FROM

DETERMINANT
TO

TENSOR
BY

W.

> V' F. SHEPPARD,

&y

Sc.D.,

LL.M.

FORMERLY FELLOW OF TRINITY COLLEGE, CAMBRIDGE

OXFORD AT THE CLARENDON PRESS

1923

OXFORD UNIVERSITY PRESS

London

New York

Edinburgh Glasgow Copenhagen Toronto Melbourne Cape Town Bombay Calcutta Madras Shanghai

HUMPHREY MILFORD
Publisher to the University

Printed in England

PREFACE
THE tensor calculus used in the mathematical treatment of relativity and concisely explained by Professor A. S.
,

Eddington in
Gravitation
',

his
is,

'

Report on the Relativity Theory of

like the various kinds of vector calculus,

a system of condensed notation which not only conduces to economy in the writing of symbols, but, what is more important, enables spatial and physical relationships to

be grasped as a whole without having to be built up from a number of components which really represent views

from

different parts of space.

is

Three-dimensional geometry
:

troublesome enough the addition of a fourth dimension made the need of a condensed notation imperative. Professor Eddington has recently pointed out that the
or physics
results, to other

tensor notation and methods can be applied, with happy and more elementary classes of problems

than those
this

which they were originally devised and an attempt to put his somewhat compressed exposition into a form in which it may appeal to a larger
for
;

book

is

circle of readers.

The book,

therefore, is not intended as

an introduction to the mathematical theory of relativity though I hope it may be of some use for that purpose
but rather as an exercise in the elementary application of methods which apart from any practical use, possess a special
,

beauty of their own.

The new notation is not introduced until the fifth chapter. The properties of determinants, which serve as the startingpoint for the application of the notation, are familiar to the mathematician but, as I hope the book may be read by
;

some who are not

entirely at

ease with determinants,

4
I have

Preface

commenced with four chapters on the elementary theory of the subject. I make no apology for doing this, instead of referring the reader to the ordinary text-book
on algebra.
;

The text-book treatment

is

not always stimu-

lating the reasons for the various stages are not necessarily clear to the student and attempt at simplicity sometimes
;

leads to loss of rigidity in proof. In such a subject it is necessary to take the reader into one's confidence ; and
this earlier part may in this respect be found helpful to some, teachers or students, to whom the later part makes

at

first a less strong appeal. I have added a chapter on some applications to the theory of statistics, to which the tensor calculus seems

specially suitable.

The

basis of this portion, so far as

concerned, is a short paper by Professor Eddington, mentioned at the end of the chapter, This is one only of
is

method

many

possible applications.
sets will
;

"What I have called double

the advanced student as matrices

be recognized by and many of the pro-

But the tensor positions will be found to be familiar. in bringing into close calculus may fairly claim that, relation various branches of mathematical study, previously
regarded as distinct,
it gives them a new life. I have to thank Professor Eddington for looking at

my

manuscript and making some corrections and suggestions.

5 June 1923.

W.

S.

CONTENTS
PAGE

INTRODUCTION

DETERMINANTS
I.

ORIGIN OF DETERMINANTS
2. Formula Solution of simultaneous equations 4. Construction of for solution ; 3. General problem
1
.

11
;

terms
II.

5.

Rule of signs.

PROPERTIES OF DETERMINANTS
1.

....
;
. .

20

Definition of determinant

2.

Elementary proper-

ties
4.
III.

3. Properties depending on the rule of signs Co factors and minors.

SOLUTION OF SIMULTANEOUS EQUATIONS

1.

29

Statement of previous results;

2.

General solution.
.
.

IV.

PROPERTIES OF DETERMINANTS (continued) 1. Sum of determinants 2. Multiplication of deter;

31

minant by a single factor 3. Alteration of column or row 4. Calculation of determinant 5. Product of determinants 6. The adjoint determinant.
; ; ; ;

V.

THE TENSOR NOTATION

2. Reciprocal properties of determinant ; 3. Elements of reciprocal determinant; determinant; 4. Set-notation ; 6. Pro5. Principles of set-notation ;
1.

39

Main

duct-sum notation
set notation
;

7.

Inner products of sets

;

8.

Unit-

Properties of the unit set minant properties; 11. Example of method.

9.

10. Deter-

SETS
VI.

SETS OF QUANTITIES
1.

57
Single sets;
3.

4.
6.

Introductory; Sets generally ;

;

2.

Double sets;
;

Inner product
;

7.
;

5. Sums and products of sets 8. Inverse double The unit set

;

sets

9.

Reciprocation

10.

Continued inner products

11. Partial sets.

6
VII.

Contents

RELATED SETS OF VARIABLES

1.

....
; ;

PAGE
72

Direct proportion of single sets 3. Reciprocal proportion of 4. Cogredience single sets and contragredience 5. Contragredience with linear
;

Variable sets

2.

relation

6.

Ratios of sets generally

7.

Related sets

of higher rank.
VIII.

DIFFERENTIAL RELATIONS OF SETS

1.

...
sum
or pro-

85

Derivative of a set
3.

2.

Derivative of
;

duct

Derivative of function of a set

4.

Transforma-

tion of quadratic form to

sum

of squares.
.

IX.

EXAMPLES FROM THE THEORY OF STATISTICS

1.

92

3. Conjugate Mean-product set sets 4. Conjugate sets with linear relations 5. The 6. Criteria for improved values frequency-quadratic 7. Determination of improved values.

Preliminary

2.

X.

TENSORS IN THEORY OF RELATIVITY

1.

.115

Preliminary
4.

sets
ties.

3. Other Single sets (vectors) Reason for limitation 5. Miscellaneous proper;

2.

APPENDIX

Product of Determinants

.122
124
125

INDEX OF SYMBOLS

GENERAL INDEX

INTEODUCTION
As
it

this is a comparatively

new

subject to most readers,

it is

may

be as well to explain briefly what

is

about.

vector in (say) 3 dimensions

a directed quantity,

determined as regards both direction and magnitude by its components, which are magnitudes measured along three
definite axes.

beforehand, a vector SI

we can
is

These axes being supposed to have been fixed take them in some definite order; and
then determined by a
call

set of 3 quantities,

which we may

^1

^2

4r

Algebraically, the idea of a vector can be extended to something which is determined by a set of m quantities ^1

^3'"^ra>

where

have any value. determinant (say)

m may

is

the algebraical sum of all the products that can be formed in a certain way according to a certain rule of signs from the set of quantities

Each element of this set has its position fixed in the set by the numbers of horizontal and of vertical steps required to reach it from the initial element a. Thus the set is extended in two directions, while the set which determines
a vector
is

'

'

a determinant

extended in one direction only. This applies to with any number m 2 of elements.

Introduction

In the same way we might have a set extended in 3 directions, the symbols being written along the edges of a cube and along lines inside the cube or on its faces and we
;

c could, in theory, increase the number of directions to 4 or more, by proper convention as to the order in which

'

the elements are to be taken.

On

quantity may be regarded as a set not extended in any direction. The tensor calculus, using the word tensor in its broad
' '

what in the language of vectors

the other hand, a single is called a scalar

sense, deals

with

all

these different kinds of sets, in relation

by which we can regard axes of reference as being determined. In the narrower sense in which the word is used in reference to the theory of relativity,
to sets of variables

only sets which satisfy certain conditions are called tensors. In this book I have treated the tensor calculus as arising
out of the use of determinants.

Chapters I-IV deal with

the elementary theory of determinants, so far as it is required for our purpose. (The student who is familiar

with determinants can skip these chapters.)

the tensor notation
is

In Chapter V introduced in successive steps, with

explanatory remarks. These latter are in small print, not because they are less important, but in order not to break the continuity of the chapter as a whole. In Chapter VI
these explanatory paragraphs (or parts of them) are brought together and amplified so as to give a general idea of the

elementary properties of sets. Chapters VII and VIII deal with some developments of the subject in its general aspect. Chapter IX shows the application of the

methods to certain problems in the theory of statistics and of error this can be omitted by any one who wishes to pass on to Chapter X, which deals very briefly with the
;

tensor in

its

more limited

sense, as applied to the theory

of relativity.

DETERMINANTS

I.

ORIGIN OF DETEEMINANTS
Deter-

I.

l.

Solution of simultaneous equations.

Suppose we have two equations = 19)

minants ordinarily
equations.

arise out of the solution of simultaneous

=
Then,
give
if

18)
could

we used only elementary methods, we

first

multiply the

by

3 and the second

by

2,

which would

= =

57)
36)'

and thence, by subtracting, we should have 7x = 21, 0=3,

whence either equation would give

Similarly, if

we had

three equations

x3y
we
could,

2z

between the first and the second and between the second and the third, obtain

by eliminating

9x

70+^= lly =
we
2,

5)
13)
'

whence, proceeding as before,

SB

should obtain

1,

y=

s=4.

This process of successive elimination is tedious, especially when there are more than two unknowns ; and it is found

12

Solution of simultaneous equations

1. 1

better to obtain a formula for the general solution apply it to the numerical values of the particular case.

and

1 can be Since such an equation as # written in the form (+ l)# + (-3)^ + (-2)s (-1), we can use positive signs throughout, it being understood that

3y2z =

the quantity represented by any symbol

positive or negative.
I. 2.

may

be either

Formula

for solution.

(i)

For completeness we

begin with one unknown.

The equation
a^x

gives

..&. a
i

(ii)

For two unknowns the equations

may be

written

Multiplying the first equation by and subtracting, we get

&2

and the second by

b lt

fcA'-'bV*
whence

^-4Ai
<*

"ASimilarly, interchanging 0's

A
's,

and

It should be noticed that the expressions for

x and

for

have the same denominator, and that the numerators are

1.

(i)

Formula for

solution

13

obtained from the denominator by replacing the as in the one case, and the 's in the other, by &'s.
(iii)

Next take the

case of three unknowns.

Let the

equations be

Eliminating z from the

first

two equations, we obtain

2 (<*&-(*&) as + (b l cz -b 2 cjy = ViSimilarly from the second and third equations

V-

Then, eliminating y from these equations, we get

As

before, the

numerator

by replacing as by
the denominator.

is got from the denominator and we therefore need only consider 's, Multiplying out, it becomes

Hence, replacing the as by

&'s for

the numerator,

_
Corresponding expressions can be obtained for y and for
z.

I. 3. General problem. (i) We might proceed in the same way for equations in four or more unknowns. But this would mean that each case would have to be considered

separately

formula, but the

practically

and not only should we fail to get a general algebraical work would soon become

impossible.

We

therefore

alter

our tactics.

14

General problem
write

1.

(i)

We

down
t

the general equations involving

un-

knowns x,y

z,.

.w

(I. 3.

A)

guess at a solution, and then verify that this solution does actually satisfy the equations.
(ii)

The values

tions will be in the form of fractions.

first

of x, y> z...w as found from these equawill consider

We

the denominators.
2, for

tained in

Putting together the results obthe cases of m = 1,2, 3, we find that the

successive denominators,
are

which we

will call

DV\

an expression ZX TO ), which we should guess to be the common denominator in the solution of the equations (I. 3. A), and of which L( l \ L^\ L^ shall

We

want

to obtain

be the particular cases for

m=

1, 2, 3.

(iii) The three D's in (I. 3. 1) have a general similarity, which enables us to obtain a formula for ZX m ). It will be seen that both in DW and in J)W some of the terms have We will see first how sign + and some have sign the terms are constructed, and then consider the question
.

of sign.
I. 4.
-#'s, for

Construction of terms.
which the values of
the product of
2.

(i)

For each of the three

each
coefficients
coefficients,

are

1, 2, 3 respectively,

term

is

factors,

which are the

in the equations in

In writing down these

1.

(ill)

Construction of terms

15

it is convenient to keep them in the relative positions in which they occur in the equations. Thus we get

For

DW

For

In each case we have a set of quantities arranged in the form of a square. The individual quantities are called the elements of the set the quantities in a vertical line
;

column, and the quantities in a horizontal line constitute a row. The columns are numbered from the left, and the rows from the top. The diagonal drawn from the top left-hand corner i.e. the diagonal through
constitute a

ax

is
(ii)

called the

leading diagonal.

Each term contains

factors,

the set in such a

way

that one (only) shall

column and that one (only) shall m contains .ZX every term which can be constructed
)

which are taken from come from each come from each row. Also
in this

way. Take, for example, D^. Since one factor is to come from each column, the factors are an #, a #, and a c. The a can be either a l or a 2 or 3 i. e. it can be taken in three
,

as has been taken, one row has been used up, and the b can only be taken in two ways and, when one of the two b's has been taken, the c can
ways
;

when one

of these three

3.2.1 = 6 possible combinations the number of terms in L&\

similarly only be taken in one way.

There are therefore

of factors ;

and

this is

(iii) Another way of stating the thing is that, if we keep to a fixed order a 6 c of the factors in each term, the suffixes of the factors are the numbers 1 or 1 2 or 1 23,

arranged in different ways, and there

of the possible arrangements.

is

one term for each

16
(iv)
is

Construction of terms

1.

(iv)

We

conclude that
1

D (m
li

contains terms, each of which

constructed by taking

m factors from the set of m x m

c

m*

quantities

ai

ifi
m
Jm

am

in such a

way that there shall be one factor (only) taken there from each column and one (only) from each row
.

1 m ways 1) being one term for each of the m (m in which this can be done. Or, which comes to the same
.
.

thing, that the terms are

suffixes

made up
!

of factors a b c

./ with

12

arranged in different ways, there being

one term
I. 5.

for each of the

possible different arrangements.

Eule of

signs.

(i)

It will be seen that, in the D's

after

DM, half of the terms are positive and half negative, and that in each case the term found from the elements in
the leading diagonal namely # x or a^b 2 or a^b^ is posishould therefore expect that half of the terms tive.

We
)

in

.ZX

m would be

positive
call

a\

&2 C

3- -fm
first

which we

and half negative, the term the leading term and usually

place

being positive.

The
<?

difficult question is
5,
*s

that

of signs. or

In the case of (say) m = whether such a term as a 3 5 2 ^4*i

how
^

are

we

to

know

nave

* Qe s ^S n

-?
(ii)

The sign

of a term must, if the letters

ale.,

./"are

kept in their original order, depend on the arrangement of the suffixes, i. e. on the extent to which this has departed

from the

initial

ment such

as

arrangement 1 2 3 ...m. Now any arrange35241 can be got from the initial arrangement

12345 by a series of interchanges of adjacent figures. We must fix a definite order in which these interchanges are to

1.

(iii)

Rule of signs

17

be made.

We therefore say that each figure is to be moved in

which ultimately comes first, then that which ultimately comes second, and so on. Thus in this particular case the successive stages would (repeating for each group of interchanges the arrangement from which we start)
turn, beginning with that

be 12345, 13245,31245; 1245,1254,1524,5124; 124,214; 1 a total of seven interchanges. Now let us look 14, 41
;
:

at the signs in
is
is

I)W and L^\

In I)W the arrangement

21

obtained from 12 by 1 interchange, and the sign for 21 In .ZK 3 ) the signs of the successive terms, and the
.

stages

by which the
:

final

arrangements of

suffixes

are

obtained, are as follows, the

numbers of interchanges being

added in heavy type

123
123, 132 .. 123, 213 ..

+
1
1

123, 213, 231

-f

123, 132, 312 123, 132, 312, 321 ..

)

2 2 3

2 and for IX 3 ) the sign is It will be seen that both for .ZX - or + according as the number of interchanges is odd

therefore adopt this as our rule in the case of ajb^c.^l^e^ for instance, seven interchanges are necessary, and the sign is therefore

or even.

We

(iii)

the

above

In order to find the sign of any given term by be necessary to perform all rule, it would

the interchanges. shorter method is to look at the it stands and to consider the reversals of order in it i. e. taking the suffixes of the term in pairs in every

term as
;

possible

way without
i.

altering their order, to see in

how

many
order.
pairs,

cases the

numbers are in the reverse of their order in

e.

the leading term,

are in descending instead of ascending

,

The term ajb^c^il^ those in which the

for instance, gives the following

is

order

reversed being printed

24, 21, 41.

in heavier type:

35, 32, 34, 31, 52, 54, 51,

18
It
in
is

Eule of signs

1.

(ill)

easily seen that each interchange, of the kind described

(ii)

above, produces one reversal of order

for,

while

we

are shifting* one number, such as the 5 in the second group of arrangements there shown, the relative order of the

other numbers remains unaltered. It follows that the number of reversals of order is the same as the number
of interchanges of this kind and therefore the sign of a term will be or + according as the number of reversals of order is odd or even.
;

(iv) The interchange of any two suffixes in a term changes the sign of the term.

and ^ [Let the two suffixes be coming before ^ in the term in question, but not necessarily being before it in numerical
</>
;
<

order.

Then the interchange of (1) First let $ and ^ be adjacent. and increases or decreases the number of reversals of order by 1, and therefore changes the sign of the term. (2) Next suppose that there are x suffixes between $ and fy. Then we can move ty in front of by x + 1 interchanges with the adjacent term, and then move into the original position of ^ by x interchanges. This is a total of 2 x + 1 interchanges, each of which in succession makes a change of sign the total
(f>
>//
< <
:

result

is

to

change the sign of the term.]

far

(v)

We

have so

assumed that the factors of a term

are arranged in the original order of the letters a

bed

Now

suppose that the order of the factors is altered in any way. How does this affect the rule of signs ? The alteration of order can be brought, about by a series
of interchanges of factors. Suppose there is an interchange and of Then, by (iv), the number of reversals of a^ 6^. order of suffixes is altered by an odd number, but the

number of

reversals of order of letters is also altered

if

by

an odd number; and therefore,

we

consider the

sum

I.

5 (v)

Eule of signs

19

of the numbers of reversals of order of letters and of

suffixes, this

sum

either

is

not altered or

is

altered

by an

It follows that, if the factors of a term have leen shifted about so that the letters a b c . are not in their original order, the sign of the term depends on the sum of the

even number.

numbers of reversals of order of

spectively,

letters

and of

suffixes

re-

being
(I

or
4

according- as this

sum is odd or even.

.

For example, in

b 5 c 2 a 3 e l there are five reversals of order

of the letters and eight of the suffixes, so that the sign is

II.
II.

PROPERTIES OF DETERMINANTS
.

