Blackjax

What changed

Breaking changes

To build a HMC or NUTS kernel in 0.2.1 and previous versions one needed to provide a `potential_fn` function:

python
kernel = nuts.kernel(potential_fn, step_size, inverse_mass_matrix)


Instead we now ask the users to provide the more commonly used log-probability function:

python
kernel = nuts.kernel(logprob_fn, step_size, inverse_mass_matrix)


where `logprob_fn = lambda x: -potential_fn(*x)`

New features

- Tempered Sequential Monte Carlo (AdrienCorenflos 40 )
- Rosenbluth Metropolis Hastings algorithm (AdrienCorenflos 74 )
- Effective Sample Size, RHat (junpenglao 66 )
- Higher-order integrators for HMC (rlouf 59 )

Bugs

* Missing key splitting in trajectory integration (wiep 53 )

What changed

* `momentum` and `position` were passed to the kinetic energy in the wrong order, leading to biased sampling as noticed in 46. We corrected this behavior and added a new test.

What changed

- The Stan adaptation scheme, including dual averaging, computing covariance with Welford's algorithm and the schedule (rlouf)
- Recursive implementation of NUTS (junpenglao)
- Many BUG fixes on NUTS (junpenglao)

New features

- `hmc` kernel
- `nuts` kernel
- Notebook with examples of how to sample one or multiple chains with HMC, NUTS