A problem of Linnik for elliptic curves and mean-value estimates for automorphic representations. Appendix: Recovering modular forms from squares by Dinakar Ramakrishnan (Q1960754)

We cite from the authors' excellent introduction: When Linnik introduced the classical large-sieve in 1941, he was motivated by the following problem: given a nontrivial pimitive character \(\chi\) modulo \(q\), how large (compared to \(q\)) can the first \(n\) be such that \(\chi(n)\neq 1\)? From the Riemann Hypothesis one can deduce \(n\ll (\log q)^2\), and the (weaker) conjecture \(n\ll q^\varepsilon\) for all \(\varepsilon> 0\) is known as Vinogradov's conjecture. Linnik's technique makes it possible to prove that the number of exceptions to these conjectures is extremely small. For example, let \(N(Q,\alpha)\) be the number of pimitive characters \(\chi\) of modulus \(q\leq Q\) such that \(\chi(n)=1\) for all \(n\leq (\log Q)^\alpha\), \((n,q)= 1\), \(\alpha> 1\) being given; then from the large-sieve inequality for Dirichlet characters, we can derive \[ N(Q,\alpha)\ll Q^{2/\alpha+ \varepsilon} \tag{1} \] for all \(\varepsilon> 0\), whereas there are about \(Q^2\) primitive characters of modulus at most \(Q^1\). Moreover, because the exponent is less than 1 for \(\alpha> 2\) and there are about \(Q\) real characters of modulus less than \(Q\), our statement also proves that there are very few exceptions for real characters, which corresponds to the problem of the least quadratic non-residue. In particular, for \(Q\) tending to infinity, the probability that two real characters take the same values for all primes \(p\leq (\log Q)^\alpha\) tends to zero. In recent years, it has been widely perceived that elliptic curves over \(\mathbb{Q}\) are a natural analogue of real Dirichlet characters. In this context, the corresponding problem would be, given two elliptic curves \(E\) and \(F\) of conductor less than \(Q\), how large (always compared to \(Q\)) can \(n\) be if \(E\) and \(F\) have the same number of points modulo \(p\) for all primes less than \(n\) and yet are not isogenous? This problem was considered by \textit{J.-P. Serre} [Publ. Math., Inst. Hautes Étud. Sci. 54, 123--202 (1981; Zbl 0496.12011)]. Assuming the Riemann Hypothesis for Artin \(L\)-functions, he showed that in this case too \(n\ll (\log Q)^2\) follows. In this paper, we are able to prove some analogues of (1). Theorem 1: Let \(M(Q,\alpha)\) be the maximal number of isogeny classes of semistable elliptic curves over \(\mathbb{Q}\) with conductor less than or equal to \(Q\) which for every prime \(p\leq (\log Q)^\alpha\) have a fixed number of points modulo \(p\). Then we have for any \(\varepsilon> 0\) \[ M(Q,\alpha)\ll Q^{8/\alpha+ \varepsilon}. \] It follows from this and a lower bound for the number of isogeny classes of semi-stable elliptic curves with conductor less than \(Q\) that the probability that two such elliptic curves have this property tends to zero as \(Q\) tends to infinity, if \(\alpha\) is large enough. We also have other results in more general cases. As in Linnik's original treatment, we attack the problem by means of an analytic inequality for a larger class of objects encompassing the elliptic curves, namely holomorphic cusp-forms of weight two. Thus we use the Theorem of Wiles, and its further expansions, which prove the modularity of many elliptic curves over \(\mathbb{Q}\), to embed the set of isogeny classes of modular elliptic curves over \(\mathbb{Q}\) in the set of primitive cusp forms. However, due to incomplete knowledge of lower bounds for the Fourier coefficients of cusp forms, it appears that this inequality is not sufficient to prove the result we are seeking. We have to supplement its use by that of another similar inequality for the coefficients of the symmetric square \(L\)-function of cusp forms, and also appeal to a result of Ramakrishnan about the possible multiplicity of the symmetric square, the proof of which appears in an Appendix to this paper. This other inequality requires the study of Rankin-Selberg convolutions of \(\text{GL}(3)\) automorphic forms, and it is actually not much harder to prove a generalization of our mean-value estimate to all \(\text{GL}(n)\), in the context of automorphic representations satisfying the Ramanujan-Petersson bound. The result is related to the large-sieve, although it is not as powerful as one could expect; roughly it corresponds to the case of sums much longer than the conductor of the forms appearing. We first study the Linnik problem for primitive cusp forms, stating the analytic results required for the proof before showing how our main result follows from this. We then apply the theorem to elliptic curves, with a short preliminary discussion of estimates for the number of isogeny classes of elliptic curves over \(\mathbb{Q}\) with conductor at most \(Q\). Both upper and lower bound are used in our theorem. The problem of finding an upper bound for a fixed level \(q\) was recently considered by \textit{A. Brumer} and \textit{J. Silverman} [Manuscr. Math. 91, 95--102 (1996; Zbl 0868.11029)]. Their individual bound can be strengthened on average, and we show how this is done. It is then time to come back to the proof of the mean-value estimate, in its full generality for \(\text{GL}(n)\). This result may be of independent interest, although it falls short of the hypothetical large-sieve inequalities which can be expected by analogy with the case of Dirichlet characters (namely, the case of \(\text{GL}(1)\)). A variant for Maass forms is used by \textit{W. Luo} in [J. Reine Angew. Math. 506, 215--235 (1999; Zbl 0969.11018)].