the same quantity, and in determining the

equation of a curve which shall suitably represent the relation
between two variables in cases where the exact law connecting
them is not known.
making any observations, we may employ the

method to determine how precise the component measurements
of a series must be in order to yield a required degree of precision in the final result; or, conversely, to determine what
Also, before
the precision of the final result will be, knowing the precision
attainable in the component measurements. This latter application of the method will be treated at length in the course
on " The Precision of Measurements."
The cause
of the discrepancies between the

results of our different observations is that every observation
Errors.
3.
that
is
a measure
is
subject to error.
These errors are of two
Constant or Systematic Errors and Accidental Errors.

4.
Constant Errors are errors which in all measures of the
same quantity, made with the same care and under the same
conditions, have the same magnitude, or whose presence and
kinds,
magnitude are due to some fixed cause. These constant

errors may be of several classes, which are designated as
follows:
First.
Theoretical Errors, such as those due to the refrac-
tion or aberration of light, the effect of a definite change in

temperature or moisture on our standards of measurement,
etc.
As soon
errors
known the magnitude of

be calculated and their effect eliminated
as their causes are
may
from the observations.

Second. Instrumental Errors, such as errors of division of
graduated scales, defects in micrometer screws, eccentricity of
Thrsc errors will he dUrovcrnl l.y ;ui r\;unin:irirrh-s, j-tc.
tion of the instruments ami tln-ir rfl'rrts rliminutril t'rom tin*
&SNBRAL miXCIPLES.
observations, either by a particular method

instruments or by subsequent computation.
using the
These are due to personal pecuof an observer, who always answers a signal too soon
Third.
liarities
of
Personal Errors.
or too late, always estimates a quantity smaller than it is, etc.

The character and magnitude of these errors may be deter-
mined by a study of the observer, his "Personal Equation"

may be obtained, and his observations thus corrected for thin
source of error.
Although of a somewhat different character,

be
considered in connection with constant errors.
these should
made
when a figure 3 is read for a figure 8, or
is
A mistake
5.
when
Mistakes.
in reading a graduated circle
which
is
numbered
in
both
read 43 instead of the complementary

directions the angle
These
mistakes are usually of such a characangle 47, etc.
detected
be
ter that they may
by an inspection of the observais
tions
and a proper correction made.
Accidental Errors are errors due to irregular causes,

whose effect upon the observations is not determined by any
6.
circumstances peculiar to that particular set of measurements,

and which cannot therefore be computed and allowed for
Such errors are those due to sudden changes in

owing to sudden and unobserved changes in tem-
beforehand.
refraction
perature; unequal expansion of different parts of an instru-
shaking of an instrument
ment with change
in temperature;
in the wind, etc.
But most important
of all are those errors
from imperfections in the sight, hearing, and other

senses of the observer, which render it impossible for him to
adjust and use his instruments with absolute accuracy.
which
arise
After a
full investigation of the
constant errors, the observer

much as possible, both
should diminish the accidental errors as

in
number and magnitude, by taking every precaution and care

measurements themselves. The problem now remains
in the
to
combine the observations so that the remaining accidental

upon the results,
errors shall have the least probable effect
METHOD OF LEAST SQUARES.
and
it is
to bring about this combination of observations that

the Method of Least Squares.
we employ
When no more
observations are made than are sufficient to

determine one value for each of the unknown quantities, we
must accept these values as the most probable

additional observations are
we can not
But
ones.
if
made leading
take any one of them
to discordant results,
as the correct value, and in
already stated, we shall probably not be able to obtain

the true values of the unknowns. All that we can do is to find
fact, as
values of the unknowns which shall remove the discrepancies

between the different observations and which shall be those
values that are rendered most probable by the existence of
the observations themselves.
On first thoughts it may seem that these accidental errors,

being due to so many different and unknown causes, will be
beyond the scope of mathematical investigation. Neverthethe theory of probability requires that these errors shall
follow in magnitude and frequency a law that is capable of
less,
exact mathematical expression, and experience confirms the

correctness of this law.
For more extended remarks on these subjects
see
Holman, " Discussion of the Precision of Measurements," pp. 1-14.

" Text-Book of Least
Merriman,
Squares," pp. 1-6.
"
Chauvcnet,
Spherical and Practical Astronomy," pp. 469-473.
Wright, "Treatise on the Adjustment of Observations," pp. 11-18.
LAWS OF ERRORS OF OBSERVATION.

7.
The
derivation of the general laws of the occurrence of

and of the processes for determining
errors of observation,
the most probable values of the unknown quantities, will l>e

based upon the following
.Vn
Axiom. If a series of n direct observations, J/i, M^
U made upon the value of a quantity M, all the observations
b-in: made with tin- Kirne care and under the same eircmii.
GENERAL PRINCIPLES.
stances, the
most probable value
arithmetical
mean
of that quantity
of the observations.
-2T
MQ =-
.^
is
the
Or
= ~-.
(1)
The Eeal Error (x) of an observation is the difference

8.
between the observed value of the measured quantity and the
real value.
The Residual (v) of an observation is the difference

9.
between the observed value of the measured quantity and
the value rendered most probable by the existence of the
observations.
10.
Example.
made upon
Eight observations are
the
a coil of wire, the true resistance being 512.

Find from these observations the most probable resistance,
resistance of
and
also the real errors

Observations.
and
residuals.
Real Errors.
Residuals.
+
+
-
+
+
-
+.30
+.10
.4
.2
.1
.20
+ .20
- .30
+ .20
.3
.2
.3
.1
.0
.20
.10
From
the observations, then, we should say that the most

probable resistance of the coil is 512.10. It will also be
noticed that the
sum
of the residuals
is
That
zero.
this is a
general result following from the assumption of the arithmetical mean as the most probable value may be proved as
follows:
If
the
observations are
M^ M^
n,
the
6
arithmetical
mean
and the residuals v l9 u 2
v w , then
we have
.
since
Sv
.-.
M=
(2)
The weight of an observation expresses its relative worth compared with other observations.
Thus, if six observations are made upon the value
11.
Weighted Observations.
of a quantity, five of which give the same result, while the

sixth differs, in combining these two different results to
obtain the most probable value of the unknown, the first

value ought to have five times the influence upon the final
result that the second has, since it has taken five times as
much
labor and time to obtain
it.
Hence
in general
we may
say
12.
The Weight (p) of an observation may be considered

number of times the observation has been
as representing the
repeated and the same result obtained.

The weights assigned to observations
may be due
to a
variety of causes, as difference in skill of observers, difference

in the instruments used or the circumstances under which
But whatever the cause, the

the observations are made, etc.
of
weighting an observation will be
effect on the final values
tin-
Karne as indicated in the preceding paragraph.
M
M
t
n
Suppose n observations, J/j,
are made upon the value of a
of weights />!, 7> 2
/>n
of the
To find the most probable value
quantity M.
13.
Example.
quantity.
From the above interpretation of the meaning of weight,
we may consider that the whole number of observations is
/> i+/> 2+
Pw
or 2j, and that the result 3/j has
PRINCIPLES.
i;i-:\i-:i!AL
obtained in
observations,
in
observations,
etc.,
Therefore, by (1)
ffiJ
3f
is
called the General
If the residuals are v

Vi
=M
MQ,
v2,
Mean.
.
= Jf
v2
we have
v n,
3f
vn
=M
Which shows
(4)
that in the case of direct observations of differ-
ent weights the sum of the weighted residuals is zero.

14. If the observations are not made directly upon the
required quantities, the method of adjusting
the results so as to obtain the best possible values of the
unknowns will depend upon the laws which govern the disvalues of the
tribution of the errors of these observations.
It is
found in
practice that the accidental errors of observations follow certain well defined laws, and what these are may best be seen
by taking an actual example.

One thousand shots are fired at a target
15. Example.
which is divided into a number of horizontal sections by lines
one foot apart, the centre line of the target being in the
The
middle of one of these spaces.

as follows
In Space.
to
"
"
shots were distributed
In Space.
Shots.
In Space.
44
4.14104-4
190
24
to
3J
79
+34
-34
"
-44
16
-j-24
10
54
+14
89
212
204
193
Shots.
" --
i
"
4
14
-14
-24
44
"
Shot*.
Tn this case the errors are evidently the distances of the

shots from the centre of the target.
Further, as far as can
be judged from these one thousand shots,

fired the
lines
is
if
another shot
probability that this shot will fall between

.001
.004
-H
" .010
" .089
3*
+
-
i and -liis.204
2
-2*
-3*
-4*
is
the
-3$
4J
-5
.193
" .079
" .016
" .002
" .212
The sum of the above probabilities is unity, and, therefore,

as far as the preceding shots show the 1001st shot will certainly hit the target.
Now
16.
lines
using as abscissas the distances of the horizontal
from the centre of the
target,
and as ordinates the num-
ber of shots falling in the corresponding spaces,

construct the following figure
we may
Figure
r
And
1.
o
if
the entire area of this figure
the area of each
is
taken as unity thru
of a
rectangle will <lnnte the probability
GENERAL PRINCIPLES.
shot, if fired, falling within the
corresponding space of the
target.
The graphical representation of the accidental errors of

observation will always give a figure similar to the above.
Hence denoting errors by abscissas, and their frequency by
ordinates, the law of error of any series of observations may
be represented by a curve whose general form is determined
"
by Figure 1. This curve is called the Curve of Error," and
is
shown
in
Figure
2.
Figure 2.
PDM
In order that this curve may represent exactly the distribution of the errors in any given series of observations it
at some definite distance to the
ought to meet the axis of
right and left of the origin and coincide with the axis from
there on, for in all actual observations there is a limit beyond
which no errors occur. But as the exact point of meeting
could not be determined for any given case, and as it would

not be possible to obtain the equation of such a curve, we
make
it
asymptotic to the axis of
A",
t-iking care that the
error thus introduced shall in any set of observations be so

small as to be negligible.
10
17.
some
An inspection of Figures 1 and 2 will now exhibit

of the general lawt of errors of observation and the
corresponding properties of the curve of error.
Laws of Error
derivfd
from an
inspection of Figure 1.
Representation of these laws by the

Curve of Error.
The maximum point of
Small errors are more
First.
the
frequent than large ones.

Second. Positive and negative
errors of the same absolute mag-
on the axis of Y.
The curve is symmetrical with
respect to the axis of F.
nitude are equally likely to occur.

The probability of the
Third
occurrence of very large errors is
axis of
very small.
Fourth. The frequency of any
error depends upon the magnitude
The equation of the curve will

be of the form
curve
is
The curve
X.
of that error.
is
asymptotic to the
y=*(0
'
(6
now, the total area between the curve and the axis
denoted by unity the probability that the error of
any given observation will fall between the magnitudes x and
x-\- dx will be represented by the area included between the
18.
of
If,
X be
curve, the axis of J^ and the ordinates of the curve at the

errors x and x -f- dx, or by
y dx=<f>(x) dx
(6)
And this probability will be known as soon as we find the

form of the function <f>(x).
19. The above expressions in (5) and (6) are the ones
we should use if we regard the curve of error as repreM -n ting the law of occurrence of errors of observation. If,
however, we look upon the curve as expressing the law to
that
which we mu>t make

values of the
tin-
unknown
conform, in order th:it the

obtained from them may IK-
n-Milualt*
<ju:iiititii'H
GENERAL PRINCIPLES.
we should
the most probable values,

the expressions
11
replace x
and use
by
= *()
ydv = <i>(v)dv
y
and
and
for (5)
(7)
(8)
(6), respectively.
THE EQUATION OF THE CURVE OF ERROR,

20.
results
Let n observations, all of the same weight, with

be made upon any function or
n
z
3/i,
functions of a
number
unknown
of
quantities z 1? z 2 ,
un
and let the residuals of M^
n be w 1? v 2
z
and the probability of the occurrence of these residuals be
<(tfi) dv,
<j>(v z )
dv,
<(y n )
P=<t>(v
4>(v 2 )
Each
different
+(v n )
Then
dv, respectively.
occurrence of
simultaneous
probability of the
residuals will be
the values of the
the
these
all
(9)
(efo)
log<j>(v n )-}-?ilogdv
(10)
method that might be adopted for computing

z would lead to a difunknowns z 1? z 2
q
,
ferent set of residuals
u 2,
Wj,
vn
but obviously that set
z q should be considered the best which

of values of z^ z 2
corresponds to the particular set of residuals v^ v 2 , . . v nj
the probability of whose occurrence is greater than that of
,
any other
set.
Therefore the most probable values of
make
The values of z l5
those that
in (9), or log
z2 ,
z x , z2 ,
in (10), a
are
maximum.
corresponding with this latter
condition are those that satisfy equations (11). It maybe

noticed that these equations also express the preliminary
conditions leading to a minimum value of log P, but the
12
nature of the problem
is
such that a
maximum
value of
evidently exists while a minimum does not, and it is therefore unnecessary to investigate further the mathematical
conditions for a
maximum.
Hence we have
Bz l
'<HV
and for convenience we may put
substituting in (11)
we have
5u
5y
These equations contain

there are
we
at*
find tin-
many
;ill
r<|u:itiniH
dv
the
:is
unknowns z^ z^
unknowns,
form of the function
(/')
\ve
z^ and
lu-ncf as soon as
can solve these
GENERAL PRINCIPLES.
13
z
Since
equations for the most probable values of z ly z 2
q
we have considered the general case, and the above results
,
number
are to hold true whatever the
unknown
of
quantities
and the form of the functions observed, we may deduce the

form of i/f(u) by solving a special example.
Example. Let n observations of equal weight be made
upon the value of a single unknown z 1? with results M^ J/,,
vn
Then the
and let the residuals be v 1? w 2
n
most probable value of z is given by
.
differentiating with respect to z 1?
auj
-L
__
.
r\
^.
dz l
But
in this case, as
"~
rv
dz l
dZi
substituting (a) in (14), changing
^(i
dv n
_
_
all
the signs,
ll
+v +
2
In order that (b) and (c)
is,
may
j,(v)
vn
,.
= cv
therefore
Integrating,
log
da
<j>(v)
cw
= |c
That
(15)
dv
^r<f>(v)cv
'
v
-=
--=
<^>(y)
=^
(c)
both be true the functional
Substituting this in (13) and (12),
dz
(b)
in (2),
symbol ^ must indicate multiplication by a constant.

in general
we have
+ *(t>a)+...*(0 )=0
was shown
ol
dz
14
Since y=<f>(v)
we may
is
the equation of the curve of error, (7),
therefore write
it
But on examination of the curve, y is seen to be a decreasing function of v, and hence the exponent of e is essentially
negative.
Accordingly we will write our equation in the
form
v = lee-*^
the values of
(16)
k and h depending upon the character
observations, but in
all
of the
observations of the same kind and
weight having the same values.

This equation represents the law in accordance with which
the residuals must be distributed in order that the best results
may be obtained from our observations. But, as before
mentioned, if we wish our curve to represent the most prob-
we should
able distribution of the real errors of observation
write the equation in the form

'
(17)
Hereafter
we
shall use
without further remark either form
of the equation according to the aspect in

considering our curve.
which we are
An
inspection of the above equation will show that it

the conditions noted in discussing the form of the
curve of error in paragraph 17.
satisfies all
21.
It is
important to notice that in
all
discussions in
tlu>
of Least Squares the number of observations is supto

be large and always greater than the number of
posed
unknown quantities. As will be illustrated later on, para-
Method
graph 'H, whenever this is the case there is a remarkable

agreement bet ue<n the results obtained in practice and those
GENERAL PRINCIPLES.
indicated
are few in
our
15
by the theory. And even when the observations

number the method still affords the best means at
command
for
their
the
adjustment,
results
obtained
merely having a smaller weight than they would have had

derived from a greater number of observations.
if
THE METHOD OF LEAST SQUARES.
We
are now in a position to see whence comes the

" Least
Squares."
In paragraph 20 it was pointed out that whenever we
22.
name
make
a series of observations, each observation of the set

having the same weight, the most probable system of values
of the unknown quantities will be that which corresponds
with the set of residuals the probability of whose occurrence
a maximum.
That is, the best set of values
unknowns will be that which gives a maximum value
is
P=
$(v,}
<t>(v l )
But from equation (16)
..*(y)
(dv)
this reduces to
P = kn e -* W +
2
v **
)
(
dv)
Since the exponent of e in this expression

will be a maximum when
evidently
(18)
is
v? -\-vf-\-
of the
to
= Sv
is
negative,
a minimum.
(19)
Hence the adjustment of observations by the Method of

is based upon the
principle that the most probable system of values of the unknowns is that which renders
the sum of the squares of the residuals a minimum.
Hence
Least Squares
the name.
The conditions
in equations
for a
maximum value
(14), and since
it
of
has been
were expressed
shown
that the
16
means multiplication by a constant, those equa\f/

tions reduce to the following, called
function
NORMAL EQUATIONS.
#Vf
3lJ n
An inspection of these equations will show that they also

express the conditions that will make the sum of the squares
of the residuals a minimum.
In the adjustment of observations the above are the fundamental equations. In order to obtain the most probable
values of the
necessary
is
unknowns
to
in any set of observations, all that is

form the Normal Equations for that set and
solve them simultaneously.

The examples already solved for
direct observations are merely special cases of the above
general solution.
CHAPTER
II.

