ANNALS OF MATHEMATICS

Vo . 38, No, 1, January, 1937

ON INTERPOLATION I
QUADRATURE- AND MEAN-CONVERGENCE IN THE LAGRANGE-
INTERPOLATION1

BY P . ERDŐS AND P . TURÁN

(Rece ved March 19, 1936)

Let

(1)

be an aggregate of po nts, where for every n

(2) 1 X2(n) x (nn)
> X(1n) > > . . . > > -1.
Let f(x) be def ned n the nterva [-1, +1] . We def ne the nth Lagrange-
parabo a of f (x) w th respect to B, as the po ynom a of degree _< n - 1, wh ch
ta es at the po nts x(n)f x(n) f(x(n)) f(x (n) ) .
p 2 1 . . . f xn
(n) the va ues f(x(n))
1 2 > > n 1

We denote th s po ynom a by L.(f) and we somet mes om t to nd cate ts de-

pendence upon x and B . It s nown, 2 that
n n
(3) L,, (,f) = J (x n) ) tn) (x) _ f(x ) (x) •

The funct ons (x), ca ed the fundamenta funct ons of the nterpo at on,
are po ynom a s of degree n - 1 and f
n
(4) w(x) = wn(x) _ (x - x)
=1

then
w(x)
(5) (x) _ ,
w (x ) (x - x )

I Th s paper was part y read at the Math, and Phys . Assoc at on, Budapest, May 26,

1934 .
2
In the symbo n ) (x) the etter n s an ndex, and does not nd cate the nth d fferent a
quot ent . In the paragraphs . and 2 . as far as poss b e we sha not exp c t y denote the
dependence upon n .
142
 It s nown that f 4(x) s a po ynom a of the mth degree, then
(6) Lm+ ( 1P) = W (x) =1,2,

When 4(x) es 1, we obta n from (6) and (3)

n

(6a) E t (x) _- 1 .
=1

The f rst prob em s that of convergence, .e . f f (x) and B are g ven, we as

whether or not at any g ven x o the sequence of po ynom a s L.(f) tends to
f (xo) . Suppose f (x) to be a cont nuous funct on ; then accord ng to the we
nown theorem of Faber,' to any g ven B we may f nd a cont nuous funct on4
f (x) such that n the nterva [ -1, + 1] the L,,(f1) parabo as do not converge
un form y to f1(x) . In 1931 Bernste n5 proved that, g ven any wB,emayfnd
a cont nuous funct on f2 (x) such that the sequence L n (f2) s unbounded at a
certa n f xed ~o where -1 <_ ~o <_ +1 . The proof s. based upon the fo ow ng
theorem of Hahn :6
The necessary and suff c ent cond t on that
m L„(f)z_ xa =f( )
n- CO

for a g ven B, at a g ven xo (-1 < x o _<_ +1), for any cont nuous f(x), s that
n
(7) A(n) _ 1 y (xo) 1 < C
P-1

where C s a pos t ve constant ndependent of n. Thus t was on y necessary

to prove that for any g ven B we cou d f nd a ~o n [-1, +1] such that the
sequence (7) was unbounded, f n --* cc .
If we are to prove the d vergence not at a certa n x o, but at a countab e aggre-
gate n the nterva [-1, +1], we obta n a suff c ent cond t on n the fo ow ng
genera sat on of Hahn's theorem .' Let
n
(8) A(x, n) _ Z I v(x)
V=1

