Feature request: Jacobian calculation for gradients

[aterenin]

I'd like to use TransformVariables as part of an HMC sampler, but my evaluating my target distribution's probability density involves solving a partial differential equation, which is done outside Julia and thus not compatible with AutoDiff. There are a special techniques for obtaining gradients for such models which involve solving an adjoint equations, but that's a digression.

For people with use cases like mine, it would be nice if there was a function which for a transformation y = t(x) transforms grad(log(f(x)) into grad(log(f(y)) with the appropriate Jacobian correction added to each individual coordinate - similar to transform_logdensity, but for gradients of log-densities.

[tpapp]

I am not entirely sure what you are asking for here. You have a transformation, you can also calculate its log Jacobian determinant, and you would like to use it for Bayesian inference?

with the appropriate Jacobian correction added to each individual coordinate

Sorry, I am not getting this --- the Jacobian is for the whole transformation, not by coordinate.

Perhaps a mock example would help.

[aterenin]

Sure.

Suppose we have two random variables, X and Y with X ~ Beta(2,2) and Y = X/2. Suppose that we do not know the density of Y.

Let f_X(.) and f_Y(.) be their densities. We are interested in evaluating d/dy* ln f_Y(y) for a giveny, so we transform x = 2y and obtain f_Y(y) = f_X(x / 2) |J(x)| where J is the Jacobian of the transformation, here just equal to the constant 2. The logarithm is just ln f_Y(y) = ln f_X(x / 2) + ln(2) which is what transform_logdensity gives you.

I am interested in d/dy ln f_Y(y) in x coordinates. In other words, my target distribution is constrained, so I transform it to be unconstrained. transform_logdensity lets me evaluate its density, accounting for the transformation. How do I evaluate the gradient of its log density, also accounting for the transformation?

Right now, the examples in DynamicHMC do this by applying automatic differentiation to ln f_Y(y), but in my use case this isn't possible, hence the desire to obtain it directly.

Does this example help?

[tpapp]

Yes, this is very helpful, and reminds me of something I encountered myself when I was coding a log density and its gradient without AD.

I will think about it an get back to you in this issue.

[tpapp]

Cf tpapp/DynamicHMCExamples.jl#18 which is pretty much asking for the same thing. And I realized I need it too occasionally. It should be easy to implement, will start when #56 is merged.