Wiki
Wiki

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

Updated

Proof of the consecutive unique-difference bound


Let B⊂ZB\subset\mathbb Z have m≥2m\ge2 elements and suppose that every integer 1≤d≤L1\le d\le L has exactly one representation d=b′−bd=b'-b with b,b′∈Bb,b'\in B. If L=0L=0, the bound is immediate. Otherwise put u=⌊(L+1)/2⌋≥1u=\lfloor(L+1)/2\rfloor\ge1, so 2u−1≤L≤2u2u-1\le L\le2u.

For every integer jj define

F(j)=∣B∩[j,j+u−1]∣,r(d)=∣{(b,b′)∈B2:b′−b=d}∣.F(j)=|B\cap[j,j+u-1]|, \qquad r(d)=|\{(b,b')\in B^2:b'-b=d\}|.

All sums below have finite support. Counting the intervals that contain two specified points gives

∑jF(j)2=mu+2∑d=1u−1(u−d)r(d)=u(m+u−1).(1)\sum_j F(j)^2 =mu+2\sum_{d=1}^{u-1}(u-d)r(d) =u(m+u-1). \tag{1}

For a point bb in [j,j+u−1][j,j+u-1] and a point b′b' in the adjacent interval [j+u,j+2u−1][j+u,j+2u-1], the positive difference d=b′−bd=b'-b lies between 11 and 2u−12u-1. There are u−∣u−d∣u-|u-d| choices of jj for this pair. Thus

∑jF(j)F(j+u)=∑d=12u−1(u−∣u−d∣)r(d)=u2.(2)\sum_j F(j)F(j+u) =\sum_{d=1}^{2u-1}(u-|u-d|)r(d)=u^2. \tag{2}

The only representation counts used in (1) and (2) are r(d)=1r(d)=1 for 1≤d≤2u−11\le d\le2u-1. Translation of the index jj then gives

∑j(F(j+u)−F(j))2=2u(m−1).(3)\sum_j\bigl(F(j+u)-F(j)\bigr)^2=2u(m-1). \tag{3}

For each a∈{0,…,u−1}a\in\{0,\ldots,u-1\} let fa(n)=F(a+nu)f_a(n)=F(a+nu) for n∈Zn\in\mathbb Z. These values count BB in a partition of the integers into disjoint intervals, so fa≥0f_a\ge0 and ∑nfa(n)=m\sum_n f_a(n)=m. Write

Ta=∑nfa(n)2,Ea=∑n(fa(n+1)−fa(n))2.T_a=\sum_n f_a(n)^2,\qquad E_a=\sum_n(f_a(n+1)-f_a(n))^2.

The bijection j=a+nuj=a+nu and (1), (3) yield

1u∑aTa=m+u−1,1u∑aEa=2(m−1).(4)\frac1u\sum_a T_a=m+u-1,\qquad \frac1u\sum_a E_a=2(m-1). \tag{4}

Fix a positive integer kk and average kk translates of each sequence:

ga(n)=1k∑i=0k−1fa(n+i).g_a(n)=\frac1k\sum_{i=0}^{k-1}f_a(n+i).

Nonnegativity and the mass identity give 0≤ga(n)≤m/k0\le g_a(n)\le m/k and ∑nga(n)=m\sum_n g_a(n)=m. Therefore ∑nga(n)2≤m2/k\sum_n g_a(n)^2\le m^2/k. Telescoping consecutive translates and applying the triangle inequality in ℓ2(Z)\ell^2(\mathbb Z) give

∥fa−ga∥2≤1k∑i=0k−1∥fa−fa( ⋅+i)∥2≤k−12Ea.\|f_a-g_a\|_2 \le\frac1k\sum_{i=0}^{k-1}\|f_a-f_a(\,\cdot+i)\|_2 \le\frac{k-1}{2}\sqrt{E_a}.

Now use the triangle inequality for arrays indexed by (a,n)(a,n), with squared norm u−1∑a,n∣⋅∣2u^{-1}\sum_{a,n}|\cdot|^2. Equation (4) gives, for every positive integer kk,

m+u−1≤mk+(k−1)m−12.(5)\sqrt{m+u-1} \le\frac{m}{\sqrt k}+(k-1)\sqrt{\frac{m-1}{2}}. \tag{5}

Take k=⌈m1/3⌉k=\lceil m^{1/3}\rceil. Since k≥m1/3k\ge m^{1/3} and k−1≤m1/3k-1\le m^{1/3}, the right side is at most (1+1/2)m5/6(1+1/\sqrt2)m^{5/6}. Squaring and using L≤2uL\le2u proves

L≤2u≤2(1+1/2)2m5/3−2m+2=(3+22)m5/3−2m+2.L\le2u \le2(1+1/\sqrt2)^2m^{5/3}-2m+2 =(3+2\sqrt2)m^{5/3}-2m+2.

This proves the claim. No step bounds the support diameter or uses uniqueness for any difference larger than LL. □\square