Scribd
Upload
38 views5 pages

On A Lemma of Littlewood and Offord: P. Erdös

This document summarizes several theorems and conjectures regarding the number of sums of complex or real numbers that fall within certain intervals or circles. It begins by improving upon a previous lemma of Littlewood and Offord, showing that the number of such sums falling within a circle of radius r is less than cr2nn-1/2. It then proves several related theorems, such as bounds on the number of sums falling within intervals of given lengths. It concludes by conjecturing generalized versions for Hilbert spaces and suggesting stronger forms that may be easier to prove.

Uploaded by

johnoftheroad
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
38 views5 pages

On A Lemma of Littlewood and Offord: P. Erdös

This document summarizes several theorems and conjectures regarding the number of sums of complex or real numbers that fall within certain intervals or circles. It begins by improving upon a previous lemma of Littlewood and Offord, showing that the number of such sums falling within a circle of radius r is less than cr2nn-1/2. It then proves several related theorems, such as bounds on the number of sums falling within intervals of given lengths. It concludes by conjecturing generalized versions for Hilbert spaces and suggesting stronger forms that may be easier to prove.

Uploaded by

johnoftheroad
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

ON A LEMMA OF LITTLEWOOD AND OFFORD

P. ERDÖS

Recently Littlewood and Offord1 proved the following lemma:


Let be complex numbers with \xi\ ^ 1 . Consider the
sums ]C*=i€&#fc> where the €& are ± 1 . Then the number of the sums
]Cft=i€&x* which fall into a circle of radius r is not greater than
cr2n(log n)nr112.
In the present paper we are going to improve this to
cr2nn~112.
The case Xi = 1 shows that the result is best possible as far as the order
is concerned.
First we prove the following theorem.
T H E O R E M 1. Let be n real numbers, \xi\ ^ 1. Then the
number of sums ]C2=i€&#& which f all in the interior of an arbitrary in-
terval I of length 2 does not exceed Cn,m where m= [n/2]. ([x] denotes
the integral part of x.)
Remark. Choose Xi = l, n even. Then the interval ( — 1, + 1 ) con-
tains Cn,m s u m s ^ i e ^ , which shows that our theorem is best pos-
sible.
We clearly can assume that all the Xi are not less than 1. To every
sum ]Qfc=i€fcxfc w e associate a subset of the integers from 1 to n as
follows: k belongs to the subset if and only if e&= + 1 . If two sums
2^J»i€^jb and ]Cfc=i*& #& are both in 7, neither of the corresponding
subsets can contain the other, for otherwise their difference would
clearly be not less than 2. Now a theorem of Sperner 2 states that in
any collection of subsets of n elements such that of every pair of sub-
sets neither contains the other, the number of sets is not greater than
Cn,m, and this completes the proof.
An analogous theorem probably holds if the Xi are complex num-
bers, or perhaps even vectors in Hubert space (possibly even in a
Banach space). Thus we can formulate the following conjecture.
CONJECTURE. Let xi, #2, • • • , xn be n vectors in Hilbert space,
||#;|| è l . Then the number of sums 22=i€fc#& which f all in the interior
of an arbitrary sphere of radius 1 does not exceed Cn,m.
Received by the editors March 28, 1945.
1
Rec. Math. (Mat. Sbornik) N.S. vol. 12 (1943) pp. 277-285.
2
Math. Zeit. vol. 27 (1928) pp. 544-548.
898
A LEMMA OF LITTLEWOOD AND OFFORD 899

At present we can not prove this, in fact we can not even prove that
the number of sums falling in the interior of any sphere of radius 1
iso(2n).
From Theorem 1 we immediately obtain the following corollary.
COROLLARY. Let r be any integer. Then the number of sums ]Qb-i € ***
which fall in the interior of any interval of length 2r is less than rCn,m.
T H E O R E M 2. Let the Xi be complex numbers, \x%\ j=£l. Then the num-
ber of sums X)i=i€&#fc which fall in the interior of an arbitrary circle of
radius r (r integer) is less than

crCn,m < ci^n-1'2.


We can clearly assume that at least half of the Xi have real parts
not less than 1/2. Let us denote them by xi, x2, • • • , xt, t*tn/2. In
the s u m s ^ î . ^ ^ we fix et+h • • • , en. Thus we get 2 ' sums. Since
we fixed et+u * • • > *nijL\;-i*kXk has to fall in the interior of a circle
of radius r. But then 2^kw.i€hR{xh) has to fall in the interior of an in-
terval of length 2r (R(x) denotes the real part of x). But by the corol-
lary the number of these sums is less than

crCt.um < 0*2*1$**.


