RNUTS

This repo contains all the code for our planned paper on Stiffness and using Implicit Midpoint in NUTS.

Outline of Code

get_nuts_samples(z0) # nuts.R

1. ham_system <- create_hamiltonian_system() # nuts.R
2. integrator <- create_integrator() # integrators.R
3. get_single_nuts_sample(z0, ham_system, integrator) #nuts.R A) p0 <- ham_system$get_momentum_sample() B) tree <- create_onenode_tree(z0, ham_system) # initialize tree from build_tree.R C) new_subtree <- build_tree(depth, tree$z_plus, tree$z_plus_1, directions[depth+1], ham_system, integrator, DEBUG)

Adding Tests

• since code is still kind of unstable it may only be worth it to run tests for outside code such as the GMRES solver
• have a top level script that runs all tests and make sure it always works

Things to Discuss

• getting in to Stan
• how to test for correctness? just look at quantiles of posterior converging?
• should numerical experiments showing correctness of samples be included in the paper? possibly in supplemtary material. showing that it's symplectic is enough
• use stepsize and mass matrix Stan decides? TBD how to determine midpoint stepsize. show a plot of work/sample vs. stepsize