?

Csl

00

:*=T
I

co

CO

QA
281

W48

A SHORT COURSE
IN

INTERPOLATION

(D 309)

A SHORT COURSE
IN

INTERPOLATION
BY

E^T.

WHITTAKER
Sc.D.

F.R.S.

PROFESSOR OF MATHEMATICS IN THE UNIVERSITY OF EDINBURGH

AND

GEORGE ROBINSON
M.A.,
LECTURER

IN

MATHEMATICS

IN

B.Sc.

THE UNIVERSITY OF EDINBURGH

BLACKIE AND SON LIMITED

50

OLD BAILEY, LONDON; GLASGOW, BOMBAY

1923

Printed and bound in Great Britain

PREFACE
A KNOWLEDGE
who make

of the

Theory of Interpolation

is

required by all

inferences from the results of observation, especially

by astronomers, physicists, statisticians, and actuaries. Until

recently it was somewhat neglected in the mathematical schools
of

many

British Universities

but of late years there has been

wider acceptance of the view that the subject is easy enough to

be put at the beginning of a student's course, that it forms an
excellent preparation for the Differential Calculus,
it

cannot be

left out.

The present
first-year

and that

text

is

offered as a short exposition suitable for

undergraduates

chapters of a larger

it is

a separate issue of the

general field of the Calculus of Observations.

23rd July 1923.

first

four

work by the same authors, dealing with the

CONTENTS
GHAPTEE

INTERPOLATION WITH EQUAL INTERVALS OF THE ARGUMENT

SECT.

1.

Introduction.

........
........
.......
.....
.......

2.

Difference tables

3.

Symbolic operators

4.

The

differences of a polynomial

5.

The

differences of zero

6.

The

differences of

7.

Representation of a polynomial

8-9.

10.

x(x-

1)

(x-

2)

by

(x-p+l)

factorials

of interpolation

The Binomial Theorem

Examples on Chapter

I.

CHAPTEK

...

The Gregory-Newton formula

.

.10
.15
.16

II

INTERPOLATION WITH UNEQUAL INTERVALS OF THE ARGUMENT

11. Divided differences

12.

Theorems on divided

differences

13-15. Newton's formula for unequal intervals of the argument

16. Divided differences with repeated arguments

17-18. Lagrange's formula of interpolation

19.

Remainder-term in Lagrange's formula

Examples on Chapter

II.
vii

.20
.22
.

24

.27
.28
.32

INTERPOLATION

viii

CHAPTEE

III

CENTRAL-DIFFERENCE FORMULAE
SECT.

20. Central-difference notations

21-22.

The Newton-Gauss formulae

of interpolation

23.

The Newton-Stirling formula

24.

The Newton-Bessel formula

25. Everett's formula

26.

Example

27.

Summary

28.

The

of interpolation

of interpolation

of central-difference formulae
of central-difference formulae

lozenge diagram

29. Relative accuracy of central-difference formulae

30. Preliminary transformations

Examples on Chapter III.

CHAPTEE

.35
.36
.38
.39
.40
.41
.42
.43
.47
.50
.51

IV

APPLICATIONS OF DIFFERENCE FORMULAE

31-32. Subtabulation

53

34. Inverse interpolation

35.

The

derivatives of a function

36. Derivatives in terms of central differences

37. Derivatives in terms of divided differences

Examples on Chapter IV.

INDEX

.57

.......
......

33. King's formula for quinquennial sums

60
62

.64
.65
.68
71

CHAPTER

INTERPOLATION WITH EQUAL INTERVALS OF THE ARGUMENT

1.

Mathematics

Introduction.

idea of correspondence
2
thus
sponds a value of x
:

e.g.

is

occupied largely with the

number x there corre-

to every

oj=l,

2, 3, 4, 5,

...
...

x* = l, 4, 9, 16, 25,

One
holds

of the

is

of that

two variables between which correspondence

called the argument and the other

is

called the function

argument.

y of an argument x is denned by an equation

where
f(x) is an algebraical expression involving only
y=f(x),
arithmetical operations such as squaring, dividing, etc., then by
If a function

the value
performing these operations we can find accurately
But if 2/ = log 10
of y, which corresponds to any value of x.

not possible to calculate y by performing simple

arithmetical operations on x (at any rate it is not possible to
calculate y accurately by performing a finite number of such
recourse to a table,
operations), and we are compelled to have
(say), it

is

which gives the values of y corresponding

values of x

of

to certain selected

e.g.

x.

log x.

X.

log x.

7-0

0-845 098

7-4

0-869 232

7-1

0-851 258

7-5

0-875 061

7-2

0-857 332

7-6

0-880 814

7-3

0-863 323

7-7

0-886 491

The question then arises as to how we can find the values

the function log x for values of the argument x which are
1

INTEKPOLATION

intermediate between the tabulated values,

e.g.

such a value as

to this question is furnished by the

of
Interpolation, which in its most elementary aspect
theory
"
be described as the science of reading between the lines

a;

The answer

=7' 15 2.

may

of a mathematical table."

In the further development of the theory of interpolation

shown how to find the differential coefficient of a
function which is specified by a table, and also to find its
integral taken between any bounds of integration.
it

will be

kind of interpolation was used by Briggs,* but interpolation of

the kind hereafter explained, based on the representation of functions
by polynomials, was first introduced by James Gregory j~ in 1670.

Suppose a function f(u) is given in

a, a + w, a + 2w, a + 3w, ... of its

Difference Tables.

2.

a table

for

the values

required to find the value of the function

when the argument has the value a + xw, where x is a fraction.
Before this problem can be solved by the method of inter-

argument

u.

It

is

polation, it is first necessary to form what are called the differThe quantity
ences of the tabular values.

f(a+w)-f(a)
is

denoted by

The

first

A/() and

denoted by A/(a + w).

denoted by A2/() and

while the quantity

denoted by

and

called the first difference of

+ w)

is

is

is

called the second difference of

is

called the third difference of /(.),

A 3/(a) and

/(),

so on.

It is convenient to
arrange

differences for

the tabular values and their

increasing values of the argument in what

called a difference table, as follows

*

f(ct).

f(a + 2w) -f(a + w], which

Moreover, the quantity

is

is

is

difference of f(a

Briggs' method
formulae.

difference

was, however, closely related to the modern centralCf. his Arithmclica Logarithm,! ca, ch. xiii., and his

Trirjonometria Britannica, ch. xii.

14, pp. 1, 73, 84, 88

is

Cf.

Journal of the Institute of Actuaries,

15, p. 312.

t Rigaud's Correspondence of Scientific

Men

of the 17th Century,

2, p.

209.

INTEEPOLATION
Argument.

Entry.

f(a)

A2

A.

A3

a+w

A 3/(r,)

a + 4:W

f(a

and similarly for differences of order higher than the third.

The first entry f(a] is called the leading term, and the differences
of /(a), that is to say A/(a), A 2/(a), ... are called the leading
differences.

(with

its

Evidently each difference in the table is the number

proper algebraic sign) obtained by subtracting the

number immediately above and

immediately below and to the left.

to

the

left

from

number

the

The sum

of the entries in any column of differences is equal to the

between the first and last entries of the preceding column.
Thus in
This affords a numerical check on the accuracy of the table.
the above table we have
difference

A 2/( +
An

example

3tt>)

= A 2/() + A3/(a) + A 3/(a + w) + A 3/( +

of a difference table

the natural sines of angles

from 25

Zw).

the following, which represents

40' 0" to 25 43' 0" inclusive at

is

intervals of 20".

Argument.

0-433134785866963

20"

0-433222179172439

40"

0-433309568404859

2541'0"

0-433396953563401

20"

0-43348 43346 47243

40"

0-433571711655565

25 42' 0"

0-433659084587544

20"

0-43374 64534 42359

40"

0-433833818219189

2543'0"

0-433921178917212

A2

A.

Entry.

2540'0"

7393305476

73892 32420

A3

-4073056
- 822

-4073878
-822

87385158542

-40 74700
-820

87381083842

-40 75520
-823

87377008322

-4076343
-821

87372931979

-4077164
-821

87368854815
- 4077985
8

73647 76830
-

7360698023

4078807

822

INTEEPOLATION

It will be seen that in this case the third differences are practically
when quantities beyond the fifteenth place are neglected, any
from constancy in the last place being really due to the

constant

departure

decimals in the original entries.

neglect of the sixteenth place of
the fourth differences are zero.

So

It will be found that in the case of practically all tabular

the differences of a certain order are all zero ; or, to

functions

than one unit in the last

speak more accurately, they are smaller
in
This fact lies
decimal place retained in the tables
question.
at the basis of the

method

of interpolation, as

we

shall

now

see.

Symbolic Operators. The formulae of the calculus of

differences may be very simply represented by the use of what
Of these we have already introare called symbolic operators.
3.

duced A, and we shall now consider another operator denoted

byE.
Let

w represent the interval between

successive values of the

E denote the operation

so
that
the
of increasing
argument by w,
E/(a) =/( + w)\ in
x
general we shall write ^ f(a}=f(a + xw], where x is an integer.
of the function /(),

argument

Now
so

definition

by

A/(

owt>)

the operators

and

let

we had &f(a + xw)=f(a + ocw + w) -f(a + xw],

= (E -

+ xw}.
It is therefore evident that
are connected by the relation A = E - 1 or

~L}f(a

E and A

E=1

+ A.

When

symbolic operators obey the ordinary laws of Algebra

they may be separated from the symbols representing the
functions to which they refer and treated independently in
the same way as symbols of quantity.
Now it may be
easily shown that the following relations are true for the

much

operator
)

= A/() + A/(&) + kf(c] +

=
&kf(a) 7jA/(), where k is a constant factor,
n
= A m + n/(), where m, n are positive
/(a)

+/(&) +/(') +

.}

.,

AA

The corresponding

Thus

in

many

integers.
identities for
are

respects the operators

algebraic symbols and may be combined

E
like

and

them.

behave

like

INTEKPOLATION
The following examples

illustrate the use of these operators

To express the nth differences of

X i.
of the successive entries.
,

a tabulated function in terms

i.e.

M (M
"/(a)

=f(a + nw) nf(a

Ex.

+ mo -w) + -

^_

1
J

'

'/(a

+ nw-2w)-.

To express the function f(a + xw} in terms of /(a) and the

when x is a positive integer.

2.

successive differences of f(a),

so that

f(a

4.

+ xw} =/() +

The Differences of a Polynomial.

table for
difficulty that the difference
as follows
:

X.

the

We

find

without

function y = x z

is

INTERPOLATION

Note that the table may be extended indefinitely when we know the
For by definition, when we add to an
third differences to be constant.
entry in a column of differences the corresponding first difference, the
sum so formed gives the next entry in the column. It follows that the
column of second differences can be formed from the leading term 6 by
repeatedly adding the constant third difference 6

the column of

first

differences being formed from the leading term 1 by adding in siiccession

The values of x3 are then obtained
the second differences 6, 12, 18,
from the leading term
by adding in succession the first differences
.

1, 7,

19, 37, 61,

...

Consider the case

when

the tabulated function f(x)

is

polynomial of degree n, say,

f(x')

Then

Now
n
)

= an + nwa n ~ 1 +

^^

'-w

an

~2

+ wn

so that

+~LW.

This
first

a polynomial of degree (n - 1) in a, and therefore the

differences of a polynomial represent another polynomial
is

of degree less

by one

unit.

repeated application of this result we see that

the 2nd differences represent a polynomial of degree

By

3rd

n-3,
.,

i.e.

the nth differences are constant.

the (n

n - 2,

0,

It follows, therefore, that

l)th differences of a polynomial of the nth degree are

all zero.

A table of values of any power

numbers may be formed by simple addition when
the leading term and the leading differences are known, in
5.

The Differences of Zero.

of the natural

INTERPOLATION

precisely the same way as in forming the table of cubes ( 4).

The differences of the leading term O p which are generally used
in forming a table of SOP, are known as the differences of zero.
,

They are

of frequent occurrence in the calculus of differences.

In order

to

form a table of reference of the

apply the result of

3 (Ex.

