The Operator Derivative in Continuous Stochastic Calculus

1 The Operator Derivative

This chapter formalizes the operator derivative \(D := \delta ^*\) as the Hilbert space adjoint of the stochastic integral, and establishes the representation layer: Clark–Ocone, GKW orthogonality, and the variance identity. All results follow from adjointness alone, without assuming any chain rule, Leibniz rule, or Gaussianity [1].

1.1 Unbounded Operator Foundation

The stochastic integral \(\delta \) is a densely defined operator, and \(D := \delta ^\dagger \) is constructed via Mathlib’s LinearPMap.adjoint.

Definition 1 Operator derivative (unbounded)
✓
#

Given a densely defined stochastic integral \(\delta : L^2(\Omega ; H) \to _{\ell } L^2(\Omega )\), the operator derivative is \(D := \delta ^\dagger \), the partially defined adjoint constructed by Mathlib [1, Definition 2.5].

Theorem 2 D is closed
✓
#

The operator derivative \(D = \delta ^\dagger \) is a closed operator, automatic from Mathlib’s LinearPMap.adjoint_isClosed [1, Proposition 2.7(b)].

Proof ▶

Immediate from LinearPMap.adjoint_isClosed applied to the dense domain hypothesis.

Theorem 3 Adjoint identity (unbounded)
✓
#

For \(F \in \mathrm{dom}(D)\) and \(u \in \mathrm{dom}(\delta )\): \(\langle D F, u \rangle = \langle F, \delta u \rangle \). This is the definition of \(D\) via Riesz, proved by Mathlib.

Proof ▶

Follows from LinearPMap.adjoint_isFormalAdjoint.

1.2 Unbounded Energy Space

The full probabilistic structure needed for Clark–Ocone in the unbounded setting.

Definition 4 Unbounded energy space
✓
#

A structure carrying Hilbert spaces \(L^2(\Omega )\) and \(L^2(\Omega ; H)\), a densely defined stochastic integral \(\delta \), predictable projection \(\Pi \), expectation functional, algebraic operations (mul, smul, pip), and bridge axioms connecting inner products to expectations [1, Definition 2.1].

Definition 5 Operator derivative D (unbounded)
✓
#

The operator derivative \(D := \delta ^\dagger \) constructed as the unbounded adjoint within an UnboundedEnergySpace [1, Definition 2.5].

Theorem 6 D is closed (energy space)
✓
#

\(D\) is a closed operator. From Mathlib [1, Proposition 2.7(b)].

Proof ▶

Delegate to LinearPMap.adjoint_isClosed with the dense domain hypothesis.

Theorem 7 Adjoint identity (energy space)
✓
#

\(\langle D F, u \rangle = \langle F, \delta u \rangle \) for \(F \in \mathrm{dom}(D)\), \(u \in \mathrm{dom}(\delta )\).

Proof ▶

From LinearPMap.adjoint_isFormalAdjoint.

Definition 8 PRP (unbounded)
✓
#

The Predictable Representation Property: every centered \(F \in L^2(\Omega )\) can be written as \(\delta (v)\) for some predictable \(v \in \mathrm{dom}(\delta )\) [1, Definition 3.1].

Definition 9 Isometry condition (unbounded)
✓

\(\langle \delta u, \delta v \rangle = \langle u, v \rangle \) for predictable \(u, v \in \mathrm{dom}(\delta )\). This is the Itô isometry [1, Definition 2.3].

1.3 Representation Layer

Clark–Ocone, GKW orthogonality, and the variance identity are proved from adjointness, PRP, and the isometry condition alone.

Theorem 10 Clark–Ocone (unbounded)
✓

For \(F \in \mathrm{dom}(D)\) with \(\Pi (DF) \in \mathrm{dom}(\delta )\): \(F = \mathbb {E}[F] + \delta (\Pi D F)\). The proof uses PRP to obtain a representing element \(v\), then isometry + adjointness to identify \(\Pi (DF) = v\) [1, Theorem 3.2].

Proof ▶

PRP gives \(v\) with \(\delta (v) = F - \mathbb {E}[F]\); inner product testing with adjointness and isometry shows \(\Pi (DF) = v\) by density.

Theorem 11 GKW orthogonality (unbounded)
✓
#

For predictable \(u \in \mathrm{dom}(\delta )\): \(\langle F - \mathbb {E}[F] - \delta (\Pi DF), \delta u \rangle = 0\). The GKW residual is orthogonal to the stable subspace [1, Theorem 3.4].

