Leapfrog, adpatve stepsize, and time reversibility

[junpenglao]

@Hamster_on_wheels

The gradient of the log-likelihood can change a lot when the gradient is evaluated at different points.

That’s almost always the case.