Closed ksnxr closed 6 months ago
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 98.87%. Comparing base (
2e25624
) to head (3cd525a
).
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
The current MALA transition energy is implemented incorrectly. In fact, the correct one should be the reverse of the current one.
To see that, note that MALA's log acceptance probability is calculated as
prev_energy - new_energy
, withnew_energy
being the transition energy frominitial_state
tostate
. As such, the correctnew_energy
should be-log p(state) - log q(initial_state|state)
, but the current implementation implements-log p(initial_state) - log q(state|initial_state)
; the implementation of RWMH seems correct.Reference: https://en.wikipedia.org/wiki/Metropolis-adjusted_Langevin_algorithm