Simone Carlo Surace

> Interestingly `ForwardDiff` also returns the wrong thing if `x` is not a `PDMat` already: I think this would be a great test to do for all the functions overloaded...

So `cholesky` is not defined for non-symmetric `x`, which means that a gradient step will lead to the subspace being left. This is an instance of the `gradient` being correct...

Especially because using a `Symmetric` wrapper fails: ```julia julia> ForwardDiff.gradient(logdet ∘ PDMat, Symmetric(x)) ERROR: ArgumentError: Cannot set a non-diagonal index in a symmetric matrix ``` whereas it works for Zygote:...

Yes, as I said I do believe that the differential is correct because it is only defined for a symmetric tangent. I think we can focus on getting the gradients...

Thanks for the suggestion. Maybe I misunderstood, but this did not solve the issue: ```julia struct LinearMean{T} a::T b::T end (f::LinearMean)(x) = f.a * first(x) + f.b using AbstractGPs: CustomMean...

Oh, I get it. Thanks, this seems to do the trick. Actually, wrapping `LinearMean` in `CustomMean` seems overly complicated. I could make `LinearMean

P.S. `CustomMean` currently does not seem to be exported or documented, for that matter.

I ended up with a general struct `FunctionOfTime{Tf}

> The reason I am asking is, because of my ignorance, so far it seems to me that ReinforcementLearning.jl is _an alternative way_ to do agent based simulations. So I...

I guess this means some of the lower bounds are outdated/no longer working.