Least Squares Approximations of Power Series: 1 20 10.1155/IJMMS/2006/53474

LEAST SQUARES APPROXIMATIONS OF POWER SERIES
JAMES GUYKER
Received 27 June 2005; Revised 12 June 2006; Accepted 22 June 2006
The classical least squares solutions in C[1, 1] in terms of linear combinations of ul-
traspherical polynomials are extended in order to estimate power series on (1, 1). Ap-
proximate rates of uniform and pointwise convergence are obtained, which correspond
to recent results of U. Luther and G. Mastroianni on Fourier projections with respect to
Jacobi polynomials.
Copyright 2006 Hindawi Publishing Corporation. All rights reserved.
1. Introduction
The ultraspherical or Gegenbauer polynomials p
n
(x) with given constant 0, nor-
malized by p
n
(1) = 1, arise in solutions to least squares approximation problems (see
[3, 11, 12]): dene an inner product on C[1, 1] by
f , g) :=
_
1
1
f (x)g(x)
_
1 x
2
_
(1)/2
dx. (1.1)
Then p
n
(x) is generated by applying the Gram-Schmidt procedure to 1, x, . . . , x
n
, and is
given recursively by
p
0
(x) =1, p
1
(x) =x, for n 1,
p
n+1
(x) =
n
n+
p
n1
(x) +
2n+
n+
xp
n
(x).
(1.2)
For each f in C[1, 1],
n
_
j=0
_
f ,
p
j
_
_
p
j
_
_
2
_
p
j
(x) (1.3)
is the unique polynomial which minimizes | f p|
2
over all polynomials p of degree
Hindawi Publishing Corporation
International Journal of Mathematics and Mathematical Sciences
Volume 2006, Article ID 53474, Pages 120
DOI 10.1155/IJMMS/2006/53474
2 Least squares approximations of power series
at most n. A consequence of a recent result of Luther and Mastroianni [8, Theorem 2.1
(Corollary 3.3)] on Fourier projections with respect to normalized Jacobi polynomials is
the following.
If /4 /4 +1/2, then
_
f (x)
n
_
j=0
_
f ,
p
j
_
_
p
j
_
_
2
_
p
j
(x)
_
_
1 x
2
_
cE
n
( f )ln(n+2), (1.4)
where
E
n
( f ) :=inf
_
_
_
_
f (x) p(x)
__
1 x
2
_
_
_
: p polynomial of degree n
_
(1.5)
and c is independent of f and n.
This extends a classical theorem on Chebyshev polynomial ( =0) approximation (see
[5, Theorem 14.8.2], [11, Theorem 3.3]). In particular, if n k 1 and | f
(k)
(x)(1
x
2
)
+k/2
|
< , then
E
n
( f )
c
(n+1)
k
E
+k/2
nk
_
f
(k)
_
, (1.6)
where c is independent of f and n, which generalizes Jacksons theorem (see [3, Theorem
4.8], [7], [8, Corollary 3.4]).
In this paper we obtain analogs to (1.4) and (1.6) for power series f dened on the
open interval (1, 1). Such functions f (especially without closed forms) arise, for ex-
ample, in solutions to dierential equations. It will be necessary to rst extend the above
least squares polynomial. This is accomplished in Section 2 by replacing the integral in
(1.3) by a sum in terms of Maclaurin coecients of f and inversion coecients of expan-
sions of monomials as linear combinations of ultraspherical polynomials. After proving
key properties of the latter coecients in Section 3, we then derive uniform or pointwise
estimates to f with these least squares extensions.
2. Generalized Fourier coecients
We rst consider a general notion of summability. The following implies the well-known
convergence tests of Abel and Dirichlet [2, Theorems 10.17, 10.18] but with modied
error estimates.
Proposition 2.1. Suppose that
a
i
and
b
j+1
b
j
converge. Then,
_
i>n
a
i
b
i
b
n+1
+
_
j>n
b
j+1
b
j
n
__
a
i
__
, (2.1)
where
n
__
a
i
__
:=max
_
_
i>k
a
i
: k n
_
(2.2)
converges to zero.
James Guyker 3
Proof. Note that
_
j>n
_
i>j
a
i
b
j+1
b
j
= lim
m
m
_
j=n+1
_
i>j
a
i
b
j+1
b
j
lim
m
n
__
a
i
__
m
_
j=n+1
b
j+1
b
j
<
(2.3)
by the hypotheses. Moreover,
_
j>n
_
i>j
a
i
b
j+1
b
j
= lim
m
m1
_
j=n+1
m
_
i=j+1
a
i
b
j+1
b
j
(2.4)
since by the triangle inequality,
_
j>n
_
i>j
a
i
b
j+1
b
j
m1
_
j=n+1
m
_
i=j+1
a
i
b
j+1
b
j
_
i>m
a
i
m1
_
j=n+1
b
j+1
b
j
+
_
jm
_
i>j
a
i
b
j+1
b
j
m
__
a
i
__
_
j>n
b
j+1
b
j
(2.5)
which converges to zero as m tends to innity.
Finally, since
m
_
i=n+1
a
i
b
i
=
_
m
_
i=n+1
a
i
_
b
n+1
+
m1
_
j=n+1
_
m
_
i=j+1
a
i
_
_
b
j+1
b
j
_
, (2.6)
we have
_
i>n
a
i
b
i
_
i>n
a
i
b
n+1
+
_
j>n
_
i>j
a
i
b
j+1
b
j
b
n+1
+
_
j>n
b
j+1
b
j
n
__
a
i
__
.
(2.7)
The quantity
n
(a
i
)) was used in [6] to approximate power series with linear combi-
nations of Legendre polynomials ( =1). By Abels theorem [4, page 325], f (x) =
a
i
x
i
is in C[1, 1] if and only if
a
2i
and
a
2i+1
both converge. In this case, we have for 0,
E
n
_
_
i
a
i
x
i
_
2
_
n+1,
__
n
__
a
2i
__
+
n
__
a
2i+1
___
, (2.8)
where
(n, ) :=
_
n
2 +n+2
_
(/2)(n/(+1))
_
2 +2
2 +n+2
_
. (2.9)
This is immediate from the proposition since for xed x < 1, b
i
= x
2i
(1 x
2
)
is non-
negative and decreases to zero, and hence
_
i>n
a
i
x
i
_
1 x
2
_
1
_
t=0
x
t
_
2i+t>n
a
2i+t
_
x
2i
_
1 x
2
_
1
_
t=0
2x
n+1
_
1 x
2
_
n
__
a
2i+t
__
=2x
(n+1)/(+1)
_
x
((n+1)/(+1))
_
1 x
2
_
_
1
_
t=0
n
__
a
2i+t
__
,
(2.10)
where |x
((n+1)/(+1))
(1x
2
)
=(n+1, ) by a calculus argument. Note that (n, )

