Вы находитесь на странице: 1из 6

J. SOC.

INDUST.

APPL.

MATH.

Vol. 12,No. 4, December,1964


Printed in U.S.A.

ON LAGRANGE-HERMITE INTERPOLATION*
J. F. TRAUBt

1. Introduction.Let the p(n + 1) numbers y m) 0 < i < n 0 < m


< p - 1, be given.It is well knownthat thereexistsa unique polynomial
Pn,p(t) ofdegreep(n + 1) - 1 such that
~~Pnm)
(xi)

(1-1)

= yi m)

O <

0 :5 m < p

i < n,

-1

A classical problemis to finda formulaforPnsp(t) in the form


n

Pn,
p( m=) yB
iP=O

p-1

z
m*(t)H

The conditionson the Cn:P(t) are that


Dt'Cmnn
(x,) =

(1.2)

6j,mar,i%

< r < n,

0 < j < p-1,

d/dtand aj,m is a Kroneckersymbol.These conditionsare used


whereDt
by Householder[5, pp. 193-195] to derivethe formulasforp = 1, 2. The
formulaforp = 3 is givenby Salzer [9]. The solutionforn = 0 is givenby
Taylor's formula.
Many authorshave reportedon the case wherep dependson i. General
fora solutionin thismoregeneralcase may be foundin Fort
prescriptions
[2, pp. 85-88], Greville[3],Hermite[4],Krylov [6, pp. 45-49], Kuntzmann
do not deter[7, pp. 167-169],and Spitzbart [12]; but these prescriptions
minethe structureofthe interpolating
ourselves
polynomial.By restricting
to the case wherep is independentof i, whichis the mostimportantcase in
practice,we can determinethe structure.Salzer [10]discoveredsome of the
propertiesof P ,, (t) by semiempiricalmeans.
We shall obtain,by a partial fractionexpansion,a solutionofsurprising
simplicity.[See (3.6), (3.7), or (3.8).] The solutiondepends upon the Bell
polynomialswhichwe now discuss.
2. Bell polynomials.Let g = q(t) and defineB. by
(2.1)

= B,(co; ga , **

eWG Dtnewg = B,(co)

qi _g9

gqn),

whichare polynomialsin gi . DeB, is a polynomialin cowithcoefficients


fineU.,k by

B.(w; gX

g.)

= E

kayO

U.,k(

g.k+l)c

* Received by the editors May 4, 1964, and in revised formJuly 2, 1964.


t Bell Telephone Laboratories, Incorporated, Murray Hill, New Jersey.

886

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions

ON LAGRANGE-HERMITE

887

INTERPOLATION

Then
Unk

!kc!DW-Bn(0).
0

The Bn( 1; 91 *
9n) were studied by Bell [1]. (See also Schlomilch
[11, p. 4].) An explicitformulaforBn is

co
xII1 (9-b

Bn = n!

wherej = EiZ- bi and wherethe sum is taken over all nonnegativeintegersbi forwhich nj=i ibi = n.
Let F(t) = f[g(t)]. Then
'n
...
Zf(k)
g(n-k+1)
q
(2.2 n)=
(g
fUnk
k-O

or

= Bn(f; ',

g(n)),

f(k)

fk.

Generatingfunctionsand symbolic recurrencerelations for the Bell


polynomialsmay be foundin Bell [1] and Riordan [8, pp. 35-38, 45-48].
The firstfiveBn are:

Bo =

12

B1 = wgq,
B2 =

W2g2+ ?42,

B33= W3+
B4 = C4O9

32

g2 +

W/3,

+ 6Cw
3q212 + co2(4g,93+

3922) + Wg4

polynomial.Let P(t)/Q(t) be a
3. The formulafor the interpolatory
properrationalfunctionand let Q(t) have a zero of multiplicityp at xi.
Let
1

PI," .+

Q(t)

j=i (t - xi)

P(t)

E (t-x__

77(t)

+ x(t).