Definition of determinant.

bine the results obtained in

4 and

I. 5.

We can now comWe suppose that

we are dealing with a set of mxm = m 2 quantities, which we can arrange in the form of a square, thus (the quantities
being denoted by crosses)
:

X X X

X X X

X X X

...

...

X X

...

expression which we have to consider is the algebraical sum of a number of terms, of which some are taken positively and some negatively. Each term (apart

Then the

from sign)
in such a

is

way
;

the product of m elements of the set, taken that one element (only) shall come from each

column and that one element (only) shall come from and there are m terms, corresponding to the each row m different ways in which this can be done. The leading term is the term containing the elements in the leading The signs of the diagonal of the square, and has sign -f
!

other terms are to be found by replacing the elements of b v etc., arranged as a key set the set by a v a 2
,

The sign of a term is then or + according as the sum of the numbers of reversals of order of the letters and of the
suffixes, as
is

compared with the leading term odd or even.

a^

l>

c^

./m

II.

(i)

Definition of determinant

21

The
a

c3

algebraical sum of the terms so obtained, namely -fm^ a 2^i c3 -fm + e tc., is called the determinant
set,

of the

and

will be denoted
it is

by D.

It should however

be

observed that
;

the determinant of the set as so

with different arrangements of the elements of arranged a set, still keeping them in a square, we may obtain
different determinants.

can therefore define the determinant as the algesum of terms of the form a b c r . ., where pqr p q are the numbers 1 2 3 ... m arranged in some order, there
braical
. .
.

We

being a term

for each of
term being

the

prefixed to the

m ! possible orders, and the sign + for the natural order 1 2 3 ... m

and

or

reversals of natural order

for other orders according as the number of is odd or even.

for

The symbol

the determinant

is

single vertical lines before and after the set

constructed by placing thus ;

means the determinant

^2

<?

etc.,

which we have

called

The terminology is the same as is given in I. 4 and I. 5 for a set. The quantities between the vertical lines are
the elements of the determinant.
are a

Those in a vertical line

column
;

those in a horizontal line are a row.

is

The

leading diagonal

hand corner

the diagonal drawn from the top leftand the leading term is that containing the

elements through which the leading diagonal passes. The leading term, as already stated, is taken positively. If the symbol for a determinant contains m columns and

rows, the determinant

II. 2.

is

said to be of the

mth order.
the

Elementary properties.

(i)

From

mode of

22
construction
it

Elementary properties

II.

(i)

follows that each element of the determinant

I)
I

appears in

(HI

out of the

terms

and there are no

the

two terms having more than

(ii)

2 factors alike.
is 0,

If each element of a column or of a row

is

determinant

= 0.

[For each term contains one of these

elements as a factor.]
II. 3.
(i)

Properties depending on the rule of signs.

is

The value of a determinant

;

not altered by'making the columns

rows and the rows columns

i
o

e.g., for

m=

4,

*i

ci
Co

fl

vn

'O

[Let D be the determinant, and R the new determinant obtained by making columns rows and rows columns in the symbol for D. Then, apart from sign, D and R obviously have the same m terms. We have therefore only to consider signs. The two determinants have the same leading term, which is positive in both. Let t be any other term of D, say ajb^c^ .... Then t also occurs in R, but, since the terms of R must be constructed according to the system prescribed in our definition of a determinant, the factors of t in R will be arranged in the numerical order of the suffixes, namely diCjuJbi .... The sign of t in D depends on the number of reversals of order in the suffixes 3421. and the sign of t in R depends on the number of reversals of order in the letters deal .. But each of these numbers is the sum of the numbers of reversals of order of letters and of suffixes as compared with the and 1234...; and, by I. 5 (v), these original orders alcd sums are either both odd or both even. It follows that the sign of is the same in both determinants. This is true for each term of D or R and the two determinants are therefore equal.]
! .

.,

determinants correspond so that the columns of one are the rows of the other, each determinant is said
to be the transposed of the other

If two

II. 3 (v)

Properties depending on the rule of signs

(i)

23

(ii)

It follows from
is

that any statement as to columns

or rows

equally true for rows or columns.

We
'

shall
f

indicate this, for conciseness,

by

'

column [row]

or

row

[column]
(iii)

'.

If any two columns [rows] of a determinant are

inter-

changed, the absolute magnitude of the determinant remains

unaltered, but its sign is changed.
e. g., that we interchange the 6's and the e's. Let $ be any two suffixes. Then, in the original determinant, corresponding to any term which contains b^ and e^, there is another term exactly similar except that the factors are b^ and e^ and these two terms, by I. 5 (iv), are of opposite sign. The effect of interchanging the &'s and the e's is that the two terms are interchanged, i. e. the sign of each is changed. This applies to every such pair of terms.]

[Suppose,

and

(iv)

If any two columns

determinant
is

[rows']

of a determinant are iden-

tical, the

= 0.

see this in either of two ways. Consider a pair of terms such as are mentioned in (iii). The one contains b^ and e.j, the other is exactly similar, except that it contains and e^ and the two terms have opposite b^

[We can
(1)

signs.

60 determinant

If

= e^ and fy = e^, the two terms cancel.

is

The whole

made up

of such pairs.

(2) More briefly, suppose we interchange the two columns which are identical. Then the determinant remains unaltered.

But, by (iii), its sign is changed. the determinant is 0.]

(v)

This can only be the case

if

Since columns and rows

may

be interchanged, a

determinant

is sometimes represented by its leading diagonal alone, if this indicates a system for insertion of the

remaining elements.
.../

The notation
I

is

am

24
it is is

Properties depending on the rule of signs

II.

(v)

then immaterial, so far as the value of the determinant concerned, whether we enter the a's as a column or

as a row.

But

it

should be mentioned that the relative

arrangement of columns and rows is of importance later on, when we come to consider properties of sets of quantities.
ex(i) In the complete which is either a^ one pression for D, each term contains or 2 or # 3 etc. We can therefore group the terms accordII. 4.
,

Cofactors and minors.

ing to the
instance,
',

's

they contain.

In the case of

D^\

for

h
b

C,

cc
c.,

~a

ci

are grouped in this way ; and Suppose that the terms of a m be denoted let the resulting coefficients of a^ a 2 3 .
. .

-w
3

\>?AiA 2 As...A m

Then

D=
Similarly, if
c's etc.

a1 ^1

+ a2 ^ 2 +

J 3 +...+a m <4 ni

we group the terms according to the 's or they contain and denote the coefficients of the &'s
by
JB1

or cs etc.
shall

B B
2

^m
C2 +

or ^i ^2 ^3

^m
f

e ^ c<

we

have

4BV^+4*s+V*+...+M
D=
The
Cl

'

C\

c2

c,

0,+

...

+ c m Cm

etc.

the cofactors of the correspondelements of the determinant thus the cofactor of 3 is ing -Z? where b z J53 is the sum of all the terms which contain b y 3
A's, ^'s, etc., are called
; ,

The terms which contain the leading element x are obtained from the leading term a^^c^ .fm by altering
(ii)
.

II.

(iii)

Cofactors

and minors

suffixes, and prefixing the proper sign to the term, in the manner already described, with the proviso that the But this process will give us factor tfj remains unaltered.

the

the

by a v of the terms so constructed from a leading term # 2 c 3 In other words, the cofactor .fm
products,
. .

of

0j is

the determinant

In

.ZX

3
),

for example, it is

b.

The determinant which is obtained from D by striking out the column and the row which contain any element of the determinant is called the minor of that
(iii)

element in the determinant.

We
We

see

from

(ii)

that

A l = cofactor
any other element, say element, with the sign
e?

of a 1

= minor of a v

might show in the same way that the cofactor of 4 is equal to the minor of that
,

or

position of the element in simpler to find the cofactor

+ prefixed according to the the determinant but it is

:

the position of a r and the rth row.

by bringing the element into Let the element be in the ^th column

Then we can make it the leading element, without altering the order of the other columns or rows, by means of q I interchanges of its column with

an adjoining column and r 1 interchanges of its row with an adjoining row. Each of these interchanges, by II. 3 (iii), 1 and the total result is multiplies the determinant by
;

to multiply

by

or

by

according as q

+r

2 is odd

26
or even.
leading-

Cofactors
1

and minors

II.

(iii)

Having got the element into the position of the element, we strike out the first column and the first

row

the result, apart from the prefixed sign, is still to give the minor of the element, since the relative positions of the
;

other columns and rows are unaltered.

Hence

the cofactor

or + preequal to its minor with the sign fixed according as the number of steps from the leading element to this element is odd or even ; it being understood that each

of any element

is

step

either horizontally from one vertically from one row to the next.
is

column to the next For example,

3,

or

J 3 = + minor
(7

of

=
etc.

minor of

c 4,

(iv)

We

have found in

(i)

that

etc.

We

have now to find the value of such sums as

etc.

but replace the &'s or the find the value of

Let us take the second and third of these as examples, c's by 0's. Then we want to

Now we

see

from

(II. 4. 1)

that this

is

the value of the

determinant

> .../,
c

Jm

II.

(v)

Cofactors
1

and minors
.

for the cofactors of

O m in this determinant are

5

a in D, i.e. are Let us replace throughout this determinant by any letter, other than fl, occurring in D, e.g. by c. Then the determinant becomes

the same as the cofactors of a l

A,

A3

k
li

/a
r

in

'

Jw

But
cal,

this is a determinant

which has two columns identi-

and

its value,

by

3 (iv), is 0.

Hence

Similarly

b^ +
1

t>

A2

...

^ + 2 52
1

+...^
becomes

+ bm A m =

= or

etc.

(v)

Now

let us

interchange the columns with the rows,

so that the determinant

cl

C2

"

cn

J\

./2

/3

f /

Then, by

(i),

the value of the

new determinant

is

the

same

as that of the old, i.e. is

e

element

Also the minor of any in the new determinant is the same as its minor

D.

4n the

old determinant, but with columns

;

changed, so that its value is unaltered

and rows interand the number of

The steps from a^ to e is the same in both determinants. cofactor of e in the new determinant is therefore equal to
its cofactor in

the old determinant.

Hence by applying

28
(II. 4. 1)

Cofactors and minors

II.

(v)

and

(II. 4. 2)

to the

new determinant we

get

new

sets of relations,

namely

=D
J

(II. 4. 3)

etc.

and

o.
etc. (vi) If all
J

(II. 4. 4)

are 0, the determinant

the elements in a column [row], except one, is equal to the product of that one

by

its cofactor.

III.

SOLUTION OF SIMULTANEOUS EQUATIONS

1.

III.

Statement of previous

results.

We

have

next to consider the solution of the simultaneous equations 3. A). Before we do this, it will be convenient to (I.
express in determinant form the results obtained in
I. 2.

These results are as follows

(1) If
(2) If

&

then* x

-f-

then

*.

(3)

If

then

= =

where

a,

b,

c,

* Here, as elsewhere, vertical lines denote a determinant, not absolute value '.

30
III. 2.

General solution

III. 2

General solution.

The general equations

of

which we require a solution are those

set out in (I. 3. A),

namely

2.

A)

The form of the

given above.

suggested by the results Multiplying the successive equations by

solution
is

A lt A 2

...

A m and
,

adding,

we have

and (II. 4. 2) the coefficient of x is equal and those of y^ z,...w are equal to 0. Also the expression on the right-hand side is what D would become

By

(II. 4. 1)

to D,

if

we

replaced the

as by

's.

Hence

x = =

IV.

PKOPEETIES OF DETEEMINANTS
(continued)
1.

IV.

Sum

of determinants.

If two determinants

are identical except as regards one column [row], their sum is a similar determinant in which the elements of that

column [row] are the sums of corresponding elements in the two determinants. [For example

= =

fl-L

+ ^j

bl

since

it is

IV.

2.

Multiplication of determinant by a single

If each element of a column [row] is multiplied factor. by the same factor, the determinant is multiplied by that
factor.

[For example

=
b2 c2

*3

C3

If the elements 3. Alteration of column or row. column [row] are multiplied by a single factor and added to the corresponding elements of another column [row], the

IV.

of a

value of the determinant

is

not altered.

[For example

3*3

32

Calculation of determinant

IV. 4

(i)

IV. 4. Calculation of determinant. (i) There are two main methods for calculation of a numerical determinant. () When m is small, we can use the formula

then

I>=3

IV. 5
This
a
b,

(i)

Calculation of determinant

33
it

is

=0
c,

if

=b
e,
e>

or if a

c,

etc.

Hence

contains
it

ac,...a
. . .

a
b

as factors.

Similarly

con-

tains b

f\ and so on.

Looking

to the

leading term,
factors;
i.e.

it will

be seen that there can be no other

[Example.

Hence prove that

mb

m ~l c
cr-ri

m~ f m ~ e
l
et

,_

x coefficient

in<

(a),

where
,r-H
1 1
i)';-

fi

f...i

r~
i

IV.

5.

Product of determinants.

(i)

The product of

and a determinant of order n can be expressed as a determinant of order m + n by placing For the leading diagonals in line and filling in with O's.
a determinant of order

example
!

*1

34

Product of determinants

IV. 5

(i)

the terms of the product of the one expanded determinant by the other are accounted for, and there are no others.]

have now to show that the product of two determinants, each of order m, can be expressed as a determinant of order m. To obtain the general formula,
(ii)

We

it will

be sufficient to take a particular case,

l

e.

g.

ft

7i
y.2

'2

ft
ft

~*

73

provided that in our reasoning each determinant.

we

retain

as the order of

By means

of the

first

sentence of

(i)

we can

write

down

DE as a determinant of order
0,

2m,

i.e.

IV. 5
the
1's,

(ii)

Product of determinants
whether

35

I's),

m
1

is

even or odd.

Then, inserting- the

we get

36

Product of determinants
is

IV. 5
called

(ii)

in the lower right-hand quarter of the above

P, we

have,

by

(i),

1
1

i.e.

\
a

<i

ft

7i
7a

wa bs

^2

ft

a3

cs

3s

7s
.

(IV. 5.1)

The reasoning
above.

is

quite general,

and the product of two

determinants of any order can be written

down from the

By

interchanging columns and rows in one or other

or both of the original determinants we get three other All four expressions, of course, are equal expressions.

when expanded
standard form.

we
It
is

shall

take

the

above to be the

to be noticed that

DE
;

or

DxE

by reversing the order of multiplication we get four other forms, but The these are only the transposed of the previous four.
t '
'

means the product of D and

in this order

eight forms are given, in the new notation (see V. 6 (vi)), in the Appendix (p. 122). They are based on the principle
that in the standard form the element in the ^th column and rbh row of the product is formed in a particular way

from the #th column of the row of the second.

IV.
6.

first

determinant and the rih

adjoint determinant. the determinant whose elements are the

The

(i)

Let I) denote

cofactors of the
i. e.

corresponding elements in the original determinant,

IV. 6

(ii)

The
'*!

adjoint determinant

37
(iv. e. A)

B^

Ct

...^;

If

we interchange columns and rows

in

D'

and then

express the product DD* as in (IV. 5. 1), we find that the elements in the principal diagonal of the product are
a.

A,

Ao

of which,

are

0^+

+ b^B* + + bm m etc., each =D, while the other elements by #2 + ... + a m m) l A l + b 2 A 2 + ... + bm Am etc., 2 +
. .

.am

Am mm

b,

B,

. . .

(II. 4. 1),

is

l>

each of which, by

(II. 4. 2), is

= 0.

Hence
(IV. 6. 1)

DD' =

I)

0...0

D
6

0...0

O...D

and therefore

D'

= $-\

(IV. 6. 2)

The determinant If was formerly called the reciprocal, but It is not is now more usually called the adjoint, of D. the true reciprocal of D, since the product of the two is
not
1

but Dl

(cf.

V.

2).
,
,

(ii)

^%

to D',

Let the cofactors of A^ B^ Ct Then, applying ft yi we have

, , . . . .

. . .

in

(II. 4.

D' be denoted 3) and (II. 4. 4)

Ti=

We

can regard these as equations /D...CIComparing them with

for

determining a l9 fl lt

38

TJie adjoint determinant

(II. 4. 3)

IV. 6

(ii)

which are obtained from

the solution
is

and

(II. 4. 4),

we

see that

etc.

the cof actor of any element of the adjoint determinant equal to the corresponding element of the original determinant, multiplied by the ratio of the adjoint determinant to the original
is

Hence

determinant.

V.
V.
i.

THE TENSOK NOTATION*

properties of determinant.
is
(i)

Main

The

notation so far used

For higher work we a more liberal use of

the ordinary one for elementary work. reduce the number of letters and make
suffixes.
;

We
.

by Xv X X ,...Xm
2, 3

shall replace x, y, z

. . .

by ml > understood that the values Also, being assignable to each of the letters q and /are 1, 2, 3, ...?#, we can use d and d to denote the determinants
'

a lt d lf cv

dn

and

so

on.

it

qr

rq

whose elements in the ^th column and /th row are respectively f

and^,

i.e.

31

...dT
,..r/,

dM "32

W2
(V.I. A)

d11
23

2m

so that the interchangeability of

columns and rows gives

I-

Kr = !<**
I

'(V.1.1)
to

(ii)

The

/'s in

(III. 2.

A) were supposed

be^nown

* The paragraphs in small print may be omitted on first reading, but should be read before Chapter VI is taken. t It is more usual to have the suffixes the other way round i. e. to use 11, 12, 13,... 1m as suffixes for the first row. But my arrangement seems to follow more naturally from the a^ ... notation, and I also find that it fits in better with the subsequent work.
;

40
quantities,

Main
so

properties of determinant
served
to

V. 1

(ii)

that the equations

determine

x>yi z >>" W'

We shall have to consider the relations between

. . .

the set of quantities which we have denoted by #, y, z, and those which we have denoted by k v 2 3 .. km
/

We
,

therefore, in altering x, y, alter k to

z,

. . .

?0

to

X Xv Xy
lt

. . .

Xm

also

2,

properties which we have to consider are Definitions are marked with capital letters propositions with arabic numbers.
(iii)

The main

set out below.

(A)

= cofactor of dp8 in
1, 2,...