1, and for each > 0, lim
n
(n, ) =0; and for each n > 0, lim
(n, ) =0. Further-

more, by Abels theorem and the proposition, (
a
i
x
i
)
(k)
is in C[1, 1] if and only if the
series

a
2i+t
(2i +t)
k
(t = 0, 1) converge, in which case we have in (2.10) corresponding
to (1.6),
n
__
a
2i+t
__
_
2
n+1
_
k
n
__
(2i +t)
k
a
2i+t
__
. (2.11)
Suppose now that

a
i
x
i
is a convergent power series on (1, 1) and p
n
(x) is ultras-
pherical with constant 0. By (1.2), since x
i
=xx
i1
, we have the inversion formula
x
i
=
i
_
j=0
m
i j
p
j
(x), (2.12)
where m
i j
=0 if i j is odd, m
00
=1, otherwise
m
i j
=
j 1 +
2j 2 +
m
i1, j1
+
j +1
2j +2 +
m
i1, j+1
(2.13)
with m
i1
:=m
i1,0
+(1/2)m
i1,2
when =0. (We assume m
i j
:=0 if either i < j or j =1.)
Clearly
i
j=0
m
i j
=1 and 0 < m
i j
1. In fact, (2.13) is equivalent to m
11
=1,
m
j j
=
j 1 +
2j 2 +
m
j1, j1
, j 2,
m
i+2, j
=
(i +2)(i +1)
(i j +2)(i + j +2 +)
m
i j
, i j.
(2.14)
James Guyker 5
This may be veried by rst showing by induction on n, where i = j +2n (with j xed),
that if (m
i j
) satises (2.14), then
m
i1, j1
=
j(2j 2 +)(i + j +)
i( j 1 +)(2j +)
m
i j
(2.15)
with m
i1,0
:=((i +1)/2i)m
i1
when =0 and
m
i1, j+1
=
(i j)( j +)(2j +2 +)
i( j +1)(2j +)
m
i j
. (2.16)
Substituting (2.15) and (2.16) into (2.13), we conclude that the matrices coincide.
By (2.14) we have the following well-known closed form for m
i j
(see [10, page 283]):
m
i j
=
( +2j)
_
(i j)/2 +1
_
(ij)/2
()
j
2
i+1
(/2)
(i+j+2)/2
_
i
j
_
, (2.17)
where m
i j
:= ((2
0j
)/2
i
)(
i
(ij)/2
) whenever = 0. (Recall the factorial function ()
n
:=
(+1)(+n1) when n 1 and ()
0
:=1 for }=0.)
We now dene the general Fourier coecient c
j
of
a
i
x
i
with respect to the sequence
p
n
) by
c
j
:=
_
i
a
i
m
i j
(2.18)
whenever this sum converges. Note that c
nj
:=
n
i=0
a
i
m
i j
is the jth coecient in the
expansion of the partial sum
n
_
i=0
a
i
x
i
=
n
_
i=0
a
i
_
i
_
j=0
m
i j
p
j
(x)
_
=
n
_
j=0
c
nj
p
j
(x). (2.19)
If c
j
exists for every j, then for 0 we also have
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
_
i>n
a
i
x
i
_
1 x
2
_
+
n
_
j=0
_
c
nj
c
j
_
p
j
(x)
_
1 x
2
_
_
i>n
a
i
x
i
_
1 x
2
_
+
n
_
j=0
_
i>n
a
i
m
i j
( j, ),
(2.20)
where ( j, ) := |p
j
(x)(1 x
2
)
. Note that p
j
(x) is bounded by one in [1, 1] since
p
j
(1) = 1, and hence by [12, page 95], p
j
is a convex combination of Chebyshev poly-
nomials. Thus ( j, ) 1. Moreover, (0, ) =( j, 0) =1; and for j }=0, since p
j
(x) =
(1)
j
p
j
(x) and
_
_
x
j
_
1 x
2
_
_
_
=
_
j
2 + j
_
j/2
_
2
2 + j
_
(2.21)
as above, it follows that
(2j t, )
p
2jt
(0)
=(1 t)
j
k=1
2k 1
2k 1 +
(2.22)
and lim
(2j t, ) =p
2jt
(0). Furthermore we have the following.
Lemma 2.2. If 0 < /4 , then there exists a constant c =c(, ) independent of j such that
( j, )
2
c
j!
()
j
(2j +)
, j =0, 1, . . . . (2.23)
Proof. Suppose that /4 . By [8, Remark 2.4],
sup
j
_
_
_
_
p
j
(x)
_
_
p
j
_
_
_
1 x
2
_
_
_
_
_
2
< . (2.24)
Since p
j
(1) = 1 and the orthonormal sequence p
j
/|p
j
| is unique (see [3, Theorem
4.2]), it follows from [12, page 82, equation 4.7.15] that for > 0,
_
_
p
j
_
_
2
=
2
2
() j!
(/2)
2
()
j
(2j +)
. (2.25)
(Incidentally, if =0 then |p
0
|
2
= and |p
j
|
2
=/2 for j > 0.)
3. Main properties of inversion coecients
We will use (2.20) to approximate series
a
i
x
i
on (1, 1) such that
(a
i
/i
s
) converges for
some nonnegative number s. Since c
j
=
i
(a
i
/i
s
)(i
s
m
i j
), we rst investigate convergence
of the sequence i
s
m
i j
for xed j. We begin with the following technical result.
Lemma 3.1. Let s 1 be real, and N := 2(1 +s)/s if 1 < s < 0 and N := 1 otherwise.
The function
h
s
(x) :=(x +1)
_
x +2
x
_
s
x (3.1)
is monotonically decreasing on [N, ) with limit 2s +1 as x approaches .
Proof. Letting y =2/x, we have that h
s
(x) 0 for x N if and only if
f
s
(y) :=(1 + y)
1s
+
s
2
y
2
(1 s)y 1 0 (3.2)
for 0 < y 2/N. Now f
s
(0) =0, and f

