The activity trajectories in Figure 4 were obtained by projecting the average population activity for each trial type into the low dimensional space that captured most variance between trial types, where trial type was defined by the 8 possible combinations of choice, second-step and outcome. To find this space, we calculated the average activity for each neuron for each trial type. We then averaged these across trial types to evaluate the component of activity that was not selective to different trial types. We subtracted the non-selective activity for each neuron from that neurons average activity for each individual trial type, and concatenated across trial types to generate a data matrix of shape [n neurons, n trial types ∗ n time point] representing how activity for each neuron deviated from its cross-trial-type average in each trial type. We performed PCA on this matrix to find the space that captured the most cross-trial-type variance and then projected the average population activity trajectory for each trial type into this space to generate Figure 4.
