2 Lévy–Itô Framework
This chapter constructs the jump-side stochastic calculus from first principles. The strategy is to avoid axiomatizing the compensated Poisson integral: instead, we build it as the \(L^2\) limit of compound Poisson finite sums (which require only elementary summation, not measure-theoretic integration). The finite telescoping identity replaces the Itô formula, the \(L^2\) Cauchy estimate \(\varepsilon ^{2-\gamma } \to 0\) guarantees convergence, and centering and isometry are derived (not assumed) via tendsto_nhds_unique. The chapter also develops the Poisson chaos orthogonality on the canonical space and constructs the compound Poisson path as an explicit càdlàg function.
2.1 Truncated Measures
The stable Lévy measure is restricted to \(\{ |z| \leq M\} \) to produce a finite measure suitable for compound Poisson approximation.
\(\nu _\gamma ^M := \nu _\gamma |_{\{ |z|\leq M\} }\), the restriction to \(\{ |z| \leq M\} \). This finite measure enables compound Poisson approximation: the Itô formula reduces to a finite telescoping sum, and the full integral is recovered as \(M \to \infty \).
\(\int _{|z|\leq M} z^2\, \nu _\gamma (dz) = 2c_\gamma M^{2-\gamma }/(2-\gamma ) {\gt} 0\). Provides the nondegenerate isometry constant for \(\delta _L^M\), ensuring that the compensated integral norm is strictly positive.
Apply div_pos to the product \(2c_\gamma \cdot M^{2-\gamma } {\gt} 0\) (from mul_pos and rpow_pos_of_pos) and the denominator \(2 - \gamma {\gt} 0\) (from linarith).
2.2 Finite Telescoping
Finite telescoping identities replace the Itô formula in the compound Poisson setting, reducing path-level calculus to algebraic sums.
\(f(y_n) - f(y_0) = \sum _{i=0}^{n-1} [f(y_{i+1}) - f(y_i)]\). Replaces stochastic integration in the compound Poisson setting, reducing path-level identities to algebraic sums.
Direct application of (Finset.sum_range_sub (f \(\circ \) y) n).symm.
\(f(v+z) - f(v) = f'(v)z + [f(v+z) - f(v) - f'(v)z]\): the linear part (odd in \(z\), detectable by \(D_L\)) plus the nonlinear remainder (contributing to \(\Gamma _L\)). This decomposition underpins both the product rule and the chaos characterization.
Closed by ring.
Telescoping for \(h = ab\) gives the compound Poisson product rule: a Leibniz part plus a defect \([a(y_{i+1}) - a(y_i)][b(y_{i+1}) - b(y_i)]\). The defect term is the discrete analogue of the carré du champ operator.
Apply finite_telescope to \(h(x) = a(x) \cdot b(x)\).
2.3 Compound Poisson Integral
The compound Poisson integral is defined as a finite random sum with compensation, avoiding measure-theoretic convergence issues.
The finite random sum \(\sum _{i=1}^{N} h(s_i, z_i) - \mathrm{compensator}\), well-defined pointwise since \(\nu _\gamma ^M\) has finite total mass. Finiteness of the sum avoids the convergence issues that arise for the full Lévy integral.
\(\sum (ch_1 + h_2) = c\sum h_1 + \sum h_2\), from linearity of finite sums. This ensures that the compound Poisson integral inherits linearity from ordinary summation.
Unfold random_sum, apply Finset.sum_add_distrib and Finset.mul_sum.
2.4 Compensated Integral Construction
The compensated integral \(I_1^M(h)\) is built as an \(L^2\) limit of compound Poisson finite sums, with centering and isometry derived by passage to the limit.
\(I_1^M(h) = \lim _{\varepsilon \to 0} I_1^{\varepsilon ,M}(h)\) in \(L^2\), constructed from compound Poisson finite sums. Centering and isometry are derived via tendsto_nhds_unique, transferring properties from the approximations to the limit.
