Figure 5.

Illustration of two approaches for lag selection: (a) Assuming a decay assumption—that is, $A^k=0\Rightarrow A^{k^{\prime }}=0\forall k^{\prime }\ge k$—the lag $d$ can be estimated by identifying the first $k$ such that $A^k=0$. (b) The lag $d$ can be estimated without assuming a decay assumption by enforcing entire lag matrices to be zero and setting $\widehat{d}={\text{max}}_k{\widehat{A}}^k\ne 0$.