Proof ▶

Expand inner products, apply adjoint identity, isometry, and self-adjointness of \(\Pi \).

Theorem 12 Variance identity (unbounded)
✓
#

\(\| F - \mathbb {E}[F]\| ^2 = \| \Pi DF\| ^2\). From Clark–Ocone (\(F - \mathbb {E}[F] = \delta (\Pi DF)\)) and the isometry condition [1, Corollary 3.5].

Proof ▶

Substitute Clark–Ocone, then apply isometry.

1.4 Bounded Working Framework

For the Itô integral setting where \(\delta \) is bounded, \(D = \mathrm{ContinuousLinearMap.adjoint}(\delta )\).

Definition 13 Energy space (bounded)
✓
#

A structure with \(L^2(\Omega )\), \(L^2(\Omega ; H)\), bounded stochastic integral \(\delta : L^2(\Omega ; H) \to _L L^2(\Omega )\), predictable projection, expectation, and algebraic operations. All downstream theorems are proved in this setting [1, Section 2].

Definition 14 D (bounded)
✓
#

\(D := \mathrm{ContinuousLinearMap.adjoint}(\delta )\). Constructed, not assumed. Everywhere-defined and continuous [1, Definition 2.5].

Theorem 15 Adjoint identity (bounded)
✓
#

\(\langle D F, u \rangle _{L^2(\Omega ; H)} = \langle F, \delta u \rangle _{L^2(\Omega )}\) for all \(F, u\) [1, Proposition 2.7(a)].

Proof ▶

Unfold \(D\) and apply ContinuousLinearMap.adjoint_inner_left.

Theorem 16 D is linear
✓
#

\(D(\alpha F + \beta G) = \alpha \, DF + \beta \, DG\) [1, Proposition 2.7(a)].

Proof ▶

From map_add and map_smul for the continuous linear map \(D\).

Theorem 17 D annihilates constants
✓
#

\(D(\mathrm{const}(c)) = 0\) for all \(c \in \mathbb {R}\) [1, Proposition 2.7(c)].

Proof ▶

Test \(\langle D(\mathrm{const}(c)), u \rangle = \langle \mathrm{const}(c), \delta u \rangle = c \cdot \mathbb {E}[\delta u] = 0\) for all \(u\), then apply \(\langle x, x \rangle = 0 \Rightarrow x = 0\).

Theorem 18 Uniqueness of D
✓
#

If \(\langle g_1, u \rangle = \langle F, \delta u \rangle = \langle g_2, u \rangle \) for all \(u\), then \(g_1 = g_2\) [1, Proposition 2.7(d)].

Proof ▶

\(\langle g_1 - g_2, u \rangle = 0\) for all \(u\); take \(u = g_1 - g_2\).

Definition 19 PRP (bounded)
✓
#

Every centered \(F \in L^2(\Omega )\) has a predictable representative: \(\exists v\), \(\Pi v = v\) and \(\delta v = F\) [1, Definition 3.1].

Definition 20 Isometry condition (bounded)
✓
#

\(\langle \delta u, \delta v \rangle = \langle u, v \rangle \) for predictable \(u, v\) [1, Definition 2.3].

Definition 21 Full isometry condition
✓
#

\(\langle \delta u, \delta v \rangle = \langle \Pi u, \Pi v \rangle \) for all \(u, v\). Stronger than IsometryCondition: \(\delta \) factors through \(\Pi \) isometrically [1].

Theorem 22 Bridge theorem
✓
#

When \(\delta \) is bounded and restricted to a dense submodule, the LinearPMap adjoint equals the ContinuousLinearMap adjoint. From Mathlib’s toPMap_adjoint_eq_adjoint_toPMap_of_dense.

Proof ▶

One-line delegation to ContinuousLinearMap.toPMap_adjoint_eq_adjoint_toPMap_of_dense.

Theorem 23 PRP from full isometry
✓
#

Full isometry + \(\ker (D) \subseteq \mathrm{constants}\) + closed range of \(\delta \) implies PRP. Reduces PRP to functional analysis: the Hilbert space decomposition \(L^2 = \overline{\mathrm{Im}(\delta )} \oplus \ker (D)\) [1, Theorem 3.3].

Proof ▶

Combines fullIso_implies_iso, fullIso_implies_range, and fullIso_implies_closed.