be an unbounded sequence of numbers for any f xed x(-1 5 x < +1) . In th s

case for any countab e aggregate Q, n [-1, +1] we can f nd a cont nuous

a G . Faber : Über d e nterpo ator sche Darste ung stet ger Fun t onen . Jahresber cht
der Deutschen Mathemat ervere n gung, Bd . 23 . 1914, S . 190-210.
' In the who e of th s paper the express on "cont nuous" denotes a funct on cont nuous
n the who e of the nterva [-1, +11 .
5 Bernste n : Sur a m tat on des va eurs etc . Bu . Acad . Sc . de 'URSS . 1931 . No . 8 .
1025-1050 .
6 H. Hahn : über das Interpo at onsprob em . Math . Ze tschr ft 1, 1918 . 115-142. H s
proof s based on a genera pr nc p e of Lebesgue .
7 Banach-Ste nhaus : Sur e pr nc pe de a condensat on des ungu ar tés Fundaments
Math . (1927) .
144 P . ERDŐS AND P . TURÁN

so, that Ln(f3) s unbounded for any e ement of 52 1 . G. Grünwa d8 proved

f3 (x)
that (8) ho ds for an mportant c ass of matr ces, the nth ne of wh ch s pro-
v ded by the rea roots of Jacob 's po ynom a J n (x, a, 0) . To put t rough y the
Lagrange nterpo at on ta en for any B s "bad" from the po nt of v ew of
convergence .
Let now B be def ned so that ts nth ne s g ven by the n d fferent rea roots n
(-1, +1) of the Tschebyscheff-po ynom a T n (x) (for wh ch Tn (cos t9) = cos no) .
Then we may eas y ver fy that'
"-1
Ln(f) = ao + r=1
Z ar cos rt
where
(9) 1 n 2 n / 2 - 1
ao = - f(x ), ar = - f(x ) cos r ",r,
n =1 n .1 2n
2 -1 =1,2, n.
x =cos
2n r=1,2, n-1 .

By a heur st c m t ng process we m ght obta n the Four er ser es of

f(cos t~), wh ch nd cates an nterest ng ana ogy between these spec a nter-
po at on parabo as and the Four er ser es of f(cos 4) . Th s ana ogy a so appears
n many other re at ons e .g. the form of (8) ta en for T, wh ch determ nes the
convergence, s comp ete y ana ogous to Lebesgue's constants, we nown n
the theory of Four er-ser es . Another ana ogy : We now that for any g ven
countab e aggregate 5Z2 we may f nd a cont nuous f (x) so that the part a sums
of ts Four er-ser es are un form y bounded n [0, 27r] and neverthe ess they osc -
ate at every po nt of 522 . For the Lagrange nterpo at on we proved that for
any g ven countab e 5Z3 we may f nd a cont nuous f4(x) so that ts Lagrange-
parabo as w th respect to T are un form y bounded n [-1, +1] and neverthe ess
they do not converge to f4(x) at the po nts of 523 . For the present we om t the
proof. We on y nd cate that t s based upon the we nown construct on-
pr nc p e of Lebesgue . Connected w th these facts and others that we d d not
ment on, the fo ow ng resu t s very aston sh ng . There s a cont nuous f5 (x)
such that the f rst-order ar thmet ca means of ts Lagrange parabo as at x = 0
ta en for T
+ - - -+L .(fb)o
on (0) - L (fb)o
n
are unbounded . We om t the proof ; we may obta n t w thout any d ff cu ty
from Hahn's theorem . It seems that the same s true for the ar thmet ca means
of any order ; as far as we now th s s not dec ded as yet . We do not even t
now w th certa nty whether or not the convergent Lagrange-parabo as of a

a Ora commun cat on .

9 L . Fejér, D e Abschätzung w nes Py on's etc. Math. Ze tschr ft 32, 1930, 426 .
 ON INTERPOLATION 14 5

cont nuous funct on ta en for any B at x o can converge to another va ue than

f (xo) . The remar s above show that the quest on s d ff cu t .* Neverthe ess a
certa n ana ogy between the Four er-ser es and these spec a nterpo at on
parabo as for the ar thmet ca means may be preserved n a way found by
Prof. Fejé.10r
The second group of quest ons s concerned w th the so ca ed "quadrature-
1

