Professional Documents
Culture Documents
legendrepaperinprintCA PDF
legendrepaperinprintCA PDF
DOI 10.1007/s00365-015-9321-3
B L. Bos
leonardpeter.bos@univr.it
A. Narayan
akil@sci.utah.edu
N. Levenberg
nlevenbe@indiana.edu
F. Piazzon
fpiazzon@math.unipd.it
123
Constr Approx
its orthonormalized version. Thus, with δi, j the Kronecker delta, the family Pn∗ satisfies
1
Pi∗ (x)P j∗ (x) dx = δi, j , i, j ≥ 0.
−1
1 ∗ 2
n
K n (x) := Pk (x) . (1)
n+1
k=0
1 1
lim K n (x) = √ , x ∈ (−1, 1),
n→∞ π 1 − x2
1 1 1
lim √ = 1, x ∈ (−1, 1),
n→∞ K n (x) π 1 − x 2
The purpose of this paper is to show that the above is actually an identity.
We expect this result to have use in applied approximation problems. For example, one
application lies in polynomial approximation of functions
from point-evaluations.
n Our
n
result indicates that the functions Q j (x) j=0 := √ K1 (x) P j∗ (x) are an ortho-
n j=0
normal set on (−1, 1) under the Lebesgue density √ 1 2 . If we generate Monte Carlo
π 1−x
samples from this density, evaluate an unknown function at these samples, and perform
least-squares regression using Q j as a basis, then a stability factor for this problem is
123
Constr Approx
given by max x∈[−1,1] nj=0 Q 2j (x) = n + 1 [2]. In fact, this is the smallest attainable
stability factor, and therefore this approximation strategy has optimal stability.
Note that (2) is equivalent to the following:
1 1 1 1
Pi∗ (x)P j∗ (x) √ dx = δi j , 0 ≤ i + j ≤ 2n,
−1 K n (x) π 1 − x 2
Written in this last form, our result (or rather, a transformed version on the unit circle
shown later in (3)) bears a strong resemblance to Berstein–Szegö approximations (also
referred to as a “Geronimus formula”) for orthogonal polynomials on the unit circle,
(e.g., [3], section V.2, p. 198). Given a measure dμ on the unit circle with associated
orthogonal polynomial family n (z), such results have the form
π π
p(z)
dθ = p(z)dμ(θ )
−π |n (z)|2 −π
for any polynomial p of degree 2n or less and z = exp(iθ ). There are also analogues
of this type of formula on the real line, e.g., [4]. However, the precise formula we
have derived has a factor of K n (z) in the denominator, instead of a squared orthogonal
polynomial modulus. To our knowledge, the formula (2) cannot be concluded in a
straightforward manner from existing Bernstein–Szegö results. Additionally, whereas
the Bernstein–Szegö formula is valid for a general class of measures dμ, it is shown
in [1] that formula (2) does not hold for any other orthogonal polynomial family.
The remainder of this document is devoted to the proof of (2).
123
Constr Approx
For z = eiθ in the integral (3), we obtain dθ = −i z −1 dz, cos(θ ) = J (z), and the
equation becomes
z −1 P ∗ (J (z))P ∗ (J (z))
1 i j
dz = δi j , 0 ≤ i, j ≤ n, (4)
2πi C K n (J (z))
In general, the coefficients of the factor polynomial, Tn (θ ) in this case, are algebraic
functions of the coefficients of the original polynomial. However, in our case we have
the explicit (essentially) rational factorization.
d n+1 z n−1 (z 2 − 1)
Then
1
K n (J (z)) = Fn (z)Fn (1/z). (6)
2(n + 1)
z n−1 z 2 − 1
Fn (1/z) = z −2n (n + 1)z n Pn (J (z)) − Pn
(J (z)) . (7)
2
Hence
2
z2 − 1
Fn (z)Fn (1/z) = z −2n (n + 1)2 z 2n (Pn (J (z)))2 − z 2(n−1) (Pn
(J (z))2
2
2
z2 − 1 2
= (n + 1)2 (Pn (J (z)))2 − z −2 Pn
(J (z) .