A nf(a) =f(a + nw)

nf(a

differences of zero

we

1),

+ nw

w)

^n(n

l]f(a

+ nw

2w)

- ...

and write

If

we now

substitute in this equation particular values for

obtain the equations

A n (F = nP - n(n

ftn-lip-l

and therefore

From

the relation

- 1) (n - Z)P .+
+ |(n _ i) n - 2)P &n OP = n A" ~ * IP ~ \

\}P

= n p-l _

n_

n(n

tion

P1

we

1P~\
(

1)

Aw

-1

+A

construct a table of values of

of this equation,

n,

= A /() +
iv)
~y() we see that
and equation (1) may be written

n
(& OP

that

and

n.lP + OP,

A"" /^ +
1

We now

i)i>

x, p,

remembering that

ft

~1

OP

A n O^

~1
(2)

).

by the repeated applica-

A01 = 0, A1 1 = 1,

and

also

INTEKPOLATION

This polynomial

we suppose the

is

denoted by

of [X]P to be unity,
[a]p

is

argument

called a factorial.

If

in the difference table

we have

= a(a-l)(a -2)

[a + !]P =(a + l)a(a-l)

(a-p + l),
(a-p + 2),
-2)
.

= a(a-l)(a-2)(a-3)
so that

and

[x]

interval of the

(a-p + 2)

&[x]P=p[x]P-

{(a

+ 1) -(a-p

It follows that

A|>]P

a result that

may now

following table

[S]P-I

[s+l]P =~ [sF

be used to tabulate the values of [x]P/p

as in the

INTERPOLATION

transformation, we obtain an expression for a polynomial of the

wth degree in terms of factorials
:

are constants and

a, (3, 7
thus obtain the result

where

We

</>

is

(a;)

= a + p[x]+ 7 [xJ + S[x] 3 +

i

<f>

Ex.

n (x)

To represent

the

function y

x4

12x3

a constant

v (say).

+ v[x]'\

+ 42a;2

30x

(2)

+ 9 and

successive differences in the factorial notation.

Using detached

coefficients
1

when

dividing by

x,

1,

2,

.,*

its

INTEKPOLATION

10

we obtain

Differencing this equation,

2
]

+ n(n-l)v[x] n -*.

(4)

Moreover,
2
1
&f(a + xw} = 2.3.3 + 2.3.4e[] + 3.4.5[>] +
+ n(n-I)(n-2) v [x] n -*,
and so on for differences of higher order. The values of
.

coefficients a,

ft,

the equations

Equation

y,

are found

...

(2), (3), (4),

(2)

may now

=0

by putting

in each of

so that

be written

x(x-l)(x-2)

fo-rc

l)

nl

This formula * enables us

in terms of the factorials
difference table

A?t/
J ^(

of the function is given.

/()}, which

is

an

x.

identity.

2,

f(a

+ Zw) =/()
+ {/(a + Zw) -

8.

''

express the polynomial f(a + xw)

^uhen a
x, x(x-~L), x(x-l)(x- 2),

=/(a)+ {f(a+w)
x

to

This general formula may be easily verified for special values of

When x = 0, it becomes /(a) =/()
When x= 1, then

When

(5)

the

Zf(a

+ w) +/()}.

The Gregory-Newton Formula of

general formula of the last section

may

The
Interpolation.
be applied to solve the

problem of interpolation.
Suppose that y is a function of an argument u and that the
values of y given in the table are /(), f(a + w), f(a
corresponding to the values a, a + w,
f(a + 3iv),
a + 3tv, ... of u. Also suppose that these values of the function
.

are entered in a difference table

n are constant.
which
y
correspond

order
of

and that the

We

differences of

are not supposed to know the values

to other values of u, such as u = a + \w.
*

Cf.

Ex.

2,

3.

LNTEKPOLATION

11

an analytical expression

It is required to find
mediate values of y.

The problem may be stated graphically

for these inter-

as follows

Draw

Let K, L, M,
the rectangular axes Ou, Oy.
be points on the u axis having abscissae a, a + w, a + 2w, a + 3w,
At these points erect ordinates KA, LB,
respectively.
.

respectively to the entries f(a), f(a + w),

MC, ND,
equal
f(a + 2w), f(a + 3iv),
.

Then the points A, B,

C,

D,

so

The

determined are points on the graph

"
"
problem of finding a smooth curve to pass through the points
has not a unique solution in fact an infinite
A, B, C, D,
number of curves satisfying these conditions can be found. As
of the function.*

our aim
solution

is

a practical one,
our problem.

we

of

naturally choose the simplest

Eemembering that the

Bx

simplest functions are polynomials,

we

if it is

inquire

possible to pass through the

points A, B, C, ... a curve

which is the graph of a poly-

nomial function of degree n.

We have already seen

L
T^TO

(
4) that for any polynomial
of degree n the differences of order n are constant and for the
set of values /(a), f(a + w), f(a + 2w), ... it has been assumed

that the differences of order

a polynomial of degree
f(a + w), f(a + 2 w),

exists

are constant.

This being

which takes the values

so,

f(a),

the argument u has the values

a
a
in
+ w, + 2w,
a,
fact, by the last section, we can write
down an expression for the polynomial. It is
.

y =f(a)

when

2!
/Yf /yi
%Aj\tAJ

n\

-/M,
* We do not know
anything about the portions of the graph intermediate
between these points, but we assume that the graph is a smooth curve for
our present purpose we can take this to mean that the function has finite
differential coefficients of all orders at every point.
;

INTEEPOLATION

12

where x
and where

A2/(a)
and

We

u by

connected with

is

the

relation

u = a + xw,

stands for /(a + w) -f(a),

stands for f(a + 2iv) - 2f(a + w) +/(),

so on.

now take the polynomial (1) to represent the function

values
y
for
of the argument intermediate between the
tabulated values.
The portions of the graph intermediate
shall

also

between the points A, B,

C,

may

therefore be filled in

by

drawing the curve

y=f(a+xw)

=/() + *A/() +

A^a) +

(2)

and in order

to compute the value of y corresponding to any

intermediate value of the argument such as a + ^w, we simply
substitute the value x = | in this formula,* which is the analytical expression required.

The fundamental problem of interpolation is thus solved.

The formula (1) is often referred to as Newton's formula of
interpolation, although it was discovered by James Gregory in
1670.1

The

application of the Gregory-Newton formula

following examples

is

illustrated

by the

Many books of logarithmic tables, etc., contain a table of the binomial

coefficients required in the interpolation formula (1), at intervals of 0-01 from
C=0 to a;=l.
Of. a

letter of Gregory to Collins of date November 23, 1670,

t
printed in
An example of the use of the formula is
Rigaud's Correspondence, 2, p. 209.
worked out on p. 211 of Rigaud. Collins was accustomed to send on to Newton
the mathematical discoveries of
Gregory (cf. Rigaud, 2, p. 335).
Newton's publications on interpolation are contained in
:

1.

The Methodus Differential

but written before

October 1676.
2. A letter written in 1676 to John Smith.
3. Lemma v. in Book iii. of the
Principia published in 1687. The above
formula is Case i.
4. Various references in the Commerdum
Epistolicum of dates 1672/3 to 1676.
These have been collected and edited by D. C. Eraser in the Journal
the
published

in

1711

of

Institute of Actuaries, 61
(1918-19), pp. 77

and

211.

INTEEPOLATION
Ex.
!C

1.

From

the table given leloiv to find

13

the entry corresponding to

= 21.
A2

A.

Argument.

Entry.

20

0-229314955248

22

0-230016702495

A3

A4

701747247
602297

- 1944

702349544
600353

0-230719052039

24

598413

0-231422001936

2G

-1937

703548310
596476

0-232125550246

28

-1940

702949897

704144786
30

0-232829695032

Here

w = 2,

= 20,

f(a

and x =

+ xw) =/(2 1),

X(X

2)

A 3/(20) +

|.
.

= 229314955248 + 1(701747247) - 1(602297) - TV (1944)

= 229314955248
_f75287-l
"
+ 350873623-5
= 229665828871-5 -

121-5

75408-6

so

/(21)

= 0-229665753463.

Ex.
4h

30 m

To find the co-ordinate X of the sun on November 10, 1910, at

G.M.T. (X is the sun's true geocentric co-ordinate measured on a line

2.

passing through the true equinox of the date).

The Nautical Almanac gives the following readings from which
construct a difference table

we

1910.

-X.

November 9-0

0-6850997

9-5

0-6787188

A2

A.

A3

-63809

-514
-64323

10-0

0-6722865

10-5

0-6658032

11-0

0-6592696

11-5

0-6526859

-510
-64833

-503
-65336

-501
-65837

We must interpolate for

is

12 b

Then

4h

30,

4 h 30 m from

November

aa a fraction of the

10-0.

The argument

argument, gives x

= 0-375.

INTEKPOLATION

14

x= 9-5740313
= 9-7958800(?i),

log
log (a;-

where

(ri)

indicates that

l)

9-7958800

is

log

the logarithm of a negative

number

= 9-6989700
= 9-0688813(?i)
log Jz(z-l)
1 = 9-5228787
log
= 0-2 108534(n)
log (a -2)
= 8-80261 34.
lx(x -!)(- 2)
log

Also
log

(-64833) = 4-81l7961(?i)

log

= 9-5740313
=
(-64833z) 4-3858274(?i)
= log (-24312-4)

log

503)

log |x(^

log

log

= log

- X = 0-67228650 - 0-00243124 +
- X = 0-6698612.

Therefore
finally
9.

An

Form

Alternative

= 2-7015680(?i)
= 9-0688813(?i)

=
|x(x- 1)(- 503) 1-7704493

= 0-3010300
log 2
=
8-8026134
2)
log lx(x-\}(x= 9-1036434 = log
log lz(:e-l)(z-2)(2)
and

1)

of

58-94

0-1.

0-00000589,

Gregory - Newton

the

Formula. The Gregory-Newton formula may be written in

an alternative form which is convenient when an arithmometer *
is used.
^Rearranging the formula of the last section in the form

and assuming the differences of order n to be constant, we may

replace the Gregory-Newton formula by
(1)
f(a + xw] =f(a) + xu lt
=
where
u^

A/(a)

1(1

x)u z

wn = A w/(), which

is

constant.

When

computing a value of the function by this method, we

the constant difference un and calculate in succession
with
begin

un _ 2
u lt finally substituting the value
.,
of Wj in equation (1).
The following example will serve as an
illustration of this method
the values of u n _ ly

When

an arithmometer

is

found useful for this purpose.

not available Crelle's Calculating Tables will be

INTEKPOLATION
= 24-46980 05207

Tofindf(6) when 6

Ex.

15

020, having given

AV

A.

0.

/(#).

24-4

0-216 198 561 343

24-5

0-216366833650

As

A*.

168 272 307

745715

169018022

768

0-216535851672

24-6

746483
169 764 505

0-216705616177

24-7

747256
170 511 761

748033

171259794

782

0-217047387732

24-9

777

0-216876127938

24-8

773

748815
172 008 609

0-217219396341

25-0

Here
Hence

w = 0-l,

a = 24-4,
x = 0-698 005 207 02.
= A3/() - |(3 - z)A4/(a) = 768-0-576x5
= 765-1,
= A 2/(a)-l(2-o> 3 =745 715-0-434 Ox 765-1
2
= 745 383-0,
M 1 = A/(a)-l(l -a> 2 =168272 307-0-150 997 4
= 168 159 756-1.
3

<w

x 745 383

Then
f(a

or

+ xiv) =/() + asMj

= 0-216 198 561 343 + 0-698 005 207 02 x
= 0-216 198 561 343
+ 117 376 385,
=
0-216
315 937 728.
/(0)

10.

The Binomial Theorem.

By

0-000 168 159 756

use of the operator E,

we

can write the Gregory-Newton interpolation formula in the form

When

thus written, the formula

obtained by expanding

is

seen to be the same as that

(1 + A)* by the Binomial Theorem in

and then operating on f(a) with the

ascending powers of A
terms of the series so formed,

i.e.

The Binomial Theorem was made known

(in correspondence)

Gregory-Newton formula in fact, Newton

seems to have discovered the Binomial Theorem by forming the
n
expansions of (1 + x) directly for integral values of n, and then
In the
writing down the powers of x in these expansions.
case of the coefficient of x 2 he would have
six years after the

INTEEPOLATION

INTERPOLATION
(b)

x.

sin x.

28 40' 00"

0-479713113250246
755651470168

10"
20"
30"
40"
50"
28 41' 00"
10"
If y

2.

= 0,

17

798
840
883
925

188
724
259
793

562
526
363
072

452
998
705
474

968325653205
0-480010857105798

= 2.c 3

x2

+ 3x +

1,

calculate the values of y corresponding to

and form the table of differences. Prove theoretically

that the second difference is 12x + 10 and verify this numerically.
3. Find the function whose first difference is the function

1, 2, 3, 4, 5,

ax3 + fix 2

yx

+ 8.

Find the successive differences of

4.

1/x, the interval being unity,

cos nx, the interval being w.