convergence" .e . the exam nat on of the sequence L n(f )dx, (n = 1, 2,

-1
St e tjes11 proved that f the ndBtshfgervobya tsf
Pn(x) = 0, P n(x) be ng the nth Legendre-po ynom a , then the ntegra s of the
Lagrange parabo as be ong ng to any bounded and R ntegrab e ("R ntegrab e"
means a funct on ntegrab e n R emann's sense n [-1, +1]) f(x) tend to
1

f(x)dx, or as we sha say n the fo ow ng pages : there s quadrature-

-1

convergence for th s matr x . S nce then Fejér" and Szegö13 gave a powerfu
genera sat on of th s theorem . Pó ya14 proved that we have quadrature-
convergence for cont nuous f(x) f and on y f the sequence
n
A (n) _
Z
=1
I ~ ~ )
1

rema ns be ow a bound ndependent of n ; here X n ) = 1( ' ) (x) dx, n = 1, 2,

-1

= 1, 2, n. A ( ) are the so ca ed "Cotes numbers" of the matr x . Fejér 12

proved that the pos t veness of a of the X n s suff c ent for the quadrature-
convergence. In the fo ow ng pages we exam ne nstead of the quadrature-
convergence the so ca ed mean convergence, wh ch requ res more than the pre-
v ous one. Mean convergence requ res for any bounded and R ntegrab e
funct on f (x)
1
(10) m [f(x) - L,t (f)]2 dx = 0.
n—o -1

In §1 we show very s mp y that for a genera c ass of matr ces there s mean
convergence .
The matr x-c ass n quest on s g ven n the fo ow ng theorem .
THEOREM I. Let p(x) be a funct on such that
(11a) p(x) >_M>0 -1 <x<+I

o L . Fejér : Über Interpo at on . Gőtt nger Nachr chten 1916 . 66-91.

11 St e tjes : Oeuvres Bd . I . 377-395 .
12 L . Fejér, Mechan sche Quadraturen m t pos t ven Cotesschen Zah en. Math. Ze tsehr.
37, 1933, 287-310 .
13 G. Szegő, Asymptot sche Entw c ungen der Jacob schen Po ynome, Schr ften der
Kön gsberger Ge ehrten Gese schaft, 1933 .
14 G. Pó ya : Über d e Konvergenz von Quadraturverfahren, Math. Ze tsehr. 37. 1933 . 264-

287 .
146 P. ERDŐS AND P. TURÁN

(1 1b)
f1
p(x) dx ex sts .

It s nown that there s an nf n te sequence of po ynom a s w 0(x), 1

W (x),

where the degree of w n (x) s n w th

1f 1
wn(x)W,n(x)p(x) dx -
; 0
0
f
n=m
n ;PC m '
• coeff c ent of xn n ton(x) = 1 .

As nown co n(x) has n [-1, +1 ]n d fferent rea roots . Then our re at on (10) s
true for any matr x formed of these roots . Or more genera y,
THEOREM Ia. Let con(x) be the above po ynom a s, A n and Bn constants such
that the equat on
(12) Rn(x) = Xn + . . . = con(x) + Anwn-1(x) + BnWn-2(x) = 0
may have n [-1, +1]n d fferent rea roots and B n <_ 0 ; then (10) ho ds a so for
the matr ces formed by these roots .
In § 1 we prove th s Theorem Ia . We commun cated our Theorem I to Pro-
fessor Szegö and he found for t essent a y the same proof as we d d .
The restr ct on on the roots of (12) s not very great, for t s ev dent that n
[-1, +1] there are a ways at east (n - 2) changes of s gn .
We prove Theorem Ia by prov ng the re at on

(13) m [f(x) - Ln (f)]2p(x) dx 0,

n-*oo -
wh ch w be shown to be a consequence of p(x) >_ M and of the ex stence of
p(x) dx . From (13) t fo ows by (11a) that

0 < - 1 [f(x) - Ln(f)] 2 dx < M f [f(x) - Ln(f)]2p(x) dx,

1
and th s by (13) estab shes Theorem Ia .
COROLLARY OF THEOREM Ia . For a bounded and R ntegrab e f (x) we have
for the matr ces g ven n Theorem Ia

( 14) m
f
1 f(x) - Ln (f) 1 dx = 0

and a fort or there s quadrature convergence for these matr ces .

