Failure of the integral Hodge conjecture for threefolds of Kodaira dimension zero (Q2179697)

Let \(X\) be a complex projective manifold. The Hodge conjecture claims that the cycles of codimension \(k\) generate the space of Hodge classes in \(H^{2k}(X, \mathbb Q)\). The stronger \textit{integral} Hodge conjecture claims that this even holds with integer coefficient, i.e. the cycles generate the Hodge classes in \(H^{2k}(X, \mathbb Z)\). By the Lefschetz \((1,1)\)-theorem the integral Hodge conjecture holds for \(k=1\), but Kollár constructed counterexamples for \(k=\dim X-1\) which are threefolds of general type [\textit{E. Ballico} (ed.) et al., Classification of irregular varieties. Minimal models and abelian varieties. Proceedings of a conference, held in Trento, Italy, 17-21 December, 1990. Berlin etc.: Springer-Verlag (1992; Zbl 0744.00029)]. On the other hand \textit{C. Voisin} [Adv. Stud. Pure Math. 45, 43--73 (2006; Zbl 1118.14011)] proved the integral Hodge conjecture for threefolds with trivial canonical bundle such that \(h^{0,2}(X)=0\), so one could speculate that the integral Hodge conjecture holds more generally for threefolds with numerically trivial canonical class. In this paper, the authors show that this is not the case: they show that if \(S\) is any Enriques surface and \(B\) a very general curve genus at least one, then the integral Hodge conjecture fails for the product \(S \times B\). For the proof the authors give a geometric reformulation of the problem: if the integral Hodge conjecture holds on \(S \times B\), then every finite étale cover of \(B\) is induced, via an algebraic correspondence, by the universal cover of \(S\). They then show by a degeneration argument that this property fails for a very general curve \(B\). As a consequence of their technique one obtains that there exist smooth projective threefolds with a \(2\)-torsion class in \(H^4(X, \mathbb Z)\) that is not algebraic.