SUPPLEMENT TO \INFERRING NETWORK STRUCTURE FROM INTERVENTIONAL TIME-COURSE EXPERIMENTS"

Furthermore this orthogonalisation can be used to improve computational effeciency when comparing a large number of models with the same X0. Define x ′ = (In − P 0)x (which can be precomputed) and then assuming that X0Xγ = 0a×b, we have x T (In − P 0 − P γ)x = x′T (In − P γ)x. This reduces the number of computations that have to be performed for each model matrix Xγ . When the only parameter common to all models is an intercept parameter the orthogonalisation corresponds to centring the predictors and x′ becomes the centred response.