I)

(B)

(C)

DD' =
If

UP*
r

(2)

in lJ cofactor of lJ ps

(3)

then
(p

(4)

1,

2,...i)

As a preliminary, concise, we may though

(iv)

to

make the statements more

introduce

this is not essential

V. 2

Main

properties of determinant

41
can be

the ordinary notation of summation. written

Thus

(1)

ii

(i)

^f- />

and

(4)

can be written

(4)

(v)

The

set out in

tensor notation involves five steps, which are 2-4, 6, 8 below. The reader will find it

helpful to copy out the statement in (iii), modified as in (iv), and make the successive alterations which are now to

be described.

Eeciprocal determinant. The first step (which modern text-books) relates to (1) and (2). The determinant D* has sometimes been called the reciBut, as has already been pointed out in procal of D.
V.
2.

will be found in

not a true reciprocal, since the product of m Since, however, is not 1 but D' contains m columns, we see that if we form a new
(i), it is

IV. 6

the two determinants

duct of this

determinant by dividing each element of it by new determinant and D will be 1.

D the proWe
there-

fore write (1) in the

form

/>=!,

2,...i;\

i,2,...*

/;T

yD

ps

(1 if ?

*~KHf
defined

and form a new determinant

D"

by
(C)

This

new determinant

will be equal to

D'/D

m=

1/D, so

42

Reciprocal determinant
;

V. 2

that the product of the two determinants will be 1 and the cofactor of /D in the new determinant will be

dpi/D

pg are so related that their product is 1 and that the cofactor of any element of either determinant, divided by the

=d

Dp8

D".

Hence the two determinants

and

D"

determinant, is equal to the corresponding element of the can therefore call each determiother determinant.

We

nant the reciprocal of the other.

V.

3.

Elements of reciprocal determinant.

is

The

next step

to have a single
(cofactor of

symbol d in D)-i-D.

for

* ~^
;

We

have already used

Dpg for the cofactor of d

new
by

and

it is

inconvenient to introduce a

letter in place of

or

D.

We
We
(3),

therefore denote the above expression

accordingly, in our statement, replace (B), and the second line of (4), by

(1), (C), (2),

= (cofactor of dp8 in

I))

+ l)

(B)

l/'=\dv\

(C)
(2)

DD" =1
d
(p

= ps
. . .

(cofactor of df* in

D")

1)"

(3)

2,

m)

Xp =
and

2 &*
(3),

YS

[2nd

line of]

(4)

The

parallelism of (B)

and of the two

lines of (4),

should be noted.

V.

4.

Set-notation.

(i)

The next

step

is

to abbreviate

(4), as altered.

This contains two statements

a hypothesis

V. 4

(ill)

Set-notation
so that

43

and a conclusion; they are similar in form, need only consider the first one, namely

we

The expression on the left-hand side of this state(ii) ment has a definite value for each value of s; we can
denote these values by
.

-^I

^2 ^3"'^Wl-

The statement then takes the form

(*=

1,

2,...^=7
c

cannot merely omit the (*=l, 2,...m)' from this, without leaving it doubtful whether we are speaking of some particular s or of each & We get over the difficulty

We

by omitting the
in
*

'

(s=

1,

2,...m)

by a Greek letter. s s that a statement of the form

E =Y

'

and replacing the The convention then

'

s
is

means that
i.e.

is s

equal to

na = L. Y for each
8

of the values of

s,

that

E\
it

^11

^2

= Y2>

^3

= Yz>~- Em = Ym
t

being understood that the values 1, 2, Z ...m which are to be given to a Greek letter have been settled before-

hand and remain the same throughout our work.

(iii)

in (4), it

Applying becomes

this convention to the

:

two statements

if

p=

2v- x P= r1

then

**=

2 <***;
.v

4)

It is immaterial

what Greek

letter

we

use in either of

these statements, provided the letter is the same on both sides. could have used the same letter in the two

We

44

Set-notation
it is

V. 4

(iii)

statements, but in the particular case the letters different.*

better to have

We have
put
(iv)

rather spoilt the

6.

symmetry of
is

(4),

but we will

this right in

The statement

in (1)

a statement as to the value

of a certain expression for all values of p and all values of q and, so far as the left-hand side is concerned, we could extend the above principle by using two Greek
;

letters.

But the right-hand

side presents difficulties

and
(

we must
V.
5.

therefore leave this over for later consideration

It is desirable at this

8).

Principles of set-notation.

stage

to consider the principles underlying the notation

which we have

just adopted.
(i)
i.

e.

Take first the case of a single set of m quantities or elements an aggregate of m quantities which fall into a certain linear

arrangement.

We denote

these by, say,

-"-a

A\

A .mA m
3

We
is

have settled that a statement such as

a comprehensive
l

way

of saying that
2

Thus

A = Ev A E. A 3 = E ...A m = Em we use Ap etc, when we are referring to a particular member

2>

s,

we use A^ etc. when we are making a statement with regard to each member of the set in turn. We may also want to speak of the set as a whole. It will be found that no confusion arises from using A\ in this sense also. We can thereof the set, and
fore say that

A^

(A!

A A s ...A m )'
2

the brackets being used in order to show that we are considering the set as a whole. In this sense we might regard the statement

A\

= E^ as meaning that

the set

A^

as a whole

is

equal to the set

* I have as far as possible used X, p, ^ a- in this chapter to correspond to p, q, r, s, reserving v for product-sums ( 6). Later on ib is better to have no fixed rule, beyond that laid down in 5 (iii).

V. 5

(iv)

Principles of set-notation

45

E^

parts.

as a whole, this equality of the wholes implying equality of the shall have to take this step later (Chapter VI) : for the E\ as present it will be sufficient to regard the statement A^

We

merely an abbreviated way of saying that A^

= Ev A = E
2

2,

etc.

(ii) The interpretation of A\ as the set as a whole recalls the idea of a vector as the resultant of a number of components. But

the operations which we shall have to perform \vith single sets do not follow exactly the same laws as those which govern operations

with vectors as ordinarily understood the analogy must not be pushed too
(iii)

(see note to (vi) below), so that

far.

Next take the case of a double set of

consisting of

single sets, each containing

m
,

quantities, i.e. a set elements. If we

denote the elements of the gth single set by ...Fqm qiJ q2 this single set as a whole can be called and the complete
,

F F

double set can be called We think of the elements as F^p. arranged in a square, the columns of which are the single sets by regrouping (e. g. lr 2r ,...Fmr ) we get the rows of the square. We have already, in 1, adopted the convention that in dqr or drq the q represents the column and the r the row and similarly we
:

Fqp

shall say that in F^ ? or betical order, means the

F^ etc. the first letter, according to alphacolumn and the second the row.
^tm
77T

Hence

fu.o\r\\

77?

77T

77*

-T-21

XT -^31

TTT

\i-Fnu.-

f\n\

^im

Fmf'F

the brackets being inserted, as before, in order to show that the set of quantities is in each case to be regarded as a whole. Then
the statements
F
**

pp~

==
fJLp

(*/J,p*

pl*>

mean

respectively that qr = Gqr and that value of q taken with every value of r.

Fqr = Hrq

for every

A double

set is

symmetrical

if it is

not altered by interchanging

columns and rows.

(iv) A particular form of double set is obtained by multiplying together every element of one single set and every element of another single set (of the same number of elements). If these two sets are B^ and C (in this order), the representative element of

46

Principles of set-notation
,

V. 5

(iv)

the resulting double set will be B Cr so that the double set can q be represented by B^ C or C B^. This double set is called the p p product of the two single sets. It should be noted that we must

not write

partly because this would not define the particular arrangement of the elements of the double set, and partly because we shall presently have to give a special meaning to these latter expressions.
it

as

'

'

B^ C^

or as

B p Cp

'

(v)

single quantities, such as a or

In addition to double sets and single sets, we have to use Te. Any such quantity is called
:

a scalar. It need not be a constant it may, as will be seen be a definite function of the elements of one or more sets.

later,

(vi) We shall for the present be dealing only with expressions which, interpreted according to the laws of ordinary algebra, are obtained from scalars, single sets, and double sets by addition, subtraction, and multiplication. The rule of interpretation is the

same that we adopted in (iv) for B^ C p we replace the Greek X p p. ..by p qr... and take the total expression to be the set obtained by giving to each of the quantities p qr... each of the values 1 2 3 ... m. For example
:

letters

(a) JcAp

means the

single set

whose elements are

',

n'A^
(b)

kA

2)

...KA m

AK

bB^ means

the single set whose elements are

A,bBvA
(c)

bBv...A m

bBm

',

A^p-aFpGp means

the double set

whose element
.

in

the qth column and the

Hh
;

row

is

A qr -aFr Gq
is

It is obvious that this system of interpretation .with the laws of ordinary algebra for instance

in accordance

and

so on.

We
*

are further restricted, in the case of expressions containing

sets follow the

(vi) (a) and (b) and from (iv) that single same rule as ordinary vectors as regards multiplication by a scalar, addition, and subtraction, but not as regards

It will

be seen from

multiplication together.

V. 6

(ii)

Principles of set-notation

47

more than one term, to (1) scalar expressions, (2) single sets arising as sums (' sum of course including difference ') of expressions which contain the same letter, e.g. aA^ + bB^ + ..., (3) double sets arising as sums of expressions which contain the same
'

'

pair of letters, e. g. aA^ v + bBv C^. consider such an expression as A^ +

set,

We do not therefore have to Ba which is really a double

,

not a single

set.

symbol have and with the ordinary meaning of the word, been placed below the line the exception being the use of dP s to mean (cofactor of dps in D) -T- D. This latter system of having upper suffixes as well as lower suffixes will sometimes be found convenient. We may, for x and in that case instance, want to denote a single set by A A 1 A 2 ,... would be members of the set. Where there is any risk
(vii)

The

suffixes

which we have

so far attached to a

usually, in accordance with the regular practice in algebra

all

of confusion, we shall not use the ordinary indices of algebra at thus the square of Ap will be Ap A not A z p p
;
,

V.

6.

Product-sum notation.

(i)

Our next simplifica-

tion consists in dropping the sign of summation in (1) and (4). But, since merely to drop it and to replace, say,

2d
s

ps

dq 8

ky

d>P

dq S

would be misleading, we use

special notation.
is

The number of alphabets

it will

at our disposal

be found not only that we can use for this purpose without risk of error, but Greek letters
limited
:

and

that there are actual advantages in doing so.

rule we adopt is that, when an expression of the (ii) The form JBpC has to be summed for the values 1, 2,...m ofp, we denote the result by replacing p by a Greek letter

in both places
expression as

and, conversely, the

meaning of such an

Bv Cv is
a
.

(V.6.A)

48
(iii)

Product-sum notation

V. 6

(iii)

of any

C be replaced V V could pair of z/s in other identical Greek letters; e.g.

The

by a pair
(V.6.1)

Bv Cv = Bp Cp.

can The v (or p) is for this reason called a dummy. think of the sum represented by v Cv as the result of link-

We

ing the elements of Bv with the corresponding- elements Cv we can therefore describe a Greek letter which occurs twice as a linked suffix, and one which occurs once
of
;

only as a free suffix.

(iv)

The

rule in

(ii)

applies

if

either or both of the expres;

Bp and C has a free suffix as well as the p e.g. A v B^ means A B^ + A 2 S^ + ...+ A m B m and AXv Bpv means AXI B + A^ B + + A Xm Bpm which is a double pl p2 B set whose typical element is A qv (v) We can also have successive summations expressed
sions
l
,

...

rv

in the

same way.

Thus

*W-.
involves summations with regard to A, with regard to n, and with regard to p. It is easy to show that these sum-

and

mations can be made in any order: e.g. we can take C^ I) together as if the A and p were free, and then

bring in
(vi)

Bx and E

ffp

As an example of the brevity effected by this we may take the expression for the product of determinants. Even for so small a value of m as 3, the
notation

determinants

expression obtained in IV. 5 is formidable.

2's,

(ii)

We

product of two condense it by introcan

for the

ducing
it

In the new notation clumsy. will be found that the method of IV. 5 (ii) gives
but the result
is

*
I
! | |

=
|

*xr

I-

(v. e. 2)

Vi 8

(ii)

Product-sum notation
:

49

The
I

process can be repeated

e.g.

X
I

X
I
I

V
1

dgr

=
I I

V V/<Ar
respectively)

I-

'

6 3)

(vii)

The

result of applying- the product-sum notation to

(iii)

the statement in
s in (4)

is

that (1) and (4) become (p and

a-

being replaced by A and

//>=1,2,...;\ v? = i, 2,...m
If
V.
in
7.

r V _(li toi
then

4^=1;,

rx = ^7,.
(i)

(4)

Inner products of

sets

The quantity

BV CV B

behaves

like an algebraical product. call it the inner product of v and Cv to distinguish it from an ordinary or outer product such as x Cp ( 5 (iv)). The inner product of v and Cv is

many respects

We

the

of the elements in the leading diagonal of the outer product of B^ and Cp .

(ii)

sum

In the same

way

B^v, and
(iii)

A Xv Bvp

A V B^V

is

the inner product of

Av

and

is

the inner product of

A Xv

and

Bvp

The process of forming an inner product,

inner multiplication.
8.

as above,

may

be

called

V.

Unit-set notation.
form of
(1),

(i)

We
l

have

finally to con-

sider the

which

is

a statement that

tf=
a statement

So

far as the left-hand side is concerned, this is

as to the values of the elements of a double set

As regards the right-hand

into
;

side,

however, the statement

falls

two first that cl?" d = 1, and next that d? a d = if pa qa and q are different. We want to replace these by a single

statement.
(ii)

We

do this by converting the statement into one as

50

Unit-set notation

V. 8

(ii)

to the equality of two double sets. For this purpose we construct a set whose typical element, in the qih column if they are and rth row, is 1 if q and r are the same and
different
;

set, in

other words, which has

for each
else.

element of

its

we

call this

leading diagonal and set *

everywhere

If

U
I

then our definition of

I

*j

|^

is

that

the function of p and ^ which

We

can therefore write

(1) in the

form

/^=1,
V0
or, in

l,

2,...m;x 2,...m )

_ V-

the set-notation,

V.

9.

Properties of the unit

is
I

set.

(i)

We
)

have defined

as

the set whose typical element

{p
__

1 if

=p

(V. 9.

A)

Hence each element of the leading diagonal

other elements are all
;

of
|

is

1,

and the

in other words

O...OY
1
1

(V. 9. B)

0...0 0...0
6... i>

* J. E. Wright The usual symbol, adopted by Einstein, is Invariants of quadratic differential forms ') uses T;^. Neither of (' these seems sufficiently distinctive and 6 already has a considerI have therefore altered the symbol to able number of other uses.
;

8;

(*

unit

Xji'),

as

an experiment.

V. 10

(i)

Properties of the unit set

51
The

This set (with any pair of letters) will be called the unit set. following are its chief properties.
(ii)

The

set is symmetrical,

i.

e.

(V. 9. 1)
(iii)

The determinant
II?

of the set
1

is

0...0
1

(V. 9. 2)

0...0

1...0

6...1

(iv) Also, if we multiply together this determinant and any other determinant, in either order, it will be found that we merely reproduce the latter, i. e.

lH Vl
(v)

x ll? 1=

Vl'

(V.9.8)

The

special importance of the set, or of

t

of

it,

lies in its effect

when combined with another

a product-sum. It will be found that, values 1, 2, 3, ...m,

any column or row set to form having any one of the

jXj

__

/*

=A

1^.4

=\^A.

=A
9. 4).

(V 9
Then

5)

[For example, take

= 3 in the first part of (V.

A1 +
0.

=
Thus the
set
effect of

0.

A +
2

A3 +

0.

AI +

...

which contains

of a single or double inner multiplication by is to alter the latter to X (or /*). /* (or X)
|

V. 10.

Determinant properties.
results, there
it is

(i)

Before

we

write

down the final we shall find

another small change which convenient to make. In the statement

8
(ii),

= |> if^dpff
upper
cr

obtained in
<r,

the linked suffixes are an

;

and a lower

which cancel one another

and the

52
free suffixes

Determinant properties

V. 10

(i)

A and p are in the same respective positions It is desirable that, whenever possible, these two conditions should exist. In each of the statements d^ a XK = Y and X^ = dx<J Yot given at the end of 6,
on the two
sides.
ff

one of the conditions exists but the other does not.

We

make them both

exist
x

by

replacing- Xx
l

throughout, by

X =(X
as explained in
(ii)

X ....X m ),
z

5 (vii).
after
8,

Our statement,

carrying out the alterations

indicated in

2-4, 6 and

and in

(i)

above, becomes

Notation.

dP*

l>=\dqr
D" =
d4
|

\.

(V. 10. A)

(cofactor of dp8 in
r
\.

D) -f- D.

(V. 10. B)

(V.10.C)
(V. 10.
1)

Properties.

dXa a = d^ \^' DL" =

1.
;
.

(V. 10. 2)

d
If

= (cofactor of dP* in i>") ^ 7>' p8 Y = d X\ then X* = (t"Ya

fT

(V. 10. 3)
(V. 10. 4)

Xff

To

these

we may add

the formula (V.

6.

2) for multiplica-

tion of determinants,
I

namely
l

Vl*l Vl = *Arl-

(V.10.5)

V.
that

11.

Example of method.
are

Toil lustrate the methods

we

means

able to use, let us verify (V. 10. 4) It is given that of (V. 10. 1).

now

by

To

find the value of

&* Yat
it

the above value as

not do to replace Yff by stands, since we should then have

it will

V. 11
three
A.'s.

Example of method

53

We
5),

must

first

replace

A.

Iff

by some other
9.

suffix,

say p.

We then have, by

in the expression for (V. 10. 1)

and (V.

which
it

is

what we wanted

to prove.

The reader

will find

instructive, for comparison, to write out the proof in the ordinary notation. In the above proof we have proceeded from dXa (d 3F) K to (d I*. It has already been pointed out, in 6 (v),

d^)

that summations in a case of this kind can be

order.

made

in

any

SETS

VI.
VI.
in V.
1.
1 1
,

SETS OF QUANTITIES
(i)

Introductory.

It

may have been noticed that

:

in deducing (V. 1 0. 4) from (V. 1 0. 1 ) and (V. 9. 5), we made no direct use of determinant properties the only indirect use being in the relations between elements and
their cofactors, from

which (V.

