integrate_stochastic

bayesflow.utils.integrate_stochastic(drift_fn: Callable, diffusion_fn: Callable, state: Dict[str, int | float | Tensor], start_time: int | float | Tensor, stop_time: int | float | Tensor, seed: SeedGenerator, steps: int | Literal['adaptive'] = 100, method: str = 'euler_maruyama', min_steps: int = 50, max_steps: int = 1000, score_fn: Callable = None, corrector_steps: int = 0, noise_schedule=None, step_size_factor: int | float | Tensor = 0.01, **kwargs) → Dict[str, int | float | Tensor]

Integrate a stochastic differential equation from start_time to stop_time.

This function dispatches to either fixed-step or adaptive-step integration logic, depending on the selected integration method and the value of steps.

Parameters:

drift_fncallable

Function computing the drift term of the SDE. It should accept the current state and time as inputs.

diffusion_fncallable

Function computing the diffusion term of the SDE. It should accept the current state and time as inputs.

stateStateDict

Dictionary containing the initial state of the system.

start_timearray-like

Starting time for integration.

stop_timearray-like

Ending time for integration.

seedkeras.random.SeedGenerator

Random seed generator used for noise generation.

stepsint or {‘adaptive’}, optional

Number of integration steps for fixed-step integration, or 'adaptive' to enable adaptive step sizing. Adaptive integration is only supported by the 'shark' method. Default is 100.

methodstr, optional

Integration method to use (e.g., 'euler_maruyama' or 'shark'). Default is 'euler_maruyama'.

min_stepsint, optional

Minimum number of steps for adaptive integration. Default is 50.

max_stepsint, optional

Maximum number of steps for adaptive integration. Noise is pre-generated up to this number of steps, which may increase memory usage. Default is 1000.

score_fncallable, optional

Score function used for predictor–corrector sampling. If None, no corrector step is applied.

corrector_stepsint, optional

Number of corrector steps applied after each predictor step. Default is 0.

noise_scheduleobject, optional

Noise schedule object used to compute alpha_t during the corrector step. Required if corrector_steps > 0.

step_size_factorarray-like, optional

Scaling factor applied to the corrector step size. Default is 0.01.

**kwargs

Additional keyword arguments passed to the underlying step function.

Returns:

StateDict

Final state dictionary after integration.