Isonormal Gaussian families turn Hilbert directions into random coordinates
Objective: For “Isonormal Gaussian families turn Hilbert directions into random coordinates,” identify the type of the derivative or operator, the hypothesis supporting the decisive proof arrow, and the first conclusion lost when Linearity and pairwise covariance alone do not define a jointly Gaussian family unless joint Gaussianity is included or constructed.
Unit 1, chapter 1 begins with a concrete question: “Isonormal Gaussian families turn Hilbert directions into random coordinates.” The index h is a deterministic direction, while W(h) is a centered Gaussian random variable. Malliavin calculus differentiates a random output with respect to a declared direction in the underlying Gaussian noise. Before manipulating D, δ, L, or a covariance matrix, we name the probability space, the isonormal Gaussian family, the Cameron–Martin direction, the random variable, and the norm in which the operation is defined.
Read the formal statement one object at a time: An isonormal Gaussian family over a real separable Hilbert space H is a linear map W:H→L²(Ω) with E[W(h)W(g)]=⟨h,g⟩_H. The derivative DF is not an ordinary derivative with a random symbol attached. It is an H-valued random object; H records admissible noise directions, while the surrounding Lp space records random integrability. Every domain such as D^{1,p}, Dom(δ), or Dom(L) is therefore part of the mathematical claim.
The bounded theorem target is: For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩). We keep equality in Lp, equality almost surely, convergence in a Gaussian Sobolev norm, and equality of distributions separate. A formula first proved for smooth cylindrical variables extends only after the relevant operator is closable and the approximating sequence is controlled in its graph norm.
The proof route is Start from An isonormal Gaussian family over a real separable Hilbert space H is a linear map W:H→L²(Ω) with E[W(h)W(g)]=⟨h,g⟩_H. Prove the identity on smooth cylindrical variables, transport it through the stated Gaussian Sobolev or operator closure, and stop exactly at For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩). Each arrow names its input type and hypothesis. Finite-dimensional Gaussian integration by parts supplies the initial identity; closure, duality, chaos orthogonality, semigroup contraction, or a tangent equation then transports it to the stated infinite-dimensional object without silently strengthening the conclusion.
A calculation makes the types visible: Let e_1,e_2 be orthonormal and h=2e_1−e_2; compute Var W(h) and Cov(W(h),W(e_1+e_2)). We retain the Gaussian coordinates, covariance or Gram matrix, derivative vectors, chaos coefficients, time grid, and exact fractions before forming a determinant, divergence, generator value, conditional projection, or approximation error. The audited endpoint is The variance is ||h||²=5 and the covariance is ⟨2e_1−e_2,e_1+e_2⟩=1.
The first invalid step under changed assumptions is also part of the lesson: Linearity and pairwise covariance alone do not define a jointly Gaussian family unless joint Gaussianity is included or constructed. A responsible argument reports the weaker statement that survives, the extra nondegeneracy, integrability, adaptedness, smoothness, bracket, or inverse-moment condition needed for repair, and the claim that the finite model cannot establish.
Keep three layers separate. The algebra layer differentiates or expands a declared smooth functional; the analytic layer proves that the operator, inverse covariance, or density estimate belongs to a stated space; the probabilistic layer interprets the result as sensitivity, representation, absolute continuity, smooth density, or an approximation bound. Passing one layer does not certify the other two.
The chapter closes with reconstruction rather than recognition. Starting from the raw model, you must rebuild the definition, justify the theorem’s decisive arrow, reproduce the exact endpoint, activate the failure mutation, and state a repaired conclusion. Merely naming a theorem or obtaining one matching decimal is not completion.
Bounded theorem for this chapter: For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩). The equality or implication has exactly the smoothness, integrability, domain, and convergence meaning declared here, and no stronger frontier conclusion is silently included.
Type every object in “An isonormal Gaussian family over a real separable Hilbert space H is a linear map W:H→L²(Ω) with E[W(h)W(g)]=⟨h,g⟩_H..” A cylindrical functional is a function of finitely many Gaussian coordinates; its gradient is transported back into H. For a limiting object, choose a cylindrical approximating sequence and state the graph norm that controls both the variables and their derivatives.
