interpolation approximation thesisinterpolation process is developed. in the second chapter a...
TRANSCRIPT
3079
INTERPOLATION AND APPROXIMATION
THESIS
Presented to the Graduate Council of the
North Texas State University in Partial
Fulfillment of the Requirements
For the Degree of
MASTER OF SCIENCE
By
Ram Lal, B.S., M.S.
Denton, Texas
May, 1977
Lal, Ram, Interpolation and Approximation, Master of
Science (Mathematics), May, 1977, 53 pp., bibliography,
6 titles.
In this paper, there are three chapters. The first
chapter discusses interpolation. Here a theorem about the
uniqueness of the solution to the general interpolation
problem is proven. Then the problem of how to represent
this unique solution is discussed. Finally, the error
involved in the interpolation and the convergence of the
interpolation process is developed.
In the second chapter a theorem about the uniform
approximation to continuous functions is proven. Then the
best approximation and the least squares approximation (a
special case of best approximation) is discussed.
In the third chapter orthogonal polynomials as dis-
cussed as well as bounded linear functionals in Hilbert
spaces, interpolation and approximation and approximation
in Hilbert space.
TABLE OF CONTENTS
Chapter Page
I. INTERPOLATION ... .......... 1
II. APPROXIMATION.... ........... 25
III. HILBERT SPACE. . ........ . . . . . . 41
BIBLIOGRAPHY..-. .--.-...... .. ..... ... 53
iii
CHAPTER I
INTERPOLATION
In this chapter we will start with the interpolation
problem and will describe some interpolation processes,
the error involved in the interpolation and finally the
convergence of interpolation processes.
Definition 1.1. Let X be a given linear space. The
set of linear functionals defined on X forms a linear space
called the algebraic conjugate space of X and is denoted
by X*.
Theorem 1.1. If X has dimension n then X* has dimension
n also.
Proof. Let X have dimension n. Let x1,...,Xn be a
basis for X. Now, if x EX, then it can be uniquely expressed
as
x = a x + a2x + + ax11 2 2 n nTherefore, for a functional L EX* we have
L(x) = a L(x ) + a2L(x2) + ... + aL(x )1 . 2 ) +n nFor any x E X, let L.(x) = a ; i = 1,2,....,n. L. are func-
tional defined on X.
We will show L are independent. Suppose the L. are
not independent. Then for some . / 0, we have
1 1 1 + 2L2++ n Ln n= 0.
1
2
Therefore,
SL )+L 2L2( j) + . i.. L.(x.)+ (1.1)
+n L (x.) =0(x.) = 0.
We know L.(X.) = 0 when i / j and L.(x.) = 1 when i = j.
Thus from (1.1) we get
0+ 0 + .. + . 1+0 + ... + 0 = 03
or . = 0, a contradiction. Therefore, the L. are inde-
pendent.
We next show that every n+1 functionals are dependent.
Let L , .0.., Ln, Ln+1 E X*. Consider the n+1 n-tuples
(L 1),L(x2),. ..,Ln(x))'
i = 1,2,.. .,n+1. Since Rn (or Cn) is of dimension n,
therefore these n-tuples are dependent. Hence we can find
YlIY2.'''''n+1 not all zero such that
n+1y. i(L(X1 ),L.(x2),...,L1 n)) = 0 = (0,0,...,0).
i=l1
Therefore,
(yL1 + Y2 L2 + --- + Yn+y1 Ln+1 i)(x) = 0 ,
for i = 1,. ..,n. By taking linear combinations, we get
(yL1+ y2 L 2 + + yn+1 L 1 ) (x) = 0 for x E X.
Therefore, L 1 ,L2,...,LA+1 are dependent. Hence X* is of
domension n.
Q.E.D.
Now we define the general interpolation problem. Let
X be a linear space of dimension n and let L1 ,L2 ,...,Ln
be n given linear functionals defined on X. For given
numbers w1 ,w2,' .'w *n, can we find an element of X, say x,
3
such that
L.(x) = w i = 1,2,..
The answer is "yes" if the L. are independent in X*.
Lemma 1.2. Let X have dimension n. If x1 ,...x
are independent in X and Li,.*.Ln are independent in X*
then the determinant ILL(x.) I 0. Conversely, if either
x1,*...xn or L ,...,Ln are independent and IL (x)I 0,
then the other set is also independent.
Proof. Let x1 ,x2 .. 'xn be independent in X and
LL2,...,OJLnbe independent in X*. Suppose that ILi(x)H 0.
Then also IL.(x.) I = 0. The systemJ31
a L (x1 ) + a2L2 (x) + --- + anLn1(x1)=0
a L (xn) + a2L2(Xn) + --- + a L (x ) =011n 22n n n n
has a non-trivial solution a1 ,...,an. Then
(aL+ a2L2 + --- + anL )(x) 0,
i = 1,2,...,n. Since xl*.*xn forms a basis for X,
(alL + a2 L 2 +--- + anLn)(x) = 0 for x E X
and hence a1 1 + a 2L2 +...+ anLn = 0. This is a contra-
diction. Therefore ILi(x )I 0. Conversely, let
x 1 ,..,xn be independent and JL.(x.) I 0. Suppose on the
contrary L1 ,..,Ln are dependent in X*. Then there exist
a 1 ,...,an (not all zero) such that
a 1L1+ ... + a Ln = 0 or
(aL + +...+ anLn)(x) = 0 for all x E X.
4
Now x1,...,xnEX and hence
(a L +. + an L )(x.) = 01 1 n n 1
i = 1,...,n or
a1 L1(x1 ) + ... + anLn (x 1) =0
a1L (x,)+ ... +anL (x) =0.u1n n n n
Since IL. (x.) I 0, the homogeneous system should have only
the trivial solution, i.e., all a. are zero. But this is
a contradiction. Hence L1 ,. ..,Ln are independent in X*.
Q.E.D.
Theorem 1.3. Let a linear space X have dimension n and
let LJ,L2 ,...,Ln be n elements of X*. The interpolation
problem Li(x) = wi, i = 1,2,...,n possesses a solution for
arbitrary values w1 ,w2 , ... ,wn if and only if the Li's are
independent in X*. The solution will be unique.
Proof. (i) Suppose the interpolation problem L.(x) =w.1
possesses a unique solution. We have to show the L. are1
independent. Let x ,x2' 'X'n be a basis for the linear
space X. Let x EX be the solution to the interpolation
problem Li(x) = wi, i = 1,2,...,n. Since x EX, x can be
written as a linear combination of x1 ,x2 ,..'n. Let
x = a1x + a2X 2 + +...+axnxn
Then L.(a1x1 + a2x2 + ... +a x =i., i = 1,2,...,n or
alL.(x1 )+ a2 L.(x2 )+ ... +aa L. (x ) = w.T ss L(x wpo2 s ni n 1
The system L.(x = w. possesses a non-trivial solution if
5
IL (X) 0. Since xix2 ' *''Xn are independent and
IL. (x.) I 0, then by the Lemma 1.2, the L. are independent.1 3 1
(ii) Suppose the L. are independent. We have to show
that the interpolation problem L.(x) = w., i = 1,2,...,n
possesses a unique solution; i.e., we have to show
JL .(x) I / 0. We know xi,x2 '.. .IXn are independent and
also the L. are independent. Therefore by Lemma 1.2,
Q.E.D.
We state without proof the following theorem.
Theorem 1.4. Let z0,z,.. *.,zn be distinct real (or
complex) numbers and
1 ~ n1 z 0...z 0
1 z.z n
V(z0,z1 ,...,)zn) (1.2)
1 z ... z nn n
Then the recursion formulan-A
V(Z 0 Z,,...,zn) =v(z0 ,z1 ,... *zn-1) .R (zn-z) (1.3)i=l
holds and hence the Vandermonde determinant,
nv(zO,z.,...,zn) . T(z.-z.). (1.4)
Now we give some examples of systems possessing the inter-
polation property.