s
(y) 0 is equivalent to
g
s
(y) :=(1 + y)
s
_
1 s(1 + y)
_
1 s. (3.3)
However g
s
(0) =1 s and
g
s
(y) =(s)(1 + y)
s1
_
s +(1 +s)y
_
0 (3.4)
on (0, 2/N]. Hence h
s
(x) is decreasing on [N, ).
James Guyker 7
Finally note that
h
s
(x) =
_
1 +
2
x
_
s
+
(1 +2/x)
s
1
1/x
(3.5)
and hence the limit follows from LH opitals rule.
The next result is the key to our approximations.
Theorem 3.2. Let p
n
(x) be ultraspherical with constant 0, let m
i j
be dened by (2.13),
and suppose that s 0. Then for each j,
L( j) := lim
n
( j +2n)
s
m
j+2n, j
(3.6)
exists if and only if 2s 1. If 2s 1, then the sequence ( j +2n)
s
m
j+2n, j
is monoton-
ically increasing. If = 2s 1, then with s
:= |s] (|] = greatest integer function) and
:=|],
L( j)
2j +
j!
(1 +)(2 +)( j 1 +)
_
1 3 5(2s
1)
_
2j +
!
( j +1)( j +2)( j +
1)
_
1 3 5(2s
1)
_
, }=0,
(3.7)
where the inequalities are equalities if s is an integer; and the equality holds in the latter
inequality if is an integer.
On the other hand suppose that > 2s 1. Then for each j,
(a) L( j) =0,
(b) the sequence ( j +2n)
s1
m
j+2n, j
is summable,
(c) there exists an integer I( j) j such that whenever i j is even the following are
equivalent:
(i) (i +2)
s
m
i+2, j
i
s
m
i j
;
(ii) i I( j);
(iii) i
s
j( j +) (i +2)i
s
( 1) 2i
s1
(i +1)(i +2)[(i +2)
s1
i
s1
], and
(d) i
s
m
i j
=O(i
[(2s1)r]/2
) for any r in (0, (2s 1)).
Proof. Let s 0 and 0 }=i j. Then by (2.14), we have
(i +2)
s
m
i+2, j
i
s
m
i j
=
A
i
_
1 j/(i +2)
__
1 +( j +2 +)/i
_
i
s1
m
i j
, (3.8)
where A
i
= j( j +)/(i +2) +h
s
(i) ( +2) with h
s
(x) given by Lemma 3.1. It follows that
( j +2n)
s
m
j+2n, j
) is Cauchy if and only if
_
i
A
i
_
i
s1
m
i j
_
_
1 j/(i +2)
__
1 +( j +2 +)/i
_
< . (3.9)
Now
0 <
_
1
j
j +2
_
_
1
j
i +2
__
1 +
j +2 +
i
_
3 + j + (3.10)
and A
i
decreases to 2s 1 . Hence if ( j +2n)
s1
m
j+2n, j
) is summable, then L( j) ex-
ists by Proposition 2.1. Similarly, the converse is true if }= 2s 1 since A
1
i
increases
monotonically to (2s 1 )
1
.
As in (3.8), we have
i
_
i
s1
m
i j
(i +2)
s1
m
i+2, j
1
_
=
(1 +2/i)
_
+2 h
s1
(i)
_
j( j +)/i
(1 +2/i)
s
(1 +1/i)
(3.11)
which converges to 2s +3 by Lemma 3.1. Thus letting i = j +2n we have
2s +3 = j(0) +2 lim
n
_
n
( j +2n)
s1
m
j+2n, j
( j +2n+2)
s1
m
j+2n+2, j
1
_
. (3.12)
By Raabes test [9, page 396],
i
i
s1
m
i j
converges when the limit in this identity is greater
than one and diverges when it is less than one. Hence L( j) exists (and (b) follows) if
> 2s 1, and fails to exist if < 2s 1.
Next suppose that > 2s 1. We show L( j) = 0 by verifying