(a)
(b)

5.

Express f(x)

= 3x3 + xz + x+
ax(x

1) (x

2)

in the form

+ /3x(x -

1)

+ yx + 8

Calculate the values of/(x) for x

by comparing
etc., and form a difference table.
Verify the equation
coefficients.

/(x)

=/(0)

+ xA/(0) +

Compute the

6.

^1^(0) +

third difference of /(51)

from the following table of entries

52
51
x

1 K- C

2)

1, 2, 3, 4, 5,

A 3/(0)

by the formula

of

3,

Ex.

1,

140608

132651

/(x)

^-

= 0,

53

54

148877

157464

table.
verifying the result by means of a difference
7. Given the table of values

y
find

by means

function of
8.

of the

3
16

-2

-1

1
1

Gregory-Newton formula an expression

for y as a

x.

Construct a difference table having given

= 0-776 701 1840

= 0-776 773 802 4
5-952 = 0-776 846 408 7
5-953 = 0-776 919 002 8
5-954 = 0-776 991 584 9

log 5-950
log 5-951

log
log
log

and determine log 5-9505.

9.

Let

p,

q,

r,

be successive entries in a table corresponding to

equidistant arguments.
Show that when third differences are taken into account the entry

INTEKPOLATION

18

corresponding to the argument half-way between the arguments of q

and r

is

+r
.

2
10. Let p,

s
q, r,

in a

+ (q

+ r)-(p + s)-.

(De Morgan.)

16

be successive entries (corresponding to equidistant

It is required to interpose 3 entries (corre-

table.

arguments)
sponding to equidistant arguments) between q and r, using third differShow that this may be done as follows
ences.
also
Between q and r interpose 3 arithmetical means A, B, and C
2s
between 3q 2p s and 3r
p interpose 3 means A', B', and C'.
Then the 3 terms required are A + ^VA', B + ^B', C + ^C'.
:

(De Morgan.)
11.

Determine log 6-0405, having given

= 0-7810369386
6-041=0-7811088357
6-042 = 0-7811807209
6-043 = 0-7812525942
6.044 = 0-7813244557

log 6-040
log
log

log
log

Using the method of

the values
12.

9, find sin

sin

B.

6.

0-410718852614
0-414693242656
0-418659737537
0-422618261741
0-426568739902
0-430511096808

24-25
24-50
24-75
25-00
25-25
25-50
13.

24-4698005207, having given

Given the values

x.

f(x).

858-313740095
869-645772308
880-975826766
892-303904583
903-630006875

2
3

calculate /(I -5) by the Gregory-Newton formula.

14. The values of a function
corresponding to the values 1, 2, 3, 4, 5
of the
argument are

0-198669,

0-352274

0-237702,

0-276355,

0-314566,

Calculate the values of the function when the

argument has the values 1-25 and 1-75 respectively.
15. Using the difference table
find the values of
given in
2,
sin 25 40' 10" and sin 25 40' 30".
Also verify the answers
respectively.

25
25
sin 25
sin 25
sin

sin

= 0-433 353 261 493

= 0-433 440 644 614
41' 30" = 0-433 528 023 660
41' 50" = 0-433 615 398 631
40' 50"

41' 10"

obtained by taking x numerically

less

416,
711,
896,
149,

than unity in the formula of

8.

INTEEPOLATION
16. Calculate log tan

log
log

log
log
log

log

24

24
24
24
24
24
24

tan
tan
tan
tan
tan
tan

0' 5",

19

given the values

= 9-648 583 137 400 95

0' 20" = 9-648 696 457 723 08
0' 40" = 9-648 809 758 267 66
1'
0" = 9-648 923 039 045 83
1' 20" = 9-649 036 300 068 75
1' 40" = 9-649 149 541 347 57.
0'

0"

f oo
17.

The following

table gives the values of I(x)

e~ s ds

/;
x.

I(x).

0-00
0-01
0-02
0-03
0-04
0-05

0-886
0-876
0-866
0-856
0-846
0-836

= 0-025 by

Calculate I(x) for

226
227
229
235
248
268

92
24
57
90
22
53

interpolation and verify your result by

use of the formula

1(0)

-!<*)

= *-

_--- +
/yO

/yO

/y>

..

CHAPTER

II

INTERPOLATION WITH UNEQUAL INTERVALS OF THE ARGUMENT

11.

We

Divided Differences.

have so far assumed that the

values of the argument proceed by equal steps but with data

derived from observation it is not always possible to complete a
;

For example, when astronomical

difference table in this way,

observations are disturbed by clouds there are gaps in the

records.

Consider the case in which the values of the argument, for

is known, are unequally spaced, and suppose
x = a 1} x = a 2>
that the values of f(x) are known for x = a
=
-a
a
a 3 -a 2
a
intervals
x
where
the
.,
n
z
.,
1}
c^-a^,

which the function

In place of ordinary differences

a n -a n -i need not be equal.
Let
as divided differences*
what
are
known
we now introduce
us form in succession the quantities

and

These are called divided

us form

so on.

"2

differences of the first order.

let

Moreover,

2>

1>

\
a7 o),

/K

"0

a a)

-A a *

gi)

~
"3

//
7 N
=J( a3> a '> a l)-

These are called divided differences of the second order.

Also

let

This is called a divided difference

The
of the third order.
divided differences of higher orders are formed in the same way,
so that the order of a divided difference is less
by unity than the

number

of

arguments required

for its definition.

* Divided differences
might fairly be ascribed to Newton,

Lemma

v.

The

term was used first by De Morgan, Diff. and Int. Calc. (1842), p. 550, and afterwards by Oppermann, Journ. Inst. Act. 15 (1869), p. 146. Ampere, Ann. de
Gergonne, 26 (1826), p. 329, used the name interpolator!/ functions.
20

INTERPOLATION
Divided differences
as follows

be expressed more symmetrically

may

- ai

f(a 2 a v

21

-a
o

ai

/K)
2

+ _J_(/K)

-a

(^

"T"

-a

__
(a Q

rtj)

(aj

2)

2)

(az

)'

K-

In

(2 ~ ao)

K-

general, as

may

(2 ~

l)

and

is

in fact the

l>e

easily

difference oftliepth order is

sum of

a s)

(1 -

__
(

fl)

a3 ~ a o) ( a S ~

l)

a3

z)

shown by induction, a divided

a symmetric function of its arguments

(p

1)

functions of the form

/K)

difference-product of ar with a

a lt a 2

.,

r -i,

a r +i>

>

ap

It is evident from this statement that when the arguments required

form a particular divided difference are arranged in a different order,
the value of the divided difference remains unchanged, e.g.

to

/Ku

n_2'

_!>

>

a l> a o) =/( a O' a V

ffl

>

2>

a i-l

ft

n)'

Divided differences are arranged in a taUe of divided

ences as follows
Argument.

differ-

Entry.

/K)
/(%, a a

f(a z a3 )
,

4,

a5 )

3,

a4 )

INTEKPOLATION

22
The following may
differences

X.

serve

as

an example of a

table

of divided

INTEEPOLATION
a homogeneous function of aQ a 1 of degree (n

23

f(a

1).

Moreover,

a v a 2)

which is a homogeneous function of a Q a v a 2 of degree (n - 2).

a l5 a z
In general /(
a^) is a homogeneous function of
ap of degree (n - p). Taking p = n, we see that
aQ av a z
,

f( a

.,

.,

an )

is

.,

a constant.

n
Corollary: The divided differences of order (n + 1) ofx are zero.
IV. The divided differences of order n of a polynomial of the
nth degree are constant.

This theorem follows immediately from theorems I., II., and

III., since the divided difference of order n of each of the terms

whose degree

is less

than n

is zero.

divided difference of order r may be expressed as the

Y.
quotient of two determinants each of order r + 1.
Consider the divided difference of the third order,

*
_ v /( ) (difference-product of a 1} a 2 3 )
difference-product of aQ a v a 2 as
,

Now
of the

a difference-product

may be

expressed as a determinant

kind known as Vandermonde's, thus

product of a lt az

Therefore

3)

24

INTERPOLATION

or

/K) /K) /K) /K)

a,

1111
so in general for differences of order higher

and

13.

than the

Newton's Formula for Unequal Intervals.

third.

Let f(u)

be a function whose divided differences of (say) order 4 vanish

and suppose its values for 4 arguments

are negligible
a,
a
so that the table of divided differences
known
are
2>
s

or

a
is

av

as follows

Argument.

Entry.

/K)

/K)

n a,.
0'
i>
J/Yrt
\
,

a "/)

/K)

Ctc.

f(a v

We may

ff

2,

obtain the value of the function for any other

Beginning with the constant

argument u

in the following way.

difference which is of order 3> we
f(u, a

have

a v a a ) =/(a c a v a 2 a 3 ).

(1)

definition of the divided difference of order 2,

By

f(u, a

j)

=f(a

av

a0>

2)

and therefore
f(u, a Q

Again by

aj =/(

-a
av a z
2 )f(a Q

a lt a z ]

(2)

definition,

Au

>

a o) =/(o>

j)

+ ( u ~ a i)f( u ao> a i}>

>

and substituting in this equation the value of f(u, a Q a^) from (2),
~
~ i) ( u - a 2 )f(a ,a v a z ,a s).
f( u a o) =f( a o> a i) + ( u a i)f( a o> a i> 2 ) + ( u
~a
u a o)'
Also by definition
(4)
o)f(
f(u) =/(a ) + ( u
~
~ i)/K> i> ^2)
~
u
or
ft
u
a
+
/( w ) =/( o) + (u
i)
o)
(
o)/(o'
(
- 2
+
!,
2;
(5)
,

>

>

rt

(w

(u

ttj)

(u

)/(a

%).

INTEKPOLATION

25

From the

equations (1), (2), (3), (4) the quantities^?/, a 0> lf a 2 ),

f(u, a
aj, f(u, a ), f(u) are now known and may be inserted
in the table of divided differences thus *
flf

Argument.

Entry.
~f{ II
I \ Uf

/]/

Uj

I
I

f( u
o

/K)

ai

/K)

o)

>

f(ai>
/(

,i)

/fa,

/K>

i,

2)

/K,

2,

03)

/K>

2)

2)

1>

2.

3)

Formula (5) may evidently be generalised to express a

function whose divided differences of order (n + 1) are
negligible
or zero, in the form

/fa)

=/K) + fa - s)/K>

i)

fa

fa

i)/K

i,

.)

This formula was discovered by Newton, f

The first term on the right-hand side of this equation represents the
The
polynomial of zero degree, which has the value /( ) at u = a
first two terms
together represent the polynomial of degree 1, which has
.

and f(a^) at
and a-^ respectively, and so on.
The remainder term which must be added to the right-hand

the values /(

the equation in order to obtain strict

accuracy
(u

But

this

-a

(u

- Oj)-

term vanishes

if

(u

- a ny(w,

is

side of

in fact
.

ffij,

).

,,

the divided differences of order

are rigorously

constant.

Ex.

From

the table given below to find the entry corresponding to

=0

3-7608.

f(x)

-3989423

- 500
a x = 2-5069

-3988169

a 2 = 5-0154

-3984408

-199
- 1499
-199

-2496
3

= 7-5270

-3978138

* In
practice the value of f(u) is usually found
divided differences in this way, as in the worked-out

t Principia (1687),

Book

iii.

Lemma

v.

Case

ii.

by forming the successive

example below.
Cf. Cauchy, (Euvres, (1) 5,

p. 409.

(D

309)

INTERPOLATION

26

successive divided differences of /(),

Forming the

we

where u =3- 7608,

find

/C,

~ 199
a)=
i>
)= -500 + 1-2539 x(- 199)= -749-526,
= -3989423 + 3-7608 x ( - 749-526).

i)=/(0'

o>

/( ;t)

'

f(u]

The
14.

calculated value

is

therefore 0-3986604.

The Gregory-Newton Formula as a Special Case of Newton's

The Gregory-Newton formula may be regarded as the special
the formula of the last section when the intervals of the argument

Formula.
case of

are equal.

For in Newton's formula

By

for

unequal intervals suppose that we put

we

constructing a table of divided differences,

see that

In the same way we find

and

so on.

we now replace
the argument becomes
If

f(a

which

u by a + xw, the formula

X

+ xw) =f(a) +
is

for

unequal intervals

of

the Gregory -Newton formula.

15. The Practical Application of Newton's Formula.

In laboratory computation from Newton's formula, we proceed
by a method which is really identical with that given above

(Ex.

13).