10. 1)

we can dispense even with this indirect use.

tions

dKa Xx

= Ta

known; and we can

namely

the values of d^ a treat the statement in (V. 10.

was derived. But In the equaare supposed to be

1),

in terms of as a set of equations giving the values of d for instance, those of dxa 20, the set dKa contains If,
.

x<J

m=

400 elements, and (V. 10. 1) is a condensed statement of the 400 equations (each with 20 terms on one side) which Xa Thus for p = 2 we should give the 400 values of d
.

have

dlz <P*

+ d^d + + d^

. . .

dlm d* m

^+
21
,

.-

+ d*m d* m =
+ *<F* =
d 23
d 2m

-which give
Similarly for

the values of d

22
,

. . .

i.e.

of d*.

",

d*

a
,

etc.

have, in fact, arrived at a position similar to (ii) that reached at the end of the first chapter. started

We

We

with the problem of solving a set of simultaneous equations,

and shrived

at

a probable solution, involving

what we

58
called determinants.

Introductory

VI. 1

(ii)

To

verify the solution,

we had

to

The determiinvestigate the properties of determinants. nant thus took the leading place, its applicability to the

A determinant
which
for

solving of equations being one only of its properties. of order m is based on a set of m 2 quantities,

convenience of reference are thought of as

arranged in a square, the determinant being expressed by enclosing the set of symbols of the quantities between vertical lines and we have reached the stage at which the
:

set of quantities

as

becomes the important thing, its existence the basis of a determinant being one only of its

properties.
(iii)

These properties we have now to consider.

The

following sections of this chapter are mainly a restatement, with obvious modifications and extensions, of results obtained in the preceding chapter.

&C~v(J l/-</C
VI.
2.

Single

sets.

(i)

We may

have a single set

Ap = (A
The
separate quantities are called its elements.

A 2 A 3 ...A m).
comprised in the set
is

A 1 A 2 ...A m

The order The of elements comprised in it.

regard to such a set
is

of the set

the number

typical statement with

of the form

Af
This, in the
first

= Ef

instance,

we regarded merely

as a short

way

of saying that

A l = El A 2 = E2 ,...A m = m
,

-,

but we must now think of

sets

it

as a statement that the

two

and

each taken as a whole, are equal, this

equality of the wholes implying the equality of corresponding elements. The analogy of a vector may help us here.

VI. 3

(ii)

Single sets

59

The statement that two

think of
is

vectors are equal implies that the

;

components are equal, each to each

but what

we

really

not the separate equalities of the components but the equality, in all respects, of the vectors.
(ii)

Single sets behave like ordinary vectors as regards addition and subtraction and multiplication by a scalar
(

(iii)),

but not as regards multiplication of one single

set

by another.
3.

VI.
order

Double
i.e.

sets.

(i)

We may have a double

A
21

set of

comprising
J / ^npl n
A

m2
A

elements
^31
J ^L m

^12
I

^22

^32
t

4*

A^ m .. A mm ^ Here the quantity in the $ih column and rth row is A qr We adopt the convention that the first Greek letter (in

\A lm

A 2m

alphabetical order) represents the column and the second

the row, so that

fit

13

m
.

*
."

and

The sets the representative element of which is A rq A are called the transposed of each other.
.

pfjt

(ii)

The determinant
the determinant of the set* the determinant of

is called
|

and similarly

A rq

is
|

* The set is usually called the matrix of the determinant. It is a singularly inappropriate name, as the symbol of the set is the inner part of that of the determinant, not something which surrounds it. The set is really the substance or core of the determinant.

60
(iii)

Double

sets

VI. 3

(iii)

The brackets

in which the elements of

A p9

A^ A

pti

have been placed are not essential,* and have been introduced partly to help the eye and partly to indicate that the
sets are
(iv)

being considered as a whole.

A double set is symmetrical if columns and rows can

if

be interchanged without altering it. Thus, and conversely. metrical, then A "=.A
;

A^ p

is

sym-

Sets generally. (i) We describe a single set as being of rank f 1, and a double set as being of rank 2. a set of rank 3 of order m is made up of m (ii) Similarly Thus we might represets of order m ; and so on. double 3 by There would have to be sent a set of rank
VI.
4.

A^
l

a convention as to the order of the symbols, so as to disfrom A VVL etc. Where, however, the set is tinguish

A^ v?

symmetrical, so that not arise.

(iii)

^
(Jivp

=A
is

,^ p

=etc., this difficulty does

The

set of

rank

a single quantity or scalar.

to its

To denote a set generally, without reference suffixes, we use a Gothic letter such as J3 or 25(iv)
*
is

An alternative method, in the case of a double set or matrix, to enclose the symbols between two pairs of vertical lines, so as

to distinguish the set from the determinant, which has two single lines. It is not a satisfactory symbolism from our point of view, as it would seem to suggest that the set is more restricted than the determinant, whereas what we are aiming at is to free the set from the bonds of the determinant. t I have been doubtful as to the appropriate word. In reference to tensors Einstein uses Rang, Hilbert Ordnung, Eddington rank, de Sitter order. The objection to either of the last two is that there is already a settled meaning for order as regards a determinant, .and (though this is not so important) for rank as regards a matrix 3 (ii)). It would seem reasonable to describe a set >(see note to

w/"

containing mf elements (i. e. composed of m sets each containing 1 elements) as being of degree f. I have, however, felt bound to keep to Eddington's use of rank.

VI. 6

(i)

Sums and

products of

sets

61
(i)

VI.

5.

Sums and
we add

products of

sets.

If two or

more
same

sets, of the same rank and the

suffixes,

tracting)

same order, have the them by adding (or subcorresponding elements. Thus
(or subtract) for sets of

and similarly
(ii)

higher rank.
set,

We

multiply a

when we multiply every element

e.g.

of whatever rank, by a scalar of the set by the scalar

;

fC ^L-in

/f^jfnn

"^^9

"^^9

Thus the determinant of kA

the determinant of
(iii)

is

m not k times, but k times,

with

If SL and 25 are two

sets,

different suffixes,

g respectively (/and g not being necessarily J& is the set of rank f+ff obdifferent), their product tained by multiplying every element of one by every

ranks

f and

&

element of the other.

all

the sets

Thus
set

(Here, as elsewhere, we assume that are considering are of the same order.) the product of two single sets and J3ff is the double

we

obtained by giving to p and o- separately each of the values 1 to m. product obtained in this way is

A p Ba

sometimes called an outer product, to distinguish an inner product as defined in 6 below.

' '

it

from

VI.

6.

Inner product.

(i)

When a suffix
C V V
,

occurs twice
in

in an expression such as

A vv

or

or,

more generally,
or JB

any

(where the
expression
suffix
;

single expression or product, e.g. A Xvvpttf vp ...Cvffttt letters may be in any order), this means that the
is

to be

summed

for the values 1, 2...m of

the

e.g.

62

Inner product

VI. 6

(i)

BV CV = JV?i + B2 C + ...+Bm Cm BV D, V = A Ai + Z>A2 + + Bm D, m

2 2
. . .

)
.

(VI.

6.

A)

etc.

a letter occurs twice in this way, each of the Where a letter occurs once only, letters is linked.

Where

two
it is

free.

The linked

suffixes are called

dummy,

as they can be

replaced by in the expression.

(ii)

a pair of any other suffixes not already occurring-

In the particular case where the expression to be summed is of the form B V CV the result is called the inner product of Bv and Cv , or of Bx and Cp etc. It is immaterial
,
,

use in this latter description, since they have to be replaced by one and the same suffix.
suffixes
(iii)

what

we

j5

From a pair of double sets A^ and B^, or A X(r and we can by a single product-summation form several

different double sets

is also

two summathe scalar quantity A B^, formed by if sucwhich can be simultaneous or successive tions, the first is a product-summation giving us the cessive,
;

A Bvp A x? B^ A^ BXp
,

etc.

There

double

set

A^B^
it

or

scalar quantity

A B

A^B^.

Strictly speaking

this

is Ike inner product

of

A^

and

Bpp

But

occurs less frequently than the double sets

obtained by a single summation, and it is therefore more convenient to call one of these latter the inner product.

We shall call A^p B^p the complete inner product of A^ and

B^p
.

By

for the product of

analogy with the expression found in V. 6 (vi) two determinants, we define the inner
sets

product* of two
*

and

B
B

or,

more

generally, of
'

The

This is what, in the case of matrices, is called the product '. true product of two sets An a and Xp is a set of rank 4. If this use of inner product seemed likely to lead to confusion with the * complete inner product', we could use a different phrase, such as interproduct '. It should be noted that the inter' '

VI. 6

(v)

Inner product

63
letters are in

two

sets

ff

and

Bx

in each of

which the
1

alphabetical order, as the set A^V V? \ the linked suffix in this latter expression being chosen so as to
their proper

be (alphabetically) intermediate between the two free suffixes. This is equivalent to saying as regards this case, that the inner product of two double sets is the double set
1

whose element in the qth column and vth row is the inner product of the t\th column of the first set and the ith row of
the second set
;

and we apply this rule to

all cases.

The

simplest way of applying it is to use the result for A^p and and alter the order of the suffixes where necessary.

Bpp

Suppose, for instance, that

we want the
.

inner product of

A^

and

JB
pfA

then by writing

inner product is that in all cases the inner product depends on the relative thus the inner product of position of the original sets
;

B =F we see that the A ^v Fvp = A ^v Bpv It should be noticed

A^ Bvp but B^ A vp
;

B^

and

A^

is

not

in (iv) There are four main forms of innerfproduct constructed accordance with (iii) and four others, which are really repetitions, can be obtained by interchanging the two sets. Denoting the

inner product of & and 33 (in this order) by <E x B.^cf. IV. 5 tb^product of two determinants), the forms are as follows
:

(ii)

as

A
tjip

xB^p

= A flv Bvp = A v^Bp V

(1)

BppxA^p
Bpp x
pfJL

= B^Ayp = A vp B^ v

(5) (6)

AppxBpp^A^Bp,, = A VpBvp x Ap/j. B^p

^Pt*.

(2)
(3)

Bpp

(4)

BppxApp^BrpAyp-ArpBvp A p ^ = B^v A pv = Ap V B^ v B x A = B A pv = A pv Bv^

pfJL VfJi

(7) (8)

(v) The transposed of an inner product such as A^ v BV p is found in the usual way (V. 5 (iii)) by interchanging the free suffixes p. and p. By comparison of (1) with (8), (2) with (7), etc., it will be

seen that the transposed of the inner product oftivo double

sets is the

mediate product-sum, for A^B^^ is not A^ v BV p but either A\pB^ p or A^B^. There are various reasons for taking the former, rather than one of the two latter, as the inner product.
*
'

64

Inner product
sets,

VI. 6

(v)

inner product of the transposed posed of the inner product of

in reverse order; e.g. the trans-

A
pfJL

and

B
pfji

is

the inner product of

Bpp and A^p.

(vi) It will be seen by comparison with IV. 5 (ii) and the Appendix that the rule for construction of the inner product of two double sets is exactly the same as that for construction of the product of two determinants so that the determinant of the inner product of two double sets whether we call them (say) A^ ? and B or A^ and B pK is equal to the product of the determinants of the two sets.
;

pfJL

VI.

7.

The unit

set.

(i)

The unit set*

is

defined as the set whose typical term

is

(VI.
so that
(I
1
1

7.

A)

0...01

(VI.

7. 1)

0...0

0...0
6...i,
is

666
(ii)

From

the definition

it

follows that the set

sym7. 2)

metrical, i.e.

!=!'
and that

(VI.

||*|=1.

(VI.

7. 3)

of this set is that, if A^ is (iii) The special property set (possibly containing other suffixes o- r...), then

any

1^ = 1^ = 4,,
*

(VI.

7.

4)

so that the inner product of the unit set and any other set This
is

by analogy with the

'

unit matrix

'.

VI. 8
is

(ii)

The unit

set

65

the same as the latter but with the suffix changed. In other words, the unit set acts, for inner multiplication, as a substitution-operator.
(iv)

In particular, the inner product of two unit

set, i.e.

sets is

a unit

7.

5)

VI.
set

8.

Inverse double

sets.

(i)

We

take any double

rrU<
and we say that there
is

another set

connected with

it

by the condition that the inner product

is

of the former and the latter

a unit

set, i.e. (see

(iii)

and

(i))

that

^^=IJz

(VI. 8.1)

This represents m equations, which are sufficient to determine the m 2 values of AT(I when those of A are known.

The

set

qr

pf

*,

as defined
.

by the above condition,

is

called the

inverse of the set A^p

that (VI.
letters
8.
ju,

1)
y

will

keep always hold, however we alter the

We

shall

to this notation, so

A,

p.
is

(ii)

The above

subject to one condition.

If

we

write

the equations which determine the elements in, say, the second row of A pf*, namely

down

A* m

66

Inverse double sets

see that
iii

VI. 8
be a solution

(ii)

we

order that there

may
\

it is
is

necessary that

we should have \A r ^
I

0,

which

the

same thing as

VIgfeO.

(vi. 8. 2)

This applies also to the other rows. It is a sufficient as well as a necessary condition for the existence of A pl*.
(iii)

(VI.
|

8. 1)

Taking it that and (VI. 7.

\

A qr
\

=
\

0,

we have, by (V.
||?

6.

2)

and

3),

AV \x\A rq =

A* v A rv

=
|

|=
0.

1.

(VI. 8. 3)

It follows that

^
(iv)

(VI. 8. 4)

The statement (VI. 8. 1) is a statement as to the by taking each value of p with each It is therefore equally true to say, by intervalue of /*.

m2

relations obtained

changing p and

p,

that

^"^ =
The

i;-

(VI. 8. 5)

expression on the left-hand side is ( 6 (iii)) the inner product of A*? and A pfjl ; and these are the transposed of Apfl and Hence, if 35 is the inverse respectively.

A^

of

#, the transposed of

51

is

the inverse of the trans-

posed of 2B(v)

The

relation in (VI. 8. 1) is a relation connecting

columns of the original set and rows of the inverse set. There is a similar relation connecting rows and columns.
For (VI.
8. 1)

gives

whence, as will be shown in

(v), it
l",

follows that

^
and hence
also

^=

(VI. 8.6)
8. 7)

4^""=

1-

(VI.

VI. 9

(iv)

Inverse double

sets

67

The expression on the left-hand side of, (VI. 8. 6) is the vX and AXv which is the equivalent inner product of A
,

of that of

A
is

pl

and

so that,

by the

definition in

(i),

the latter

the inverse of the former.

of

Hence,

if

25

is

the inverse

#, then Si

is

the inverse of 25.

The

relation (VI. 8. 7) is similar to (VI. 8. 5), and shows that the transposed of 25 is the inverse of the transposed of Si.

VI.
sets

9.

Reciprocation.

(i)

Suppose there are two single

Zx and Yx connected by the relation

Then, as in V. 11, we have

i. e.

If

Y = A^X\
ff

then

XA = A Ka Ya

(VI. 9. 1)

to be each single set, in turn, Similarly by taking of a set of second or higher rank, with JA to correspond, we find that

If
(ii)

YaTttt = AXff X$ tttt

then

X^ ttt = A^ YffT>t%

(VI.9.2)

Thus the operation represented by A Ka is annulled Xa the operation A ; and conversely. The sets A Ka and by Ka A will be said to be reciprocal to one another and the
:

process adopted in (VI. 9. 1) have to use very frequently

(iii)

and (VI.
will

9. 2)

which we

shall

be called reciprocation.
is

We

see

from

8 that the reciprocal of a set

the transposed of the inverse of the set, and conversely. If a set is symmetrical, its inverse and its reciprocal are
identical.
(iv)
1

As an example

of the application of (VI. 9,

1),

suppose that

68
for all values of
a-.

Reciprocation
Then, by reciprocation,

VI. 9

(iv)

Xx = A Xa O =
provided that
|

(This is practically the same A^ x by m inthing as saying that, if the inner products of
r
=
\

0.

dependent

single sets are all 0,

then

Xx is

0.)

(v) Similarly, suppose that

A A \v r

v(j

4 n A\v Uva
of
A.

identically,
|

i.e.

for

all

values

and

cr,

and that

Aq r

is
|

not

= 0. Then, by reciprocation, cva = A XV A Xp Dpa = ; DP = ir.

|

Thus we can divide both

This supplies the missing
(vi)

sides of the equation 8 (v). step in

by

AXv

The two

definitions,

and the proposition, used in

are
}

the establishment of (VI.

9. 1)

a)
and from these we deduce (VI.
altered
9.
1).

We

could

have

the

order

in

various

ways.

For instance, we

could have defined |* by (1);

successive values,

thence,

by giving

A.

its

Xx = Axa A^ a JF. have A^A^ = haveZx = l^,

|

and equating coefficients, we should have Ka Also we might have defined A got (A). by (VI. 9. 1), instead of by (B), as the coefficients of Y when the equax x This would give tions Y = A Ka X are solved for X
ff
.

ff

Then,
;

if

or, if

we defined we defined

by (1), we should by (B), we should

which

is (1).

VI. 10. Continued inner products. (i) We can construct continued inner products without ambiguity, provided we adhere to the rule laid down in 6 (iii). Suppose, for instance, that we

VI. 10

(v)

Continued inner products

69
That of

want the inner product of

A Xff BXff
,

and
.

ff

\.

A Xa
X(J ,
ffv
;

and
i.e.

If we a according to the rule, is by (2) of 6 (iv), the inner product of Xa and the inner product of A Xa BXo and C a \ is that of AM, B^ff, CffX and DX(, is
,

BXa

A^B^
,

call this

then,

CffX is

FXv C

A^B^Cw
,

Similarly

A^B^C^D^.
ATA

(ii)

On

is
is

not

A ap Bpv C^D^.
(iv),

the other hand, the inner product of A aX> B aX C\ fft For, by (4) of 6 (iv), that of A aX and
Calling this

BaX
is,

by

BffpApfr 6 (3) of

Cva GvX =

G a\

the inner product of

G aX

and CXa

C^B^A^.

Similarly that of

(iii)

sets is
e. g.