2
123
Constr Approx
2
−2 z2 − 1 1 1 2
z = z−
2 4 z
1 2 1
= z +2+ 2 −4
4 z
= J 2 (z) − 1,
so that
2
Fn (z)Fn (1/z) = (n + 1)2 (Pn (J (z)))2 − (J (z)2 − 1) Pn
(J (z)) .
1
2
K n (x) = (n + 1)2 (Pn (x))2 − (x 2 − 1) Pn
(x) .
2(n + 1)
Proof First, we collect the following known identities concerning Legendre polyno-
mials [5]:
(Christoffel–Darboux formula)
n
n+1
Pk2 (x) = Pn+1 (x)Pn (x) − Pn+1 (x)Pn
(x) , (8a)
2
k=0
(Three-term recurrence)
(n + 1)Pn+1 (x) = (2n + 1) Pn (x) + x Pn
(x) − n Pn−1
(x), (8c)
(x2
− 1)Pn
(x)
= n (x Pn (x) − Pn−1 (x)) , (8d)
1
K n (x) = Pn+1 (x)Pn (x) − Pn+1 (x)Pn
(x) .
2
123
Constr Approx
(x)P (x) − (n + 1)P
(n + 1)Pn+1
(x)P (x) = −nx P (x)P
(x)
−n 2 (Pn (x))2 − n Pn−1 n n n
(x)
x Pn
(x) = n Pn (x) + Pn−1
(8c)
Lemma 2 Let Fn (z) be the polynomial of degree 2n defined in (5). Then all of its
zeros are simple and lie in the interior of the unit disk.
has a zero at z = 0 and its other zeros are those of Pn (J (z)), namely, those z ∈ C for
which J (z) = r ∈ (−1, 1), a zero of Pn (x). But
J (z) = r ∈ (−1, 1)
⇐⇒ (z + 1/z)/2 = r
⇐⇒ z 2 − 2r z + 1 = 0
⇐⇒ z = r ± i 1 − r 2.
In particular |z| = 1 for the zeros of z n Pn (J (z)). It follows then from the Gauss-Lucas
Theorem that the zeros of Fn (z) are in the convex hull of z = 0 and certain points on
the unit circle, i.e., are all in the closed unit disk.
123
Constr Approx
1
K n (J (z)) = Fn (z)Fn (1/z),
2(n + 1)
so that K n (J (z)) also vanishes. But |z| = 1 implies that J (z) ∈ [−1, 1], and K n (J (z))
thus cannot vanish. Therefore, no zeros of Fn lie on the unit circle.
To see that the zeros are all simple, an elementary calculation and the ODE for
Legendre polynomials gives us
Fn
(z) = 2n(n + 1)z n−1 Pn (J (z)) + nz n − (n + 1)z n−2 Pn
(J (z)).
Hence Fn (z) = Fn
(z) = 0 if and only if
z 2 −1 z n Pn (J (z))
n+1 0
2 = .
2n(n+1)
z nz − n+1 z
z n−1 Pn
(J (z)) 0
Hence Fn (z) = Fn
(z) = 0 if and only if z n Pn (J (z)) = z n−1 Pn
(J (z)) = 0 if and only
if Pn (J (z)) = Pn
(J (z)) = 0, which is not possible as Pn (x) has only simple zeros.
As all the zeros of Fn (z) are in the interior of the unit disk, the zeros of G n (z) are all
exterior to the (closed) unit disk.
The following formulas for Fn (z) and G n (z) will be useful.
Lemma 3 We have
z n
Fn (z) = (2n + 1)z 2
− 1 Pn (J (z)) − 2nz Pn−1 (J (z)) ,
z2 − 1
123
Constr Approx
and
z n
2
G n (z) = z − (2n + 1) Pn (J (z)) + 2nz Pn−1 (J (z)) .
z2 − 1
z2 − 1
z2 − 1
z2 − 1
z n+1 z n+1
z n−1 Pn (J (z)) = 2n 2 J (z)Pn (J (z)) − 2n 2 Pn−1 (J (z)).
