Will Tebbutt

> With which method it is not a bound? Pretty much anything that's not the ELBO. EP (FITC) / SoR / DTC etc are all just approximations, with no particular...

I was wondering about that too. On the face of it, it does make more sense, as they're certainly approximations. The only downside I can see is that you'll have...

To my mind `ApproxPosteriorGP` ought to store only what it needs in order to construct the GP corresponding to the approximate posterior. If it needs to hold onto the likelihood...

> The one really annoying thing is that LaplaceGP requires an explicit jitter argument, and it seems weird to have to access lfx.fx.Σy - which should be an implementation detail...

When I say abstraction, I think I should really say it's an implementation detail. In particular, we implement the LatentFiniteGP in terms of a FiniteGP and a likelihood, and the...

> to hide the implementation detail for now? Sounds good to me. > Oh, the other part that's missing is that the only API implemented for LatentGP so far is...

@devmotion can we close this?

Hi Seth. Yes, this is almost certainly what's going on. The doubling of the dimensionality is being caused by the output type being complex for non-integer exponent. I'm not entirely...

> even if not making changes, may still warrant a mention in the docs as a "gotcha"? Definitely. > Can we detect this case, or a subset of this case,...

Might make sense to cast it explicitly wherever it's used to the element type of the thing that it's being added to. This way if (for some reason) we ever...