Author response image 2. Gradient updating with randomness in the starting location and each updating step does not generate the biased end-point pattern seen in the data.

(A) Simulation of reward-gradient updating trajectories starting from random locations (white circles) on the reward function. Each updating step = scaling factor x gradient + noise (noise along the me and z dimensions were generated independently from the same Gaussian distribution). The updating process stopped when reward exceeded 97% of the maximum (red circles indicate end points). Note that the end-points of the updating process were located all around the reward-function plateau. The reward function was from the LR-R blocks in an example session of monkey C. (B) Polar histograms of the angle of the end-points relative to the peak of the reward function, showing end-points scattered all around the peak with clusters in two locations.