Cons der ng on y the quadrature convergence or rather the va d ty of the
more r gorous (14) we sha prove the fo ow ng more prec se
THEOREM II . Let

(15a) p(x) >_ 0 [-1, -}-1] ;

Suppose further the ex stence of

(15b)
f1~ p(x)dx and
-I1 1
p(x)
dx.
 ON INTERPOLATION 147

If the po ynom a s (12) are formed w th the orthogona sed po ynom a s con(x) be-
ong ng to such p(x), (14) ho ds for the correspond ng matr x .
In Theorem a and II we have-as far as we now-the f rst genera theorem
for mean and quadrature-convergence .
Now we ment on some nterest ng spec a matr x-c asses . Let

-1 <a-<-<0
p(x) _ (1 - X) C , (1 -}- x) 0
-1<o<0

and A n = B n = 0. Then ev dent y (12) s sat sf ed .e . there s mean con-

vergence . The w n (x) be ong ng to th s moment-funct on are the Jacob -po y-
nom a s for the parameters a and (3 . If a = (3 = 0, we have the case of the
Legendre-po ynom a s Pn(x) ; for a = a = -2 the Tscheb scheff po ynom a s
Tn(x) .
Let now a = = 0, and A n and Bn such that (12) be sat sf ed . Fejér 12
proved the quadrature-convergence for th s matr x c ass ; as we see, here we
obta ned mean convergence. If a = 0 = 0, A n = 0, B n = - 1 we have Rn(x) _-
P n (x) - P n-2 (x) and obta n a matr x consp cuous by ts nterest ng extrema
propert es ." Now suppose
-1<a<1
- x)a (1 + x),8
-1<, <1

and An = B n = 0 and cons der on y (14), wh ch-as Theorem II shows- s sat sf ed .

Szegö 13 proved that for max (a, #) < we have quadrature-convergence but for
max (a, 0) > we have not . Thus, by the spec a case of our Theorem II, we
obta n a new proof of Szegö's theorem for -1 < a < 1, -1 < 0 < 1 ; we even
1
obta n more, for accord ng to Szegö (f (x) - Ln(f ))dx -~ 0 whereas we have
-1
1
= á we have the case of the
f 1
1 f (x) - Ln(f) I dx 0, as n --> co . For a =
Tscheb scheff-po ynom a s Un(x), where U n (cos z~) = s n(n + 1)t~/s n z~ .
If we cons der the nterpo at ng parabo as, nstead of n [ -1, + 1], on y n
[-1 + e, 1 - e], then, as here (1 - x)a (1 - - x)0 >_ M > 0, we obta n for a
a,0>-1
1-e
m [f(x) - Ln (f ) ] 2 dx = 0 .
n-- ao -1-}-e

Th s s nterest ng from the po nt of v ew of Szegö's resu t . It shows that

n the case of max (a, 0) > the d vergence s due to the rap d growth of the
parabo as on the marg n .

15 L. Fejér, Best mmung derjen gen Absc ssen etc ., Anna de a R . Scuo a Norma e
Super ors d P sa, ser e 11, Vo . I, 1932 .
14$ P. ERDŐS AND P. TURÁN

We mmed ate y obta n Theorem II from (13), v z.

dx
0 < 1 I f (x) - Ln (f ) I dx = 1 I f (x) - Ln(f ) 11/p(x)
-1 -1 1/p(x)

Í (x) _ Ln(, )]2 p(x) dx .