Rearranging the formula of

+ (u ~

,/H =/K)
+ (uThis equation

o)

[/K>

!){/(<,,

may
f

Vl

see that

i)

a lt a 2 ) + (u -

2 )(/(a

a v a 2 a3) +
,

.)}].

be written in the form

f( u )

where

we

13,

=/K

=/K)
tt

+ ( u ~ oK>

l)

'

vr

= ?-th

divided difference

v n =f( a o> a v

>

(u

- a r r+1
)v

a n)> a constant.

The v's are computed in the following order v n -i, v n -z,

The value of f(u) is then obtained from equation (1).
:

v1

>

Entry.

150

INTERPOLATION

28

In terms of these divided differences with repeated arguments

the formula of Newton becomes
)
0>
)/K, a ) + (u a Yf(a
/(u) =/(a ) + (u- a 3/K- o> o> o) +
+
,

(u

This formula will be used later to obtain an expression for

the derivatives of a function in terms of its divided differences.*
a;

Ex.

Given the

values^
- 3).

11

27

QQQ

mi5

34

42

,,

356Q6 6851Q

MI terms o/ powers of (x

and extending
Constructing a table of divided differences
for x = 3, we obtain
repeated arguments
X.

42

it to

include

INTERPOLATION
a polynomial of degree n,
1) are zero, i.e.

Since f(x)
order (n

al

/K>
the

Arranging

tt

of

factors

a n>

2>

divided differences of

its

is

29

= 0.

denominators

the

the above

in

fractions so that the first factor in each denominator

form

(x

is

of the

- a we obtain
f ),

A*)
(x -fl. )(aj-

1)

(x-an )

-a
Q ) (a
a^ (a

(x

-a z ).

- an
)

(a

/K)

(x-a1 )(a 1 -a

f(a n )
(x

- an n - a
)(a

(aw

-a w _i)'

which is Lagrange's formula in a form suitable for computation.*

Another way of writing this formula is obtained by multiplying both sides of equation (A) by

(x-aQ )(x-a l )(x-a 2 )

(x-a n ),

when we obtain

(x-a^(x-a^
/w ~K-i)K),

(a;

a2)

(x

(x

an )

,,

- ---

(a l -a

)(a 1 -a 2 )

}(x-a l )
+ 7 (x-a w
-a
-a,
(an

)(a n

(ai -a^

1)

f( a

\T\
(an -an _ 1y
1

(a;-a M -i)

K-r

(x-a n )

,~

n}'

(&)

when

a set of experimental data obey a

important
law which can be expressed algebraically as a polynomial of degree n,
then not less than (n + 1) observations are required in order to construct
If only n values were used, the resulting polynomial
the polynomial.
would be of degree (n - 1). Before applying the Lagrange formula it is
therefore necessary to ascertain the order of the divided differences which
to note that

It is

are of constant value

and thus find the proper value

for n.

'

Ex.

I.

Given the values

of f(x) corresponding
*

to

6g> ^ 64>Q 44>Q 3 Q-1

to

calculate the value

x = 27.

Lagrange's formula was

first

sur Us
published in his Lemons dUmentaires

matlilmatiques, in 1795, reprinted in his (Euvres,

7, p.

286.

INTERPOLATION

30
Applying formula

we

(A),

obtain

/(27)

(27

14) (27

17) (27

31) (27

35)

68-7

(27

14) (14

17) (14

31) (14

35)

64-0

+
(27-

17)

(17-

14) (17

-31) (17 -35)

44-0

*
(27

31) (31

- 14) (31 - 1 7) (3 1 - 35)

39-1

f
(27
/(27)

_
~

68-7

/.

/(27)

is

49-3.

The required value

Given the data

2.

35) (35

13923

4160

Ex.

A} =

(B),

14) (35

64-0

7560

= 49-317

<5

J.

^i

17) (35

44-0

_
~

3808

31)

39-1
12~096'

(approx.).

.,

J\3C)

Applying formula

to form the cubic function

~t

of x.

we have

(z-l)(s-2)(a!-5)
(0-1) (0-2) (0-5)

--

a<a;-2)(a-6)

x(x

1) (x

5)
'

1(1

-2) (1-5)

2(2-1) (2 -5)

5(5-1) (5 -2)

=
18.

a;

+ x z - x + 2.

An

alternative proof of Lagrange's formula

is the following

use of determinants
Let

Pn

denote a polynomial of degree

Substituting in succession the values a

obtain

=A
2
=
/(aj A + Ba x + C^ +

f(a

n,

and put

av

(an)

= A + ~Ba n

by the

+ La^,

.,

an

for x,

we

INTERPOLATION
Eliminating A, B, C,
ally we have

'o)

from these equations determinant-

Expanding
first

row,
1

/K)

11

/K) /K)

..

X
X2

ft*

xn

a n

31

ft,,

ft,

this determinant according to the elements of the

we

see that

...

INTEEPOLATION

32
,

f
yi

71

-g
-a

(a?

) (a?

((!

+/K);

which

-a

'(an -a

(a

(x
n

- a 2 ) (a; - fl 8 )
- 2 ) x - as )

(x-a n )

(a^-a

(x

)(an

-a1)

(an

n -i)'

the formula required.

is

x, each of which has the

a n respectively.
In the
we
consider
of
functions,
mathematics,
however,
only
practical applications
a n are sufficiently close together, any one of the
such that if a a lt
.,
functions may be represented with tolerable accuracy by the polynomial
an
The
P n for the range of values included between a Q a T
.,
formula may thus be used for interpolation.

There

is

values /(a

an

infinite

/(e^),

),

.,

number

of functions of

/(<%) at

a lt

.,

19.

The

Term

Remainder

in

Formula

Lagrange's

Let f(x] be a function of the real

Interpolation.*
variable x defined in an interval to which belong the values
of

xv

of order

.,

xn

and possessing in

this interval the derivative

n.

Consider the function g(x], where

x w-l

f(x]
-PI

71-1

Xn

J v^o;

X n-l

The determinant vanishes

the differential calculus
(n
x.

for the values

we

xv

see that since g(x)

.,

xn

By

vanishes for

derivative g'(x) vanishes for n values of

the second derivative for (n-\] values, and so on; the nih

1)

values of

x, its

Thus
derivative vanishing for one value of x in the interval.
xn such
there exists a value x intermediate between x x lt
.,
,

that

n
(fi

\x)

Forming the nth derivative

of the

entiating the variable elements of the

*

= 0.

Pcano, Scritti

offcrli

determinant by
row, we have

first

ad E. D" Ovidco (Turin,

1918), p. 333.

differ:

INTEEPOLATION
fn (x)
f(x

...

nl

If

we

we expand

determinant according to the elements of

solve for f(x ) in the
resulting equation,

this

column and

first

=0.

x nn

the

33

find

^0
T^
w-i

__

ov

/yi

/ /yi

tX./o /

tl

1.

//y.

^37

w ^i

~~

/j-

^2} \^Q

(rvt

'

'

'

(^0-^1)^0-^3)
~
~X

(X Q

(^0

Xj) (X Q

- X2
)

(^o

^2)

ff]j

^nj

//

v\

i/

fco-a

s)

^'ij (,"^2

""

~~ ^v I
(&-I
11 1
\
1

^oj
o/

^0
/

X2 ~

-%n -l)

(ff

2)

(
3-

X ~

^M)

nl

where x
This

is

some number intermediate between x

is

Lagrange's formula with a remainder term.

EXAMPLES ON CHAPTER II
1.

If /(x)

= -g, find the divided differences/^, &),/(',

Ju

2.
,

3.

= g(x) + h(x\ where g(x) = x*

= 0(5, 7, 11, 13) + A(5, 7, 11,

If /(x)

7, 11,

and

13)

h(x)

and /(a,

6, c),

= x3

6, c

verify

d\

that

13).

Given the values

a;

f(x]

10

11

13

48

100

294

900

1210

2028

form the table of divided differences and extend

of the function for x
3 and x
1 4.

it to

include the values

4.

Find the function

11

13

14

18

19

21

f(x)

1342

2210

2758

5850

6878

9282

(a)

(&)

f(x) in each of the following cases

16

17

19

23

29

31

f(x)

65536

83521

130321

279841

707281

923521

by means of a

table of divided differences.

INTEEPOLATION

34
5.

Calculate /(I), given the values

704969

729000

804357

830584

884736

SB

658503

f(x)

6. Assuming /(x) to be a function of the fourth degree in

value of /(1 9) from the values

11

17

21

23

31

f(x)

14646

83526

194486

279846

923528

x,

find the

7. The values of a cubic function are 150, 392, 1452, 2366, and
5202, corresponding to the values of the argument 5, 7, 11, 13, 17
respectively.
Apply the Lagrange formula to find the function when
the argument has the values 9 and 6 "5 respectively.
8.

and

Find an expression for the function in each of the examples

using the Lagrange formula of interpolation.

(7),

(6)

CHAPTER

III

CENTRAL-DIFFERENCE FORMULAE
20.

In this chapter we
Central -Difference Notations.
formulae of interpolation which employ

shall consider certain

differences taken nearly or exactly from a single horizontal line

In order to express these simply it is
of the difference table.

convenient to modify the notation of the calculus of differences.

Several systems of modified notation are in use.
One,
which we shall frequently employ, was introduced by W. F.

Sheppard* and
It is

table.

equivalent to

AE

from any number

table,

i.e.

Since

2
,

where

to the

E = 1 + A.
S = AE~^

as usual denotes the transition

number next below

and

it

in the difference

A = SE^, we may

therefore

"2

Ee writing

from the following difference

which may be regarded as

will be understood

based on a symbol

IS

"2

the ordinary difference table,

Argument.

write

we

obtain

Entry,

2w

u_ 2

u_,

8u_

8u

u_j
8

-1

-i

a
Su

a+

If we suppose each row of the difference table to be numbered with

the suffix p of the corresponding entry up , or, in the case of a row
situated midway between two entries up and up+l to take the number
we see that A 2rM the differences of even order of UQ, are reprep+
,

1,

sented in the central -difference notation

*

Proc.

London Math.

by

8 2r u n since

they are situated

Soc. 31 (1899), p. 459.

35

CENTKAL-DIFFEKENCE FOKMULAE

36
in the row

The

r.

differences of

the expression 8 2r+1 ur+

odd order

since they lie in the

It is often required to find the

A 2r+1 u
row

are represented

by

+ %.

mean

arithmetic

of

two

In the
adjacent entries in the same column of differences.
8 system of notation we indicate this mean by the symbol //,.

Thus

fj.8u

3
u is |(3 3 w _ , +
to be %(8u _ +
^
8u^), p8
for the mean differences of the other entries.

is

and so on

mean

denned

differences

may be

inserted in the table to

fill

#X),
The

in the gaps

that occur between the symbols of the quantities from which

they are derived.

In another notation which was suggested by

symbol

is

used instead of

following difference table

Argument,

8.

*2w

u_ 2

u_i

A-a

Joffe

the

in the

+w

A M_j
2

AM$
a

A.

illustrated

Zcy^-j

S.

is

Entry,

The notation

& 2u
A

A
A

M_j.
3

A4

^Q

MJ

+ Zw

tt

u2

The Newton -Gauss Formula of Interpolation. Suppose that a function f(a + xw) is given for the values
.21.

a - w,

a,

a + w, a + 2w,

of its argument.
If in the Newton formula for

aQ =

unequal intervals we take

= a + w, a z = a w, a3 = a + 2w, a^ = a 2w, and so on,

a,
and denote a + xw by u, we obtain
~ a a a + w + u - a u ~ a ~ w a a + w,a=
f( u ) f( a + ( u
)f(
)
(
) (
)f(
w)
+ (u a)(u-a w] (u a + w]f(a, a + w, a w, a + 2?/ )
a1

>

>

(u

a) (u

-a

+ (u-a)(u-a-

a - 2w)

a + w) (u
f(a, a + w, a-w, a + 2vv, a
2iv)
-a+
w) (u
w) (u-a- 2w) (u -a + 2w)
- 2w, a +
f(a, a + w, a-w, a + 2iv, a
3w).
w} (u

'

(1)
* Trans. Act.
Soc.

Amer. 18

(1917), p. 91.

CENTKAL-DIFFEKENCE FOEMULAE

37

The divided differences contained in this equation may be

written in the ordinary notation of differences as follows
:

f(a,a + w)=f(a,

a + w, a - w) =

f(a, a

+ iu,a- w, a + 2w) =

^- A /(a
3

- w),

etc.

