[ 415 ]

XV T. Functions of Positive and Negative Type, and their Connection the

Theory o f Integral Equations.

B y J. M ercer, .A
B, Trinity , Cambridge.
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

Communicated by Prof. A. It. F o rsyth , S c.D., . , F.P.S.

Eeceived December 21, 1908—Bead May 13, 1909.

Introduction.

T he present memoir is the outcome of an attem pt to obtain the conditions under

which a given symmetric and continuous function ( s , t ) is definite, in the sense of
H i l b e r t .* A t an early stage, however, it wTas found th at the class of definite
functions was too restricted to allow the determination of necessary and sufficient
conditions in terms of the determinants of § 10. The discovery th at this could be
done for functions of positive or negative type, and the fact that almost all the
theorems which are true of definite functions are, with slight modification, true of
these, led finally to the abandonment of the original plan in favour of a discussion of
the properties of functions belonging to the wider classes.
The first part of the memoir is devoted to the definition of various terms employed,
and to the re-statement of the consequences which follow from H il b e r t ’s theorem.
In the second part, keeping the theory of quadratic forms in view, the necessary
and sufficient conditions, already alluded to, are obtained. These conditions are then
applied to obtain certain general properties of functions of positive and negative type.
P art III. is chiefly devoted to the investigation of a particular class of functions of
positive type. In addition, it includes a theorem which shows that, in general, from
each function of positive type it is possible to deduce an infinite number of others of
th at type.
Lastly, in the fourth part, it is proved th at when k ( , t ) is of positive or negative
type it may be expanded as a series of products of normal functions, and that this
series converges both absolutely and uniformly.
* ‘Gott. Nadir.’ (1904), Heft I.
(456.) 18.10.09
 416 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

P a r t I . — D e f in it io n s and D e d u c t io n s from H il b e r t ’s Theorem .

§ 1. Let k (s ,t) be a continuous symmetric function of the variables t

defined in the closed square a<s <
which are continuous in the closed interval b). W hen the function 6 ranges
through the class ©, there are three possible ways in which the double integral

k (s,t) 6 (s) dt
may behave:—
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

(i) There may be two members of ©, say and such th at

k (s, t) 0X( s)( ) ds dt, k( s,t) 02(s) 02 ds dt

have opposite signs ;

(ii) Each function 6 may be such th at

[ [ (s) 0 {t) ds 0;
J a Ja

(iii) Each function 0 may be such th at

[ [ k( t) 6 (s) 0 (t) ds dt < 0.

a Ja

This suggests a classification of continuous symmetric functions defined in the

closed square. W e shall speak of those which have the property (i) as functions of
ambiguous type, whilst the others will be said to be of positive or negative type,
according as they satisfy (ii) or (iii).
§ 2. From the point of view of integral equations this classification is of considerable
importance. H il b e r t has proved* th at

[ [ x(s,t) 6(
s) 6 (t) ds dt = tJ- [ \fjn (s) 0
Ja *a 7i= l a _

where (s), \fj2(s), (s),•••> are a complete system of normal fu

the characteristic function k (s , t) of the integral equation

f (s) = (f>(s)—X (" (£) dt,

Ja

and \ u k2, ..., \ n, ..., respectively, are the corresponding singular values. It follows
at once from this that, when the singular values are all positive, k {s, t) is of positive
* ‘ Gott. Nachr.’ (1904), pp. 69-70. See also Schmidt, ‘ Math. Ann.,’ Band 63, pp. 452, 453. We
shall refer to the result given above as H ilbert’s theorem. The theorem stated by H ilbert on p. 70 of
the paper referred to can be deduced by writing 6{s) — x (s) + y (s) in the equation written above.
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 417

type in accordance with the above definition. Conversely, we may prove that, for
every function of positive type, the above integral equation has only positive singular
values. For, if we multiply along the homogeneous equation

'I’n(s) = fK k (. s, (t)
da

by if*n ,)sa( nd integrate with respect to s between the limits and we obtain
rb rb
K/
c(s, t) x\}n(s) xfjn( ) ds dt = 1.
Ja Ja
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

Since the double integral on the left cannot be negative, and \ n is a finite number,
it appears th a t
K > o.
Thus the necessary and sufficient condition that a continuous symmetric function
should be o f positive type is that the integral equation o f the second kind o f which it is
the characteristic function should have all singular values positive* '
In a similar manner it may be proved th at this statement remains true when we
replace the word positive by evitagn,in both places where it occurs.* M
since a function must be of ambiguous type when it is of neither the positive nor the
negative type, we conclude th a t the necessary and sufficient condition fo r a continuous
symmetric function to be o f ambiguous type is the existence o f both p
negative singular values o f the integral equation o f the second hind o f which it is the
characteristic function.
§ 3. I t is easy to see that, corresponding to a function k ( , t ) whose type is
ambiguous, there exists a function 9 (s)which is not zero in the
and satisfies the relation
[ f ks,( t) 0 (s)

For, if we employ the notation of (i) above, and suppose th at k is any real constant,
we shall have

f I K(s’ 0 [0i (*) + ^ 2 (s)] [0i (0 + &02 (*)] i I k (s, t) 6X(s) (t) ds dt
J aJ a J aJ a
+kf [ k5,( t) [#! ( s)92(t) + 9^(s) 9X(i)] ds
JaJa JaJa
* It follows from these results that, unless k (s, t)si identically zero, we can
| k (s, t) 0 (s) 6 (/) ds dt = 0,

for all members of 0. We shall prove this result in a different manner further on (§ 12), but it is useful to
make the remark at this stage, since it shows conclusively. that a function which is not identically zero
cannot be both of positive and negative type.
VOL. CCIX.— A. 3 H
 418 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

The coefficient of k2 on the right has a sign opposite to th at of the term independent
of k;accordingly, when we equate the right-hand member to zero, the resulting
quadratic has its roots real. It follows that, if we suppose one of them to be a, the
function
0 (s)= 01 (+ a02

will satisfy (A), and it cannot be identically zero, because this would imply th at 01(
is a constant multiple of 02 ,)as( nd hence th a t the two integrals m
the same sign.
The converse of this theorem, however, is not true, for there are functions both of
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

the positive and of the negative type which agree in this property with those of
ambiguous type; these are known as the semfun
called definite snoitc,uf and have the property th a t (A) can only be satisfied by
function 0 (s) which is zero at each point of
The two classes of functions we have just mentioned have distinctive properties in
the theory of integral equations. For, if t) is of positive or negative type, it is
evident from H ilbert ’s theorem th at (A) can only hold when

f
da
i pn()s6 (s)ds = 0 = 1, 2, ...).

By a known theorem* we must, therefore, have

| k( s, t) 6 (t) dt =0 (a < s < b).

da

Thus the necessary and sufficient condition that a function o f positive or negative
type should be definite is that it should be perfect.

P art II. — T h e N ature of F u n c t io n s of P o s it iv e and N e g a t iv e T y p e .

