Approximation Theory - Lecture 6 6 Best Approximation in C (A, B) (Cont.)

Part III - Lent Term 2005
Approximation Theory Lecture 6

6 Best approximation in C[a, b] (cont.)
6.1 Chebyshev systems
The proof of the alternation theorem is based on the following two properties of Pn :
1) For any set of m distinct points (zi ), where m n, there exists a polynomial q Pn
that changes its sign exactly at these points;
2) Any p Pn has at most n distinct zeros in [a, b].
So, it is suggestive to make some generalizations.
Definition 6.1 (Chebyshev systems) A set = (u0 , . . . , un ) from C(K) is a Chebyshev system, if
it satisfies the Haar1 condition: each polynomial
p = a0 u 0 + + an u n ,
with not all coefficients equal to zero, has at most n distinct zeros on K. The (n+1)-dimensional
space Un spanned by such a is called a Chebyshev space.
Lemma 6.2 The following conditions are equivalent.
(i) (ui )n0 is a Chebyshev system.
(ii) For any n + 1 distinct points (xi )n0 K, the following determinant is not zero:

u0 (x0 ) un (x0 )

D(x0 , x1 , . . . , xn ) :=
.

u0 (xn ) un (xn )
(6.1)
(iii) If (xi )n0 are distinct points of K and (yi )n0 are arbitrary numbers, then the interpolation problem
a0 u0 (xi ) + an un (xi ) = yi ,
i = 0..n,
has a unique solution for the unknonwns (aj ).

Proof. Condition of Definition 6.1 can be expressed as follows. If (x0 , . . . , xn ) are distinct points
of K, then the system of n + 1 equations with n + 1 unknowns aj s
a0 u0 (xi ) + an un (xi ) = 0,
i = 0..n,
has only the trivial solution aj = 0. This is well-known to be equivalent to (ii) and (iii).
Lemma 6.3 For any m distinct points (zi )m

1 in (a, b), with m n, there is a polynomial q Un that
vanishes exactly at these points (except perhaps the end-points) and changes sign at each of these points.
Proof. For m = n we take q(x) := D(x, z1 , . . . , zn ) (see (6.1)). For m < n, we take a certain
combination of similar determinants.

Theorem 6.4 Let Un be an (n + 1)-dimemsional Chebyshev subspace of C[a, b]. Then p Un is the best
approximant to f C[a, b] if and only if there exist (n + 2) points a t1 < < tn+2 b such that
f (ti ) p (ti ) = (1)i ,
|| = kf p k,
(6.2)
i.e., if and only if the difference f (x) p (x) takes consecutively its maximal value with alternating signs
at least (n + 2) times.
1 Alfred
Haar, 1885-1933, Hungarian mathematician, studied in Gottingen
by Hilbert.
6.2 Haars unicity theorem

Lemma 6.5 Let K contains at least (n + 2) points. If p is a best approximation to f C(K) from a
Chebyshev subspace Un , then the set Z of all points for which |f (x)p (x)| = kf p k contains at least
(n+2) points.
Remark 6.6 This lemma is not covered by the Chebyshev alternation theorem. The latter is valid
only for the real-valued functions on an interval [a, b] or a circle T, while the lemma is applicable
to the complex-valued functions as well.
Proof. Suppose that Z = (xi )m
1 , where m n + 1. Then, by Lemma 6.2 (iii), we can find a
polynomial q such that q(xi ) = [f (xi ) p (xi )], all i, so that
max[f (x) p (x)] q(x) = max {|f (xi ) p (xi )|2 } < 0,
xZ
1im
a contradiction to Kolmogorov criterion.
Theorem 6.7 Let Un be a Chebyshev subspace of C(K). Then each f C(K) possesses a unique polynomial of best approximation.
Proof. Assume that, for a function f , there are two polynomials of best approximation, p and q:
kf pk = kf qk = E(f ). By the triangle inequality, we see easily that also r := 21 (p + q) is a best
approximation. By the previous lemma, there are at least (n+2) points for which |f (x) r(x)| =
E(f ). At each such point x, for the real numbers
:= f (x) p(x),
:= f (x) q(x),
we have | + | = 2E(f ), || E(f ), || E(f ). But this is possible only if = . Thus, p and q
coincide at least at (n + 2) points, and because of the Haar condition they are equal identically.
This theorem has a remarkable converse (given here without a proof).
Theorem 6.8 (Haar [1918]) Let (ui )n0 be linearly independent continuous functions on K that contains
at least (n + 2) points. Then each f C(K) has only one polynomial of best approximation if and only if
is a Chebyshev system.
The next theorem shows, however, that all the remarkable properties of the real-valued Chebyshev systems are restricted, in a sense, to the functions given on an interval.
Theorem 6.9 (Loss of Haar2 ) If a compact K contains a fork (a Y-shaped curve), then there is no Haar
subspaces of C(K) of dimension > 1.
Proof. Indeed, we can put x0 and x1 on the two arms of the Y, any other points being on the trunk
of the Y, and then the points can be kept apart and moved continuously so that x0 and x1 change
places, and so that the remaining points return to their original positions. Thus the sign of the
determinant of D(x0 , x1 , . . . , xn ) is reversed. It follows that a zero determinant must occur during
the moves, which excludes the Haar condition.