Then it is easy to show that


(3.1)

i
13P'P-j=

Zap

k=O

p-k

p (*-k)(Xi

(-)

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions

888

J. F. TRAUB

This resultis thekey to the solutionof the Lagrange-Hermite


interpolation
problem.It permitsus to writethe interpolatorypolynomialas a linear
combinationof the yjm)*
Let
n

r = 11
= (t
i=O

(3.2)

Q(t)

xi),

RiR(t)tt= t - ( )txi

L()
LiMt)-R(()

Ri(xi)

We calculate the contributionto Pnp(t) due to xi and thensum on i. We


have
Pnp( t) _Q ( t) P~Q(t)
Opi P5

k3.0p-

j5
Z

-Q)

(t (X)

( j-k)
-

(j

P(t)

)+

kRi
A
~RPx) (xi)

1aptip-k

k)!

Using
p(Fk)(Xi)

yi=

we obtain,aftersome manipulation,
E
Pnp,(t)= LiP(t) m=0

(t

(in)

(33i

V=O

Mn!

(t-x ) R (xi)DtvRi-P(xi) + p(t).


v!

Let
(3.4)

Sv(xi)

(-1)v(v-1)!

r=O;r i
7

(xi

Xr)

It followsfrom(2.1), (3.2), and (3.4) that


(3.5)

RiP(xi)DtvRi P(xi)

B,(p; Si, ***,).

fromall the xi, we obtain as a


Using (3.5) and adding the contributions
solutionto our problem

(3.6)

Pnp(t)

__
(
Li(t)
i=O
m=0

) Yi

V=O

_*

_1

*B,(p; SI,

S,S).

formulais contained
Thus theessenceofthe pthorderLagrange-Hermite
in the B,(p; SI, ** , ), O < v p - 1. Let
Gpim =

v=O

(t-x)
v.

B(p; Si I

S,).

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions

889

INTERPOLATION

ON LAGRANGE-HERMITE

may be obtainedfromthepolynomialGpio by truncatObservethat Gp,i,m


ing the highestm terms.Hence foreach p, Pnp(t) may be easily obtained
fromGp Gpio. The firstfiveG, are:
Ci

G2=

1,

1 + (t

x)2S1,

+ (t - xi)3S1 +
1 + (t - xi)4S, +

G3 = 1

2(t

Xi)2[32S12

+ 3S2],

G4 =

2(t

xi)2[42Si2

+ 4S2]

+ I(t
G=

1 + (t
+

1(t

-_xi)3[43S13

-_xi)5S1
-

4S3j,

3.42S1S2

Xi)3[53S13

352S1S2

5S2]

2(t -_xi)2[52SI2

5S3]

+ 3S2
+ 52(4S1S3
+ 653S12S2
2) +
(t -_xi)[54S14
Equation (3.6) may be writtenin a numberof otherways. Let

T, = Tv(x*)= (v -1)!

x_-

r=OfrpziXi

5S4].

Xr

Then
n

p-i

t-X

(3.7) Pnp(t) = E Lip(t) E (tMIx)


m=O mV.
i=O
Let
Hp,m,l =

i()

P-1-m

r
v=k

V.I

*..*

(Ti),

P-i-rn

E,

B,(p; T,
i!v

Tv-k,

T,).

Then
n

P(t)

X mn

p-1 (t

L* (t) Yj
m=0
i=o

p-i1-m
y(m) k=O Hpimk

m.I

are polynomialsin the


A formulaforPnw,(t)in which the coefficients
Li(j)(xi) may be obtained as follows.Let
Ri-p(t) = f[g(t)],

Then using (2.2), and withL'j)

Li'j)(xi),

i (-1)kk!C(p

Rp(xi)DevR-P(xj) =

g(t) = Ri(t).

f(U) =u-P,

k=O

+ k - 1, k)Uv,k(L',

. L(v-k+1)

Hence
n

PBt,(t)

(3.8)

P-1

= j Lip(t) ,
p-1-m
izO

(t -xi)^
VI

(t

, (

k=O

yi(m)Epim

-Uvk(

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions

L (P-k~rl)