2 z −1 z −1
Lemma 4 We have
1. The polynomial y = Fn (z) is a solution of the ODE
(n − 2)z 2 − n
(1 − z 2 )y
+ 2 y + 6ny = 0.
z
z(1 − z)y
+ (c − (a + b + 1)z)y
− aby = 0,
123
Constr Approx
n
6. If we write Fn (z) = k=0 ck z
2k , then
n
G n (z) = cn−k z 2k
k=0
n
2(n − k) + 1 2k
= ck z .
2k + 1
k=0
Proof Equation (1) is easily verified using the ODE for Pn (x) and the definition of
Fn (z), (5).
(2) follows by changing variables z
= z 2 .
(3) follows as 2 F1 (a, b; c; z) is the only polynomial solution of the hypergeometric
ODE. The constant of proportionality is calculated by noting that the leading coefficient
of 2 F1 (a, b; c; z) is 2n + 1, whereas that of f n (z) is (2n + 1)/2n times the leading
coefficient of Pn (x), i.e., (2n + 1)/2n × 2n n /2 .
n
n
(a)k (b)k z k
2 F1 (a, b; c; z) = ,
(c)k k!
k=0
so that
n 2
1 (a)k (b)k (2k + 1) k
= 2n .
k! (c)k
2k
n
(6) follows easily from the fact that G n (z) := z n Fn (1/z) = k=0 cn−k z
k and that
cn−k is easily computed from the formula for ck given in (5).
2n
Pi∗ (x)P j∗ (x) = ak Pk (x)
k=0
123
Constr Approx
for certain coefficients ak . From (a) and (b), we then conclude that
2(n + 1)z 2n−1 P ∗ (J (z))P ∗ (J (z))
1 i j
dz = 2a0 .
2πi C Fn (z)G n (z)
But for i = j,
1
1
a0 = Pi∗ (x)P j∗ (x)P0 (x)dx = 0,
2 −1
as P0 (x) = 1 and Pi∗ (x) and P j∗ (x) are orthogonal. While for i = j, we have
1
1 1
a0 = (Pi∗ (x))2 P0 (x)dx = .
2 −1 2
We will now compute the partial fraction decomposition of the integrands in (a)
and (b),
with
(n) (n) (n)
(n + k + 1)Ak+1 := (2(n + k) + 1)J (z)Ak (z) − (n + k)Ak−1 (z), k = 1, 2, . . . ,
and
(n)
(n + k + 1)Bk+1 := (2(n + k) + 1)J (z)Bk(n) (z) − (n + k)Bk−1
(n)
(z), k = 1, 2, . . . .
Further, we let
123
Constr Approx
and
(n) (n) (n)
(n − k)Dk+1 (z) := (2(n − k) + 1)J (z)Dk (z) − (n − k + 1)Dk−1 (z),
1 ≤ k ≤ n − 1.
Proposition 2 We have
(n) (n)
2(n + 1)z 2n−1 Pn+k (J (z)) = Ak (z)G n (z) + Bk (z)Fn (z), k = 0, 1, 2, . . . , (10)
so that
and
(n) (n)
2(n + 1)z 2n−1 Pn−k (J (z)) = Ck (z)G n (z) + Dk (z)Fn (z), 0 ≤ k ≤ n, (11)
so that
as desired.
We now prove (10) for the case k = 1. We calculate, using Lemma 3,
(n) (n)
A1 (z)G n (z) + B1 (z)Fn (z)
= z n−2 G n (z) + z n Fn (z)
= z n−2 (G n (z) + z 2 Fn (z))
z n
2
= z n−2 2 z − (2n + 1) Pn (J (z)) + 2nz Pn−1 (J (z))
z −1
+z 2
(2n + 1)z 2 − 1 Pn (J (z)) − 2nz Pn−1 (J (z))
123
Constr Approx
z 2n−2
= (2n + 1)(z 4 − 1)Pn (J (z)) − 2nz(z 2 − 1)Pn−1 (J (z))
z −1
2
= z 2n−2 (2n + 1)(z 2 + 1)Pn (J (z)) − 2nz Pn−1 (J (z))
= z 2n−1 {(2n + 1)(z + 1/z)Pn (J (z)) − 2n Pn−1 (J (z))}
= z 2n−1 {(2n + 1)2J (z)Pn (J (z)) − 2n Pn−1 (J (z))}
(8b)
= 2(n + 1)z 2n−1 Pn+1 (J (z)).