Equation

(1)

thus takes the form

-? A 2

f(a + xw) =f(a) + *A/() +

(x
^

l)x(x
v

/(

1) (x

2)

- w)

A 4/(a - 2w)
.,

/AN
(A)

This formula, which

Newton,

is

one of the group of formulae

often called the Gauss formula.
is

known

to

The differences used in this formula are as nearly as possible in the

horizontal line through /(a) in the original difference table. The formula
is therefore convenient for use when the value of the
argument for which
This
is required is near the middle of the tabulated values.
formula may be represented more simply by using the symbol (ri) r to
denote the binomial coefficient
the function

n(n

1) (n

2)

(n

-r+

1)

~^T
so that it

/(a

may

be written

+ xw) =/(a) + xA/(a) + (z) 2 A 2/(a - w) + (x +

l) 3

A 3/(a - w)
.

22.

The Newton-Gauss Backward Formula.

formula of the last section another

is

often

from

used

when x

is

may

From

(B)

the

be derived which

measured in a negative direction

towards decreasing values of the argument.

write f(a - xw) in the form f{a + x(- w}} and change
the sign of w in the discussion of the last section.
The
f(a),

Suppose we

i.e.

CENTRAL-DIFFERENCE FORMULAE

38

arguments and corresponding entries is then

Instead of A/(a) in the Newton-Gauss formula we
have f(a- w) -/(), or -A/(-w); A 3/( - w) in the

order of the
reversed.

now
-

A /(a - 3w),

f(a

and

xw} =f(a]

becomes

above formula

A 3/(a - 2w)

We

so on.

A 5/(& - 2w)

becomes

thus obtain the formula

3
xhf(a -w) + (x)^f(a w) (x + l) 3 A /(
2w)
5
4
+ (x + l) 4 A /(a 2w) -(x+ 2) 5 A /(a

which has been called the Newton-Gauss formula for negative

interpolation, or the Newton-Gauss backward formula.
In the Gauss formula
23. The Newton-Stirling Formula.

f(a + xw) =f(a) + a?A/(a) + %x(x

-

1)

A 2/(a - w]

+ l(x + l}x(x l)A /(a w)

4
~
+ ^( x + l) x x ~ 1)
2)A /(
(

the terms

may

be rearranged thus

f(a + xw) =f(a) + x{kf(a)

tff(a

w}}

~f(a - w)

Suppose we replace the differences of even order within the

brackets by differences of odd order, using the identities
(a

A /(a
4

and

so on.

,,

We
//

2w)

I 2 ) (s2
5!

= A/() - A/(a - w),

= A 3/(a -w)- A 3/(a -

2w),

obtain the result

A/() + A/( - w)
-I

f(a + xw) =f(a) + x^--t

x(x

w)

- 22
)

^p

A 5/fo -

2;) +

xz

.,
wAzf(a - w)
>

A 5 /(a -

^~

(A)

CENTRAL-DIFFERENCE FORMULAE

39

This formula, which was first given by Newton,* was afterwards studied by Stirling f and is called the Newton- Stirling

formula.
The mean-differences i { A/() + A/(a - iff) } i { A3/( ,

2 10)

+ A3/(

->)},

are completely symmetrical with regard to increasing and decreasing

This fact enables us to express the formula very concisely
arguments.
by means of the central-difference notation of 20
etc.,

5!

where

fjiSu

and

= -^(8_ j + Sttj),

so on.

24 The Newton-Bessel Formula.

In the Newton-Gauss

formula
2
f(a + xw) =/(.) + zA/() + lx(x l)A /(a w)

+ \(x + l)x(x + J-4 (+ I)a;(a5 us substitute for i/(a), iA 2/(-w), -iA 4/(-2w),
values obtained from the identities
let

etc.,

their

A 2/( - w) = A 2/() - A 8/(a - 10),

A4/(a - 2w) = A4/(a - w] etc.

The above equation becomes

= |{/(a) +/( + w)} +
/( + xw)

(a;

l)g(a;

1) (g

2)

(*

w)}

(A)

is symmetrical with respect to the argument (a + ^

This formula, which was first given by Newton \ and later
used by Bessel, is called the Newton-Bessel formula.

which

Newton, Mcthodus Di/ercntialis (1711), Prop. iii. Case

Methodus Differentials (1730), Prop. xx.
Methodus Differentialis (1711), Prop. iii. Case ii.

i.

f Stirling,

Diffcrentialis (1730), Prop. xx. Case

ii.

Stirling,

Mcthodus

CENTKAL-DIFFEKENCE FOKMULAE

40

If in this formula

f( a

+ i w + yw ) =

we

i {/(,)

x-\=y,

write

+f(a + w)} +

becomes

it

7/A/(rt)

i
(a

- w)

(B)

25.

When

Everett's Formula.

required to inter-

is

it

of tables
polate between /() and f(a + iv) in the construction
use a
statisticians
the
of
subdivision
intervals,
frequently
by
formula due to Everett,* which is generally written in the form

denotes (1 -x), and where as

where u x denotes f(a + xw], and

usual

denotes

A2 E - \

Thus

= l-f = i.

= i,

for
u^,

This formula involves only even central differences of each

of the two middle terms of the series between which the
interpolation has to be made.

To prove this formula we eliminate from the Newton-Gauss

formula
f(a + xw) =f(ci) + zA/() +

+ (x+

(z) 2

l) 4

A /(

the differences of odd order

AA) =/(
A 5/(

A 2/( - w) + (x+

2w) +

(x

l) 3

A 3/( - w)

A /( -2w) +
5

2) 5

by means of the relations

2
- 2
Ay( - w) = A /() A /(
- A4/( - 2w)

+ w) -/(),
= A 4/(r6
2w)

w)

w),

The Newton-Gauss formula becomes

a ~ w)
f(a + xw} =f(a) + x{f(a + w) -/()} + (x)z^/(
4
2
2
A
+
+ (x+ l) 3 {A /(a)
(x+ l) 4 A /(
/( -w)}
'

+ (x + 2) 5 {A /(a - w) 4

Using the relation

may
*

(p

l) q+1

A 4/(a - 2w)} +
=

(p) q+1

(p) q ,

2io)

this

equation

be written
J.I. A. 35, p. 452 (1901). Tables of the coformula have been published in Tracts for Computers, No. V.

Brit. Assoc. Rep. (1900), p. 648

efficients iu this

CENTBAL-DIFFEEENCE FOBMULAE
+ xw) =

41

+ xf(a + w} + (x + l) 3 A /() - (z) 3 A 2/(a w)

4
4
(x + 2) 5 A /(
w) -(x + l) 5 A /( - 2w) +
Introducing central differences and rearranging the terms,

/(

(1

x)f(a)

f(a + xw}

(1

x}f(a}

+ xf(a + w} +

(x) 3 S*f(a)

(x

we now transform the

If

relation

!-& =

so that

-(x + l
+ w) +

l) 3 8 f(a

(x

by means of the
(x+l) 5 = -( + 2) 5>

coefficients of f(a)

(x) 3

= -( +

l) 3

we have

etc.,

f(a + xw}

which
26.

+ l}^f(a} + + 2)^f(a) +.
f(a} +
+ xf(a + w) + (x + l) 3 8*f(a + w) + (x + 2) 5 34/(rt + w) +.
(

is Everett's

formula for equal* intervals of the argument.

The follow-

Example of Central-Difference Formulae.

ing example illustrates the application of the various centraldifference formulae

:

To compute the value of log cosh 0'3655,

w
having given a
of loglo cosh x at intervals 0*002 of the argument.

table of values

Forming the difference table, we see that the differences of the third
order are approximately constant.
The differences of the fourth order
will, however, be taken into account since such a difference may affect
the accuracy of the last figure of the result.
Argument.

Entry.

0-360

0-0275 5462 3980

0-362

278 5523 7805

30061 3825
152 8035

30214 1860
281 5737 9665

0-364

-2122

-2135

30366 7773
284 6104 7438

0-366

-3

1523778

-2138

30519 1551
0-368

-13

1525913

287 6623 8989

152 1640

30671 3191
0-370

290 7295 2180

= and =0-0281 5737 9665.

formula put = f,
=
5737
/(0-3655)
1(281
9665) + (- T -^)(152 5913) + g^(- 13)
+ f(284 6104 7438) + (-^5 (152 3778) +
(- 3)
= 283 8513 0494-75 - 14 2937-59 - 0-13 = 283 8498 7557-03.
In

Everett's

o;

|-,

^^

.-.

log cosh (0-3655)

= 0-0283

8498 7557.

Corresponding formulae for unequal intervals have been given by R. Todhunter, J.I. A. 50 (1916), p. 137, and by G. J. Lidstone, Proc. Edin. Math. Soc.
40 (1922), p. '2Q.
(D

309)

CENTKAL-DIFFEBENCE FOKMULAE

42

In the Newton-Bessd fornuila put x =

/(0-3655)

.-.

=$

HI HI]

\.

+ 1(30366

7 773)

= 2830921 3551-5 + 7591 6943-25 = 283 8498 7557-02.

cosh (0-3655) = 0-0283 8498 7557.

log

By

142954-27

16-68

- 0-14

the Newton-Gauss formula

/(0-3655)

= 281

5737 9665

+ f(30366

= 281

5737 9665

+ 2277

= 283

8498 7556-95.

7773)

+ (- -A)

(152 5913)

50829-75 - 14 3054-34

+ 116-76-0-22
.-.

log cosh (0-3655)

By

= 0-0283

8498 7557.

the Newton-Stirling formula

/(0-3655)

= 281

5737 9665 +

= 281

5737 9665

= 283

8498 7556-94.

+ ^(152 5913)

f-

+ 2 2717

8612-38

+ 42

9163-03

+
.-.

log cosh (0-3655)

= 0-0283

116-40

8498 7557.

The Formulae of the preceding Sections may be

more concisely by means of the Central-

27.

expressed

Difference Notation of
Everett's formula

20.

(x

The Newton-Bessel formula

ux =

,
,

(X

+ -13

(x

+r-

l} Zr

Zr

ui +

(x

a;

+ r-

The Newton-Gauss formula

ux = w + a;8w, + (x) 2 8 2 + (a? + l^S 3 *^ +

_i

^?'

i(,

(a;

J.

CENTKAL-DIFFEKENCE FOKMULAE
The Newton- Stirling formula

43

4!

|{(a

+ r)*. +

(a?

+r-

2r

I) 2r }8

u +

Newton-Gauss backward formula

u_ x = U - x8u.i +
S2
Q

28.

(z) 2

to

find

(x

We

The Lozenge Diagram.

which enables us

a large

now give a method

number of formulae of
shall

interpolation.

Let

(p)q

pi

denote the quantity

entry f(a + rw}.

We

and

let

ur denote the

obtain at once the relations

(i)
(2)

and, combining these equations,

. r}

we

{(p

see that

or
_r

(3)

"
"
Suppose we arrange these terms in the form of a lozenge so
that the terms on the left-hand side of the equation lie along

the two upper sides of the lozenge and the terms of the right-

'q-H

AA
q+i

FIG. 2.

CENTRAL-DIFFEBENCE FORMULAE

44

We obtain

hand side along the lower sides.

in which a line directed from left

the above diagram

to right joining

two quantities

denotes the addition of those quantities.

Equation (3) may be expressed by the statement that

in

travelling from the left-hand vertex to the right-hand vertex of

the lozenge in the diagram, the sum of the elements which lie

along the upper route is equal

along the lower route.
It

is

to the

sum

evident that this statement

let

example,

of the elements which

may

be extended.

lie

For

us place in contiguity the lozenges correspond-

ing to

p = n-

upper vertices of the lozenges, which are of the form

form a sort of difference table

so that the
_ r,

(n

We

1)^

obtain the following diagram

(n-i),

Au,
FIG. 3.

Applying the rule given by equation

the

sum

the same

(3), it is

evident that

of the elements along either of the following routes is

:

(n

l) 2

A 2 t_ 1 +

(w

CENTRAL-DIFFERENCE FORMULAE
Since

MO +

AM = w1 +(w-l) 1Aw

(w) 1

we may form

45
three

other expressions beginning with the term u { instead of u and

equivalent to those already given, namely,
Wj

+(n-

l)]Aw +

(?z.) 2

A 2 w_ 1 +

(n

l) 3 h?u_ 1

and two similar expressions.

If we examine the structure of this diagram, it will be seen
that the values of q and r in the expression (p) q&?u_ r are
arranged in precisely the same way as for the differences
A 9/(w - no) in an ordinary difference table. The values of p are
constant along any diagonal descending from left to right of
the diagram, while along a diagonal ascending from left to right

The first value

these values increase by unity at each vertex.
of p along either line radiating from U Q is taken to be p = n.