§ 4. The double integral

I [ ks,( t) 9 (0 ( ) ds ■ ( 1)
Jn Jn

in which k ( s,t) is an assigned symmetric and continuous function, and 6

member of the class @, may be regarded as the limit of a certain set of quadratic
expressions. For, let au a2,..., an be points of the interval (
way th at the distances between consecutive members of the set of points consisting of
a, b and these n are all equal. Then, by the theory of double integration, and in
virtue of the symmetry of k ( s,t), (1) is precisely equal to

( b - a )2 L t[* °h) ^2(a i) + K(a 2>a a)+ ... + *(«„, a») hfiafi + 2K(au aa) Ojaffi+ ... ]

* Cf. Schmidt, op, dt., pp, 451, 452,

 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 419

The quantity inside the square brackets is evidently a particular value of a

quadratic form whose coefficients are K(a1 K(a2, a2)>
and, when 0 ranges through the class ©, the numbers (<q), 0 (a2), ..., 0 (an) will
assume all possible real values.
I t is thus suggested th at we are to look upon the double integral (l), when Granges
through ©, as the limiting case of a quadratic form whose variables assume all possible
real values. The function k (s, t) clearly
form. Moreover, when k (s , )t is of positive type, the double integral (1)
to a quadratic form which cannot take negative values for real values of the variable ;
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

and similarly in regard to the case when k (s, t) is of negative type.

Now the question, whether a quadra,tic form does, or does not, take both signs, as
the variables assume all real values, has been shown to depend on the signs of certain
determinants whose elements are coefficients of the form.# The considerations we
have ju st indicated seem, therefore, to point to the existence of properties of the
function k (s ,t) which will decide its type, without directly considering the integral
(1). I t is the object of the present section to show that this is actually the case.
§ 5. Let us, for the present, confine our attention to a function k ( s , t) of positive
tvpe, so th at
"r rb cb
k (s , t) 6(
JaJa

for all functions 0 belonging to ©

W e shall, in the first place, define a particular class of the functions ©. Let q be
any point of the open interval ( a,b), and suppose th at e and rj are a
numbers which are so small th at the points q±(i7 + e) also belong to the interval.
Then the continuous function which is zero for and q + 77 + e < s < b,
which is equal to unity for q — jr< s<
intervals ( q - ^ - e , q —77), (q + ??, q + 7?+ e), will be denoted by q). The values
of the function in these latter intervals will be given by

S—(sl—rj —e) (q + 77+ e)—

€ €

respectively, and will evidently be positive numbers less than unity at interior points.
Consider now the values of the function

Si) si)
at the various points of the square a ^ t ^ b o± the (s, t) plane. In the
accompanying figure this large square, which we shall denote by Q, is intersected by*

* See, for example, B romwich, ‘ Quadratic Forms and their Classification by means of Invariant
Factors’ (1907), chap, ii., where necessary conditions are obtained. It is not difficult to obtain conditions
which are both necessary and sufficient.
3 h 2
 420 ME. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYhE*

two sets of four lines drawn parallel to the axes of s and these are the lines
t = s1±rj, t = s1±(r) + e); s = s ^ r j, s = s1±(r) +
resp
identified by observing th at the number at the point where any one of them
intersects an axis is the value of the corresponding variable which is constant along
it. It will thus be seen th at the square denoted by is bounded by the four lines
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

A X I S OF S

s = s1±r), t — Si + 7) ; while the area dn,w hich is shade

be referred to as the border of qn, is the part of the square bounded by (77+ e),
t — s1±(rj + e) exterior to qn. A little reflection will show that, at points of Q which
do not belong either to qu or to dn, one or other of the functions 0e>r,(s s^, 6e ri{ t ; s)
is zero ; that, at points of qn, each of these functions is unity ; and, finally, that in du
neither function exceeds unity. It follows then th at

Sx) 0e,„(£; sx) = 1 in qn,

< 1 in
= 0 elsewhere.
§ 6. The integral
ff
Ja Ja
k (s,t ) 0etV(s ; sx) ; ds dt
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 421

may be looked upon as j k (s, t) 0e>v(s ; sx) 0(<r)( t ;taken over Q, or, as it
usually w ritten,#
I * s;( t ) 0e>r)(s;
JQ
and, from what has been said in the preceding paragraph, th a t portion of the latter
which arises from the part of Q exterior to d n is zero, while th at arising from qn is
simply
k ( s , t ) dt).
*hn
W e have, therefore,
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

( f k (s , t ) 0 e>ri(s] Sx)0e>ri( t ; s j d s d t
•J a J a

= f K (s,t) (ds dt) + f

J qn J du

Again the total area of dn is 4e(2i7 + e), and so, it M is the maximum value of
| k ( s,t) | in Q, we have

f k (s, t) 0etV(s ; 5X) 0e>r)(;t (ds dt) < 4e (2^ + e

J <£u
also the remaining integral on the right-hand side of (2) can be replaced by

*si-v s*-7!
which is evidently equal to
( k (sx + u, Sx + v) du dv.
J — y] J — yj

Thus it follows from (2) th at

f f/c (s, t ) 0t>n(s; s1) 0et71(t; s ^ d s d t - ^ P ( s x + u , S x + v) < 4e(27?+ e)M. . (3)

Ja Ja J —ri J —y)

Now let us suppose it possible for k(si, tSo have a*negat

because k (s, t) is continuous, we can choose a value of so small that
k (sx + it, Sx + v) < —
for all values of u and v whose moduli are not greater than rj. We shall therefore
have
[V f1?
— k (sx + u, Sx + v) d.u dv > 2i72a.
J * - 7] - 7]

Recalling our hypothesis that k (s , t) is of positive type, it follows from this and
(3) th at
rfa 2e (2^ + e) M,
C
*f. H obson, ‘ The Theory of Functions of a Real V ariable’ (1907), p. 416.
 422 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

for all values of e which are less than a certain positive number (§ 5). But this is
evidently impossible, because, when e tends to zero, the right-hand side tends to zero,
and we arrive at the contradiction th at a fixed positive quantity (viz., 7 is less
than, or equal to, zero.
W e conclude th at k (sx, sx)cannot be negative when sx lies
(a, ;)b and hence, since k (s, s )is continuous in the same inte
closed, we have the result th at every function k ( 5 , t) 'which is o f positive type in the
square a < s < b, ^a t<6 satisfies the inequality

k (sl5 Si) 5 : 0* < < b).

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

§ 7. This is a first condition which must be satisfied by these functions, and we may
obtain a second on similar lines. Let sx and be any two distinct points of the open
interval ( a,b), and, as before, let e and 77 be two positive numbers ; the latter will no
be supposed so small th at the intervals [sx—(77 + 6), sx+ (77+ e)], [s2—(17+ e), s2+ (1?+ e)]
are both contained within (a, b) and do not overlap. We now propose to consider the
values of the function

O A ,, (s ; $1) + x f i ^ ;s( s2)] [xxf v ; + ;

at points interior to Q, when xx and x 2 are any real constants. For this purpose we may
make use of a diagram (fig. 2) which is an obvious extension of the one employed in
the previous paragraph. The square Q is divided in this case not by eight, but by sixteen
lines, viz., those whose equations are s = sa± 77, .9 —sa± (
(a, /3 = 1, 2). By giving a and (3 all possible values in the equations just written, it
will be seen th at we obtain four sets of eight, for each of which we can distinguish a
square q# bounded by the lines s = sa ± 77, t = s^ 7±7 ; mor
evidently have borders dag of width e. I t is not difficult to see that, in those parts
of Q which are exterior to the borders daf} (a1, 2),

0e,v(s
or
^e,;r(t 5 ,<?l) = ^e,r, 5 ^2) = ^ 5 I

th a t in the square qaP we have

sa) = sp) = 1,

0e,1) (s ; s3_a) = 9^ (t ; s3_p) = 0 ;