The rest of the proof proceeds by induction. Assuming that (10) and (11) hold from
0 up to a certain k, we prove that they also hold for k + 1. To this end, we calculate
(n) (n)
Ak+1 (z)G n (z) + Bk+1 (z)Fn (z)
(n) (n)
(2(n + k) + 1)J (z)Ak (z) − (n + k)Ak−1 (z)
= G n (z)
n+k+1
(n) (n)
(2(n + k) + 1)J (z)Bk (z) − (n + k)Bk−1 (z)
+ Fn (z)
n+k+1
1
(n) (n)
= (2(n + k) + 1)J (z) Ak (z)G n (z) + Bk (z)Fn (J (z))
n+k+1
(n) (n)
− (n + k) Ak−1 (z)G n (z) + Bk−1 Fn (J (z))
1
= 2(n + 1)z 2n−1 {(2(n + k) + 1)J (z)Pn+k (J (z))
n+k+1
− (n + k)Pn+k−1 (J (z))} (by the induction hypothesis)
= 2(n + 1)z 2n−1 Pn+k+1 (J (z)),
123
Constr Approx
by the three-term recursion formula for Legendre polynomials (8b) with degree m =
n + k.
Similarly,
(n) (n)
Ck+1 (z)G n (z) + Dk+1 (z)Fn (z)
(n) (n)
(2(n − k) + 1)J (z)Ck (z) − (n − k + 1)Ck−1 (z)
= G n (z)
n−k
(n) (n)
(2(n − k) + 1)J (z)Dk (z) − (n − k + 1)Dk−1 (z)
+ Fn (z)
n−k
1
(n) (n)
= (2(n − k) + 1)J (z) Ck (z)G n (z) + Dk (z)Fn (J (z))
n−k
(n) (n)
− (n − k + 1) Ck−1 (z)G n (z) + Dk−1 Fn (J (z))
1
= 2(n + 1)z 2n−1 (2(n − k) + 1)J (z)Pn−k (J (z))
n−k
− (n − k + 1)Pn−(k−1) (J (z)) (by the induction hypothesis)
= 2(n + 1)z 2n−1 Pn−(k+1) (J (z)),
with m = n − k.
Due to the J (z) factor in the recursive definitions of Definition 1, the functions
A(n) (n) (n) (n)
k (z), Bk (z), C k (z), and Dk (z) are all Laurent polynomials. It is easy to verify
that they have the forms:
(n)
n+(k−3)
• Ak (z) = a j z j , k ≥ 1,
j=n−(k+1)
(n)
n+(k−1)
• Bk (z) = b j z j , k ≥ 1,
j=n−(k−1)
(n)
n+(k−1)
• Ck (z) = c j z j , 1 ≤ k ≤ n,
j=n−(k+1)
(n)
n+(k−1)
• Dk (z) = d j z j , 1 ≤ k ≤ n.
j=n−(k+1)
123
Constr Approx
Hence, for 1 ≤ k ≤ 2n − 1,
1 2(n + 1)z 2n−1 Pk (J (z)) 1 p(z) q(z)
dz = dz + dz
2πi C Fn (z)G n (z) 2πi C Fn (z) C G n (z)
p(z) p(z)/cn Rj 2n
= = ,
Fn (z) Fn (z)/cn z − zj
j=1
2n
1 p(z) 1
dz = 2πi Rj.
2πi C Fn (z) 2πi
j=1
1 2(n + 1)z 2n−1 Pk (J (z))
dz = 0, 1 ≤ k ≤ 2n − 1.
