The inverse Goldbach problem (Q4435534)

Let \({\mathcal A}\) and \({\mathcal B}\) be sets of natural numbers, \({\mathcal P}\) be the set of all prime numbers, and define \({\mathcal A}+{\mathcal B}=\{a+b:\;a\in{\mathcal A},\;b\in{\mathcal B}\}\). Also write \(A(x)=| {\mathcal A}\cap[1,x]| \) and \(B(x)=| {\mathcal B}\cap[1,x]| \), where \(| {\mathcal S}| \) means the cardinality of a set \({\mathcal S}\), and \(x\) is to be regarded as a sufficiently large real number. Now suppose that \(| {\mathcal A}| \geq2\) and \(| {\mathcal B}| \geq2\), and that \({\mathcal A}+{\mathcal B}\) coincides with \({\mathcal P}\) except only for finitely many elements, that is, one has \(({\mathcal A}+{\mathcal B})\cap[x_0,\infty)={\mathcal P}\cap[x_0,\infty)\) for some \(x_0\). Then, it is proved in this paper that \[ x^{1/2}(\log x)^{-5}\ll A(x)\ll x^{1/2}(\log x)^4, \] and the same for \(B(x)\). The previously best result in this direction, which was obtained by the same author very recently, was even weaker than bounds of the form \(x^{\varepsilon}\ll A(x)\ll x^{1-\varepsilon}\), for any fixed \(\varepsilon>0\), so astonishing progress is made in this short article. The proof of this remarkable theorem is, however, based on a rather simple idea, and the structure of the proof may be summarized as follows. For each large \(a\in{\mathcal A}\), one finds that \(a+b\) must be prime for all \(b\in{\mathcal B}\), so that \(a\not\equiv-b\) (mod \(p\)) for all smaller primes \(p\). Therefore, on writing \(\nu(p)\) for the number of residue classes \(b\) (mod \(p\)) for \(b\in{\mathcal B}\cap[1,x]\), large elements of \({\mathcal A}\) are forbidden to belong to \(\nu(p)\) residue classes modulo \(p\). Then the new methods of the author come in. If \(\nu(p)\) would be ``not so large'' for ``many'' primes \(p\), then \(B(x)\) is small by Gallagher's large sieve, but \(B(x)\) cannot be too small in view of an existing lower bound for it. So \(\nu(p)\) must be ``large'' for ``many'' \(p\), and then the above sharp upper bound for \(A(x)\) is shown via Montgomery's large sieve. The remaining bounds follow immediately by symmetry and the trivial bound \(A(x)B(x)\gg x/\log x\). In combination with a known result, the above theorem implies solution of the inverse ternary Goldbach problem. In fact, it is deduced from the theorem that there do not exist sets \({\mathcal A}\), \({\mathcal B}\) and \({\mathcal C}\) of natural numbers with all of \(| {\mathcal A}| \), \(| {\mathcal B}| \), \(| {\mathcal C}| \) exceeding 1, such that the set \({\mathcal A}+{\mathcal B}+{\mathcal C}\) coincides with \({\mathcal P}\) except only for finitely many elements.