* The reader may compare this with the fact that, when we have a quadratic form which only assumes
non-negative values, and we put all the variables save one (say xx) equal to zero, we deduce that the
coefficient of aq2 must be >: 0.
f Both these pairs of equalities will hold in certain parts of the square, but we only require that at
least one of them should be true.
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 423

and th a t in the border daf}the last pair of equations still hold, but 0ttr) ;
are each less than, or equal to, unity. From this it appears th at the function

lx Av(S ; *0 + x20 ^ (s; s2)]( t ;

= 0 outside the borders daf},
and th a t in the border d afi its modulus is < j x aXp | .
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

§ 8. Let us now write

0 (s)= x x9 ^

for the sake of brevity. I t follows from the remarks of the preceding paragraph th at
rb rb 2 2 r
k ( s , t) 0k(s) 9 (t) 'ds dt = Xx aXp k (s, t) d
*a J a a-1 ^= 1

+ a=i ^=i2 uf af} . W

Now the area of each of the borders dafi is 4e and so we have

2 2
t % f k (s, t) 9 (s)0 (
a = l /S—1 *J1d ~
 42 4 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE

moreover, it is easily proved that, in virtue of the symmetry of k

2 2 r
2 2 x ax p ( , t) (< dt)
a = l /3= 1
^a/3
J <7 «

can be w ritten as

[x ^ ksj( + w, Sj + v) + 2ayr2/c(s1+ w, s2+ v) + a;2V (s2+ ^, s2+v)] dudv

-17 ' -17?

From this and the equation (4) we finally obtain the inequality

If P P F2(w, <
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

k (s , t) 0(s) 0(t)dsdt- du)v

' Cl * Cl J — yj J —77

where F2(u, v) is the integrand of (5).

The function F 2(u,v) is, of course, dependent on the real constants x x an
us suppose it possible to choose them in such a way th at

F 2(0, 0) = xxk (su sx)+ 2 ( s

takes a negative value, say —a.Owing to the fact th at k ( ,t

then clear th at we can chose 17 so small th at

F 2{u, v) <

for | u|< r),| v|< 77. From this we deduce the inequality

yfa < 2e (217 + e) ( | |+ | | )2M,

as in the corresponding place in § 6 ; and hence, as this is impossible for sufficiently

small values of e, it follows that, when and s2 lie in the open interval and
and x 2 are real, F2 (0, 0) is not negative. Accordingly, since k ( , t) is continuous, it is
easily seen th at every function k (s ,t) which is o f positive
a ^ t ^ b is such tha,when x ax nd x2 are any rea

fa < < b\
xxk (sx, sx)+ 2xxx2ks(1? s2)-t- x 2 k (s2, s2) >: 0
\a < s2< b) ‘

§ 9. The reader will now be prepared for a general theorem of which those already
considered are particular cases. After having been through the latter in detail it will
be sufficient to sketch the general proof.
Take any n distinct point sx,s2,..., sn in the open interval (
and y are so small th at the intervals [sa—(17+ e), sa+ (^?+ e)] (« = 1, 2,..., form a
non-overlapping set contained within (a, b). Now let

0{s) = t x a6 ^ ( s \ ),
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 425

where x u x 2, . . . , x na re any real constants; and consider the values of the function
0 (s) 0 (t)in Q. I t will be seen on consideration th at in this general case Q must be
regarded as divided by 8 nl
ines, and th at there are 2 squares each
border da/3 (a, (3 =1, 2 ,..., n).I t will also be seen th at

0 (s) 0 (t) = x ax pin qafi 1 , 2 n),

= 0 outside the borders

and th at in the border dafi we have

| 0 (s) 0 (t )|< | x ax p| .
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

Proceeding then as in the case n —2, we obtain the inequality

f f k (s, t)6 (s)0 (t)ds d t — f I F„ du dv < 4e + e) ) M,

J —r\ J —7] \a = l /

where
F„ ( u , v ) = x ? k (sl + u, sx+ v) +x £ k
+ (<§!-f + v') + ... ;

and hence we establish th a t Fre (0, 0) is always >: 0. Eventually we obtain the
general theorem :—
Every function k ( s,t) which is o f positive type in the square a < .9 < b, < 6
must be such tah, when sx, s 2, ..., sn awe any points o f the
have
X \ K (§i, Si) + X 2 K ( S 2,S2) + ... + X n2K ( , Sre) + (<9x, 2) 4" ...

fo r all real values o f x u x 2, ..., x n.

§ 10. In accordance with the notation employed by F uedholm, let

k ( s u S i ) K ( S i , S 2) K ( S i , S n)
S-y, S 2, Sn
K
^*2? • • ' J S]l K ( S ,2 Si) K (<92, ^2) (s2, 6 n)

K(Sn, Si) K S2) ... K(sn,Sn)

Then, by the theory of quadratic forms, it is known that, in virtue of the inequality
which has just been obtained, we must have#

k ( Su S‘2, s n\ 0.....
Vk, ^2, V
* Vide Bromwich, ‘ Quadratic Forms and their Classification by means of Invariant Factors’ (1906),
pp. 19, 20.
VOL. CCIX.—A. 3 I
 426 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

and this is true independently of the number of points, su and their situation
in the interval ( a, b).
Conversely, by an appeal to the theory of integral equations, we may prove th at
any continuous symmetric, function k (s,t)defined in Q
is of positive type. For it will be remembered that, according to F r e d h o l m ’s
theory,# the singular values of the equation

f ( s ) = <j>( s ) - \
Ja
f K t) (t) (0
are the zeros of the integral function
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

d (x) = i - x f k (#„«,) d s ^ v r r K h> y ds^ds,...

Ja A I J a da \^1> ^ 2 /
rb rb
+ (-V k ( Sl’s 2' ’ ' ‘ ‘ ’ Sn) ds ds2 ...
1$i, s2, ..., snl

Applying our hypothesis th at (6) holds for all values of s1} s2, ..., it appears th at
the coefficient of ^ — din the series on the right cannot be nega
H il b e r t has proved th at every continuous symmetric function has its singular values
all real. It follows, therefore, that, if Xr is any one of the zeros of D (X), we shall
have

k (sx, s x) dsx+ If| k

O I Ja da da \^1? ’->2? ^3/
dsx ..

2 rb rb b rb
^1) ^2? • • • 5 S2r
1+ *' ' -• ] dsx ds2... ds2n
2! \Si ,s 2J 2 >a a \^1? ^2? • • • ) ^2n

where the series in the square brackets on the left is not negative and that on the
right is positive; and hence, that \ r must be positive. Since we have seen that, for
k ( s ,t ) to be of positive type, it is sufficient th at all the singular values of (7) should be
positive, we may now state the following theorem :—
In order that a continuous symmetric function k (s , t) defined in the square
a< s< b, <
a t < b may be o f positive
functions
/ Si, S2\ ..,s n\
•K («1, Si), K \Si,S2/ ’ ” K\Si, s2, •. s j ( 8)

should never take negative values when the variables s2, ..., sn each range over
the closed interval (a, b).
Vide ‘ Acta Mathematical XXVII (1903).
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 427

I t may be remarked that, as a corollary of this theorem, we have the notable fact
th at, if any continuous symmetric function is such th at the integrals

K(«., S.) ds„ f f k h ’ s: ) dSl ds2, P f kf ds, ds2... ds„ ...

