Why would that matter? I was thinking “If you’d use 4-bit ‘IEEE’ floats, it wouldn’t be very surprising to see an anomaly like this. Would it be for 5-bit ones?

Why would that matter? I was thinking “If you’d use 4-bit ‘IEEE’ floats, it wouldn’t be very surprising to see an anomaly like this. Would it be for 5-bit ones? 6-bit? What’s a large enough number of bits?” I also said “For this problem, I don’t see how that would give different results as picking any number in [-1, +1]” because I think 64 is (more than) large enough, but still, I think you’d need an argument if you want to draw a mathematically sound conclusion “this is worth looking into” from such a simulation.