Spike Train Cumulants for Linear-Nonlinear Poisson Cascade Models
Spiking activity in cortical networks is nonlinear in nature. The linear-nonlinear cascade model, some versions of which are also known as point-process generalized linear model, can efficiently capture the nonlinear dynamics exhibited by such networks. Of particular interest in such models are theoretical predictions of spike train statistics. However, due to the moment-closure problem, approximations are inevitable. We suggest here a series expansion that explains how higher-order moments couple to lower-order ones. Our approach makes predictions in terms of certain integrals, the so-called loop integrals. In previous studies these integrals have been evaluated numerically, but numerical instabilities are sometimes encountered rendering the results unreliable. Analytic solutions are presented here to overcome this problem, and to arrive at more robust evaluations. We were able to deduce these analytic solutions by switching to Fourier space and making use of complex analysis, specifically Cauchy's residue theorem. We formalized the loop integrals and explicitly solved them for specific response functions. To quantify the importance of these corrections for spike train cumulants, we numerically simulated spiking networks and compared their sample statistics to our theoretical predictions. Our results demonstrate that the magnitude of the nonlinear corrections depends on the working point of the nonlinear network dynamics, and that it is related to the eigenvalues of the mean-field stability matrix. For our example, the corrections for the firing rates are in the range between 4 % and 21 % on average. Precise and robust predictions of spike train statistics accounting for nonlinear effects are, for example, highly relevant for theories involving spike-timing dependent plasticity (STDP).