By

extending this diagram we arrive at the following, which

FIG. 4.

CENTKAL-DIFEEBENCE FOKMULAE

46

be called a lozenge or " Fraser" diagram since

modification of one due to D. C. Eraser.*

may

Now

un

the Gregory-Newton formula for

is

it

sum

the

is

of the

elements from UQ along the downward sloping line to the line

So un = the sum of the elements from u
of zero differences.
along any route whatever to the line of zero differences.
Erom the identity u + n&u = u 1 + (n - 1) Aw it is evident
that the value of

un

is

unaltered

if

a route

selected starting

is

In general the sum of the elements

along any route proceeding from any entry u r whatever to the
from u^ instead of from u

line of zero differences is equal to

Applying

this rule,

we have

= uo + n\& u _ i + fa +

un

at once from the lozenge diagram

A2

A3w _ 3
+ + 3) 4A4 u_ 4 +
= o+ OOiAtt.j. + (n + l) 2 A 2 tt_! + (n + 1) 3 A 3 M_ Z
+ (w + 2)4A4 it_ 2 +
2
a
4
=
A
W
u
n
un
+ 2
_2 +
_i + (+l )sA - 1 + (n + 1 4A
)iA
o +
2
3
4
u n = Ui + (n- 1)^0+ () 2A tt + (w) 3 A M_ ! + + l) 4A -w_ 1 +
(

1 )2

-2

2) 3

(?i

it

Rewriting equations

un = UQ +

(rij-fiu

_$

1 ) 2S

+ (n +

1)

u_j

+ (n+

2) 4 S

(4)

(5)
(6)

in the central-difference notation,

(5), (6)

+ (n+

(7)

we
.

find
.

and
7t n

which
If

is

we

+ (TO^SMJ + (n) 2 3 2 M + (n +

l) 3 8

3
j

+ (w +

the Newton-Gauss formula.

now take the mean of these values of

whose differences are along the row corresponding

un = u o + ( n)i ^'u - * + <^) +

-3-

{ (

1)

1)4 S

we obtain the formula

to u
:

the Newton-Stirling formula.

of u n from equations (6), (7) may be expressed either
as Everett's formula or as the Newton-Bessel formula.
Writing (6), (7)

which

is

The mean value

in the central-difference notation,

un = U

+
+

(n

(71

+r-

!) 2r S^tt

+ (n + r)2r+1S2^ +1MJ +

+ r- IVSa^ + n + r (

Taking the arithmetic mean

* J.I.A.

l) 2 , +1 S

of these values of

43 (1909),

p. 238.

2 '- +1
j

(8)

(9)

u n we may eliminate
,

CENTRAL-DIFFERENCE FOEMULAE

47

by applying the relations (p) q = (p+ 1),7 + 1 (p) q+l

The coefficient of S 2 '"^ takes the form
- l") 2r + (n + r) 2r+1 + (n + r- l) 2r+1 } or (n + r) 2 , +1
The co-

differences of odd order

2 '- +1 (t.

and

+r

(n

or

of

efficient

= S 2r M 1 -8 2r M

- (n + r -

& Zr u Q
l) 2r +i,

becomes 1 { (n

+r-

- (n + r) 2r+1 -(n + rfor (1 - n) we see that

l) 2r

and by substituting

The arithmetic mean

of equations

(8), (9)

may

l) 2 r+i}

thus be written in the form

is Everett's formula.
mean of the values of u n
Suppose, however, we find the arithmetic
in (8) and (9) and simplify the coefficients of differences of odd order in
the resulting expression by means of the relation

which

We now

obtain the result

Wl ( ff)

"I

-M

J[

(n

n
which

is

the Newton-Bessel formula.

Relative Accuracy of Central-Difference Formulae.

necessary to use approximate formulae which

29.

It is frequently
terminate before the column of zero differences is reached. From
the last section we have seen that the sums of the elements along
any two routes which terminate at the same vertex are identical.

two adjacent vertices (p) q ^u_ r+1 and

in
the
are
same "lozenge," the sums of the
which
&iu_
r
(p) q
elements along these routes differ by (^) (A%_ r+ i - A^_ r), i.e.
by (p) 3 A9+1 w_ r Extending this result to routes terminating in
4
the same column of differences, for example, at A%_ 3 and A w
it is evident that the sums of the elements along these routes
If the routes terminate at

(y

differ

(n + 2) 4 A
shall now

by

We

-w

_3

(n

l) 4

A 5 w_ 2 + (n)^u_ v

consider routes

that

lie

along horizontal

these yield the formulae containing mean-differences. In

the last section it was shown that a mean-difference formula is

lines

obtained by taking the arithmetic mean of the elements along

two adjacent routes. From the mode of formation we see that
the sums of the elements along such routes are identical as far
For example,
as the vertices at the intersections of the routes.
the
the Newton-Gauss formula is equivalent to
Newton-Stirling

CENTRAL-DIFFERENCE FORMULAE

48

formula as far as differences of even order, and it is also

equivalent to the Newton-Bessel formula as far as differences
When a formula is curtailed, the question arises
of odd order.
as to

whether

it is

terminates at a

more advantageous to select a route which

difference or at an ordinary difference.

mean

The following diagram represents the portion of the lozenge

diagram along the row corresponding to U Q and adjacent to

the differences of order 2r. Let

denote the mean difference (n + r)z r+1 Zr + l u and let
denote the mean difference
[j!.8

(n

+r-

2
l) 2r /xS 'wi.

/>

(n+r)r

(n+r)'ar 1

2r

(n+r+i)2r+2

FIG. 5.

The route along the dotted line through A represents the

Newton-Stirling formula and the route along the dotted line
through B represents the Neivton-Bessel formula. The NewtonGauss formula, which is represented in the diagram by a zigzag
intermediate route,
vertices 8 2ru

and

is

equivalent to the Stirling formula at the

and it is also equivalent to the Bessel

8 2r+2 u

formula at the vertices

S 2r

-%

and

S 2r+1 %i.

Consider the three routes representing the Gauss and the

Stirling formulae and the formula which contains the differences
S 2 '- 1

^!

S 2 'V

S2

^ ^,
1

+zu

If we suppose these
Q
formulae to be curtailed so that the last difference of each is of
order 2r + 1, we may compare the accuracy of these formulae
,

and

by ascertaining the magnitude of the neglected terms of order

(2r + 2). The sum of the elements along either of the routes from
the

common

vertex

8 2r+2 u

to the line of zero differences being

CENTRAL-DIFFERENCE FOEMULAE

49

the same, the most accurate formula

is the one in which the

term
of
order
is
the
+
smallest.
These terms are
neglected
(2r
2)
+2 u
Q
%{(n + r) 2, +2 + (n + r + l) 2 , +2

''

respectively,

magnitude.

and they are also arranged in ascending order of

The Newton-Gauss formula is therefore more

accurate as far as

mean

differences

(2r+l), when

of order

further terms are neglected, than the corresponding NewtonStirling formula passing through the same differences of even

and both are more accurate than the formula containing

S 2r+1 w_.
In precisely the same way we see
that the Bessel formula is more accurate than the Gauss
order

the difference

formula as far as differences of even order when further terms

are neglected.
In general, a central -difference formula terminatat
a
mean
ing
difference of the entry up is more accurate than
a formula which is curtailed at the corresponding centraldifference of
is

up _i, and

it is less

accurate than a

curtailed at the corresponding difference of

formula which

up+ i*

We

shall now illustrate by an example the superiority which centraldifference formulae generally have over other interpolation formulae.
Let it be required to find u m where
If we employ for

^x<|.

purpose an

interpolation formula which proceeds according to

central differences of
and stop at the (2r+l)th term, the result is

this

the same as

u -rt u -r+i,

if

we employed Lagrange's formula with given values

-1

un

so tna t

19 the error

by

of

is

i),fr

where
denotes some number between a
rw and a + rw. If, on the
other hand, we employ the Gregory-Newton formula, and stop at the
(2r+l)th term, the result we thereby obtain is the same as if we
u zn so
employed Lagrange's formula with given values of UQ u^
.,
that the error is
,

afo-1)
where

77

denotes some

(a;

2r

number between a and a+2rw,

Now

does not, in most cases, differ greatly from / (2r+ll (^), but (x + r)(x + r - 1)
much smaller than x(x 1)
(x
r) is
(x
2r) in absolute
value when
Thus the error is smaller in the former case than
.

<x<.

A detailed discussion of the accuracy of interpolation formulae is given in

papers by W. F. Sheppard, Proc. Lond. Math. Soc. 4 (1906), p. 320, and 10 (1911),
p. 139 ; D. C. Fraser, J.I.A. 50, pp. 25-27 ; G. J. Lidstone, Trans. Fac. Act. 9 (1923).

CENTBAL-DIFFEBENCE FOEMULAE

50

For

in the latter.

this reason central-difference formulae are prefer-

able to the ordinary formulae for advancing differences.

The following remarks* are of general application
" Formulas which
to constant differences are
:

proceed

exact,

and are

true for all values of n whether integral or fractional.

" Formulas which
stop short of constant differences are approximations.
"
formulas which terminate with the same difference are

Approximate

identically equal.

"

Approximate formulas which terminate with distinct differences of

The difference between, them is
the same order are not identical.
expressed by the chain of lines necessary to complete the circuit."
30.

Preliminary

formulae of

In

Transformations.

interpolation

should

not be

preliminary transformation has been effected.

trate this
Ex.

I.-

e.

10"

cases

"We shall

illus-

by two examples.
Suppose that

it is

required to find

a table of logarithms the following entries

0'

certain

used until some

sin 15".

We

have from

CENTRAL-DIFFERENCE FORMULAE
It

best to interpolate in the sequence of numerators

is

1, r,

and

51

r(r+l),

r(r

l)(r

2),

to interpolate separately in tlie sequence of

denominators

We

then divide the former result by the latter, in order to obtain the
required interpolated value.
Stirling (Methodus Differentialis (1730), Prop. xvii. Scholium) says:
" As in common
algebra the whole art of the analyst does not consist
in the resolution of the equations bxit in bringing the problems thereto
;

dexterity required in the

performance of the process of interpolation than in the preliminary
determination of the sequences which are best fitted for interpolation."

so

likewise

in

this

analysis

there

is

less

The general rule is to make such transformations as will

interpolation as simple as possible.

make

the

EXAMPLES ON CHAPTEK III

Given

1.

sin 25
sin

find the value of sin

= 0-433
= 0-433
20" = 0-433
40" = 0-433

41' 40"

25 42'

571
659
746
833

711
084
453
818

655
587
442
219

565
544
359
189

25 42' 10" by the Newton-Gauss formula.

Find the value of log

2.

0"

log sin

16' 8"-5

sin

having given

= 7-670 999 750

8" = 7-671 448 629 9
9" = 7-671 897046 4
10" = 7-672 345 0002

16' 7"

using the Newton-Gauss formula.

16' 8"-5 from the following
Check your result by obtaining log sin
data
16' 6" = 7-670 550 405 5
log sin
8"= 7-671 448 629 9
:

10" = 7-672 345 000 2

12" = 7-673 239 524 3
3. Apply the Newton-Stirling formula to compute
from the table of values

25 40' 0"
20"
40"
sin 25 41' 0"
20"

sin

= 0-433134785866963
= 0-433222 179172439
= 0-433309568404859
= 0-433396953563401
= 0-433484334647243

and verify your answer, using the Newton-Bessel formula.

sin

25 40' 30"

CENTRAL-DIFFERENCE FORMULAE

52
4.

Given
log

310 = 2-4913617
320 = 2-5051500
330 = 2-5185139
340 = 2-5314789
350 = 2-5440680
360 = 2-5563025

3375 by the Newton-Bessel formula, verifying the

by one or more other central -difference formulae and comparing

find the value of log

result
it

with the true value.

5.

Show

that the

[3-5282738.]

which

is

method really derives all the

summation by parts, i.e. by the use

lozenge-diagram

interpolation formulae by repeated

of the formulae

the analogue in the Calculus of Differences of the formula

fudv
in the Integral Calculus.

= uv

fvdu

CHAPTER

IV

APPLICATIONS OF DIFFERENCE FORMULAE

An important application of interSubtabulation.

the
extension of tables of a function.
formulae
is
to
polation
we
Thus, supposing
already possess a table giving sin x at
31.

we might wish

to construct a table giving

This operation is called subSubtabulation might evidently be performed by
tabulation.
of the new values by ordinary interpolation,
each
calculating
but when the new values are required in this wholesale fashion

intervals of

1'

of x,

sin x at intervals of 10" of

x.

better to proceed otherwise, forming first the differences of

of values of the function, and then calculating
the latter from those differences.*
it is

the

new sequence

T T p T 2 T3 ...

be a given sequence of entries in a

table corresponding to intervals w of the argument, and let
2
their successive differences be AT = Tj - T A T = T 2 - 2T t + T

Let

etc.