-1 - p(x)
and by (13) th s proves (14) .
From the above ment oned theorem of Szegö we see that there are B matr ces,
for wh ch we have no quadrature-convergence and thus a fort or no mean
convergence . The most mportant prob em n th s connect on wou d be to
g ve the necessary and suff c ent cond t on of the mean convergence . Our
Theorem III g ves a necessary cond t on for the mean convergence . It asserts
THEOREM III. If the sequence
n 1
C(n) _ E (x) 2 dx
=1 -1

s unbounded as n ---> ~, there ex sts a cont nuous fs(x) such that for our matr x

m [f6(x) - Ln(fs)] 2 dx = -}- oo .

n- o0 -

§1
As exp a ned n the ntroduct on we have to prove (13) for the fundamenta
po nts g ven by the roots of the Rn (x) po ynom a s of (12) . F rst we prove16 that

(16a) (x) p(x)dx >_ 0 = 1, 2, • • • , n

and
n 1 1
(16b) E
=1 f1 (x) 2 p(x)dx <
f1 p(x)dx.

Cons der the express on

[ (x) 2 - z(x)] p(x)dx .

But 1,(x) 2 - 1, (X) = Rn (x)F(x), where F(x) s a po ynom a of degree (n - 2),

n wh ch the coeff c ent of the h ghest term s ev dent y 1/Wn(x,) 2. Thus f
F(x) = COWO(x) + -- Wn-+ 2(x)/Wn(x ) 2 , by the orthogona ty of the CJ n (x)'s
we have
B
L z(x) 2 - (x)] p(x)dx - W n-2(x) 2 p(x)dx < 0
-1 W n (x )2 -1

s For p(x) ~ 1 the proof of (16a) s to be found n Fejér's 12 paper .

1
 ON INTERPOLATION 1 49

.e.
1 1
(16c)
f1 (x) 2 p(x)dx <_
f1 (x) p(x)dx

wh ch mmed ate y estab shes (16a) ; by summat on for = 1, 2, n we

obta n (16b) n consequence of (6a) .
Let now Q4 be an aggregate n [ -1, + 1] formed of c osed non-over app ng
nterva s . We prove that
1 1
(17)
x n) {94
2: 1:
x n)
{94
f 1
(x) (x) p(x)dx < 2 ~
X (n)
X 14
f 1
( X )2 p(x)dx .

F rst we assert that for every , w th 1 < < n, 1 <- < n

1
(18) ( - 1) + +1 I = (-1) + + (x) (x) p(x)dx > 0 f 2 .
-1
For by (5)
Rn(x)
h 1 1 Rn(x) p(x)dx .
Rn(x ) Rn(x ) -1 (x - x ) (x - x )

As ,
Rn(x)
= d0 co0(x) + + dn-3 Wn-3(x) + Wn-2(x)-
(x - x ) (x - x )
Hence cons der ng the def n t on of Rn(x) we have
1
Bn
I = Wn-2(x) 2 p(x)dx
Rn(x ) Rn(x ) -1

wh ch proves (18), as B n <_ 0 and s gn Rn (x ) Rn (x ) _ ( -1) + Thus we

have
1
(x) (x) p(x)dx
-1
x ( n ) {SZ4 x 12 4
1 1
_ (x) 2 p(x)dx ( -1) + (x) (x) p(x)dx
-1 -1
Xy 194 x~n~ {~ 4 x( n) 1 4
1 1
2 (x)2 p(x)dx - [ ( - 1) (x)]2 p(x)dx
-1 -1
x n ) 194 x n) {24
1
2 ~ (x) 2 p(x)dx7
-1
n)
x J Q4
thus (17) s proved .
150 P . ERDŐS AND P . TURÁN

If 9 4 denotes the who e of the nterva [-1, +1], n consequence of (16b) we

have
n n 1 1
(19) E=1 E=1 f-1 (x) (x) p(x)dx S 2
-1
p(x)dx .