INDIRECT OBSERVATIONS.
23.
In the determination of the values of quantities by
means of observations the functions
of the
unknowns
that
it
necessary to observe may be of any form, but if they are

not linear the normal equations derived from them are likely
is
be complicated and
to
difficult, if not impossible, to solve.

the observations are not upon linear functions of
unknowns, the first step will be to reduce them to
Hence
the
if
equivalent linear expressions by transformations depending

on the character of the observed functions; see paragraphs
43 and 44. It will be necessary to consider, therefore, the
method of adjusting observations on linear functions alone,

and the procedure in cases of this kind may be illustrated by
the following simple example.
24.
Example.
8^
/#2 ,
s,
are three solids
whose masses
Not having standard weights enough to obtain

are required.
all these masses directly, by varying the distribution of the
solids in the
are
made
pans of the balance the following observations
AS'
tf,
=^
=
=#
-f-
1.7
grams.
2.4
--
3.0
8 , are repreIf, now, the most probable values of S^ S^

sented by Zj, z 2 , 2 s> an d the corresponding residuals of the
v n the values of these residuals in
observations by v^ v 2
terms of
z 1} z 2 , z 8
may
be found from the above observations
IS
by transposing all terms to the first members of

tions, and we obtain at once the following, called
the equa-
OBSERVATION EQUATIONS.
21
22
1.7
2.4
za
_2
+ +
22
23
z2
23
Applying equations (20)
1.0
3.0
=V
=
=^
=v
l
u2
as formulas
we have
the
NORMAL EQUATIONS.
(
2l -2
2
1.7)
+ (_ + +
22
Zl
z3
+ (z
(2,-
2.4)
+ (-z + +
l
z,
z3
-f(2 2
Simplified, these
_2z 2
2X(b)
substitute
-2
-3.0)
(B)
-1.0)
-2 -3.0)(_l) =
3
23
+32
(=)
become
2z 1
X()
-1.0
0.7=0
(a)
2.3
(b)
=
0.4 =
(c)

25.
Rule.
"
For each
19
observation write an " Observation
form a "Normal
by inultijtlyiity the first member of each observation equation by the coefficient of that unknown in tlmt
equation, adding the results and placing the sum equal to
/
then for each imkiioirn
"
Solve these equations simultaneously for the values
zero.
of
the
unknowns.
In solving for the most probable values of the unknowns

the second members of the observation equations are very
vn
For this is
instead of v t
2
which the equations naturally appear, and if the
observations were exact the residuals would actually all be
zero.
The method of solution is the same in either case.
commonly written zero

the
form
26.
in
If the observations
Observations of Unequal Weight.
of equal weight the same method will apply,

in
the formation of the normal equations each
that
except
observation equation will be used the number of times denoted
are
not
all
its weight.
Thus in the last example if the observations
have the weights 4, 9, 1, 4, the normal equations will have
the same form as in (B), page 18, but each part of each
equation will be multiplied by the weight of the observation
equation from which it is derived. This will give the
by
NORMAL EQUATIONS.
9(2 3
2.4)
+ (-
2L
=
-3.0)(-l) =
28
+ + - 1.0)
22
23
_L4(2 2
-2
3.0)
or reduced
52!
-5z, -
5z l
2l
the solution of which

z,
= 7.07
92 2
32 2
23
32 3
+142
5.80
(a)
6/2
(b)
10.6=0
(c)
22
= 5.42
;1
4-2
20
Further
it
will at
once be seen that
if
p p
l9
z,
pn
are the
weights of the corresponding observations, equations

take the general form:
(20)
WEIGHTED NORMAL EQUATIONS.

BVz
Bv n
(21)
Hence for the formation of the normal equations in weighted

observations on linear functions of the unknowns, we have
the following:
For each observation write an " Observation
27. Rule.
"
Equation / then for each unknown form, a "Normal
Equation" by multiplying
the first
member of each
observa-
tion equation by the coefficient of that unknown in that

equation and by the weight of that equation, adding the
results
and placing
the
sum equal
to
zero.
Solve these
equations simultaneously.
result will be obtained if we begin by
each
observation
equation by the square root of
multiplying
its weight and then proceed according to the first rule
28.
The same
(paragraph 25).
This result illustrates the important principle that multiplying a set of equations by the square roots of their weights
reduces them all to equivalent equations of weight unity.
29.
Relation between the Weight of an Observation and

If in paragraph '20 the n observations
the Value of h.

have weights
hi,
A2
^w
p n an(l tne quantity

/> 1? /> 2
tnen equation (18) becomes
.
^ (dv)
h
most probable set of values

a maximum, and
makes
which
+ hfvf-\-
7iiHV
The conditions
h n vn
z
for a
minimum
is
is
h values
Vn^
(dv)
unknowns
of the
Tlie
'l
21
(22)
is
that
maximum when
a minimum.
(23)
value of this expression are
the following, which are then for this case the
NORMAL EQUATIONS.
But equations (21) are also the normal equations

Hence (21) and (24) must be identical, and
for
this case.
Pi
P*
...
Pn
/*i
7i 2
2
:
/*,/-
(25)
A is proportional to the weight of

Accordingly, since h increases in value as
the quality of the observations is improved, it is called "The
That
is,
the square of
the observation.
Measure of Precision."
22
(23) and (25) that the most
follows from
it
Further,
unknowns
probable system of values of the
will
be that in
which
Pi ^\-\-PI v2
minimum.
2
Pn Vn is a
(26)
And this is the most general form of statement of the

" Least
Squares." The same principle is repreprinciple of
sented in equations (21).
If large numbers occur
better to compute the most probable
corrections to apply to the observed values rather than the
most probable values of the unknowns themselves. In this
30.
way we can
31.
of Corrections.
Computation
in the observations
often avoid a large
Example.
altitudes
it is
amount
P P P
P^
l9
8,
of numerical work.
6,
are five points whose
above the mean level of the sea are to be determined
from the following observations
P = 573.08
P -P
P
P P
P -P
Z
An
4,
of difference of level.
P - P = 170.28
P P = 425.00
P = 319.91
P = 319.75
= 2.60
= 575.27
= 167.33
= 3.80
!>
I>
inspection of these observations shows that
P = 573 +
l
/>f =676
P = 745 + z
P = 320 + z
1,
we may put
(A)
z, z s, Z H z s are small corrections whose most

now have for
values
are to be determined.
probable
where
Zj,
We
OBSERVATION EQUATIONS.
573
-573
575
Zj
- 573.08=0
or
2.60=0
or
*!_|_ Za
575.27
or
z,-.08
z2
z l -.(>()
,
.27
=
=
=
- 575- z
742 _ z
745 _j_ Z4
_
575 _ z
745 _|_ z
_
_L.
z
320 -z
745
742
z3
320
320
From
these
+
+
167.33
3
a
c
z6
zc
3.80=0
_ 170.28 =0
- 425.00=0
- 319.91 =
- 319.75=0
we now form
or
za
or
24
or
z4
or
z4
23
-z - .33 =
-z .80=0
- z - .28 =
2
=0
z6
+ .09=0
=
or
z6
or
z 6 -j-.25
the
NORMAL EQUATIONS.
+
z4
z4
z2
z3
-f3z 4
z4
=
1.08 =
=
-f
-f
z5
+ 3z
.52=0
.26=0
.47
.34
and solving,
2l
=-.19;
z2
= .14; = .05; = .43; =

z3
Substituting these in equations (A)
z4
z5
we have
.03
for the
most
probable altitudes,
^ = 572.81
P = 575.14
2
P = 742.05
P = 745.43
P = 320.03
If the original observation equations had been retained,

the independent terms in the normal equations would have
been
570.48
32.
240.26
Significant Figures.
163.53
599.08
The adjustment by
214.66
the
Method
of Least Squares of observations which occur in practice,

although not difficult, is apt to be long and laborious. Hence
much as possible it is of great importance that careful attention should be given in the solutions
to reduce this labor as
to the proper use of significant figures.
When
in
doubt,
24
however, as to the proper number of figures to retain it is

many rather than too few, as the superfluous
while
figures can be rejected at the end of the computation
better to keep too
too few are retained the results obtained from the computations will be worthless.
if
For a general discussion of the subject of significant

Holman's " Precision of Measurements," pages 76
see
but for the present the following rules will
figures
to 84,
most
suffice for
cases.
Rule
1.
In casting off places offigures increase by 1 the

when the following figure is 5 or over.
In the precision measure retain two significant
last figure retained,
Rule
2.
figures.
In any quantity retain enough significant figures

place in which the second significant figure
of its precision measure occurs.
Rule 4. When several quantities are to be added or
subtracted, apply Rule 3 to the least precise and keep only
Rule
3.
to include the
the corresponding figures in the other quantities.

Rule 5. When several quantities are to be multiplied or
divided into each other, find the percentage precision of the
If this
least precise.
is
per cent or more, use four significant figures.

"
.1
"
jO I u
"
"
"
..
five
8 {x
"
"
u
in all the work.

conflicts with
If the final result obtained in

Rule 3, apply the latter.
this
way
6.
When logarithms are used, retain as many
in
the
mantissce as there are significant figures
places
retained in the data under Rule 6.
Rule
The application of these rules is not always possible in the

course of the work, since the precision measures may not be
known
until the
it IH
end of the computation.
But as
;i
<_'*
aufficient in direct observations to retain <>nc
;u-r:il
more
25
place of figures than is given by the individual observations,

in indirect observations to retain two additional places.
and
CONDITIONED OBSERVATIONS.
33.
Conditioned Observations
are
those
in
which
the
unknown
quantities must be determined not only so as to

satisfy as closely as possible the observation equations, but
also so as to satisfy exactly certain other conditions.
These
number than the unknown quantities,

unknowns could be determined from the con-
conditions must be less in
otherwise the
ditions alone.
The adjustment of observations of this class may be reduced

method already used for unconditioned observations in
to the
the following way.

The observations are represented by " Observation Equations," and the conditions by certain other equations, called
" Condition
Equations."
Between these two sets of equations we will eliminate as

many unknowns as there are conditions. From the resulting
equations, which will be the same in number as the observations, we will form in the usual manner the "Normal
Equations" for the remaining unknowns. Having solved
these normal equations and substituted the results in the condition equations, we shall obtain the values of the unknowns
first
eliminated.
All conditions of the problem are now fulfilled, for the

condition equations are satisfied exactly and, moreover,
according to the principle of Least Squares our results
are those rendered most probable
by the existence
of the
observations.
As
in
the
example
last
considered,
it
is
often
more
advantageous to compute corrections to the observed values

of the unknown quantities rather than the values of the
quantities themselves.
26
34.
Find the most probable values of the angles

from the observations,
Example.
of a quadrilateral
= 101
B= 93
C= 87
D= 11
^4
13' 22"
39
"
52 40
"
58"
0'
The condition
weight 3
49 17
to be satisfied
is
in this
problem
(B)
Let
gj,
z2 > z s
^ e the most probable corrections
24
to
add
This gives for
to the observed values.
OBSERVATION EQUATIONS
=:
zl
=0
^0
=0
*2
z3
z4
weight 3
2
2
1
and for the
CONDITION EQUATION
Zi
z2
+Z8
+ + 58=0
*4
(D)
Eliminating z 4 between (D) and (C), the equations from

which the normal equations are to be derived become
zt
z,
=
=
=0
weight H
2
"
TllK
ADJUSTMENT OF OBSERVATIONS.
Applying the
rule in
27
paragraph 27, these give the
NORMAL EQUATIONS
-
4 Zl
22
58:=0
23
(F)
in equation (D),
Solving, and substituting the results
we
find
z,= 22
8.29
28
=- 12.43
=- 12.43
z<=- 24.85
Applying these corrections to the observations (A), the

most probable values of the angles are
A = 101
13'
93
49
4.57
87
26.57
77
52
15.15
B=
C=
D=
13".71
In eliminating unknowns between the observation

and condition equations care must be taken that the observation equations are not combined with each other or multiplied by any quantity. For if this is done the weights of
Note.
the observation equations will be altered.

(See 28.)
35. In the above example it is evident that the corrections
be applied to the different observations are inversely as

their weights.
And, in general, when there is but one
to
equation of condition, the observations expressing direct

determinations of the unknowns, the corrections will be proportional to the coefficients of the unknowns in the equation
of condition divided by the weights of the corresponding
observations.
proof of this
is
given in paragraph 117.
28
The most common

are
all
is that in which these coefficients

example just solved, and we may then
case
unity, as in the
derive the
Find
Rule.
the difference between the theoretical

and divide this correction among
observed results
and
the
observations in the inverse ratio of their weights.

In the last example the sum of the observed angles exceeds
360 by 58".
~ 58 X
Therefore the correction to be applied to
- -=
^~
58
is
8.29
EMPIRICAL FORMULAS AND CONSTANTS.

36.
In the work so far considered the observations are
supposed to be made either directly upon the values of the
unknown
quantities or upon some function of the unknowns

whose form, and the constants entering into it, are definitely
known. But another sort of problem frequently occurs,
in which observations are made upon the values of a certain
variable and the corresponding values of some function of it,
the exact form of the function not being known. The object
in this case is the determination of the most probable form of
the function and the values of the constants involved that
;
the derivation of the algebraic expression best representing

the law connecting the variable and function.
is,
This expression
may be
looked upon as the equation of a
curve, abscissas denoting values of the variable and ordinates

values of the function, and for all values of the variable
within the range of the observations we may determine from

it the most probable values of the function
corresponding.
But except in special cases, where tin- number of observations
i>
large,
where the law connecting variable and function is

and where the equation obtained is an aeeurate
well detim-.l,
29
representation of this law, it cannot be assumed to apply

beyond the range of the observations. And in no case would
it
be safe to make use of the curve very far beyond the limits
of the observations.
37.
The Method
of Least Squares will not assist in deter-
mining the form of the function. This must be settled upon

beforehand, either from theoretical considerations or by
constructing a plot, using values of the variable as abscissas
and of the function as ordinates, when the smooth curve
drawn through the points thus obtained will indicate the

form of equation to be used.
It is to be observed that this is a method of trial and will
not necessarily give the most probable form of the function
and in fact we may not be able to obtain the form that would
be absolutely best. Further, several forms of equation may
be known which would represent well the plotted points. In
such a case that should be considered the best in which the
sum of the squares of the residuals is found to be the least.
38. As soon as the form of the function is decided upon it
should be reduced to the linear form, and the determination
;
of
the values of the constants involved
then a simple
is
application of the preceding methods.

As the " Observation Equations " in any given problem will
all be of the same kind, it is
usually advisable to write out
the general form of the "Normal Equations" and arrange
the computations in tabular form, while the retention of
proper number of significant figures

importance in this work.
tlu'
is
of
particular
39.
case that frequently occurs is that in which the
quantity y is a constantly increasing function of the variable &, or where the plotted curve is approximately parabolic
in form.
Here the equation

...
may be
(27)
taken to represent the relation between the variable
30
and the function. The larger the number of terms taken in

the second member, the more accurately may the equation
obtained be made to represent the results of the observations
but the labor involved increases rapidly with increase in the
number
40.
and if the plot shows a very nearly

two terms alone may suffice.
of terms,
line the first
straight
In measuring the velocity of the current

were obtained
Example.
of a river the following results
Depths.
Velocities.
4.86
5.14
5.15
4.85
4.24
3.36
2.16
0.67
The velocity at the surface is 4.250. Find the equation of

a curve which will express the relation between x and V.
Plotting the observations we find the curve
V
f
31
This is approximately parabolic in form and passes through

the fixed point (0, 4/25).
Therefore the relation between x
and
be
expressed by the equation
may
V= 4.2
6V
(A)
and substituting in this the corresponding values of x and

as given by tlu- observations, we shall have eight observation
and C
equations from which the most probable values of
are to be computed.
All of the observation equations being of the form (A) we
have the
NORMAL EQUATIONS
.25
2 Vx 2
(a)
.25
- 2 Vx =
(b)
For computing the coefficients in these equations

convenient to arrange the following table.
it is
most
32
Substituting these results in (a) and (b)
8772
we have
C+ 1296 B-\- 341.95 =
(c)
41.17 =
(d)
1296(7+ 204 ./?-{and solving,
= .7465
(e)
V = 4.25 + .7465x - .1493*'
(f )
C =-.1493
J?
Therefore the required equation
is
Whenever the quantities in the observations are so large

that the use of logarithms is desirable, these can best be put
in the same columns directly over the natural numbers.
41.
It
is
to
be remarked that in solutions like the above
we
assume, from our method of forming the observation and

normal equations, that the observations on the values of the
function are alone subject to error, the observations on the

variable being supposed to be exact or to have errors so small
as to be negligible.
42. Periodic Phenomena.
If as the variable increases the
function passes through recurring values, that is, if y is a

periodic function of a?, some form of trigonometric equation
would be the proper one to select. For instance, a good form
to use is
y
where
= A + B sin
-4, jB,
C cos^fx
(28)
are the constants whose most probable values
are to be found, and ra is the number of units of x comprised

in the entire cycle of values of y.
The quantity
is to be
determined from an inspection of the observations, and if it

might be used, all should be
appears that several values of
tri. <1, and that which leads to the smallest sum for the
squares
of the residuals
is
to be considered the best.
If the several
33
cycles are not similar and regular, additional terms involving

multiples of x will have to be added to equation (28).
43. Special Treatment of Exponential Equations. If the
equation selected is not in the linear form as regards the

unknown constants the general method of procedure is
given in paragraph 44, but in some special cases a more
simple reduction
is
possible.
quite
common
case
is
the
following:
Suppose the relation between x and y

equation
is
expressed by the
= kxm
(29)
the problem being to obtain the best values of

Taking the logarithms of both members of (29),
log
Denoting log k by
= log k
k',
-\-
m log x
this equation
m log x -j- k'
k and m.
log y
becomes
which is in the linear form as regards the unknowns m and

in the usual
k', and the normal equations may now be formed
manner.
The most convenient way of determining whether equation
(29) is a suitable one to select or not is to plot the corresponding values of x and y on logarithmic cross-section
paper.
will lie
If (29) is a
proper equation to use the plotted points
upon a straight line.
REDUCTION OF OBSERVATION EQUATIONS TO THE

LINEAR FORM.
When
the observation equations are not linear as

regards the unknowns, the only practicable method of pro44.
cedure, as already mentioned in paragraph 23,
is
to reduce
METHOD
34
them
follows
This reduction
that form.
to
LEAST SQUARES.
O/*'
may be
effected
as
Let the observation equations be

,/UZ,,
t,
Z
Z
2,
2,
Zq
Zq
= M,
= J/
8
.....
which
in
Zj,
J/u J/2
known form.
,
the unknown quantities and

g represent
the observations, the functions being of
2,
(A)
Z9
Zq be approximate values of Z15 Z2
Let Z/, Z2
found by trial or the solution of a sufficient number of the
'
',
observation equations, and let the most probable values of
Z Z
2,
t,
Zq
be
ZS
+ ZH
'
z2
,...Zq
'
(B)
z2
z
small corrections whose values are to be
q being
determined by the Method of Least Squares. The first
observation equation in (A) may then be written,
Zj,
Expanding the
.
)
.
member by
Taylor's Theorem, denoting

and
neglecting terms containing
by &!,
z
the first, this becomes
than
q higher
first
Zq
/i(Z,', Z/,
powers of z n z a ,
.
Ok,
If
*!
now we
î
represent the coefficients of

also
*i
Jl/i
by m,, and
z,,
z2 ,
treat
tf
the
oilier
35
equations in (A) in the same manner, our observation equations will take the form
The second members reduce
the
since
the residuals
to
the most probable

quantities
7 represent merely
values of the corrections.
Equations (C) can now be solved
z l5 z 2 ,
in
the
Z^
Z.2 ,
usual
.
Zq
the most probable values of

found by substituting the results in (B).
manner and
From the following observations find the

45. Example.
most probable values of x and y.
sin
cos x
cos 2y
1.5
-f-
3 sin
1.7
a*
By
trial
it
is
=
=
y
+ 5y =
-\-
(A)
2.1
and 18 are approximate

accordance with paragraph 44
found that 42
values of x and y.
Then
in
we put
1
and expanding the different functions, we

&!
*2
= 42 + cos 36,
= 1.48
= cos 42 + 3 18,
= 1.67
sin
sin
18ir
= 2.11
=
^
JT/!
-^
= cos 42,
^-
= -sin42,
9
J/2
sin
36
3 cos 18
=2.85
42?r
=-2
~=
=-.67
dlc 8
&2
^--
=-1.18
d Ar,
=5
=1.47
.02,
find
=.74
2
%j
.03,
J/3
.01
86
Therefore
we have
for
J425J
1.18z 2
.02
_ .67z! -f
2.85z 2
.03
1.47 2l
From
_|_
=v
=v
=v
5z 2 -f .01
(C)
t
z
these are obtained the
NORMAL EQUATIONS
Solving,
z
+ 4.57z + .020 =
+ 34.51z - .012 =
= 00147
1= =-. 00845
3.16 2l
4.57z x
z2
These results are

measure we have
z
l= - 29'.0
Reducing
z2
(E)
measure.
in circular
to degree
5'.1
(F)
Substituting these in (B),
x
46.
= 41
y=18
Sl'.O
5'.1
(G)
If the observations are conditioned, precisely the
same
method
will be followed in reducing all the equations to the

linear form.
The rest of the solution will then be as usual.
If the values
found for z^
z2 ,
should turn out to be
zq
so large that the terms involving their second and higher

powers can not be neglected as assumed, the process must
repeated using the values of
Z Z
approximations.
In a few cases the equations
t,
2,
Zq
first
may be reduced
obtained as
to the linear
form by some special artifice of a simple* character, as in

paragraph 43. In such cases the method of expansion by
Taylor's
Theorem should not be
used.
CHAPTER
III.

