Look at every possible two particle collision. To simplify, consider one of the particles as the initial frame of refernce so you now just have impact angle, which is random.

Now play back time by looking at that particle's post collision path as the frame of reference and play time backwards. You will find a bias in input angles for these collisions.