i
(i
s
m
i j
)
p
< for some
positive constant p. By Raabes test as above,
i
_
_
i
s
m
i j
(i +2)
s
m
i+2, j
_
p
1
_
=(iX)
(1 +X)
p
1
X
, (3.13)
where
X =
+2 h
s
(i) j( j +)/(i +2)
(i +1)(1 +2/i)
s
. (3.14)
Hence
lim
i
X =0, lim
i
iX = (2s 1). (3.15)
Therefore with i = j +2n,
p
_
(2s 1)
_
= j(0) +2 lim
n
n
_
_
( j +2n)
s
m
j+2n, j
( j +2n+2)
s
m
j+2n+2, j
_
p
1
_
. (3.16)
By Raabes test
i
(i
s
m
i j
)
p
< if p is chosen such that p[ (2s 1)] > 2. Thus L( j) =0.
By (3.8), (i +2)
s
m
i+2, j
i
s
m
i j
if and only if
j( j +)
i +2
+h
s
(i) +2, (3.17)
where h
s
(i), i 1, decreases to 2s +1 by Lemma 3.1. If 2s 1, then
+2 2s +1 h
s
(i)
j( j +)
i +2
+h
s
(i) (3.18)
so ( j +2n)
s
m
j+2n, j
) is increasing.
James Guyker 9
Finally if > 2s 1 (and thus 2s +1 < +2), then there exists an integer I( j) j such
that (3.17) holds if and only if i I( j). Furthermore, (c)(iii) follows from (3.17) and the
identity
h
s1
(x) =3 +
2
x
+
_
x +1
x
__
x +2
x
__
(x +2)
s2
x
s2
x
s3
_
. (3.19)
In order to show (d), let b
n
:= [( j +2n)
s
m
j+2n, j
]
p
, where p is chosen above. Let q be
in the interval (1, p[ (2s 1)]/2). There exists an integer N
0
such that if n N
0
, then
qb
n+1
< n(b
n
b
n+1
). Thus if N =maxI( j), N
0
, then
q
m
_
i=1
b
N+i
< Nb
N
+
m1
_
i=1
b
N+i
(N +m1)b
N+m
(3.20)
so by (c)(i) it follows that
(q 1)mb
N+m
(q 1)
m
_
i=1
b
N+i
< Nb
N
(N +m)b
N+m
. (3.21)
Therefore for every m=1, 2, . . . ,
b
N+m
<
Nb
N
N +qm
<
Nb
N
N +m
(3.22)
and if r := (2s 1) 2/ p, then r satises (d).
Finally suppose that =2s 1. By solving (2.15) for m
i j
when i j is even, we have
i
s
m
i j
=
2j +
j!
(1 +)(2 +)( j 1 +)
_
j1
k=0
i k
i + j + 2k
_
_
i
s
m
ij,0
_
, (3.23)
where m
ij,0
= 1 3 5(i j 1)/(2 +)(4 +)(i j +). Note that if s is a posi-
tive integer, then
m
ij,0
=
1 3 5(2s 1)
(i j +1)(i j +3)(i j +2s 1)
(3.24)
and hence L( j) is given with s
=s and equality in both inequalities in (3.7). Assume that

s is not an integer and i j = 2n. Since i
s
m
i j
is increasing and i
s
m
i j
/i
s
m
ij,0
is bounded
by and converges to ((2j +)/ j!)(1 +)( j 1 +) by (3.23), it suces to show that
( j +2n)
s
m
2n,0
converges and is bounded by 13(2s
1). This will follow from the

integer case once we show for xed j and n that
f (s) :=
( j +2n)
s
(2s +1)(2s +3)(2s +2n1)
(3.25)
is monotonically increasing for s 1/2 since =2s 1 0. But f

(s) 0 if and only if
ln( j +2n) 2
_
1
2s +1
+
1
2s +3
+ +
1
2s +2n1
_
. (3.26)
Since the right side is decreasing and ln( j +2n) ln(2n), the inequality will follow if we
show ln(2n)
n
k=1
(1/k). Now for every integer m in (1, n),
ln(n) =
_
n
1
dx
x
ln(m) +
n
_
k=m+1
1
k
ln(m1) +
n
_
k=m
1
k
ln(1) +
n
_
k=2
1
k
=
n
_
k=1
1
k
1.
(3.27)
Thus,
m1
_
k=1
1
k
ln(m1) +ln(n)
m
_
k=1
1
k
ln(m) +ln(n)
n
_
k=1
1
k
. (3.28)
Since
4
k=1
(1/k)ln(4)<ln(2)<
5
k=1
(1/k)ln(5), we have that ln(2)+ln(n)
n
k=1
(1/k)
for n 5.
Hence for xed n 5, f (s) is increasing and f (s) f (s
). Therefore ( j +2n)
s
m
2n,0
is
bounded and thus i
s
m
i j
converges. It follows that
( j +2n)
s
m
2n,0
=
i
s
m
i j
_
i
s
m
i j
_
/
_
i
s
m
ij,0
_
(3.29)
converges and has limit bounded by 13(2s
1). Finally
and the second in-

equality follows by multiple cancellations when }=0.
If s is an integer, then I( j) may be estimated easily.
Proposition 3.3. Let s be a nonnegative integer, > 2s 1, and let I( j) be given as in
Theorem 3.2. Consider the following expressions:
k
=2
s+1k
__
s +1
s +2 k
_
+
_
s +2
s +2 k
__
, k =0, 1, . . . , s 2,
s1
=4
_
s
2
+
_
s
3
_
+
_
s +1
3
__
,
s
=
s
( j) = j( j +) 2
_
s
2
2s +1
_
,
s+1
= (2s 1).
(3.30)
If s = 0, then I(0) = 0 and I( j) = |
s
/
s+1
] when j 1; and if s = 1, then I(0) =
|
s1
/2
s+1
]. Otherwise, let s =3k
+t
for nonnegative integers k
and t
2, and dene
R( j) :=1 +
max
_
s
( j),
k
+1
_
s+1
. (3.31)
Then I( j) < R( j) +1, and if
s
( j)
k
+1
, then R( j) 1 < I( j). In particular, if =2s, then
I( j) =R( j).
James Guyker 11
Proof. By the binomial theorem and the identity (
s
k
) +(
s
k+1
) = (
s+1
k+1
), a straightforward
computation shows that (c)(iii) of Theorem 3.2 is equivalent to
s+1
i
s+1
s
_
k=0
k
i
k
. (3.32)
The special cases s = 0 and s = 1, when j = 0, follow easily from (3.32). Thus let s =
3k
+t
in the other cases. By the following argument (which is usually given to bound
the zeros of a polynomial [1, Theorem 6.1]) we have if
i M( j) :=max
_