The work
47.
so far considered has treated solely of the
methods by means of which the most probable values of the

unknown quantities may be determined from a series of
But in general something more than this is
observations.
desired.
We wish to know, if possible, how much reliance
can be placed upon the results obtained, and how they compare in precision with other determinations of the same
quantities.
our results
Curve
Preliminary to the discussion of the precision of

will be necessary to consider more fully the
it
of Error.
The Constant h. The law of distribution of errors of

shown to be represented by a curve
48.
observation has been
whose equation
is
= ke-
y
In
this, if
0,
=k
(30)
Therefore the constant k represents the intercept of the

Curve of Error on the axis of Y. It is not, however, an
independent quantity but is determined by the value of A,
as will
now be shown.
h. Since, as shown in paragraph 18, the total area between the curve and the axis
of
is denoted by unity, we have
To Find
49.
in
Terms
of
or
k C*e- h 'W dx
*/0
This
may be
written
= ~*
(a)
38
Let t
and when
Aic,
a;
.-.
0,
dt =. hdx.
0.
6-*
<ft
tf
"
Also,
when x
C e-M* h dx
GO,
oc,
(c)
*ô
/o
Multiplying this equation by
we have
^ * T = J]" J!"
[JT
= f"Jo
1
_
+
2(1
~"2(1
<fo
x2 )
X2
re-*
Jo
dx
Jo
Substituting (31) in (c) and (b)
^
2
k
2&
"
y/^
Therefore equations (16) and (17) become
(33)

50.
It
39
has been shown that the quantity A is a "Measure

and hence a determination
of Precision" of the observations,
would enable us
of its value in each case
relative reliability of different
to
compare the
measurements.
In practice,
not found convenient to compute the value of

h directly, and so this quantity is used only in developing
the theory of the subject, while in the comparison of observa-
however,
it is
tions certain other quantities now to be derived are used as

These latter quantities are called the
precision measure's.
of the Errors," the "Mean Error," and the "Probable

Error," respectively, and maybe computed directly from the
observations.
Further, as it will be found that they all bear
"Mean
a definite relation to A,
the value of this quantity can be
determined from them if desired.

In the following discussions no distinction
is
made between
positive and negative errors of the same numerical magnitude,

and unless otherwise stated the observations are all of the
same weight.
THE MEAN OF TKE ERRORS OR AVERAGE DEVIATION.

51.
The Average Deviation
mean
the arithmetical
an observation
of
(a.d.)
is
of the errors all taken with the positive
sign.
Since from (6) the probability that the error of a single

observation will
fall
between x and x-{-dx

<f>(x)
if
dx,
n observations are made, the number
between these limits
the
sum
4>(x) dx.
of all the errors of the observations
J- x
is
xoo
of errors falling
is
Hence
is
x<j>(x)
dx
or
2n
Jo
<j>(x)
dx.
40
Dividing this
a.d.
last
/*>
Jo
n we have
expression through by
2A
x+(x) dx
e~h
^'x dx
VTTÔ
2
*'
-^z f V* (-2A jc)
2
<fo
AVirô
a.d.
or
(34)
THE MEAN ERROR.

52.
The Mean Error
root of the arithmetical
The
total
number
of an observation
(/A)
mean
is
the square
of the squares of the errors.
of errors being n, the

is, as just shown,
number
falling
between x and x-\-dx
<j>(x)
dx,
and the sum of the squares of these errors
n x $(x)
1
dx.
Therefore the sum of the squares of
Ja
x2
<t>(x)
the errors
all
as
shown
in
is
dx
>
But
is
dx
paragraph 49,
dx==l
or
-hW dx
-La,
Differentiating (b) with respect to
-2A
^.
(b)
A,
^Vefr;
= -~
c)

and replacing the integral in
from equation (c), we have
(a)
by
41
value as determined
its
THE PROBABLE ERROR.

53. The Prohable Error (r) of an observation is an error
such that one-half the errors of the series are greater than it
and the other half
less
than
Or
it.
an error of such a
it is
magnitude that the probability of making an error greater

than it in any given observation is just equal to the probability
of making one less than it, both probabilities being one-half.
The probability that the error of an observation will fall
between x and x-\-dx being <j>(x) dx, the probability
r is
that the error will fall between the limits r and
P =
If
is
J-r
4>(x)dx
the probable error,
Ce-Wdx
yVô
is
wfe
(36)
one-half, or
(37)
and from this definite integral r is to be found.

Let t
hdx.
we have
Also when x
hx, .'. dt
r,
t :=
=.
and
when
x
t
these
results
0.
Ar,
0,
Substituting
in (37), we have
(38)
42
Denote hr by p. Then by interpolation in a table of

values of this integral, the value of hr in (38) is found to
be
hr
When
is
(39)
.47694
54.
.47694
small the values of
may be
e~^ dt
found by expanding e~* into a series and integrating the

successive terms.
Thus, by Maclaurin's Theorem,
= Jo/<
dt
(1
v
/4
..
dt
IL
J!_
5[^
I
3
RELATIONS BETWEEN
55.
From
(40) r
= -,
r
or
Also from (35)
=
=
or
r, a.d.,
and from (34)
Ik,
AND
p.
a.rf.
p a.d. \f*
,8453a.^.
|i,
7[3_
(41)
=
=
n\
.6745
\i
(42)

The relation between the values of /x,
be conveniently expressed as follows:
u.^2
"~
===
a.d.,
r,
43
and h may
== ~ == a '^' 0f
(^)
or arranging in tabular form,
From
1.0000
1.4826
1.2533
0.6745
1.0000
0.8453
a.d.
0.7979
1.1829
1.0000
this
it
a.d.
will be seen that
p.
>
>
a.d.
56. Further, since in (25) it

follows at once from (43) that
a.d.*
JL
2
K-
(44)
was shown that
QC
JL
r2
oc
<x
h\
it
fc
That is, the weights of different determinations of a

quantity vary inversely as the squares of their Mean Errors,
their Probable Errors, or their Average Deviations.
It is to
be observed, however, that the determination of the
relative weights of quantities from a comparison of their

precision measures according to (45) applies only when the
quantities are of the same kind and subject to the same constant errors, if any of the latter exist.
The
98.)
(See
applications of (45) are numerous and important.
57. Example A.
Suppose n direct observations, all of
the same weight, be made upon a quantity, and that the
probable error of a single observation is r. Then since the
44
weight of the arithmetical mean

will be given by
*
Or
its
n,
probable error r
'
or
in general, suppose
observation of weight
is
/>,
(46)
,.
is any precision measure of an

and suppose p is the weight of a
second similar quantity or observation, then the corresponding precision measure S of the latter will be
The
p=
common
case of most
occurrence
is
that in
which
and then we have
1,
A line
Example B.
is
= -=
(48)
measured
five
times and the aver-
age deviation of the mean (A. D.) found to be .016 feet.

How many additional measurements are necessary in order
that the A. D. of the mean may be reduced to .004 feet ?
Let x be the
total
number
.000256
x
Consequently
required
the
of observations required.
number
Then
.000016
80
of
additional
measurements
75.
is
Example
C.
In two determination! of the quantity
there were obtained
1^
Find
of
427.320
tln-ir
and
its
0.040,
relative weights,
probable error.
Z,
427.30
0.16
and the most probable
v:ilue

The above
Note.
method commonly employed
the
is
45
to
denote that the probable errors of the observations are 0.040
and
0.16.
From
From
(3), the
16
~
most probable value of
A=
and the weight of
427
16
will be given
-
by
30
"g
427.319
being 17, by
= -^ =
vTT
Therefore
16 2
(45)
we should
(48),
.039
write the result
= 427.319
Z
REPRESENTATION OF
a.d.,
\L,
.039
AND r ON THE CURVE OF
ERROR.
58, To
we have
find the points of inflection of the
y
-
For a point of
inflection,
2 A2
e-^
of
ke-M#
2 h*k
].,
Curve
a
(2 A
su
0,
or
1)
"
=
by(35)
"
Error
46
That is, the Mean Error is represented by the abscissa of

the point of inflection of the Curve of Error.
See
in
OM
9.
page
Next, for the abscissa of the centre of gravity of the area
we have
to the rifirht of
t
figure 2,
OD
x dx
y
= ^ry dx
I
7=
VTT
for,
Cy dx=
18),
/0
Integrating,
a;
Finally,
if
an ordinate
PP'
a.d.
by
(34).
be drawn so as to bisect the
area to the right of the origin between the curve and the
axis of X, the Probable Error will be represented by the dis-
For the probaoccurrence of an error less than the amount OP
tance of this ordinate from the axis of Y.

bility of the
is then equal to the probability of the occurrence of an error

is the
greater than OP. This being the case,
probable
error by definition.
OP
CHAPTER
IV.

DIRECT OBSERVATIONS.
Observations of Equal Weight.
59.
vations
Given n direct obser-
the same weight on a single quantity
M, to
Mean and Probable Errors and Average Deviation
all of
find the
of a single observation
and of the Arithmetical Mean.
M^ M^
Let the observations be

"
"
"
*'
"
"
arithmetical
real errors
mean be 3IQ
be
residuals be
n.
x lt
ic 2 ,
xn
v l5
v 2 ...
vn
Denote the mean and probable errors and average deviaby /n, r, and a.d., respectively,
tion of a single observation
and the corresponding quantities for the arithmetical mean

by /MO, 7* and A.D. Then by definition,
,
Vj.
= MI
MQ,
and
If
V2
MQ
= MI
...
=M
MQ
= \\*&
represented the true value of
would be the same
Vn
as the real errors,
M,
the residuals
and we should have
/
!
and if n is large this formula is practically exact. But when

n is small a more accurate expression is necessary. To
48
obtain this let
3f
be the true value of
<>)
ic
Siquaring, adding,
=
exactly, but
Jf
that
is,
of
it
is
by
vl
vn
Therefore
XQ
and dividing by n
The value
M.
for
by (2)
Sv
known and can not be found

approximately equal to the mean error of
not
is
je
to
(48),
Ji..
^n
Substituting this value in the above
n
(n
by (42)
and
\)p?
.6745
n=
.6745
,.
we have
2u 2
(49)
---
(52)

60.
In order to avoid the use of the squares of the
we may proceed
residuals
On
as follows
From
(49)
the average, the values of the residuals will then be
n-
Adding and dividing by
n,
neglecting the signs of the
residuals,
2u
n
"S,x
ff.ff.
by (48)
A.D.
by (41)
rn
and
49
In -^~T
n
50
purpose. Results derived from (50) are to be regarded

more accurate than those obtained from (54), the latter
being a second approximation.

61.
Example.
From
the following measurements on the

and the
length of a base line find the most probable length

values of the various precision measures.
51
and we should write for the most probable length of the

base line
Jf
455.330
.042
Using the same
Observations of Unequal Weight.
62.
notation as above, with slight modifications,
Let
"
"
"
will
represent the General Mean.

Pn be the weights.
P\i Pit
a.d.t, a.d.i, . . . a.d. n be the average deviations.
MO
"
"
we
/*!?
/*a
r it
<f>
a.d., fr r,
/*n
rn
be the mean errors.

be the probable errors.
and v refer
to observations of
weight unity.
Then by (48)
11
i/
etc.
If the " Observation
Equations
" are
formed for this case
they will be
MI
M = VD M
MQ
v 2,
M=
Q
v n.
And, as was shown in paragraph 28, if these equations are

each multiplied by the square root of the weight of the corresponding observation, they will all be reduced to equivalent
equations of weight unity. On performing this operation
residuals of the new equations
it will be seen that the
become
And
of
evidently to these reduced observations the formulas
paragraph 59 apply.
52
Therefore
V/
^ V
M.
Also,
= V
a.d.
(58)
rt
.6745
|i.
(59)
n==
.m
(60)
similar to that used in paragraph 60,
Sl?
=
=
l6745
it
that
a.d. k
v^lD
by a method
may be shown
n-"l
n(n
.8453 a.d.
rk
.8453 a.d. A
(62)
.8453 î.D.
(63)
(61)
1)
Formulas (61), (62) and (63) are not of much value
in
practice unless all the weights are perfect squares, for otherwise no labor is saved in the computations, since the square
root of each weight will have to be determined, and the results

obtained can not be considered as reliable as those given by
formulas (58), (59) and (60).

63.
Example.
Given a
series of observations
difference in longitude between

a.d., A.D., r, r4 , r
/x,
/IQ,
.
two
stations.
on
To
Jf/J
find
53
the
54
FUNCTIONS OF OBSERVED QUANTITIES.

Theorem. Given any number of quantities and their
Mean and Probable Errors and Average Deviations, to find
the Mean and Probable Errors and Average Deviation of any
64.
function of the quantities.
Let the quantities be

"
"
mean
JH/j,
j^,
/^,
errors be
z,
.
pq
M=f(MM
"
"
function be
"
"
mean
"
"
probable error of
"
"
average deviation of
i,
M be E.
M be R.
M be D.
error of
..
q ).
The derivation of the general formula will be simplified

we consider first a few special forms of functions.
Case
65.
I.
if
M = MI M
Suppose
2.
The number
of observations from which

and M^ and
and /^ have been determined is not necessarily
known, but we may assume that for each quantity it is any
large number n, and that the real errors of the observations
hence
are
for
J/i,
z,
Then
the real errors of
observations on
or
*/, */',
--
a,', a*", *,'",
and M^
E>
V,
computed from the separate

will be
m*+m
(64)
since in the most probable case the term
appear, as there most likely will be as
Sxjg,
many
will dis-
positive as
55
negative products of the same absolute magnitude of the
form x
ic 2 .
By successive applications of the above the same principle

may be extended to cover the algebraic sum of any number
So that
of quantities.
if
M= M M
&= ^+ +
then
...
}V
n,
Sfi
(ttf)
Since probable errors and average deviations differ from

errors merely by a constant factor, we shall likewise
have
=
D =
12*
66.
ri
a.d.?
Example,
Cambridge west
Omaha
(C)
Springfield east
Find
L,,
for
"
+ a.d.S +
Case IT.
-f-
7?
=
=
r*
.
Sr
(66)
Sa.tf. 2
(67)
results,
h.
m.
sec.
sec.
of Greenwich,
44
30.99
0.23
"Cambridge,
Omaha,
39
15.04
0.06
25
8.69
0.11
"
y/(.23
Suppose
JS
and
its
probable error.
= 5* 58 m 37*.34 0*.26
= .26
+ +
Jf" = ai-Mi
.06 2
Using the same notation as in

errors of J/ will be
or
the longitude of Springfield,
by (66)
67.
Given the telegraphic longitude
(A)
(B)
?V
.II
paragraph 64, the
real
=
aiî
(68)
56
68.
Case III.
aq
(65) and (68)
By combining
also
J2 2
and
J> 2
69.
M= a^Mi
Suppose
=
=
Sa
r2
(69)
S a 2 a.d.*
The length
Example.
of a bar at 20
is
Centigrade
0".0037, and its expansion per

0".0018.
What is the length
centigrade degree 0".0036
of the bar at 56 Fahrenheit ?
found to be 75".0041
20
C.
20
+ 32 =
5
The expansion
F. will be
for 1
9
.0020
or
68
F.
.0036
.0018
.0010
= 75.0041 - 12 X .0020
= 74.9801
E = V.0037 + (12 X .001
= .013
L = 74".980 0".013
L
70.
Case IV.
Let
MI
=a
where a t a s
,
Suppose
-f-
lt
=a
(Jtfi,
q ).
M=
wi a
a q -\- m^
q
assumed
quantities very
arbitrarily
M^
a q are
M=f
a -f-
so that
. . ni
are
ly m^
nearly equal to JHJ, M^ . . .
qt
q
so small that their second and higher powers may be negThen the errors of Jf/i, M^
lected.
q may be con.
sidered to be in
m m^
lt
mq
and hence
/i t ,
/x.,,
/x y

may be regarded
/;/,,
>/>.,,
mq
M=
/(i,
2,
errors
2,
-|-
9)
by Taylor's Theorem, and

by M', we have
dM
mi
negligible terms in
higher powers of
'
Then the mean

error of terms
is
error of
(A), and by
practically the
ly
2l
=
=
this is
thing,
mq
and
.
same as the mean

given by
by
sides of a right triangle are
ured with results
'
will be the
(69)
same
Two
Example A.
'
the second
m m
71.
expression
.
dM'
what
quantities
M'
-
or
the
of
now have
+m
ai
this
Expanding
denoting
mean
the
as
We
57
49.53
0.59
50.38
0.93
meas
58
Find the length of the hypotenuse