The transposed of the inner product of any number of double the inner product of the transposed sets, in reverse order; the transposed of the inner product of A\ ff , B^ a , Cff\, D\ a is

the inner product of ffA C\e, B ff^ A aX [For we have shown, in 6 (v), that this is true for the inner product of two sets and
, >

thence
(iv)

it follows,

by induction,

for

any number of sets.]

of double
;

The inverse of the inner product of any number

,

inner product of the inverse the inverse of the inner product of B aff
sets is the

sets, in reverse order

e. g.

Caff

a<r ,

Eaa F

is

the inner

product of
aff

E, D, C, B a = B pC0y Dj5 E then A UA

,

i.

e.

Sa ,

= V* &*&&&.
by
a
,

(VI. 10.1)

[Denote

this latter expression (right-hand side)

and
-

alter

^y

d in it to

v p.

Then the inner product of A aff and

F00 is

Hence, by reciprocation,
so that
(v) It follows
,

Faa == A
from
(iv)

ffa
.]

is the transposed of that the reciprocal of the inner product of any number of double sets is the inner product of the reciprocal sets ; e. g.
(iii),

and

since

A aa

aa

then

= Bl>* &*&*&*.
(VI. 10. 2)

70
VI. 11.

Partial sets
Partial sets.
(i) When we are dealing with A^ = (A 1 A 2 A 3 ...A m )

VI. 11
a set

(i)

to consider the separate or mutual relations of groups of the A's. The simplest case is when the set divides into two groups, one group consisting of k elements, which we shall take to be the first k, and the other group consisting of the

we sometimes want

other

m-k elements.

If

we use

suffixes

apy... in reference to

the

i n reference to the second, reserving group, and <f>x^ Xpv... for the set as a whole, we may treat the two groups as partial single sets of orders k and respectively, and write

first

A a ^(A A
l

2 ...A

k ),

m-k A ^(A k+l Ak+2 ...A m

()

).

(ii)

In the same

by

division

respectively.

way a double set A^ p may fall into four groups by two lines cutting off k columns and k rows We could denote these groups by

jj.7 "O.w
being regarded as

^rj, "d>w
m

coming before y. The groups A a y and (f> k respectively. The would be partial double sets of orders k and would each have different numbers of columns groups a and

A^

A^y

a 1, the set so constructed will not in general be reciprocal set as the set made up of the corresponding elements of the the same

and of rows, and therefore would not be double sets but this would usually not matter, as we should be specially concerned with A ay and A^. The important point to notice is that, if we take A ay, say, as a partial set and construct the inverse set A^ a or the
;

set inverse or reciprocal to

is

A^ p

The

inverse set

A7 a

for instance,

given by
1

with summations made only from

to

k instead of from
a
.

1 to

m.

mistake, we may write it (A~* ) k inverse to the partial set may be written

To avoid

Similarly the set

A^

[A^] m _ k

(iii) If,

however,

all

the elements in the portions

Aa^

and A^y

are

0, so that the set

a y and all the elements in A reciprocal set ; the elements of A*l and are just the
partial sets

AW

A^ A^

two A^p practically consists only of the this is also the case for the complete t af and

and A^f are 0, same whether they are

VI. 11

(v)

Partial sets
set

71

regarded as obtained from the complete

sets

A ay

A^ or from the partial

and

A^.
or

(iv) If

we had two

more

sets,

single or double, divided in the

manner described above, we could take portions from different sets to form new sets. If, for instance, we had divided A^ into A a and and B^ into B a and B^ (orders again k and m k), we could A<p, construct a new set consisting of A a and B^.
(v) It

being

add together two sets of different order. We could, however, look at the matter in another way. Consider the two sets
(A,
(0

A a + B<p,

would, of course, be incorrect to describe this or the old set as being Aa + for

A^:

new set as we cannot

z ...A

...0

),'

...0

A k+l A k +

2 ...A

m ).
;

The sum
and in

of these, if this sense, if

regard a and

as

we regard each as having the suffix X, is A k we denote the two sets by A a and A^ and connoting X, we could say that

If

to the projections of

we compare AX with a vector, we see that A a and^ correspond A^ on a plane ', i. e. a surface of the first
'
'

degree, passing through the first k axes, and on a 'plane through k axes, respectively. the last

The extreme form,

which A^ was having m 1 O's in
in
it.

if

we made
into

further divisions, would be that

describe the set

m component single sets, each up would only be in this sense that we could as being the sum of its m components.
split It

VII.
VII.
1.

BELATED SETS OF VARIABLES

Variable sets. (i) In the earlier chapters we manner in which determinants arose in
m) d l8 X^ + dzs X*

considered the

solving a set of equations of the form

(*

2,

. . .

. .

+ dms X =

Y,

(1)

and in the chapter preceding- this we have considered the general aspects of the system under which we express these equations and their solution in the form
l'

=d,a X\

& = #'?..
'

(2)

According to the definitions we gave to the notation, Z and Yff are each used in different senses in the two places

where they occur in (2) X^ means the elements of A' Xt on its on its first occurrence and 'each element of X and similarly for Ya but in the resecond occurrence
'

verse order.

have, however, by this time practically reached the stage of treating a set as a whole, so that we can now regard (2) as a pair of statements, one of which

We

gives an expression for the set a in terms of the set J x in terms of Ya . while the other expresses The form
,

of either expression determines the nature of the relation between the two sets.
(ii)

The

special features of the particular case were that

unknown quantities which we wanted to find, that the ^'s were coefficients, more or less accidental, and
the X*s were
that the T's were known quantities arising from the application of these coefficients to the JTs and, more important,
;

VII. 1

(iv)

Variable sets

73

that this was merely an isolated set of equations, for which we had no further use when we had found the X's.

The relations with which we have to deal in this (iii) and subsequent chapters are of a different nature. We have a set Si and a set 25, each consisting of a number (the same for both) of elements, which we will call the In each set the elements need not be all A's and the .Z?'s.
of the same kind.
it either has, or

Each of the A's

(as in

is

a variable

i.

e.

the theory of statistics or of error) be regarded as having, a very large number of These variables, the values actual or possible values.
of which are algebraically independent,* together constiIn the same way the .Z?'s are tute the variable set Si.
variables,

can

and constitute another variable

of variables

set

25.

But

the two sets

other
:

are

not independent of each

of which the
versely.

they are connected by certain relations, by means J5's are known if the A's are known, and con-

Thus the

-B's

are functions of the A's, in the

i?'s.

ordinary sense of the word, and the A's are functions of the In this case we say that 25 is a function of Si, and
i.e.

Si a function of 25.

think

it;

But we must not only say it, but we must treat the functional relations of

the A's and the B's rather as interpreting the nature of the functionality of Si and 25 than as actually consti-

tuting this functionality.

In the particular case we have been considering, x and 25 were the single sets X and J"A and the relation between them was linear i. e. the 7's were linear functions of the X's, and the X's were therefore linear functions of
(iv)

Si

each

m
is

By algebraical independence of m quantities we mean that may have any of its values, whatever the values of the other l may be. This does not imply statistical independence, which
K

a different thing.
2611

74
the
of
J's.

Variable sets

VII.

1 (iv)

AT*,

In such a case we say that Jx is a linear function x and it follows that X is a linear function of 1\.

(v)

from
the
ff

its applications,

In dealing with the theory of the subject, as distinct we are concerned not with the actual

values of elements of sets but with the relations between

Thus in the case of the linear relation dX(rX^ where the variable single set Y is expressed x in terms of the variable single set X and the fixed double set dX(T the elements of dX(r form a kind of framework
sets.

Y =

ff

A A
A
into

=
==

U A + r/21 A +^31 A "12 A + "22 A + "32 A +

c

+" +^nl
'

+ f m2 A + "mm A
fitted
;

= d\m A + ^Zm A +

^3m

A +

which the values of the X's and the Y's can be

and what we are really investigating are the properties and In the present mutual relations of such frameworks.
chapter we shall consider certain simple relations between two such frameworks, namely relations between the linear relation of one pair of sets and the linear relation of another
pair of sets.

VII.
quantity
1
,

2.

Z is

Direct proportion of single sets. a linear function of m X's, which we

it is
l

(i)

If a

will call

X X z ,... X,
This
is

of the

form
//

Z=t X
l

+ ttX* + ...+A m X*n =

A J\

(1)

the simplest form of statement of a linear relation. is the 3rd difference of the Suppose, for instance, that

Ps, formed in the usual way,

i.

e.
1
.

Z= AAAA'
This
//

is

equivalent to
//

l==-l,

= 3,

Z= J ~3 J + 3 J - A' so that ^=-3, =l, J = A =...=J = o.

4 3 2
1
,

7/

TO

VII. 2

(iii)

Direct proportion of single sets

alter,

75
(1) will

The

^'s

having these values, the X's may

but

always give the 3rd difference.

(ii)

is

suppose that there is another set A and that the same function of the A^s that Z is of the X's

Now

as in the above example, that it is the 3rd difference. Then the ^'s are the same, so that
e. g.,

C=hx A\
We can
write (1)
in the form

(2)

on the understanding that a suffix in a denominator is linked with a similar suffix on the other side and implies an inner
multiplication.

Similarly

we can

write (2) as

Equating the two values of hx we have

,

J*~^
as our

- Z
is

(3)

way

of the A's that

of stating that C is of the -Z's.

the same linear function

(iii)

Next suppose
.1*, so

that a set

is
is

a linear function of
a linear function of

the set

that each of the Y's

the X's.

Then we can take

:

of (1) to be each of the

but the sets of ^'s will be different, so that the relation will be of the form
in turn

Js
J

If there
*

is also

a set

.Z?

p,

each element of which

is

the same

The set might be called either Y? or p ; we choose the latter as giving a convenient symbol dun for the coefficient of X^ (see V.
10

(i)).

76

Direct proportion of single sets

VII. 2

(ill)

linear function of the A' a that the corresponding-

element

of

is

oi^s, then

the two sets of d's being the same.

We

therefore have *

as a statement of the fact that the j#'s have the

relations to the A's that the Y's

same

linear

have to the X's.

This would

be the case, for instance, if the Y's were the successive differences of the X's, and the .Z?'s were those of the A's

according to the same system. In view of the variety of ways in which we are able to
deal with sets according to algebraical laws, it is perhaps permissible to describe this as a case of direct proportion,

and to say that

bears to X**.

J3
p
*

bears the

same ratio

to

that

Yp
A'**

This
is

the operator that

into

ratio', here denoted by required to convert A* into

d^

is really

Bp

or

Yp

(iv)

If the linear relation of the

.#'s to

the

As

is

the

same as that of the

to the -B's is the
(or, if

i"s to the X*s, then that of the A's

X**,

same as that of the X's to the Y's * to A is the same as that of Y to then the ratio of A? to J? is the same as that of p
the ratio of

Bp

d A*. Then A [Let Up Therefore A*/B = X*/Y .]

f
p
*

=&

Similarly

X*

= ti

'

This expression B /A^ must not be confused with B /A^ as p p the double set whose typical element is B r /Ai. The limitation in V. 5 (vi) excludes double sets of this kind from consideration.

VII. 3
(v)

(i)

Direct proportion of single sets

sets of relations
.

77

The above

can be expressed by the

diagram in Fig. 1 The crosses may be taken as representing

either coefficients of the X*s

A3
A2

in the values of the Y's and

of the A' a in the values of the

j?'s 2.
;

or

in virtue of (VII.

1)

coefficients of the

Fs

in the values of the X's and

of the

.5's

in the values of

the

a.

(vi)

Ratios of the kind considered above can be com;

bined according to the laws of ordinary algebra

e.

g.

78
(ii)

Reciprocal proportion of single sets

If the ratio of

VII. 3

(ii)

Bp
i.

to

yp

A^

is

*o

of that of

then the ratio of

to

Yp
-

the reciprocal of that of to X** is the reciprocal

Ap

=
[Let

then

= -.
A..

Then
Therefore

k..

VII. 4

(ill)

Cogredience and contragredience

79

Cogredience and contragredience. (i) In we havenot assumed the existence of any relation x x between A and X or between Bx and YK Where a relation
VII.
2
4.

and

does exist, the important cases are those of cogredience and conx start with a set and a set which tragredience.

We

is

derived in a definite

a function of
tion of

X\ We

Xx
is

way from X*, in other words is then take Yx to be any linear func-

and

5X to

be derived from

Yx in
is

that
A*.

derived from
cases in

Xx

Then
:

BK

the same way some function of

Of the
(1) If

which

this is a linear function,

we

are

concerned with two special classes

always said to be cogredient.

(2)

Bf /A*

is

= Yp/X^ then A x
,

and

XK

are

always = are said to be contragredient.

If

Bp /A*

is

K X*/Yp then A and X*

Examples of cogredience

are

contragredience in VIII. 3 (iv)

(ii)

given in IX. 6 and IX. 4 (v).

(x),

and of

A x and XK are cogredient or x contragredient, we might say that A in the one case varies
Instead of saying that
directly as

X^ and

When we
if

say that

in the other case varies reciprocally as Ax varies directly as we mean that,

X\

X\

multiplied by any double set involving A, A^ is when we say that A x varies multiplied by the same set
is
:

Xx

reciprocally as
this set.
(iii)

Xx

we mean that,
A,

if

Xx is multiplied by any

double set involving

is

multiplied by the reciprocal of

If

^ and 1*
X
.

are

I
{
(

and 1* and

^
;

contragredient )

are also

<
(

>.
)

then

Px and Xx
,

are cogredient

contragredient
are
(

if

and

but
are

P*

aud

J
...

are

contragredient
^.^ P
,

contragredient) &
..

^
(

>

then

,,

and

v A\

contragredient^

cogredient

80

Cogredience

and contragredience

VII. 4

(ill)

[Suppose, for instance, that

K and P* and A

A and X

are contragredient,

are also contragredient.

Then we have

relations of the

form
,

Sf /A" = X"/Yf
whence, by (VII.
3.
1),

qf /P

so that

P* and
5.

Xx are cogredient.]

VII.

Contragredience with linear relation.

(i) simplest case of contragredience is that in which the contragredient sets are connected by a linear relation.

The

(a) Suppose that, with the notation of x is between A and

3,

the relation

Then,
(VII.

if

denotes the inner product (cross-product.) in

AP = ay = OnPP + K + Si)* J?
1

3. 2),

Thus Q

a quadratic in since each of the four sets

is

J J2 J
1
,

3
,

. . .

Jm

i.

e.

in

y.

Also,

X Y^ A^ W"
1

*,

is

a linear function

of each of the others,

can be expressed in a good

other ways,
or

e.

g. as a quadratic in IF, or in the form

many k^A^Y^

4* 44,
(b)

If the relation between

i.

e.

if

= Mp
3.

pM
1),

it

A^ and A can be shown that we

rA*

is

symmetrical,

shall, in addition

to (VII.

have the further relations

*-?- ll-^ Y -X' A^~ W

p
(ii)

Conversely, suppose that

we

are dealing with a set

VII. 6

(ii)

Contragredience with linear relation

81
that

of quantities or coordinates

X* = (X l X*...X m ), and

we come

across

an expression

where a^ v
given by

=a
vfji

Then we may

construct a

new

set

Xp
and we
shall

= a^X",
the system of coordinates linsome linear functions of
;

have

Now
early

suppose
or

we change

replace

the X's by

#'s will normally have some definite meaning and this meaning, though not their actual values, will remain unchanged when the X'a are changed. Suppose

them.

