Beruflich Dokumente
Kultur Dokumente
direct numerical experience with (6), we have had a good deal of experience
with the analogous free boundary problem. In this case, Levi-Civita's
parametrization and use of K = K(t) leads to an integral equation es-
sentially of the form
X(o) = M,(o.)eT(')K(fX(o.) do). (8)
Here v(or) is known and M is a parameter. By letting v(oa) involve other
parameters, the cases of symmetric walls, cusped cavities, Riabouchinsky
flows, etc., can be treated in the same way.
Using 5-25 points on barriers whose angular extent was as large as 1550,
very good approximations were obtained to the curvature, and extremely
good approximations to the profiles of the obstacles. This problem has
already been adapted for explicit computation on the computing machine
at the Institute for Advanced Study.
Generalization of the preceding formulae to regions with corners and to
annular regions can be easily made, but we have not explored their prac-
ticability for accurate computation on high-speed computing machines.
A detailed discussion of the preceding formulae and their arithmetiza-
tions, together with the accuracy which may be expected if they are used,
will appear in a longer article.
*
This work was supported by the Office of Naval Research through Contract N5ori-
76 Project 22.
1 Theodorsen, T., and Garrick, I., "General Potential Theory of Arbitrary Wing
ticular cases of the general theory. Throughout this note f(x) will denote
a real even integrable function of period 2r, a. its Fourier cosine coefficients,
and s.(x) the partial sums of its Fourier series.
THEOREM 1. If f(x) is constant in a neighborhood of 0, each of the in-
equalities sn(O) _ f(O) and sn(O) < f(O) holds for infinitely many values of
n.
Much sharper statements can be made; for example, we shall
establish Theorem 1 by proving the following theorem.
THEOREM 2. Under the hypotheses of Theorem 1, s.(O) - f(O) and
Sn + l(O) - f(O) cannot have the same sign for a sequence of values of n which
have density 1, or even P6lya maximum density 1.
One might expect that in some sense about half of the partial sums would
satisfy each inequality in Theorem 1, but this is not necessarily true.
THEOREM 3. For any E, 0 < f < 1, there is an f(x) satisfying the condi-
tions of Theorem 1 and with Sn(O) < f(O) at most for a sequence of integers of
density e.
If we place less restrictive conditions on f(x) in the neighborhood of 0
we can obtain other results of a similar character, of which the following
theorems are examples.
THEOREM 4. If f(x) is analytic for xI 6, a > 0, and Sn(O) - f(O) has
the same sign as sn + 1(O) - f(O) for all n except for a sequence of zero density,
then an = O(e - E)E > 0.
In other words, it is not possible for an f(x) which is analytic at 0 to
have Sn(O) _ f(0), or even Sn(O) - f(O) of the same sign as Sn + (O) -f(O),
for all n except a sequence of zero density, unless f(x) is analytic for all x.
This shows that the sign of Sn(O) - f(O) is influenced by the behavior of
f(x) outside a neighborhood of 0, so that we are not dealing with a local
phenomenon.
THEOREM 5. If f(x) has a continuous ph derivative for xj . a and
Sn(O) - f(O) has the same sign as sn + l(O) - f(O) for all but a finite number
of n, then an = O(n-P).
The proofs of Theorems 1, 2, 4 and 5 which I shall outline here use the
theory of entire functions; however, A. Zygmund has pointed out to me
that proofs can also be obtained by methods more usual in the theory of
Fourier series. The use of entire functions makes it easy to obtain sharper
theorems when something is known about the size of 5 in Theorems 2, 4
and 5; roughly speaking, the larger 6, the smaller the density of perma-
nences of sign in the sequence { Sn(O) - f(0) } which needs to be assumed.
Furthermore, it is possible to weaken Sn(O) - f(0) _ 0 to Sn(O) - f(0) >
- hn, where { hn} is a sufficiently rapidly decreasing positive sequence.
Our starting point is the usual formula for the partial sums,
In the case of Theorems 1 and 2, f(t) -f(O) = O for tI < 6,6 > O, and so
Sn(0) - f(O) = 7r '-fT[f(t) - f(O)] sin (n + '/2)t csc '/2t dt
= - + 'j [f(r - u) - f(O)]
X cos (n + 1/2)U sec '/2u du.
Consider the function
F(z) = oT-5[f(7r - U)- f(O)] cos (z + '/2)U sec '/2u du.
This is an even entire function of order 1 and type r - 6 <Tr, real and
bounded on the real axis; if Sn(0) - f(O) and Sn + l(O) - f(O) have the same
sign, F(z) changes sign between n and n + 1; thus F(z) has zeros whose
density, or P6lya maximum density, is at least that of the n's for which
Sn(O) - f(O) has the same sign as Sn + ,(O) - f(O). By results of Levinson3
F(z) vanishes identically if this density exceeds 1 - 6/r; in that case
f(- u) = f(O) almost everywhere and the Fourier series of f(u) reduces
to its constant term. This establishes Theorem 2 and hence Theorem 1.
For Theorems 4 and 5 the discussion is similar but somewhat more com-
plicated. We now have
(-1)n[sn(0) -f(O)] = F(n) + (-1)n7r1f);g(t) sin (n + '/2)tdt (*)
where g(t) = [f(t) - f(O)]csc 1/2t. We express the sine in terms of ex-
ponentials and write (-l)n as ein. In the case of Theorem 4, g(t) is
analytic at t = 0, and we can replace integration along the real axis in
(*) by integration around three sides of a small rectangle extending into
the upper or lower half plane, according to which exponential is under
consideration. The integral on the right of (*) then becomes the sum of
two terms which are O(e - ) ee > 0, as z = x -- + co, and of four terms which
are entire functions of order 1 and (small) type e. Thus (-_l)n[Sn(O) - f(0)]
= G(n) + H(n), where G(z) is an entire function of order 1 and type
less than xr, while H(x) = O(ex- ) as x -X + o . If sn(0) -f(O) and Sn + 1(0)
- f(O) have the same sign, it follows that G(Xn) = O(e-Xn) with n < Xn <
n + 1, and with the hypothesis of Theorem 4 we then have4 G(x) = O(e-x)
as x -- c. In particular, G(n) = O(e-") and so Sn(O) - f(O) = O(e-")
and hence an = O(e-`).
The proof of Theorem 5 is much the same except that now we integrate
p times by parts in the integral on the right of (*) and use a theorem of
Duffin and SchaefferP to deduce the order of G(x) from that of G(Xn).
We turn now to the proof of Theorem 3. Let k be an integer greater
than 2, e a positive number less than 7r/(2k). The function
F(z) = z- sin4 z 11(1 -Z2/n2)
where the product extends over all positive integers n not of the form mk
VOL. 37, 1951 MA THEMA TICS: J. M. COOK 417