We study Granger Causality in the context of wide-sense stationary time series. The focus of the analysis is to understand how the underlying topological structure of the causality graph affects graph recovery by means of the pairwise testing heuristic. Our main theoretical result establishes a sufficient condition (in particular, the graph must satisfy a polytree assumption we refer to as strong causality) under which the graph can be recovered by means of unconditional and binary pairwise causality testing. Examples from the gene regulatory network literature are provided which establish that graphs which are strongly causal, or very nearly so, can be expected to arise in practice. We implement finite sample heuristics derived from our theory, and use simulation to compare our pairwise testing heuristic against LASSO-based methods. These simulations show that, for graphs which are strongly causal (or small perturbations thereof) the pairwise testing heuristic is able to more accurately recover the underlying graph. We show that the algorithm is scalable to graphs with thousands of nodes, and that, as long as structural assumptions are met, exhibits similar high-dimensional scaling properties as the LASSO. That is, performance degrades slowly while the system size increases and the number of available samples is held fixed. Finally, a proof-of-concept application example shows, by attempting to classify alcoholic individuals using only Granger causality graphs inferred from EEG measurements, that the inferred Granger causality graph topology carries identifiable features.

The work is concerned with the first-order linear autoregressive process which has a rectangular stationary marginal distribution. A derivation is given of the result that the time-reversed version is deterministic, with a first-order recursion function of the type used in multiplicative congruential random number generators, scaled to the unit interval. The uniformly distributed sequence generated is chaotic, giving an instance of a chaotic process which when reversed has a linear causal and non-chaotic structure. An mk-valued discrete process is then introduced which resembles a first-order linear autoregressive model and uses k-adic arithmetic. It is a particular form of moving-average process, and when reversed approximates in m a non-linear discrete-valued process which has the congruential generator function as its deterministic part, plus a discrete-valued noise component. The process is illustrated by scatter plots of adjacent values, time series plots and directed scatter plots (phase diagrams). The behaviour very much depends on the adic number, with k = 2 being very distinctly non-linear and k = 10 being virtually indistinguishable from independence.

We consider the model Zt = φ (0, k)+ φ(1, k)Zt–1 + at (k) whenever rk−1<Zt−1≦rk, 1≦k≦l, with r0 = –∞ and rl =∞. Here {φ (i, k); i = 0, 1; 1≦k≦l} is a sequence of real constants, not necessarily equal, and, for 1≦k≦l, {at(k), t≧1} is a sequence of i.i.d. random variables with mean 0 and with {at(k), t≧1} independent of {at(j), t≧1} for j ≠ k. Necessary and sufficient conditions on the constants {φ (i, k)} are given for the stationarity of the process. Least squares estimators of the model parameters are derived and, under mild regularity conditions, are shown to be strongly consistent and asymptotically normal.

We consider the model where φ1, φ2 are real coefficients, not necessarily equal, and the at,'s are a sequence of i.i.d. random variables with mean 0. Necessary and sufficient conditions on the φ 's are given for stationarity of the process. Least squares estimators of the φ 's are derived and, under mild regularity conditions, are shown to be consistent and asymptotically normal. An hypothesis test is given to differentiate between an AR(1) (the case φ1 = φ2) and this threshold model. The asymptotic behavior of the test statistic is derived. Small-sample behavior of the estimators and the hypothesis test are studied via simulated data.