Skip to content

MAKLA-BCSS2 sampler #913

Description

@paulindani

Algorithm category

MCMC (gradient-based)

Paper reference

Title: Divide, Interact, Sample: The Two-System Paradigm
Authors: James Chok, Myung Won Lee, Daniel Paulin, Geoffrey M. Vasil
Year: 2026
Link: https://arxiv.org/abs/2509.09162

Existing implementations

Official repo: https://github.com/paulindani/MAKLA_JAX
Implementation is already in JAX.

Benefit and motivation

In the paper, there is
(1) a new sampler MAKLA-BCSS2, which is a variant of GHMC, using deterministic 2 gradient and 1 likelihood evaluation per step.
(2) There are new ensemble adaptation methods for finding good mass matrix to get better preconditioning.

https://arxiv.org/abs/2509.09162 has shown significant improvements in efficiency over BlackJAX NUTS on 48 different posteriors from PosteriorDB.

Comparison to existing BlackJAX algorithms

Closest existing method: NUTS
Advantage over it: deterministic computation cost per iteration (more efficient VMAP, better implementation efficiency), better statistical efficiency in terms of grad/ESS.
Disadvantage over NUTS: none that I am aware of.

Estimated JAX implementation effort

S — standard ops, fits existing pattern

JAX-specific implementation notes

No response

Are you willing to open a PR?

Yes — I can implement this

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions