Neural Operators for Jump-Diffusion Control
This note checks, on mathematical grounds, whether the neural-operator framework is a valid approach to the jump-diffusion stochastic control problems in this repository — the jump-kernel market-making model, the Hawkes family, and the jump-to-default portfolio. It is the companion to the neural-architectures plan, which pivots onto this question. The notation for the jump-to-default problem follows the correlated price-default note.
The verdict, stated up front, is: the approach is valid, and in fact better motivated than the diffusion case, because the jump generator is a Fourier multiplier — the exact class of operator a spectral neural operator represents natively. The validity carries four specific caveats, named in the last section.
The control problem
Let the state be \(x \in \mathcal{X}\), with \(\mathcal{X} = \mathbb{R}^d\) for a diffusion, a lattice \(\mathbb{Z}^m\) for an inventory, or the group \(\mathbb{Z}_2^m\) for default indicators. A controlled jump-diffusion evolves as
\[ dx_t = b(x_t, u_t)\,dt + \sigma(x_t, u_t)\,dW_t + \int_{\mathcal{Z}} j(x_{t-}, u_t, z)\,N(dt, dz), \]
where \(N\) is a Poisson random measure with intensity \(\nu(dz)\,dt\), so that the jump size \(z\) is drawn from \(\nu\). The value function \(V(t,x) = \sup_u \mathbb{E}\!\left[g(x_T) + \int_t^T f(x_s,u_s)\,ds\right]\) satisfies the HJB PIDE (partial integro-differential equation)
\[ 0 = \partial_t V + \sup_{u\in U}\Bigg[ b(x,u)\cdot\nabla V + \tfrac12\,\mathrm{tr}\!\big(\sigma\sigma^\top D^2 V\big)
- f(x,u) + \int_{\mathcal{Z}}\!\big(V(x + j(x,u,z)) - V(x)\big)\,\nu(dz) \Bigg], \]
with terminal condition \(V(T,x) = g(x)\). The distinguishing term is the last one, the nonlocal jump generator
\[ \mathcal{J} V(x) = \int_{\mathcal{Z}}\big(V(x + j(x,u,z)) - V(x)\big)\,\nu(dz). \]
For the repository's problems this term is concrete. In the market-making jump kernel, \(x = q \in \mathbb{Z}^m\) is inventory, \(z\) is the fill amplitude drawn from the kernel \(\mu_i^{b/a}\), and a bid fill shifts \(q \mapsto q + z e_i\). In the jump-to-default problem, \(x = (S, D)\) with \(D \in \mathbb{Z}_2^m\), and the generator is \(\lambda_i(S,D)\,\big(u(t,S,D+e_i) - u(t,S,D)\big)\).
What a neural operator learns
A neural operator approximates a map \(G : \mathcal{A} \to \mathcal{U}\) between function spaces, rather than a single function. Here the natural input is the data that varies across instances — the terminal condition \(g\), or the coefficients and jump measure — and the output is the value function \(V\). The two reference architectures are the DeepONet branch/trunk factorization [@lu2021deeponet] and the Fourier neural operator (FNO), whose layer is a spectral multiplier
\[ v \mapsto \mathcal{F}^{-1}\!\big(R_\theta \cdot \mathcal{F}(v)\big), \]
with \(R_\theta\) a learned Fourier symbol and \(\mathcal{F}\) the transform on \(\mathcal{X}\) [@li2021fourier, @kovachki2023neural]. Universal-approximation theorems guarantee that such networks approximate a broad class of nonlinear operators [@lanthaler2023parametric, @cohen2026hilbert]; the question here is whether the specific operator defined by a jump-diffusion HJB sits inside a class that these theorems — and, more importantly, the practical architecture — can reach efficiently.