Prove the finite-coordinate identity first. Write the Gaussian density or covariance matrix explicitly, differentiate only ordinary smooth functions, integrate by parts with a justified boundary limit, and translate the coordinate gradient into the H inner product. This is the inspectable origin of the infinite-dimensional notation.
Establish the first implication in the declared route: Start from An isonormal Gaussian family over a real separable Hilbert space H is a linear map W:H→L²(Ω) with E[W(h)W(g)]=⟨h,g⟩_H. Prove the identity on smooth cylindrical variables, transport it through the stated Gaussian Sobolev or operator closure, and stop exactly at For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩). Record whether the step uses density of cylindrical variables, an adjoint definition, orthogonal chaos decomposition, a contraction estimate, a commutation rule, or a linearized stochastic equation.
Control the passage to the target space. If an unbounded operator appears, prove convergence of both input and image; if an inverse matrix appears, declare the event of invertibility and the moment of its determinant; if a stochastic integral appears, distinguish adapted Itô integration from anticipative divergence integration.
Compute the finite checkpoint independently: Let e_1,e_2 be orthonormal and h=2e_1−e_2; compute Var W(h) and Cov(W(h),W(e_1+e_2)). Preserve exact covariance entries and derivative components, then verify The variance is ||h||²=5 and the covariance is ⟨2e_1−e_2,e_1+e_2⟩=1. The computation checks the typed identity, but it does not replace the closure or uniform-integrability argument required for a general limit.
Test the mutation before concluding: Linearity and pairwise covariance alone do not define a jointly Gaussian family unless joint Gaussianity is included or constructed. Locate the first proof line that loses meaning or justification. Keep earlier identities, discard only unsupported consequences, and name one concrete hypothesis that would repair that line.
The preceding steps establish exactly “For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩)..” They do not prove a full Hörmander theorem, a rough-path lift, an SPDE regularity structure, jump-noise calculus, or an application-specific calibration claim unless those additional objects and estimates have separately been constructed.
Worked model to reconstruct from exact inputs: Let e_1,e_2 be orthonormal and h=2e_1−e_2; compute Var W(h) and Cov(W(h),W(e_1+e_2)). Keep the named Gaussian coordinates, basis directions, covariance entries, time cells, and parameters visible until the final comparison.
- Write the complete finite input for “Let e_1,e_2 be orthonormal and h=2e_1−e_2; compute Var W(h) and Cov(W(h),W(e_1+e_2)).”: list Gaussian coordinates, covariance entries, functions, time cells, and every declared parameter. Verify covariance positivity and all probability normalizations before differentiation.
- Apply the chapter definition to the raw functional. Differentiate one coordinate at a time, keep basis vectors attached, and assemble the H-valued derivative or matrix without erasing indices.
- Follow Start from An isonormal Gaussian family over a real separable Hilbert space H is a linear map W:H→L²(Ω) with E[W(h)W(g)]=⟨h,g⟩_H. Prove the identity on smooth cylindrical variables, transport it through the stated Gaussian Sobolev or operator closure, and stop exactly at For h_1,…,h_m, the vector (W(h_1),…,W(h_m)) is centered Gaussian with covariance equal to the Gram matrix (⟨h_i,h_j⟩). Compute the intermediate inner product, divergence, chaos coefficient, semigroup factor, tangent solution, bracket direction, or conditional projection requested by the model.
- Recompute the endpoint by a second route, such as an ordinary Gaussian moment, covariance calculation, direct conditional expectation, or matrix identity. Compare types as well as numbers.
- Both valid routes meet at The variance is ||h||²=5 and the covariance is ⟨2e_1−e_2,e_1+e_2⟩=1. State the exact hypothesis that made them agree, then activate the boundary mutation and record the first discrepancy rather than forcing the old formula to survive.
Result: Audited endpoint: The variance is ||h||²=5 and the covariance is ⟨2e_1−e_2,e_1+e_2⟩=1. This is the result of the displayed finite model under its stated hypotheses; it checks the chapter identity but does not by itself prove an unrestricted infinite-dimensional theorem.