Methods

We use artificial null data generated from the model, with $\beta=0$ and $a_p=0$. The stimulus is taken to be a square wave with period of $60$ seconds. We have $T=200$ scans, a TR of $3$ seconds and use an HRF resolution of $\rho=6$. Using these values we generate time-series for $10000$ voxels. We then fit the model with an autoregressive order of $P=4$ and use a basis set of the top 3 eigenHRFs shown in figure 3. We use two different models, one with no HRF constraints ($m=0$ and $C=I$) and one with HRF constraints ( $m=\tilde{m}$ and $C=\tilde{C}$ as given in equation 17). We use an f-contrast of $[1,0,0;0,1,0;0,0,1]$.