mancusolab/susiepca

Numerically Stable PIP computation

Closed this issue · 2 comments

If L is large, computing PIPs can be numerically unstable. Let's switch to something like,

pip = 1 - jnp.exp(jnp.sum(jnp.log1p(-alpha), axis=-1))

Should #13 be axis=-1 or axis=0 ?

axis = 0 is correct, fixed!