Mesoscopic fluctuations of the zeta zeros (Q707603): Difference between revisions

Selberg's celebrated central limit theorem for \(\log \zeta(\frac 12+it)\) states that \[ \frac{\log \zeta(\frac 12+it)}{\sqrt{\frac 12\log\log t}} \] converges in law to a two-dimensional Gaussian, that is if \(\mathcal R\subset \mathbb C\) is a rectangle, then as \(T\rightarrow \infty\) we have \[ \frac 1T meas\{ 1\leq t\leq T : \frac{\log \zeta(\tfrac 12+it)}{\sqrt{\frac 12\log\log t}} \in \mathcal R \} \longrightarrow \frac 1{2\pi}\int_{\mathcal R} e^{-\frac{x^2+y^2}2} \,dx\,dy. \] This result has been generalized by several authors [\textit{C. P. Hughes, A. Nikeghbali} and \textit{M. Yor}, ``An arithmetic model for the total disorder process'', Probab. Theory Relat. Fields 141, No. 1--2, 47--59 (2008; Zbl 1144.60020); \textit{W. D. Joyner}, Distribution theorems of L-functions. Pitman Research Notes, vol. 142 (1986); \textit{K.-M. Tsang}, The distribution of the values of the zeta function. Thesis, Princeton University, October 1984], who studied multidimensional analogues, as well as the distribution of \(\log\zeta(\frac 12+\varepsilon_t+it)\) where \(\varepsilon_t\rightarrow 0\) is a small shift. In the case where \(\varepsilon_t\gg 1/\log T\), the normalization \(\sqrt{\frac 12\log\log T}\) has to be changed to \(\sqrt{-\log \varepsilon_t}\). The paper under review generalizes both approaches, by studying the distribution of the vector \[ \frac 1{\sqrt{-\log \epsilon_t}} \left( \log \zeta(\tfrac 12+\varepsilon_t+if_t^{(1)} + i \omega t),\dots, \log \zeta(\tfrac 12+\varepsilon_t+if_t^{(\ell)} + i \omega t)\right), \] where \(\varepsilon_t\gg 1/\log t\) are small shifts tending to zero, \(f_t^{(\ell)}\) are certain functions of \(t\) satisfying \[ \frac{\log|f_t^{(i)} -f_t^{(j)}|}{\log \varepsilon_t} \rightarrow c_{i,j} \] \((i\neq j)\) for some constants \(c_{i,j}\geq 0\), and \(\omega\) is uniform on \((0,1)\). (The normalization by \(\sqrt{-\log \varepsilon_t}\) has to be changed to \(\sqrt {\log\log t}\) when \(\varepsilon_t \ll 1/\log t\).) It is shown that this vector converges in law to a multivariate Gaussian distribution \((Y_1,...,Y_{\ell})\), where the \(Y_i\) are complex Gaussians with mean zero and the following covariance structure: \[ Cov(Y_i,Y_j)= \begin{cases} 1 &\text{ if } i=j \\ \min(1,c_{i,j}) & \text{ if } i\neq j. \end{cases} \] This non-trivial correlation structure is quite unexpected, and is due to the fact that the shifts are not too far apart from each other. Moreover, the result has several quite interesting consequences, one of which is what the author calls the mesoscopic fluctuations of zeros of \(\zeta(s)\). Indeed, the result can be used to study \[ \Delta(t_1,t_2) := S(t_2)-S(t_1), \] where \(S(T):=\arg \zeta(\frac 12+it)\) is the error term in the Riemann--von Mangoldt formula for the number of zeros of \(\zeta(s)\) of imaginary part at most \(T\). The author shows that on a ``mesoscopic'' scale of \((\log T)^{\delta}\) with \(\delta<1\), \(\Delta(t_1,t_2)\) exhibits an interesting non-trivial covariance structure, conforming numerical experiments of \textit{M. Coram} and \textit{P. Diaconis} [``New tests of the correspondence between unitary eigenvalues and the zeros of Riemann's zeta function'', J. Phys. A, Math. Gen. 36, No. 12, 2883--2906 (2003; Zbl 1074.11046)]. Finally, it is shown that analogous results hold for characteristic polynomials of large unitary matrices, strengthening once more the link between zeros of the zeta function and eigenvalues of large unitary matrices.