Suppose

it is

desired to find the values of the function

in question at intervals w/m of the argument so that (m - 1)

intermediate values are to be interpolated between every two

consecutive

members

km = ^2, km = T 3

etc.,

new sequence be
AJQ = tj
where

is

of the set

sequences thus required by

and

tp,

0>

2,

T 1; T 2

so that

.,

Denote the

=T

m = T 1;

let the successive differences in the

Aj

used instead of

'o

~
2

new sequence.
now
be found in terms
sequence may
differencing in the

"

~^~

o>

kc.,

denote the operation of

The differences in the new

to

of the differences in the

old sequence by the use of operators in the following way.

*

Lagrange, (Euvres,

53

5,

p.

663 (1792-3).

54

APPLICATIONS OF DIFFERENCE FORMULAE

Denoting the initial value t or T by

Gregory-Newton interpolation formula

f(a),

we have by

the

\ and A are thus connected by the relation

and the operators

A! = (1/m^A +

for simplicity that

Suppose
of

the

If

+ (l/m) 8 A +

is

(1)

the last non-zero difference

A5 T = 0, A 6 T = 0,

that

etc.

(1) gives

V - {(1/m^A +

we now

A4 T
so

sequence,

original

Equation

we

(l/m) 2 A

(l/7/i) 2

A2 +

substitute the values

s
(2)
(l/m),A + (l/m) 4A*}
= 1,2,3,4 in the last equation,
.

are able to determine all the differences of the

in terms of the differences of the old sequence

new sequence

A 1Au = i

in

(l-m)(l - 2m) (1 3m)

~*~

lo

'

(5)

(6)

When
may

the differences are thus calculated, the entries

be derived in the usual

values of

sm
with the tabulated values
Ex.

The

zm

logs of the

way by

formed in

T1? T 2 T 3
,

1}

No.

z , t3

numbers 1500, 1510, 1520, 1530, 1540 being

given to nine places of decimals, to find the logs of the integers between

and 1510.
The difference

The
simple addition.
this way should agree

table of the original values

is as

follows

A3

A*.

249

-4
245

1500

APPLICATIONS OF DIFFEKENCE FOEMULAE

Here
.-.

Aj

m= 10.
_(

~10 3

4)=

0-0004 in the ninth

'

2-10*

19047)+

place,

which

is

negligible,

55

APPLICATIONS OF DIFFEEENCE FOEMULAE

56

An

It is frequently convenient when

32.
Alternative Derivation.
dealing with a function whose degree is known to insert values of
the function, intermediate to those already tabulated, by the following

method

Suppose, for example, that a function f(x) may be represented by a

polynomial of the third degree, and that values of the function are
tabulated at intervals w = 10 of the argument. Let it be required to insert
values at an interval w=l.
Using the notation of the last section, we

have (by the Gregory-Newton formula)

T =

Ti =

4=

T 2= 'ao =
T 3 = 30 =
^

+ 30A^ + 435 A!% + 4060A 1 %.

Differencing these equations,

we

see that

AT =
,,

Similarly

+ 145^% + 1 020A a %,

AT 2 = 10A1* +245A1%
A 2 T = 1 00 A x % +
A 2 T 1= lOOA^ +
A 3 T =1000A 1 %.
.-.

The

leading term and

seen to be

differences

its

for the subdivided intervals are

T
Aj% = -OlA T - -009A
A^ =-lAT --045A 2 T
3

from which the values

Ex.

w=5,
the

t-^

z,

s,

are formed

by addition.

Having given a

table of values of log x at intervals of the argument

to insert between log 6250 and log 6255 the intermediate values of

function at intervals

w=

1.

A.

Entry.

Put

A2

T = log 6250 = 3-7958800

3473

T^ log 6255 = 3-7962273

-3
3470

T 2 = log 6260 = 3-7965743

-2
3468

T 3 = log 6265 = 3-7969211

differences of the second order are approximately constant, so we
assume log x to be a polynomial of the second degree.

The

T =f = 3-7958800,

AT =5A 1 +10A 1 %=3473,

f

A 2 T =25A 1%=

-3.

These are precisely the set of equations of

31 when A 3 T
differences of the tabulated function, are assumed to be constant.

the third

APPLICATIONS OF DIFFEEENCE FOEMULAE

From

these equations

we

57

obtain the values

A 1 %=-0-12,

A^

expressed in units of the seventh decimal place.

Forming the difference table for the subdivided intervals,

Ar

Entry.

log 62

50 = 37958800-00

A]

2
.

694-84
log

62 51

= 37959494-84

-0-12
694-72

log 62 52

= 37960189-56

-0-12
694-60

log 62 53

= 37960884-16

-0-12
694-48

log 62 54

= 37961578-64

0-12

694-36
log

62 55

= 37962273-00

We may now insert these values of the function in the table of values, thus
6251 = 3-7959495
6252 = 3-7960190,

log
log

etc.

We may obtain without difficulty formulae for subtabulation based

on central-difference formulae, or on Everett's formula.
These are
frequently to be preferred to the subtabulation formulae based on the
Gregory-Newton formula.

Owing

to the rapid

differences, care

accumulation of error in the higher orders of

must be taken

to include additional places of digits in

the computations, as in the above examples.

Estimation of Population for Individual Ages when

Populations are given in Age Groups. We shall now find
33.

the values of a statistical quantity, such as the population of

a given district, for individual years, when the sums of its

values for quinquennial periods are given.*

Let
uv u 2
be the values of the
w_ 2 u_ v u
.,
.

quantity for individual years, and let the quinquennial sums

be
so that
W_j,
Wj,
.

.,

"Wj

=u7

i(

+u 5

It is required to find the value

+u

it

* G.
King, J.I. A. 43, p. 109 (1909).

(D

309)

+ u3

in terms of the
See also 50,

Ws.

p. 32.

APPLICATIONS OF DIFFEEENCE FOEMULAE

58

The Newton-Stirling formula may be written

n(n 2 - 1) A 3 w _ 9 +~A% _
-

If

--

we denote u + u_ n by y n and neglect the

we may write
tl

differences of

the fourth and higher orders,

Therefore

and

=u +y +
= 5u Q

+ W_ =

2
Eliminating ^. u_ 1 from the two last equations, UQ
expressed in terms of the

Ws

writing

or,

A2 W_ X

for

(W _ 1 - 2W + W 1 we

be

obtain the result

),

w = 0-2W -0-008A2 W_ 1

or

may

(1)

To find the value of the quantity for the middle year of the second
quinquennium, when the following are three consecutive quinquennial
Ex.

sums: 36556: 39387: 41921.

Denote the given quinquennial sums by
and form a difference table.

W_j = 36

W_

x,

W W
,

respectively,

556

W =39387
W = 41 921

2831

-297
2534

The required quantity u

is

therefore,

= 0-2 x 39 387 = 7877-4 + 2-4

= 7879-8,

by

-008

(1),
(

- 297)

=7880.

so

The above formula may be extended to include the fourth

Ws when we neglect the differences of the u'a
the sixth and higher orders.* We have now

differences of the

of

When

the groups are unequal,

divided differences.

we can proceed

in a similar way, using

APPLICATIONS OF DIFFERENCE FOEMULAE

59

yn = un + u_ n

= 2.
.

y +

7/2

A% _

W_

= 10w + SlOA 2

Eliminating U Q from the three

(2)

2,

+ 377A%_ 2
+ 4627A%_ 2
,

last equations,

we have

AWj - AW_ 2 = S^BA^.j + 4250A%_ 2

and

and eKminating

A 2 w_j

from these two equations we find that

A 2 ^.! = 0-008A 2 W_ 1 - 0-00096A4 W_ 2

and

we now

If

substitute these values in equation

we obtain

(2),

the result

UQ

= 0-2(W - 5A 2 M_ t - A4 ,Y_ 2 ),

u = 0-2W - 0-OOSA2 W_ 1 + 0-000896A4 W_ 2

or

This value of UQ was also given by G. King.*

The following demonstration of a more general formula

due

to

these definitions

we

is

G. J. Lidstone.
r

Let

= 2tt

s,

-r

and

= 2

us

etc.,

r+l

yx =

let

3r+l

us

where p is some number independent of

have at once

x.

From

and
In

Bessel's formula,
TO (m2
,

Form

the difference y% +m

- y*_ m
*

J.I.

- j)

an(l in the result substitute

A. 43,

p.

114.

W and

its

APPLICATIONS OF DIFFEKENCE FOEMULAE

60

differences

for

and

Ai/

its

We

differences.

thus

obtain the required

formula.

The

which,

result

when
M

is

2r + 1 = 5, becomes
=0-2W -0-008A 2 W_ 1 +

as found above.

Inverse Interpolation. We shall now consider the

process which is the inverse of direct interpolation, namely, that
34.

of finding the value of the argument corresponding to a given

value of the function intermediate between two tabulated values,

when

a difference table of the function

is

This

given.

is

known

as inverse interpolation.

Let f(a + xw) denote a particular value of the function of

which the differences are tabulated. We now wish to find the

value of the argument x corresponding to f(a + xw} for this
it is best, if - \<x<^, to use
formula *
;

purpose

Stirling's

2 2
f(a + xw} =f(a} + z|{A/(a) + A/(a w)} + f z A /(a w)
3
w) + A /(a
+
2w)
.

2io)}

(1)

Dividing throughout by |{A/() + A/(a w)}, the

of x, equation (1)

may

coefficient

be written in the form

x = tn-$3?D 1 -%x(a?-l)T) 2 - 1fea?(a?-l)n 3 -.

(2)

where ra = {/(a + xw) -/(a)}/i{A/() + A/(a - w)},

2
D, = {A /( - i*}/H4fa) + A/(a - w)},
D 2 = {A3/(a -w) + A3/( - 2^)}/{A/(a) + A/(a - w)},

and

so on.

We

have now to solve equation

(2)

by successive

approximations.
1st approximation

= m.

Substituting this value in equation (2)

approximation

x = m-

we obtain the 2nd

WDj - im(m

This value of x

is

now

3rd approximation for

* If

1)D 2

-^m (m 2 - 1)D 3
2

substituted in equation (2) to form the

and so on for further approximations.

x,

|<a;<f,

Bessel's formula should be used.

APPLICATIONS OF DIFFEEENCE FOEMULAE

Instead of solving equation
it in the form

(2)

by

successive approximations

61

we may

arrange

We
may

have merely to reverse this series to obtain a formula from which x

be found by direct substitution, namely,

A.

62

APPLICATIONS OF DIFFEEENCE FOEMULAE

From

35. The Derivatives of a Function.

Newton formula

f(a + xw)

the Gregory-

=/() +
+ X(X

X~

2}

^f(a) +

(1)

we have at once
f(a + xw)-f(a)
xw

If

a;

is

taken very small so that xw->Q, the left-hand side of

The limiting
the equation is of the form {/(a + h) -f(a)}fh.
the
when
is
derivative
of the
this
7t->0
of
value
expression
function f(x)
obtain

for

the

a of

value

its

argument.

We

thus

...}.

(2)

The successive derivatives of the function may be obtained

by the use of the differential calculus in the following way.
Differentiating (1), we obtain

wf(a
4a;

+-

- ISx2 + 22x -

-jp

Also
xw)

and so on

= A 2/(a) +

(x

fir 2

Sr

-4-

1 1

for derivatives of higher order.

Putting x =

we obtain the
+ ^f(a) - ^f(a) + ^f(a)

in this set of equations,

wf(a) = A/(a)

i-A /(rO

results

- f A4

/(a)

+ fA

APPLICATIONS OF DIFFERENCE FOEMULAE

To find

Ex.
X.

the first

and second

derivatives of

locj e

x at

x= 500.
A4

63

64

APPLICATIONS OF DIFFERENCE FORMULAE

and in general
,/<>() = (A
36.

- 1A2 + 1A 3 |A4 +

.)-/().

