Feature request: logdensity_and_hessian

[DilumAluthge]

The logdensity_and_gradient function is very useful for using e.g. ForwardDiff to compute the gradient of the log density function.

Could we also add a new function logdensity_and_hessian that uses ForwardDiff to compute the Hessian of the log density function?

Although, I suppose that, in the process of computing the Hessian, ForwardDiff will also compute and store the gradient. So it would probably make the most sense to add a new function logdensity_and_gradient_and_hessian that returns the value, the gradient, and the Hessian.

[tpapp]

I am happy to support this (incidentally, what's you use case? I am just curious).

But I want to think a bit more about the API. I think I would prefer

struct Derivatives{N} end # FIXME check n ≥ 0 etc
Derivatives(n::Int) = Derivatives{n}()
logdensity(ℓ, x, derivatives = Derivatives(0))

And once we have a breaking API change anyway, I would prefer solve too #56 by returning any object that supports the properties .value, .gradient, and .hessian (each where applicable, when called with the relevant Derivatives(...)). I think more values than 2 is error-prone.

I appreciate suggestions about the API, bikeshedding is welcome before we finalize it.

I agree that I don't think there is a use case for having the value and the hessian but not the gradient.

[DilumAluthge]

LogDensityProblems is a convenient way to formulate maximum likelihood problems. If I don't feel like writing down the derivatives by hand, I can use AD to find the MLE using gradient ascent.

So in order to find the MLE, I only need the gradient of my log-likelihood function. But if I also have the Hessian of my log-likelihood function, I can negate it to get the observed information matrix, and I can use that to estimate standard errors (based on the asymptotic normality of the MLE).

[DilumAluthge]

I like the API that you have outlined here. People that only need the value would do this:

foo = logdensity(ℓ, x, derivatives = Derivatives(0))
foo.value

People that need the gradient would do:

foo = logdensity(ℓ, x, derivatives = Derivatives(1))
foo.value
foo.gradient

People that need the Hessian would do:

foo = logdensity(ℓ, x, derivatives = Derivatives(2))
foo.value
foo.gradient
foo.hessian

[tpapp]

Thanks for the comments. I will prepare PR soon.

[tpapp]

Just a quick heads-up: this is still on my radar, I just want to experiment with the practical viability of calculating Hessians via AD for medium-sized models (100-1000 parameters) first.

[sethaxen]

I'd also like this. My use case is described in mlcolab/Pathfinder.jl#115. In Pathfinder we look for a MAP estimate using some optimizer, and we'd like to support users passing second-order optimizers.

[tpapp]

Happy to add an API for this, the only thing that was holding this back is AD support for efficient Hessians.

Will do a PR next week.

Also, I am going ahead with logdensity_and_gradient_and_hessian, unless people object. The API outlined above would be breaking, that's for the future. But of course bikeshedding the name is welcome.

[tpapp]

Or maybe logdensity_gradient_and_hessian.

[sethaxen]

Will do a PR next week.

Sounds great!

Or maybe logdensity_gradient_and_hessian.

I prefer this one.