Example 1.1. Let there be given n+l distinct (real or
complex) points z0 ,z1 ,. .,zn and n+l (real or complex)
6
values w0 ,w1 ,. . .,wn. There exists a unique polynomial
pn (n for which pn(zi) = w , i = 0,l,...,n. Here
L (pn n 1n(z - 1- - ,=,,...,n and 1,z,...,zn is a
basis for Yn(=X). To show there exists a unique poly-n
nomial pn(z) EP , we know by Theorem 1.3 and Lemma 1.2
that we have to show JL.(z3) I 0. Now,
n
n1 z 0 ... zn
L (z )I = . =V(zO,z,...,z = II (z.-z.).
n
Since z0 , z,. ..,Zn are distinct,
n
IL.(z) = II (z.-z.) 0.j >i
Hence, there exists a unique polynomial pn EP for which
Pn(z i) = w, i 0,1,...,n.
Example 1.2. Taylor Interpolation
Let X = Yn, L0 (f) = f(z 0),...,Ln(f) = f(n)(z 0) Let
fn)) n n .-f(z) =Z2 a.z2 . We have Li( Z a.z )(ZO0'
j=0 3 j=0 2
i = 0,1,...,n. Then
a0L0 (1) + a1 LO(z) + + anL0 (z n
a + a1z + + a zn0 10 nO =
a L (1) + aL (z) + ... + a L (zn) = n!aO n l n n n n
and
7
1 21z 0 Z 0 ''' Zn
0n -0 1 2z0 .. nz0
IL (z) I = 0 0 2...n(n-)zn-2 2!...n!0.
0. 0 0... n!
Since ILi(z 3 ) I 0 and zi are independent, by Lemma 1.2
the L. are independent. Hence by Theorem 1.3 the inter-
polation problem
Li(f) = fUO (z0), i = ,1...,n
possesses a unique solution.
Example 1.3. Full Hermite Interpolation
Let X = .'N To avoid indexing we use the functional
information employed without using the symbol L as follows:
(me)0' ' (zn)
(m)f(z ),f'(z ),...,f (nn n (n
z. Z-,1N = m0 + ... +m +n).
If we can show the homogeneous problem has only the zero
solution, then the nonhomogeneous problem posseses a unique
solution.
If p E.N and satisfiesN (m0)
p(z0) = 0, p'(z0 ) =0,...,p (z0) = 0
(ml)p(z1 ) = 0, pt (z1) =0,... ,p (z1) = 0
0 1 (m n)p(zn) = 0, p'(zn) = 0,...,p n(zn) = 0
8
(mn)then p must vanish identically. Let p (zn) / 0 for the
time being. By the factorization theorem,
p(z) =A(z)(z-zo mO+1 (z-z )Ml+1... z-zn-1Mn-l+1 (z-z n~p) z)1= n
The degree of p(z) is N = m0 +m + .. + mn + n. Therefore,(in) 0 1' nM +1A is constant and p n(z) = A. mn,(z-zO)m0+ ..(z-z) )n-i)n n.zn- 0 n. -
Since the z are distinct, p(mn) (zn) = 0 which implies A = 0.
Thus p = 0. Hence the homogeneous interpolation problem
possesses only the zero solution.
Definition 1.2. A linear combination of 1, cos x,
cox 2x,...,cos nx, sin x, sin 2x,...,sin nx is known as a
trigonometric polynomical of degree n. The corresponding
linear space will be designated by Y . It has dimensionn2n+l.
Example 1.4. Let X = 6, L (f) f L( =
f(x),..,L2nf(x 2n) where -1 x0 <X1 <.'' <x 2 n <iT'
n nLet f(x) = a0 + Za cos jx+ 2Z a2j+1sin jx,
0 =1 4j j=
Lk k), k = 0,1,...,2n.
Let 11 cos x, sin x, cos 2x,, sin 2x,,
G = ILk ) I =
..cos nx0 sin nx0cos x sinx cos 2x sin 2x
...cos nx1 sin nx1
cos x 2n sin x 2n cos 2x2n sin 2x2n
...cos nx 2n sin nx2n
9
Multiply the 3rd, 5 ,..th columns by-i (iota) and add
them, respectively, to the 2nd 4th, columns. Then
ix. 2ix. nix.G = 1 e 3 sin x.e 3 sin 2x. ... e 3sin nx.j.
33 13
Multiply the 3 rd, 5th columns by -2i and to them add
the 2nd 4 th columns, respectively. Hence
ix. -ix. 2ix. -2ix. nix. -nix.(-2i)nG = 11 e 3 e e 3e ... e ie 31.
Interchange the columns,
(-l)n(n+l)(-2i)nG = le.nix. 1 enix
Mul tiply the j throw by enix , j = 0,1,. ..,2n to obtain
ni(x0+.. n+x2n)( n(n+l) ne (-1) (-2i) G
ix. 2nix
which is the Vandermonde. Henc
eni(x0+. x2n) (-1)
e,
n(n+l)n
2n ix. ixk= II (e - e k
j>kix. ixk
Since the x, are distinct, e 3 eik, for j / k. There-
fore, ILk.(xj) I = G / 0. We know the x. are distinct; there-
fore by Lemma 1.2 the Lk are independent and then by Theorem
1.3, the interpolation problem
Lk(f) f(xk) k = 0,1,...,2n,
possesses a unique solution.
We now proceed with how to represent a unique solution
guaranteed by Theorem 1.3. We will be dealing with the
10
Lagrange Formula and the Newton Formula. Let z 0 ',1,...,zn
be distinct points and define
n (z-z.)
lk(z) = 1 = ,,i=0 Tzk-z )i k
It is obvious that
1 k (z.= .=0Oif k j
lk(Z kif1 if k = j
For given values w0 ,w1 ,. .W ,wthe polynomial
Pn(Z) = J wklk(z) (1.6)k=0
is in andn
pn(z.) = w i = 0,l,...,n.
Formula (1.6) is called the Lagrange Interpolation Formula.
If we define w(z) = (z-zo)...(z-zn), then
1k(z) = wzz) (1.7)k ~(z-zk) wt( zk)
and (1.6) can be written as
n
pn (z) = w w(z(1.8)k=0 k (z-zk)w (zk)
If w = f(z.) for some function f, then
n WZpn(z) = f(zk) w(z) (1.9)
k=0 (~kw(k)
and pn (zk) = f(zk) k = 0,1,...,n.
Definition 1.3. We shall designate the unique poly-
nomial of class P that coincides with f at z0,.,.,zn byn0
Pn (f;z).
Suppose that q E -n . Then q is uniquely determined
by the n+1 value s q-(zi) , i =0 ,1,. . .,n. Hence we mus t have
11
Pn (q;z) E q(z). (1.10)
Note: nE 1k(Z) 1
k=0and (..
E (zk-z)lk(z) E 0, j = 1,2,..k=O
We know 1j(zk) = 6jk j,k = 0,1,...,n.
Definition 1.4. Let L1 ,...Ln be a set of functionals
and f1 ,'. fn be a set of polynomials such that L.(f.) = 6..
i,j = 1,...,n. Then the polynomials f. and the functionals
L. are said to be biorthonormal.
Let LO(f) = f(z 0), ... ,Ln(f) = f(zn). Obviously,
L.(l.) = 6... Therefore, the L. and the 1. are biorthonormal.1 3 13 1 1
We will state without proof the following theorem which says
that for a given set of independent functionals, we can
always find a related biorthonormal set of polynomials.
Theorem 1.5. Let X be a linear space of dimension n.
Let L1 ,L2 ,...Ln be n independent functionals in X*. Then
there are determined uniquely n independent elements of X,
say xx *a''' *, such that L(x6*) = .. For any x E X,
we haven
x = 2 L.(x)x.*.i=1
For every choice of wl,w2 ,...,wn the element
nx = 2w.x.*
i=l11
is the unique solution of the interpolation problem
L (x) = wi, i = 1,2,...,n.
12
The Lagrange formula has one drawback. If we want to
pass from a space of dimension n to a space of one higher
dimension, we must determine an entirely new set of ele-
ments y1*,y2 '*'' y 1 that are not related in a simple
fashion to the old set x*,x2..'n,x*. Newton's Represen-12 n
tation of the solution gets around this difficulty.