0
s+1
, 1 +

k
s+1
: k =1, . . . , s
_
, (3.33)
then i satises (3.32),
s
_
k=0
k
i
k
s+1
+
s+1
(i 1)
_
+
s+1
(i 1)
s
_
k=1
i
k
=
s+1
_
1 +(i 1)
i
s+1
1
i 1
_
=
s+1
i
s+1
.
(3.34)
Thus I( j) < M( j) +1.
We will show that M( j) = R( j) for all j, which will imply the proposition since in
the case
s
( j)
k
+1
we will then have that M( j) = 1 +
s
( j)/
s+1
and x = M 1 fails to
satisfy (3.32). If
s
(M1)
s
=
s+1
x
s+1
s
_
k=0
k
x
k
, (3.35)
then 0
s1
k=0
k
(M 1)
k
which is impossible since M 1,
k
0, and
0
= 2
s+1
> 0.
Thus by (3.17), since h
s
(x) decreases on [1, ), it will follow that M( j) 1 < I( j).
Thus it remains to show M( j) =R( j). The following results may be readily established
from the given denitions:
(a) if s =1 ( j }=0), 2, or 3, then M( j) =R( j);
(b)
0
/
s+1
1 +
1
/
s+1
for all s 2.
We therefore assume henceforth that s = 3k
+t
4, where k
1 and t
= 0, 1, or 2,
and we seek to verify that
max
_
1
, . . . ,
s
_
=max
_
s
,
k
+1
_
. (3.36)
We rst show that
s1
<
s2
which may be rewritten as
s
2
<
_
s +2
4
_
+2
_
s +1
4
_
+
_
s
4
_
(3.37)
or upon further simplication as
24s
s 1
< (s +2)(s +1) +2(s 2)(s +1) +(s 2)(s 3). (3.38)
This inequality is true when s = 4, and for s 4 the left side decreases and the right side
increases. Hence
s1
<
s2
.
Next we prove that
for i =1, . . . , s 3, it follows that
i+1
>
i
i s 3i. (3.39)
As above, we have that
i+1
>
i
is equivalent to
_
s +2
s +1 i
_
> 3
_
s +1
s +2 i
_
+
_
s +2
s +2 i
_
(3.40)
which in turn is equivalent to
(3i s 1)(i +s +2) +2i < 0, i =1, . . . , s 3. (3.41)
If s 3i, then (3i s 1)(i +s +2) +2i i s 2 5 since i s 3. And if s < 3i, then
s =3i k for some integer k 1. In this case
(3i s 1)(i +s +2) +2i =(k 1)(i +s +2) +2i 0. (3.42)
Hence (3.39) follows.
Finally since s = 3k
+t
3i for i = 1, . . . , k
, we have from (3.39) that

k+1
>
k
(k =
1, . . . , k
). Moreover, since s < 3i for i =k
+1, . . . , s 3, it follows from (3.39) that

k
+k

+k+1
(k =1, . . . , s 3 k
). Since
s1
<
s2
above, we have the desired identity (3.36).
Remark 3.4. For each s in Proposition 3.3, there exists J =J(s) such that
R( j) 1 < I( j) < R( j) +1 (3.43)
for all j J.
4. Approximating power series
We consider power series

a
i
x
i
such that the series

a
2i+t
2i+t
(t = 0, 1) converge for
some positive sequences
2i+t
that are either both monotonically increasing or both satisfy
lim
i
(
2(i+1)+t
/
2i+t
) = 1. If they are both increasing, then by Proposition 2.1 and the
ratio test,

a
i
x
i
converges uniformly on [1, 1], and as in (2.10) we have the estimates
for 0:
_
i>n
a
i
x
i
_
1 x
2
_
1
_
t=0
x
t
_
2i+t>n
_
a
2i+t
2i+t
_
_
x
2i
_
1 x
2
_
2i+t
_
1
_
t=0
2x
n+1
_
1 x
2
_
n+1
n
__
a
2i+t
2i+t
__
2x
(n+1)/(+1)
(n+1, )
n+1
1
_
t=0
n
__
a
2i+t
2i+t
__
.
(4.1)
James Guyker 13
On the other hand, if the sequences satisfy lim
i
(
2(i+1)+t
/
2i+t
) = 1, then

a
i
x
i
con-
verges pointwise on (1, 1) and for each x in (1, 1), there exists N = N(x) such that
inequalities (4.1) hold for n N.
We begin with the uniformly convergent case.
Theorem4.1. Suppose that
a
i
x
i
is a power series such that the series
a
2i+t
2i+t
(t =0, 1)
converge for some positive, monotonically increasing sequences
2i+t
, let 0, and let p
n
(x)
be ultraspherical with constant 0. For every j and t, the general Fourier coecient c
2j+t
of
a
i
x
i
with respect to p
n
) exists and satises
c
2j+t

n
_
i=0
a
i
m
i,2j+t
2
_
2
n+1
lim
i
1
2i+t
_
m
maxI(2j+t),n+1,2j+t
n
__
a
2i+t
2i+t
__
, (4.2)
where m
maxI(2j+t),n+1,2j+t
=O((n+1)
(+1r)/2
) 1 for any r in (0, +1).
Moreover, for each x in [1, 1],
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
1
_
t=0
2
x
(n+1)/(+1)
(n+1, )
n+1
+
_
2
n+1
lim
i
1
2i+t
_
_
_
2j+t
n
m
n+1,2j+t
(2j +t, )
+
_
n
<2j+tn
m
I(2j+t),2j+t
(2j +t, )
_
n
__
a
2i+t
2i+t
__
,
(4.3)
where
n
:=
1
2
__
2
+4(n2)( +1)
_
,
_
2j+t
n
m
n+1,2j+t
(2j +t, ) 1.
(4.4)
If 1/4 /4 , then there exists a constant c independent of n such that for n 2,
_
n
<2j+tn
m
I(2j+t),2j+t
(2j +t, )
c
_
1
_
2t +
+
1
_
4 +2t +
+
1
2
_
4
_
nt
2
_
+2t +
_
4 +2t +
__
(4.5)
and hence convergence is uniform in (4.3) if
lim
n
n+1
n
__
a
2i+t
2i+t
__
=0, t =0, 1. (4.6)
Proof. By Theorem 3.2 with s =0 and Proposition 2.1, we have that
c
2j+t

