Optimal hypothesis testing with combinatorial structure: Detecting random walks on graphs

Abstract

Suppose we have a time series of observations for each node in a network, and we wish to detect the presence of a particle undergoing a random walk on the network. If there is no such particle, then we observe only zero-mean Gaussian noise. If present, however, the location of the particle has an elevated mean. How well can we detect the particle at low signal-to-noise ratios? This is a special case of the problem of detecting hidden Markov processes (HMPs). The performance metric we analyze is the error exponent of the optimal detector, which measures the exponential rate of decay in the miss probability if the false alarm probability is held fixed as the observation time increases. This problem exhibits deep connections to a problem in statistical physics: computing the free energy density of a spin glass. We develop a generalized version of the random energy model (REM) spin glass, whose free energy density provides a lower bound for our error exponent, and compute the bound using large deviations techniques. The bound closely matches empirical results in numerical experiments, and suggests a phase transition phenomenon: below a threshold SNR, the error exponent is nearly constant and near zero, indicating poor performance; above the threshold, there is rapid improvement in performance as the SNR increases. The location of the phase transition depends on the entropy rate of the Markov process.