In this example
and
its
probable
error.
.59 2
r49.53~l 2
4-
.93
r50.38~l*
|_70.65j
[_70.65j
.78
and
70.65
0.78
.Example B. If the probable error of x is r, what is

the probable error of the common logarithm of x ? In this
case
M=
Iog 10
^ =
Iog 10 x
d Iog 10 x
Iog 10 e
dx
Example
Po of sin
C.
If the
weight of x
is
/>,
what
is
the weight
x?
Denoting the mean errors of x and sin x by

respectively, we have, by (45) and (70),
E*
ua
c>*
ft
cos*
as
/t
and
-*,
Po
t^_
j&
J)
Po
59
cos x
sec-
72. Equation (70), which expresses the law of propagation of error in functions of observed quantities, is one of the
most important
whole theory of the Method of Least
in the
in particular is based the discussion of th*

Squares. Upon
" Precision of Measurements."
This subject treats, in the
it
first
place, of the
methods of finding the precision of a
quantity obtained by computation from a series of measured

snd, in the second place, it investigates the prequantities
cision with which the component measurements of a series
;
must be made
in order to obtain a required
illustrate the character of the solutions

73.
Example.
degree of pre-
The following simple example win
cision in the final result.
In the determination of a current bv a
tangent galvanometer
we have
rr
10 -- tan
<
I is the current in amperes, II the horizontal component of the earth's magnetic force, G the galvanometer
the angle of deflection.
Given the errors
constant, and
in -f-fy G and tan <, to find the error A in I.
81,
25 8 3
where
<f>
By
A2
(70)
=
100
60
Dividing this equation by I*
That
to the
is,
sum
and tan
Hence
H
tan
is
determined within
"
^2
"
"
.1
v/
.16 -+- .04
Next, suppose the value of

of
(r,
H,
.01
-f
is
"
"
.46 per cent.
required to within
(c)
.1
per
find the necessary accuracy in the determinations

and tan
when the error in each of these quan<
have the same influence upon the

(b) we shall now have
tities is to
From
we have
per cent.
"
"
.4
<f>
=
To
<
if
cent.
- tan 2
the square of the percentage error in I is equal

of the squares of the percentage errors in J3J 6r,
<.
then
=
8,
Sa
83
=
=
=
total error.
.000577
.00058^
.00058
.00058 tan
(e)
<
It is comparatively easy to obtain the necessary accuracy

and tan <, but difficult in the
measurements of
in the
case of
77.
For additional work of

of Measurements."
this
kind see Ilolman's u Precision

74.
It is to
Combination of Functions of the Same Variables.

be noticed that equation (70) applies only when
M^ M^
If
a function of independent quantities.
are merely different functions of the same
is
must proceed as follows

Let
Mi
<
M =
M =
Jlf
for
"
ic 2 ,
JCj,
we
quantities
z,t)
z2 ,
z*)
z 1? z 2 ,
2 fc
are subject to
the corresponding errors in
ic^
and
will be
M
M
ly
XL
2y
A^
Where
J/i,
(z x , z 2 ,
$ (Zu
any single observations of
errors
If
61
= a^
=
!,
and a/, a 2
'
with respect to z^
will then be
ak
at
Where
X=
z2 ,
Then
if
and
^1'
Jl/i
(Aai
the
2 'z 2
a k xk
a k 'x t
The corresponding
zk .
AX,
(a)
(b)
M^
error in
A'X,
(c)
are the differential coefficients of
and 3f2
Substituting in (c)
A'a\) x v
number
-\-
the differential coefficients of

.
denoted by
a z xz
are the differential coefficients of

;
X=
with respect to
and (b),
-|-
oj'aj! -j-
(Aaz
\Xk
-\.
Aa\)
from (a)
of observations or values of
X be
w,
XV*
(d)
62
since
in
the most probable case the product terms will
cancel out.
Expanding (d) we have
E*
+ A'a,')* rf +
(Aa,
+ A'a,'}
(Aa,
^' 2
tf
KV +
.'V
As
Example.
J/i
Jf2
^4
(71)
^4'
1,
r= 4
.01
or
/n
325!,
0.1,
= + J/
=
a =
a =
9
2
2
3
+ X + X X X =
E = 0.5
Jf
Then
By
a very simple problem take
22i,
and
(70
...)
75.
Jlft
'
1,
2,
3.
-01
.01
.25
In this particular example the result may be found directly

from (68) by substituting at first in
the values of M^
and JH/j.
M=
Thus
22 X -f 32!
E=
If
or
J/i
(69)
and
we
5^
52 t
0.5
J/2 had been independent quantities, by (64)

should have had
E=
V(2 X
0.1)
+ (3
0.1)
0.36
INDIRECT OBSERVATIONS.
76. The determination of
unknown quantities in case
the precision measures of the

the observations are indirect
involved a knowledge of the weights of the unknowns, and

consequently the method of computing these weights must

first
be demonstrated.
It will
63
be assumed at present that
all
the observations are of weight unity.
METHOD OF COMPUTING THE WEIGHTS.
FIRST
Let the observations be
which M^ M,
n denote the actual observations,
and 2j, 2 2 ... zq the most probable values of the unknown
Let
quantities.
in
Then
3/
M=
&2
1?
7??
2,
M =m
... kn
the above equations give rise to the
si
fat
^2^
^2
^
(A)
By
we now form
the rule, paragraph 25,
the
NORMAL EQUATIONS
2
zl
2a
2!
2 a&
-f-
22
22
2 ai
2
*2
+
+
2 aq
2 bq -f 2
^>m
-f-
=
=
(B)
2!
2 a?
-f-
za
2 6?
q*
+2 </w =
64
Multiply the
.
by &, the second

Then
of (B)
first
and add the
results.
...
+ Q^V +
-f-
&S am
-f-
of
(B) by
Qqlaq)
...
2 bm
=
of
Let
zl
in
19
ZH ... z q
Qq
QZ, ...
shall
be determined so that the coefficient

be unity, and the coefficients of z a
(C)
each equal to zero.
2ai
That
is,
let
=
...
Qq *q*
Then (C) becomes

zv
Qâm
Equations (D)
(B),
2 am
ftSJm
may be
in
(E)
derived from the normal equations
them we replace z z 2 ...

1, and 2&m, 2cm, ...
by
if
Qq X qm
t,
,... Qq
by Q^
by zero. Hence
,
2qm
the solution of the normal equations with these changes will

give the values of Q 19 (> 2 , ... Q q
.
TO SHOW THAT Qi
Expanding the
IS
THE RECIPROCAL OF THE WEIGHT OF XL
coefficients of
(E) we have
65
=
m m
or collecting the coefficients of
ly
2,
mn
(E')
=
For convenience
ai
a,
in writing
we
will let
=
=
.
Q q qn
=
Multiply the
and add the
first
Multiply the
first
results.
(F)
by
a l9
the second
first
of
(D)
by
b ly
the second
by
a2
Then
results.
=
and add the
of
(G)
by
of
(F)
Then
(H)
by
6 2 , ...
66
ca
=
=
2?a
Likewise
Multiply the
and add the
=
=
of
(D)
(H')
of
by
(F)
the second
04,
by
o^,
This gives
results.
2a 2
Let
Let
Let
first
by second
&S Oa
Q,
ftS Aa
Q g ^ ga
and
by (H)
(H')
.
(I)
ftg,
be the mean error of an observation of weight unity.

be the mean error of z^
be the weight of
ft
Zl
The mean
mean errors
equal to
m^ m
errors of
M, M^
of
z.
2,
.
mn
are the same as the
and each
is
accordingly
Therefore from (G), by (69)
ft.
*y
But by (48)
Comparing
'
ft zi
this
(i)
(J)
with
Pzi
J)
we
see at once that
(K)
Therefore, for the First Method of computing the weights

we have the following
:

Rule
77.
for
In
I.
normal equation for
the
the absolute term
and
Sara,
in
zl
67
write
the other equations
zero for each of the absolute terms 2 bm, 2 cm, ... 2 qm.
The value of z l found from these equations, is the reciprocal of the weight of the value of z l obtained by the solution
of the normal equations.

To find the weights of z%, z 8
?/wy, forming a corresponding
unknown.
,
proceed in a similar
of equations for each
g,
set
SECOND METHOD OF COMPUTING THE WEIGHTS.

Write equations (B) of paragraph 76
78.
z-i
zl
+
2 ab +
2
zz
z2
2 ab
2
-f-
... *q 2 aq
-\-
2 bq
in the
form
am
2 bm
==
(L)
Then
in the solution
by the preceding method, equation
(E) becomes
0i^
C.l?+
0,0
(M)
in which, as was proved in (K), Q

is the reciprocal of the
weight of Zi. Whatever method of elimination is employed
in the solution of the normal equations, the coefficient of
in
the value of
must necessarily be always the same.
Hence we have
Write A, B,
79. Rule II.
Q instead of zero in the
second members of the normal equations and carry out
(//> ir solution
in any convenient /'">/.
Then the most
.
z
are given by those
probable values of Zj, z 2 ,
in the results v/V^VA are independent of A, 7?.
.
68
reciprocal of the coefficient of A

The weight of 2 2 is the reciprocal of
of B in the value of z 2 etc., etc.
The weight of
is the
in the value of z^
the coefficient
THIRD METHOD OF COMPUTING THE WEIGHTS.

.
80.
From
the second, third,

. .
equations of (L) find
2 2 , 2 8 , ... z
in terras of z
and substitute
q
.
the values of
in the first of
Then
without reduction.
(L)
the
first
of
(L) becomes
Rz^
Where
is
T+ A
the
sum
which
T
-
is
ft
shown
J?,
<7,
Dividing through by
the most probable value of
in deriving the
...
of all the numerical quantities result-
ing from the substitutions.
in
terms in
2 1?
jR,
and, as was
second method,
!-'.*
R
From
81.
Rule
III.
the
Substitute in the
za , z 8 ,
of
remaining equations.
or introducing
any
(O)
once
this follows at
the values
zq
normal equation for
in terms of
Then
z^
as
zl
found from
before freeing offractions

reduction factor, the coefficient of ^

is the
in this equation
weight of the value of
zt
69
obtained
in the solution.
in a similar
z
To find the weights of 2 2 2 3
q proceed
the normal equations for each of these unknowns.
.
way with
For the solution
an example by the three different
of
im-thods see paragraph 84.
THE MEAN ERROR OF AN OBSERVATION.

82. The next step will be to derive /*, the mean error
of an observation of weight unity.
In the following demonstration the equations referred to by letters are those in
paragraphs 76 to 81.
Let the real values of
2!
-f
and substituting
z t , 22 ,
22
!,
...
JBj,
be
Xq
(A) we have
in
xq )
+ m^ =
A2
(P)
where A 1? A 2
Or of
zl
An
of
results.
xâ^
are the real errors of
J^ M
22
-\-
xâb
first
of
(P) by
t,
the second
by
a2
...
This gives
-f-
But by the
J!/i, J!/2,
mn
tlie first
Multiply
and add the
-\-
TWj, 771 2 >
+
-\-
(B)
equal to zero, and therefore
...
xq
the
2 aq
-\-
âq
first line
^ am
2aA
in this equation is
70
+ x 2 aq 2 aA =
x 2 &?
2i +
2 *A =
2 a* -f
.........
+x
x 2 a2
l
Also
2 al
a-2
2 a?
a-,
2 fy
xq 2 ? 2
2 ?A
(Q)
These being of the same form as the normal equations

(B), the value of x resulting from their solution will be
of the same form as that of z l in the solution of those equaA for m. From
tions, with only the substitution of
(G) we
shall therefore
ajAi
Multiply the
i\j,
Then
2 fly
Also
2?y
=
=
=
=
zl
2 a2
+z
2 ab
by the
-f-
first
of
Substituting these values in the above,
2 my
Now
.
(z
the second
a^
by
Say
2 Ay
and multiplying the first of (A) by

and adding the results
a2
.
(R)
(P) by v^ the second of (P) by
results.
=
.
... On& n
cujAjj
of
first
and add the
have
multiply the
and add the
first
results.
=
of
2 aq
2 am
(B).
we
find that
2 AU
(A)
-|-
by v
This shows that
(S)
the second by

2 av
zx
z2
2 mu
2?y
2 yu
71
a
and as above,
= 2y
Say
this result
Combining
with
we have
(S)
2 my
...
2y 2
2 Ay
(T)
Next multiply the first of (A) by A

the second by
add the results and compare with (T). This gives
t,
Aa
zl
2 aA
-f-
22
2 #A -f ... zg 2 ?A
2w 2
2 Aw
And
finally,
by A 2
-f-
i=
multiplying the
first
and adding the
results,
(U)
(P) by A 1? the second
of
we have
zl
a?!
S A
Therefore from
a?!
a? a
2 Â
S'A
2 A2
(U)
2 A
a? 2
2 *A
xq 2 ?A
2A 2
...
We
jc.j
must now find the mean values of the terms X

2 #A, ... a* g S Â. Expanding 2 A,
2 aA
from (R)
jcj
=
=
a x Ai
ai
+
+
a2 A 2
a 2a 2 A 2 2
A2
+
+
aA n
Â^
Multiplying,
!
2 aA
aÂi 2
-[-
terms
in
AjA 2 AxAs,
,
a anA
tt
(V)
72
In the most probable case these product terms vanish, and

a
2
the mean value ^\ we
substituting for A^, A 2 , . . . A n
have
a;2
Similarly
From (W)
=
=
=
2 #A
83.
may
By
2
/u
^+s
=
=
rz
2 Oa
q being the number of unknowns,
then,
V-z
2
2
/x
.6745^
method
ViT^
(>
()
V/i^r^
Pz(n
q)
similar to that used in paragraph 60,
we
derive
a.d.
y
(7.3)
q)
Ztf
84.
Example.
In
illustration
of
(76)
q)
the
take the example in paragraph 24, where
above
we found
processes
for
73
-
2i
22
1.7
za
Zl
z,
z8
22
--
_
--
2.4
1.0
3.0
;{
=
=
=
=
vl
vt
.,
v,
V4
and for
NORMAL EQUATIONS
-
2 zl
2 22
2 2X
3 22
2.3
=
=
(b)
3 z 3 -- 0.4
==
(c)
0.7
23
-
zl
(a)
SOLUTION FOR THE WEIGHTS BY THE FIRST METHOD.
For finding the weight

would be written
2 22
2 zl
-
2i
of
2t
the above normal equations
28
3 22
+82,
zi
=0
=0
(a')
(b')
(c')
Solve for z l
3
62
(a')
(V)
-62 -32 -3 =
2
5 zl
- -
6 22
6 22
2l
zi
(c')
3 28
3
z,
*i
At

2
2l
2 2t
of
22
2 22
=
=
=
=
= y
(d')
the equations are

23
1
2,
+32.
=0
=0
(a")
(b")
(c")
74
Solve for
3
z,
6 zt
(a")
(c")
Substitute in
3 g
z,
za
(b")
z,
.-.
z8
of
2^
--
2za
2 zx
3 z2
(d")
the equations are
=0
=0
z4
z,
=
=
jt> 2 ,
zt
3 z8
--
--
=
=
=
(a"')
(b'")
(c'")
Solve for z
from (b'")
zx
substitute in (a"')
2 sa
(c"')
--
3 z8
6s8
3z8
Zi
2s
^ =
(d-j
SOLUTION BY THE SECOND METHOD.
The normal equations

in the
following form
2z,
2zj
z,
will
now be
modified so as to appear
--
2zt
3 z,
z,
3 z8
--
0.7
2.3
0.4
= A
= B
= C
(a)
(b)
(c)

Solving,
3
(a)
2i
_2i
(c)_
5 zl
6 2 a -- 3
--
--
2.1
--
0.4
=
=
=
4.6
-7.1
21
32,
6 2 2 -- 2.5
4- 6 2 2
2t
2,
and
7.1
==
_
+
27?
(7
3^ + 2^+tf
Xl
75
(8)
(d)
Substituting (S) in (b)
3 22
16.5
22
5.5
-f6^
and
5J?
jt>^
2(7
-3
(e)
Substituting (8) in (c)
32 8
7.5
28
2.5
3JL
and
2.B
zs
2<7
(f)
SOLUTION BY THE THIRD METHOD.
The normal equations

2 2t
2 z1
zl
are
2 22
-\-
3 22
now taken
28
in their original form.
0.7
2.3
3 2,
0.4
=
=
=
(a)
(b)
(c)
76
To
obtain
and
Zj
weight we proceed as follows
its
from (c)
z,
from (b)
za
'6
= ^
+
o
Substitute in (a)
2%
33
_* % _i?_*_:i_o.7
3
3
=
Collecting terms,
Zl
For
3X
and
7.1
+ (c)
- 6z
2l
Zj
-6
--
--
12
Substitute in (b)
za
(d)
za
3 zt
2.3
3.3
z,
0.5
1.0
Collecting terms,
2.5
za
Zl
za
(a)
For
and
5.5
-3
/>
(e)
z,
(a)
+ 2x(b)
2z, -- 3
z,
--
6.7
Substitute in (c)
---
zt
6.7
IT
z,
--
0.4
COMPUTA HO* OF THE PRECISION MEASURES.

3
z8
z,
and
2.5
It is evident that the three
same
results
=0
7.5
Collecting terms,
and that the work
is
77
Zi
(f)
methods give identically the

about the same in each case.

