Technical Summary: Signature Kernel and Schwinger-Dyson Kernel Equations as Two-Parameter Rough Differential Equations

1. Problem Statement

The paper addresses the mathematical formulation and numerical approximation of two specific kernel equations—the Signature Kernel and the Schwinger–Dyson Kernel—when driven by highly irregular signals that fall outside the classical bounded-variation regime.

While these kernels were previously well-understood for paths of bounded variation (where classical integration applies), many signals of interest in applications (e.g., Brownian motion, solutions to stochastic differential equations) are too irregular for standard two-variable integration. The authors aim to:

1. Develop a rigorous rough-path framework for two-parameter differential equations on rectangular and simplicial domains.
2. Interpret the Signature and Schwinger–Dyson kernel equations as instances of Two-Parameter Rough Differential Equations (2P-RDEs).
3. Establish well-posedness (existence, uniqueness, and stability) for these equations in spaces of jointly controlled rough paths.
4. Derive numerical schemes for the rough Schwinger–Dyson equation and analyze their computational complexity.

2. Methodology

2.1 Theoretical Framework: Jointly Controlled Rough Paths

The theory is built upon the space of jointly controlled rough paths, extending the work of [GH19] and [CP23].

• Definitions: A two-parameter path $Z$ is $(X, Y)$-controlled if its increments in the first parameter are controlled by the rough path $X$ and in the second by $Y$, with specific remainder terms satisfying joint regularity conditions.
• Stability under Composition: A critical technical step is proving that the class of two-parameter controlled paths is stable under composition with smooth functions. This allows nonlinear terms in the kernel equations (e.g., products of the kernel with itself) to remain within the controlled space, enabling the definition of nonlinear 2P-RDEs.
• Rough Integration: The authors utilize a robust framework for rough integration over rectangles $[a, b] \times [c, d]$, defined as the limit of discrete sums over rectangular partitions.

2.2 Extension to Simplicial Domains

A key methodological innovation is the extension of rough integration to two-dimensional simplices (triangular domains), which is required for the Schwinger–Dyson equation.

• Symmetrization: The authors define a "symmetrized" path $\hat{Z}$ across the diagonal of a square domain.
• Fubini-type Theorem: They prove that the integral over a simplex can be recovered from the integral over a rectangle via this symmetrization. Specifically, the simplex integral is defined as half the integral of the symmetrized path over the square. This bridges the gap between the rectangular integration theory of [CP23] and the simplicial requirements of the Schwinger–Dyson equation.

2.3 Formulation of the Kernel Equations

The paper formulates the kernels as fixed points of 2P-RDEs:

• Signature Kernel: Defined on a rectangular domain $D_{t,v} = [a, t] \times [c, v]$. The equation is linear in the kernel $K$:
  $$K(t, v) = 1 + \int_{[a,t]\times[c,v]} K(s, u) \langle dX_s, dY_u \rangle$$
  This is shown to be a 2P-RDE where the integrand is a smooth function of the controlled path.
• Schwinger–Dyson Kernel: Defined on a simplicial domain $D_{t,v} = \Delta_{[t,v]}$. The equation is quadratic:
  $$K(s, t) = 1 - \int_{\Delta_{[s,t]}} K(s, u) K(u, r) \langle dX_u, dX_r \rangle$$
  The authors extend this from bounded variation paths to geometric rough paths of arbitrary low regularity ($p \ge 1$).

2.4 Smooth Regime and Integro-Differential Reformulation

When the driving signals are smooth rough paths (admitting a time derivative along the diagonal), the authors connect the 2P-RDEs to classical PDEs and integro-differential equations:

• For the Signature Kernel, the 2P-RDE recovers the linear PDE derived in [LLS24].
• For the Schwinger–Dyson Kernel, the authors derive a system of integro-differential equations. This is achieved by exploiting combinatorial identities involving non-crossing partial matchings (related to free probability) to relate the trace component of the kernel to its higher-order tensor components.

