Subgradients and Subdifferentials #
Subgradients of functions f : E → F defined
via the non-negativity of the Bregman divergence 0 ≤ D_[f](y, x, g) on an explicit
domain s (i.e. the linearization error is non-negative on s).
Carrying the domain s : Set E explicitly avoids using indicator functions
while supporting constrained convex analysis.
Main definitions #
HasSubgradientWithinAt f g s x:g : E →L[R] Fis a subgradient offatxons.subdifferentialWithin R f s x: The subdifferential set∂[R, s, x](f).
Main results #
- Monotonicity in the domain:
HasSubgradientWithinAt.mono. - Chain rules:
HasSubgradientWithinAt.comp_affineandHasSubgradientWithinAt.comp. - Equivalence with classical inequality:
hasSubgradientWithinAt_iff_le. - Fermat's rule:
hasSubgradientWithinAt_zero_iff_isMinOn. - Suprema and maxima:
HasSubgradientWithinAt.sup_left,HasSubgradientWithinAt.sup_right,HasSubgradientWithinAt.finset_sup', andHasSubgradientWithinAt.ciSup.
Notation #
∂[R, s, x](f): Scoped notation inBregmanforsubdifferentialWithin R f s x.
g is a subgradient of f at x on domain s (non-negativity of Bregman divergence).
Equations
- HasSubgradientWithinAt f g s x = ∀ y ∈ s, 0 ≤ bregDiv f y x g
Instances For
The subdifferential ∂[R, s, x](f) of f at x on s.
Equations
- subdifferentialWithin R f s x = {g : E →L[R] F | HasSubgradientWithinAt f g s x}
Instances For
Scoped notation for subdifferentialWithin.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Chain rule: g₂ ∘ g₁ is a subgradient of f₂ ∘ f₁ when g₂ is monotone.
Equivalence with the classical definition.
Eta-expanded form of HasSubgradientWithinAt.add
Eta-expanded form of HasSubgradientWithinAt.add_isMinOn
Eta-expanded form of HasSubgradientWithinAt.smul
Eta-expanded form of HasSubgradientWithinAt.sup_left
Eta-expanded form of HasSubgradientWithinAt.sup_right