Mario Lezcano Casado

You are completely right. I got a bit lost in the notation. IMO, it makes complete sense to support this, and it may be even possible to support it within...

It may be best to create a new class then, I think. @egaznep would you be keen on having a go at this one?

There are a number of issues with the current implementation, (e.g. the computation of `R_zz` is wrong, i.e., it needs tests, and you are accepting `Sequence`s but then you check...

That is a very good point. You can also have a convenience helper to transform the complex parameters to real parameters, which would also fit this pattern.

Also, I am inclined to think that complex 2n-dimensional real normal distributions are in correspondence to n-dimensional complex normal distributions because of a dimensionality count argument. The Wikipedia page shows...

@fritzo made in https://github.com/pytorch/pytorch/pull/92241#pullrequestreview-1250501932 some very valid points that should be taken into account by whoever wants to push this forward.

Closing this one as it's low prio and I don't see a clear path to landing it. This behaviour can already be implemented in core via `torch.zeros((m,n)).diagonal(k).fill_(1)`.

Open a new PR, as now this PR has everyone subscribed to it

That check does not check that a matrix is SPD, as the matrix may not even be symmetric. When looking closer at what you want to do, actually the best...

LGTM!