2.5 Numerical Scheme

A numerical scheme is proposed for the rough Schwinger–Dyson equation based on the derived integro-differential system and combinatorial identities.

• Algorithm: The method uses dynamic programming on a grid, solving a linear system at each step. It leverages the structure of non-crossing partitions to compute higher-order terms efficiently.
• Complexity Analysis: The authors provide rigorous estimates for runtime and memory complexity, showing exponential dependence on the tensor algebra truncation level and polynomial dependence on the grid dimension.

3. Key Contributions

1. Unified Rough Path Framework: The paper establishes a general theory for 2P-RDEs on both rectangular and simplicial domains, unifying the treatment of the Signature and Schwinger–Dyson kernels under the umbrella of jointly controlled rough paths.
2. Simplicial Rough Integration: It introduces a rigorous definition of rough integration over two-dimensional simplices at low regularity, extending previous results that were limited to smooth approximations or specific variation regimes.
3. Well-Posedness and Stability: The authors prove existence, uniqueness, and continuous dependence on the driving rough paths for both kernel equations. This provides a solid theoretical foundation for approximating these kernels with irregular data.
4. Combinatorial Reformulation: For the Schwinger–Dyson kernel, the paper derives a novel integro-differential reformulation in the smooth regime using combinatorial identities, offering an alternative analytic description.
5. Numerical Implementation and Analysis: A concrete numerical algorithm is presented with detailed complexity analysis. The paper validates the theoretical complexity estimates and demonstrates the method's performance across different regularity regimes.

4. Results

• Theoretical: The Signature Kernel is proven to be the unique solution to a linear 2P-RDE. The Schwinger–Dyson Kernel is proven to be the unique solution to a quadratic 2P-RDE on a simplex. Both solutions are shown to be stable with respect to perturbations in the driving rough paths.
• Smooth Regime: In the case of smooth rough paths, the 2P-RDE for the Signature Kernel reduces to the known PDE. The Schwinger–Dyson kernel is shown to satisfy a specific system of integro-differential equations derived from combinatorial properties of the kernel's tensor components.
• Numerical Experiments:
  • The authors tested the proposed scheme (denoted $SDK_\kappa$) against a randomized CDE method ($SDK_r$) from [CT26] using fractional Brownian motion with varying Hurst indices ($H$).
  • Findings: As the regularity of the driving signal decreases (lower $H$), the discrepancy (Mean Absolute Error and Standard Deviation) between different numerical schemes increases.
  • The results confirm that schemes with higher approximation orders (higher $\kappa$) maintain better agreement with each other as regularity decreases, consistent with the theoretical expectation that lower regularity requires higher-order approximations.
  • Runtime experiments confirmed the exponential scaling with grid size and block size, and polynomial scaling with dimension, matching the theoretical complexity estimates.

5. Significance and Claims

The paper claims to provide a robust rough-path framework that allows for the interpretation, analysis, and numerical approximation of kernel identities (Signature and Schwinger–Dyson) driven by signals that are not of bounded variation.

• Theoretical Significance: By framing these kernels as 2P-RDEs, the authors leverage the powerful stability and well-posedness results of rough path theory. This is particularly relevant for applications involving stochastic processes (like Brownian motion) where classical integration fails.
• Practical Significance: The development of a numerical scheme for the rough Schwinger–Dyson equation, complete with complexity analysis, enables the computation of these kernels in settings previously inaccessible to standard methods. The experiments suggest that the proposed methods are consistent and that the theoretical regularity requirements are necessary for accurate approximation.
• Limitations: The authors note that some numerical comparisons involve driving signals outside the theoretical justification of certain schemes (e.g., using bounded-variation methods on rough paths), treating these comparisons as heuristic benchmarks rather than rigorous convergence proofs.

In summary, the work bridges the gap between the algebraic structure of rough paths and the analytic properties of kernel equations, providing both a theoretical foundation for irregular data and a practical tool for their computation.