Substituting the values found for
observations equations (A), we have
7.1
7.1
5.5
5.5
+ 2.5
1.0
3.0
5.5
2.5
(72)
By
(74)
By
(73)
2.4
4,
,,
Mz ,
=
= +
=
vl
v2
va
v4
q =.
zs
.1
.01
.1
.01
.1
.01
.0
.00
=
=
=
=
=
.17
.6745;*
Ll
in
3.
and
.03
In this example
By
=
=
=
=
1.7
2.5
z ly z 2
.12
.30
rw
.20
.U
rza
.10
.03
By
</.^.
(75)
r
=
.8453
=
a.rf.
.16
.13
the
vf
u22
v3 2
v42
78
Observations of Unequal Weights. If the observations

all of the same weight the formulas and operations
are merely modified in the usual manner and equations (72)
85.
are not
and (75) take the more general form
(77)
a.d.
-S v V
V n(n
(78)
q)
CONDITIONED OBSERVATIONS.
Suppose there are given n observations, n condiThen by paragraph 33,
tions and q unknown quantities.
the method of solution is to eliminate n' unknowns between
"
and " Condition " equations, leaving
the " Observation
1
86.
"
"
independent unknowns in the Normal equations.

Consequently formula (77) now applies to this case and it
q
n'
would be written
=
al80
__
n q
(79)
n'
*W*
^ = J~
y
Pz(n - q +
(80)
n')
n' unknown quantities can be

The weights of the q
found by any one of the three methods already given and
then the mean error of each unknown may be computed by
using formula (80). If the mean errors of the n' quantities
that were first eliminated is wanted, their weights must be
determined by eliminating a different set of n quantities

from the original observation equations and solving the neces1
sary seta of equations for these weights.
The
first
method of

solution for the weights
as the
unknowns have already been found.
actual values of the

87.
would perhaps be best here
79
Given the
Example.
=
=
=
=
3.0
22
Zl -}.
,+
-
z2
Zl
1.5
2.2
3.4
28
weight
"3
"2
and the
CONDITION EQUATION
--
zs
To
mean
Eliminating
Zl
z,
(B)
of
z 1} z 2 ,
and
23,
and
between (A) and (B) there remains
z3
_J_
22
3.0
--
22
1.0
2.2
2.9
we have
the
z2
2l -}-
these
0.5
errors.
z2
From
--
most probable values
find the
also their
22
=
=
=
=
weight
"3
"2
NORMAL EQUATIONS
7
2l
_
-f-
22
10
_
-
4.8
19.4
Solving,
2l
z2
from (B)
z3
=
=
0.98
/? Z1
2.04
jo 2l
2.54
=
=
6.9
9.9
(D)
80
Now
eliminating za between (A) and (B),

Zl _J- Zg
3.5
zs
1.4
2l
2l
From
these
_L.
we
z8
Z8
__ 3.4
=
=
=
=
2.7
derive the
new
weight
we
find
"3
"2
set of
NORMAL EQUATIONS
7 zl
and solving for
zt
z3
10 z 3
=
=
4.3
24.4
z8
z8
2.54
Substituting the values of z^

the residuals
za
2s
an(i
z2
9.9
i
(G)
n equations (A),
we have
v2
pv*
.02
.0004
.0004
.06
.0036
.0144
.16
.0256
.0768
.12
.0144
.0288
.1204
Since in this example
By
4,
n'
(79)
08
=
1,
2jw 2
3,

The
first
mean
significant figure in the
errors
81
so large
is
not worth while to retain the second place as usual.

If desired, the probable errors and average deviations can
now be computed by the usual formulas.
that
it is
88.
made
In case the observations are
directly
upon the
values of several quantities subject to certain conditions,

and equation (79) reduces to
have n
,
we
v^
from which the mean error of any observation may
be computed from its weight.
89.
Taking the example
Example.
in
at
once
paragraph 34 on
we had for
the measurement of the angles of a quadrilateral,
=
=
=
=
z1
z2
z3
24
weight 3
"2
"2
1
for the
CONDITION EQUATION
z2
z3
z*
58
(B)
and for the
NORMAL EQUATIONS
4
z,
+
+
3 z2
2l
_j_
22
21
z2
+
+
+
28
z3
3 28
+
+
+
58
58
58
=
=
=
(C)
Solving these equations for the values of the

quantities
and
we have
also for their weights,
Zi
z2
23
=
=
8.3
12.4
12.4
p
p
p
zi
st
Zi
=
=
=
3.5
2.5
2.5
unknown
82
and forming a new
set of
normals containing
we
24 ,
find,
on
solving for that quantity,
z4
These values of
2,,
24.9
Pz4
z 2 , z 8 , z4
residuals of the observations,
1.75
are also in this case the
and therefore
to
compute the
precision measures we have
n'
By
v2
pv*
8.3
68.9
206.7
12.4
153.8
307.6
12.4
153.8
307.6
24.9
620.0
620.0
1441.9
(81)
,*
- =
We
==
"
38
20
rzi
13
24
r^
r Z9
29
r^
20
16
shall accordingly write for the most probable values of

the angles of the quadrilateral
CHAPTER
V.
THE DISTRIBUTION OF ERRORS.
90. Having developed the processes for the adjustment of
observations according to the Method of Least Squares, it
will now be interesting to show how closely the distribution
of errors
found in actual practice corresponds to the theoupon which our methods of solution are
retical distribution
based.
By formula (36), the probability that the error of a single

observation will be numerically less than a is
P =
~ fV^Vx
VTT
Let
t
x,
and when
stituting in
dt
..
Table
I.
h dx.
a,
Also when
ha
x
.
0,
Sub-
(82),
P =
Values of
(82)
Jo
p
4= C 'e-*dt
for values of the
argument
(83J
are given in
Also for any series of observations this quantity
will represent the fraction of the entire number

should have errors less than the amount a. Hence if
which
is
multiplied by the whole number of observations the result

will be the number of errors which should be less than the
limit
a.
84
91.
Example. Forty measurements on the diameter of

ring were made by Bessel, with the following
Saturn's
results
85
The fourth column contains
the differences between the sucand

P,
by multiplying each of these
differences by 40, the number of observations, we have the
quantities in column five, which are the numbers of errors
cessive values
of
that according to the theory should fall within the corresponding limits. Column six shows the actual number of
residuals occurring
between these
limits.
86
with the question of the detection of a mistake or constant

error, which a consideration of the circumstances of the
observations or of the instruments might reveal, but it is
assumed that there is nothing whatever to guide us except
the mere fact of the unusual size of the residuals of the
observations under discussion.
To
reject an observation
differs considerably from the others is

entirely unjustifiable, while to retain it without any investigation is a neglect of the evidence furnished by the observa-
merely because
it
tions themselves.
The adoption of any rigid criterion based upon the magnitude of the residuals is perhaps more satisfactory from a
mathematical standpoint than from that of a practical observer,
and some of the latter are of the opinion that no observation

should be rejected entirely, even the most widely discordant
ones being given a certain weight. In this latter case,
however, the Theory of Probability will furnish a guide as to
the proper weights to assign to the different observations.
Of the various criteria that have been proposed, that
developed by Pierce (see Chauvenet, page 558)
is
perhaps
the most complete.

The derivation and application of this
criterion is, however, somewhat long and complicated, and
for all ordinary cases the following simple
practically as good results.
93.
Criterion for the
methods
will give
Rejection of a Single Doubtful
was shown in (83) that in a series of n

observations the number of errors numerically less than a
should be nP, and therefore the number of errors greater
than a should be
Observation.
It
- nP
n(l
- P)
(84)
If the value of the expression in (84) is less than one-half,

the occurrence of an error of magnitude a will have a
greater probability against it than for it, and hence the observation corresponding may be rejected.
87
Accordingly the limit of rejection, a, of a single doubtful

is obtained from the equation
observation
P =
or
94.
Example.
angle
are made,
rejected
In
2 t&
(85)
Fifteen observations on the value of an
Ought any
of
the observations to be
88
all
Using
M =
9
the observations
we
24'.20
find
t/ 4 256
-
.6745
.37
14
By
-By Table
As
90
(85),
.967
30
I,
the residual
3.17
1.40
.-.
larger than
is
1.17
we
a,
reject the
last observation.
From the remaining observations we now compute

mean value and a new set of residuals. And we find
'
By
24'.31
The
I,
third observation
From
.-.
V
V
.964
.84
.6745
we
y i^.
find
.962
26
.27
accordingly be rejected.
r"
24'.23
new
13
the thirteen observations that remain
M =
"
Q
3.11
may
.6745
P =
(85),
By Table
r'
3.08
.-.
.68
Therefore no more observations are to be rejected.
.22
The Huge Error.
95.
observations
is
In
89
where the number of

a simple and safe criterion
cases
not unusually large,
for the rejection of a doubtful observation

of the " Huge Error."
is
found
in tin
DM
an error of such a magnitude that 999 out of every

1000 errors are less than it and only 1 as large as or greater
This
than
is
it.
Therefore the probability that the error of any given

"
is .999, and
observation will be less than the " Huge Error
from Table
I,
when
Then
in
Huge Error
any limited
also,
=
=
=
4.9
4.9
3,3
|i
4.1
a.d.
series of observations,
greater than the huge error

observation corresponding.
See
.999,
Holman, page 30
is
found,
(86)
if
we should
an error
reject the
Wright, page 131.
CONSTANT ERRORS.
96. Throughout our discussion of the methods of adjusting
observations so as to obtain from them the most probable
values of the unknown quantities, all constant errors are
supposed to have been eliminated before the Method of

Least Squares is applied in deducing the results.
If this is not done, and each observation is
subject to the
same constant
error, the final result will be affected by an

in short, the Method of Least Squares is
and
equal amount,
not capable of removing or reducing the effect of errors of
this kind.
is
All that
to reduce to a
is
accomplished
minimum
l.y
the use of the nu-thod
the effect of the Accidental Errors.
90
Hence it will be seen that although by increasing the

number of observations of a given kind we may increase the
precision, that is, reduce the probable error, of our final
result as much as we choose, yet we do not in this way
necessarily increase the accuracy of the determination.
But if the unknowns can be determined in several ways, or

under a variety of different circumstances, with various
instruments, or by different observers, then it is most probable that the constant errors of the different sets of measure-
ments
be grouped about the true values of the unknowns

according to the exponential law of error. Accordingly a
combination of such observations will enable us to increase
will
not only the precision, but also the accuracy of the final
result, the constant errors of the different sets tending to
cancel each other in the same
way
that the accidental errors
of a single .set do.

It is for this reason that determinations of a quantity from
observations made in a variety of ways are more valuable
than those obtained merely from different

ments of the same kind.
sets of
measure-
The probability of the existence of a constant error

often be expressed in the following manner.
97.
may
Example.
standard 100
ohm
coil
is
compared with
a Wheatstone's bridge and the mean result found to be

100.90
To find the probability that there is an error
0.20.
in the bridge
Suppose
between
the result
observation, and we
-(-
0.80 and
100.90
-j-
0.20
is
1.50 ohms.
treated as a single
by an application of (83) the probof

the
error
observation is numerically less
that
this
ability
than 0.60 ohms.
Here
Hence, as far as
bility that
find
100.90
3.00
.-.
P =
.957
.20
shown by the
ohms is within
is
observations, the proba0.60 ohms of the true
value
i>
th:it
and
is
But since
.957.
it is
known
91
that the true resistance
00 ohms, it follows that there is the same probability

there is a constant error in the bridge between -|- 0.30
1.50 ohms.
Combination of Determinations having Different Constant Errors. In case two or more determinations of a
quantity, together with their probable errors, are obtained,
-f-
98.
method of combining them so as to secure the best final

was considered in paragraph 56, and in Example C,
rraph 57. But it was there assumed that all the results
were subject to the same constant errors, while if this is not
tlu-
result
true the probable errors of the separate determinations bear

relation to their weights, and accordingly in such cases
no
another process must be adopted.

To determine whether the different measurements
fairly
be considered to have the same constant errors
proceed as follows
Let the determinations of the quantity
may
we may
and
let the difference
Then
(a)
rz
(b)
the probable error of
E
If
MI
JtT.
between these
is
of such a
M be
be
MI
J[/i
results
is
(c)
by (66),
= yV +
ra
(d)
magnitude that an accidental error as
great as it may reasonably be expected, we may assume that

the constant errors of (a) and (b) are the same, and proceed as in Example C, paragraph 57.
But if the probability of making two determinations which
differ
J/!
by the amount d
and
MI
to
is
very small, we had best consider
have the same weight, provided there are no
92
special reasons for regarding one better than the other.

final value ot
will then be the arithmetical mean of
and MS, and
its
An
measured by a theodolite
is
angle
transit with results
By
By
What
MI
probable error will be found by (53).
Example A.
99.
and by a
The
Theodolite,
24
13'
36".0
Transit,
24
13'
24"
3".l
14"
the most probable value of the angle and

able error?
is
its
prob-
Referring to the preceding paragraph,
rt
rz
3.1,
and the probable error of d
3.1
v/
14,
12,
is
14 2
14
Then from Table I the probability that the accidental

error of a determination will be at least as large as 12 is
found from
That
is,
i?
.86
,.
- P
.57
14
there
is
more than an even chance
that
two such
determinations of the angle will differ by as much as 12.

Hence it is fair to assume that the two determinations are
not affected by constant errors of different magnitudes, and
they would be combined as in Example C, paragraph 57.
Example B.
Suppose the zenith distance, JtfJ of a

is found to be
observed at two different culminations,

J/J
j*fa
What
is
Here
and for
=
=
14
53'
12".10
0".30
14
53'
14".30
0".50
the best final value ?
=
- =
r
2.2,
It
=
.58
=
3.8,
v^9
1
+
-
=
P =
.25
.58
.01
star,
93
tin* diffrn-nre in the two deterarridental

due
to
errors, will be as large as 2.2 is
minations,
one
in
a
hundred.
It
is
to be concluded then that the
only
constant errors of observations at the two culminations differ
by about 2.2, and as there is nothing to show that one measurement is more accurate than the other we will give them
both the same weight and take the mean. Then the best
Therefore the chance that
value for the zenith distance
M =
For
By
"
JU
=
=
ra
(53)
14
53'
is
13".20
0".74
18> 10
14
53'
14
53'
13".20
.6745
\J
^^
For a more extended treatment of this

The Theory of Errors and Method
chap.
'
14
.74
subject see Johnson,

of
Least Squares,"
vii.
THE WEIGHTING OF OBSERVATIONS.

100.
In case the relative worth of observations
is
not
by methods already discussed, the proper weight to

each quantity in the final adjustment can only be
to
assign
determined from a full knowledge of all the circumstances of
settled
the measurements.
Even then considerable experience
in
the particular work in hand is required before the best values

The weight given to a
for these weights can be assigned.
quantity should never be considered final, but always subject

whenever new information with regard to the
Thus an observation which at first is
obtained.
is
quantity
to revision
supposed to deserve a high degree of confidence is often

found on later investigation to possess very little value, and
vice versa.
See also Wright, page 118.
94
OTHER LAWS OF ERROR.

101.
Although
in the great majority of cases the distribu-
tion of errors follows the exponential law thus far considered,

there are a few special cases in which some of the suppositions
made
in deriving that law do not hold, and hence for the

adjustment of such observations the corresponding special
laws of error must be determined.

.
For instance,
in
applying the exponential law we assume a
number
of observations, that each observation is sublarge

law of error, that small errors are more
the
same
to
ject
than
to
occur
large ones, and that positive and negative
likely
Now
it is
errors are equally probable.
easy to conceive of
cases where only positive errors can occur, or where the
probability of the occurrence of a small error may not be
greater than that of a larger one, etc. If we can determine

the different sources of error in any case and the relative
effect of each upon the quantity sought, we shall arrive at
the law of error for that particular set of observations. The
case of most common occurrence is the following.
102.
Suppose
errors between the limits
all
a and
are equally probable, and that there are no errors beyond

Then if y
these limits.
<t>(%) is the equation of the
Curve of Error, and its area is represented as in para-
graph
18,
we have
J%(0 & =
2 4>(x) C dx =
Jo
(a)
or
(b)
by the supposition made <(x) must be a constant.

Integrating and solving for <(), we have
since
<|>(x)
= -2
a
(87)
MISCKL L
To
find the
. \
M:<> i -s
TIII-:
Mean Error we have
graph 52
on KMS.
l>y <k-iinition
= Jr-a x^(x)
P*
95
as in para-
dx
f* a
a Jo
(88)
7?
The Probable Error
is
derived from the equation
Cr
a Jo
(89)
T
Finally, for the
Average Deviation we have

a.d.
=
=
Ca x
!
J -a
<j>(x)
ax
a
j
-I IC x dx
a***
j
And
the Curve of Error has the form
(90)
96
That the average deviation and probable error are in this

a may also be seen from the defini-
case equal to one half of

tion of these quantities.
Example. In taking a logarithm from a four place table,

what is the probable error of the mantissa ?
In this case the maximum error is .00005, and all
errors between
.00005 and .00005 are equally probable.
Therefore
r
103.
The only other
.000025
special case of
common
occurrence
is
which the error of a quantity is due to two sources,

each of which can with the same probability assume all
a.
Here it may be shown that
values between a and
the curve of error consists of two straight lines whose equathat in
tions are
2a
y
Also
|i"
~+
2a
oc
y
r
For a more extended discussion
(2
\/2")
oc
(92)
of special laws of error, see
Wright, paragraphs 31 to 39. See also Example 151.