The

that,

when

becomes
1

Y^ X^
~~~

as based

of the #'s

becomes

Y Then we shall Y" 7-yH u* u

*.
-trfji

on this meaning have

and

also
Vrj/

~~
~~VJA

"""

^VV

"l/M
I

'

VII.

6.

Katios of sets generally.

(i)

The

word, ratio

has so far only been used in reference to single sets connected by a linear relation ; if the relation between Y and X**
is

of the

form
call

J"

=rf

JF*,

we

call

the ratio of

Yp

to

*,

and we

d^p the reciprocal of this ratio. extend the use of the word to sets other than single
(ii)

We

can

sets.

We
,

involves a scalar.

have already had examples of a ratio which x Thus in 2 we had relations Z=fr^X
,

C=hx A x

and we said that

-h A K- X-

^A'
x x
;

Here we can quite well

call

hK the ratio of C to A or of Zto X

82
i.e. it is

Ratios of sets generally

the ratio of a scalar to a single
set.

VII. 6
5

(ii)

Similarly in

(ii)

the statement

Y
may be
is

Q
*

1 regarded as a statement that the ratio of Q to Jf Y X v In each case the ratio of one equal to that of Y^ to
.

set to another is the set

by which the

latter has to be multi-

plied in order to obtain the former.

(iii)

The more important

case is that of the ratio of

a variable set of any rank to another variable set of the same rank. If SL is a variable set of any rank, and 25
is

a set which is a function of SL and

if

is

of the same rank,

of the form

and

the relation between 25 and Si

is

B=pa,
a constant set whose symbol contains all the suffixes occurring in Si and 25, then we can call p the ratio

where p

is

of 35 to Si, and denote this ratio

(iv)

by

relation of 25 to Si as linear.
tions,

In these cases we can continue to speak of the Also, by solving the equawe find that there is a relation of the form

a=p'jB,
so that, if }5 is a linear function of Si, then Si is a linear

function of 25.

Here, p being the ratio of 25 to Si, p is the ratio of Si to 25, and we can call each ratio the reci-

procal of the other.

(v;

As an example, suppose
is

that Si and 25 are of rank

3,

the product of three double sets, each of which has inner multiplication with Si. Then the relation might

and that p

be of the form

VII. 7
It is easy to

Ratios of

sets generally

83

show that in

this case

Thus the
is

reciprocal of the product of the three double sets the product of their reciprocals.

VII.

7.

Belated sets of higher rank.

With

the

preceding explanation, there is the ideas of equality of ratios,

sets, to sets

no difficulty in extending and of related systems of

of higher rank. Suppose, for instance, that we have a set SL which is a function of three single sets, and a set 25 which is a
If the three latter sets function of three other single sets. were functions of the three former, 75 would be a function
of SL.

The

cases analogous

to

those considered in

would be the cases in which there were linear relations between corresponding sets. Suppose that X^, Y^ ZK are
X linear functions of ?7 ,

TA

= V7*. r,
that Si
of
is

WK respectively, ^ = 6^, z = c w,
,

e.g.
vr

U\

a certain function (in the most general sense) and that 35 is the same function of 7*,

W\

XH YKy ZK
pose that,

Then 75

is

some function of

SL.

If
are

when

the values of

xp ,^M(r ,^ T

we supmade to

and the ratio of vary, 75 is always a linear function of SL, 75 to SL is always compounded of the ratios of x to IT", of 7X to Fx , and of K to x , each taken directly or reciproin 4. cally, we get an extension of the cases considered

Thus we might have

so that

^=%
6.

each of these latter expressions representing a set of rank

84

Related

sets

of higher rank

VII. 7

as regards

In this particular case we can say that jf is cogredient Ux and 7h and contragredient as regards v or

we can say

that

or that it is directly proportional , reciprocally as regards x and F'* and reciprocally proportional to ^T". to

it varies directly v

as regards

UK

and V* and

VIII.

DIFFEKENTIAL KELATIONS OF SETS

1.

VIII.

sider the cases in

Derivative of a set. We have now to conwhich two sets vary together continuously,

so that there can be a derivative (differential coefficient) of

one with regard to the other, this latter being a single set. The derivative will in all cases be a partial one, since the
elements of the single set vary independently.
(i)

The simplest

case is that of a scalar linear function

Here

Giving p

all

values

to

m we
y

can write this

M_ -

~ Z

and we can regard hx as the derivative of the is of rank 0) with regard to the set X\
(ii)

set

Z (which

Similarly, if

Yr =
then*

^>

'

however, be noticed that in this statement the sign has not the same meaning- in the two places in which it is = ^ rfiX<* used. When we say that d Y^fiX* d^ we mean that dqr for all values of q and r but, when we say that Y p /X^ = d^p, = we mean that for all values of r.

It should,

'

dX^

86
(iii) r

Derivative of a
case of
(ii)

set

VIII. 1

(iii)

<)

A particular X fi X is = 1 or
q

is

where

according as r is

= q or ^= q,

J^=Z\

Here

since the

Xs

vary independently.

Hence

(see (VII. 2. 3))

(iv)

Taking

also sets of

ourselves to linear functions,

higher rank, and not limiting we see that the derivative of

a set with regard to a single set by 1 than that of the original set.

XK is a set of

rank higher

Derivative of sum or product. (i) The sums and products of sets follow the ordinary laws of derivatives of sums and products e. g.
VIII.
2.

derivatives of

=C B. M
X

*AK

(VIH.2.2)

(ii)

As a

the scalar particular case of this last result, take

quadratic form

considered in
account,

VII.

5.

Here, taking

(VIII.

1.

1)

into

we have

This can be verified by expressing Q in terms of the X's and to X^ in the finding the partial derivative with regard
usual way.
(iii)

For an application of

this,

suppose that

for all values of the X's.

By

taking adjoining values of

VIII. 3

(iii)

Derivative of
the result which
v

sum

or product

87

*,

we get

is

expressed by differentiation,
.

namely

v 2a^X = 2b^X

Differentiating again, or equating coefficients,

we

find that

V
VIII.
is 3.

= V*
set.
(i)

Derivative of function of a

If z

a function of y, and

is

a function of

a?,

then we know

that

dz dy dz ~~ dx dx dy

'

and, total and partial differential coefficients being in this

case identical,

^y^x
(ii)

<)%'
is

Now

suppose that

a function of
r

Ax

and Cp

is

a function of
respectively,

Then, p and
that
r

being values of X and p

we know

^C ^Bl
r

^C ^B 2

W
5e.

Hence, giving p and r

all their values,

(vm. 3.1)
rank.
<r...

The argument applies to sets of higher we are dealing with Cpa '~ where the
y

If, e.g.,

relates

to

aspects independent of

B^

then

(iii)

As

a particular case (see (VIII.

.

1.

1)),

'*

.. (VIII3

88
(iv)

Derivative of function of a set

VIII. 3

(iv) is

Suppose that the relation between B^ and

say

AK

linear,

Then, replacing

"

or

Cp by C, we

have

Hence

Thus
this

^ and ^C^^

are contragredient.

We

can express
are contra-

X by saying that A^ and the operator ^/<)A

(v)

The determinant
;

of

is

the Jacobian of

with regard to A^

i.e.

(VIII.

3.

4)

From (VIII. 3. 1), taken with (V. 10. 5), we have the ordinary formula for the product of two Jacobians
:

(VIII.

3. 5)

VIII.

4.

Transformation of quadratic form to

sum

case in

of squares. (i) For an example of a Jacobian, take the which a quadratic form is to be expressed as the sum
Let the
/
1 \

of the squares of linear functions of the variables. quadratic form be s\ -y-u vp

where

= a^ATX?,
=
-

(I)

Let
(2)

VIII. 4

(i)

Transformation of quadratic form

89

be a set of linear functions of the X's, so that

XH
Suppose that the
1

WY

a.

(3)

#'s

are chosen so that (1) gives

2

Q=Y Y + Y,Y +... + Ym Ym = Y

1

Y,.
/

(4)

Then, by substitution from

(2),

Hence, by comparison with (1)

(see

(iii)),

The Jacobian which we should

that of

usually

want to

find is

ZM with

regard to

ff ,

i. e.

J=^

%
and

(3),

therefore

But
and

(5)

gives

therefore

Hence, combining this with

(6),

Similarly the Jacobian of

7>

-s-M-^2"'
-Mlt/

ff

with regard to
\

X^

is

_/^

I-

f\
I

There are a good

many

different

in (4), but they all give the

ways of expressing same two Jacobians.

as

90
(ii)

Transformation of quadratic form

Hence we

VIII. 4

(ii)

easily obtain the value of the multiple integral

pOO
pOO
...

pOO

D=\ -00 J
Using

e-* Q

dX dX*...dXm.
1

(9)

J -00

1-00

(\)

to denote

...(m times),

we have

(10)

(iii)

We

shall require, in the next chapter, the value of

JT/D,

where

and

is as above. We find the value of N, as we have found that of D, by expressing everything in terms of Y's. By (2) and (4), and (VIII. 3. 1) and 2 (ii),

and therefore, by

(3),

.dY^
Let us write

(12)

M^jl^^Y^Yt^dT^T^^T^
so that

(13)

VIII. 4

(iii)

Transformation o/ quadratic form

91

Then M t consists of m terms due to the m values of o-. Since Q is the sum of the squares of the T's, the only term which counts in the integration is that for which o- = t. Also we know that

JHence
it

follows that

M
and
therefore,

=Jw(^
(14),

e-WdY dY,...dYn
l

by

N/D =

= !?

(15)

IX.

EXAMPLES FKOM THE THEORY

OF STATISTICS *
1.

IX.

Preliminary.

(i)

The
3

special feature of a statis-

tical set

AA
of the kind which
that each

= (X X X
1

...

Xm

),

we have

to consider in this chapter, is

X has

or possible values,

one only of a very large number of actual which together constitute the field from
;

which the

is drawn and the fundamental facts with which we are concerned are the relative frequencies of occurrence of the various possible combinations formed by

taking an
variables,

X from each of the m fields. Thus the X's are and the expression for the relative frequency of
X2 ...Xm
)

joint occurrence of a particular set (Xl the X's of the set, with certain constants.

involves
class

In a large

of cases the constants depend on the

mean values

of the X's

and the mean squares and products of their deviations from their respective means. It is to these cases that the new notation is specially applicable. It may be that some
of the X's are

drawn from the same

field

we

shall proceed

as if the fields were all different, but this does not affect

the validity of the reasoning.

(ii)

We shall only
The
. .

consider two kinds of cases.

is

(a)

first

kind of case

where our

statistical

information relates to a large number of individuals, and X, the measures of specified attributes, such . are 2 l 6

XX

as height, head-length, chest-expansion, intelligence, etc.,

* I have dealt with the problems of this chapter in as general terms as possible. The explanations in small print may help to show the statistical student the way in which the problems

actually arise.

IX.

(iii)

Preliminary
' '

93

of any individual. Here the variability of any has reference to the different values that it takes for different
individuals.
lar values of

1 2 may be a complicated function of 3 these values and of certain constants which are to be

X X X ...

The frequency of

joint occurrence of particu-

determined.

The other kind of case is that in which the quesone of graduation or fitting- '. Here ... l 2 3 are observed values of different quantities (e.g. rates of
(b)

tion

is

'

'

'

X X X

mortality at successive ages), or are the results of observation of one quantity at different times or by different
of any particular lies in the fact that the observed value contains an unknown error ;
variability

observers.

The

'

'

and our treatment

is

based on the assumption that a rela-

tion of a particular kind holds between the true Xs. In cases of this latter kind it should be noticed that

the only things which we treat as variables are the errors in the X's. may take, as the typical case, the observed

We

rates of mortality

XXX
1

3 ...

at

ages

^^ 2 # 3 ...

If

we

denote the true rates by fi^fs"*! ^ nen the assumption which we make is really an assumption that is a certain
function, with constants to be determined, of
this function is concerned, f
t.

So

far as

and

might be

called variables.

But, for our purpose, they are not variables. We are concerned with the fixed values t^ t^t z ... and the corresponding fixed, though unknown, values fx 2 3 .. ; the real variables
.

are the differences between the observed values

X X2 X
1

. .

and the true values

f2

. . .

(iii) There are two reasons why the same mathematical methods apply to subjects so different as relativity and statistical theory* One is that the number of X's in a statistical set may be very large in the second kind of case mentioned in (ii) it may be as many as 20 or 30. The need of a condensed notation is therefore even
:

94

Preliminary

IX.

(iii)

not exceed

greater than for relativity, where the number of dimensions does 4. The other reason is that, as in the relativity theory,

we are, to a certain extent in the first kind of case, and very largely in the second kind of case, concerned with sets constructed from the original sets by means of linear relations.
(iv) We denote the Xq and Xr by (Xg Xr
.

mean product
or,

of the deviations of

),

more

briefly,

by fqr (Xr

i.e.

fqr = (Xq'X = mean value of

r)

(Xq
It

mean

Xq

mean

Xr

).

must be clearly understood that, though (X .Xr ) q depends on q and r it relates to the complete fields from which X and Xr are drawn, and, for any parq ticular values of q and /, is not a variable, like X and Xr
t

but a constant.
(v)

Although we have defined (X^.Xr)

as a

mean

value,

our dealings with it ultimately depend on the algebraical laws which it follows. These are, first that it satisfies the

Xr

ordinary laws of multiplication of two expressions i.e. that, c being a constant as regards the A"s,
,

X^ and

(J

I = (Xr .Xg
r)

),

(Xq .(Xr + Xe )) = (Xq .Xr ) + (Xq .Xi ), (Xq .cXr ) = e (Xq .Xr ); (IX. 1.1)
any
linear function of the X's, then
i.e.

and next
(u.u)
is

that, if u is

positive unless u
(u
.

= 0,
>
if

that

u)

if

u^
0,

0.

(IX.

1.

2)

It is clear that (u.u)

u= 0;

for (IX. 1.1) gives,

by

putting

0,

(*g .0)
below).

(IX. 1.3)
1.

whence (?/.0)=0 follows by means of (IX. These are the only properties we
.

1)

(cf. (vi)
;

shall use

and

our results will therefore

(X

Xr) that

satisfies

be true for any meaning of these laws, provided, of course, that

IX.

1 (viii)
is

Preliminary
a constant as regards all the X's.
shall

95

(Xy.Xr)

And, con-

versely, only be dealing with sets for which. .Xr) has a meaning and a value for each value of q with (Xq each value of ?*, and satisfies these laws. It will be assumed

we

that the values of

our results are

values.

(Xq final

.Xr ) are known when expressed

;

or, at

any

rate, that

in terms of these

Whether we
products or not,
.

are dealing with

mean

mean
and

we can

call

(X .Xr) the
(ii)

squares or . of ( )

Xr

In the

first

kind of case mentioned in

from its mean, i. e. would be the mean square of deviation of q be the square of the standard deviation and (Xq r ) would be the mean product of deviations of q and r from their respective
;
.

(Xq

Xq

would usually

means, and would therefore, in the case of normal correlation, be and r equal to the product of the standard deviations of

Xq

multiplied by their coefficient of correlation. In the second kind is the mean square of error of of case (Xq and (Xq r ) is q) q
.

the

mean product

of errors of

Xq and X
.

r.

(vi)
.

It follows from (IX. 1. 1) that

)
.

(a^ Xr = a^ X
(vii)

r)

(a^
set

b
p

X =
p)

a^ X
.

p ).

(IX.
If

1.

4)

X^

is
is

a set of the kind considered in this section,

then so also
of

any other

which

is

a linear function

AA

Suppose, for instance, that

Then, by (IX.
(Y*
.

1.

4),

70

(^

vr

X = bx n vr (Xx Xv
v]
.

),

which has a definite meaning for each value of q with each value of r, and can be shown to satisfy the laws stated in (v).
(viii)

Our

results are also subject to the condition that

none of the determinants of the double sets we have to Xr) 0, whether the range deal with are i.e. that ( X
;
.
|

96
of values of q and r a part of it only.

Preliminary
is

IX.
1

(viii)

the whole of the range

to

or

IX.

2.

Mean-product

set.

(i)

The quantities (X q

r)

constitute a symmetrical double set

J uo -Jou.
fJLp

(Xl .X3 )...(Xl .Xm )

(IX2 A)

We

call this

the

mean-product

set.
is

(ii)

fW =fP*
(iii)

Corresponding to this there

reciprocal

set

given by
.

2. 1

If (X

X = 0,
s)

we can

for the

purpose of this
independent.

chapter

describe

X, and
is

as

statistically

a loose description, since the and statistical independence of two variables complete p would imply a good deal more than that the mean 8
Strictly speaking, this

product of their deviations from their respective means But we are only concerned, here, with mean should be 0.
squares and

mean

products.

from the point of view (iv) The simplest class of cases of algebraical treatment consists of those cases in which the X's are statistically independent of one another and the

mean square of deviation of each Xis this by duplicating the set, thus
:

1.

We

can express

X^ ^i
.

X X
2

Y A2

Xm 3 ... Y Y ^S'-'^m

and saying that the ( ) of corresponding elements of the two sets is 1, and that the (.) of elements which do not
correspond
is 0.

IX. 3

(ii)

Mean-product
kind we have

set

97

For a

set of this

/, r
so that the

= (x4

!,)

!?.
set.

mean-product

set is

the unit

It follows
its

that in this class of cases the mean- product set and

reciprocal set are identical.
(v)

The next kind of

case, in point of simplicity, is that in

which the X'a are

statistically

independent of one another but the

mean

squares of deviation are not all 1. This would be the case, 's were independent observations, not all of for instance, if the

the same weight, of a single quantity. For practical purposes a case of this kind can be brought under (iv) by expressing each in terms of its standard deviation (square root of mean square of

deviation) as the unit.

(vi)

There

fall into

is also an important class of cases in which the X's in one group is statistically two groups, such that each

in the other group. If, as in VI. 11 (i), we independent of each denote the two groups by A a and A^ t then the property is that
0.

IX.

3.

Conjugate

sets.

(i)

When
we
is

a set

XK

is

not of

2 (iv), the simple kind described in to introduce another set X* which (1)

shall find it useful

a linear function of

Jx and

(2) is
:

such that,

if

we

place the sets opposite one

another, thus

^1

-^2

^3"'^W

A2

X*...X m

the (.) of corresponding elements of the two sets is 1, and that of elements which do not correspond is 0. This new set X^ is said to be conjugate to X^.
(ii)

The second

of the above conditions can be written

(2?.Jg )=l2,
2511

(IX. 3.1)

98
or

Conjugate sets

IX. 3

(ii)

(iii)

Each element A^ of the new

set will contain

terms, with m coefficients which have to be determined from the m equations given by

There are altogether

m2

equations to determine

A\
x
.

By