Let now f (x) be cont nuous, 9(x) the po ynom a of degree n - 1 that g ves
the best approx mat on to t n Tscheb scheff's sense for the nterva [-1, +1] .
Wr te
o
(20) f (x) - (P(x) = 0(x)

(21) max f (x) - cp(x) = En-1

I=1~1
1
In - [f(x) - Ln(f )]2 p(x)dx.
-1
Then by (6) we have
1 1
In = [A (x) - Ln( ,á)]2 p(x)dx <_ 2 0(x)2 p(x)dx
-1 -1
(22) 1
+2 A)2 p(x)dx = In
Ln(,
-1

We have ev dent y
1
(23) I n <= 2 En-1 p(x)dx.
-1

Further by (3)
n n 1
2 E 0(x ) 0(x ) (x) 1 (x) p(x)dx
=1 -1 -1
(24) n n 1 1
:5 2
En-1 (x) 1 (x) p(x)dx <_ 4 En-1 p(x)dx
=1 =1 -1 -1

by (19) ; from (22), (23) and (24) we have

1
(25) In <_ 6En-1 p(x)dx
-1

wh ch by We erstrass' theorem estab shes (13) for any cont nuous Ax) .
Now we requ re a Lemma .
Fejér's 12 theorem asserts : f for any aggregate of po nts B the "Cotes numbers"
1
(x)dx are non-negat ve for any and n, then we have quadrature con-
-1
vergence . It may be proved n the very same way that f for a g ven matr x
the "Cotes numbers be ong ng to the non negat ve and R ntegrab e p(x)"
1
(26)
f 1
(x)p(x)dx
 ON INTERPOLATION 151

are non-negat ve for any and n and f(x) s bounded and R ntegrab e, then
1 1

m Ln(f)p(x)dx = f(x)p(x)dx.
n—o -1 -1

It wou d be superf uous to repeat the proof . We requ re th s n the proof of

our
LEMMA . Let B, be a matr x sat sfy ng (26), and Or, be a set of a f n te num-
ber of non-over app ng nterva s n [ -1, + 1 ] ; then for n > n,, we have
1

(27) (x)p(x)dx < 2 p(x)dx .

-1 fU6
2 (n) Iob

PROOF . We eas y obta n th s resu t f we cons der the funct on 4(x) hav ng
the va ue 1 for po nts of 05 and 0 e sewhere . >,(x) s ev dent y bounded and R
ntegrab e, so that accord ng to Fejér's theorem
1 1

(28)
n- 00
m
f1 Ln(,,)p(x)dx =
f 4,(x)p(x)dx .

But by the def n t on of 1,(x) we may wr te

1 1

(29)
f 1
Ln()p(x)dx =
X
(n)
(S2 s
-1
(x)p(x)dx,

further
1

(30)
f 1
~(x) p(x)dx =
S1s
p(x)dx .

(27) s an ev dent consequence of (28), (29) and (30) .

Now we cons der the matr x B def ned as n Theorem Ia . In consequence of
(16a) the Lemma s app cab e ; we obta n from (16c) and (27)
1 1

E
(31) E
y (n) [or
-1
(x)2 p(x)dx _<_
Z
t f1 (x)p(x)dx < 2
fob
p(x)dx,

and f na y from (17)

1

(32) z E
f 1
(x) (x)p(x)dx < 4 p(x)dx .
z(n){jtb
z ) (~ s

Let now f (x) be any bounded and R ntegrab e funct on. Then n v rtue
of the R emann ntegrab ty, to any Ewe can f nd a f n te aggregate of non-
over app ng open nterva s of tota ength < E such that f we exc ude these
nterva s, the osc at on of the funct on s <_ Eat any po nt of the rema n ng
aggregate U6 . We now def ne f7(x) as fo ows : 1. n U 6 et f, (x) _- f (x) . 2. f
 152 P . ERDÖS AND P . TURÁN

we denote the exc uded nterva s by (p , q 1 ) . . . ( p,, q v ), (v f n te), the funct on

