`shrinkage.HybridStep`

Shrinkage for Python · python · 0.2.0

HybridStep(
    kind: typing.Literal["backtracking", "fixed_bound"],
    initial_curvature: float | None = None,
    max_backtracks: int = 100
)

Curvature policy for hybrid proximal-gradient transitions.

Use backtracking or fixed_bound. A fixed bound uses the design's Frobenius bound. Backtracking doubles curvature after rejected trials. Logistic outer step halvings have a separate budget on HybridCoordinateDescent.

`kind`

Source: packages/python/src/shrinkage/solver.py

kind: typing.Literal["backtracking", "fixed_bound"]

`initial_curvature`

Source: packages/python/src/shrinkage/solver.py

initial_curvature: float | None = None

`max_backtracks`

Source: packages/python/src/shrinkage/solver.py

max_backtracks: int = 100

`backtracking`

Source: packages/python/src/shrinkage/solver.py

backtracking(
    *,
    initial_curvature: float | None = None,
    max_backtracks: int = 100
) -> shrinkage.HybridStep

Use a positive seed or the largest averaged column curvature.

`fixed_bound`

Source: packages/python/src/shrinkage/solver.py

fixed_bound() -> shrinkage.HybridStep

Use a computed global curvature bound without a spectral estimate.