n
_
i=0
a
i
m
i,2j+t
_
2i+t>n
a
2i+t
m
2i+t,2j+t
2m
maxI(2j+t),n+1,2j+t
n
__
a
2i+t
__
, (4.7)
where m
maxI(2j+t),n+1,2j+t
satises the given identity,
n
__
a
2i+t
__
=max
_
_
2i+t>k
_
a
2i+t
2i+t
_
_
1
2i+t
_
: k n
_
, (4.8)
and 1/
2i+t
decreases with nonnegative limit. Thus by Proposition 2.1,
n
__
a
2i+t
__
max
_
k
__
a
2i+t
2i+t
__
_
2
k+1
lim
i
1
2i+t
_
: k n
_
_
2
n+1
lim
i
1
2i+t
_
n
__
a
2i+t
2i+t
__
.
(4.9)
Therefore (4.2) follows from (4.7).
By (2.20), we have
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
_
i>n
a
i
x
i
_
1 x
2
_
+
1
_
t=0
_
2j+tn
_
2i+t>n
a
2i+t
m
2i+t,2j+t
(2j +t, ).
(4.10)
Hence (4.3) follows from (4.1) and (4.2) where, by Proposition 3.3, I( j) = ( j( j + )
2( +1))/( +1), and therefore I(2j +t) n+1 if and only if 2j +t
n
.
Suppose that 1/4 /4 . By Lemma 2.2, ( j, )
2
c/(2j +) for all j, and
m
I(2j+t),2j+t
1, so the estimate follows by the proof of the integral test since for m2,
m
_
j=0
1
_
2(2j +t) +
1
_
2t +
+
1
_
2(2 +t) +
+
1
2
__
2(2m+t) +
_
2(2 +t) +
_
.
(4.11)
Remark 4.2. If
lim
n
n
n+1
n
__
a
2i+t
2i+t
__
=0, t =0, 1, (4.12)
then convergence in (4.3) is uniform since m
i j
( j, ) 1 for all i and j.
James Guyker 15
The quantity
n
(a
i
)) was approximated in [6] for the standard Maclaurin series of
calculus. We will illustrate the other parts of our estimates. For Legendre polynomials, we
have an estimate that is comparable to (1.4).
Example 4.3. Let = 1 in Theorem 4.1. By [6], m
i j
2/( j +1) for all i and by the proof
of the integral test,
m
_
j=0
2
2j +t +1
2 t +ln
_
(2 t)m+1
_
. (4.13)
Hence in (4.3), it follows that
_
n
<2j+tn
m
I(2j+t),2j+t
(2j +t, ) max
n
<2j+tn
(2j +t, )
_
2 t +ln
_
(2 t)
_
nt
2
_
+1
__
.
(4.14)
Since
lim
n
2 t +ln
_
(2 t)|(nt)/2] +1
_
ln(n+2)
=1, (4.15)
we have as with (1.4), if
lim
n
max
n
<2j+tn
(2j +t, )
n
__
a
2i+t
2i+t
__
ln(n+2)
n+1
=0, t =0, 1, (4.16)
then the approximations in (4.3) converge uniformly.
However, if 1/4 =/4 , then by Lemma 2.2
m
I(2j+t),2j+t
(2j +t, ) c
2
2j +t +1
1
_
2(2j +t) +1
(4.17)
and for m2,
m
_
j=0
1
(2j +t +1)
3/2
2
(3/2)t
_
2 +t
1 +t
_
(2 t)m+1
_
. (4.18)
Therefore convergence is always uniform in this case.
Finally we consider pointwise convergence. The next result with s = 0 is identical to
Theorem 4.1 with
i
1.
Theorem 4.4. Suppose that

a
i
x
i
is a power series such that the series

(a
2i+t
/(2i +t)
s
)
(t =0, 1) converge for some nonnegative number s, let 0, and let p
n
(x) be ultraspherical
with nonnegative constant 2s 1. For every j and t the general Fourier coecient c
2j+t
of
a
i
x
i
with respect to p
n
) exists and the following estimates hold.
(a) Assume that =2s 1. Then
c
2j+t