f7 (x) s represented n (p , q ) by the stra ght ne connect ng the po nt (p , f(p ))
and (q , f(q )) . Thus we def ne f7 (x) for the who e of [-1, +1], and ts osc a-
t on s at any po nt _<_ e. But then f7 (x) may be un form y approx mated by a
po ynom a ~p (x) to w th n 2E. Let the degree of (p(x) be m = m(E) . Then we
have
1
I n = [f(x) - Ln(f)]2p(x)dx < 2 [f7(x) -- Ln(,f7)]2p(x)dx
-1 -1
1 1
(33) [f -f7 - Ln(f - f7)] 2 p(x)dx < 2
+ 2 fp(x)dx [f7(x) - Ln(f7)] 2
1 -1
1 1
+4 [f - f7] 2 p(x)dx + 4 Ln(f - f7) 2 p(x)dx =- Jn + Jn + J ;a',
-1 -T

say. As the degree of approx mat on to f7 (x) s 2E, we have by (25) for n > m(E)
1
(34) Jn 1 <_ 24E2
f 1
p(x)dx .

Further as f (x) - f7 (x) d ffers from 0 on y upon nterva s, of wh ch the tota

ength s -<_ E and as I f (x) - f7 (x) < 2 max I f (x) I =- 2M, we have
IXI-<1

qZ
(35) Jn <_ 16M2 p(x)dx .
'-1 fp
For J'„. we may ev dent y wr té
n n 1
J"'n - 2: E (f(x ) - f7(x ))(f(x )
=1 =1
- f7(x ))
-1
c (x) (x)p(x)dx .

In consequence of the def n t on of f7 (X) the terms of th s sum d ffer from

0 on y when x; and x e n nterva s (p , q ) and (pp , qp) respect ve y .
Hence
1
/1/ I
11 n < 4M 2 E E
z (p , 41) x UP4, 4p)
=1, 2, . . . v p=1,2, . . .
f 1
(x) (x)p(x)dx

(36) X,

Qq
<_ 16M2 ~ p(x)dx
=1 fp
by (32) .
As the tota ength of the range of ntegrat on s < E, t s ev dent by (33), (34),
' (35) and (36), that In --> 0 as n -+ oo . Hence the resu t .
 ON INTERPOLATION 153

§2
In th s paragraph we sha prove Theorem III . Let us wr te
n 1
S(n) _ ~ 1Z(x)2 dx,
=1 -1

and suppose th s to be unbounded as n ~ We sha prove that we can f nd

a cont nuous funct on f(x) w th

m sup [ f(x) - Ln (f)1 2 dx = -{- oo .

n-+ao -1

By hypothes s there ex sts an nf n te sequence n1 < n 2 < . . . w th S(n1 ) <

S(n2) G • • • --> cc . For the sa e of s mp c ty of notat on we denote by m
the mth e ement of th s sequence n

FIG . 1

Let the mth fundamenta po nts be 1 >_ . . .

> ~(2 ) > > ~ m ) _>_ - 1 .
We regard them as absc ssas and to any we adjo n an ord nate E , where
f y E2, Em have arb trar y the va ues + 1 or -1 . Thus we have m po nts ;
we connect them as n F g . 1 and obta n a cont nuous funct on 4, E (x) w th

(37a) I ~E (x) < 1 f or -1 <_ x < + 1

I

and
1 m m 1
(37b) Lm( f)2 dx = ~ ~ Eµ Ev (x) (x)dx .
- µ= v= -I

By var at on of the E's we obta n 2m d fferent 4, E (x) funct ons . For these
funct ons we have by form ng the sums of (37b)

(38)
2m
~f I
Lm (,~ E ) 2 dx =
y~ -1
v (x)2 dx = 8 (m),
 154 P . ERDŐS AND P . TURÁN

hence we may choose E's, so that for the correspond ng 4,E (x) wh ch we s mp y
denote by 4(x), we have
1
(39) > S(m) .
Lm.(t.)2 dx _-
-1

