We treat the problem of searching for hidden multidimensional independent auto-regressive processes. First, we transform the problem to Independent Subspace Analysis (ISA) Here: the unknown mixing matrix A ∈ R D×D , the hidden components s m ∈ R d , andGoal of IPA: estimate s(t) and A (or W := A −1 : separation matrix) by using observations z(t) only. Specially:when ∀F m = 0 and d = 1.
AssumptionsFor an AR process, the innovation is identical to the noise that drives the process ⇒ IPA models(t + 1) = Fs(t) + e(t), (6) z(t) = AFA −1 z(t − 1) + Ae(t − 1), (7) z(t) = Ae(t − 1) = As(t). Assuming that W ICA (z) is unique (up to permutation and sign of the components), then it is W ISA (z) (up to permutation and sign of the components). In other wordswhere P ∈ R D×D is a permutation matrix to be determined. (Proof in [5], e.g., for elliptically symmetric sources) ⇒ IPA estimation steps: 1. observe z(t) and estimate the AR model, 2. whiten the innovation of the AR process and perform ICA on it, 3. solve the combinatorial problem: search for the permutation of the ICA sources that minimizes the cost J. Thus IPA needs only two (more) steps: (i)Ĥ, and (ii) optimization of J in S D (permutations of length D).
Assistants 3