The load-bearing result: the jump generator is a Fourier multiplier
Take additive, state-independent jumps, \(j(x,u,z) = z\) (the market-making kernel and Hawkes fills are exactly of this form). Then the jump generator is a convolution:
\[ \mathcal{J} V(x) = \int_{\mathcal{X}}\big(V(x+z) - V(x)\big)\,\nu(dz) = (\nu \ast V)(x) - V(x). \]
Its Fourier transform diagonalizes the operator. With \(\widehat{V}(\xi) = \int e^{-i\xi\cdot x} V(x)\,dx\) and the characteristic function \(\widehat\nu(\xi) = \int e^{i\xi\cdot z}\nu(dz)\), the shift identity \(\int e^{-i\xi\cdot x} V(x+z)\,dx = e^{i\xi\cdot z}\,\widehat{V}(\xi)\) gives
\[ \widehat{\mathcal{J} V}(\xi) = \big(\widehat\nu(\xi) - 1\big)\,\widehat{V}(\xi). \]
That is, \(\mathcal{J}\) is a Fourier multiplier with symbol \(m(\xi) = \widehat\nu(\xi) - 1\). This is exactly one FNO layer with the symbol set to \(R_\theta = \widehat\nu - 1\); the jump generator lies in the native representational class of the spectral operator, and is recovered exactly by a single layer, not merely approximated.
The same statement holds on each of the repository's state spaces, with the appropriate Fourier transform:
- Inventory \(\mathbb{Z}^m\) (market-making kernel, Hawkes): the fill generator is a convolution on the lattice, diagonalized by the discrete Fourier transform on the torus, with symbol \(\widehat\mu(\theta) - 1\) where \(\widehat\mu\) is the characteristic function of the amplitude kernel.
- Default state \(\mathbb{Z}_2^m\) (jump-to-default): the generator \(\mathcal{J}_i u(D) = u(D+e_i) - u(D)\) is translation-by-\(e_i\) minus the identity on the group \(\mathbb{Z}_2^m\). The Walsh–Hadamard transform \(\widehat u(\omega) = \sum_D u(D)(-1)^{\omega\cdot D}\) diagonalizes it, since translation by \(e_i\) multiplies by \((-1)^{\omega_i}\), giving the symbol \((-1)^{\omega_i} - 1\). The jump operator annihilates modes even in bit \(i\) and rescales odd modes by \(-2\).
So in every relevant case the nonlocal term — the part that makes classical grid methods expensive — is the part a spectral operator represents natively. The diffusion term \(\tfrac12\mathrm{tr}(\sigma\sigma^\top D^2 V)\) is, by contrast, a local differential operator, which the FNO handles less naturally. The jump-diffusion problem is therefore structurally better suited to a spectral operator than a pure diffusion is, not worse.
Where the operator framework is exact
Separate the HJB into three pieces, each with a clean status.
Linear PIDE (fixed policy). At a fixed feedback policy \(u = u(t,x)\) the \(\sup\) disappears and the equation is a linear PIDE whose solution operator is the linear semigroup \(e^{\tau(\mathcal{L}^u + \mathcal{J})}\), with \(\mathcal{L}^u\) the diffusion generator. The jump part is a Fourier multiplier (above), and the diffusion part is a second-order operator; both are within the approximation class of a neural operator [@kovachki2023neural]. For a state-independent jump kernel this semigroup is, at the level of the jump generator, exactly representable.
The control supremum. The \(\sup_u\) is a pointwise nonlinearity, not an integral operator, so the full solution operator \(g \mapsto V\) is nonlinear. Two routes handle it, both already in the literature. Policy iteration alternates a linear value solve (operator-friendly, above) with a pointwise policy update, exactly the structure of DeepONet-plus-policy-iteration [@lee2024hamilton]. Alternatively the nonlinear operator is learned directly, which the universal-approximation results for nonlinear operators cover [@lanthaler2023parametric, @cohen2026hilbert]. Either way the nonlinearity is not an obstruction to validity.
This decomposition — jump = spectral layer, diffusion = differential operator, control = pointwise max — is the rigorous content of the validity claim.
Caveats (where the framework must be extended)
Four points require explicit treatment; none invalidates the approach, but each must be named rather than silently assumed.