Jet Ja ^2/ Ja Ja Ja \^1? ^2? •••? &n/

are none of them negative, then the functions (8) have the same property.
§ 11. The properties of the determinants (8) may be used to obtain some idea of the
nature of functions of positive type. Let us suppose, in the first place, th at there is
a point (cq, cq) belonging to Q at which one of these functions k vanishes. The
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

determinant k ^ (^ \ evidently reduces to —[#c (s, eq)]2; hence, because it can never
be negative,
k( s,cq) = kcq,
( s) = 0.

In other words, if we draw the square Q and the diagonal the existence of a
point (cq, cq) on this diagonal at which k ( s , )t vanishes in
vanishes everywhere on the lines drawn through this point parallel to the axes of s
and t. In particular, we deduce from this th a t function k (s, t) is o f positive
type, and is not zero everywhere Q
in , cannot vanish everywhere on the
More generally, let us suppose th at there are points cq, cq, ..., an of the interval
(cq h) such th at
dl.j CC2y •••) CLn
d \ y d 2y CLn • • •?
(0

By considering the determinant whose elements are the first minors of the four
elements belonging to the first two rows and columns of

S5 di, ( 10 )
Sy CLi? 0^2,
we obtain the equation4

Sy d \ y 0/2) ,, dn aa, ^3? •••i \

Sy d \ y d 2) *p dn g 2, a 3, .. ., a j

. /^5 ^2) •' af d d2p

1, Sy a 2 d
' \s, a2, anj d \ , do 1••? CLn

Becalling th at the first term on the right vanishes in virtue of our hypothesis, and
th at neither of the terms in the product on the left can be negative, it is clear that
we have
s, cq,
cq, cq, ••• 5 a}

Vide S cott and Mathews, ‘ Theory of Determinants’ (1904), p. G2.

3 I 2
 428 ME. J. MERCER: FUNCTIONS OF POSITIVE ANI) NEGATIVE TYPE,

at each point of the interval {a, b); and it can be proved

remainder of the functions

•K••5 Ctr 1-, S, CCr + i, Cl’n

^'27 •••? Ctr—
i, O',., . •«,

have the same property.

Again, because the determinant (9) and the functions ( ! l) all vanish, it is easily
seen th at the function (10) vanishes identically. Accordingly, if any one of the
functions (8) vanishes for all values of the variables, so must all those which follow it.
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

It appears, therefore, that, when k ( s ,t ) is of positive ty

integral equation (7) is either an infinite power series in X whose coefficients are
alternately positive and negative numbers, or else it is a polynomial whose coefficients
obey the same law.
Another property which is worth noticing is that, if L is the upper limit of the
function k (.s , s)in the interval ( a,

—L < k (s, t) S. L

in the whole of the square Q. This follows immediately from the fact that, since

we have
k (.9, s) k ( t, > [_k £ ) ] 2.

§ 12. W e have so far confined ourselves to the consideration of functions of positive

type, but the reader will easily perceive th at the results obtained for these functions
may be made applicable to those of negative type by a simple device. In fact, if
k (s, t) is of negative type in the square Q, and we suppose th at

K' (s, t)= —ks( t),

it is evident th at k.s,
( t)is of positive type in Q. Ap
about functions of positive type to k ( s , )t , we may deduc
k (s, t); o
f r instance, the necessary and sufficient condition a continuous symmetric
function k (s ,t) defined in the square a ^
that the functions

should never be negative when the variables 6*x, . sn, ... each range over the closed
interval ( a,b).
W e may remark th at this result and that of § 10 prove the classes of functions of
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 429

positive and negative types to be mutually exclusive, save for the trivial case when
k (.s1, )t vanishes everywhere. For, if k(s, t) belongs to both

k (su s1,)^
0 5x) > 0

for all points of the interval (a, b ) ;and hence k (su must be
this interval. I t follows, then, from a remark made in § 11, th at k (s, is zero in the
whole square Q.

P art III.— C e r t a i n F u n c t io n s of P o s it iv e T y p e .
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

§ 13. In the present section we propose to investigate certain species of functions

which are of positive type. The remark made at the end of the previous section
(§ 12) will make it plain th at there is no loss in thus limiting ourselves, since the
corresponding results for functions of negative type may be at once deduced by the
device there explained.
Let us again consider the square Q of the (s, t) plane which is bounded by the lines
s= a, s = b, t = a, t = b ;a nd let us suppose th at it is divided into two triangles by
the diagonal whose equation is s — t.The most direct method of defining
symmetric function in Q is, evidently, to define a continuous function in one of the
triangles, say th at in which s< t ;a n
remaining portion of the square by defining its value at a point for which s > to be
th at at its image by reflection in the diagonal. For example, if is a continuous
function of s in the interval (a,b), and we define k ( , t) to be equal
triangle s< £, then the continuation of this function into the triangle s > is
evidently 6 (t).
The theorem of § 10 may be applied to the function we have just defined, and hence
the condition th at it should be of positive type deduced. Instead of doing this,
however, we shall consider the more general function*

K(
.9, t)=
= <f>(s)0 (t)

where 6 (s) and <j>(s) are both continuous in the interval (a, b). It will be remembered
th at functions of this kind occur as G r e e n ’s functions of certain linear differential
equations of the second order, and th at it is therefore of some interest to know when
they are of positive type. Accordingly we shall seek necessary and sufficient
conditions which will ensure th at this is so.
§ 14. In the first place, let us suppose th at and (,v) are any continuous
functions whatever ; and let t be the set of points belonging to (a, b) at which
neither of them vanish. This set will evidently be dense in itself in virtue of the
* Cf. Bateman , ‘ Messenger of Mathematics,’ New Series, 1907, p. 93.
 430 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

continuity of the functions ; but it cannot be closed, unless it contains every point of
the interval. Moreover, it can be proved th at a and /3, its lower and upper limits
respectively, do not belong to the set, unless they coincide with the end points of the
interval.
A t each point of the set X the quotient

will have a definite value, because (f>(.s) is never zero. We may therefore define a
single-valued function f(«), whose domain is and whose value at any point is th at
of this quotient. I t will appear in the sequel th at the properties of k (s, depend
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

very largely on the nature of f (s), and accordingly, in anticipation of this, we shall
Speak of it as the discriminator o f k(

continuous in its domain, hut it will never have the value zero.
§ 15. Let us now suppose th at k (s , t) is of positive type, and is not zero everywhere
in the square Q. We have proved (§ 11) that, under these circumstances, the function
k (sx, which in the present case is simply 0 (f> cannot be zero in the whole
of (a, ;)b also, at points where it does not vanish, we know th at k (sx, $i ) is positive
(§§ 6, 10). It follows that, for a function of positive type, the set X certainly exists,
and th at in it the discriminator only takes positive values.
Again, when s1 and s2 are any two points of X, and > we have

