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), unnumbered corollary, printed p. 899 (published scan). The printed strict inequality is corrected below.

Statement. Let N≥1N\ge1, let r≥1r\ge1 be an integer, and let x1,…,xNx_1,\ldots,x_N be real with ∣xi∣≥1|x_i|\ge1. The number of assignments whose signed sum belongs to an open interval of length 2r2r is at most

rBN=r(N⌊N/2⌋).rB_N=r\binom N{\lfloor N/2\rfloor}.

Proof. Write the target interval as (u,u+2r)(u,u+2r). It is contained in the disjoint union

⋃j=0r−1[u+2j,u+2j+2).\bigcup_{j=0}^{r-1}[u+2j,u+2j+2).

Each piece is a half-open interval of length two and hence contains at most BNB_N assignments by Theorem 1. Summing over the rr pieces proves the bound. The possible extra endpoint uu can only enlarge the counted set; every internal division point is included in exactly one piece. □\square

Printed endpoint correction. The source says the count is less than rBNrB_N. For r=1r=1, even N≥2N\ge2 and x1=⋯=xN=1x_1=\cdots=x_N=1, exactly BNB_N assignments have sum in (−1,1)(-1,1). The weak inequality is therefore necessary. This is a correction supplied by the compilation, not a cited author erratum. Radius zero and negative integers are not part of the geometric statement.

The sharper Theorem 3 replaces this coarse multiple by the sum of the rr largest binomial coefficients, truncated when r>N+1r>N+1.

Bears on. Problem 498: an input to the order bound of Theorem 2, not the problem's exact bound.