Let Z0 ,Z1 5,0...,zn be n+l distinct points and form
the n+l independent Newton polynomials 1,z-z0 ,(z-z0 )
(z-z ),...,(z-z )(z-zl)...(z-zn-1). For given values
w0 ,w 1,...,wn there is a unique member p of P for which1 n n
p(z.) = w i = 0,1,...,n (1.12)
and is given by
p(z) = a0 + a1(z-z0) + a2 (z-z0(z-z1 )
+ ... + an(z- z0 )(z-z 1 )...(z-zn- 1). (1.13)
If we solve (1.12) for ai's, we will get
a0 = w 0
w1- w 0 _ w 0 1a1 z -z -z -z + z -z (1.14)
10 0 1 1 0
1 w2 -w0 w1 -w 0
(z2 -z1 ) z2z0 Z1 Z0
w 0+w1
z0- 1 0)(z-z2) (zl-z0)(zl-z2 )
+ ()w 2
(z 2- z0)(z2-z 1)
13
Definition 1.5. The constant a. is called the dividedJ
difference of the jth order of w0 ,w1 ,...,w with respect
to z0 ,z1, ,...,z . It is designated by
a = [wol i' '...' ' j ]. (1.15)
On comparing the coefficient of zn in (1.8) and (1.13) we
getn
an = Ew0,...,wn kO E k (1.16)k=0 w'(zk)
i.e.,
n wan= w k
k=0 Z
and this gives the set defined in (1.14). Now, if we take
w to be some functionf at zy, i.e., f(z) = w i = 0,1,...,n
thenn
pn(f;z) = Z [f(zO),f(zI),...,f(zk)Jk=0
(z-z 0) ... (z-zk-1) (1.17)
and is called finite Newton series for f(z). If we define
LO(f) = f(z0)
f(z0) f(z1 )
L1 (f) %z-z1 + (z -z0 )
k f(z.)L (f)= ZLk .= k
j=0 (z.-z.)i=0 J
ifj
nthen, pnP(f; z) =2; Lk=f0 wk(z).
k=0
14
Now,
w.(z) EP if 0 j n -1,
and it follows that
Pn(wj(z);z) = w (z)
hencen
w.(z) = ZLk(w.(z))Wk(z). (1.19)J k=0
By putting j = 0,1,2,..., in (1.19), we get
Lk(w(z)) = k.
Therefore, the functionals L, L ,. . Ln,.. and the poly-
nomials w0 (z),w1 (z),...,wn(z),..., are biorthonormal.
In the Newton representation, we add an additional point
and increase the degree of the interpolating polynomial with
no extra work as in the case of Lagrange representation.
We say that the Newton representation has a permanence
property. We can achieve this type of biorthonormality and
permanence by the following theorem.
Theorem 1.6. Biorthonormality Theorem or Generalized
Newton Representation.
Let X be a linear space of infinite (countable) di-
mension. Let x1 ,x2 ,..., be a sequence of elements of X
such that for each n, x1,x 2''''n are independent. Suppose
further that L1,L2,..., is a sequence of linear functional
in X* such that for each n, the n xn determinant
ILi (xi)Ajn =0 0.Le ae i,j=0 o
Then, there are determined uniquely two triangular systems
15
of constants a..,b.. with a.. 0 (all i) such that if13 13 11
L* = allL1
L *= a L +a L2 21 1 22 2
L= a31L1 + a32 L2 +a 33 L3
.IL.* = Ea..L .
j=113 3
and
X.* =Xi1 1
x* b2 1x +x 2
x 3* b31x 1 +b 2x +x 3
i-1x.*=E b..x. + x.
1 3j=1 13 3 1
then we have L.*(x1) =S.. ij = 1,2,....1 3 13
Proof. We use mathematical induction. When n = 1, we
want to have L1*(x*) = 1. We get a 1L1(x1) = 1
a =f0.11 -L _____
Therefore, it is true for n = 1. Let it be true for n. We
will show that it is true for n+l. That will imply that it
is true for any n.
Now, since it is true for n, we have the following
constants.
a 11
a 2 1 a 22
anl an2 0 ann b bni n2 'n,n-1
wi th a1 1 *a 22...an/ 0 and such that
L *(x *= 6. i,j = 1,2,...,n.
We have to show that we can obtain first
bn+1, 1, bb+1,I 2, .. ,1n+l, nand from a knowledge of these values can then obtain
an].+ l11,0 . . .,a n+l, n+l
with an+ln+l / 0 and such that
.L *(x *= 6 i, j = 1,2,.. .,n,n+l.
The conditions not contained in (1.20) are
S n+ = 0
L*x1*)(
i = 1,2,...,n and
= 0 i = 1,2,...,n,
1.
Fro L.(x*) = 0, i =12From Li (xn+l) = i , ,...,n, we get
n+,l1*(x ) + ... + n+1,Ln 1 n 1 n+l
(1.22)
b L n(xi) + ... + b L *(x ) = -L *(xn+1,nn n n n+l
iLY*(xj) = ZaikLk(x) ,and using the fact Y
k=1
implies IY 1 =
(L *(x.))
IAI-B (where A,B,Y are matrices) and
i= aikLk j) we will eventually get the
k=1
following:
16
1
b 21
1
(1.20)
(1.21)
We know = A- B
17
L1l(xl) ... Li*(xn) a1 1 0 ... 0 L )...L n)
= a2 1 a2 2 ... 0
Ln ).Ln n) n 1)..n n
anl an 2 ... a
Hence, JL.*(x.) " = a1-a2 2 .. annIL. (x.) K._1 0.13 'i,3=1 1022 nfl 1 3 i1j-l
Now, L .* are independent and IL4*(x .) J1 . / 0. Thereforei 1 3 1,j=1
by Theorem 1.3 the system in (1.22) possesses a unique solu-
tion. Thus we know the b's and hence can determine the
xn+.Consider,1+1
L* (x*) = 0 i =l,2,...,nn+1
and
L* (x*) = 1.n+1 n
an+,iL 1 (x ) + .. + an+in+iLn+i (x = 0
n~ (x*)1 an+ln+l n+l~x
an L (x *) +...+ a L (x ) = 0 (1.23)n+1,1 I n n+1,n+1 n+1
an+ I, 1 n+l n+ln+1Ln+1 n+l
This system has a unique solution if
L (x J I'=, 0 or IL,.(x.i,'0.
Using the same arguments as above, we will eventually get
L *x-) .*. .L 1)L ..L (x+1)1 b21.bn+,1.1 Ln11 (x1 1(ixn+i) 1 - 9
0 1 ...b
L- .L* n+,2
n+l 1* ...Ln xL((xn+l) n+l 1)..(L+1n+1 .0 0...
i.e., 1 IL x Ijl Li-xj)ijl 1 0.
18
From (1.23), we have
an+l,n+1
0
1
L 3...n+l1 1
L (x* ) ...L (x* )1 n+1 n+l n+l
L (x * ).L (x* )
1 n+1 n+1 n+1
IL 3 x i 1
IL. (x.)jIn+11 3i,j=1
Since IL.(x.) ij=l 10and IL. (x.)ij= 0, then
an+1,n+l / 0. Hence it is true for n+1. Therefore, it is
true for any n.Q.E.D.
Definition 1.6. Let there be given a sequence of
values yOy 1 . . . . . . The difference of adjacent values is
designated by
Ayk Yk+l -Yk k = 0,1,...,
and is called the first order forward difference. Similarly,
A 2 AAkL ) =Ayk+1[ - Ayk
(yk+2 ~-Yk+l - (k+l Yk= Yk+2- 2 yk+l +yk
L (x* )o...L(x* )1n+1 n n+1
19
In general,
Yn1k = "( yk n 2k + 1 nyk
and we define Ayk = k'
We state without prdof the following theorems.
nTheorem 1.7. For n>-O , Anyk = n-r=nyk+r,=(n
r=0
n!
r!(n-r7!'