In particular, no polynomial space of dimension > 1 in more than one variable can be Chebyshev,
and that makes the construction of uniform best approximations to functions of several arguments
something of an art.
2 Haar
in German means hair
6.3 Exercises
6.1. Prove that Tn , the space of all trig. polynomials of degree n, is a Chebyshev space on
T = [, ).
P
Hint. Write tn (x) = nk=n ck eikx , substitute z = eix , and reduce the problem to the algebraic case.
6.2. Prove that, for distinct k , Un = span(ek x )nk=0 is a Chebyshev system on any [a, b].
P
k x
Hint. For p(x) = c0 + n+1
, p Un+1 , we have p0 Un . Use this fact with induction.
k=1 ck e
6.3. Using arguments similar to those used in the proof of Theorem 6.7 prove that, on the circle
(i.e. on the period T = [, )), there is no Chebyshev space of even dimension.
Remark. Chebyshev spaces on the circle of odd dimension do exist, e.g. Tn .
6.4. Repeat the Haar construction for
= {1, x2 } on [1, 1].
Find appropriate f and a set of its best approximations.
Theorem 6.8 (Haar [1918]) Let (ui )n0 be linearly independent continuous functions on K that contains
at least (n + 2) points. Then each f C(K) has only one polynomial of best approximation if and only if
is a Chebyshev system.
Proof. Half of this theorem has been already established. For another half, suppose that is
not a Chebyshev system. Then there exist points (xi )n0 such that the matrix [ui (xj )] is singular.
Let non-zero vectors [a0 , . . . , an ] and [b0 , . . . , bn ] be selected orthogonal to the columns and rows,
respectively, of this matrix. Thus
P
P
1)
j = 0..n; and 2)
i = 0..n.
i ai ui (xj ) = 0,
j bj ui (xj ) = 0,
P
P
Set q := i ai ui and let p = i ci ui be any polynomial. Then it follows that
P
10 ) q(xj ) = 0, j = 0..n,
20 )
j bj p(xj ) = 0 p Un .
A) Let F be the class of functions f C(K) such that f (xi ) := sgn bi = 1 for all non-zero bi s,
and kf k = 1. Then, for any f F, and for any polynomial p Un , at some xj we must have
kf pk |f (xj ) p(xj )| 1 .
Indeed, if |f (xj ) p(xj )| P
< 1 = |f (xj )| for all such xj , then sgn p(xj ) = sgn f (xj ) := sgn bj ,
contradicting the equality j bj p(xj ) = 0. Hence, En (f ) 1 for all f F.
B) Now, select any f F, suppose that kqk 1, and set
F (x) = (1 |q(x)|)f (x) .
This F is also in F , hence En (F ) 1. But for any [0, 1], the polynomial q is a best approximation to F because
|F (x) q(x)| |F (x)| + |q(x)| 1 |q(x)| + |q(x)| = 1.

Approximation Theory - Lecture 6 6 Best Approximation in C (A, B) (Cont.)

Uploaded by

Copyright:

Available Formats

Approximation Theory - Lecture 6 6 Best Approximation in C (A, B) (Cont.)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Approximation Theory - Lecture 6 6 Best Approximation in C (A, B) (Cont.)

Uploaded by

Copyright:

Available Formats

Part III - Lent Term 2005

Approximation Theory Lecture 6

has a unique solution for the unknonwns (aj ).

Lemma 6.3 For any m distinct points (zi )m

Haar, 1885-1933, Hungarian mathematician, studied in Gottingen

6.2 Haars unicity theorem

a contradiction to Kolmogorov criterion.

in German means hair

You might also like