* & 3 = [* W * (*)]*/(*•) [ / ( * ) - / ( * • ) ] ;

hence, since f (sx)is a positive number, it follows by the theorem of § 10 th at

/(* i) - / ( * i)-
This result may be combined with the previous one in the statement th at the
discriminator of k ($, t ) is a non-decreasing function whose values are all positive.
We have next to consider the points of ( , b) at which one or both of the functions
0 (s), </>(s) vanish. These fall naturally into three sets, according as they belong to
(l) the closed interval ( a,a), (2) the closed interval (/3, or (3)
(a, /3). As regards (l), it is not difficult to show th at 0 (s)vanishes in the whole
interval. For, if ax is any point of ( a,a), one at l
must be zero ; and hence, since k (au ax) is zero, the function k ax) is zero at each
point of ( a,)b§( 11).
Now when s > axwe have
k (s,ax) = </>(s),

and, at points of X, <j> (s)does not vanish ; we must therefore have 0

be proved in a similar manner th at (f) ( s)vanishes eve
Finally, we can show that, at points of the open interval (a, /3) which do not belong
to X, both 6 {s) and <f>(s) vanish. In fact, if ax is any one of these points, there are
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 431

clearly points of t both on its right and on its left. The argument we have ju st
employed will then establish that, by reason of the former, (ay) is zero, and that,
by reason of the latter, <£ (cq) is zero.
§ 16. Conversely, let us suppose th at kis defined in
functions 9 ( s ) , (j> (s)which have the properties mentioned in the preceding paragrap
and let us consider the function
Su s2, .,sn
sl> s2,... , sn

where su s2,. .., sn

a re variables each confined to the interval ( , b). We m
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

th at, as this function is symmetric, it will take all possible values in the domain
Si < s2< s3< ... < sn.T hus, since we are only concerned with the sign of the func
we may always suppose the variables to satisfy these inequalities. Firstly, let us
suppose th a t one of the variables has a value not belonging to the domain of the
discriminator of k ( s,t). If such a value belongs to (a, a), the point sx must eviden
lie in this in te rv al; hence, since

k (s15 sr) = 0(Si) <j>{sr)(r = 1, 2, ... ,

and 9(sx) vanishes by our hypothesis, it is evident th a t all the elements of the first
row of (12) are zero. In a similar manner it may be proved that, when one of the
variables has a value belonging to the interval (6, /3), all the elements of the last row
vanish. Again, if one of the variables, say sm, has a value belonging to the open
interval (a, /3), but not to we shall have

0 ( O = <£ (sm) = 0

by our hypothesis. It is thus easily seen th at the elements of the mth row of (12) all
vanish. Summing up our results so far, we conclude th at the function (12) can only
take values different from zero when the variables sx, ..., are each confined to
the set 2.
§ 17. Let us next consider the case when the variables are restricted in this manner.
The function (12), when expressed in terms of the functions 6 and <£, is

9 (sx) (f)(sx), 6(
sx) (f>(s2),..., (sx)
0 (sx) (j>(s2), 9 (s2) <j)(s2), ...,
9 (sx) <f> (s3),9 ( s2)(f>(s3), ..., 9 (s3) (j) (

9{ s l) 4>(sn), 9 ( s2) <f>(sn),...,

 432 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

hence, by dividing through both the r th row and the r th column of this determinant
by ()<
■r — L 2, n ),its value is seen to be

/ ( * l). /(*l)> •••» /(«■)

/ O ) . / ( * » ) . •••. / G )

/ ( » i). /( * > ) . /W
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

/(* ). /(* ). .... /(* .)

The determinant ju st written can he evaluated without difficulty, and thus we find
th a t (12) is

W («.)*(«.) •• ^(s«)]2/'(si) [/(*>)-/(«l)] [/(*»)-/(*»)] ••• [/(s.)-/(s»-i)]-

Now, according to our hypothesis, /*(s1) is positive and each of the factors
[ / ( S n ) —J (s n-i)~] is positive or zero. I t follows, then, th at (12) cannot take negative
values when the variables are each restricted to the set 2. Taking this in conjunction
with what was said in the previous paragraph, we see th at the functions

can never take negative values, when the variables su s2, ..., ... each range over
the interval ( a, b),and hence, by the theorem of § 10, th at
W e may, therefore, state our results in the following theorem :—
If 0 (s)and )s( are each continuous functions the interval ( ,
necessary and sufficient conditions that the function

k (s, t)=
= </>00 0 W (s - 1),
should be o f 'positive type are (1) that the discriminator o f the function should be
positive and non-decreasing in its domain %, and (2) if
and upper limits o f 2, 0 (s)should be zero in the interval a), <
interval (/3, b), and both 6 (s)and <f>(s) zero at points o f the open interr
do not belong to %.
As a corollary of this, by supposing th at (s) 1 < < b), the reader may
deduce the corresponding conditions for the function defined in § 13.
§ 18. Let us now investigate under what circumstances a function which
satisfies the conditions stated in the enunciation of the theorem of § 17, is definite.
If the domain of its discriminator is not dense everywhere, it will be possible to find
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 433

an interval (c, cl),lying within ( a , b ) , such th a t at each of i

0 (sx) <}>(sx) is zero. W e shall, therefore, have (§ 11)
k (s, t) = 0(c < .s- < d, « < £< h)

= 0 ( c :< £ < d, a < ,s*< b);

in particular, k ($, t ) will vanish everywhere in the square < s c < £ < d.
Now, if x ( s) any continuous function of s defined in the interval b), which is
zero in the intervals a < s < c, lc<
we shall have
f f f
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

I k («, *) X (s) X ( 0 dt = K(.9, £) x (.9) x (£) cfe

da da Jc d c

= 0,

by the properties of y (s) and k (,9, £). I t follows from this that, if k (s, £) is definite,
the domain of its discriminator must be dense everywhere in (a, b).
Again, let us suppose th a t the discriminator of k (.9, t) has a constant value
throughout a certain interval (c, d). I t will then be seen th at within the square
c < .9 < d , c < t < d
K (s, t) =p<j>(s) 4> ;

and hence, if y (s)is defined as before, th at

rb rb T rd
K (s ,t)x {s)x {t)ds(s)
•Ja J a LJ c _

It. may be proved without difficulty th at there exists a function y (s) which is not
everywhere zero, and is such th at

f <Ms) x ( sM 5 = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . (13)
Jc

For, let Xi (s)and X2( sh) e any two functions which are not m
another, and which satisfy the conditions imposed on y ($)• Then, if either of the
integrals
r d r d
<t>(s)Xi{s)ds, (s) \ 2(s) ds
JC Jc

is zero, we shall have an obvious solution of (13). O11 the other hand, if their
respective values /xx, be different from zero, it is easily seen that

X(S) = X ^ ) _ X ^ )

satisfies (13); and, in virtue of our hypothesis, y 18 not zero everywhere in (a, b).
W e conclude, therefore, th at we can always find a function which is such th at

( f
Ja Jfl
K(
.9, £) y (.9 ) y (£) ds dt = 0.

VOL. CCIX.— 3 K
 434 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

I t thus appears th at the discriminator of a definite function of positive type cannot

be constant throughout any interval.
§ 19. Conversely, we may show th at every function of positive type, whose
discriminator (1) has a domain which is dense everywhere in ( , b), and (2) has not a
constant value in the whole of any interval, is definite. For, if this were not so, we
would be able to find a continuous function (s) other than zero, such th at

[ k (s, t) xjj (t) dt = 0<?>).*

Ja

Supplying in the value of ( s,t), this equation may be writ

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

s) I <j) (t) xp (t) dt = 0 <

a( . .