(5 )

The Derivatives of a Function expressed in Terms

same Horizontal Line.

of Differences which are in the

By

differentiating Stirling's formula,

f(a + X 'w]

2 2
=/() + 4{A/(ft) + A/( - w) + iz A /( - w)
2 2
3
- 2w)}
l )|{A /(a - w) + A 3
+ x(x
/(

+ T^( 2 - I 2 (* 2 + ^^x 2 (x2 - I 2 ) (x 2 )

22)

the differential coefficients may be represented

by a rapidly
converging series in terms of the horizontal differences. Thus

wf'(a

+ xw)
+ zA 2/( - 2w)
3
w) + A /(
2w)}

w)}

A 2/(

2w) +

x^ f(a - w} + A /(
3

=
Putting x

2^a)

in these equations,

2w) +

2w)}

3w)} +

we have

wf(a) = i{A/() + A/(

w*f"(a)

A 5/(

1.)}
H(A 3/( - + A 3/( + aVi(A5/( - 2w) + A 5/( - 3w)} +
= A 2/(a - w) - jf&fta - 2w) +
^tff(a -3w)+.

2w)}

')

(1)

(2)

These equations give the value of the derivatives in terms of

differences which are symmetrical as
regards the direction of
increasing and decreasing arguments.
In order to extend these results to derivatives of
higher
order we shall write Stirling's formula in the central-difference
notation of
20 as far as differences of the eighth order.

f(a + xw)

2
2
= u + X(jiSu + IX^ Z U Q + lx(x z 1)^S% + '^x (x - 1)S%
+ ^x(x 2 -l)(x 2 - 4)^/0 + r ix2 (x2 - 1) (x 2 - 4)S%
+ Wau(^ - 1) (* 2 - 4) (x2 - 9)^/

1) (z

4) (x

9)5%

APPLICATIONS OF DIFFEEENCE FOEMULAE

When

the right-hand side

powers of x, we obtain

f(a + xw]

arranged according to ascending

= u + x(fj.8u

both sides of

If

is

65

this

substitute the value x

aio/^X) + ^(rW^o - m^X)-

3)

equation are differentiated and we

0, we obtain the value of wf'(ct) as in

and the higher derivatives of /(.) are formed by

z
differentiating wf'(a + xw), w f"(a + xw), and so on.

equation

(1)

The

successive derivatives of f(a) correct to differences of

the eighth order are given by the following equations
:

We
in

see that wf'(a) is equal to the coefficient of x in

(3) and,
n n
n in the
general, iv f( \a) is equal to the coefficient of x

This result might have been

equation (3) multiplied ~by n
obtained at once by comparing (3) with Taylor's expansion
\.

+ xw}.

of/(

To express the Derivatives of a Function

37.

of its Divided Differences.

We

in

Terms

shall first find the derivative

of a function f(x) for the particular value a of the argument x

in terms of its divided differences.
As shown at equation (3),
we
write
13,
may

f(u,a

=/(

>i)
.

+ ( u ~ i)/(Vi>
- a2
(u
a^) (u
)

rt

2)
.

+ ( u ~ a i)( u ~ 2 )/K.i>2.3)
- n
.,),
-i)/K. a v
(u
<

where the divided differences of order beyond the wth are

supposed negligible. If we put u = a we have
,

/K>

o)

=/K'
+
+

i)

+ ( a o ~ a i)f( a o>

K -- (o -~ aa)/K
K a K a
i)

i)

a)

i.
i.

a 2-

a)
s)

K-n-i)/K' a

i>

a n ).

(1)

66

APPLICATIONS OF DIFFERENCE FOEMULAE

But

in

16 we found that

f(u) =/(

(w

)/(o>

o)

+ ( u ~ 'o) 2/K. a r a o)
+ ( u ~ o) 3/K. rt o> a o> a o) +

and by Taylor's expansion,

so that

/'() =/K'

ft

\a )/n =/(
which gives the nth derivative
ence of the

Equation

0>

ff

.,

and in general

o)>

.),

in terms of the divided differ-

th order with repeated arguments.

ft

/K) =/K

a o>

=/K>

i/"(o)

o)>

(n

(1)
i)

(a

+
-

thus becomes

K- ~
K

rt
il/( o.i.
rt

i)

ff

2)

K-

i)(-o

2)/(o.i. a 2.

K--i)/K'i>

a)

>

)>

which gives f'(a ] in terms of its successive divided differences.

As a special case of this formula when a = a + w, a 2 = a Q + 2>, etc.

which

is

the formula of

35.

more general expression

for the derivatives of a function

terms of its divided differences may be obtained from

Newton's formula
- ^/K' i- 2
-a
) (x
}f(a flj) + (x)
f(x) =/(a ) + (x
+ (x - .) (x - a x ) (a; - 2 )/(0' a v <%> c's} +
Denoting the factor (x a n ) by a n this equation becomes
in

=/(

f(x]

a /(rt

ttj)

a oai /(rt

+ ^0^2

1;

2)

<%_!/(<),

aoOjajj/fao, Oj,
rf

1;

Differentiating both sides of this equation,

we

!)

rt

3)
(3)

see that

+ (a + ^}f(a a lt a 2 )
+ (aQaj + a 2 + a^'^MS' %' a 2> ffs) +
a
+ 2)/(S' i. a z> as}
=, /K, i.
2 + ( o +
i
+ (aQaj + a a 2 + a Q a 3 + a x a 2 + a^g + a 2 a 3 )/(^ ftj, z 3
=/(

f'(x)

2,

a n ).

.,

a,

//2

+
!

=/(a

a lf
"1

and

so on.

(a

2,

a2

aa

3,

a2

a 3 )/(

1(

1>

2'

Ct

C!

2,

3,

4)

'3>

"4>

s)

4)

+ a 4)/K'

4)

(5)

APPLICATIONS OF DIFFERENCE FORMULAE

67

In these equations the coefficient of the divided differences of order r

In
a. r _
a symmetric function of the quantities a aj, 2
.,
l
differentiaequation (3) this coefficient is of r dimensions, and after each
so we see, therefore, that the
tion its dimensions decrease by unity
a r) in the equation for f(r\x)/rl is unity
ax
coefficient of /(a
.,
a
a r _ ), and all differences of lower
zero dimension in a
.,

is

(i.e.

order vanish.

we suppose

If

above

/'(a

)=/(a

3
),

/"(a

= 3=
= 2/(

= a,,, we
),

Substituting in equation (4) the value x

and

obtain the values given

so on.

we

obtain equation

(2),

namely,

The latter equation is used when the derivative of a single value of

the function is required but when the derivatives of several values of
the function are to be computed, we use equation (4).
;

Ex.

the following table of values compute the third and fourth

6 has the values 5, 14, and 23
of f(B) when the argument

From

derivatives
respectively.

13

16

21

29

57

1345

66340

402052

1118209

4287844

21242820

We

first

form a table of divided differences

e.

57

o=

2=

1345

=16

a 6 =29

21242820

APPLICATIONS OF DIFFEEENCE FOEMULAE

68

From

equation

we have

(5)

at once

+ a l + 2 + 3)45
If "(9) = 556 +
+ a O a l + aO a 2 + a O a 3 + a O a4 + a l a 2 + a l a3 + a l a 4 + a 2 a 3 +
(

so

/'"(5)=1626,

From equation

/"'(14)=12 102,

a 2a4

a 3 a 4)l|

/"(23)= 32 298.

we have

(6)

= 45 + (a + a t + a 2 + a 3 + a 4)l,
1V
IV
/ (5) = 624, / (14)=1704, / (23) = 2784.
/"(&)

IV

so

EXAMPLES ON CHAPTEK IV
1. The logs of the numbers 400, 410, 420, 430, 440 being given to
seven places of decimals, find the logs of the integers between 400
and 410.

400 =
410 =
=
log 420
=
430
log
log 440 =
log

log

If A r T is the last non-zero

Ar+1 T =0, A'' +2 T = 0,

2.

that

tabulation are

2,

difference of the original sequence, so

show that the formulae for sub.,

mr ~ z

A,-2 To +
r

The

2-6020600
2-6127839
2-6232493
2-6334685
2-6434527

-Zm?- 1

/(r-2)(l
~

differences of

- m) (1 - 2m)

(r

~~ -m)^~

2) (r

3) (1

order higher than the rth in the

new sequence

are, of course, all zero.

3.

The following

are three consecutive quinquennial

sums

44133, 41921 and 39387.

*
Mouton, an astronomer of Lyons, in 1670 noticed that if in a sequence
whose ? th differences are constant, say = c, intermediate terms are inserted
-

equal parts,
corresponding to a division of each interval of the argument into
r
its rih difference constant and equal to c/m

then the new sequence has

APPLICATIONS OF DIFFEPtENCE FOEMULAE

Find the value

of the

69

quantity for the middle year of the second

quinquennium.
4.

The populations

for four consecutive age groups are given

by the

table of values

Age Group.
25 to 29 years (inclusive)
30 to 34 years (
)
35 to 39 years (
)
40 to 44 years (
)

Population.

458572
441424
423123
402918

Estimate the populations of ages between 32 and 33 years, and between

37 and 38 years respectively.
5.

Show

that

if
\\ o

=u +u +
0/t

l/t

+u

(t

_ l)lt ,

and in general

then the individual value ullt

W W W

vidual values

x,

2)

may
.

and

be found from the groups of t inditheir differences by the formula

u rlt =
where third

differences are neglected.*

6. In the following set of data h is the height above sea-level, p the

Calculate by a difference table the height at which
barometric pressure.
=
when 7i=5280.
and
29
the
pressure
2?

2753

4763

^ = 30

27

25

7i

7.

Form

pressure in

20

a difference table from the following steam data, where

per square inch.

is

Ibs.

9C

93-0
11-38

p
Calculate

10593

6942
23

96-2
12-80

104-2
17-07

100-0
14-70

p when 0=99-l and determine by

temperature at

which p =

108-7
19-91

inverse interpolation the

1 5.

Calculate the real root of the equation

8.

x3

+x- 3=

by inverse interpolation.
* C.
tion,

H. Forsyth, Quarterly Publications of

December 1916.

the

American

Statistical Associa-

APPLICATIONS OF DIFFERENCE FORMULAE

70

differential coefficient of log e x at x

Find the

9.

= 300, given

the table

of values
log e x.

x.

5-703782474656
5-707110264749
5-710427017375
5-713732805509
5-717027701406
5-720311776607
5-723585101952
5-726847747587

300
301
302
303
304
305
306
307
Find from the above table the
10.

differential coefficient of log e x at

x= 302.

Given the values

x.

y.

858-313740
869-645772
880-975826
892-303904
903-630006

3
4
find the value of ---

095
308
766
583
875

when x = 0.

dx*
11.

Find

J
~

when z=

d?s

1.

given the following values

z.

y.

0-198669
0-295520
0-389418
0-479425
0-564642
0-644217

3
4
5

36 to compute the
Apply the central-difference formulae of
and second derivatives of log<, 304, having given the table of values

12.
first

of Ex. 9.
13. From the following data compute the first four derivatives of the
function y corresponding to the argument x = 1 1
:

x.

y.

108243219
121550628

13
15
21

141
163
174
214

158
047
900
358

164
364
628
884

INDEX
(The numerals refer

Argument,

to

pages.)

King's formula for quinquennial sums,

generalised, 59

Binomial Theorem, 15
Lagrange's formula, 28, 30
remainder term of, 32
Lozenge diagram, 43

Central-difference formulae, 35

example

of,

41

relative accuracy of,

- summary

47

Central-difference rotations, 35

Newton's formula for unequal intervals,

24
- for repeated arguments, 28

Derivatives of a function, 62

Newton-Bessel formula, 39,

of,

42

in terms of central-differences, 64

in terms of divided differences, 65

Difference tables, 2, 21, 27

Differences of a factorial, 7

42,

47,

42,

46,

49

Newton-Gauss formula,

36,

49

backward formula,

of a polynomial, 5
of zero, 6

37, 43

Newton-Stirling formula, 38, 43, 46,

49

Divided differences, 20, 65

theorems on, 22
with repeated arguments, 27

Operators, symbolic,

4,

63

Polynomial, differences

(operator),

of,

Everett's formula, 40, 42, 47

Repeated

arguments, divided

differ-

ences with, 27
Factorials, 8

Function,

Gregory-Newton

Sub tabulation, 53
formula, 10, 14,

26

Symbolic operators, 4

Unequal

Inverse Interpolation, 60

for,

King's formula for quinquennial sums,

57

intervals,

Newton's formula

24

Zero, differences

Printed and 'bound in Great Britain.

of,

S P 304860

lot)

DO NOT REMOVE
CARDS OR SLIPS FROM THIS POCKET
PLEASE

UNIVERSITY OF

QA.

281

P&ASci

TORONTO

LIBRARY

Whittaker, (Sir) Edmund

Taylor
A short course in
interpolation