-
Absorbing (killing) generator. The jump-to-default generator \(\lambda_i\,(u(\cdot, D+e_i) - u(\cdot, D))\) is sub-Markovian: the \(-\lambda_i u\) part is a killing term (asset leaves the market), and the state space loses a dimension. The standard operator/universal-approximation theory is stated for conservative or parabolic semigroups; the killing extension is straightforward (it contributes only a diagonal \(-\lambda_i\) term, which the symbol above already captures via \(\widehat\nu(0) - 1 = \nu(\mathcal{X}) - 1 \neq 0\) for a sub-probability \(\nu\)), but it is a technical assumption that must be stated, not left implicit. This is a model assumption of the jump-to-default problem [@kraft2007bankruptcy, @bielecki2004credit].
-
Measure-valued input and parametric complexity. If the operator is to amortize over the jump kernel — i.e. the input is the measure \(\nu\) — then \(G\) must be continuous in \(\nu\) under some topology for the approximation rates to bite, and the dimension of the measure enters the complexity. Lanthaler & Stuart [@lanthaler2023parametric] proved that general operator learning suffers a curse of parametric complexity and that HJB structure can beat it; whether the jump-measure input reintroduces it is open and is the single most important risk to quantify. This is a technical (and partly numerical) assumption.
-
Regularity of the value. A jump value function is less smooth than a diffusion one: jump-to-default induces a kink in the derivative at the default boundary, and arbitrary jump kernels reduce Hölder regularity. This does not break validity — \(V\) remains the viscosity solution and is continuous — but it degrades approximation rates relative to the diffusion case. This is a quantitative caveat, an approximation rather than a structural failure, and it predicts that jump problems will need more operator capacity than their diffusion counterparts.
-
Monotonicity / comparison principle. Convergence to the viscosity solution is guaranteed for monotone schemes [@barles1991convergence], and a spectral operator is not automatically monotone. Enforcing monotonicity (e.g. input-convex blocks) may conflict with the spectral structure, so a rigorous convergence statement requires additional design; a practical approximation does not. This is a technical assumption for rigor, not a validity failure.
Validation targets
Each claim above is checkable on the repository's exact-reduction lattice, and the checks are ordered from strongest to weakest:
- Spectral recovery of the linear jump generator. For a fixed policy (or no control) and a state-independent kernel, the jump semigroup is a Fourier multiplier; an FNO should recover it exactly, and a learned symbol should converge to \(\widehat\nu - 1\). This is the cleanest and most direct test of the load-bearing result.
- Reduction to diffusion. As the jump intensity tends to zero the problem converges to the diffusion HJB already handled by the literature [@lee2024hamilton, @han2018solving]; the jump-aware operator must reproduce that limit. This is a limiting-behaviour check.
- Repository reductions. The jump-to-default no-loss and zero-diffusion
limits and the market-making equal-\(k\) case have closed forms (see the
correlated price-default note and the
multi-asset plan);
the Hawkes family reduces to constant intensity. Matching these is the
exact-reduction check from
solver/06-exact-reduction-consistency. - Manufactured solution. Choose an analytic \(V\), derive the jump measure and running cost that make it a solution, and verify recovery to a documented tolerance.
Conclusion
The approach is valid, with the strongest possible grounding: the jump generator — the nonlocal term that motivates jump-diffusion in the first place — is a Fourier multiplier on the state's group (\(\mathbb{R}^d\), \(\mathbb{Z}^m\), or \(\mathbb{Z}_2^m\)), and is therefore represented exactly by a spectral neural operator layer. The diffusion term and the control supremum are handled by the same mechanisms as in the diffusion literature, so nothing new is obstructing them. The residual risk is concentrated in a single open point — whether amortizing over the jump measure reintroduces the curse of parametric complexity — plus the routine extensions of killing generators, reduced value regularity, and monotone approximation. These are the precise objects to resolve in the implementation, and the exact-reduction lattice is the natural place to resolve them.