Now, as xf/ )s(is continuous, and is not zero everywhere, we can find an inter
(c, d) of (a, h) within which it does not vanish; also, as the domain of the discrimi­
nator is dense everywhere, it will be possible to find a point, and, therefore, a whole
interval (y, 8), belonging both to (c, d) and the domain. The interval (y, 8) will thus
be such th at in it the functions i jf (s ), 6 (s), (J) (s do not
this interval the function of s
| </>(t) (t) dt . (15)

has a derivative which does not vanish; and hence, by a well-known theorem of the
differential calculus, that this function cannot be zero more than once in (y, 8). It is,
therefore, evident th at by contracting (y, 8) sufficiently we can ensure for it the
additional property th at (15) vanishes at no point belonging to it.
Returning now to the equation (14), and supposing th at s is confined to the
interval (y, 8), we see th at

f ( s ) = —| 0 (t) xp(t) dt ^I j

Hence, since both the numerator and the denominator on the right are
differentiable, and the latter does not vanish in (y, 8), the function
differentiable in this interval. In fact, by applying the ordinary rules, we obtain

/'(« ) = o (r - 5 - §)•
But this is impossible, because by our hypothesis f ( s ) cannot be constant in any
interval. VFe conclude, therefore, th at k ( s , t) is a definite function.
§ 20. I t may be remarked th at the conditions (l) and (2) of the preceding
paragraph may be stated in another and more convenient form. For, if a discrimi-

* Vide§ 3.
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 435

nator satisfying these conditions had the same value at two distinct points, it would
necessarily have th a t value at all points of its domain which lie between them (§ 15).
Thus, since the condition (l) and the continuity of 0 (s), <f>(s) assure us of an interval
of the domain which lies between these points, the condition (2) would be violated.
Hence a discriminator of this kind must be a steadily increasing function; and,
conversely, a steadily increasing discriminator satisfies (2). W e may, therefore,
combine the 'results of the two preceding paragraphs in the theoreiq :—
Thenecessary and sufficient noitdc , that a function k tfi sat
requirements ofthe theorem of § 17, should he , is that its discriminator
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

should he a steadily increasing function whose domain is dense everywhere in ( , h).

As an application of this theorem we may consider the function*

k (s,t) — ( s —a)( ) ( < t),

— {t—a ) ( h —s)(s>t).

The discriminator has the open interval ( a,h) for its domain, and its value at any
point is
(s-a )/(h -s),

which steadily increases with s. I t follows from § 17 and the theorem just stated
th at k (s, t) is a definite function of positive type.
§21. Leaving the particular class of functions with which we have been dealing,
let us now suppose th at k (s, t)is any function of positive type defined
a< s< h,a < £ :< 6. Let cq, a2, ..., am be any m points of the interval ( , h) which
are such th a t
. ^ 1) tq, ‘••5
\0q, a2, •.., am
Then the function
,scq, a2, ..., am \/K( ai, (h, •••, «»
f tq, a2i ^mji V/Uj ^2>

will evidently be symmetric and continuous in the square a ^ s ^ h , a ^ t ^ h .

Again, when the function

K1
/ ^1? ^2) •••5 ^?i? ^2? *•*5
\Si, s2, ..., s„, cq, ct2, ..., am

is expressed as a determinant, it is easy to see th at the minor obtained by suppressing

all but the ith of the first n rows and all hut the of the first n columns is

^ tq, Cq, •••» \

K \S„ cq, a2> ..., amJ

* This is the generalised form of H ilbert ’s classical function, vide ‘Gott. N acho/ p. 227 (1904).
3 K 2
 436 ME. J. MEECER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

The determinant of n rows and columns, whose elements are these minors, will
therefore he
^ {^ \ i •••> /'*lj **2? •••) '^"n\
\a u cq, ..., a j \ ..., s j ‘

But, by the theory of determinants, we also know th at it is equal to#

/ cq, cq, .. •? \ 71 1 K (si ..• cq, .. •? ^m\

1-- ?
\8i, ..’•> Cq, •• •? d'mj

Thus, equating these two values, we find

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

>^2? •••? ^ 1? ^2? •••? ^1? •••? •••, cq

\Si, S2,••., \Sl5 6*2, ..., Cq, ..., ..., cq91/ • ( 16)

Now, in virtue of our hypothesis th at k (s, t) is of positive type, it follows from § 10

th at the quotient on the right-hand side of this equation has a denominator which is
positive and a numerator which is not negative. Hence we have

hh’ S» M 2: 0 ;
Wl, #2) •••>

and thus, as this is true for all values of n ,the theorem

o f positive type.
§ 22. In the light of this result, it appears th at each function of positive type can
be used to generate an infinite series of such functions. We might, therefore, expect
to obtain other species of functions of positive type by taking k t) to be of the kind
considered in §§ 14- 20.
For simplicity, let us consider the function

h (s, t) = k ^ j k (cq, cq),

where
k (cq, cq) = 0 (cq) cf) (cq) ^ 0.

Confining our attention to the triangle s < it will he seen that the variables
6* and t can be related to the constant <q by either of the inequalities :—

(i) s<
(ii) s ^ cq <
(iii) cq — s <

The reader may find it convenient to refer to the accompanying diagram, in which

* Fide Scott and Mathews , op. pp. 67, 68.

 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 437

the square sa< b, ta< bi

s drawn, and the portion o
corresponds to each set of inequalities is marked with its number.

(a,#) (&,$)
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

By expressing h (s, t) in terms of the functions 0 and <j>, it is easily seen th at at each
point of the region (i)
h (s, t) = (f) (s) <P(0 0 (0
(cti) (ax)_
th at in (ii) h (s,t) is everywhere zero, and th at in (iii)

h (s, 0(0 _ (s)

t) — 0(<%) <j>(t)
_6 (ax) <j) (a1)_
In a similar way, or hy a mere interchange of the variables s and t, the values of
h (.s, t) in the corresponding divisions of the triangle s > t can be obtained.
Now, let 0X(s),be continuous functions defined by
0i (s) = <f>(»i) 0 (a < .s < b );
</>M 0M (a < s < cq),
4>i (0
4>(«i) 0(«i)
0 («i < s < 6);
also let (5), 4>z (s) be two others defined by
02(s) =0 (a < 6' < a x),
_ 0 (0 __ 4>(Q («i < 6' < 6) ;
(ax) <£ (oh)
4>2(s) = 0 (ai)4>(s) (a < s < 6);
and, finally, let two functions hr (.s, £) (r = 1, 2) be defined in the square < 5 :f~ 6,
a < t < l> by
/C (*, 0 = (s) <l>r (t) (s < «),
= 4>r (0 0r (0 O'
 438 ME. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

On comparing these latter functions with h (s, it will be seen th at we have

hi (*, t) = h(
s, t)
=0 elsewhere;
and
h2 (s,t) = h (s,
= 0 elsewhere.

It follows from this th at we have

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

h ( s , t) = hi (s, + (s, t)

at each point of the square in which these functions are defined. But it is easily seen
that, as k ( s , t) is of positive type, the functions lir t) satisfy the requirements of
the theorem enunciated in § 17. Thus h t) is merely the sum of two functions of
the same nature as k (s, t), and hence, as it is obvious a priori th at the sum of any
number of functions of positive type is a function of positive type, it appears th at we
do not in this way obtain any new species of these functions.
The reader may convince himself in a similar manner th at the same conclusion
holds in regard to the more general function considered in the preceding paragraph.
§ 23. Although the result of § 21 proves to be so barren in this respect, it may be
applied to obtain an interesting property of the symmetrical minors of the deter­
minant of the integral equation

f ( s ) = (f> (s)-\\ K(s,t) d t , ...

