Approximating Novikov-Shubin numbers of virtually cyclic coverings (Q1691960)

Let \(G\) be a discrete group and \(\mathcal NG\) its von Neumann algebra. Self-adjoint elements in matrix rings over \(\mathcal NG\) have a spectral density function \(F\), and their Novikov-Shubin invariant measures its decay at 0 along the scale of power functions: if \(F(\lambda) \sim_{\lambda \to 0} \lambda^{\alpha + o(1)}\) for some \(0 \leq \alpha \leq \infty\) then its Novikov-Shubin number equals \(\alpha\). It is a priori possible that no such equivalence exists (i.e. \(F\) oscillates between two power functions) but no example of a group \(G\) and a spectral density function \(F\) where this behaviour occurs is known (see [\textit{W. Lück}, \(L^2\)-invariants: Theory and applications to geometry and \(K\)-theory. Berlin: Springer (2002; Zbl 1009.55001)], Remark 2.42). Through combinatorial Laplacians these numbers give rise to topological invariants of \(G\)-CW complexes, as it turns out that they are invariant under \(G\)-homotopy equivalences. When \(G\) is a finite group, all matrices over the group ring \(\mathbb C G\) have a spectral gap and thus there are no interesting Novikov-Shubin invariants. However a situation where it is particularly interesting to have a quantitative measurement of the spectral gap is the following: if \(G\) is residually finite, \(A \in \mathbb CG\) and \(G_i\) is a residual chain of quotients, and \(A_i\) is the reduction of \(A\) in \(\mathbb C G_i\) then one can ask whether the decay of the spectral gap of the \(A_i\) as \(i \to +\infty\) is related to the decay of the spectral density function of \(A\) at 0, i.e. its Novikov-Shubin number. For example if \(A\) itself has a spectral gap then we can ask whether the \(A_i\) have a uniform spectral gap (this is not true in general), but in general there are many ways to give a quantitative formulation of the question above. The paper under review studies such a formulation, whose precise statement is as follows: for an \(n \times n\) Hermitian matrix \(B\) let \(\sigma^+(B)\) be its smallest positive eigenvalue, and \(m^+(B)\) its multiplicity. Then define \[ \alpha(B) = \frac{\log\left(\frac{m^+(B)} n\right)}{\log\sigma^+(B)}. \] Then for \(A\) a self-adjoint matrix over \(\mathbb C G\) the question is: does \[ \alpha(A) = \limsup_{i \to +\infty} \left( \alpha(A_i) \right) \] hold, where \(A_i\) is the reduction of \(A\) in \(\mathbb C G_i\). The limit (as opposed to limit superior) does not hold in general (and this seems to be the rule rather than exception) and it is also interesting to ask whether, if \(\alpha(A) > 0\), we have \[ \liminf_{i \to +\infty}\alpha(A_i) > 0 \] as this has immediate consequences for the problem of ``determinant approximation'' (see [Lück, loc. cit.], Question 13.52). The author proves that both these questions have a positive answer when \(G\) is \(\mathbb Z\), or more generally a virtually cyclic group. This is also the only case where determinant approximation is known (see [Lück, loc. cit.], Chapter 13), and the proof of the authors' result uses the same tools as the latter, namely Baker's theorem in Diophantine approximation together with elementary arguments.