Computation of denoising posterior precision matrix in JAC method (score_fn_iid)

[touronc]

I am working with the JAC score estimation (for i.i.d. data) inspired by the paper "Diffusion posterior sampling for simulaton-based inference in tall data settings" from Linhart J. et al (algorithm 2).

In the estimation of the marginal denoising posterior precision matrix, why is there an addition between the 2 terms ?
Shouldn't it be a multiplication instead ?

Here is the link to the specific line of the function :
(

sbi/sbi/inference/potentials/score_fn_iid.py

Line 727 in dac03db

denoising_posterior_precision = m**2 / std**2 + torch.inverse(cov0)

)

Thanks !

[janfb]

Thanks for raising this @touronc ! @manuelgloeckler any thoughts?

[manuelgloeckler]

Thanks for checking I think you are right, this should be a multiplication, good catch!

Will test and adjust.