4 The Complete Calculus
This chapter assembles the assembled Itô formula, stochastic Fubini, the representer rigidity theorem, and the capstone: the complete stochastic calculus theorem. The only axiom in the formalization is stated here [1].
4.1 Assembled Itô Formula
\(\langle Y \rangle := \mathrm{pip}(DY, \Pi DY)\). The quadratic variation in the bounded setting [1].
\(\delta (\varphi '(Y) \cdot \Pi DY) = \varphi '(Y) \cdot \delta (\Pi DY) - \varphi ''(Y) \cdot \langle Y \rangle \). Proved from product rule + chain rule + pip algebra [1, Theorem C].
Apply product rule, substitute chain rule into the pip term, identify the correction via pip_smul.
\(\varphi (Y) = \mathbb {E}[\varphi (Y)] + \varphi '(Y) \cdot \delta (\Pi DY) - \varphi ''(Y) \cdot \langle Y \rangle \). Combines Clark–Ocone with the Itô decomposition [1].
Substitute Clark–Ocone, apply chain rule to rewrite \(\Pi D(\varphi (Y))\), then ito_formula_bounded.
4.2 Stochastic Fubini
\(\int _T \delta (u(t)) \, d\mu (t) = \delta (\int _T u(t) \, d\mu (t))\). In the bounded framework, \(\delta \) is a CLM and CLMs commute with Bochner integration [1, Remark 3.6].
One-line: ContinuousLinearMap.integral_comp_comm.
\(D(\int _T F(t) \, d\mu (t)) = \int _T D(F(t)) \, d\mu (t)\). Same principle: \(D\) is also a CLM [1].
One-line: ContinuousLinearMap.integral_comp_comm applied to \(D\).
4.3 Itô Process
\(Y_t = Y_0 + \int _0^t a_s \, ds + \delta (\mathrm{integrand}_t)\) with drift \(a\) and stochastic integrand [1, Section 6].
For each \(t\), the Itô decomposition holds pointwise in time. Applies ito_formula_bounded at each \(t\) [1, Theorem C].
Instantiate ito_formula_bounded at each time \(t\).
\(\delta (\int \varphi '(Y_s) \Pi DY_s \, d\mu ) = \int \varphi '(Y_s) \delta (\Pi DY_s) \, d\mu - \int \mathrm{correction} \, d\mu \). Combines pointwise Itô with Fubini [1].
Push \(\delta \) inside via stochastic_fubini, apply pointwise Itô, split the integral.
4.4 Gaussian Volterra and fBM
A Gaussian Volterra process: an energy space with Hurst parameter \(H \in (0,1)\), kernel \(k\), and a Gaussian cylindrical structure [1, Section 4].
Gaussian processes admit cylindrical structures via Cameron–Martin; apply cylindrical reduction [1, Theorem 5.3].
One-line: apply cylindrical_implies_leibniz to the Gaussian cylindrical structure.
Leibniz holds for fBM with all \(H \in (0,1)\), including \(H \neq 1/2\) where fBM is not a semimartingale [1].
Delegate to leibniz_gaussian.
4.5 Representer Rigidity
With deterministic representers: \(\mathbb {E}[\langle DF, u \rangle ] = \sum _i \mathbb {E}[\partial _i F] \cdot \mathbb {E}[\langle \kappa _i, u \rangle ]\). The expectation factorizes [1, Theorem 6.1].
Expand \(DF\) via IBP rep, distribute pip and expectation, apply the deterministic hypothesis.
If \(\sigma \) is stochastic and Brownian representers are nondegenerate, then \(M\)-representers are not deterministic. Contrapositive: deterministic \(M\)-representers force \(\mathrm{pip}(\kappa _i^W, u) = 0\), contradicting nondegeneracy [1, Theorem E].
Assume deterministic, derive \(\mathrm{mul}(\sigma ^{-1}, \mathrm{pip}(\kappa _i, u))\) is constant, cancel by nondegeneracy, obtain contradiction with nontrivial space.
4.6 Spectral Properties
\(N := D \circ \delta : L^2(\Omega ; H) \to L^2(\Omega ; H)\). The Ornstein–Uhlenbeck generator [3].
\(\langle Nu, v \rangle = \langle u, Nv \rangle \). From \(\langle D(\delta u), v \rangle = \langle \delta u, \delta v \rangle = \langle u, D(\delta v) \rangle \) [3].
Unfold \(N\), apply adjoint identity twice, commute via real_inner_comm.
\(\langle Nu, u \rangle = \| \delta u\| ^2 \geq 0\) [3].
Unfold, apply adjoint identity, conclude with real_inner_self_nonneg.
4.7 Log-Sobolev and the One Axiom
\(\Gamma (f, f) \leq \Gamma _2(f, f)\): the Ornstein–Uhlenbeck generator has curvature \(\kappa = 1\). Proved: \((f')^2 \leq (f'')^2 + (f')^2\) [1].
Unfold definitions, conclude with sq_nonneg.
The Bakry–Émery log-Sobolev inequality for the Gaussian measure: \(\mathrm{Ent}_\gamma (f^2) \leq 2 \int (f')^2 \, d\gamma \). This is the ONE axiom in the formalization. The curvature bound \(\Gamma _2 \geq \Gamma \) is proved; the semigroup interpolation argument (Bakry–Émery 1985) requires the OU semigroup which is not yet in Mathlib [1].
\(\mathrm{Ent}_\gamma (f^2) \leq 2 \int (f')^2 \, d\gamma \) with sharp constant \(C = 2\). Derived from the Bakry–Émery axiom [1].
Direct application of bakry_emery_log_sobolev.
4.8 The Capstone
Given an isonormal process with Gaussian structure: the adjoint identity, Malliavin derivative, Itô isometry, centeredness, norm preservation, Gaussian moments, and \(L^p\) multiplication all hold. This is the paper’s thesis in one Lean theorem: \(\delta \to D = \delta ^* \to \) adjoint identity \(\to \) Clark–Ocone \(\to \) Leibniz \(\to \) Itô [1].
Seven-fold conjunction assembled from separately proved theorems.
4.9 Density Extension and Concrete Assembly
If a jointly continuous bilinear map \(T\) vanishes on a dense subspace \(S\), then \(T = 0\) everywhere. The abstract content of the Leibniz closure step [1].
Fix one argument in \(S\), use isClosed_ker + density to extend to all of \(V\); then flip and repeat.
The concrete Brownian energy space: given an isonormal process \(W\), compatible multiplication, scalar action, and pointwise inner product on Mathlib’s \(L^p\) spaces, all axioms are proved from concrete computations [1, 6].
\(\langle v, Pv \rangle = \langle Pv, Pv \rangle \) for an idempotent self-adjoint \(P\) [1, Lemma 6.3].
Show \(\langle v - Pv, Pv \rangle = 0\) using idempotence and self-adjointness, then rearrange.