Thus the total number of sums which fall in the interior of a circle
of radius r is less than
c2r2n/n1'2t
which completes the proof.
Our corollary to Theorem 1 is not best possible. We prove:
T H E O R E M 3. Let r be any integer, the Xi real, \xi\ ^ 1. Then the num-
ber of sums X ) 2 - I € * * A which fall into the interior of any interval of length
2r is not greater than the sum of the r greatest binomial coefficients {be-
longing to n).
Clearly by choosing Xi = 1 we see that this theorem is best possible.
The same argument as used in Theorem 1 shows that Theorem 3
will be an immediate consequence of the following theorem.
T H E O R E M 4. Let Ai, A2, • • • , Au be subsets of n elements such that
no two subsets Ai and A3- satisfy AOAj and Ai—Aj contains more
than r — 1 elements. Then u is not greater than the sum of the r largest
binomial coefficients.
Let us assume for sake of simplicity that n — 2m is even and
r = 2j+l is odd. Then we have to prove that
900 P. ERDÖS [December

+3

Our proof will be very similar to that of Sperner.2 Let A i,A 2, • • -,AU
be a set of subsets which have the required property and for which
u is maximal. It will suffice to show that in every A the number of
elements is between n—j and n+j. Suppose this were not so, then
by replacing if need be each A by its complement we can assume that
there exist A's having less than n—j elements. Consider the -4's with
fewest elements; let the number of their elements be n—j--y and let
there be x A's with this property. Denote these A's by A i,A 2, • • • Ax.
To each Ai, i = l, 2, • • • , x, add in all possible ways r elements from
the n+j+y elements not contained in A, We clearly can do this in
Cn+i+y,r ways. Thus we obtain the sets Bi9 B% • • • , each having
n+j—y+1 elements. Clearly each set can occur at most Cn+j-v+ur
times among the B's. Thus the number of different B's is not less than

Hence if we replace Ait A^ • • • , Ax by the B's and leave the other


A's unchanged we get a system of sets which clearly satisfies our
conditions (the B's contain n+j—y+1 elements and all the A's now
contain more than n—j—y elements, thus B—A can not contain more
than r — 1 elements and also B<X.A) and has more than u elements,
this contradiction completes our proof.
By more complicated arguments we can prove the following theo-
rem.
THEOREM 5. Let Ai, A^ • • • , Au be subsets of n elements such that
there does not exist a sequence of r+1 A's each containing the previous
one. Then u is not greater than the sum of the r largest binomial coeffi-
cients.
As in Theorem 4 assume that n — 2mf r = 2j+l, and that there are
x A's with fewest elements, and the number of their elements is
n—j—y. We now define a graph as follows: The vertices of our graph
are the subsets containing z elements, n—j—y^z^n+j+y. Two ver-
tices are connected if and only if one vertex represents a set containing
z elements, the other a set containing z+1 elements, and the latter
set contains the former. Next we prove the following lemma.
LEMMA. There exist C2n,n-/~y disjoint paths connecting the vertices
containing n—j—y elements to the vertices containing n+j+y elements.
Our lemma will be an easy consequence of the following theorem
Ï945] A LEMMA OF LITTLEWOOD AND OFFORD 901

of Menger: 3 Let G ie any graph, V\ and V2 two disjoint sets of its ver-
tices. Assume that the minimum number of points needed for the separa-
tion of V\ and V2 is w. Then there exist w disjoint paths connecting Vi
and V2. {A set of points w is said to separate V\ and V2, if any path
connecting V\ with V2 passes through a point of w.)
Hence the proof of our lemma will be completed if we can show
that the vertices Vi containing n— j—y elements can not be separated
from the vertices V2 containing n+j+y elements by less than
C2n,n-3-v vertices. A simple computation shows that Vi and V2 are
connected by

C2n,n~j-y(n + j+ y)(n + j+ y — 1) • *- (n — j — y+ 1)
paths. Let z be any vertex containing n+i elements, —j—yi^i ûj+y-
A simple calculation shows the the number of paths connecting V\
and V2 which go through z equals
(n+i)(n+i—l) • • • (n— j — y+l)(n— i)(»—i—1) • • • (n—j—y+1)
^•(n+j+y)(n+j+y-l) • • • (n—j—y+l).
Thus we immediately obtain that Vi and V2 can not be separated by
less than C2n,n-j~y vertices, and this completes the proof of our lemma.
Let now -4i (1) , ^ 2 ( 1 ) , • • • , Axa) be the A1 s containing n~j—y ele-
ments. By our lemma there exist sets A^l), i = l, 2, • • • , x;
Z = l , 2, • • • , 2j+2y + l, such that Ai<2t+2v+l) has n+j+y elements
and A^l)CA^l+x) and all the A's are different. Clearly not all
the sets A^l), Z = l, 2, • • • , 2j+2y + l, can occur among the
A1, A2, • • • , Au. Let AiM be the first A which does not occur there.
Evidently s^r. Omit Aia) and replace it by A^K Then we get a new
system of sets having also u elements which clearly satisfies our con-
ditions, and where the sets containing fewest elements have more
than n—j—y elements and the sets containing most elements have
not more than n+j+y elements. By repeating the same process we
eventually get a system of ^4's for which the number of elements is
between n—j and n+j. This shows that
+ƒ

which completes the proof.


One more remark about our conjecture: Perhaps it would be easier
to prove it in the following stronger form : Let \xt\ à 1, then the num-
8
See, for example, D. König, Theorie der endlichen und unendlichen Graphen, p. 244.
902 P. ERDÖS

ber of sums 5^2,, !€*#* which fall in the interior of a circle of radius 1
plus one half the number of sums falling on the circumference of the
circle is not greater than Cn,m. If the Xi are real it is quite easy to
prove this.
We state one more conjecture.
(1). Let \xi\ = 1. Then the number of sums 23ï-i€*xfc with
E*-i € ***l ^ 1 is greater than c2nnrl9 c an absolute constant.
UNIVERSITY OF MICHIGAN

You might also like