n
_
i=0
a
i
m
i,2j+t
L(2j +t)
n
__
a
2i+t
(2i +t)
s
__
, (4.19)
where L(2j +t) is bounded as in (3.7). Moreover, if n2x
2/s
/(1x
2/s
) (:=0 when
s =0), then
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
1
_
t=0
_
2(n+1)
s
x
(n+1)/(+1)
(n+1, )
+
_
2j+tn
L(2j +t)(2j +t, )
_
n
__
a
2i+t
(2i +t)
s
__
,
(4.20)
where
2j+tn
L(2j +t)(2j +t, ) is bounded by a polynomial in n with degree term
(1 3(2s
1)/(
+1)!)n
+1
if }=0, and degree term 2n if =0. In particular,
if
lim
n
n
+1
n
__
a
2i+t
(2i +t)
s
__
=0 (t =0, 1), (4.21)
then
n
j=0
c
j
p
j
(x)(1 x
2
)
converges pointwise to
a
i
x
i
(1 x
2
)
on (1, 1).
(b) Assume next that > 2s 1. Then for each j,
c
2j+t
n
_
i=0
a
i
m
i,2j+t
2max
_
I(2j +t)
s
, (n+1)
s
_
m
maxI(2j+t),n+1,2j+t
n
__
a
2i+t
(2i +t)
s
__
=O
_
(n+1)
[(2s1)r]/2
_
n
__
a
2i+t
(2i +t)
s
__
(4.22)
for any r in (0, (2s 1)). And if n 2x
2/s
/(1 x
2/s
) (:=0 when s =0), then
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
1
_
t=0
2
_
(n+1)
s
_
x
(n+1)/(+1)
(n+1, ) +
_
2j+t
n
m
n+1,2j+t
(2j +t, )
_
+
_
n
<2j+tn
I(2j +t)
s
m
I(2j+t),2j+t
(2j +t, )
_
n
__
a
2i+t
(2i +t)
s
__
,
(4.23)
where 2j +t
n
if and only if I(2j +t) n +1; thus,
n
may be solved from the
quadratic inequality (c)(iii) in j of Theorem 3.2. Furthermore,
_
2j+t
n
m
n+1,2j+t
(2j +t, ) 1. (4.24)
James Guyker 17
Proof. Inequalities (4.19) and (4.22) follow from Proposition 2.1 and Theorem 3.2 since
c
2j+t

n
_
i=0
a
i
m
i,2j+t
_
2i+t>n
a
2i+t
(2i +t)
s
_
(2i +t)
s
m
2i+t,2j+t
_
. (4.25)
Moreover, by (2.10) and (2.20),
_
_
a
i
x
i
n
_
j=0
c
j
p
j
(x)
_
_
1 x
2
_
1
_
t=0
x
t
_
2i+t>n
a
2i+t
(2i +t)
s
(2i +t)
s
x
2i
_
1 x
2
_
+
1
_
t=0
_
2j+tn
_
2i+t>n
a
2i+t
m
2i+t,2j+t
(2j +t, ).
(4.26)
Therefore (4.20) and (4.23) follow from Proposition 2.1, (4.19), and (4.22), since for
xed x in (1, 1) and t, the sequence (2i +t)
s
x
2i
(1 x
2
)
is decreasing for 2i +t > 2x

2/s
/
(1 x
2/s
), and lim
n
(n+1)
s
x
(n+1)/(+1)
=0 by the ratio test.
Suppose that =2s 1. Then by (3.7),
_
2j+tn
L(2j +t)(2j +t, )
|(nt)/2]
_
j=0
L(2j +t) (4.27)
which is bounded by 2(|(nt)/2] +1) if =0, and is otherwise bounded by
1 3(2s
1)
!
|(nt)/2]
_
j=0
(4j +2t +
)(2j +t +1)(2j +t +
1)
=
1 3(2s
1)
!
2
+1
_
n
+1
2
+1
(
+1)
+ lower terms
_
.
(4.28)
(This follows from the observation that for any nonnegative integer k,

n
i=1
i
k
is a poly-
nomial in n with degree term n
k+1
/(k +1).)
Remark 4.5. I(2j +t) may be dicult to approximate in (4.23) when s is not an integer
(see Example 4.9(c)). If s
:=min|s], ( +1)/2(s) and
(a
2i+t
/(2i +t)
s
) (t =0, 1)
converge, then so do
_
a
2i+t
(2i +t)
s
=
_
a
2i+t
(2i +t)
s

1
(2i +t)
s
s
;
n
__
a
2i+t
(2i +t)
s
__
2
(n+1)
s
n
__
a
2i+t
(2i +t)
s
__
(4.29)
by Proposition 2.1. Thus, although the resulting estimates with s
(instead of s) would
be less accurate, we may use (4.20) when s
= ( +1)/2 and (4.23) with Proposition 3.3

when s
=|s].
Example 4.6. Let s = =1. Then L( j) =2j +1 by (3.7), (2j +t, ) 1, and in (4.20) for
n 2x
2
/(1 x
2
), it follows that
_
2j+tn
L(2j +t)(2j +t, )
__
nt
2
_
+1
__
2
_
nt
2
_
+2t +1
_
(4.30)
and thus we have pointwise convergence when
lim
n
n
2
n
__
a
2i+t
(2i +t)
1
__
=0, t =0, 1. (4.31)
If /4 , then by Lemma 2.2,
_
2j+tn
L(2j +t)(2j +t, ) c
__
nt
2
_
+1
_
; (4.32)
so pointwise convergence follows in this case when
lim
n
n
n
__
a
2i+t
(2i +t)
1
__
=0, t =0, 1. (4.33)
Example 4.7. Let s =1 and =2 in Theorem 4.4. Then I(0) =2, and by Proposition 3.3,
I( j) =( j +1)
2
for j 1. Therefore, since (2j +t, ) and m
I(2j+t),2j+t
are bounded by one;
if n 2x
2
/(1 x
2
), then in (4.23),
_
n
<2j+tn
I(2j +t)m
I(2j+t),2j+t
(2j +t, )
_
t +1
3
___
nt
2
_
+1
__
2
_
nt
2
_
+3
__
(2 t)
_
nt
2
_
+t +1
_
;
(4.34)
so we have pointwise convergence when
lim
n
n
3
n
__
a
2i+t
(2i +t)
1
__
=0, t =0, 1. (4.35)
However, if /4 , then by Lemma 2.2,
_
n
<2j+tn
I(2j +t)m
I(2j+t),2j+t
(2j +t, ) c
_
n
<2j+tn
(2j +t +1); (4.36)
so as in Example 4.6 pointwise convergence follows in this case when
lim
n
n
2
n
__
a
2i+t
(2i +t)
1
__
=0, t =0, 1. (4.37)
Example 4.8. Let s = 2 and = 5 in Theorem 4.4. By Proposition 3.3, I( j) < R( j) +1 for
all j, and R( j) =( j( j +5) +6)/2 < I( j) +1 for j 3. It follows that I( j) =R( j) when j 3
since R( j) is an integer in this case. Therefore for n max14, 2x/(1 x), we have in
(4.23),
_
n
<2j+tn
I(2j +t)
2
m
I(2j+t),2j+t
(2j +t, )
1
4
_
n
<2j+tn
_
(2j +t)(2j +t +5) +6
_
2
,
(4.38)
James Guyker 19
where the latter sum is a polynomial with degree term n
5
/10. If /4 in this case, then
by Lemma 2.2 the corresponding polynomial estimate is of degree four.
Example 4.9. Let us consider nonuniform approximation with Chebyshev polynomials,
that is, =0 2s 1 in Theorem 4.4.
(a) Let s =1/2. By (3.7), L( j) 2, and thus if n 2x
4
/(1 x
4
), then in (4.20),
_
2j+tn
L(2j +t)(2j +t, ) 2
__
nt
2
_
+1
_
; (4.39)
so we have pointwise convergence if
lim
n
n
n
__
a
2i+t
(2i +t)
1/2
__
=0, t =0, 1. (4.40)
(b) Let s =0. By Proposition 3.3, I( j) = j
2
2 and in (4.23),
_
n
<2j+tn
m
I(2j+t),2j+t
(2j +t, ) n
n+3. (4.41)
(c) Let 0 < s < 1/2. By (c)(iii) of Theorem 3.2, i I( j) if and only if
j
2
i
s
(i +2)
2
i
s
(i +2)
1+s
(i +1) (4.42)
(which may be used to nd
n
but can only be solved numerically for i, with j xed,
to approximate I( j)). However we may obtain estimates to (4.23) by replacing s
with s
=1/2 and using (4.20) as in (a) above and Remark 4.5.

