aquakin.LaplaceConfig#

class aquakin.LaplaceConfig(method='fd', ridge=1e-06, eig_keep=0.01, fd_step=0.001, dtmax=None)[source]#

Bases: object

Laplace-posterior settings.

Pass an instance to laplace= to enable the posterior with tuning; laplace=True uses these defaults and laplace=False disables it.

Parameters:

  • method ({"fd", "gauss_newton"}) – How the unconstrained-space Hessian is formed: "fd" (default) finite-differences the gradient; "gauss_newton" uses J^T J with the residual Jacobian by AD (PSD by construction, Fisher for nll).

  • ridge (float) – Ridge added to the Hessian before inversion.

  • eig_keep (float) – Relative eigenvalue floor for the eigen-truncated covariance.

  • fd_step (float) – Relative step for the finite-difference Hessian (method="fd").

  • dtmax (float, optional) – If given, rebuild the reactor with this dtmax cap for the Hessian pass only – a tighter solve for the second-order-sensitive covariance.

__init__(method='fd', ridge=1e-06, eig_keep=0.01, fd_step=0.001, dtmax=None)#
Parameters:

  • method (str)

  • ridge (float)

  • eig_keep (float)

  • fd_step (float)

  • dtmax (float | None)

Return type:

None

Methods

__init__([method, ridge, eig_keep, fd_step, ...])

Attributes

dtmax

eig_keep

fd_step

method

ridge