regrouping these according to the values of /u in (A A^), we see that if x is conjugate to x then A"A is conjugate to A\ This is in fact evident from the symmetry of (IX. 3. 2).

(iv)
let

To
first

express

Xx

in terms of

us

take

W to be any linear function of AA W= a-X,.

.

AA

or

Xx

in terms of
,

Ax

say
(I)

Then we want to find an expression for a**. As we know the value of (AA A^), we take the

W and X\

of

By

(IX.

1. 4)

and (IX.

3. 2)

we

find that

whence

^=(
Substituting in
(1),

W=(W.W)X^
Taking

(IX. 3. 3)

W to be each element of A
A

in turn,
.

we have
(IX.
3. 4)

A = (A\A^)A
We
do not yet

/X
7/x

know

started with
larly

W as

the values of (A

A
.

a linear function of

A But, if we had X\ we should simi).

have got

W=(W.X )X,
ll

(IX.

3. 5)

whence
(ix. 3. e)

IX. 3

(vii)

Conjugate sets
form

99

Writing

this in the

^ =A^"'
we
have, by reciprocation,

(IX.

3. 7)

A*

=/**rx

(ix. 3. s)

which gives JP* in terms of X^. Further, comparing (IX. 3. 8) with (IX. 3. 4), we see that
If (Jx
.

jy

=f^

then (2* Z")

.

=/**

(IX. 3. 9)
is

and, of course, the converse also holds.

This result
1
(viii),

dependent on the assumption, made in is not 0.

|

that

fqr

(v)

We

could have obtained (IX.

3. 4)

and (IX.

3. 6)

in

fewer steps by considering the set as a whole instead of element by element. If we assume

then we get

(A* . X")
so that

X
(

"A; JT)
.

This gives (IX. the same way.

(vi)

3. 4)

and (IX.

3. 6)

can be obtained in

We

can write (IX.

3. 3)

in the form

and similarly from (IX.

3. 5)

(IX.
(vii)

3. 11)

If the set

Xx is

of the special kind considered in

(iv), i.e. is

such that

then

(*x.V=tf.

100
so that the set
is

Conjugate sets
identical

IX. 3

(vii)

with

its

conjugate.

The

set is

said to be self-conjugate.
(viii)

In a case of the kind mentioned in

is

(v),

where
.

fpq =
that

if

f pp =
(ix)

q^tp, but fpp

not necessarily

=1,

it

Vfpp, and/P<z

may be shown

if

p, so that

XP

Xp /fpp

Next consider a case of the kind mentioned in 2 (vi), consists of two portions, the elements in each portion being independent of those in the other portion. As before, we take one portion to consist of the first k elements, and the other of the remaining m k, and we denote the two portions by Xa and
where

X\

XQ.

Then the

special property is that

/
whence, as in VI. 11

a</)

EE(*a .J^)

= 0,

(1)

(iii), it

follows that

/* =
get, according as X belongs to the the two separate results

0.
3. 7)

(2)

Breaking up the right-hand side of (IX.

first

into

two

portions,

we

or to the second portion,

Xa =fay X7, X^-f^X*.

Similarly,

(3)

from

(IX. 3. 8), a

^/aTXy,
,

& ~f&Xp

(4)

In finding/ ? and/^ from/ A/I it is (see VI. 11 (iii)) immaterial as a whole or the sets f whether we take the set a y and separately, so that (4) may equally well be written (see VI. 11 (ii))

f^

f^
(5)

Xa = (/aT)^, X* = [f*+] m - k X.
IX.
(i)

4.

Conjugate sets with linear relations.

any
;

Let

Yp be

set

which
let

is

a linear function of
its

therefore of

If

and

Yp

X^

and

be

taking we have

W in

conjugate

set.

Then,
turn,

(IX. 3. 10) to be each element of

Yp in

Similarly

(Yf .X

fl ),

ryj = (P.Z"), F/^ = (J"..TJ.

/J

(IX. 4. 2)

IX. 4
(ii)

(v)

Conjugate

sets

wiih linear relations

101

By combining.

ratios,
ff

we get
.

such, results as
.

(E,

X") = 3/J, = Z /Yp Yf /^ = (S,

P) (Tf

X")

(IX. 4. 3)
(iii)

To

find

Jf f ,

suppose that

*,***,
Let the conjugate
set

d)

be

?f =
Then
If

y.

(7". 7,)

= (V*".4W *)
set is

whence, by reciprocation,

Thus the conjugate

(iv) Similarly, if

F = ^A^.
y*

(2)

then
etc.

in other words, the conjugate of an inner product

is

the

inner product of the

f the factors '

Let us write

Then
(C*.

Ts

since(X

.X

<T )

= IJ.

=h
Hence

=Y

f
.

We might,

alteraatively, have deduced this from

(iii)

by means of

(VI. 10. 2).]

(v)

From

(IX.

4. 1)

= YP/XK
and

(*P X")

(**

ty

= X"/F

therefore, by VII. 3

(iv),

7p F = A^.
Thus conjugate
sets are contragredient (VII. 4)

(IX. 4. 4)
;

and the

102

Conjugate

sets

with linear relations

its

IX. 4

(v)

all linearly related sets.

inner product of a set and If

sets

conjugate
T**

is

the same for

we denote

this inner product

by Q, then the
relations of the

X^ X*, T^

are connected

by

four

form

We

can express

in such forms as
(
.

x x q = x^x" = JA jg z J" = z y) jx A;,

(
.

the last of which,

when

written out in

full, is

Xm Xm
or in

more general forms such as

and

where

DK are any

linear functions of

XK

It

must

be remembered that the invariability of Q only applies for the particular values 19 2 , m If there were a dif-

X X

ferent set of A"s there would (in general) be a different Q.

IX.
cases

5.

The

frequency-quadratic.

(i)

In most of the
or of the second

we are considering, whether of the first

kind mentioned in 1 the frequency of joint occur(ii), rence of values lying within limits

X +
1

%d

X X +
lt

%d X2

Xm +

is

proportional to

xm are the deviations of where, if xlt x 2 . from their respective means, P is of the form
,

Xv X

Xm
.

P = a ll x

sc

+ 2a l2 xl x2 + a^x^ + 20 13 #1

ff

+ ... + a mm xm xm

(1)

In our notation this becomes

PSa^V
where
a**

(2)

= a *.
1

(3)

IX. 5
(ii)

(iv)

The frequency-quadratic

103

Let us write

Then

it

can be shown

(see

VIII. 4

(iii))

that

mean
.

value of

EPx a

= \\^ f

q =

*\

(5)

Taking (Bp Cq ) to mean, for these cases, the mean product x of B and C it follows from (5) that E is conjugate to qJ p xx, i.e.

Ex =

A
a?
.

(6)

Hence, by

(4),

^ = ^,
3.

and therefore, by (IX.

(IX. 3. 4),

10),

or
x
(x

by comparison with
.3f).
(7)

a Xf

x /x

=
/JL

are the

in the expression for given in (1) and mean products of the elements of squares in terms the conjugate set. Similarly, if we expressed
(iii)

Thus the

a's

mean

of the conjugate set, the coefficients would be the mean squares and mean products of the elements of the original
set;
i.e.

r^m^m x

\\t
j)

ivhere

of two quantities (iv) Take, for example, the case standard deviations and coefficient of correlation are Then it is well known that

19

X^ whose
,

elt c2

and

r.

104

TJie frequency-quadratic
set,

IX. 5

(iv)

This gives, for the members of the conjugate

c2(l-

It is easily verified that the

mean

l square of x

is

equal to the

coefficient of

x^

in P,

and

so on.

(v) If we express the a?'s linearly in terms of a new set can put this yx the value of will remain unaltered. as follows. that yK is any linear funcdifferently Suppose
,

We

tion of of h Ki .

#v

Let /'^EE (yK

if

y^

and

x/ let ^

be the reciprocal

Then,

we write

we

shall

have

(vi)

The

(Xq
is

2)

or (x

# r ) which we have
of

so far been

considering
of

the

mean square
;

x^

or the

mean product

and x r) without regard to the values that each of the Xy i. e. the mean other #'s may have square or mean product
taken for
all possible

values of these other

their relative frequencies.

We may

also

xs according to want to know

what happens when some of the #'s have definite values ascribed to them and are not allowed to vary from these In these cases we follow the principle of VI. 1 1 (ii). values. Suppose that all the #'s after x are fixed. Let the #'s up Then our methods apply to Xfr be denoted by xa or x^ etc. The principle, to the set of order k formed by these #'s. we want to study the therefore, is as follows. Suppose k quantities variation of the k quantities xa when the m
fc

XQ

are fixed.
f r

We first
i

nP) >/v reciprocal set /

(= a

^e m quantities x

construct the mean-product sety^ then construct the ; a and

x^

(=a^), the elements of which are the in the terms in P then take out the partial set coefficients ay corresponding to xa ; and then find the set (f )k which ay ay The result is the mean-product is the reciprocal off
;

IX. 6
set of

(i)

The frequency-quadratic
when x^
is fixed.

105

so

but

is

(This is a well-known theorem, in terms of determinants.) usually expressed

a

If the partial sets xa and x^ in (vi) are independent, set so that (xa x^ 0, the elements of the mean-product
(vii)
.

outside the portions corresponding- to (xa ay ) and (x^ . o^,) 3 (ix)) the values of (/ ) ft will be will all be 0, and (see
.

a7

the same whether

or

we

construct

them from the whole

;

set

from the partial set. This is otherwise obvious for, if the %'s of xa vary independently of those of o^, they vary
in the same

way whether

the latter are fixed or vary.

IX.

6.

Criteria for

improved values.

(i)

Our next

problem, considered in this and the following section, is that of reduction of error, in a case of the second kind mentioned
in
1 (ii).

We have a set of quantities A = (#x A A-.-O.)

errors
;

which contain

the

mean

products of error being

(IX. 6.

<^=(-Z>x-Zy.
The whole
set

A)

D^

consists of

two portions

All the D's are the results (direct or indirect) of observaare negligible (within tion ; but the true values of the the degree of accuracy to which we are working), and, if U is any one of the J^'s, or any linear function of them, we without altering, can add to it any linear function* of

D^

D^

except to a negligible extent, the true value which it repreIf the sum of U and an indeterminate linear funcsents.
tion of

DQ

is

represented by U' the problem of reduction of

y

error is to determine this linear function so as to

*

make

We could replace is, of course, an incomplete statement. any function of U and D^ -which would be equal to U if the DQ were all 0. But we are only considering linear functions. 2611 O
This

U by

106
(

Criteria for improved values

IX. 6

(i)

U' . U') a minimum.

The
t

resulting value of
1

is

called

the improved value of U and will be denoted by 1 U. The can replace are called the auxiliaries. elements of

D^
.,

We

by
(ii)

/3,

y,

and

<

by

x>

Vo

>

as occasion requires.
arises is as follows.

The way

in

which

this

problem

We

start

. . with a set of observed quantities 2 lt m which correspond to a series of values of some other quantity t at equal or unequal intervals the Xa might, e.g., be rates of mortality at
,

X X

different ages.

The

Xs

contain errors; and our fundamental

assumption, based on general experience and on inspection of the particular data, is that the true values are so nearly of the form
k l that their (in ordinary notation) c + c l t + cz t + . . + c^._j t differences (divided differences if the values of t are at unequal l)th are negligible. may therefore add intervals) after the (k
.

We

to each

X any linear function of these differences, which are what

Z)fc +1
,

we

are calling

Dk+v -An-

The problem

is

to

determine
square of

the coefficients in this linear function so that the

error of the

mean

sum

of the

and the

linear function shall be

a minimum.

(iii)

We

have

first

the two portions of

what relations hold between and the two portions of its conD^
to see

jugate set

when similarly divided. Denoting the conas usual, by jD\ let the two portions be jugate set,

Here
the
since

it is

to

set

(order k]

each

is not (in general) conjugate to the set a (order k) t element of it is a linear function of the

be

observed that

Da

whole .Z\; and similarly

conjugacy
is

for Ifi.

Now

the condition of

But, if LP and ZL belong to non-corresponding portions of the two sets, q cannot be equal to /?. Hence we get the relations

IX. 6

(vi)

Criteria for improved values

107

(iv) First let

be an element of

tion of

D^

or a linear func-

say

#^Z.

Then

its

improved value must be

(IX. 6.1)

For this makes (U'

(IX.
(

V'} = (0

= 0. = 0, 0)
if
?7'

1. 2),

(W

U'} would be

>

by (IX. 1.3); and, by 0. Hence U' were not

f
.

U') is a

minimum when

0.

(v)

The next most simple

case is that in

which

U is an
(1)

element of

Da or a linear function of .#",

V=aJT.

say

Let the value of U' be

U'= U+u,
where
u

= a+V r

(2)

Then, by (IX.

1. 1),

But, by

(1)

and

(2),

and by

(2) of

(iii)

above,

Hence (/' J7') is a minimum when (u. u) is a minimum and this, by (IX. 1 2), is when u = 0, so that
.
.

U'

r
.

Hence the improved value

i.e.

is

the same as the original value

l(
(vi)

V) = aa I).
U

(IX. 6. 2)

simplicity of the results obtained in (iv) and (v) that we should in all cases regard as expressed suggests a in terms of and D^ . There is no difficulty about this,

The

108

Criteria for improved values

IX. 6

(vi)

in theory, whatever linear function of the D's may be. for instance, is given as a linear function of If, D^, then we obtain our result by eliminating a (k values) between

this formula for

terms of D^, that

i.

e.

and the k equations which give D in in terms of Suppose then a and D^.
a

U= V+W,
where

Then

is

formed from
so that

U by

adding some linear func-

tion of

Dp

U'

= 7+ W,
and

where

F = ba D

as before,

is

of the form

Hence

(V .V) =
But
(V.

(7. 7) + 2(7.

W')+(W. W).
0,

W) =

(6a lT.e+l)j

= 6a <P(lT.Dj =

by

(2) of

(iii).

Hence

But

this is a

minimum when I(bu lT +

if

W=
U

0.

Hence
(IX. 6. 3)

&Dj = *-<-.
in terms of

In other words,
involving D^.
(vii)

we

express

Da

the improved value of

and Z^,

is

found by omitting the part

a
,

Since,
(2) of

by (IX.
(iii), (

6. 3), a
.

IU

is

a linear function of -#

nnd,

by

DJ = 0,

it follows,

by (IX.

1. 4),

that

In other words, the ( of the auxiliaries is 0.

= 0. (IX. 6. 4) (7tf.Zy of any improved value and each )

.

IX. 6

(xii)

Criteria for improved values

109

(viii)

differ

It also follows from (IX. 6. 3) that if two quantities by a linear function of the auxiliaries they have the
value. in (vi) to be each

same improved
(ix)

By taking U

member, in turn,

of a set

4ta
we
find that

IB
Also

(IX. 6. 5)
i.

e.

the improved value of any linear function of the

_Z?'s

is

the same linear function of their improved values.

(x)

Altering

in (IX.

6. 5)

to >&,

and writing C^

=
e)

k^

BK we

find that

x
i. e.

-Jj

(xx.,

related in the
briefly,
(xi)

the improved values of two linearly connected sets are same way as the original sets; or, more
a set and
its

improved values are cogredient.

Since we know that the improved values of D^ we have really only to determine those of k other In view of (IX. 6. 5), we can choose these to quantities. be any linear functions of Da that we like, with or added and similarly we without linear functions of
are 0,

D^

can replace D^ by any linear functions of D^ provided, in both cases, that none of the mean-product determinants The functions so chosen can be called are 0. a and
:

D^
ing

so that

we

need only consider the problem of find-

/..
The
result stated in (IX. 6. 2) gives us the extension, to the

(xii)

general case in which the errors of the original observations

may

110

Criteria for improved values

IX. 6

(xii)

moments'
(

have any mean squares and mean products, of the 'method of ordinarily applied to the case of a self-conjugate set
3
(vii)).

We

take X^, as in

(ii),

to be the original observations,

and

Z)^ to

be their differences of successive orders.

Then we have

found that the improved value of any linear function of

Da
.

is

the

1 . . IP as the original value. are , D\ But, by VII. 3 (v), successive sums of the elements of the set conjugate to X^ ; and the first k moments of x are linear functions of these sums.

same

X\

original values

Hence the improved values of these moments are the same as their and this, by (ix), is the same thing as saying that
;

x the moments of the improved values of are equal to the moments of the original values. Thus the method of moments

But it should be observed that it does not apply to still applies. the original set of observations, but to the conjugate set. As a simple example, suppose that X^ is a set of independent
being cp cp
observations of a single quantity, the mean square of error of Then ( 3 (viii)) the conjugate set is
. ,

Xp

"V
2

V
fnw
\

will all have the same improved value, which we will IX, there is only one moment to be considered, namely, the Oth moment, or sum, of the conjugate set. Hence, equating the sums of original and of improved values,

As the X's
call

which gives the familiar

result.

IX.
(i)

7.

Determination

of
1

improved values.

From the results obtained in the preceding section we deduce three methods of finding the improved value of any element of a say

(1)

We can

Zy. express

D*

in

terms of If and

D^

and then

omit the part involving D^. (2) We can say that

The
is

result is

IDj.

IDj

some

linear function of J9.

;

This linear function has k coefficients to be determined

they are determined by the condition that,

if

the linear

IX. 7

(ii)

Determination of improved values

is

111

function

expressed in terms of D^, the coefficient of

Df

is

practical

and those of other elements of Da are 0. The application of this method depends on the cir-

Dj by adding which we have called b^ 1)^. This linear function has m k coefficients to be determined. We have found in (IX. 6. 4) that (II)f D^ = this gives m k equations, from which the coefficients in question can be determined. Thus the necessary and sufficient conditions
IDf
a linear function of

cumstances of the particular class of cases. is obtained from (3) We can say that

D^,

for IDjf are that it differs from D* by a linear function of D^ and that (IDf Z> ) = 0. These three methods are exhibited in (ii), (iv), and (v) below, and a fourth method is given in (vii).
.

(ii)

To apply the
a<

first

method,

let

us write

We
e^

do not need
at once,

^ since the only part of Da that counts

is e

for the

improved value
(2)

a^

IT

we

therefore get rid of

.

This,
sides,

by

by means of something whose ( D^) is 0. 6 (iii), is D7 of Taking the (D^ ) of both

.

we have

Also,

by

(1) of

(iii),

Hence
Here a, ft, y relate to the partial set (A l A 2 and the statement is limited to this set. Dealing only with this set, let us denote the reciprocal of d^ by
.
.

(d/3 7 )&

this, as pointed
is

out in VI.

1 1

(ii) (cf.

5 (vi) of the as

present chapter)

not ordinarily the same thing as d

112

Determination of improved values

IX. 7

(ii)

obtained from the whole set of order m.

We

have then,

by

reciprocation,

a and dropping the Substituting in the expression for in order to get the improved value, we have
,

(IX. 7.1)

(iii)

Although, in the above, we have not needed e^, we

ought to find its value in order to satisfy ourselves that, 6 (vi), any linear function of ~DX as has been stated in
say

Da

g a Da + g^Dty
'

and

D^

can be expressed as a linear function of to do this, it is only necessary to prove the

-Z?

proposition for follow at once.

a , since

the formula for ga

Da +ff^^

will

We have

written

a To find e ^ t we must get rid of the and have found ea/3 so we again use (2) of 6 (iii), getting first term
.

4*X

Hence, by reciprocation,

where

x [^ ] m -^

f-Wm+tft,
is

partial set (-#& +1 for n is therefore

^+2

obtained from the the reciprocal of d^x AJ- ^ ne complete expression

existence of (d0a) k and [^x ] w -^ d assumption that the determinant

The

is
\

and the determinant

|

d^

formed
\

qr for

dependent on the formed for a