104. Contradictory Observations. Suppose three observaIt is obvious that
tions of a quantity give results 55, 56, 91.
to take the arithmetical mean as the most probable value in
this case would be contrary to the evidence furnished by the
measurements, while in so small a number of observations it

would not be allowable to reject the value 91 entirely. With
such a series of observations no satisfactory solution can be
obtained, but probably the best thing to do would be to take
the value which has as
has greater, or 56.
many
observations less than
it
as
it
CHAPTER
METHOD OF
GAUSS'S
105.
Method
The most
VI.
SUBSTITUTION.
laborious part of the application of the
of Least
Squares to the adjustment of observations

consists in the formation and solution of the "Normal Equations," and this labor increases enormously with increase in
the
number
of observations, of unknowns and of conditions.

unusual to find that the adjustment of a single
It is not at all
even with all the aid

from tables of logarithms, of squares, of
products, and of reciprocals, and also from the use of calcuIn general the computations can be perlating machines.
formed more rapidly and with less fatigue by using a machine
set of observations takes several weeks,
that can be obtained
or a table of products than by using logarithms, but a combination of methods is often desirable.
106.
In any case however
that the formation
of the utmost importance

of the normal equations
it is
and solution
should be effected in a systematic way, and that as far as

possible checks on the numerical work be carried along in
the computations.
By a slight amount of additional work
checks upon the results at successive stages of the solution

may be obtained by means indicated in the following demonstrations, while the most satisfactory form for the solution
of the normal equations is given by the "Method of Substitution" proposed by Gauss.
This method will now be explained and it is to be observed
;
that
it is
brackets
kind
is
customary in this subject to enclose a quantity in
when
to
the
sum
be denoted.
For the sake
of
number of quantities of the same

Thus [##] means the same as âb.
of a
simplicity in
assumed that the observations are
demonstration
all
it will be
reduced to weight unity.
98
107
If,
Checks on the Formation of the Normal Equations.
we
as in paragraph 76,
take for
(A)
we
shall
have for
NORMAL EQUATIONS
[aa]
3l
+
+
22
[aft]
[a?]
2^
+
+
*+...
.........
[ft^] 2,
[ftft]
[qq]
Let
i
*i
Si
+
=
[am]
[ftro]
=
=
(B)
[?
m]
si
(C)
[]
Multiplying the
and adding, there
[am]
[*]
[?]
of
(C) by
first
+ [] =
m
l9
[]
the second
(D)
by m,,
results
+ [6m] +
[gm]
+ [mm] = [m]
(93)
Next multiplying each of equations (Cj by its a and

and adding, etc., we have
adding, and then each by its
ft
GAUSS'S
[am~\
[a6]
[66]
Equation (93) will be
99
[60]
the absolute terras in
satisfied if
the normal equations are correct, and equations (94) when

the coefficients of the unknown quantities are correct. These
check the formation of the normal equations.

108.
The Reduced Normal Equations and the Elimination
Equations.
The value of
zt
derived from the
first
'
in terms of the
of equations
[aft]
2
[aa]*
"
remaining unknowns,
(B),
[ac]
8
is
"
'
[aa]*
Substituting this in the remaining
become
And
letting
,
1]
[M]
[oa]
[am]
[âj
equations, they
100
the above equations take the following form, which, being

the same as that of the original normal equations, they are
called the
NORMAL EQUATIONS.
FIRST REDUCED
[W,
1] 2,
&clz
An
1] z 2
1] z 8
[ftc,
cclz
+ \bm, 1] =
...clz
ml =
.
[fy, 1] zq
+ [c0, 1J +
f.
. . .
[??, 1] s,
+ [0m, 1] =
inspection of equations (F) will render
a rule for writing out any one of them.

by means of the first of equations
Now
it
easy to form
(G), eliminating
from each of the others in the same way that z l was

eliminated from the normal equations, there results the
z2
SECOND REDUCED NORMAL EQUATIONS

[cc, 2] z 3
[eg, 2] z a -f-
[c0, 2] z q
[cm, 2]
[gg, 2] z q
-f-
[gra, 2]
-/l
In which
[cc, 2J
[ cc
lj]
(i)
Continuing this process we shall
finally arrive at the single
equation
from which the value of
zq
is
determined.
GAUSS'S
The value
101
will then be obtained
by substituting the
preceding >H of
z
obtained
from the
until
and
so
is
l
on,
equations,
finally
first of the original normal equations.
The equations from
which the unknowns are actually determined are then the
of
_l
numerical value of z q
the
in
of the
first
following, called the
ELIMINATION EQUATIONS.
Itr ft] Zi
[bb,
It
\_al>~\
1]
may be
*,+...
lag] z q
[bq,
\]
zq
[am]
[bm,
1]
.........
seen from the rule
in
=
=
(9.5)
paragraph
81
that
the weight of z g , and the weight of any

\_qq, q
1]
unknown might be found at the same time as its value by
is
it the last in the order of elimination, but except in

the weights had best be obtained by the general
cases
special
of
paragraph 115.
process
making
of the Normal Equations.

first of the Observation Equations (A) ly
the
Multiplying
and adding the results, we have
the second by ra 2
Check on the Solution
109.
t ,
[mo]
But
[am]
Zj_
in equation
-|-
(T),
[bm]
-\-
paragraph 82,
= [vv]. Therefore
[ray]
[bm]
[vv] = [am]
Zi
Z2
z2
-f
\_qrn]
it
[nun]
was shown
[qm]
from the
Substituting in this the value of z l

get the result
that
[mm]
first
of (95),
[m;//, 1]
we
[vv]
in
which
[bm, 1]
ga
+ [cm, 1] +
z3
[,ym, 1] z q
102
[6m, 1]
\_ab~]
[am]
(K)
_
[mm,
1]
[mm]
fam"]
being similar in form to equations (F).

Next eliminating 2 2 in a like manner,
[wt>]
and continuing
[cm, 2]
28
[?m, 2]
this process it finally
[w]
[am]
we
2<?
get
[mm,
2]
appears that
[mm,q]
[96]
In computing
110. Arrangement of the Computations.
the coefficients that appear in the "Auxiliary" or "Reduced
Normal Equations" it is most convenient to arrange the work
The arrangement of the solution will be

an example containing four unknowns but it
will be evident that the process can be extended to cover any
in tabular form.
illustrated for
case.
Let the Observation and Normal Equations be represented

by equations (A) and (B), there being only four unknowns
and arrange a table as on the next page. In this
2,, 2 2 2 8 2 4
scheme of solution the upper lines of the rows in the first
,
compartment contain all the quantities that appear in the

Normal Equations, together with [mm] and the quantities
in the column headed 8 which are used in checking the
The
results in accordance with equations (93) and (94).
other compartments contain the corresponding quantities
for the Reduced Normal Equations, and the first line in
each compartment gives the coefficients in the Elimination
Equations.
GAUSS'S
SCHEME A.
M
log[aa]
SOLUTION OF THE NORMAL EQUATIONS.
103
104
The logarithms
compartment are
of the quantities in the first row of each

also written in, and from these by proper
subtractions are obtained the logarithms in the margin, where
[*]
[c]
[am]
Am
[oF'
[5aJ'
*
(L)
Now
in each
compartment adding the logarithms at the

row of that com-
to each of the logarithms in the first
margin
partment we obtain the corresponding logarithms written in
the other rows. The numbers represented by these logarithms
are next written above them, and if each of these quantities
is then subtracted from the one above it the result will be
the corresponding quantity in the compartment below. Some
of the squares in each compartment are left vacant as the
quantities belonging to them have already appeared above.
111. Application of Checks. Also, by (93) and (94), in
compartment the quantities in the first lines of the

column should be equal to the sura of all the quantities
in the first lines of the corresponding rows plus the quantities
Similar
similarly situated above the first terms of the rows.
if
from
the
for
in
each
will
checks
compartment;
apply
the
first
last
we subtract the product

by Ab we have
second of equations (94)

equation multiplied
[M, 1]
[Ac, 1]
[ftro,
1]
of the first
[fc, 1]
(97)
In the same manner we may show that a corresponding

check holds throughout, so that finally we shall have
[mm,
4]
[w,
4]
(98)
GAUSS'S METHOD OF SUBSTITUTION.

The
last
of the table
compartment
check and the value of [uu]
final
tion (96).
If the multiplications
in
is
105
added to give
this
accordance witli equa-
and divisions are simple or
if
a table
of squares or products or a computing machine is used the

logarithms will of course be omitted from the scheme of
solution.
112.
mation
Example.
In order to illustrate the systematic for-
of the coefficients that
appear in the Normal Equations

by the above method we
as well as the solution of the latter

will take the
=
=
=
=
=
=
0.1
0.6
0.1
0.8
0.1
0.4
First form a table containing the coefficients in these equaand also the sums s. As a first check the sum of the
tions
quantities in
column
should be equal to the
sum
of all the
quantities in all the other columns.
I.
No.
COEFFICIENTS IN THE OBSERVATION EQUATIONS.
METHOD OF LEAST
106
From
these
we now compute
the coefficients in the Normal

and
also
the
Equations (B),
necessary quantities for the
check equations (93) and (94).
II.
No.
COEFFICIENTS IN THE NORMAL EQUATIONS.
GAUSS'S
Substituting these quantities
now
107
in the general tabular
scheme of paragraph 110 we have the results on (he following

The work of the first compartment is performed
page.
without the use of logarithms, as the numbers are simple.
The quantities in the last column should all be zero according to the check equations, what small differences there are
Ting due to the rejection of figures beyond the third place
in the decimals.
The decimal points in logarithms to which
correspond negative numbers have been replaced by the
I
letter n.
that have been made now enable us to

from an inspection of the results in this table
The demonstrations
see at once
that
2
[iw]
0.238
.007
Therefore substituting in equations
(73)
and (74) we
have
fi^
If
.047
the two values of
[],
.032
obtained in the solution had
we
[bb, 1],
[cc, 2],
#s
And
\_vv~]
should have taken the mean of the two.

If the Elimination Equations
(95) are divided by
differed at all
113.
rZi
\_dd, 3],
23
the solution for the
respectively, they
ltd**
become
Am =
+ Em =
Cm =
Dw =
unknowns can be
effected
conveniently by arranging the computations in the

illustrated on page 109.
most
manner
108
SCHEME A.
a
b
b
C
c
d
U
1/1
-m
d
*d
m
?
m
c
d
V
m
'm
m
>m
SOLUTION OF THE NORMAL EQUATIONS.

SCHEME
-
B.
109
SOLUTION OF THE ELIMINATION EQUATIONS.
110
115. The Weights of the Unknowns. In order to determine the precision measures of z 2 2 and 2 3 it would
next be necessary to compute the weights of the latter quantities.
The demonstration of the processes by which these
weights may be found will not be taken up here, as the best
method to adopt varies a good deal with the character of the
example, but a statement of the results in the general form
t,
of solution will be given.
By treating the Elimination Equations in a way similar to

that used in deriving equation (E) of paragraph 76 from
(B) of the same paragraph, we may show that
equations
Zl
Am
-f JBmai
m
+H
2,
z.
where the
a's,
/J's,
+
+
+
d a,
d a,
a,
a,
Then by an
graph (77),
it
+ J)m a =
+ 7>m & =
+ Dm =
s
(100)
73
are determined from the equations
y's,
+ J3 + C + =
-4,+JSe0l + =
A + =
Ad
Cm a
Cm ft
Cm
a,
J? d
J? c
Cd
Cd
&+& =
+& =
+ =
(101)
73
application of the principles of Rule

may be shown that
1,
para-
GAUSS'S
\\\
These equations can of course be extended to cover any

of unknown quantities, and tabular schemes for the
and of the weights
computations of the a's, /?'s, y's,
... can readily be arranged.
Pzj PZJ PZJ
For a general demonstration of these results, and also for
a discussion of special methods of solution, consult
number
Johnson,
"The Theory
of Errors and Method of Least Squares,"
chap. ix.
" Treatise on the
Wright,
Adjustment of Observations," chap. iv.
"
Chauvenet,
Spherical and Practical Astronomy," pp. 530-649.
THE METHOD OF CORRELATIVES.

The method
116.
tions
"
where
is
of
"Conditioned Observa-
adjusting
explained in paragraph 33 is perfectly general, but

there are many conditions to be satisfied the solution
apt to be very laborious.
For the case that occurs most
fre-
quently in practice, in which the observations are direct and

equal in number to the number of unknown quantities, the
"
process of solution devised by Gauss and called the Method
of
Correlatives
"
is
derived as follows
the most convenient.
Let q observations,
weights
PU PK
pq
JtfJ,
,
M^
q unknown quantities, and

of the unknowns be
apply
to
q,
of the respective
let the
v ? , are the most probable corrections to

observed values as well as in this case the
v 1? v 2 ,
the
is
be made directly upon the values
of
Where
This method
residuals of the observations.

If the n' condition equations are not linear they may be
reduced to that form by the method of paragraph 44, so that
we may assume for our
112
CONDITION EQUATIONS
+m =
m =
aq v q
b q v q -f
.........
lll
l*V*
In which the quantities m^ 7W 2 ,

the observations were exact.
It
coefficients
a, #,
IqVq
.
if
mn
mn
(A)
>
would
',
all
be zero
to be observed that the
is
are not arranged in the same order in
these equations as they are in the observation equations of

paragraph 107.
The values
must be determined so
of w 1? v 2 , ... v q ,
as to
satisfy the above equations and also by the principle of Least

Squares, so as to make
2*
Pq v g*
Corresponding with a minimum value of
+ PtVidVt +
PlVl<tl
minimum.
PqV q dv q
for all possible simultaneous values of dv^ dv z ,

is, for all values which satisfy the equations,
a l dv l
-\-
a 2 dv z
bidvi
-\-
t>zdu z -f-
a q dv q
b q dv q
.........
lidVi
I*dv 2
-[-
-h
dv q
we have
this last
(B)
.
dv q
that
=
=
(C)
obtained by differentiating equations (A).

Therefore, di-ii..tiiiL tla- lirst member of (B) by R and the
first members of
AVn .,
it will be IUT(C) by A\, xS'2 ,
r
CHsary that
=
\\hcre
X*j,
/*,,
/., an-
uiidi'tiTiniiifd
c.,,
-ilici-nts.
(D)
GAUSS'S
This last equation will be satisfied

differential in it is
made equal
â2
if
the coefficient of each
to zero, that
&2
-f-
-f-
113
is, if
... kn>l2
(103)
kâ q
All that remains therefore
is
k z b q -f ...
Jc n ,l
v lt V 2 , ... v q
to find values of
& 2 ... k n ,> which will satisfy simultaneously equations (A) and (103), and that this may be done is easily
seen from the fact that we have the same number of equations
as unknowns.
Substituting the values of v^ v z ... v q from equations
(103) inequations (A) we have the following
and
&!,
(id
ab
(ib
K*
A- 2
S^bb
~+
_ dl
kn S
_ bl
Av S
+m =
l
+m =
2
(104)
S
The
M
/?'
solution of these equations gives at once the values of

Ar n ',
which are called the "Correlatives" of the
/, ...
v q are then
Condition Equations. The values of v^ v^
found by substituting the values of the tfs in equation (103).
117. As equations (104) are of the same general form as
a set of Normal Equations, Gauss's Method of Substitution
can be advantageously employed in the solution.
.
114
When there is but a single equation
of condition the second
members
of equations (103J reduce to their first terms, and

equations (104) reduce to the single equation
k^
and the values of v l9
CLd
2,
-f
(105)
... v q in (103) become
"
a,
V*-
aa mi
p
It is
from these
results that the rules in
paragraph 35 are
derived.
118.
Suppose we have given the observations
Example.
M =
M =
l
2.02,
4.13,
r=
2.52,
2.67,
2.84,
Jf,
=
=
M.
Jf6
and
let the
weights
"
2
(a)
most probable values of the unknowns be repre-
sented by
Z,
J/i
+W
lf
2,
MI
+V
2,
26
MS
+V
(b)
Also suppose that the unknowns are subject to the conditions
z*
2.
=
=
14.0
1.5

Then expressing these conditions in terras of
by means of (a) and (b), we have the
115
the corrections
CONDITION EQUATIONS
Vi
v<t
va
+
-
Referring to paragraph 116,

n'
2,
For the purpose
m
of
v.
we
.18
.04
=
=
see that in this
0.18,
computing the
(104) we next arrange
v4
example
0.04
coefficients in equations
the following table.
116
Substituting these results in equations (104),
Cl
*l
14
"
14
14
**
*'
=
h =
&j
Solving,
'^
"
.1647
,.
.1537
Then from equations (103) we get
=
=
=
Vl
va
vs
.0549,
.0055,
v&
.0455,
.0412.
(g)
-.0329,
is
at once
ss
And by substituting these results in

values of z x , z 2 , z 3 z 4 , z &
The above
we have
(b)
we can
obtain the
the solution of
Example
70,
page 127.
EXAMPLES.
An
1.
two white
urn contains
balls.
1'
five
black
red balls and

drawn from the urn what
and what is the probability
balls, three
three balls are
different combinations
may
result,
of each ?
In a single throw with a pair of dice what
2.
probability that neither ace nor doublets will appear
Four cards are drawn from
3.
From
the
is
the
What
Of getting one
probability of getting four aces?

suit?
4.
a pack.
is
of
each
a lottery of thirty tickets, marked 1, 2, ... 30,

What is the probability that numo
four tickets are drawn.

bers
and 15
will be
among them?
145
Find the odds against the appearance of 7 or 11 in a

2
7
single throw with a pair of dice.
I toss up n coins.
6.
What is my chance of getting just
one head ?
In a single throw what are the relative chances of
7.
24 25
throwing 9 with two dice and with three dice ?
8.
From 2 n counters marked with consecutive numbers
two are drawn. What are the odds against having an even
sum ?
n n
1
In two trials with a single die what is the probability
9.
5.
throwing (a) an ace the

ace?
of
10.
first
Find the probability
more times
in three trials
of
time only?
(b) at least one
throwing doublets one or
with a pair of dice.
91
216
118
11.
Find the probability of throwing exactly three aces

125
in five trials
with a single
die.
3888
be won by the first person who
throws 5 with a die of twelve faces. What is the chance of
12.
certain stake
is
to
the sixth person ?
A and B play chess.

of three.
What
out
games
13.
four games out of the
is
wins on the average two

A's chance of winning just
80
first six ?
243
and B shoot alternately at a mark. A hits once in

1 times.
n times and B once in n
Find their chances of
first hit, and the odds in favor of B if A misses on his first
shot.
Even, n n
2
14.
In
how many
be a wager of 4 to 3 that
double five will be thrown with a pair of dice ?
30
Find the probability of throwing one and only one
16.
15.
trials will it
ace in two trials with a single die.

17.
If I
have three tickets
eight blanks,
18.
what
is
my
chance of drawing a prize
Find the probability
six trials
with a single
in a lottery of four prizes
of
throwing
and
41
55
at least four aces in
die.
23328
an average seven ships out of eight return to port.
Find the chance that out of five ships expected at least three
16121
19.
On
will return.
16384
In a lottery containing a large number of tickets,
where the prizes are to the blanks as 1 9, find the chance
20.
81 J.R
of
drawing
at least
two prizes
in five trials.
100000
EXAMPLES.
119
In a purse are ten coins, all nickels except one which

in another are ten coins, all
a five-dollar gold piece
nickels.
Nine coins are taken from the first purse and
21.
is
placed in the second, and then nine coins are taken from the
If you now had your choice
latter and placed in the former.
which purse would you take ?