J n

when k (s, t) is of positive type. Adopting the notation and hypothesis of the
paragraph referred to, let A (A.) be the determinant of the above integral equation
when h (.<?, t) replaces k ( s, t). Then, since

A(X) = 1 —X j y (s„ Sl) d h + £ k *’) dSl ds2 - . . .

+ V V j ... f I h ’ *2’ M <fo, ... ...

fl ! Ja Ja Ja yAL? ^2? • • • ? ^nj

it is easily seen from (16) that

Gja, Go (17)
A (A.) = I) A. ; • > I K l * * * ’

•3Omj j \C^i3 , ..., OmJ

where
O'i, 0. < 0 )_ x h 1, «1, a2,
D X; «1, •®
ox, .a o,x, o 2, . GmJ Ja ^ 1 ? Gj) G<2 )

and is, therefore, a symmetrical mth minor of D (X), the determinant of (7), in accord­
ance with F redholm’s definition. But, as we have shown that h (s, t) is of positive
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 439

ty p e, the function A (X) has its zeros ail real and positive. I t follows, therefore, from
(17) th a t all the zeros of the minor

JQ / ^ . Cl\,Ctg, Cf'm\
\ ? ^2? •• •> ®m/

are real and positive. Since the minor must be identically zero il

^1? ^2? •••?

^2? •••?
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

(cf § 11), we have thus proved the theorem :—

Thezeros o f all symmetrical minors o f the determinant o f an integral equati
the second ,dnik whose characteristic function is o f positive type, are all reed and
positive.
In particular, as K A(s, t),the solving function of (7), is defined by

K a(s, t) = D (X ; 5, D (X),
where

D (X; t) = K (., t) - x j \ (';* ) * . + g O f t f t Z t) ^ + ••• >

it appears that, when s = t, the solving function only vanishes for positive values of X.
P art IV.—T he E xpansion of F unctions of P ositive and N egative T ype .

§ 24. I t is to be remarked th at H ilbert and S chmidt have been able to give very
little information about the expansion of a given symmetric characteristic function in
a series of products of normal functions. H ilbert * has indeed shown incidentally
that, if the number of singular values is finite,

K (s, t) = $ ; .....(18)
n=1

and S chmidt f in his dissertation has established th at this equation remains valid
when the series on the right is uniformly convergent. The latter theorem is, of
course, much wider than the former as regards its generality; but it has the defect
th at the uniform convergence, which it postulates, is not connected with any other of
the properties of k ( s,t). In the present section we shall attem pt to remedy
some measure by proving th at the equality (18) certainly holds when t) is of
positive or negative type.

* ‘ Gott. Nachr.,’ 1904, p. 73.

f Printed with additions in ‘ Math. Ann./ Band LXIII. The theorem referred to will be found on
pp. 449, 450. From a remark made on p. 453 1 gather that it is originally due to H ilbert.
 440 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

§ 25. In the paper referred to above, S chmidt* has proved that, if k (s , t) is any
continuous symmetric function, the solution of

/ (s) = 4>(«) - ^ Jfa * ('%t) <f>(0 dt

is given by

4* (s) =f (s) +2 f f ( x ) 'Pn (x)

provided th at X is not one of the singular values Xl5 \ 2, ... , X„, ... ;t moreover, the
convergence of the series on the right is both absolute and uniform. Now, when we
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

take
/ ( S) = K (S, t),

it is known that, in virtue of one of the characteristic relations,

4> (s)= K a
I t follows, therefore, from the above expansion and the homogeneous equations

= 4*n{x) K (x, t) dx 1, 2, . . . ),
Ja

th at
K, (s, t) = K ( s , t ) + i ............ ( 1 9 >
n= 1 Aw VArc — A /

I t should be remarked th at S chmidt’s theorem only allows us to assume th at the

series on the right of (19) is uniformly convergent with respect to < < ), for
each assigned value of t ; and hence, by symmetry, that it is uniformly convergent
with respect to t<
(a t< 6), for each assigned value of s. W hen k (s
type, we may establish the uniform convergence of the series in the whole of the square
a< s< b, a< t< 6, as follows. If we write in (19), it is clear that the terms
of the series on the right become functions of s, which, with the possible exception of a
finite number, are all of the same sign as X; accordingly, by D in i ’s theorem,J this series
is uniformly convergent in the interval a<,9 < b. Bu

2 | </»* (s)xfjn

the terms of the series on the right of (19) are never greater in absolute value than
those of
1 £ k [i/C2 (s)+ rpn (Q]
2 n =l X?l(Xn—X)
* Pp. 453, 454.
t We shall always suppose this to be the case in what follows.
J D ini , “ Fondamenti per la teoria delle funzioni di variabili reali ” (Pisa, 1878), §99. See also Y oung,
“ On Monotone Sequences of Continuous Functions,” ‘ Proc. Camb. Phil. Soc.,’ vol. XIV., pp. 520-3.
 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 441

Hence, as the latter converges uniformly for a < s < < t < by what has just
been said, the result follows.
§ 26. Let us denote the sum of the first m terms of the series on the right of (19)
by Sm(X ; .v, t),and the remainder after these terms by R m(X; t). We have

SV
H(X ; s, t) — 5 ^ x^n ^ •
n= 1 X„ X n=l X„

and hence, keeping fixed,

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

L «S„(\;s, t ) = - S
A - > • oo ?i —l Ajo

L^ (X 5 s, £) *Pn(s) $n(t)
Av.

Thus, since (19) can be written

H A('<L 0 R?» (X j S, t) — K+ ibwl(X * 6\

we obtain the equations

L«[Kx(s,0-B4.(A ;*,<)]= Lt [K»(*, ,, <)] = «(»,«)- 5 (20 )

A oo A G? oo n —1 An

This relation holds for any continuous function k t),but we now add the further
limitation th at the function shall be of positive type. Then, since

R . e ;* ,* ) = s
n=m + 1 An \AW A;
we shall have
R ni (X ; 5, < 0, • • (21)

for each negative value of X.

Let us, in the next place, investigate the values of K A s) for negative values of X,
it being supposed, as above, th at k ( s , t ) is of positive type. If is any continuous
function defined in the interval (a, b) , it follows from (19) and th
the end of the preceding paragraph th at

ff
Ja J i
Kx(.s*, t) 0 ( s ) 0(t) dsdt =( [ / <■ (
*a Ja A?l yA n Aj
\pn (s) 0(s) s,
dst)

Recalling H il b e r t ’s theorem, it will be seen without difficulty th at this reduces to

( f k, («, t)e(s) e (t) ds

•a Ja n —1 K n A a ^

VOL. CCIX.— A. 3 L
 442 ME. J. MERCEE: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE,

Now, when X is negative, the terms of the series on the right must be either zero
or positive. W e conclude, therefore, th at for all functions of the class ®

lb tb.
| K a( s , t) 6 (s) 6 (t) 0 (X < 0).
• a Ja

In other words, K k (s ,t) is of positive type for these values of X. A

the theorem proved above (§§ 6, 10), we see th at

K x(s, s ) > 0 (X < 0)................................. (22)

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

§ 27. Returning to the formula (20) and writing s = t, we obtain

Lt [K„ (*, s) - R „ (X; *)] = k (s, ,s) - 5 I ' h M l .