are both

D^

D ^

(see

(viii)).

IX. 7(vi)
(iv)

Determination of improved values

use the second method,

113

To

we might have proceeded

as follows.

"We write

To

find ea p

we

express

by means of (IX.

3. 4),

if in terms of 1)^, and we have

i.e.

of

Dy and D

From

the condition stated in (2) of

(i), it

follows that

<v^ = IL
and
therefore,

by

reciprocation,

'* M i!WK
Hence we get the same

('*)*'

result as before, namely,

z.-(jy.D.
(v)

For the third method, we write

and we have to
namely,

find e a(f>

The condition

stated in (3) of

(i),

gves
rf.

= (D

.2)

)^(^.2>x
k values of

This

is

true for all the

x.

By

reciprocation

f*=l**] m ->dax

This agrees with (IX. 7. 1) and (IX. 7. 2). As w /& will usually be a good deal greater than k, the method is rather of theoretical than of practical interest.
(vi)

The elements which we have found

to be important

in the above processes are the m k auxiliaries whose D^, a values are all 0, and the k elements of the improved

114

Determination of improved values

IX. 7

(vi)

conjugate set which correspond to the remainder of

Dx

These elements together constitute a we have in fact, in (iii), expressed

set.

set of order

and

Da

As the
is its

set is

important,

it is

in terms of this worth while to see

what

We

conjugate. write

where the
orders

& k and m
*

'

means that the elements of the two sets of k are combined to form a set of order m.
statistically independent. x

These two partial sets are

follows,

It

by

3 (ix), that the set conjugate to

is

(vii)

But

(daB

and d^a being

identical)

we have

already

found that

Hence we get a
set conjugate to a conjugate to
(viii)

D &D
IDa
is

Let the concise formula for finding 2Da a be 7> and let the set Ifl\ a D^
.

D &

&

be

Fa &

F*.

Then

Fa = IDa

Since

of the form
.

the form

eW,

and (IF

Da -e^D^ and ID = 0, it follows that D)

is

of

and similarly

[NoTE. This chapter is based on (1) a paper by myself in Phil. Trans. (1920), ser. A, vol. 221, pp. 199-237, in which the old notation was used (2) a paper by Professor Eddington in Proceedings
;

of the London Mathematical

Society, ser. 2, vol. 20, pp. 213-221, showing how the notation and methods of the tensor calculus can be applied, and making some abbreviations and improvements in the work and (3) a note by myself, following the above, ibid., pp. 222;

224.

have altered the notation a good

deal.]

X.
X.

TENSORS IN THEORY OF RELATIVITY

1.

Preliminary.

(i)

Tensors are sets* which

(1) are

functions of a set of co-ordinates (xl x2 #3 ...) and (2) are subject to certain conditions of transformation when the
co-ordinates are transformed.

In the theory of relativity there are four co-ordinates so that all the sets are of order 4, and any inner (#! multiplication with regard to a suffix p involves addition
(ii)

#2 #3 #4 ),

of the products for the values 1, 2, 3, 4 of /ut. But this fixing of the number of elements in a set does not affect the

general reasoning with regard to the continue to treat them as of order m.

(iii)

sets,

and we can
and a

In VII. 4 we started with a

is

set

X\

set

Ax

which

a definite function of

X\

and we supposed

A** to

be changed as the result of and the cases substitution

;

being changed by linear we considered were those in

K

X^

such changes, A varies either directly In VII. 7 we extended the inquiry or reciprocally as X\ a set Si to be a function of two or more single by taking

which, throughout

all

sets,

and considered cases in which, when these

sets are

changed by

procally as consider cases in

linear substitution, Si varies directly or recieach of the sets. For tensors we have to

which the primary substitutions are not

If in place of the original set of coset sc\, which is a function
,

necessarily linear.

ordinates XK we take a new

*

but not necessarily a linear function of #A the ratio which

must be remembered that the elements of a set are not necessarily numbers, but may be quantities and that a set as a whole is something different from its elements. What we usually mean by a tensor is some physical phenomenon represented by a set but no confusion arises if we call the set itself a tensor.
It
;
:

116

Preliminary

X.

(iii)

we have now to consider is not the ratio of x' to XK but p the ratio of their differentials, i. e. the partial differential coefficient cto' /d#A When the substitution is linear, this p
.

* to #' so that our treatment of the general equal p /tfA case is consistent with our previous treatment of the paris
,

ticular case.

We will begin with the

of higher rank.

single set,

and then go on

to sets

(iv) In the case of sets of higher rank, we sometimes have to deal not only with inner products but also with inner sums. By the inner sum, in such an expression as

A^ va we mean
,

the result obtained by replacing

for

/>

by

<r

and summing the values

0-=

1, 2, 3, 4.

It will be seen

presently ( 5 (iv)) that, as the result of the particular notation adopted, the inner products or sums have only to

be considered

when one

of the
is

two

letters
suffix.

concerned

is

an

upper

suffix

and the other

a lower

X.
sets,

2.

Single sets
start
,

(vectors).-

(i)

Beginning with single

with a pair which we call a?x (the set of and A x or xa and A a A K or A a being a function of xx or #a (The a here, like the A, denotes a complete Connected with set, not, as in Chapter IX, a partial set.) K these, or arising out of them, is a plurality of pairs x\ and A' But our purview is limited to the cases in which the relation x between A* and A' is linear, and in which, further, this linear relation is of one oftae two following kinds (a) where

we

co-ordinates)

mentioned in the note to VIII. 1 (ii) doew not does not occur absolutely, but (directly or reciprocally) as one of the factors in inner multiplication with
difficulty
<>x'

The

arise,

because

p /<)x^

regard to X or

p.

X.

3(1)
(b)

Single sets (vectors)

replacing

117

Ax

and

A fX

by

AK

and

A\

where

In the cases under () ^/ x vector in the cases under

;

is

said to be a contravariant

(b)

Ax

is

said to be a covariant

vector.

Here 'vector'
a tensor.

is

used as meaning a single set

which
(ii)

is

It will be seen
a?'

from VIII.

(ii)

that, if the relation

between #x and

is linear,

these become respectively

so that in these particular classes of cases is contravariant X and #x are cogredient, and J A is covariant if J A and a?x are contragredient.*
if ,4

X.

3.

Other

sets.-(i) For sets of higher rank, we are

d#' -r-^

similarly concerned with pairs of partial derivatives

and r-5
,

do?

^etc.;

and
r^

and a set depending on o?a ,^, ... is not a tensor unless, when xa X... become tf'^X^..., the 'ratio' of the new
value of the set to the old value

is the product of these The particular one from each pair. partial derivatives, derivatives are indicated by the position of the letters

these are upper suffixes if, so far as the particular variable is concerned, the relation is of the
a/3...
:

or juy...

It seems desirable to call attention to these classes of cases, as otherwise the tensor terminology may be found rather confusing.

118

Other

sets

X. 3

(i)

contravariant type, and lower suffixes if the relation is of the covariant type. Thus for double sets (tensors of the second rank) we should have such relations as

A'

1>x\

J^

(contravariant tensor),

Mf

^
a.0
\

^ ^T~
A
&
'T*

(covariant tensor),

/*

y/

-j
and
tion as

vtZ/

^-f ^

(mixed tensor)

for a tensor of

higher rank we might have such a

rela-

Two
if

tensors SL

the ratios
(ii)

and 2B are said to be of the same character #'/# and }5'/35 are of the same form.
set

For a scalar function of a

or sets the above

condition becomes

A = A,

so that a scalar (in the general sense) is not a tensor unless it remains constant for all changes in the system of co-

Such a function as A 1 + A 2 + A 3 + Ai, for instance, would not usually be a tensor. For tensor purposes, therefore, scalar' practically means 'invariant'.
ordinates.
*

X.

4.

Keason
'

for limitation.
tensor
'

The
is

object of limiting
to ensure that the

the definition of
result of

in this

way

any number of

steps, all of the

same kind, pro-

duced by successive transformations of co-ordinates, shall be the same as if we had passed in one step, also of the

same kind, from the set. That this is in

initial set

of co-ordinates to the final

fact ensured is seen

from the properties

X. 5

(ii)

Reason for limitation

set.

119

of the partial derivative of a that

Suppose, for example,

and that

Then, by (VII.
a

2. 2)

and (VIII.
a '

3. 1),

^_-^^ .^^" " * A ^ to *x

~

A'*

'

which
X.

is

of the same form as (l) and (2).

5.

Miscellaneous properties.

The following

are

some miscellaneous properties which are useful in determining whether a set is a tensor.
(i)

The sum

(or difference) of

two tensors of the same

suffix is

rank and character and the same

[Suppose, for instance, that tensors. Then

a tensor.

Ax

and 1^ are contravariant

and therefore

is

of two tensors is a tensor whose character (ii) The product the combination of the characters of the two. For example,

from

we

see that

120
(iii)
(

Miscellaneous properties

X. 5

(Hi)

An

inner product of two tensors, or an inner

is

sum

(iv)) of a tensor, taken with regard to suffixes of

opposite character,

a tensor.

[Take, for example,

,P

^ - *^
by
v

<)Xn

fog

l>Xy j>X'p

*/ a^;

a,

*
3. 3),

If

we

replace p

we have, by (VIII.

and therefore
,,*

~x
_

*X

which
(iv)

satisfies

the requirements.]

If in this last example

instead of p

by

v,

we had replaced p by the expression would have contained

A.,

which has no general

(v)

significance.

The

derivative of a scalar
is
(

(ii))

is

a covariant
,

vector.

Suppose, e.g., that A whose value remains constant

tions.

a scalar function of #x 3 (ii)) for all transforma-

Then

so that

^Afix^

is

a covariant vector.

If the inner product of a set Si by each of m ( 4) ~ (contra variant) is a tensor, and f vectors is a tensor, then
(vi)
.

covanant

&

X. 5
.

(vi)

Miscellaneous properties
.
,

121
is

covariant

is

} >

(contravanant)

as regards

%, where u
**'

the linked

suffix.

[Take the case in which the vectors are contravariant, . and suppose that SI = A^v Let the m vectors be
. . .

.#w/t or, the A denoted by J5^, collectively, by not indicating any tensor character. Then, by hypothesis, jB*'* is a tensor as regards vttt is a tensor as x^ and
. . . ,

^,

&\

^A^

regards xv

. .

.,

so that

From

these

we deduce

Xf whence, by division by B'

(see

VI. 9

(v)),

Hence

v ...

is

a tensor and

is

covariant as regards x^.

dealt

The

which the vectors are covariant can be with in the same way.]
case in

2611

APPENDIX
PEODUCT OF DETERMINANTS
In IV. 5 we have taken as the standard form
of two determinants
to 2, for
for product reduced from 3

the order being

now

economy of printing a^ X

In this form, the element in the gth column and rth row c inner product of the $th column of the first determinant and the rth row of the second. By of columns and rows, and also by changing interchanges
of the result is the
'

the order of multiplication, we get seven other forms, all constructed according to this rule. The eight forms can be set out as follows :

0)
(2)

a 2 /3 2

(3)

(*)

(5)

(6)

(8)

Appendix

123

first four,

It will be seen that the last four are the transposed of the but in the reverse order ; i. e. the transposed of

(1) (2) (3) (4) are (8) (7) (6) (6).

In the double-suffix notation these become, the order

of (5)-(8) being reversed
qr
:

x X

rq

= \dqK e r\ -rq = e qr\

qr

fj

q\

a e\r
I

(2)

d
I

rq

=
I

exa

(8)

(3)
qr

(6)

rq

VI s*
;

(5)

must be borne in mind that in each of these statements refers to the column and r to the row e. g. (6) means q
It

that
'

INDEX OP SYMBOLS
dqr determinant 39 dps (in Chapter V) ratio of cofactor of d ps to A^ single set 44, 58
|
| |

d qr 42
\

AW double set 45,

?VL
|

59
of

A transposed of A^p 45, 59 BV CV product-sum (inner product)

unit set 50, 64
<E set

B^ and Cp

47, 62

generally 60

SB

product of <& and B 61 A^Byp inner product of A^p and -** inverse of 65

63

A^ ^irreciprocal of A^ 67
P ratio of
ratio of
.B to p

inverse of partial set A ay 70 70 _* inverse of partial set

^76
82

A^

to

derivative of

B U

^x set conjugate to X^ 97
727 improved value of 106 A% vff inner sum derived from
-4^ contravariant vector,

(Xq

Xr

mean product

with regard to of deviations of

Xq and Xr 94

A^

85,

86

A^ vff

116

co variant vector 117

GENERAL INDEX
Under any heading, separate entries are usually in the order of occurrence, not in alphabetical order.

Addition of sets 61 Adjoint determinant 37 of element of 38

Auxiliaries 106

cofactor

Character of tensor 118 Cofactor 24 expressed as determinant 25 determinant conCofactors, structed from 36-38 Cogredience 79, 84 ; of statistical set and improved values 109 ; relation to contra variance 117 Cogredient sets 79, 84 Column of double set 15, 45 ; of determinant 21 Complete inner product 62 Conjugate set 97 ; determination of 98-99 original set as conjugate of 98 ; of inner product 101 Conjugate sets with linear relations 100 Continued inner product 68 transposed of 69 ; inverse of 69 reciprocal of 69 Contragredience 79, 84 ; with linear relation 80 ; of variables
; ; ;
;

Derivative of set 85 ; of sum or product of sets 86 of function of a set 87 Determinant 21 notation 21, 23, 39 ; elements 21 ; column 21 ; row 21 ; leading diagonal 21 ; order of 21 ; leading term 21 calculation of 32-33 (See also
;

separate headings below)

Determinant, construction of: factors of terms 14-16 rule of sign of term signs 16-19 dependent on number of reversals of order 17-18 sign of term changed by interchange
;

of suffixes 18 ; effect of altering order of letters 18-19; final definition 20-21

Determinant, main properties of, in ordinary notation 40; in tensor notation 52 if Determinant, properties of: each element of column or row 22 not altered by interis change of columns and rows 22 sign changed by interchanging if two columns [rows] 23
; ;
;

and

partial differential operators 88 ; of conjugate sets 101-102 ; relation to covariance

their

two columns [rows] identical

23 expression in terms of elements of column [row] and their cofactors 24, 26, 28 sum of products of elements of column [row] by cofactors of 27, 28 parallel elements is
;
;

117 Contragredient sets 79, 84 Contravariance, relation of, to cogredience 117 Contra variant vector 117 ; tensor 118 Covariance, relation of, to contragredience 117 Co variant vector 117 tensor 118
;

multiplication by single factor 31 ; not altered by increasing

elements of column [row] by multiples of parallel elements 31

126

General Index
Leading term of double of determinant 21
set 16;

Determinants, sum of 81 ; product of, in ordinary notation 38-36 ; product of, in tensor notation 48, 49 ; forms of product of 122 Direct proportion 76, 84 Dummy suffix 48, 62

Linear function of single set 74 Linear relation between single sets 73 ; between sets generally 82 Linked suffix 48, 62

Elements of double set of determinant 21

set 44,

15, 45,

59

of single

58

Error, reduction of 105

Fitting 93

Matrix of determinant 59 n. Mean product of deviations, notation for 94 algebraical laws satisfied by 94 Mean-product set 96; of selfconjugate set, is unit set 97
;

Free suffix 48, 62 Frequency-quadratic 102 ; meaning of coefficients in 103 Functional relation between sets
73

(See also separate headings below}

Mean-product set, reciprocal of 96 is mean -product set of

;

conjugate set 99

Mean-product set for partial variation 104

Gothic letters 60 Graduation 93

Minor 25; relation

factor 25-26
seqq.
;

of,

to

co-

Greek letters for sets 43 product-sums 47

for

Mixed tensor 118 Moments, extension of method

109-110

of

Improved values 106

criteria for

105-110 ; determination of 110114 Independence, algebraical 73 n. statistical 96

Multiplication of set by scalar 61 by set 61

Notation

see

INDEX OF SYMBOLS
;

Order of determinant 21
set
;

Independent

observations,

set

constituted by 97, 110 Inner multiplication 49 Inner product 49, 62, 63

posed of 63 tion of 64

trans; rule for construc-

58 of double set 59 generally 60 Outer product 49, 61

Partial
differential

of single of set ;

coefficient,

Inner product, complete 62 Inner product, continued Continued inner product Inner sum 116
;

see

Inverse of double set 65 ; condition for 65-66 of transposed set 67; original set as inverse of inverse set 67

reason for appearance of 115-116 Partial double set 70 ; inverse or reciprocal of 70 Partial single set 70 Product, inner see Inner product Product, outer 49, 61
:

Product of determinant by single factor 31 of two determinants, of two sets see Determinants
;

45-46, 61

Jacobian 88

Key

set

20
set

Product-sum notation 47 Proportion, direct 76, 84 procal 77, 84

reci-

Leading diagonal of double 15 of determinant 21

;

Quadratic form, expressed as inner product of two single sets 81 ;

General Index
derivative of 86; expressed as sum of squares 88-89

127
14-19
;

denominator
solution 30

general

Rank

of set 60 Ratio, as operator 76 ; of single sets 76 ; of sets generally 81, 82 reciprocal of 81, 82 Ratios, equal 75, 76, 82 83; inversion of 76, 78 ; cross-multiplication of 78
; ?

Statistical independence 96 Statistical set, nature of 92-93 Substitution-operator, unit set as

64-65
Successive summations, order immaterial 48 Suffixes, upper 47 Sum of determinants 31 Symbols : see INDEX OF SYMBOLS

Reciprocal determinant 41-42 Reciprocal of double set 67 original set as reciprocal of reciprocal set 67 Reciprocal proportion 77, 84 Reciprocation 67 Reduction of error 105 Relativity and statistical theory, similarity of methods for 93 Reversals of order 17
;

Symmetrical double

set 45,

60;
reci-

identity of inverse procal sets 67

;

and

Row

of double set determinant 21

15,

45

of

Tensor 115 conditions for 116118; contravariant 118; covariant 118 mixed 118 reason for limitation 118 Tensors, properties of 119-121 sum or difference 119 product 119; inner product or inner
; ;
;
;

sum
Scalar 46, 60; as invariant 118; derivative of 120 Self-conjugate set 100
Set,

120
; ;

Transposed of determinant 22 of double set 59 of inner product

63 ; of inverse set 67 procal set 67
;

of reci-

double 15, 45, 59; determinant of 59 ; transposed of 59 ; inverse of 65 reciprocal of 67 as sum of comSet, single 44, 58 ponent single sets 71 Set-notation 42 ; principles of 44
; ;

Unit
51,

set 51,

64
;

properties of 49;

64-65

symmetry of 51, 64

action as substitution-operator 64-65 ; inner product of two

unit sets 65
Variable sets 72 Variation, direct 79. 84; reciprocal 79, 84 Vector, ordinary, relation to single set 46 n., 59 Vector, tensor 117

Sets, functional relation

between

73

Simultaneous equations

method

of individual solution 11 ; formulae for particular cases 12-13, 29 ; statement of general problem 13-14 ; trial formula for

PRINTED IN ENGLAND

AT THE OXFORD UNIVERSITY PRESS