Popescu6. Spectral methods and phase estimationCross- and auto spectral densities of a time series, assum<strong>in</strong>g zero-mean or de-trendedvalues, are def<strong>in</strong>ed as:ρ Li j (τ) = E (︁ y i (t)y j (t − τ) )︁S i j (ω) = F (ρ Li j (τ)) (12)Note that cont<strong>in</strong>uous, l<strong>in</strong>ear, raw correlation values are used <strong>in</strong> the above def<strong>in</strong>itionas well as the cont<strong>in</strong>uous Fourier transform. Bivariate coherency is def<strong>in</strong>ed as:C i j (ω) =S i j (ω)√︀ S ii(ω)S j j (ω)(13)Which consists of a complex numerator and a real-valued denom<strong>in</strong>ator. The coherenceis the squared magnitude of the coherency:c i j (ω) = C i j (ω) * C i j (ω) (14)Besides various histogram and discrete (fast) Fourier transform methods availablefor the computation of coherence, AR methods may be also used, s<strong>in</strong>ce they are alsol<strong>in</strong>ear transforms, the Fourier transform of the delay operator be<strong>in</strong>g simply z k = e − j2πωτ Swhere τ S is the sampl<strong>in</strong>g time and k = ωτ S . Plugg<strong>in</strong>g this <strong>in</strong>to Equation (9) we obta<strong>in</strong>:⎛⎞K∑︁X( jω) = ⎜⎝ A k e − j2πωτ S kk=1⎛⎞K∑︁Y( jω) = C ⎜⎝ I − A k e − j2πωτ S k⎟⎠k=1⎟⎠ X( jω) + BU( jω) + DY( jω) = CX( jω) (15)−1(BU( jω) + DW( jω)) (16)In terms of a SVAR therefore (as opposed to VAR) the mix<strong>in</strong>g matrix C does notaffect stability, nor the dynamic response (i.e. the poles). The transfer functions fromith <strong>in</strong>novations to jth output are entries of the follow<strong>in</strong>g matrix of functions:⎛⎞K∑︁H( jω) = C ⎜⎝ I − A k e − j2πωτ S k⎟⎠k=1−1D (17)The spectral matrix is simply (hav<strong>in</strong>g already assumed <strong>in</strong>dependent unit Gaussiannoise):S ( jω) = H( jω) * H( jω) (18)The coherency as the coherence follow<strong>in</strong>g def<strong>in</strong>itions above. The partial coherenceconsiders the pair (i, j) of signals conditioned on all other signals, the (ordered) set ofwhich we denote (i, j):50
Robust Statistics for <strong>Causality</strong>S i, j|(i, j)( jω) = S (i, j),(i, j) + S (i, j),(i, j)S −1(i, j),(i, j) S (i, j),(i, j)(19)Where the subscripts refer to row/column subsets of the matrix S ( jω). The partialspectrum, substituted <strong>in</strong>to Equation (13) gives us partial coherency C i, j|(i, j)( jω) andcorrespond<strong>in</strong>gly, partial coherence c i, j|(i, j)( jω) . These functions are symmetric andtherefore cannot <strong>in</strong>dicate direction of <strong>in</strong>teraction <strong>in</strong> the pair (i, j). Several alternativeshave been proposed to account for this limitation. Kam<strong>in</strong>ski and Bl<strong>in</strong>owska (1991);Bl<strong>in</strong>owska et al. (2004) proposed the follow<strong>in</strong>g normalization of H( jω) which attemptsto measure the relative magnitude of the transfer function from any <strong>in</strong>novations processto any output (which is equivalent to measur<strong>in</strong>g the normalized strength of Grangercausality) and is called the directed transfer function (DTF):γ i j ( jω) =H i j ( jω)√︁ ∑︀k |H ik ( jω)| 2γ 2 i j ( jω) = ⃒ ⃒⃒Hij ( jω) ⃒ ⃒ ⃒2∑︀k |H ik ( jω)| 2 (20)A similar measure is called directed coherence Baccalá et al. (Feb 1991), later elaborated<strong>in</strong>to a method complimentary to DTF, called partial directed coherence (PDC)Baccalá and Sameshima (2001); Sameshima and Baccalá (1999), based on the <strong>in</strong>verseof H:π i j ( jω) =H −1i j√︁ ( jω)∑︀ ⃒k ⃒Hik −1(jω)⃒ ⃒2The objective of these coherency-like measures is to place a measure of directionalityon the otherwise <strong>in</strong>formation-symmetric coherency. While SVAR is not generallyused as a basis of the autoregressive means of spectral and coherence estimation, or ofDTF/PDC is is done so <strong>in</strong> this paper for completeness (otherwise it is assumed C = I).Granger’s 1969 paper did consider a mix<strong>in</strong>g matrix (<strong>in</strong>directly, by add<strong>in</strong>g non-diagonalterms to the zero-lag matrix), and suggested ignor<strong>in</strong>g the role of that part of coherencywhich depends on mix<strong>in</strong>g terms as non-<strong>in</strong>formative ‘<strong>in</strong>stantaneous causality’. Notethat the ambiguity of the role and identifiability of the full zero lag matrix, as describedhere<strong>in</strong>, was fully known at the time and was one of the justifications given for separat<strong>in</strong>gsub-sampl<strong>in</strong>g time dynamics. Another measure of directionality, proposed bySchreiber (2000) is a Shannon-entropy <strong>in</strong>terpretation of Granger <strong>Causality</strong>, and thereforewill be referred to as GC here<strong>in</strong>. The Shannon entropy, and conditional Shannonentropy of a random process is related to its spectrum. The conditional entropy formulationof Granger <strong>Causality</strong> for AR models <strong>in</strong> the multivariate case is (where (i) denotes,as above, all other elements of the vector except i ):H GCj→i|u = H(y i,t+1|y :,t:t−K ,u :,t:t−K ) − H(y i,t+1 |y ( j),t:t−K,u :,t:t−K )51