A and B engage in a game
22.
What is A's chance
B's as 2:3.
games out of five ?

23.
If A's skill at a certain
in
which A's
of
winning
is
game
skill is to
at least
two
double that of B,
what are the odds against A's winning four games before B
112
131
wins two?
24.
A party of twenty-five take seats at a round table.
What are the odds against any two specified persons sitting
:
next to each other
has three shares in a lottery in which there are three

B has one share in another where
prizes and six blanks.
there is but one prize and two blanks. What are their
25.
relative chances of getting a prize ?

26.
Expand through the terms involving
A B
:
16
A 8 and & 8, the
expression
-
When
diminish
is
and
+
is
(y
f (x,
*)
does -
-,
when x and y begin
Given
27.
sion
a;
-\-
increase
^*
same
to increase at the
y) = x (a
4- y + &)
2
+ hY(a
Find the value of Iog a
= 0.1. Give the result
-f-
y)
-\-
cos
expand
or
rate
the expres-
(x
28.
10
#,
when a
1001
places and then

to seven places of significant figures, in each case without the
4.000433
aid of tables.
and b
29.
tion,
Transform
x2
if
to the
z*
new
4 x
first to five
origin,
6 y
(2,
2 z
3, 1),
11
the equa0, the
120
The equations
axes remaining parallel to the original ones.

a;'
2
of transformation are x
y
2=1 +
iC
Find the minimum value
30.
2'.
of
= 3
+y + =
x 2 -|- xy
-{-
-\-
ax
y*
25
by.
a*
(ab
y'
tf)
y and z that render a maximum

22
or a minimum the function ce 2 -|- y 2 -|- z 2 -f- a;
av/.
32.
Find the values of a? and y that render a maximum
or a minimum the expression sin x -\- sin y -f- cos (x -\- y).
Find the co-ordinates of a point the sum of the squares
33.
of whose distances from three given points, (a^, y,), (a*2 y 2 ),
Find the values of
31.
a:,
(xs y 3 ),
,
is
minimum.
a;
( a?i
-
-f-
a? 2
-I-
a; 8 )
Given the volume, a 8 of a rectangular parallelepiped,
34.
when
is a minimum.
Find the volume of the greatest rectangular
find its shape
35.
its
surface
parallelo-
piped that can be inscribed in the ellipsoid

av*
"
a2
ya
ft
2*
"
In the Physical Laboratory apparatus for illustrating
36.
the Estimation of Tenths, the reading of the micrometer head

for a certain setting is computed to be 2.3038, which may be
taken as exact. Setting the apparatus by the eye the follow-
ing readings are obtained
2.314
2.324
2.310
2.519
2.320
2.302
2.313
305
What
2.326
are the accidental and real errors and the residuals?

there
Are
any constant errors or mistakes? If then- arc
.. n^taiit
errors are they of the first, second, or third class?
<
I!
HT
do you
tell?
EXAMPLES.
1:M
Are the following observations such

application of the Method of Least Squares
37.
as to call for an
their
in
:i<lju>t-
mriit?
x
x
-\-
y
2y
-\-
z-j-w
u
-f-
2z
3ic
y
y
-\-
=
=
4z
6u
In the case of direct observations, what other quantiarithmetical mean might reasonably be
38.
besides the
ties
assumed to give plausible values of the unknowns ?

the arithmetical mean preferred to these?
39.
Find the most probable value of a quantity
Why
is
M from
the observations
216.27
216.16
216.04
216.19
216.44
.29
.43
215.99
.39
.51
.33
.09
216.23
.14
215.94
Also
test the result
40.
by finding the sum
of the residuals.
M=
216.251
In the determination of a certain wave length, Row-
land made the following observations.

able
216.58
wave
Find the most prob-
length.
4.524 4.515 4.513 4.507 4.501 4.485 4.517 4.493 4.505

.500
.508
.511
.497
.502
.519
.504
.492
4.5055
41.
0.0017
Ten measurements
results as follows.
of the density of a body gave

Find the most probable density.
9.662
9.664
9.677
9.663
.673
.659
.662
.680
9.645
.654
9.6639
0.0022
In a triangulation of the U. S. Coast Survey an angle

42.
was measured twenty-four times with results
122
116
43'
44".45
50 .55
50 .95
51
51
51
EXAMPLES.
47.
made
123
In sixty-six determinations of the velocity of light
Washington the percentage of errors

magnitudes was found to be as follows:
at
Over
of different
124
Find the altitudes in Example 49 if the observations

52.
have weights 5, 3, 6, 2, 4, respectively.
53.
Solve the example in paragraph 31, giving the observations the weights 25, 25, 4, 4, 4, 4, 4, 4, 1, respectively.
Elevation of
54.
Find the most probable values
of
P =
B
2 lt 2 2 ,
and
320.25
z8
from
the observations
zl
z2
4- 2 8
28
22
= 552.10 wt. 16
= .15 " 9
= 551.23 "4
= 551.30
4
22
21
2X
22
2l
28
=
=
552.05
"
.70
"
28
22
55.
.75 wt. 1
551.2345
In the triangulation of Lake Superior there were
measured
at station
the angles
EXAMPLES.
125
Adjust the following observations of differences
57.
in
level:
Altitude of
above
401.3
wt. 16
220.8
16
^4
"
150.2
B
B
Altitude of B
C above
72.5
"
wt. 9
222.0
"
180.7
"
In "Conditioned Observations" can the number of
58.
observations required be less than the number of unknown

must the number of conditions be less
quantities?
Why
number
than the
of
unknowns?
From
the following measurements of the angles formed

at the centre of a disk by four radial lines, find the most
59.
probable values of the angles.
104
B =
98
=
D =
C
25' 13"
13
47
86
33'
20"
70
48
23
104
Also solve giving the observations the weights
25' 2".25
5, 2, 1, 4,
respectively.
Four observations on the angle
60.
of a triangle gave
mean of 36 25' 47", two observations on B gave a

mean of 90 36' 28", and three on C gave 52 57' 57".
a
r
36 25' 44".2
7". 7
Adjust the triangle.
61.
Five angles at a station are measured, and also their
sum. The observed sum differs from the sum of the five
observed parts by the amount d. What are the adjusted
;
values of the angles ?

62.
The three angles of a spherical triangle are measured
with results
46
B =
17' 38".32
60
73
35' 16".15
7' 5".16.
Adjust the triangle, knowing that the spherical excess is

2".475.
1".6
39".3; ^
126
63.
At the station Pine Mountain the following angles
were observed between surrounding stations
:
65
66
Deepwater
Deepwater
Deakyne
Burden
Deakyne
Burden
Jocelyne
Jocelyne
11' 52".500
87
141
wt. 3
15 .553
"
24 .703
"
21
21.757
"
24
Find the most probable values of the angles.

04.
Solve Examples 55 and 56 by the method of "Con.
ditioned Observations."
65.
feet,
From
.lisa
station
C below A
720.1
wt. 3
200.3
"5
520.4
"
66.
z l9
z,,
za
z,
2z x
3z a
Zl
z8
to be 5240.1
B below A
719.7
J)
520.9
JB
wt. 3
"
=
=
=
3.0
2z 2
4.5
Zj
= .52
= .27
Au =
Pb
Ay
.11
z8
2z 2
=
=
1.0
5.1
3.8
The chemical composition
by several observers
From
known
Given the following observations, subject to the conz8

find the most probable values of
z l -\- z a
and z,.
22j
67.
is
and D.
dition
altitude
able altitudes of JB,
whose
is a
C are floats on a lake, and
signal point.
the following observations determine the most prob-
JB and
of a specimen
to be as follows
Other substances
Pb and Ag
Pb and impurities
=
=
=
was found
.09
Au
.78
Impurities
.62
Au
and
Ag
=
=
=
.39
.10
.12
these observations find the most probable composition
the specimen.
EXAMPLES.
1-27
From the following observations what are the best

68.
values of the unknowns, supposing that y and z must be
equal?
=
=
=
y
x
z
5.2
wt. 4
3.0
"
1.1
"1
=
=
wt.
4.2
" 4
2.0
In determining the difference in longitude between

various cities the results obtained were
69.
Cambridge
(2) Cambridge
(3) Cambridge
(4) Washington
(1)
(5) Cleveland
Washington
23
4K041
Cleveland
42
Columbus
Columbus
Columbus
47
14.875
27.713
46.816
12.929
23
5
Adjust these observations.

The capacity of a condenser
70.
is
It is divided into five sections, a, 5,
known
c, </, e,
wt. 30
7
"
"
"
to be 14.0 m.
and
f.
known
it is
between b and d
Find the
is 1.5 m. f.
most probable capacities of the sections from the observathat the difference
tions
a
b
c
=
=
=
2.02
wt. 3
4.13
"2
2 52
"
unknowns
in
If the
4.3
wt.
1,
=
=
2.67
wt. 7
"
2.84
subject to the condition

their adjusted values?
71.
72.
ball
1.9651
the following observations are

=: 36, what are
-f- 2y -|- 3z
5.7
wt. 4,
7.3
a bluff.
wt. 9
3.77
cannon is discharged horizontally from the top of

Observations on the time, and distance of fall of the
gave the results

t
=
=
0.5
1.0
1.5
2.0
1.2
4.0
9.1
15.0
seconds
metres
128
What
ball, will
curve, passing through the point of departure of the

represent the above observations ?
73.
An Argand burner shows the following efficiencies
with varying rates of gas consumption
:
E =
2.0
2.3
2.8
3.3
4.0
4.5
5.0
2.1
2.4
2.5
3.0
3.2
3.8
4.1
feet
Find the equation of the straight line which best repg and E. The measurements
resents the relation between
on g are without appreciable error.

74.
Observations are made upon the expansion of
alcohol with change in temperature as follows
Amyl
V
t
=
=
cm.
1.04
1.12
1.19
1.24
1.27
cu.
13.9
43.0
67.8
89.0
99.2
C. degrees
V =
expresses the law connecting the volume and temperature, find the most probable
values of
and C.
If
-f-
-\-
75.
is
In a Hooke's joint where the angle between the axes
45, x being the angular
rotation of the driver,
and y
that of the follower, from the following measurements find

the equation of a curve that will represent the relation
between x and y
x.
EXAMPLES.
129
the measurements are given below.

From them derive an
equation to express the temperature at any time of tin- year.
4.66
5.42
Mar. 6.77
Apr. 8.59
May 9.83
June 10.09
9.71
July
Aug. 9.14
Jan.
1-Vb.
7.369
+
+
-
Sept.
Oct.
Nov.
Dec.
0.9854 sin 30*

0.0100 sin 60s
0.0133 sin
8.16
6.55
5.10
4.41
2.7084 cosSOa;
0.1950 cosGOa;
90x+ 0.1783
cos 90s;
In this answer the values of x begin at the 15th of January, and represent the time in months.
77.
The law connecting the time of vibration of a pendulum with its length is assumed to be of the form, T ==. m L n
From the following observations find the most probable values
of m and n.
.
L
L
is in
=
=
12.9
11.6
10.4
9.7
5.3
4.6
164.4
132.9
107.6
93.5
28.4
20.6
centimetres,
in tenths seconds.
m
78.
=
=
0.5000
1.0044
Determine the equation of a curve which will repre-
sent the following observations
X
0.0
130
80.
At a
station
the angles between a straight line

and the direcparallel to the axis of
passing through
of four points P^
the
coordinates, (a,
Having given
tions from
find the coordinates of
Point.
P P P
P.
Coordinates.
2,
8,
),
4,
of the
are measured.
four points,
EXAMPLES.
131
84.
If twenty measurements of an angle give a result with
an A.D. of 0".38, ami it is required to timl the niiLrle so
that the A.D. shall be only 0".25, how many more observations must be made?
27
85.
From
5674
the following determinations of the area of a
the most probable area and
field find
5680
12,
4,
5685
its
probable error.
5682
3,
5678
1,
0.84
5681.41
From
the following measurements by Fizeau and

most probable value for the velocity of light
with
its
together
probable error. Measurements are in kilometers.
86.
others, find the
298000
1000
299990
200
298500
1000
300100
1000
299930
V =
87.
Two
100
299917
88
different instruments give for the value of an
angle,
34
What
is
55' 33".0
34
4".l,
55' 36".0
the best value to take for the angle
34
6".3
55' 33".9
3".4
Determinations of the difference in longitude between

Washington and Key West made on seven different days
88.
gave the results

I9
What
is
19
1 .42
O .044
1 .37
.037
1 .55
.045
1 .38
.036
1 .57
.047
1 .45
.036
the best value and
its
0'.046
l .60
probable error
r.4GO
0'.016
132
In the triangulation of Lake Ontario two different
89.
instruments gave for an angle, 74 25' 5".429

0".29 from
sixteen readings, and 74 25' 4".611
0".22 from twentyfour readings. Find the most probable value of the angle
and
probable error.
In each of Examples 39-45 find the mean and probable errors and average deviation of each observation and of
its
90.
the most probable value, using formulas from (50) to (63)
according as they apply.

91.
In Example 42 divide the observations in their order
into six groups of four observations each and compute the
of each group.
Then determine the probable error of
the first of these means: (1) considered as a single measure
mean
of four times the weight of those in Example 42 ; (2) directly

as one of six observations of equal weight; (3) as a determination from its four constituents.
0".67 ; 0".72 ; 1".00
92.
The following twenty-nine measurements on the denmade by Cavendish, give as a mean result
sity of the earth,
5.48.
What
is
the probable error of an observation ? Solve

also by taking the residual that
by the usual method and
occupies the middle position.

5.50
0.14
133
Find the probable errors and weights for a single observaand also the adjusted length of the lino
and its probable error.
741.146
0.015
of
a
95.
determinations
chronometer
correcTwenty-one
tion in each cusr,
tion
gave results
- 8.78
- 8.78
- 8.68
- 8.80
.76
.51
.63
.75
.64
.70
.90
.85
.64
.58
.78
.65
.64
.93
8.96
- 8.79
8.83
Find the probable error of the mean by using both formula*

(53) and (57), and also determine the probable error of ;i
single observation by taking the middle residual.
0.017; 0.018; 0.09
rz
1.95,
1.31,
/*o
0.40,
^=
0.27,
49.64,
0.87,
0.59.
48.81
48.76
49.53
51.56
50.38
49.84
/>
M=
In the following observations show that
96.
97.
Observations on the time of ending of a transit of
Mercury are made by different observers with a variety of

instruments and under more or less favorable circumstances.
weights assigned by the computer are as indicated, find
the best value for the time and its probable error.
If the
5* 38
23*
wt.
37
55
"
38
10
"
38
26*
38
21
38
18
wt. 3
38
38
21
38
15
"
tQ
19*
5* 38
wt. 3
2
2
19*.9
An angle is measured five times with a theodolite, and

98.
seven times with a transit, giving results
Theodolite,
31".7,
39".8,
40".7,
28".6,
3'J
.8
Transit,
32
36
38
-J!)
41
.6
.8,
.7,
.1,
.8,
35".3,
36".2
134
by the two instruments

what
is
the
most
2,
probable value of the angle ?
the mean error of the result?
If the relative values of readings
are as 3 to
What
is
Given
99.
MI = 49.64
3 M + 2 J/3
t
Jtfi
65.58
.59,
.27,
and of
M*
35.15
.93,
probable errors of 4
find the
^+^
3.69
2*
J/
0.43
The
three angles of a triangle are measured, and the

error
of each observation is r
What is the probprobable
able error of the triangle error ?
r J~%~
100.
The
101.
zenith distance of a star on the meridian
observed to be
tion of the star
What
is
z
is
tion
is
The
=
=
observation
d
z
declina-
30' 14".8
0".8.
probable error ?
40 47' 35".l
2"4.
its
of a star at
upper culmina-
times, and its zenith distance z

n' times.
If the latitude is given
what
z'),
and the probable error
at
by
an
of
the probable error of the latitude ?

horizontal force necessary to start a 100-pound
is
The
19
is
The
2".3.
observed
103.
given as
zenith distance
lower culmination
90
| (z
17' 20' .3
the latitude of the place and
L
102.
21
r,
is
0.2
weight sliding along a table is observed to be 15.5
Find
the
error
of
the
coefficient
of
friction.
probable
pounds.
If a line is measured by the continued application of
104.
a unit of measure, and r is the probable error of the placing
and reading of
the length
105.
this measure,
what
is
/ ?
If the average deviations of
Zj,
z2 , z8 ,
y/T
are a,
c,
2
a
respectively, what is the average deviation of z-f -\- z 2 -\- z 8 ?
of
a
circle is measured with result
106.
If the radius
1000.0
2.0,
expressed
Two
how
should the circumference and area be
a and b, and the included angle

252.52
a triangle are measured with results a
107.
sides,
of
.06
KX. \MTLES.
135
30".
42 13' 00"
.06 feet, C
9
25452
and its probable error?
Measurements of adjacent sides of a rectangle gave
108.
a
rz
What is the probable error of the
and b
ri,
an
and for what kind of a rectangle will this probable
b
feet,
What
300.01
the urea
is
:i,
error be
least?
tlie
If the
109.
what
'/.'/.,
them
is
measured sides of a rectangle have the same

a.d. of the diagonal determined from
the
Same
110.
If
the
a rectangle are measured in the
of
sides
manner indicated in Example 104 and found to be a and b,

what is the probable error of the area ?
111.
The correction to be applied to a chronometer is
m
Ten days later the corfound to be
12 13*.2
0*.3.
m
found to be 12
rection
What
again determined and

the mean daily rate and
21*.4
is
is
its
probable error ?
0'.820
Measurements of the compression of the
112.
0*.3.
0'.042
earth's
meridian have resulted in

.000046
294
What
the probable error of the denominator 294 ?

3.98
in a circuit is due to two
is
The current flowing
113.
sources whose electromotive forces are determined to be

ej_
200
circuit
is
e2
2,
30
1.
400
The resistance of the
3.
Find the current and its probable error.
20
114.
The
and angles
side
measured with
results
29
39'
1',
120
probable value of the angle

30 14'
115.
is
B
106
7'
of a triangle are
.06
metres,
is the most
B =
and of the side

2'.2;
0.68
What
2'.
The distance between two

measured by
and
c ?
185.35
0.15
divisions on a graduated
a micrometer.
Show that the average
136
deviation of the
mean
of
two
results
is
the same as the aver-
age deviation of a single reading.