Theorem 1.8. In the case of interpolation at evenly
spaced points a,a+h,...,a+nh with values Y 0 'y ' ''in, then
the coefficients of interpolation polynomial are given by
kyo
ak -k k = 0,1,. .. ,n.k 1h
Theorem 1.9. Let p(z) be the unique polynomial of n
that takes on the values y0 'y1 '''.''n at the n+1 points
a,a+h,...,a+nh. Then
Ay0pn(z) = y0 + h (z-a) +
Any0+ -(z-a)(z-a-h)...(z-a-(n-l)h). (1.24)
n !h
If pn(f;z) interpolates to f at a,a+h,...,a+nh, then
pn(f;z) = f(a) hf a)(z-a) +n n
Anf(a)+ - (z-a)...(z-a-(n-1)h) (1.25)
where Af(a) = f(a+h) - f(a)
A2 f(a) = f(a+2h) - 2f(a+h) + f(a), etc.
Formula (1.24) and (1.25) are known as the Newton forward
difference formulas.
20
Note: If n = o, then the series on the right hand side
of (1.25) is called the Newton series for f.
Until now, we have been discussing the interpolation
process. Once the interpolation process has been carried
out, we would like to know how good are the approximations
that result. Remainder theory deals with this question and
we start with the Cauchy Remainder for Polynomial Inter-
polation.
Definition 1.7. The class of functions continuous on
[a,b] will be designated by C[a,b]. It is a linear space.
We state without proof the following theorem.
Theorem 1.10. Let f EC[a,b] and suppose that f(n+l)
exists at each point of (a,b). If as;x0 <XI '< ' <xn !b,
then
(x-x0)xx)f(x) - p (f;x) =- (n+l1) n.. (x-xQ f(n+l)(M
where min(x,x0 ... x '') < z<max(xx0 x) .The point
depends upon x,x0 '..'Ixn and f.
Corollary 1.11. Let Rn (f;x)=f(X) - Pn(fx), If
f ECn+l[a,bJ, we have
max jf(n+1) Ix-x0 1. ~x-xI1.. Ix-xnIIn fxl a !g t :g b (t)(+)
Definition 1.8. Let f be defined on [a,b] and assume
that at each point x0 E [a,bJ there is a power series ex-
pression of f valid in some interval such that
f(x) =a0 +a 1 (x-x0) + a2(x-xO 0) 2 + ...
21
Then f is said to be analytic on the inteval [a,b]. We
write f EA[a,bJ.
Definition 1.9. Let R be a region of the complex plane
and let f be a single valued function of the complex vari-
able z defined in R. If z0 ER, f is said to be analytic at
z0 (or regular at z0) if it has a representation of the form
C0
f(z)a= an(z-z) nn=0 n 0
A function is analytic (or regular) in R, if it is analytic
at each point of R. We write A(R) for the class of such
functions. A(R) is a linear space.
Theorem 1.12. Let R be a simply connected region and
let f EA(R) . Let z0 lie in R and suppose that C is a
simple, closed,rectifiable curve which goes around z0 in
the positive sense. Then
f(n) (z) n! jC f(z) dz.0 2ri C(z-z 0 n+1
We now prove a corollary to Theorem 1.10.
Corollary 1.13. Let f EA(R) where R is a region that
contains [a,b]. Let C be a closed curve that contains [a,bJ
in its interior and let L(C) be the length of C,
MC a zCf(z) I, 6 be the minimum distance from C to [a,b].
L(C)MThen, jRn(f X)If Pn(fX)IgLCn+ 01
(1.26)Proof. By Theorem 1.10,f(n+l)
JR (f*x) I=1f(WX>Pn(f;x) I jI M 10 1X-X0 ..- lx-xnJin n (1)!n (1.27)
where min(x, x 0 ' ' ''n ) < ( <max(x, x 0 ' ''n) By Theorem 1.12,
22
f(n+l) (n+1) f (z) dz2i c (z-)n+2
|f(n+1) )(n+l) !IS jf(z) dz, E E[a,b]2 IF c z-EjIf+ 2
maxCf(z) 2 L(n+)1
McL(C)n+2 -z(n+ 1).
From (1.27), we have
jf(x) -p (f;x)j M 2ML(C)2 n+1) 0 n I'n 2,F6n+2 T(n+1)!
Therefore,
M L(C)2f(x) - pn ( n+2 0 I nl.
Q.E.D.
Definition 1.10. (Triangular Interpolation Schemes)
Let there be given a triangular sequence of real or
complex points.
z 00
z 1 0 z11
T: z20 z21 z22
Suppose that. a function f has been defined on a region
containing the points of T, and let pn (f;z) be that element
of Y for whichn
pn(f;zni) = f(zni) i = 0,1,...,n; n = 0,1
In other words, pn (f;z) interpolates f at the points
of the (n+l)st row of T. The numbers in the rows of T may
23
or may not be distinct. This is an interpolation scheme of
great generality.
If we drop the row subscript, we get a scheme
z0
z 0 z
S: z0 z z 2
For a scheme of this type, we have
npn(f;z) = Z [f(zo),f(zl),...,f(zk)](z-zo)...(z-zk-l)'
nk=0
Note: The existence of lim pn (f;z) is identical with the
convergence of the interpolation series: symbolically,
f(z) ~ Z [f(zO), ... ,f(zk)](z-zO) ... (z-zk-1 'k=O
Definition 1.11. Let S be a set and (pnJ be a sequence
of functions defined on S. Let f be a function defined on
S. Now, either (a) if x ES and F->0, then there exists a
positive integer N such that if n>N, then
Jpn(x) - f (x) I < C.pn
or (b) if lim pn(x) f(x) for each x E S, then the sequence
{pn I is said to be convergent to f over S.
Definition 1.12. A sequence [pnI of functions defined
on S is said to be uniformly convergent to a function f over
S, if for each eF>0 there exists a positive integer N such
that if x E S and n >N, then
Ipn(x) - f(x) I <
24
Theorem 1.14. Let R, S, T be bounded and connected
subsets of IR. Let Rc Sc T and S is a proper subset of T.
Let A = maximum distance from any point in R to boundary of S.
6 = minimum distance from any point in R to boundary of T.
and assume that <1. Let the points of a triangular system
lie in R and let f be analytic in T. Then, pn (f;x) con-
verges to f uniformly in S.
Proof. Let
00
x1 0 x 1 1
X2 0 x 2 1 x 2 2
be a triangular system in RcT. Let C = T. Therefore,
L(T) = length of the interval T. Let MT = maxlf(x)I. LetxET
We can find a positive integer N such that
MTL(T) (A) N+l
2r6 6
Let xES and n>N. By Corollary 1.13,
MTL(T)f (x) - Pn (f2Xn I ':' n
2 6n+2
MTL(T) A n+l TL(T) N+l
Therefore, jf(x) - pnp(f;x) I <c. Hence pn(f;x) converges to
f uniformly in S. Q.E.D.
CHAPTER II
APPROXIMATION
In this chapter, we will start with uniform approxi-
mation and will discuss the best approximation and finally
the least squares approximation (a special case of the best
approximation).
The uniform approximation problem is as follows:
Let f be a continuous function defined on [a,bl; then for
a given s >0 can we find a polynomial p such that
f(x) - p(x)I<c, asx s b?
The answer to this question is yes and is assured by the
following theorem.
Theorem 2.1. (Weierstrass Approximation Theorem.)
Let f EC[a,bj. Given e>0, we can find a polynomial p for
which
If(x) -p(x)<, a:gx sb. (2.1)
Definition 2.1. Let f be defined on [0,1]. The nth
(n -1) Bernstein polynomial for f is given by
Bn kf(-)( )x (l n-k. (2.2)k=0
Note that Bn(f;0) = f(0), Bn(f;l) = f(l).
We state without proof the following:
Theorem 2.2. (Bernstein). Let f be bounded on [0,1].
25
26
Then
lim Bn(f x) f(x)
at any point at which f is continuous. If fECC[0,1], then
B n(f;x) converges uniformly to f in [0,13.
An immediate consequence of the Theorem 2.2 is the
following corollary.
Corollary 2.3. If f ECIO,l], then given an E >0 we
can find an integer N>0 such that for n>N we have
f(x) - Bn (f;x) I <6, 0 :!5 x :! 1.
