Lower and Upper bound not working for multivariate distributions

[sachinruk]

This is bit of a feature request than anything. I was wondering if it would be possible to implement the bounds of a distribution similar to what is done in stan. eg:

parameters{
  real<lower=0,upper=1> sigma;
}

The only way I can think of doing this in pymc3 is using densityDist. Lastly if this were to be implemented would it affect ADVI?

[fonnesbeck]

You can already do this by wrapping any continuous distribution in Bound.

[sachinruk]

How would this work in the Multivariate case? I wish to change the following line:

innov_like = pm.MvNormal.dist(mu=x_im1 + mu, tau=tau).logp(x_i)

to

BoundedNormal = pm.Bound(pm.MvNormal, lower=0)
innov_like = BoundedNormal.dist(mu=x_im1 + mu, tau=tau).logp(x_i)

The above chucked a dimensionality mismatch error. Any thoughts?

Also with regards to using VB, anything I ought to be cautious about when using bounded distributions?

[twiecki]

We don't have tests for bounds on multivariate distributions so I'm not surprised it isn't working.

There should be no issue with VB though.

[domenzain]

#1843 There are some pitfalls when sampling.