chi-feng / mcmc-demo Goto Github PK

Hi, @chi-feng, i have a question:
We know equation log(det(A)) = 2*sum(log(vecdiag(R))) is true, where A be your matrix and R = root(A) be the Cholesky root of the matrix A, but in line 30 of Multivariate.js, we see this.logDet = this.covL.diagonal().map(Math.log).sum(); , no need to multiply a constant 2 here?
Thanks.

https://developers.google.com/web/updates/2018/08/offscreen-canvas

Hi, @chi-feng, thanks for this insightful demo, it helps me a lot for understanding MCMC.
There is a related and interesting paper using Stein's method to approximate a target distribution, would it possible to rewrite it in js for online demo?

It used to be possible to link to
https://chi-feng.github.io/mcmc-demo/app.html#RadFriends-NS,donut
to pull up that specific algorithm and problem.

This was very useful.

At some point this stopped working and now it always shows the default (first algorithm, first problem).

Is it possible to restore this functionality?

Hi, I was trying the RadFriends-NS algorithm on the multimodal distribution. I find that after some time, it seems that no new samples are drawn (at least at the two lower peaks), but still the histograms are rapidly changing. Is this a bug or am I missing something here?

The project was started before ES2015 was widely supported, and is in need of some updating to follow modern javascript development best-practices.

If it's easy to add new distributions, it'd be nice to have Neal's funnel, even if only in 2D. The density is given by

a ~ normal(0, 3)
b ~ normal(0, exp(a / 2))

where the normal is taking a scale parameter (std deviation). The funnel is originally described in this paper:

• Neal, Radford M. 2003. Slice Sampling. Annals of Statistics 31 (3): 705–67.

P.S. Thanks for producing this---I direct people to the app all the time.