By using the' transformation y awe can easily
transfer the result of Corollary 2.3 from [0,1] to [a,b]
and hence we will get Theorem 2.1.
By this method, we not only prove the existence of
polynomials of uniform approximation, but we get a simple
explicit representation for them.
In contrast to other modes of approximations, Bernstein
polynomials yield smooth approximants. If the approximated
function is differentiable, not only do we have Bn(f;x) -* f(x)
but
Bnp)(f;x) + f(P)(x) p = 1,2....
We state without proof the following.
Theorem 2.4. Let fECP[0,l]. Then lim B(P)(f;x) = f Ex)n~
uniformly on [0,13.
Note: The Bernstein approximants mimic the behavior of the
function to a remarkable degree. But there is a disadvantage
27
that the convergence of the Bernstein polynomial is slower
than other methods. This fact seems to have precluded any
numerical application of Bernstein polynomial from having
been made. Perhaps they will find application when the
properties of the approximant in the large are more impor-
tant than the closeness of the approximation.
Now we consider the following problem. If fE[a,b] and
x~x 2''4'9xnE[a,bJ and e>O, then we can find a polynomial such that
If(x) -p(x) I <., xE[a,b] and p(x ) = f(xi), i = 1,2,...,n?
The answer to this problem is yes, and we prove the
following theorem.
Theorem 2.5. Let f EC[a,bJ and xj,x 2',''xn be n
distinct points in [a,bJ. Then there is a polynomial p
which uniformly approximates f on [a,b] and satisfies the
auxiliary conditions
p(x1) = f(x ), i = l,2,...,n.
Proof. Recall
lk(x) w(xk (x-xk wEk)n n
where w(x) = k=l k Let kM = xEa,b k I. Let
c >0. Then -' = >0. There is a polynomial p, such that1+M1
f(x) - p,(x) j < ', x E [a,b-].
nLet q(x) = Z (f(xk) Pl(xk))lk(x). Therefore, q is the
k=l
unique element of "W-6 withn -1
q(xk) = f(xk) - pl(xk), k =1,2,...,,n.
28
Now,
nmx q() m XaxxE[a ,b] k If(xk) pl(Xk) xE[a b] Ilk(x) I
n<l'-kEmax llk(x) = -M.
k=l E k
Set p = pl+ q.
p(xk) = pl(Xk) + kq(x) = fk), k =,2,
and also
If(x) -p (x) I ! gf(x) -p1(x) I + q(x) I
< F_ + F_ 'M =_ '(1+1,4) = C
i.e., lf(x)-p(x)I <C.
Therefore, f is uniformly approximated by p on [a,b] and
f(xi) = p(xi) i = 1,2,...,n. Q.E.D.
Theorem 2.6. Every continuous function can be approxi-
mated uniformly on [a,b] by continuous piecewise linear
functions.
Proof. Let f be continuous on [a,bI. Therefore, f
is uniformly continuous on [a,b]. Let . >0. There is a
6 >0 such that if x,y E [a,b] and Jx-y 1 <6 ', then
If(x) - f(y) I <c.
Pick 6 >0Oso that 6 <7 and for some positive integer
n, n 6 = b - a. Divide [a,b] into n intervals so that the
i th interval I. = Uc i,c ], where c =,a+ iS i = 0,l,...,n.
Since f is continuous on [a,bI], it i continuous on
I c[a,b]. Hence f is bounded on Ii. Thus there exist m
and M. such that mi = f(t.) f(x) s f(hi) = M for xEI ,
where t,h EI .
29
Consider on the ijth interval the following linear
functionM. -M.
L(f;I. ;x) - '6 (x-c. 1 ) + m..
L(f;I.;x) is piecewise linear on [a,b]. Clearly,1
m. L(f;I.;x) M. for x EI., i = 1,2,,...,n. Let x E [a,b].1 1 1 1
Then for some j
-e <m.i- f(x) L(f;I.;x) - f(x) M. - f(x) <6.
Thus,
L(f;I;x) - f(x) I <E.
Hence, every continuous function can be approximated uni-
formly on [a,bj by continuous piecewise linear functions.
Q.E.D.
Now if we try to approximate f with some polynomials
of fixed degree, we have to define the closeness of approxi-
mation. Frequently we measure the closeness of approximation
over the interval by taking the maximum deviation between
the function and its approximant.
Let D be the criterion for closeness of approximation
and f be a function. If p E JP such that the closeness ofn
p to f is not exceeded by any other element of P n, then p
is known as a best approximation to f (under D) . If we
change the criterion of closeness of approximation, the best
approximation will change.
Note:: lxil will designate the norm of x.
Let X be a normed linear space and x1 ,...,xn be n
linearly independent elements of X. Let y be an element of
30
X. We define the closeness of two elements as the norm of
their difference. We say that the linear combination
alx1+ ... + ann X'of'x'xn is a best approximate to y if
we have
y - (a1x + .+..+anxjfj| |y-(b1 x1 + .+bn) (2.3)
for every choice of constants b ,..,b . The element
y - (a1x + ... + anxn) is called. the error and
y - (alx1 + ... + anxn)| the norm of the error.
The best approximation problem is: can we find
al,. ..,an such that the norm of the error is minimum?
We state without proof the following theorem.
Theorem 2.7. Given y and n linearly independent ele-
ments xi, .... ,xn. The problem of finding
min|y- (a1x1 + . .. + anxn)|I has a solution,where y anda.
x ,...,xn EX.
We state without proof some corollaries to Theorem 2.7.
Corollary 2.8. Let f ECa,b] and n be a fixed integer.
The problem of finding min max if(x)-(a0+a x+. . .+an)a. a5xgb
has a solution.
Corollary 2.9. Let f EC[a,bj and n be a fixed integer.
Let p :1. The problem of finding min a bIf(x)(a+ax+.. .+axn)IPdx. al1
has a solution. Such a solution yields a best approximation
to f in the sense of least pth powers. We need only assume
that f ELP[a,b].
31
Corollary 2.10. Let x0 'xl'''.'Xk be k+l distinct points
k n 2and k 2 n. The problem of finding mi n E(f(x.) - (a +. . .+a x
1 =
possesses a solution. (This is the common problem of least
squares data fitting by polynomials.)
Definition 2.2. A norm on a linear space X is called
strictly convex if |jx|Is r, |y|fJ r imply Jx+yjj<2r unless x=y.
We state without proof the following.
Theorem 2.11. In a normed linear space X with a
strictly convex norm, the problem of best approximation
(posed in Theorem 2.7) has a unique solution.
Theorem 2.12. (Tonelli) Let S be a closed and bounded
set in the complex plane that contains more than (n+l) points.
Let f be continuous on S and set
M = min maxlf(z) - p(z) . (2.4)pE.P zES
n
Let pn (z) be any polynomial that realizes this extreme
value and set r(z) = f(z) - pn(z). Then,
(a). The number of distinct points of S at which ir(z) I
takes on its maximum value is greater than n+l.
(b). There is a unique solution to the problem (2.4).
Remark. If S contains n+l points, then M.= 0 and (b)
holds but not (a). If S contains fewer than n+l points,
then M = 0 and the solution is not unique.
Let fEC[a,b. We know by Theorem 2.12 that the pro-
blem of finding min max Jf(x) - p(x) I has a unique solution.pE nagxsb
32
Designate the solution by pn(x) and set
En (f) a x b If(x) Pn(X)
We state the following theorem without proof.
Theorem 2.13. Let f EC[a,bJ and p En with
max jf(x) - p(x) I = 6 . Then p is a best uniform approxi-arx!bmant to f of degree n and 6 = En(f) iff there are at least
n+2 points a!x1<..,<x n+2 b such that f(x ) -p(x) = *
i = 1,2,...,n+2 in an alternating fashion.
Until now, we have been dealing with the best approxi-
mation by linear functions. Now, we will study the best
approximation by non-linear functions. The situation here
is more complicated than in the case of linear functions.
We will show a case of best approximation by non-linear
functions.
