3 The Gaussian Extension
This chapter connects the abstract framework to Mathlib’s Gaussian infrastructure. The isonormal process, cylindrical functionals, Stein’s lemma, and the Malliavin derivative are formalized. The concrete Itô formula is proved from Stein’s lemma and adjointness [1, 3].
3.1 Isonormal Process
\(W : H \to _L L^2(\Omega , P)\) with isometry \(\langle W(h), W(k) \rangle = \langle h, k \rangle _H\), centeredness \(\mathbb {E}[W(h)] = 0\), and Gaussian marginals [3, Definition 1.1.1].
\(F = f(W(h_1), \ldots , W(h_n))\) with orthonormal \(\{ h_i\} \), smooth \(f\), and \(L^2\) partial derivatives \(\partial _i F\) [3, Definition 1.2.1].
Class asserting that each marginal \(W(h)\) has Gaussian law \(\mathcal{N}(0, \| h\| ^2)\) (grounded in Mathlib’s ProbabilityTheory.IsGaussian), per-coordinate Stein identity, and cylindrical orthogonality [3, 4].
3.2 Stein’s Lemma
\(\varphi '(x) = -x \cdot \varphi (x)\) where \(\varphi \) is the standard Gaussian density. The key identity underlying all Gaussian IBP [3].
Differentiate \(\exp (-x^2/2)\) via HasDerivAt.exp, multiply by the normalizing constant.
\(\int f(x) \cdot x \, d\gamma (x) = \int f'(x) \, d\gamma (x)\) for \(\gamma = \mathcal{N}(0,1)\). Proved via integration by parts using \(\varphi ' = -x\varphi \) [3].
Convert to Lebesgue with density \(\varphi \), rearrange via \(\varphi ' = -x\varphi \), apply integral_mul_deriv_eq_deriv_mul_of_integrable.
Class: \(\langle F, W(h) \rangle = \sum _i \langle h_i, h \rangle \cdot \mathbb {E}[\partial _i F]\) for cylindrical \(F\) and all \(h \in H\) [3, Proposition 1.3.1].
From IsonormalIsGaussian, construct a SteinLemma instance. Uses per-coordinate Stein, cylindrical orthogonality, and the orthogonal decomposition \(h = h_\parallel + h_\perp \) [3].
Decompose \(h\) via projection onto \(\mathrm{span}\{ h_i\} \), use per-coordinate Stein for \(h_\parallel \) and orthogonality for \(h_\perp \), expand in the ONB.
3.3 Malliavin Derivative
\(DF := \sum _i (\partial _i F)(\omega ) \cdot h_i \in L^2(\Omega ; H)\). Constructed using L2_smul_const [3, Definition 1.2.1].
\(\langle DF, \mathrm{const}_1 \cdot h \rangle _{L^2(\Omega ; H)} = \langle F, W(h) \rangle _{L^2(\Omega )}\) for cylindrical \(F\). The adjoint identity derived from Stein’s lemma [1, 3].
Expand LHS via ibp_algebraic_step, match with RHS via Stein, convert inner product with constant 1 to expectation.
\(D(\varphi (W(h))) = \varphi '(W(h)) \cdot h\) for smooth \(\varphi \) and \(h \in H\). Special case of the cylindrical construction with \(n = 1\) [3].
Unfold malliavin_derivative_of_cylindrical with \(n = 1\), simplify the singleton sum.
\(D(\varphi (F)) = \sum _i (\varphi '(F) \cdot \partial _i F) \cdot h_i\) for cylindrical \(F\) and smooth \(\varphi \). The chain rule for compositions of cylindricals [3].
Unfold malliavin_derivative_of_cylindrical for the composed cylindrical; the result is definitional (rfl).
3.4 Concrete Itô Formula
\(\mathbb {E}[\varphi (W(h)) \cdot W(k)] = \langle h, k \rangle _H \cdot \mathbb {E}[\varphi '(W(h))]\). The covariance form of the Itô formula [1, Remark 3.6].
Apply malliavin_adjoint, substitute malliavin_of_smooth, decompose via L2_smul_const_inner.
\(\mathbb {E}[\varphi '(W(h)) \cdot W(h)] = \| h\| ^2 \cdot \mathbb {E}[\varphi ''(W(h))]\). Apply level 1 to \(\varphi '\) with \(k = h\) [1].
Instantiate ito_adjoint_level1 with \(\varphi ' \mapsto \varphi ''\) and \(k = h\).
For \(\| h\| = 1\): \(\mathbb {E}[\varphi '(W(h)) \cdot W(h)] = \mathbb {E}[\varphi ''(W(h))]\). The \(\frac{1}{2}\varphi ''\) correction for unit-norm Gaussians [1].
Apply level 2, then \(\langle h, h \rangle = \| h\| ^2 = 1\).
3.5 Gaussian Moments and Itô Isometry
If \(\mu \) is Gaussian (Mathlib’s IsGaussian), then \(\mathrm{id} \in L^p(\mu )\) for all finite \(p\). This is Fernique’s theorem from Mathlib [3, 6].
One-line: ProbabilityTheory.IsGaussian.memLp_id.
\(W(h) \in L^p\) for any finite \(p\). For \(p \leq 2\): monotonicity. For \(p {\gt} 2\): Fernique’s theorem via the Gaussian marginal [3, 6].
Case split on \(p \leq 2\) (monotonicity) vs. \(p {\gt} 2\) (Fernique + comp_of_map).
\(\| \sum _i c_i W(h_i)\| ^2 = \| \sum _i c_i h_i\| ^2_H\). From the isonormal isometry property [1, 3].
Expand both sides via bilinearity, apply WP.isometry termwise.
3.6 Skorokhod Integral
A simple process \(u(\omega ) = \sum _i f_i(\omega ) \cdot h_i\) with \(f_i \in L^2(\Omega )\) and \(h_i \in H\) [3].
\(\delta (\sum _i c_i h_i) := \sum _i c_i W(h_i)\). Constructed from \(W\) linearity [3].
\(\langle DF, \sum _i c_i h_i \rangle = \langle F, \sum _i c_i W(h_i) \rangle \) for cylindrical \(F\) and deterministic simple \(u\). Extends the constant-\(h\) adjoint identity by linearity [1, 3].
Expand inner products by linearity, apply malliavin_adjoint termwise.