2πi C Fn (z)G n (z)
2n−3 2n−1
However, A(n)
n (z) = a j z j has a z −1 , while Bn(n) (z) = b j z j is still a
j=−1 j=+1
Bn(n) (z)
polynomial, of degree at most 2n−1. Therefore it is still the case that dz = 0
C G n (z)
(the zeros of G n (z) being all outside the unit disk). We need to show that
(n)
An (z) (n)
dz = 0. We write An (z) = q(z) + c/z, where q(z) is a polynomial of
C Fn (z)
degree 2n − 3. Then
(n)
An (z) q(z) 1
dz = dz + c dz.
C Fn (z) C Fn (z) C z Fn (z)
123
Constr Approx
The first integral on the right is zero, as the coefficient in q(z) of z 2n−1 is 0. For the
second integral, decompose
1 1 1 (Fn (z) − Fn (0))/z
c dz = c − dz
C z Fn (z) C Fn (0) z Fn (z)
1 1 (Fn (z) − Fn (0))/z
= dz − dz.
Fn (0) C z C Fn (z)
The first integral is trivially 2πi, while the second is 2πi times the coefficient of z 2n−1
in (Fn (z) − Fn (0))/z divided by the leading coefficient (of z 2n ) in Fn (z), i.e., 2πi × 1.
Hence, indeed,
A(n)
n (z)
dz = 0.
C Fn (z)
Lastly we calculate
1 2(n + 1)z 2n−1 P0 (J (z)) 1 2(n + 1)z 2n−1
dz = dz.
2πi C Fn (z)G n (z) 2πi C Fn (z)G n (z)
2(n + 1)z 2n−1 Pn−n (z) = Cn(n) (z)G n (z) + Dn(n) (z)Fn (z).
2n−1
2n−1
Cn(n) (z) = ck z k , Dn(n) (z) = dk z k ,
j=−1 j=−1
i.e., are both of the form p(z) + c/z, where p(z) is a polynomial of degree 2n − 1.
(n) (n)
Specifically, we write Cn (z) = p(z) + c/z and Dn (z) = q(z) + d/z. We thus have
the following expression:
(n) (n)
1 2(n + 1)z 2n−1 P0 (z) 1 Cn (z) 1 Dn (z)
dz = dz + dz
2πi C Fn (z)G n (z) 2πi C Fn (z) 2πi C G n (z)
1 p(z) c
= dz + dz
2πi C Fn (z) C z Fn (z)
q(z) d
+ dz + dz .
C G n (z) C zG n (z)
123
Constr Approx
1 p(z) 1 d
calculate dz and dz. But the first of these is just the
2πi C Fn (z) 2πi C zG n (z)
(n)
(leading) coefficient of z 2n−1 in p(z), i.e., in Cn (z) divided by the leading coefficient
of Fn (z). From Lemma 4, we have that
2n 2n
Fn (z) = 2−2n (2n + 1) z + ··· ,
n
2n 2n
and it is easy to verify by induction that also Cn(n) (z) =2 −2n
(2n + 1) z + · · ·.
n
Hence
1 p(z)
dz = 1.
2πi C Fn (z)
The second integral above is zero, as all the zeros of G n (z) are outside the closed unit
disk. The first integral is trivially d/G n (0). From Lemma 4 parts 4 and 6, we have
that
2n
G n (0) = (2n + 1)Fn (0) = 2−2n (2n + 1) ,
n
whereas the coefficient of z −1 in Bn(n) (z) is easily verified by induction to have the
same value. Hence
1 d
dz = 1,
2πi C zG n (z)
as claimed.
References
1. Bos, L., Ware, A.F.: On the uniqueness of an orthogonality property of the Legendre polynomials (2015)
(preprint)
2. Cohen, A., Davenport, M.A., Leviatan, D.: On the stability and accuracy of least squares approximations.
Found. Comput. Math. 13(5), 819–834 (2013)
3. Freud, G.: Orthogonal Polynomials. Pergamon Press/Akadémiai Kiadó, Budapest (1971)
123
Constr Approx
4. Lubinsky, D.: Applications of new Geronimus type identities for real orthogonal polynomials. Proc.
Am. Math. Soc. 138(6), 2125–2134 (2010)
5. Szegö, G.: Orthogonal Polynomials, 4th edn. American Mathematical Society, Providence (1975)
123