Theorem 2.14. Let fECLa,b] and let m and n be fixed
integers 20. The problem of finding
n n-la0xn+a1x +...+an
mmn max Jf (x) -0 1a.,b . axeb b0 xm+b xm-1+. ..+bnl
i'j 1 mhas a solution.
Proof. There is some redundancy in the coefficients
of the rational function. We can adjust them so that
b2 +2 2b0 + b2 + + bm = 1. As b's vary, we shall certainly
obtain some polynomials that do not vanish in [a,b].
If we seta xn +.+a
A= inf max If(x)- 0na.,b. axgb bOxm+...+bm
33
then 0 L\<oo. By the definition of L, we can find a se-
quence of rational functions
R Ak(X)k(X)T=- x
k Bkwhere
Akn. (k) n-iAk(x) Z a.i=O
and
Bk(x) = $ b~(k)m-j.
Bk j=O
so that if
= a bIf(x) - Rk(x) I (2.5)
then
lim A= .k+oo
The coefficients b k) are bounded due to the normalizing
condition. From (2.5)
- fK f(x) -Rk(x) <k'
Hence,
Rk(x)I k + max f(x)I M (2.6)
for some constant M. Therefore,
Ak(x) I MiBk (x) (2.7)
Since the b k) are bounded, the polynomial Bk
are bounded on La,b] and therefore Ak(x) are bounded. Now,
if a family of polynomials of bounded degrees are bounded,
then its coefficients are bounded. Therefore a k) are1
bounded.
34
(k)(k (k()Consider the points Pk = (a0 k .. n. ,ank),bk),.,bk
in the space Rm+n+2. They lie in a bounded portion of that
space. Hence, we may select a subsequence Pk that converges
to a point
P'= (as,.. .,a',bb,.. .,b').
Consider the rational functions corresponding to this
subsequence, and reindex the subsequence so that we have
lim a. k)= a! i = 0,1,...,nk-*oo11
(2.8)
lim bk) = b' j = 0,1...m
Form, n n-lafxn+a x +...+a '
R'(x) = 0 M-1 n (2.9)bxm+bxm+...+b'0 1m
If we can show that
max If(x) - R'(x)I = a(2.10)asxsb
then R' will be a best approximant and this will complete
the proof.
R', being rational, can have at most a finite number of
asymptotes. Let D(x) be the denomiator of R'(x). Select
an x E [a,b] such that D(x) 0. At such a point we must have
lim Rk(x) = R'(x).k+oo
R'(x) = f(x) + Rk(x) - f(x) + R'(x) - Rk(x)
IR'(x) I < If(x) + If (x) - Rk(x) I +.IR'(x) - Rk(x)
max If(x)I+ 'k + -k'anxsb
35
Let y = sup . As k +' -o,6k+ 0. Therefore,k
IR'(x) 1:5 max If(x)1+ p. (2.11)
The bound holds uniformly for any x E [a,bJ for which
D(x) 0. This implies that R'(x) cannot have any asymp-
totes on [a,b], because if it did, there would be values of
x in a neighborhood of the asymptotic point where the bound
would be exceeded.
Let x be any point of :[a,b].
(i) Suppose D(x) / 0. Then for k = 1,2,...,
If(x) - R'(x) 1 If(x) - Rk(x) 1+ IRk(x) - R'I(x) I
'k+ '
as k-+o, k + 0 and Ak + A. Thus
If(x) - R'(x) I sgA. (2.12)
(ii) Suppose D(x) = 0. Then, we may find a sequence of
points of [a,b],xI ,2..., such that x. + x, D(x.) 0. Then
by (2.12), we have
If(xi) -R'(x )I i = 1,2,...
By continuity, we get
f(x) -RI(x)1 A. (2.13)
Therefore, from (2.12) and (2.13) we have
f(x) .-R'(x)I A for x E Ea.b]
or
max If(x) - R'I(x) sAas x b
By the definition of A, max if(x) - R'(x) A. Henceasxsb
max If(x) -R'(x)I = A.asxgb Q.E.D.
36
J. C. Mason (Ref. Approximation Theory by A. Talbot)
has used the term "near-best approximation" to f by an
absolute distance or by a relative distance. He uses this
terminology to deal with those approximants which are not
best but are close to best. He uses two types of approxi-
mation methods in L1 , L2 , L norms. The two methods are
defined by the unique interpolation criteria and the unique
series expansion criteria. He has demonstrated that the
resulting approximations of explicit functions are either
best approximations or "near-best approximations" in the
relative norm.
We now turn to the most commonly used approximation
process. The least square approximation is the best approx-
imation in an inner product space.
We state without proof the following theorem.
Theorem 2.15. Let x1 ,x2 ,..., be a finite or infinite
sequence of elements such that any finite number of elements
xl,x2,.'xk are linearly independent. Then, we can find
constants
a 11
21 22
a31 a32 a33
such that the elements
x*=a x1 111X* a x + a x
2 211 +22 2x*= a3x + a3x + a x.3 31 1 32 2 33 3
37
are orthonormal
(x*,x* i,j = 1,2....
This result is very useful for orthonormalization.
Definition 2.3. Let x*,x*..., be a finite or infinite11' 2'
sequence of orthonormal elements. Let y be an arbitrary
element. The series E (y,xn )x* is the Fourier seriesn=l
for y. (If the sequence is finite we use finite sum). The
constants (y,x*) are known as the Fourier coefficients of y.
We write00
y ~E (y,x*)x* . (2.14)n=l n n
Note that the' symbol is used to indicate that the right-
hand side is associated in a formal way with the left-hand
side.
Note: If X is an inner product space, then |x|J = (x,x)
normalizes X.
If xx 2'''''In' 0 are orthogonal, but not
necessarily normalized, then
x* = k k =1,2,0...,Xk = xk1
are orthonormal so that (2.14) becomes
00 x xk (ylxk)y Z xk k___k=1 7'fIxk1 xk 11 k=l(xkxk) xk'
We state without proof the following theorem.
Theorem 2.16. Let x1 ,x2 '.'xn be independent and let
x * be the xi's orthonormalized. If w = a x1 +.. .+an n, then
38
nW =1 (w,X *)X *.
k=l
The least squares problem can be formulated in terms of
Nfinding min||y- Z a xll in an appropriate inner product
ai 1=lspace.
Truncated Fourier expansions have the following minimum
property. This helps in solving the least squares problem.
Theorem 2.17. Let x*,x be an orthonormal system1' 2
and let y be arbitrary. Then,
N N|fy - Z (y,x *)x *|l1:|fy - Z a xi* 11
i=1 i=l
for any selection of constants a1 ,a2 ,...,aN'
Proof. ConsiderN N N
|y - Za.X.*|11.i==i=
N N N= (yy) - a.(x.* ,y) - Ei.(y;x.*)+ Z a.Ta.(x*,x.*)
i= V i=1 i, j=1 13i
N N N= (yy) - Za.(x.*,y)- -d (y,x*) + L 1a. 2
i=l1 i=l1 i=lN N
+ E (x.*,y)(y,x.*) - 2 (x.*,y)(y,x.*)i=l i=l
N N
= (y,y) - iZlI1(yx*) 12+ Z (a-(y,xp)(dT- (x*,y))
= (y,y)- N1(yx*Z)1 2 + a -(y,x*)I 2.
Since the 'first two terms of the last line are indepen-
dent of the a's, it is clear that the minimum of
N 2fly- Z a x /| is achieved when and only when
i=1
39
a.= (y,x-*) i = 1,2,...,N
i.e., when the a's are the Fourier coefficients of y.
Q.E.D.
