Delayed feedback causes non-Markovian behavior of neuronal firing statistics

Abstract: The instantaneous state of a neural network consists of both the degree of excitation of each neuron, the network is composed of, and positions of impulses in communication lines between neurons. In neurophysiological experiments, the neuronal firing moments are registered, but not the state of communication lines. But future spiking moments depend essentially on the past positions of impulses in the lines. This suggests, that the sequence of intervals between firing moments (interspike intervals, ISIs) in the network could be non-Markovian. In this paper, we address this question for a simplest possible neural "net", namely, a single neuron with delayed feedback. The neuron receives excitatory input both from the driving Poisson stream and from its own output through the feedback line. We obtain analytical expressions for conditional probability density $P(t_{n+1} | t_n,...,t_1,t_0)$, which gives the probability to get an output ISI of duration $t_{n+1}$ provided the previous $(n+1)$ output ISIs had durations $t_n,...,t_1,t_0$. It is proven exactly, that $P(t_{n+1} | t_n,...,t_1,t_0)$ does not reduce to $P(t_{n+1} | t_n,...,t_1)$ for any $n \geq 0$. This means that the output ISIs stream cannot be represented as Markov chain of any finite order.