A->-— oo n —1

Accordingly, from (21) and (22), it follows th at

k (s , s)>
5 t t M Z ......................
n~l A.n

This is true, of course, for all values of m which are sufficiently g re a t; and, further,
when we increase m we only add positive terms to the right-hand side. By a well-
known theorem of the elementary theory of series, we thus see th at

converges for each value of s in the interval (a, ; and hence, since

a = [ * .« ] * + [+40T .
th at the series
V (s) 'K (t)
n—1 Xn

converges absolutely for each pair of values of the variables satisfying the inequalities
a < s 6 , a ^ t ^ i b . From this last result it follows th at the function

/( * , ■«).= K m - s A M M ) .(24)
n —1 An

has a definite finite value when the variables are restricted in the manner ju st
mentioned. In the paragraphs which follow we shall consider the properties of
f (s, t), and eventually prove th at it is everywhere zero. I t may be remarked th at
the inequality (23) proves the relation

0 < / ( s , s) < (s,

 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 443

§ 28. I f € is any arbitrarily assigned positive quantity, it follows from the absolute
convergence of the series on the right of (24) th a t we can choose m great enough to
ensure the inequality
£ 1*/>»(s) (t) 1
?i=m +1 X. < 3 ..................................................< ">

And, when this is done, it is easily seen that, since

I (*)$» M l > | X^w(s) xf/n | (n > m)

Xn (Xn- X ) (XK< 0 ) ’
we have
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

| (X; s, t)| < | < 0).

Again, from (20), we see th at a negative number 1/ can be chosen with so great an
absolute value that, when X < I/,

k(m J
R-a. M t) (X; s, t^j
< !;
while, from (24) and (25), we deduce

II LL m - n =s\ W An (t):_ -/(* » o <!•

Adding the three inequalities ju st written, we obtain
|K x ( « .t ) - / ( M ) l < ‘ (X < L').
In other words, we have proved the theorem
L t KA( s , t) = f ( s , t ) < < 5, < £ < 6).
— co

§ 29. I t may be proved# that, if c is any constant and ax any point of the interval
(a, b), then the solving function corresponding to the characteristic function

h(s, t) = k (s , t ) - K(ai>s) * ( a ' ’ *)

is
H ,( s, 0 = K a(s,«)------ xK O ..................................... (26)
v ' v ’ c+
whilst the corresponding determinant is easily seen to be
K AM i , a x ) - K ( a u ax)
A(X) = D(X) 1 +
c
* Cf. B ateman, ‘ Messenger of Mathematics’ (1908), p. 184. The result in question follows from
equations (24), (25), and (26), by writing /(g ) = g(t) = K(ah t) and observing that

+{s) = El& iA , X(0 = Ka (ax,t), X

rn -^
3 L 2
 444 MR. J. MERCER: FUNCTION'S OF POSITIVE AND NEGATIVE TYPE

Now, if we write s = t = cqin (19) and (24), it is easy to see th at

K, K a,) = / ( « „ « , ) + s t f e M C ,
n~\ An A

and hence th at K A (a1} cq)constantly increases with X, so long as the latter i

Consequently, when e is any positive quantity, and we take c to be

k (cq, cq) —f ( a u °h) +€,

it follows from the theorem of the preceding paragraph th at

Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

K a (cq, cq) k (cq, cq) + c

can only vanish for positive values of X. Thus, as D (X) has no negative roots, h (s,
is of positive type,* and, therefore, in virtue of the remark at the end of § 27,

*L£ H a(,s, s)>: 0 < s<

— qo

Using the formula (26), it will be seen th at this becomes

f (s, s) — > o (a < < 6).

But, as e may be taken as small as we please, this is evidently impossible unless

/( « i, s) vanishes. It follows that, as cq and may each have any assigned values
belonging to (cq b),we must have

f ( s,t) — 0

We have thus shown that, in the case of a function of positive type, the series

2 VA ( s ) (t)
K=-l X„

has k ( s,t) for its sum-function. I t was shown in § 27 th at the convergence of this
series is absolute, and, by an application of D in i ’s theorem, it may be shown th at the
convergence is also uniform in the square < < Hence, xpx ( ),
\p2 (s),..., rf>n(s), ... cire a complete system o f normal functions relating to a function
k (s,t) o f positive type and Xl5 X2, ..., Xn, ... are the corresponding singular ,
then the series
< ^n(s)^n
*
n=l
X

converges both absolutely and ,uniform

ly one/ its sum-function is (s, t).

* Owing to the fact that A (X) has only positive roots.

 AND THEIR CONNECTION WITH THE THEORY OF INTEGRAL EQUATIONS. 445

§ 30. From this theorem several interesting results may be deduced. For example,
replacing k (s, t) by the series (27) in (19), we obtain

K a ( s t) ^ ^ (0 ( 28 )
n= 1 A.

where the series on the right is uniformly convergent. Again, if we write s in

(28), and integrate with respect to s between the limits a and we obtain

f K x ( s, s) ds t —..... (29)
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

•J a n= 1 A7l A

Provided th a t X is not positive, the terms of the series on the right are all positive
and less than those of the series
t■
± \- >
?z= 1 i\n

which, by writing X = 0 in (29), is seen to converge. I t thus follows that, for X < 0,
the former series is uniformly convergent. Integrating (29) between the limits 0 and X,
where the latter is negative, and recollecting F redholm’s formula

- A p o g D(X)] = £ k x(s,

it is easily seen th at
D(x) = n ( i - X ) M O ),
w=l \ AnJ

since D (0) = 1. It now follows that, as the right-hand member of this equation is
an integral function of X, we may drop the restriction X21- 0. We have thus expressed
D (X) as an infinite product.
Finally, we may remark th at if | X| is less than the least of the numbers Xl5 X2,..., Xn,...
the right-hand side of (29) may be expressed as a power series in which the coefficient
of Xw is
2 ——
\ m+ 1
■.
ix—
—1

Also, by employing N eum ann ’s expansion for KA it is easily seen that the
coefficient of Xm on the left is
[b
Km+1 (s, s) ds,
where in the usual notation
rb rb rb
Km+l(s, t) =... k (sm, i) ds1 ds2...(m > 1),
Ja Ja Ja
and
Ki (S, l ) = K (S, t).
 446 MR. J. MERCER: FUNCTIONS OF POSITIVE AND NEGATIVE TYPE.

I t follows th at
rb -j
I Km s) ds — S r—- (tm = 1,2,...).
*a n=l An

§ 31. In conclusion, it may be pointed out th at the theorem ot § 29 holds also when
k (s, t) is of negative type. This may be deduced from the theorem mentioned by
employing the usual device, or it may be proved directly by commencing with the
equation
u [ K , ( i , «)] = * (* ,* )- 5 [A M !!
A qo n= 1 A n
Downloaded from https://royalsocietypublishing.org/ on 22 November 2021

instead of th a t at the beginning of § 27, and proceeding by a method similar to th at

which we have used above.
I t may also be of interest to remark th at by a very slight modification of these
proofs we may show th a t (27) represents * (s, t) when the latter has only a finite
number of singular values of one sign, but an unrestricted number of the other.