Corollary 2.18. Let x1 ,...,XN be an independent set of
elements. The problem of finding that linear combinationN
of x1 ,...,xN which minimizes ||y - Z a xi|1 is solved byi=l
NZ (y,x.* )X.*
i=l 1 Nminfy aNxyy 2Corollary 2.19. minly - Z ax. i=|y||1-l {yx ) I .
a. i=1i=1
Corollary _2.20. if x1',2* is a sequence of ortho-
normal elements, then00
(y ,x.* 1) 2 : l 12
i=1I
and
lim (y,x.*) = 0.i-+>
Corollary 2.21. Let x1,x2 '' '.'xn be independent. Let
x*,x2 9x' be the xk's orthonormalized according to the
triangular scheme of Theorem 2.15. Then, for all selections
of constants a1 ,...,an-1, we have
x,*Iln I1a n 11sa 1x1 + a2x2 + .. + an-Xn- + XnJ'nn
Theorem 2.22. Let xV,x2 '.. ' 'Xn be independent elements
and let x*,x2..'.,'x' * be the x.'s orthonormalized. Then, for1l 22 n 1
any element y,
n((y - (y,xk*)xk*),x.*) 0.
k=l k
The proof follows on the next page.
40
Proof.n n
(y - (y,xk*k)xk*'kk(Y'.)k=l k=1 k
= (y x*)- (y,x.*) = 0.
Q.E.D.
Corollary 2.23. Let y be arbitrary and x1 ,x2 '''',x'n
be linearly independent elements of X. Let
a x + a2x +.. + ax1 l 2 2 n n
be best approximation to y. Then the error,
y - (a1x 1 + a2 +.. + ann)
is orthogonal to xj..
CHAPTER III
HILBERT SPACE
In this chapter we will discuss orthogonal polynomials,
bounded linear functionals in Hilbert spaces and interpol-
ation and approximation in Hilbert space.
We start with the general properties of real ortho-
gonal polynomials.
We state without proof the following:
Theorem 3.1. Real orthonormal polynomials satisfy a
three term recurrence relationship.
P (anx + b -)p cnp
n = 2,3,.. . (3.1)
The following form is particularly convenient for machine
computation
P-1 = 0
PO = 1
(3.2)pn (x) = n n n(xpP)Pi*(x) - (p p 2p xpn+n)- (n'n n'n)Pnl-I
p*(x) = pn(x/(p )2 n = 0,1,2,...n Vn )~.n n
n = 0,1,2,...
Theorem 3.2. Let p*(x) = kn n + snxn-l + ... be ortho-n n nnormal polynomials. Then the coefficients in (3.1) are given
by
41
Th
are rea
Th
must co
Pr
42
a k Snb =a ( n _sn-1
n-l n n-ik k kkn-2 k n kn-2
cn an kn k2 n = 1,2,... (3.3)n-i
eorem 3.3. The zeros of real orthogonal polynomials
L, simple, and are located in the interior of [a,b].
eorem 3.4. Let fEC[a,b]; then the Fourier segmentn23,(f,p*)p*,(x)k=0 n n
incide with f in at least n+l points of (a,b).n
oof. Let Rn(x) = f(x) - Z (f,pj)p*(x). Now,n k=0
n(R (X) P*~(x)) =(f (X) Zff- P( *()
n k ~3=033kn
= (f,p*) - (f,p*)(p*,p*)j=0 3 3
= (f,p) - (f,p*) = 0.
Hence, (Rn,p) = 0 for k = 0,1,...,n. In particular,
(Rn(x),p*) = 0 = (Rn(x),l). Hence Rn(x) must vanish some-
where in (a,b). Suppose now that it vanishes at
a <x<... <X<b and at no other points of (a,b). Then,
R n(x) is of constant and alternating sign in the segments
(a x ),a~2 '''jb)
and this is true of the polynomial (x-xl) ...(x-x ). Thus,
the product Rn (X)(x-x...~(x-x) has constant sign in (a,b)
and (Rn (x)(x-x1)...(x-xj),) = (Rn (x),(x-x1)...(x-x))
cannot vanish. But by orthogonality it must vanish for
j : n. Hence j >n and the theorem follows.Q.E.D.
Now we turn to Hilbert space.
43
Definition 3.1. A complete inner product space will be
called a Hilbert space, H, if the following additional re-
quirements are fulfilled: (a) H is infinite dimensional;
that is, given any integer n, we can find n independent
elements, (b) There is a closed (or complete) sequence of
elements in H.
or
A complete inner product space will be called a Hilbert
space, if there is a complete orthonormal infinite sequence
of elements in H.
We state some examples of Hilbert space.
Example 3.1. The set of all infinite sequences [a.}1
for which
i=l
augmented by the usual definitions for addition and scalar
products and by
00 7(a,b) = Z a.~, a = [a.1J, b = [bi
i=l
as the definition of an inner product, constitutes a
Hilbert space.
It is called the sequential Hilbert space and is desig-
nated by 12
Example 3.2. L2 [a,bJ with inner product
(f,g) = a b f(x) g xdx,
is a Hilbert space.
Now we discuss bounded linear functionals in normed
linear spaces and Hilbert spaces.
44
Definition 3.2. Let L be a linear functional defined
over the elements of a normed linear space X. L is said to
be bounded if there exists a constant M such that
IL(x) I M|jxlI, for all x E X. (3.4)
If no such constant exists., the functional is called un-
bounded.
We state without proof the following.
Theorem 3.5. A linear functional L defined on a normed
linear space X is bounded if and only if it is continuous.
Definition 3.3. Let L be a bounded linear functional
defined on a normed linear space X. Then JILII is defined as
the minimum value M for which (3.4) holds. We have, ob-
viously
IL(x) I s ||L ||x|l, x E X (3.5)
and for every E >0, we can find an x0 E X for which
IL(x0) I > (|ILI|- c)|jx ff. (3.6)
An alternate formula for ||L|1 is given by
IL| =1 -sup jL(x) IxEX (3.7)
We state without proof the following.
Theorem 3.6. The set of all bounded linear functionals
defined over a normed linear space X is a linear space. The
quantity iL|| = up IL(x)I makes this linear space into a
normed linear space.
Theorem 3.7. If L is a bounded linear functional over
a complete inner product space X, then there exists a unique
45
element x0 EX such that L(x) = (x,x0), x E X, where x0 is
known as the representer of the linear functional L.
Corollary 3.8. Let L be a bounded linear functional
over a complete inner product space. Let x0 be its repre-
senter. Then
J|L 11 = |x 11
and JL(x0) I = |IL|I l|x0 11.
The representer of a functional has a simple formula
in terms of a complete orthonormal sequence.
Theorem 3.9. Let H be a Hilbert space and x*,x ,...,
be a complete orthonormal sequence of elements. If L is a
bounded linear functional on H, then L(x) = (x,y) where y
has the Fourier expression00
y ~ L(xj)x*. (3.8)k=1
Moreover,00
L(x) = Z (x,x*)L(x*) x E H (3.9k=1
and00
IL||2 Z L(x*)12(3.10)kk=1
Proof. Let y be the representer of L. Then
y ~ Z(y,xj)x* = Z (x,y)x = L(x*)x.k=1 k=1 k=1
We knowCO
(x,y) = (x,xj) (x,y)k=1
and00 00
L(x) = (x,y) = Z (x,x*) (xZ,y) = 2 (x,x)L(xf).k=1 k=1
46
2 2Also, |IL|| = I21
By Parseval's Identity, we have00 CO
|Y11 2 Z2I(yx*) 12 = Z J1L(x) 12k=1 k=1
orC0
|y||2 - L(x*) 12.k=l1
Q.E.D.
We state without proof the following.
Theorem 3.10. Let H be a Hilbert space and tx*J a com-n
plete orthonormal sequence. Let L be a linear functional
defined on H and suppose that for all x EH we have
00
L(x) = (x,xj)L(xj).k=l
Then L is bounded on H and
JIL||2 - Z IL(x*) 12
k= k
Theorem 3.11. Let H be a Hilbert space. Let H* be
its normed conjugate space. Then H* can be made into a
Hilbert space in such a way that H and H* are essentially
the same. More precisely, we can find a one to one corre-
spondence (++) between H and H* such that
(a) x1*-+Lx2 +-+L2 implies a1x1+a2 2+a1L1+a2L2'
(b) x++L implies JJxJ| = JLIJ.
(c) An inner product can be introduced in H* by writing
(L1 ,L2) = (xl'x2 ) where x*-+L1, x2++L2
(d) The norm arising from this inner product coincides
with the original norm in H* (i.e., JJLJ = suHlx).
