Partial Directed Coherence and the Directed Transfer Function

Both PDC and DTF are frequency-domain directed-connectivity measures derived from a single fitted multivariate autoregressive (MVAR) model of a set of channels. The directed transfer function is built from the model's transfer function, so it reflects the total (direct plus indirect, cascaded) influence from one channel to another as a function of frequency. Partial directed coherence is built instead from the MVAR coefficient matrix itself, which emphasizes direct influences and, unlike DTF, tends to suppress connections that are merely relayed through an intermediate node.

Because they share the MVAR foundation, they inherit all of Granger causality's assumptions: stationarity, correct model order, whitened residuals, and observation of all relevant sources. Numerous normalized variants exist (generalized PDC, full-frequency DTF) that change how the measure is scaled across frequency and outflow, and results can differ substantially between variants, so the specific estimator and its normalization must always be stated.