Skip to content

mutation

Ensemble mutation for the BlackJAX SMC sampler modes.

A mutation moves every particle at one temperature. The particles first go to coordinates fitted to the current ensemble: periodic angles are re-centred onto the real line and, when a flow is configured, mapped through a normalizing flow trained on the particles. A kernel (a Gaussian random walk or t-pCN) then takes Metropolis-Hastings steps in those coordinates. The acceptance ratio includes the Jacobian of the map, so the particles still target Jim's posterior. BlackJAX runs the mutation through its update_strategy hook.

EnsembleMutation ¤

Moves all SMC particles at one temperature, in coordinates fitted to them.

Hand read_parameters, mcmc_init_fn, update_strategy and mcmc_parameter_update_fn to blackjax.inner_kernel_tuning, with extend_params(initial_parameter_value()) as the initial parameter value and max_steps as num_mcmc_steps. The only parameter carried from one mutation to the next is the kernel's step size; a flow is trained afresh at every mutation.

Parameters:

Name Type Description Default
n_dims int

Dimension of the sampling space.

required
periodic Optional[dict[int, tuple[float, float]]]

Periodic dimensions as {index: (lower, upper)}, or None.

required
kernel_config SMCKernelConfig

Config of the kernel that moves the particles.

required
flow_config Union[SMCFlowConfig, Literal[False]]

Config of the normalizing flow, or False for none.

required

Methods:

Name Description
initial_parameter_value

Carried parameters for the first mutation (pass through extend_params).

mcmc_init_fn

BlackJAX's mcmc_init_fn; never called, as there is no per-particle state.

mcmc_parameter_update_fn

BlackJAX's mcmc_parameter_update_fn: carries the parameters forward.

read_parameters

BlackJAX's mcmc_step_fn, used only to read the carried parameters.

summary

One mutation's numbers as Python values, for the mutation history and logging.

update_strategy

BlackJAX's update_strategy: returns (update_fn, n_particles).

initial_parameter_value() -> dict[str, Array] ¤

Carried parameters for the first mutation (pass through extend_params).

mcmc_init_fn(position: Array, logdensity_fn: Callable) -> Any ¤

BlackJAX's mcmc_init_fn; never called, as there is no per-particle state.

mcmc_parameter_update_fn(key: Key[Array, ''], state: Any, info: Any) -> dict[str, Array] ¤

BlackJAX's mcmc_parameter_update_fn: carries the parameters forward.

read_parameters(**parameters: Array) -> dict[str, Array] ¤

BlackJAX's mcmc_step_fn, used only to read the carried parameters.

BlackJAX binds the carried parameters to mcmc_step_fn; update_strategy calls it with no arguments to get them back.

summary(update_info: dict) -> dict[str, Any] ¤

One mutation's numbers as Python values, for the mutation history and logging.

update_strategy(mcmc_init_fn: Callable, logposterior_fn: Callable, shared_mcmc_step_fn: Callable, num_mcmc_steps: int, n_particles: int, batch_size: int = 0) -> tuple[Callable, int] ¤

BlackJAX's update_strategy: returns (update_fn, n_particles).

update_fn runs one mutation of all particles and returns the moved particles and the mutation's info dict.