(3.11)I
47
Proof. Let {x*J be a complete orthonormal system in
H. Let L EH*. By Theorem 3.7, we have L(y) = (y,w) for
a unique w EH and for all y EH. The quantities (w,xt) are
the Fourier coefficients of w and hence,
00
11 12 = 1 (w ,x ) 12 < co.
The quantities (x ,w) = w-,x0) satisfy the same inequality
SI(x*,w) 1 ci 1 '<, and hence by Theorem 8.9.1(F) and Theorem
8.9.2 (Refer Interpolation and Approximation by P. J. Davis,)
they are the Fourier coefficients of a unique element of H
which will be designated by W. Note ||W| = |lwil.
Make the correspondence L--++w. This correspondence is
one-to-one between the whole of H and the whole of H*, for
each L E H* determines w E H and each w determines a w.
If L1 - W1 and L 2 + w2 then L1 L2 implies w1 w21for
we can find an x E H such that L1 (x) L2 (x). Therefore,
0 / L (x) - L2(x) = (x,w1) - (x,w2 ). Hence w1 I w2. Now
w / w2 . For otherwise,
(w ,x*) = (w,xtY) i = 1,2....
Then,
(w1 ,'*) = ~ x*) i = 1,2,...,
implying w1 = w2 .
Conversely, let v EH. Consider V as above and define
L by means of L(x) = (x,-v). By the above, the element v
corresponds to L. But v = v. Thus v corresponds to some
L in H*.
48
(a) Let xi++L , x2 ++÷L 2 . Then,
L (x) = (x,x1 ), L2(x) = (x'X 2 )
so that
(a1L1+a2L2)(x) = aL(x)+a2L2(x)
= (x,a x, 2 2
Now
(aX1+ a2 2 1 1+a2 2'
Hence
(a1 L1+a2L2)(x) = (x,a x1+a2x2)
and therefore,
alL+ a2L2 -+ax +a2 2.
(b) If w++L then L(x) = (x,w); thus ILl = ||W| = jw|j.
(c) The inner product properties of (L1,L2) follow
from those i-n H.
(L1 +L2 ,L3) (x 1 +x 2 'X 3 ) = (xl'x 3 )+(x 2 ,x 3 )
- (L1,L3) + (L2 ,L3)'
(L1 ,L2) ' 2) =x 2,x1 2 = 2'L9
(aLPL2)= (ax ,x2) (xlx2) = a(L 1,L2).
(L1 ,L1) = (x1,x1) = 1fx 1112 > and = 0 if and only if x1 = 0,
hence if and only if L = 0.
(d) (LL)1 = (x,x)- = |1[x11 =11lx1,
since L(y) = (y,x), y EH, iLl = . Hence fLit = (LL) .
Thus, H* is an inner product space. To show complete-
ness we need to prove that iLmLn li , m,n >N implies the
existence of an L with iL-Ln|l+0*.
49
Let L -+--xn. Then flxn-xmII = L -L Lmfs efor m,n> N6.
Thus [xnI is a cauchy sequence in H. Hence there is an x
such that x-xnll+ 0. If x+-+L, then ||L-LnI1 = lxxn il*0.
Finally, there is a complete orthonormal sequence in H*,
for if {xj} is complete and orthonormal in H and x*++L*,k n n
then L*j} is complete and orthonormal in H*. H* is there-n
fore a Hilbert space.
Q.E.D.
Note: In virtue of (3.11), the spaces H and H* are known
as isomorphic and isometric.
We state without proof the following.
Theorem 3.12. Let H be a Hilbert space and {xfJ be
an orthonormal sequence that is not complete. Then we can
find a sequence of elements [y*} (finite or infinite) such
that {x*J and fyg} together form a complete orthonormal set.
Now we will deal with interpolation and approximation
in Hilbert space.
Definition 3.4. Let x 1 ,x 2 '.*''Xn be n independent
elements of a linear space. The set of all linear com-
binationsn
X0 + a x
i=l
is known as a linear variety of dimension n.
Definition 3.5. Let x,,x 2''''Xn be n independent
elements of an inner product space and let c,c2 ,. ..,cn be
n given constants. The set of elements y that simultaneously
satisfy the n equations
50
(y,x.) = C., i =
is known as a hyperplane of codimension n.
Note: In a linear space of finite dimension n, the concepts
of linear variety and hyperplane are equivalent.We state without proof the following.Theorem 3.13. Let X be an inner product space. Let
X1 ,X2 ''.'''Xn be n independent elements. Then given any set
of n constants cc 2 , . . . , cn, we can find an element y such
that
(y,x ) = cl, i = 1,2,...,n.
Theorem 3.14. Let [xjj be an infinite sequence of
independent elements of a Hilbert space H and let constants
{c } be given. A necessary and sufficient condition that
there exists an element y EH such that
(y,x.) = c., i = 1,..., (3.12)
is that00
ZIak 2 <00 (3.13)k=l
where a1 = c / gXj),
(x1,x1) ... (xxl)
an -0(3.14)
g(xl1,-..,x n- 1)g8 l'''''n (x xn ) ... (XnX -)
n>l 0i... dn
If there is a solution, it is unique if and only if [x) is
complete (or closed) in H.
Let X be an inner product space and xl, ... xn be n
independent elements. Let S be the linear subspace spanned
51
by the x's. Let x* be the orthonormalized xi. Take an11
element z EX which is not orthogonal to all the x*'s, and
let V consist of all elements y ES for which (y,z) = 1.
Now, can we find a y EV such that ||y|| = minimum?} (3.15)
Theorem 3.15. The unique solution to the problem (3.15)
of finding a y such that ||yf| = minimum is given by
n ny = (z,xt)x* I(z,x*) 12. (3.16)
i=l =
The minimal distance is given by
Y12 n 1(3.17)
E J(z,x*) 2
i=l
Proof. Let y ES. Therefore, y can be expressed as
linear combination of x's.1
y = a x* + + ax*'11 n n
We know (y, z) = 1 where y E S and z E X and
a(x*,z) + ... + an(x*,z) = (y,z) = 1.1 1 n n
n (z,x*)
Set Z I(z,x*)12 =s 0, and write a = s + b where
i=l
the b. are now to. be determined. Now,1
n
1 = (y,z) = 1 + E b.(x*,z),i=l1
so thatnZ b.(x*,z) = 0.i=l 1 1
But
52
n n (z,x*) (x*,z)||y||2 2 I' 1 + b i)( + 1)
sn 1in n 2= + 2 b.(x*, z) +- Z B.(z,x*) + ZJb. 12
S2 s si=ls i l
1 n 2- + Ebi.i=1
The selection leading to the minimum |IIy| 2 is uniquely given
1by b. -0 and the minimum value is -. Now,
1s
n n (z,x*) n n 2y= a.x = - x*1 = (z,x*)x*1I(z,x*)1.
Q.E.D.
We state without proof the following.
Theorem 3.16. Let H be a Hilbert space. Let
x*,x*, .. ,x* be n independent elements of H. Then,1' 2' n
nw = 2 a.xt
i=l1
solves the problem
min||w|!w
subject to (w,xt) = a , i = 1,2,...,n. Moreover,
2 n 1min||w||2= E Ja2w i1l
BIBLIOGRAPHY
Books
Boas, Mary L., Mathematical Methods in the Physical Sciences,New York, John Wiley and Sons, Inc., 1966.
Cheney, E. W., Introduction to Approximation Theory, NewYork, McGraw-Hill BookCompany, Inc., 1966.
Churchill, Ruel V., Introduction to Complex Variables andApplications, New York,7McGraw-Hill Book Company,Inc., 1948.
Davis, Philip J., Interpolation and Approximation, NewYork, Blaisdell Publishing Company, 1963.
Articles
Mason, J. C., "Orthogonal Polynomial Approximation Methodsin Numerical Analysis," Approximation Theory, (byA. Talbot), New York, Academic Press, Inc., 1970.
Pruess, Steven, "The Approximation of Linear Functionalsand h2 -Extrapolation, SIAM Review, XVII (1975) .
53