Theorem B — subdifferentials of lsc proper convex functions are exactly the maximal cyclically monotone operators, unique up to a constant
ProvedRockafellarMaxMono.Cyclic.maximal_cyclically_monotone_iff_subdiffLet be a real Banach space with dual , and let be a multivalued mapping. Then:
- There exists a lower semicontinuous proper convex function on with if and only if is a maximal cyclically monotone operator.
- In this case determines uniquely up to an additive constant: if and are lower semicontinuous proper convex functions on with and , then there is a real constant with
Theorem B characterizes the subdifferential mappings of lower semicontinuous proper convex functions on a Banach space intrinsically, as the maximal cyclically monotone operators, and shows that such an operator has an essentially unique convex "potential". It completes an argument of Rockafellar's 1966 paper in which the maximality and uniqueness steps had a gap in the nonreflexive case.
Formalization Note is equality of multivalued mappings, for all . Maximality is among cyclically monotone operators (see the definition). Both parts of the theorem are stated as one conjunction. The constant is real, so on points where the uniqueness equation reads .
import Mathlib import Definitions.Def_RockafellarMaxMono_Shared_ProperConvex import Definitions.Def_RockafellarMaxMono_Shared_Subdiff import Definitions.Def_RockafellarMaxMono_Cyclic_CyclicallyMonotone
namespace RockafellarMaxMono.Cyclic
theorem maximal_cyclically_monotone_iff_subdiff {E : Type*} [NormedAddCommGroup E]
[NormedSpace ℝ E] [CompleteSpace E] (T : E → Set (StrongDual ℝ E)) :
((∃ f : E → EReal, Shared.ProperConvex f ∧ LowerSemicontinuous f ∧ ∀ x, T x = Shared.subdiff f x) ↔
IsMaximalCyclicallyMonotone T) ∧
∀ f g : E → EReal, Shared.ProperConvex f → LowerSemicontinuous f →
Shared.ProperConvex g → LowerSemicontinuous g →
(∀ x, T x = Shared.subdiff f x) → (∀ x, T x = Shared.subdiff g x) →
∃ c : ℝ, ∀ x, g x = f x + (c : EReal) := by sorry
end RockafellarMaxMono.Cyclic
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Here is a real Banach space: a real normed vector space that is complete in its norm. is its continuous dual, the space of continuous linear functionals . is the extended real line. The statement involves a single fixed set-valued map
which assigns to each point an arbitrary set of continuous linear functionals. The set may be empty. The statement uses three notions defined in this bundle, and it depends on their definitions. Their code is not reproduced here, so they are named but not unfolded:
- "proper convex" (
Shared.ProperConvex) is a property of a function . - "the subdifferential " (
Shared.subdiff f x) is a subset of attached to a function and a point . - "maximal cyclically monotone" (
IsMaximalCyclicallyMonotone) is a property of the map .
"Lower semicontinuous" is the standard notion for a function . carries its norm topology and carries its order topology.
The statement asserts that two claims both hold.
(1) Characterization. The following are equivalent:
- there exists a function that is proper convex (in the bundle's sense), lower semicontinuous, and satisfies
- is maximal cyclically monotone (in the bundle's sense).
(2) Uniqueness up to an additive constant. Take any two functions that are each proper convex and lower semicontinuous, with and for every . Then there is a real number such that
where the addition is in . Because is finite, , so this forces and to take the value at exactly the same points, and likewise . It also forces to equal the same constant at every point where both are finite. Part (2) does not assume that is maximal cyclically monotone. It is stated for the same arbitrary and applies to any pair meeting the listed hypotheses.
On degenerate cases: is never empty, since it contains . may be the zero space . Then , each is either or , and part (2) says that any two admissible differ at the single point by a real constant. Whether such exist, and whether values or everywhere- functions are allowed, depends entirely on the bundle's definition of "proper convex". The map may have empty values at some or all points. Whether such a can equal a subdifferential, or be maximal cyclically monotone, again depends on the bundle's definitions of and of maximal cyclic monotonicity. If no proper convex, lower semicontinuous has , then part (2) holds vacuously, and part (1) says exactly that is then not maximal cyclically monotone. The addition never meets the undefined form , because is a real number.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.