If the weights of the determinations of three angles
116.
117.
are 3, 3, 1, respectively,
of the three angles ?
If the weight of x is />,
118.
If
A, B,
the
(7,
sum
weight of
x
x
and the weight
what
is
the weight of
what
is
the weight of
of
is
/>,
0.6
what
is
the
c*p
In Example 107, how closely must the parts be

measured in order to obtain the area within 0.5 per cent?
119.
From
120.
to be
observations on
computed by
the
What changes in g
t
of
and
able errors
1
per cent
82
in
of inertia of a cylindrical bar
d.
The
its
mass
8.000,
1.200
what
is
0.10,
to
to
to
be
12
ratio of
is
m,
length A,
error in the determination of
negligible, the precision of the determination of

If the measurements give
times that of h.
is
its
is
diameter
its
the value of
will be
obtained from measurements on
and
pendulum formula
produced by changes in I and

units, respectively, and what are the allowI
and t if g is to be determined within
The moment
121.
and
determine
is
four
48,
and
16
7J
and what should be the

most accurately
<!
-j.-irt
A
9
':
'
EXAMPLES.
122.
137
observations give for a certain quantity x the

mean error of 2, what is tin- mean error
If
value 303, with a
of the expression 3x -\- Iog 10 2 x ?

123.
The probable error of the
angle
determination of the
maximum
the
is
probable error of
?
4~ cos
124.
If the probable error of an observation on an angle
10", is there any difference between the probable error
sin
is
What
20".
is
of the function
tion
sin
sin
cos 7?
-\-
cos
-|-
-f~
sin
6Y
and of the funcsupposing A and J? are

sin
same magnitude ?
125.
Given the observations,
of the
z1
2 Zl
-2z
4- 3z 2
+
-
z3
4z 8
=
=
3z!
25!
4-
z2
-f-
4z 2
Find the most probable values of

weights and precision measures.
2l
3.541;
4-2z 8
+ 3z
10
and also their
z 1? z 2 , z 3 ,
p zt
=
=
17
29;
r tl
.024
Find the weights and precision measures of the

unknowns in Examples 48 to 57.
127.
Determine the probable errors of the constants in
126.
Examples 72 to 79, inclusive.

The length of a pendulum which beats seconds
128.
given by
I
where
I'
is
I'
-f
-- s\
I'
the length at the equator,
sin 2
is
the ratio
of
289
the centrifugal force at the equator to the weight, and s the
compression of the meridian regarded as unknown. Putting
I'
991
x,
(f
observations in different latitudes gave in
*'
V>
millimetres
the
138
following equations, from which we are to determine

8 together with their probable errors
:
+
x +
x +
x +
0.969y
0.749y
0.426y
0.095y
=
=
=
=
=
5.13
3.97
2.24
0.56
0.19
and
EXAMPLES.
139
On
the average how many observations must In- made

before an error as large as three times the mean error will
133.
occur
134.
two
In Example 46, assuming that all errors between any

way between those limits, compute the
limits fall half
average deviation and mean error of an observation and compare their ratio with the theoretical value given in the table
in
paragraph 55.
measured 500 times and the probable error

is 0.6 c m.
Plow many errors should
?
0.4
c.m.
and
0.8
c.m.
between
occur
136.
Show how the value of TT could be determined
experimentally from observations such as those in Example
135.
line is
of each observation
131.
In a system of observations all equally good, r being

137.
the probable error of a single observation, if two observations
are taken at random, what quantity is their difference as
likely as not to exceed, and what is the probability that the
difference will be less than
r?
r^T\
0.367
In the following measurements of an angle, ought

observations to be rejected ?
of
the
any
138.
12' 51".75
47".85
47".40
48".90
48 .45
51 .05
48 .85
50 .95
50 .60
47 .75
49 .20
50 .55
139.
Determine
whether any of the
Example 44 should be
140.
below.
M=
44".45
observations
in
rejected.
is measured with the results given

quantity
all the observations to be retained ?
Ought
236,
251,
249,
252,
248, 254, 246,
257,
243, '274
certain angle has been laid out with such accutrue value may be taken as exactly 90. Twentylive observations are made upon it with a transit that it is
141.
racy that
its
140
desired to
What
test,
and the
result obtained
is
59' 57"
89
0".8.
are the odds in favor of a constant error in the instru-
ment between
1" and
5"?
Between 0" and
6"?
908 92
86 1
of
142.
measurements
a
standard
bar
metre
Repeated
with a decimetre scale gave a result 10.032
0.010.
What
are the odds in favor of a constant error in the scale between
:
0.010
143.
683.4
and
0.054
-?
43
v/To
Two determinations of the length of a line gave
0.3 and 684.9
Show that the
0.3, respectively.
best value for the length is 684.15

0.51, and that the
error
of
each
determination
is 0.65.
probable systematic
144.
Two men
and J3 observe an angle repeatedly
with the same instrument with results
A.
47
B.
23' 40"
23' 35"
23 45
23
23
30
40
47
23' 30"
23
40
23
50
24' 00"
23
20
any relative personal error, and what is the best

value?
47 23' 38".2
1".6
145.
Three independent determinations of the capacity
of a condenser made with three different instruments gave
results 42.22
.15 and 44.20
0.18.
0.21, 43.40
Is there
final
What
is the most probable value of the

capacity?
For extended treatment of the subject illustrated
in
Exam-
ples 143 to 145 see Johnson, "The Theory of Errors and

Method of Least Squares," chap. vii.
In an estimation of tenths what is the probable error
146.
of an observation ? What is the average deviation ?
0.025
In obtaining the angle of deflection of the needle of

147.
a tangent galvanometer by the usual method what is the
probable error of the result ?
EXAMPLES.
If all the errors of a series of observations
148.
and
Itctuccn
portional to
What
ami
fri-queiicy of
magnitude, what is the
a,
its
are the values of
49.
141
tlu-
a.<t.,
r,
In Example 148 what
is
and
is
fall
pro-
Curve of Error?
r
/x
must
any error
the probability that the error
of a single observation will be as large as 0.5 a.

If all values of x between
150.
and a are possible,
and their probabilities are proportional to their squares, find
mean value
the
large as
151.
What
x and the probability that x
of
will
be as
Also draw the Curve of Error.
0.5 a.
the greatest probable error of a logarithm
is
found by interpolation in a seven place table?

.000000015
152. Given the following set of Normal Equations, together
with
[mm]
unknowns and
the
1.3409, find the most probable values of

their weights and probable errors.
There
were sixteen observations.

3.1217
.5756
z,
zl
- .1565 z l
- .0067 z x
.5756 z 2
4-
2.9375 z 2
4-
.0067 z 4
1.5710
.1103 z 3 -
.1103 z 2 4- 4.1273 z 3
.0015 z 2 4-
.1565 z 3
.0015 z 4
+
+
+
.9275
-2051 z 4
.0652
.2051 z 3 -f 4.1328 z 4
4- .0178
zl
z4
=
=
0.583
0.018
0.004
0.015
pz<
153.
[mm]
From
4.12
ten observation equations, for which was found

there resulted the normal equations
2.6322,
5.2485 z l
-
=
=
=
=
1.7472 z 2
2.1954 z s
1.7472 z x 4- 1.8859 z 2 -f 0.8041 z 8

2.1954 z l
0.8041 z 2
4.0440 z 3
0.5399
1.4493
1.8681
=
=
142
Find the most probable values of
and
z 1? z 2
z8
their probable errors.

zt
154.
Find the most probable values of the
together with
0.42
0.11
unknowns
in
the normal equations
459 z,
308 Zi
389 z l
244 Z!
+
+
-
308 z 2
464 z a
408 z 2
269
z2
+
+
-
389
408
676
zs
331
z8
z8
z4
155.
result
17.50 z l
show
zl
--
+
=
244
269
331
469
z4
z4
z4
z4
507
695
-f 653
- - 283
[mm]
0.488
=
=
=
=
1129
Zt
If thirteen observation equations give rise
[mm]
z3
6.50
Zj
6.50 z x
that the
0.67
+
-
the
100.34 and to the normal equations

6.50 z 2
17.50 z a
6.50 z 8
6.50 z 2 -f 20.50 z 8
z,
1.17
of the
0.60,
2.14
13.96
6.50 z 8

0.60,
to
281
5.40
=
=
=
unknowns
z8
0.32
are
0.55
APPENDIX.
ELEMENTS OP THE THEORY OF PROBABILITY.
200.
Definition.
b ways,
fail in
and
If an
all
event can happen
these
in
a ways, and
are equally likely to occur,
ways
the probability of the
happening of the event
and the probability of
its
failure is
-\-
is
^-,
Since the event must either happen or fail, the sum of the
above probabilities must represent a certainty. But
That
the probability of a certainty is expressed by unity.

3
the
Also,
probability, J , of the happening of an event is
of
its failure is given at once by
the
known,
probability
is,
if
I-
P.
201.
dice.
is
turned up will be 5
Number
Number
of
of
single throw is made with a pair of

the probability that the sum of the spots
Example A.
What
ways
ways
of throwing the dice

of throwing five is
is
Probability of throwing five
Example B.
coin
is
is
tossed
chance that three heads and three
36
4
4
..
=
=
36
up
six
tails will
times.
be the
Find the
result.
ii
Number
Number
of
of
28
ways
of throwing the coin
ways
of throwing three heads
is
6x5x4
1X2X3
-
is
20
Probability of throwing three heads
is
=
=
64
20
64
16
202. Compound Events. A certain event can happen in

a ways and fail in b ways a second independent event can
happen in a' ways, and fail in b' ways, all of these ways
:
being equally likely to occur.
To find the probability of the simultaneous occurrence of

the two events.
The
total
number
of
ways
in
which the events can take

1
place together
is
(a
b)
-f-
(1) Both events can happen in a a' ways.

(2) Both events can fail in b b' ways.
(3) First event can happen and second fail in
(4) First event can fail and second happen in
The
probability of
(1)
is
(a
b) (a'
h
The
probability of
(2)
is
(a
The probability
of
is
The
+
;
r,
b')
b'
ways.
a' b
ways.
b')
b')
b)
b')
(a'
probability of
But the probability

a
b) (a
(a
-\-
h'
ab
(3)
(a
(4)
+^
is
(g
of the happening of the
and of the second event
is
'
-\-
.,
b"
+^
first
etc.
event
is
Hence
it
will at once be sei-n that the probability of the simultaneous

occurrence of two independent events is equal to the product
Am-: \'D IX.
iii
of the probabilities of the occurrence of the component events.

J
/' M arc the
Or, in LC'iicr;il, if / ,, 7\,
probabilities of
.
the occurrence of
probability of
any number,
/>,
of
Independent event*, the
simultaneous occurrence of
tin-
P,
all
the events
Pn
is
(A)
meant those such that the manner

no influence upon the manner of
By independent events
is
of occurrence of one has
occurrence of the others.

203.
2
tain
problem
is
The chance
C.
Example
-,
3
that
can solve a cer-
and the chance that
can solve
it
is
Find,
12
The probability
(b) The probability
that both will solve
(a)
For
(a).
This
is
it.
that the problem will be solved.
a question as to the probability of the
concurrence of two independent events.

application of
problem
Therefore by an
(A), the probability that both will solve the
is
'
For
(b).
The
The problem
A
12
A
18
will be solved unless both fail.
probability that both will fail
17
y
is
The probability
Example D.
of getting a solution
pack of cards
then replaced. In how

that an ace will be cut ?
many
is
cut,
8
is
12
7
86
29
36
36
and those taken
trials will it
be an even u
off
iv
Let
from
n be the number
of trials.
Then
is
to be found
1
i-(}"=
^52
/
where the
first
bility that
we
member
of the equation represents the proban times in succession.
shall not fail
Solving for n,
=
In nine
trials
8.7
then there
is
of cutting an ace.
204. Dependent Events.
whose modes
log 48
log 52
little
If
more than an even chance
we have
number
of events
dependent one upon another,

the probability of their concurrence will be found by the same
a' now denoting the number
method as in paragraph 202
of occurrence are
of
ways
in
which
after the first event has
happened the
second will follow, and b' the number of ways in which after
the first has happened the second will not follow, etc. Accordingly, the general formula (A) of paragraph 202 applies to
dependent events as well as to independent ones.

Also, if an event can take place in a variety of ways, the
total probability of its occurrence will be the sum of the
probabilities of its occurrence in each of the different ways.
205. Example E.
Suppose two purses contain respect-
coin is taken
ively five dimes and a copper, and six dimes.
at random from the first purse and placed in the second, and
then a coin is transferred from the second to the first. What
is
the probability that the copper will remain in the first purse ?
The probability that the copper will be taken from the first
purse and placed in the second, and then returned to the

purse
is
_L
49
first
and the probability that the copper

tlu- first
purse at all
will not be taken
from
is
Therefore the probability that the copper will finally remain

in the first purse
is
3C
J_
"
42
"
6^
_5^
42
6~
FUNCTIONS OF SEVERAL VARIABLES.

206.
For the application
of
Taylor's
Theorem
to
the
expansion of a function of several independent variables, see

Osbor tie's " Differential and Integral Calculus," page 145.
And
for the conditions that lead to
maxima and minima
values of such functions, see page 155 of the same work.
BIBLIOGRAPHY.
The following brief list of treatises, dealing with the
Method of Least Squares, is appended for the benefit of those
whose professional work requires such constant application
of the process as to render desirable a more detailed knowledge of various special methods of solution. In connection
with some of the titles attention is called to the subjects the
treatment of which
is
particularly
full.
" The
Johnson,
Theory of Errors and Method of Least Squares."
Probability of Errors. Systematic Errors. The Method of Substitur
tion.
Wright,
" Treatise on the

Adjustment of ObservationSpecial Methods of Solution.
neering J'roblemt.
Applications to Geodetic
and Engi-
VI
" Text-Book on the Method of Least

Squares."
" Treatise on the Method of Least
Chauvenet,
Squares."
Merriman,
Development of the Theory.
Applications
to
Astronomical Observa-
tions.
Bobek,
"Lehrbuch der Ausgleichsrechnung nach der Methode

der Kleinsten Quadrate."
General Synopsis of the Method, illustrated by Numerous Examples.
Roll,
" Pie Methode der Kleinsten

Quadrate."
Applications to Geodesy.
Hansen,
"
Von
der Methode der Kleinsten Quadrate."
Applications
Helmert,
to
Geodesy.
Die Ausgleichungsrechnung nach der Methode der Kleinsten Quadrate."
" Calcul des Probabilites."

" Discussion of the Precision of Measurements."
Holman,
Problems in Physics and Electrical Engineering.
" Handbuch der
Weinstein,
Physikalischen Maassbestimmungeu."
Liagre,
Applications to Physical Problems.
Oppolzer,
" Lehrbuch zur

Bahnbestimmung der Kometen und Plane-
Jordan,
" Handbuch der

Vermessungskunde."
ten."
For a complete
list
Squares published up
Merriman,
"A
of
works on the Method
of
Least
to 1876, see
List of Writings relating to the
Method of Least
Squares, with Historical and Critical Notes."

J'ublished in the Transactions of the Connecticut Academy,
vol. iv,
1877.
Notice of works published since 1876 may be found in

periodicals devoted to the progress of Mathematical Science.
Such as
" Jahrbuch liber die Fortschritte der Matheraatik."
" Bulletin des Sciences
Mathematiques."
TAB1
TAHLK
2
Values of the Intf-ral
a
r
i.
rt
i
N?r
vn
Jo
_rument
0.4769
or
Vlll
TABLE
II.
Common
Logarithms.
TABLES.
TABLE
n
II.
Common
IX
Logarithms.
TABLE
III.
Squares of Numbers.
TABLES.
TABLE
III.
Squares of Numbers.
ana Prescott
rinciples of the
st squares
'ET

Bartlett General Principles of The Method of Least Squares, With Applications (1915) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bartlett General Principles of The Method of Least Squares, With Applications (1915) PDF

Uploaded by

Copyright:

Available Formats

ta

METHOD OF LEAST SQUARES,

PROFESSOR or MATHEMATICS. MASSACHUSETTS INBTTTUTB

volume was undertaken with the

view of presenting in as simple and concise a manner an

attempt has been

believed that everything essential to

ordinary problems has been included, no

to develop at length those special

methods and forms that are so useful and almost necessary in

measurements, are to be adjusted.

expected that this

will in such cases

THE ADJUSTMENT OF OBSERVATIONS.

23. Indirect Observations.

29. Relation Between the Weight of an Observation

42. Periodic Phenomena.

THE PRECISION OF OBSERVATIONS.

COMPUTATION OF THE PRECISION MEASURES.

C6-89. Conditioned Observations

The Distribution of Errors.

The Rejection of Obser-

93, Criterion for the Rejection of a Single Doubtful

98. Combination of Determinations having Different Con100. The Weighting of Observations.

101-103. Special Laws of Error.

104. Contradictory Obser-

107. Checks on the Formation of the Normal Equations.

200. Definition; Simple Events.

THE METHOD OF LEAST SQUARES.

measurements either with or without the aid

unknown quantities or upon certain functions

the values of the

measurement, or set of measurements, the observations are

ways and under vary-

Under these circumstances it will invariably be found that

and the question that now presents

determine from these discordant observations the true values

From the nature of the case,

which are rendered most probable

after all the observations

are taken into account, and, further, to determine the degree

the above results constitutes the

primary object of the Method of Least Squares. The method

measurements of the same quantity, and in determining the

making any observations, we may employ the

of the discrepancies between the

These errors are of two

Constant or Systematic Errors and Accidental Errors.

magnitude are due to some fixed cause. These constant

Theoretical Errors, such as those due to the refrac-

tion or aberration of light, the effect of a definite change in

known the magnitude of

as their causes are

from the observations.

observations, either by a particular method

or too late, always estimates a quantity smaller than it is, etc.

mined by a study of the observer, his "Personal Equation"

Although of a somewhat different character,

in reading a graduated circle

read 43 instead of the complementary

and a proper correction made.

Accidental Errors are errors due to irregular causes,

circumstances peculiar to that particular set of measurements,