SUMS OF DISTANCES BETWEEN POINTS OF A SPHERE

Given N points on a unit sphere in k + 1 dimensional Euclidean space, we obtain an upper bound for the sum of all the distances they determine which improves upon earlier work by K. B. Stolarsky when k is even. We use his method, but derive a variant of W. M. Schmidt's results for the discrepancy of spherical caps which is more suited to the present application.


INTRODUCTION.
In this paper we shall consider the following interesting problem of a geometrical nature: Given a function which measures the distance between points on the unit spher in m-dimensional space, for what set of N points on the spher is the sum of all distances between points a maximum, and what is the maximum?K. B. Stolarsky made important progress towards the solution of this problem in [3].
We efer the rader to his paper for the history of earlier work.He showed (see Theorem 2 below), for a large class of distance measuring functions, that the sum of the distances between points plus a measure of how far the set of points deviates from unifor distribution is equal to N 2 c(f,m).Here c is a constant depending G. HARMAN only on the function f used to measure the distances, and the dimension m of the space.The sum of distances is lhus maximised by a well distributed set of points.
By examining carefully the estimation of the discrepancy of the point distribution we shall obtain results which, in certain cases, are very near to being best possible.The best possible result in this context is that the maximum sum of distances between N points is c(f,m)N 2 h(f,m)N l-I/(m-l).
We now introduce some notation following [3] to make the above statements more precise.Write U for the surface of the unit sphere in k + 1 dimensional Euclidean space E k + 1 Let M be a sequence of N points PI' "'''PN e U.For where the maximum in (2) is taken over all sequences M K. B. Stolarsky has shown P (see Theorem 2 below) that the sum in (i) plus an integral which measures the dis- crepancy of M equals a constant depending only on d, when d belongs to a P certain class of functions.This class includes the usual Euclidean distance function We write Po for the vector (I,0,...0) in E k +I vector p e U we write for points pl,P2 e U, whenever the integral in (3) exists (i.e. when the inner integral is intergrable over the orthogonal group).Here p.q denote the standard inner poduct of two vectors.We call g the kernel of Let o(U) denote the surface area of U Write do(ql,q2) for the great circle metric and dl(ql,q2) for the great circle metric defined by (3) with kernel -1/2 It is shown in [3] that dl(ql,q2) <__ do(ql,q2) Henceforth d(ql,q2) will be the usual Euclidean metric.Our main result is as follows:

"
The main result of Stolarsky's paper which we use is THEOREM 2 (Stolarsky).We have S(o;N,k,Mp) + g(x) (f(Mp,T,x) -No (x)) 1 dTdx N2 II p(p,q)do(q).2o(U) 2 Here f(M ,T,X) is the number of points of M in the spherical ap: (5) (6) {p e U: p. XPo <= x}.Also *(x) denotes the normalized surface aea of a spherical cap, radius x, and d(q) represents an element of surface area on U The second temm on the left of (6) is clearly a measure of the discrepancy of M P In the second section we shall show how to obtain an estimate for a related measure, and in our section 3 we shall prove Theorem i.

SCHMIDT'S INTEGRAL EQUATION METHOD
In the following, constants implied by the << notation shall depend only on k.We will make two changes to Schmidt's method.At one stage (p.69 of [i]) for k even he uses the inequality 1 + log(l/r) << r -I which suffices to prove his results, but which is wasteful in our pesent context.This produces the improve- ment in the exponent of N over Stolarsky's results.To improve N e to a power of log N we do not allow constants introduced to depend on a parameter e.This enables us to choose as a function of N, whereas Schmidt could only choose as a function of k and e As in [i] and [3] (section 4) we let be the nornalized Lebesgue measure on U (so (u) =i).We write C(r,p) for the spherical cap of all points on U whose spherical distance from p e U is no more than r.Put (C(r,p)) for the number of points of M in C(r,p) and D(r,p) N(C(r,p)) (C(r,p)).

P
We write E(r,s) The main result of this section THEOREM 3. We have, for N >_ k even, that 0 (log(N k) + log(6/r)) E(r,r)dr >> N l-I/k 6k (8) We note here that all the results in [i] may be improved by the replacing of (Nk) e with log(N6 k), or (log(N6 k))1/2 That these results are very ncar to being best possible is shown in [3].
The idea of Schmidt's method as we use it here is to show that for 0 < < i, for the right hand side of (9) (see p.82 of [I]).The choice of as 1-(log Nk) -I then gives (8).
We now outline how to obtain (9), referring to [i] (the reader may also follow the argument using [2]).Put 8 as r + O, uniformly in 6.The tracing of the dependence of all the constants in [i] on e is tedious but straightforward.They are all bounded above and below by positive constants independent of (this is not true for k odd, but in that case we get fo(r) >< (i e)-ir which is good enough to prove the corresponding result).We see that (9) follows from (ii), (12) and (13), and the proof of the theorem is complete.

PROOF OF THEOREM i
The lower bounds in (4) and (5) are established in [3]; we have included them here for completeness.

For a function f of
a I (Tp) d for the integral of f over the special orthogonal group acting on U.Here T represents an orthogonal transformation.For a positive function g of a real variable x [0,I] we define a distance function (which can be shown to be a metric) by P(Pl,P2 P(Pl,P2;g)=I[
is a suitable small constant.The use of the apparently trivial inequality E(r,r) >> llN(C(p,r))II 2 then gives a lower bound (l-)k+-2 N l+e/k-2/k log(l/r)) E (6r,6r)dr -i + log(6/r)) E(r,r)dr.Now, by Lemma 6 of [i], r) f(r)dr where f satisfies a certain integral equation.It is shown in [i], pages 79