Interesting, but I guess I'd have to learn APL or J to compare. Without understanding what those symbols mean, it seems much less like "math" than the numpy version.
1 + np.min(np.argwhere(running_sum[1:] == 0), axis=0)⊃ is 'first', ⍸ is where, f⍤g is like function composition, and ⍤1 means apply to each rank 1 (vector), so ⊃⍤⍸⍤1 means 'the first true for each vector'
The modification you mentioned is straightforward in any array language. Here is what you mentioned in APL (I know APL more than J but I'm sure the J would be basically the same):
This is the number of steps to return to 0 for each of 100 random walks of length 1000. (It is then easy to analyse the frequency/average/etc)