Wiki
Wiki

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

Updated


Source. The remark on printed pp. 377–378, including equation (7), and the introductory comparisons on p. 375 (PDF pp. 1–3). This is a complete rewritten proof of the comparison. Attribution of the earlier condition to Selfridge is as reported by Berger, Felzenbaum and Fraenkel through their reference to Churchhouse; that separate historical source has not been compiled here.

The power-series inequality

For 0≤ri<10\le r_i<1, put D=∏i(1−ri)D=\prod_i(1-r_i) and R=∑iriR=\sum_i r_i. Then

2−RD≤2+∑iri1−ri.(1)\frac{2-R}{D}\le 2+\sum_i\frac{r_i}{1-r_i}. \tag{1}

To prove this, expand the finite product of the absolutely convergent geometric series ∏i∑k≥0rik\prod_i\sum_{k\ge0}r_i^k. For a monomial r1k1⋯rnknr_1^{k_1}\cdots r_n^{k_n}, let jj be the number of its positive exponents. Its coefficient on the left of (1) is 2−j2-j: multiplying by −ri-r_i subtracts one exactly when ki>0k_i>0. On the right, the constant coefficient is 22, a monomial supported on one coordinate has coefficient 11, and every other monomial has coefficient 00. The coefficients agree for j=0,1,2j=0,1,2 and satisfy 2−j≤02-j\le0 for j≥3j\ge3. All monomials are nonnegative, so summing proves (1). Absolute convergence justifies the coefficient comparison even when some ri=0r_i=0.

Let xi=ri/(1−ri)x_i=r_i/(1-r_i) and F(x)=∏i(1+xi)−∑ixiF(x)=\prod_i(1+x_i)-\sum_i x_i. Equation (1) gives

F(x)−2=1D−∑iri1−ri−2≤R−1D.(2)F(x)-2 =\frac1D-\sum_i\frac{r_i}{1-r_i}-2 \le\frac{R-1}{D}. \tag{2}

Finite and limiting covering conditions

For an odd integer N=∏ipisiN=\prod_i p_i^{s_i}, take

ri=∑j=1sipi−j.r_i=\sum_{j=1}^{s_i}p_i^{-j}.

Then 0<ri<10<r_i<1 and xi=ri/(1−ri)x_i=r_i/(1-r_i) is exactly the parameter in the product-set theorem. Thus ∑i,jpi−j<1\sum_{i,j}p_i^{-j}<1 forces F(x)<2F(x)<2 and excludes a cover with distinct cardinalities. Equivalently, such a cover requires ∑i,jpi−j≥1\sum_{i,j}p_i^{-j}\ge1. Passing to the strictly larger infinite geometric sums shows that an integer covering with these odd prime divisors must satisfy

∑i1pi−1>1.(3)\sum_i\frac1{p_i-1}>1. \tag{3}

One can compare the exponent-free conditions directly as well. In (2) take ri=1/(pi−1)r_i=1/(p_i-1), so xi=1/(pi−2)x_i=1/(p_i-2) and D−1=∏i(pi−1)/(pi−2)D^{-1}=\prod_i(p_i-1)/(p_i-2). Then

∏ipi−1pi−2−∑i1pi−2−2≤∏ipi−1pi−2(∑i1pi−1−1).(4)\begin{aligned} &\prod_i\frac{p_i-1}{p_i-2}-\sum_i\frac1{p_i-2}-2\\ &\qquad\le \prod_i\frac{p_i-1}{p_i-2} \left(\sum_i\frac1{p_i-1}-1\right). \tag{4} \end{aligned}

The new necessary condition makes the left side positive, so it implies (3). Finally, for nonnegative ui=1/(pi−1)u_i=1/(p_i-1), ∏i(1+ui)≥1+∑iui\prod_i(1+u_i)\ge1+\sum_i u_i. Thus (3) implies the weaker direct-density condition

∏ipipi−1>2.(5)\prod_i\frac{p_i}{p_i-1}>2. \tag{5}

For completeness, the latter condition also follows directly from coverage: each distinct modulus is a distinct divisor m>1m>1 of NN, so the density union bound gives

1≤∑m used1m≤∑m∣N, m>11m=∏i(∑j=0sipi−j)−1<∏ipipi−1−1.1\le\sum_{m\text{ used}}\frac1m \le\sum_{m\mid N,\,m>1}\frac1m =\prod_i\left(\sum_{j=0}^{s_i}p_i^{-j}\right)-1 <\prod_i\frac{p_i}{p_i-1}-1.

The strict last inequality uses finite positive exponents. These implications compare necessary conditions only; none is sufficient to construct a cover.

Bears on. The hierarchy of obstructions for Problem 7.