890

J. F. TRAUB

Observe that EpXm may be obtainedfromthe polynomialEv Xo by truncating the highestmtterms.Hence foreach p, Pn, (t) may be easily obtainedfromEp - Epi ,o . The firstfiveEp are
El =

1,

E2=

1 + (t

xi)[-2L'],

E3 =1

+ (t

xi)[-3L'J + 2(t

xi)2[-3L"

?4 =1

+ (t

x)[-4L'J +

X)2[-44L"

+ 6( -xs)3[-5L"'

(t

+ 20(L')2]

3[-4L"' + 60L'L" - 120(I')3],

6 (t -x)

E5 = 1 + (1 -x,)[-5L']
+

2(t

+ 12(L')2],

+ '(t

+ 90L'L"

+ 30(L')2]

210(L')3]

+ 120L'L"' + 90(L")2

xl)4[-5L(4)

x)2[-5L"

1260(L')2L"

+ 1680(L')4].
As faras calculationwiththeseformulasis concerned,observethat
Lij'(Xi) = Rij(Xi)-)
Ri (xi)-

The Ri() (x,), j _ 0, may be obtained from7r(t) by repeated synthetic


division.
4. Some applications.The interpolationformulamay be used to generalize the Cauchy relations,
n

ZxLit),

i=O

1,

*,

n.

Correspondingto the case j = 0, we have the followinggeneralization.

(4.1)

E Li(t)
i=O

v=O

Y.

B(p; Si, *,Sr).

of t on the rightside of (4.1) vanishes,


Since the leadingcoefficient
n

E
i=O

(4.2)

[r(i]

BP-,(p; Si, *

, SV-l)

= 0.

This generalizes
n

Ei= 7'(xi)
/()= O.
We can derive a formulaforthe confluentdivided difference
with the
same numberof repetitionsof all arguments,f[xo, p; x1, p;
; Xn
p].

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions

ON LAGRANGE-HERMITE

891

INTERPOLATION

(This notationis introducedin Traub [13,pp. 241-242].) Since this divided


is the coefficient
of the highestdegreetermin (3.6), we obtain
difference

fAXO;x I ; S., ;n
(4.2)

Sp-p-m

Bp-i -m( p

f=0

Mm)

i)) n

(m)

(x

j-r (X.)
E
raz

REFERENCES
[1] E. T. BELL, Exponential polynomials, Ann. of Math., 35 (1934), pp. 258-277.
[2] T. FORT, Finite Differences,Clarendon Press, Oxford, 1948.
[3] T. N. E. GREVILLE, A generalization of Waring's formula, Ann. Math. Statist.,
15 (1944), pp. 218-219.
[4] C. HERMITE, Sur la formuled'interpolationde Lagrange, J. Reine Angew. Math.,
84 (1878), pp. 70-79.
[5] A. S. HOUSEHOLDER, Principles ofNumerical Analysis, McGraw-Hill, New York,
1953.
[6] V. I. KRYLOV, Approximate Calculation of Integrals, Macmillan, New York, 1962.
[7] J. KUNTZMANN, Mgthodes Numeriques Interpolation-D6riv6es, Dunod, Paris,
1959.
[8] J. RIORDAN, An Introductionto Combinatorial Analysis, Wiley, New York, 1958.
[9] H. E. SALZER, Formulae for hyperosculatoryinterpolation, direct and inverse,
Quart. J. Mech. Appl. Math., 12 (1959), pp. 100-110.
[10]
, Hermite's general osculatoryinterpolationformula, this Journal, 8 (1960),
pp. 18-27.
[11] 0. SCHL6MILCH,
Compendium der Hbheren Analysis, II, Friedrich Vieweg und
Sohn, Braunschweig, 1895.
[12] A. SPITZBART, A generalization of Hermite's interpolationformula, Amer. Math.
Monthly, 67 (1960), pp. 42-46.
[13] J. F. TRAUB, Iterative Methods for the Solution of Equations, Prentice-Hall,
Englewood Cliffs,New Jersey, 1964.

This content downloaded from 128.59.160.233 on Thu, 10 Oct 2013 14:38:40 PM


All use subject to JSTOR Terms and Conditions