Statement
Setting (p. 263, Nr. 1). For real nodes −1≤xn<xn−1<⋯<x2<x1≤1,
lk(x) is the Lagrange fundamental function of the node xk: the
polynomial of degree exactly n−1 equal to 1 at xk and 0 at the other
nodes. Pm denotes the m-th Legendre polynomial.
Main theorem (pp. 264--265, formulas (6)--(9'); the paper gives it no
number and calls it a theorem of Chebyshev type, p. 265).
Formula (6): over all node sets −1≤xn<⋯<x1≤1, the least value
of max−1≤x≤1{(l1(x))2+⋯+(ln(x))2} is 1:
−1≤xn<xn−1<⋯<x1≤1min −1≤x≤1max{(l1(x))2+⋯+(ln(x))2}=1.
Formulas (7)--(8): for n≥2 the only node set x1,…,xn with
max−1≤x≤1{(l1(x))2+⋯+(ln(x))2}=1 is the set of the n
roots of
(1−x2)Pn−1′(x)=0,
where Pn−1′ is the derivative in x of the (n−1)-st Legendre
polynomial. These are the endpoints ±1 and the n−2 zeros of
Pn−1′.
Formulas (9) and (9'): for this node set and every value of x,
(l1(x))2+⋯+(ln(x))2=1−n(n−1)(1−x2)(Pn−1′(x))2(n=2,3,…),
or, with x=cosθ, the sum equals
1−n(n−1)1(dθdPn−1)2.
Equivalent descriptions of the nodes (p. 267, (27)--(29)): the roots of
∫−1xPn−1(t)dt=0, of (1−x2)Pn−1′(x)=0, or of
Pn(x)−Pn−2(x)=0. Footnote 6 (p. 267) identifies them as the zeros of
the paper's Jacobi polynomial Jn(0,0,x), in a parametrization where
Jn(α,β,x) satisfies
(1−x2)ω′′+[2(α−β)−2(α+β)x]ω′+n[n+2(α+β)−1]ω=0,
and records
Jn(0,0,x)=(1−x2)Pn−1′(x)=−n(n−1)∫−1xPn−1(t)dt=−2n−1n(n−1)(Pn(x)−Pn−2(x)).
Further forms of the sum (p. 269, (45) and (47)), for the same nodes:
k=1∑n(lk(x))2=1−1−x2n(n−1)(∫−1xPn−1(t)dt)2=1−(2n−1)2n(n−1)1−x21(Pn(x)−Pn−2(x))2.
Limit (p. 270, (49)), stated as following easily from (45)--(48): for
these nodes, limn→∞∑k=1n(lk(x))2=1 for
−1≤x≤1, uniformly on every interval −1+ε≤x≤1−ε
with ε>0, and not uniformly on the whole interval −1≤x≤1.
Proof pointer
Pp. 265--269 (Nr. 3--6). The sum equals 1 at every node, which gives the
lower bound (11). If the sum is at most 1 on [−1,1], then
∣lk(x)∣≤1 there, so each interior node is a maximum point of its own
lk and lk′(xk)=0. With ω(x)=∏k(x−xk) this reads
ω′′(xk)=0 at the interior nodes, which forces x1=1, xn=−1 and
the differential equation (1−x2)ω′′+n(n−1)ω=0; its polynomial
solution vanishing at −1 is a multiple of ∫−1xPn−1 (pp.
265--267). Conversely, setting all values to 1 in Hermite's step-parabola
interpolation gives the identity ∑kvk(x)(lk(x))2≡1 with
vk(x)=1−ω′(xk)ω′′(xk)(x−xk) (p. 268, (33)--(34)).
For these nodes vk≡1 for 2≤k≤n−1, while
v1(x)=1+2n(n−1)(1−x) and vn(x)=1+2n(n−1)(1+x) are at
least 1 on [−1,1], so the sum is at most 1 there (p. 268, (39)--(40)).
The same identity yields the closed forms (45)--(48) (p. 269).
Read depth
Claims checked: the statement, (6)--(9'), (27)--(29), footnote 6, (45),
(47) and (49) were read clause by clause on the page images of the print,
and the proof of Nr. 3--6 was followed. The paper gives no proof of (49).
A second reader checked the statement, hypotheses, label and page against
the print; the proof was not independently reviewed.
Dependencies
None in the corpus. External inputs: Legendre's differential equation and
Hermite's step-parabola interpolation formula, (30)--(32), both standard.
Source. L. Fejér, Bestimmung derjenigen Abszissen eines Intervalles, für
welche die Quadratsumme der Grundfunktionen der Lagrangeschen Interpolation im
Intervalle ein Möglichst kleines Maximum Besitzt, Ann. Scuola Norm. Sup. Pisa
Cl. Sci. (2) 1 (1932), no. 3, 263--276; the edition read is named on the
source card.
Bears on
- Problem 1131: the theorem
minimizes the maximum of ∑k(lk(x))2 on [−1,1], not the integral
I the problem asks about, and the paper says nothing about the least
value of I. Its extremal nodes are the roots of the integral of the
Legendre polynomial that the problem page names. Integrating (9) with
∫−11(1−x2)(Pm′(x))2dx=2m+12m(m+1) at m=n−1 gives
I=2−2n−12 for these nodes, the upper bound the problem page
records from Erdős, Szabados, Varma and Vértesi; that integration is made
here, not in the paper.