Accord ng to We erstrass, 4,(x) may be approx mated by a po ynom a fm(x)

of degree µ(m) so that
(40a) 3 -1<=x<----+1
I f. (X) I z

and
1
(40b) Lm(fm)2dx > 28(m)
.
-1

Now we se ect a part a sequence m17 m2 , . . . of the sequence f (x), f 2(x),

f f

and def ne a sequence of constants c , c 2 , n the fo ow ng way . Let fm, ( x) _

f1(x) and c = 1 . Suppose m r_1, that s f ,.,-,(x) and cr-1, a ready def ned, then
we def ne
c r-1 1
(41) cr = m n mE
4 ' maX I cmr- ) (x)
IxI<1 =1

and mr as the east nteger sat sfy ng the fo ow ng cond t ons :

(42a) Mr ? µ(mr-1) + 1
1 1
(420) C2 Lmr (fmr )2 dx - 8c r Lmr(fmr)2 dx > 4r;
-1 -1

these 2 cond t ons can ev dent y be sat sf ed n consequence of (40b) and

m.'00 S(m) = ao .
We now form w th these Cr and fmr (x) the funct on
00
(43 ) f(x) = Crfmr(x) •
r=1

We sha prove that th s s the funct on postu ated n our Theorem III .
By (41)

(44) cr <
4r

and n consequence of (44) and (40a) t s ev dent that the nf n te ser es for
f(x) un form y converges n [-1, +1] .e . f(x) s cont nuous .
Now we cons der Lm,.(f) for a f xed va ue p of r. Accord ng to (42a)
• P-1 ao
L.P(f) = Z Crfmr(x) + u cr LmP(fmr)
 ON INTERPOLATION 155

hence
1
ImP = [L ., (f ) - f 2dx
-1
(45) 00 00 2
cr Lmp(fmr) - crf.,,(x) dx .
-1 [c, Lmp(fmp) + r=p+1 r=p

But n consequence of (44)

00
(46a) E Cr fmr(x) < 37 1 + 4 + 42 + . .
r=p

and n accordance w th (40a) and (41)

op oo mP
(46b) cr L..p(fmr) < E cr . 3 E J( vm p, (x) < (1+ + . .)=2 .
r=p+1 r=p+1 v=1

From (45), (46a) and (46b)

1
(47) I.P = f [Cp L ..P(fmP) - 4 0]2 dx w th 101 =<1 .
1
Further
1
I,np > C22 Lm.P(fm,p)2 dx - 8cp
I L..p(fmp) I dx - 16
f1 f
1 1 ~
> C22 Lmp(fMP)2 dx - 8cp [2f Lm .p (fmp)2 dx - 16,
-1 1
and by (42#)

Im,p>V -16 . p=1,2,3,

Hence Theorem III s estab shed .

In conc us on, we ta e the opportun ty of express ng our deep grat tude to
Professor Fejér for h s va uab e he p .

Note added November ,27, 1936 . The prob em, whether the Lagrange-para-
bo as of a cont nuous funct on ta en on a g ven B matr x can converge to any
other va ue than the funct on tse f s, as ment oned, undec ded . Recent y
1 . Marcz n ev cz proved for the fundamenta po nts g ven by the roots of

-x- 1) 9
Un(x) = 0 U„ (cos 0) = s n (n s n 9

that f the parabo as are convergent they a ways converge to the funct on
tse f . If the fundamenta po nts are the roots of Tn(x), we proved the same

for x p ,r , (p, q) = 1, p _= q = 1 (mod 2) . On the other hand P . Erdős

q
succeeded n show ng that there ex sts a cont nuous funct on such that ts

Lagrange-parabo as ta en upon th s matr x converge to + o at x = 3 .

BUDAPEST, HUNGARY .