Acknowledgment
The author is grateful to the referee for suggesting several improvements.
References
[1] O. Aberth, Precise Numerical Analysis, Wm. C. Brown, Iowa, 1988.
[2] T. M. Apostol, Calculus. Vol. I: One-Variable Calculus, with An Introduction to Linear Algebra,
2nd ed., Blaisdell, Massachusetts, 1967.
[3] K. E. Atkinson, An Introduction to Numerical Analysis, 2nd ed., John Wiley & Sons, New York,
1989.
[4] R. G. Bartle, The Elements of Real Analysis, 2nd ed., John Wiley & Sons, New York, 1976.
[5] K. R. Davidson and A. P. Donsig, Real Analysis with Real Applications, Prentice Hall, New Jersey,
2002.
[6] J. Guyker, Legendre expansions of power series, The Rocky Mountain Journal of Mathematics 29
(1999), no. 4, 12711284.
[7] D. Jackson, The Theory of Approximation, American Mathematical Society Colloquium Publica-
tions, vol. 11, American Mathematical Society, Rhode Island, 1930.
[8] U. Luther and G. Mastroianni, Fourier projections in weighted L
-spaces, Problems and Methods

in Mathematical Physics (Chemnitz, 1999), Oper. Theory Adv. Appl., vol. 121, Birkh auser, Basel,
2001, pp. 327351.
[9] J. M. H. Olmsted, Advanced Calculus, The Appleton-Century Mathematics Series, Appleton-
Century-Crofts, New York, 1961.
[10] E. D. Rainville, Special Functions, Macmillan, New York, 1960.
[11] T. J. Rivlin, The Chebyshev Polynomials, John Wiley & Sons, New York, 1974.
[12] G. Szeg o, Orthogonal Polynomials, American Mathematical Society Colloquium Publications,
vol. 23, American Mathematical Society, New York, 1939.
James Guyker: Department of Mathematics, Bualo State College (SUNY), 1300 Elmwood Avenue,
Bualo, NY 14222-1095, USA
E-mail address: [email protected]
Submit your manuscripts at
http://www.hindawi.com
Operations
Research
Advances in
http://www.hindawi.com Volume 2013
Mathematical Problems
in Engineering
Abstract and
Applied Analysis
ISRN
Applied
Mathematics
Volume 2013
International Journal of
Combinatorics
Journal of Function Spaces
and Applications
International
Journal of
Mathematics and
Mathematical
Sciences
ISRN
Geometry
Discrete Dynamics
in Nature and Society
Volume 2013
Advances in
Mathematical Physics
ISRN
Algebra
Probability
and
Statistics
Journal of
ISRN
Mathematical
Analysis
Journal of
Applied Mathematics
Advances in
Decision
Sciences
Stochastic Analysis
The Scientifc
World Journal
ISRN
Discrete
Mathematics
Differential
Equations
Volume 2013

Least Squares Approximations of Power Series: 1 20 10.1155/IJMMS/2006/53474

Uploaded by

Copyright:

Available Formats

Least Squares Approximations of Power Series: 1 20 10.1155/IJMMS/2006/53474

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

Least Squares Approximations of Power Series: 1 20 10.1155/IJMMS/2006/53474

Uploaded by

Copyright:

Available Formats

LEAST SQUARES APPROXIMATIONS OF POWER SERIES

=(n+1, ) by a calculus argument. Note that (n, )

(n, ) =0. Further-

:= |s] (|] = greatest integer function) and

=s and equality in both inequalities in (3.7). Assume that

1). This will follow from the

and the second in-

for nonnegative integers k

, we have from (3.39) that

). Moreover, since s < 3i for i =k

+1, . . . , s 3, it follows from (3.39) that

is decreasing for 2i +t > 2x

:=min|s], ( +1)/2(s) and

= ( +1)/2 and (4.23) with Proposition 3.3

=1/2 and using (4.20) as in (a) above and Remark 4.5.

-spaces, Problems and Methods

You might also like