\(\mathbb {E}[I_1^M(h)] = 0\): each approximation is centered, and centering passes to the \(L^2\) limit. This zero-mean property is inherited by all divergence operators built from \(I_1\).
Each \(\mathbb {E}[I_1^{\varepsilon ,M}(h)] = 0\) (compensated sum). The constant sequence \(0\) tends to \(\mathbb {E}[I_1^M(h)]\) by approx_converges. By tendsto_nhds_unique, \(\mathbb {E}[I_1^M(h)] = 0\).
\(\mathbb {E}[|I_1^M(h)|^2] = \| h\| _{L^2(\nu _\gamma ^M \otimes dt)}^2\): the isometry of finite-sum approximations is preserved in the limit. This identity is the analogue of the classical Itô isometry [4, Theorem 4.2.3] for the compensated Poisson setting.
Same pattern as centering: each approximation satisfies the isometry exactly, so the constant sequence converges to the limit isometry by tendsto_nhds_unique.
Packages the constructed integral into the axiomatic interface for LevyItoDecomposition, converting proved properties into structure fields. This bridge allows downstream results to use the abstract API without depending on construction details.
2.5 Lévy–Itô Decomposition
The Lévy–Itô decomposition splits the process into compensated small jumps and compound Poisson large jumps, with truncation convergence enabling passage to the full divergence.
\(L_T = X_T^{\mathrm{small}} + Y_T^{\mathrm{large}}\): compensated small jumps plus compound Poisson large jumps. This canonical splitting separates the martingale and finite-variation components of the process.
\(\mathbb {E}[F \cdot \delta ^M(u)] \to \mathbb {E}[F \cdot \delta (u)]\) as \(M \to \infty \). Essential for the chaos characterization (Theorem C), since it allows passage from truncated to full divergence.
Rewrite \(\delta ^M(u) = I_1^M(u \cdot z)\) via \(\delta \)_trunc_eq, then apply I\(_1\)_convergence.
\(\mathbb {E}[\delta ^M(u)] = 0\), inherited from centering of \(I_1^M\). Centering of the truncated divergence is a prerequisite for the even-annihilation argument at each truncation level.
Rewrite via \(\delta \)_trunc_eq and apply I\(_1\)_centered.
If \(\Phi _M \equiv 0\), then \(\mathbb {E}[F \cdot \delta ^M(u)] = 0\) for all deterministic \(u\). Even kernels have \(\Phi _M = 0\) since \(f_1 \cdot z\) is odd and \(\nu _\gamma \) is symmetric, connecting measure symmetry to chaos structure.
Apply chaos_identity with pointwise vanishing \(\Phi _M(s) \cdot u(s) = 0\) from zero_mul.
\(\mathbb {E}[(cF) \cdot \delta ^M(u)] = c \cdot \mathbb {E}[F \cdot \delta ^M(u)]\). This scalar-linearity property ensures that \(D_L\) extends to a module map over predictable scalars.
Rewrite \(cF \cdot \delta = c(F \cdot \delta ) + 0\) pointwise (by ext; simp; ring), then apply \(\mathbb {E}\)_linear and \(\mathbb {E}\)_const.
2.6 Poisson Chaos
The Poisson inner product and centered variable \(\chi \) are defined on the canonical space, and chaos orthogonality is established for the witness argument.
\(\langle f, g \rangle _r = \sum _n f(n)\, g(n)\, P_r(n)\): the \(L^2\) inner product on \((\mathbb {N}, \mathrm{poissonMeasure}(r))\). This discrete inner product underlies the Fourier–Charlier expansion that decomposes \(L^2\) into chaos subspaces.
\(\chi (n) = n - r\): the first-chaos generator. The centered witness \(G_0 - \mathbb {E}[G_0]\) is \(\chi \) with \(r = \lambda T\). As the simplest nonconstant orthogonal polynomial, \(\chi \) spans the first Poisson chaos.
\(\mathbb {E}[\chi ] = r - r = 0\), from the Poisson mean identity. Centering is one of the three hypotheses required by the representability obstruction.
