Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated


Source. Erdős (1945), Theorem 3, printed p. 899 (published scan).

Statement. Let N≥1N\ge1, let r≥1r\ge1 be an integer, and let x1,…,xNx_1,\ldots,x_N be real numbers with ∣xi∣≥1|x_i|\ge1. The number of sign assignments whose sum lies in any open interval of length 2r2r is at most S(N,r)S(N,r), the sum of the largest min⁡(r,N+1)\min(r,N+1) coefficients of (1+x)N(1+x)^N. The bound is sharp.

Proof. As in Theorem 1, make all inputs positive by reindexing their signs. If AA is the set of positive coordinates, its sum is ZA=2∑i∈Axi−∑ixiZ_A=2\sum_{i\in A}x_i-\sum_i x_i. For A⊊DA\subsetneq D with ∣D∖A∣≥r|D\setminus A|\ge r,

ZD−ZA=2∑i∈D∖Axi≥2r.Z_D-Z_A=2\sum_{i\in D\setminus A}x_i\ge2r.

Such a pair cannot give two sums in the same open interval of length 2r2r. The subsets corresponding to qualifying assignments therefore satisfy Theorem 4, which bounds their number by S(N,r)S(N,r).

For sharpness when 1≤r≤N+11\le r\le N+1, let all xi=1x_i=1 and use

L=⌊N−r+12⌋,U=L+r−1.L=\left\lfloor\frac{N-r+1}{2}\right\rfloor, \qquad U=L+r-1.

The open interval

(2L−N−1,  2U−N+1)(2L-N-1,\;2U-N+1)

has length 2r2r and contains precisely the rank sums 2k−N2k-N with L≤k≤UL\le k\le U. Their assignment multiplicities add to ∑k=LU(Nk)=S(N,r)\sum_{k=L}^U\binom Nk=S(N,r). For r>N+1r>N+1, the interval (−r,r)(-r,r) contains every sum of the all-one inputs, so the truncated value 2N2^N is attained. □\square

Conventions. The integer-radius and truncation conventions are explicit in the notation page. This is the source's sharper replacement for its coarse interval corollary.

Bears on. Problem 498: at r=1r=1 this is Theorem 1, the problem's bound for real inputs.