Write \((n-r)P_r(n) = nP_r(n) - rP_r(n)\). Apply HasSum.sub to the mean \(\texttt{HasSum}(\ldots , r)\) and \(r \cdot \texttt{HasSum}(\ldots , 1)\). The difference is \(r - r \cdot 1 = 0\) by ring.
\(\| \chi \| ^2 = \mathrm{Var}(N) = r {\gt} 0\). Implies \(\chi \neq 0\) in \(L^2\), which is the nonzero condition needed for the obstruction.
Unfold definitions; the sum \(\sum (n-r)^2 P_r(n)\) is exactly poisson_variance_identity.
\(\langle f, \chi \rangle = 0 \Rightarrow \langle f, a\chi \rangle = 0\) for all \(a\). Orthogonality to \(\chi \) implies orthogonality to \(\mathrm{span}(\chi )\), reflecting the module structure of the first chaos.
Rewrite \(f(n)(a\chi (n))P_r(n) = a \cdot f(n)\chi (n)P_r(n)\) by ring, factor out \(a\) via tsum_mul_left, and multiply by zero.
Zero first-chaos coefficient implies orthogonality to \(\mathrm{span}(\chi )\). This is the concrete form of chaos orthogonality used in the Poisson witness argument.
Extract \(\langle f, \chi \rangle = 0\) from the coefficient via div_eq_zero_iff, then apply first_chaos_orthogonal.
2.7 \(L^2\) Estimates
The \(L^2\) Cauchy estimate \(\varepsilon ^{2-\gamma } \to 0\) and related bounds ensure convergence of the compound Poisson approximations to the compensated integral.
\(2c_\gamma \varepsilon ^{2-\gamma }/(2-\gamma ) \to 0\) as \(\varepsilon \to 0^+\). Proves the approximations are \(L^2\)-Cauchy, guaranteeing convergence to \(I_1^M(h)\). This estimate is the analytic core of the compensated integral construction.
Factor as \((2c_\gamma /(2-\gamma )) \cdot \varepsilon ^{2-\gamma }\). Apply tendsto_const_nhds.mul with rpow_tendsto_zero_at_zero_right.
\(\varepsilon ^\alpha \to 0\) as \(\varepsilon \to 0^+\) for \(\alpha {\gt} 0\). This basic real-analysis fact is used in the \(L^2\) Cauchy estimate to control the truncation error.
Use continuous_rpow_const for continuity of \(x \mapsto x^\alpha \) at \(0\), and zero_rpow for \(0^\alpha = 0\).
\(2c_\gamma M^{2-\gamma }/(2-\gamma ) {\gt} 0\) for fixed \(M\): the isometry constant for \(\delta _L^M\) is positive and finite. Finiteness at each truncation level is what makes the compound Poisson approximation scheme viable.
Same as second_moment_pos: div_pos applied to positive numerator and positive denominator.
2.8 Compound Poisson Path
The compound Poisson path \(Y_t = \sum _{s_i \leq t} z_i\) is constructed as an explicit càdlàg function, providing the large-jump component of the Lévy–Itô decomposition.
\(Y_t = \sum _{s_i \leq t} z_i\): a pure-jump càdlàg path defined as a deterministic function of the jump data. This explicit path representation avoids the need for general semimartingale theory.
Compound Poisson path restricted to \(|z_i| {\gt} 1\). The jump count \(N_T \sim \mathrm{Poisson}(\lambda T)\), and the centered version \(N_T - \mathbb {E}[N_T]\) is the obstruction witness that drives the main theorem.
\(Y_0 = 0\): the sum over an empty index set vanishes. This initial condition anchors the telescoping identity used in the Itô formula.
All jump times satisfy \(s_i {\gt} 0\), so each indicator \(\mathbf{1}_{s_i \leq 0} = 0\). Apply Finset.sum_eq_zero.
\(Y_T = \sum _i z_i\), since all jump times satisfy \(s_i \leq T\). At the terminal time, the path accumulates all jumps, recovering the full random sum.
All indicators are \(1\) by if_pos. Apply Finset.sum_congr.