Generative Path-Law Jump-Diffusion: Sequential MMD-Gradient Flows and Generalisation Bounds in Marcus-Signature RKHS

Imagine you are trying to predict the future path of a chaotic storm, a stock market crash, or a wandering animal. You don't just want a straight line; you need a realistic, messy, "jagged" path that accounts for sudden gusts of wind, unexpected jumps, and changing weather patterns.

This paper, written by Daniel Bloch, introduces a new, super-smart computer program called ANJD (Anticipatory Neural Jump-Diffusion) that can generate these realistic, chaotic future paths.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: Predicting the "Jagged" Future

Most standard prediction models are like smooth, straight roads. They are great for driving on a highway, but they fail miserably when the road turns into a bumpy dirt track with sudden potholes (jumps) and sharp turns (regime shifts).

The Challenge: Financial markets and physical systems often have "black swan" events—sudden, massive jumps that break the smooth flow. Old models get confused by these jumps or try to smooth them over, losing the true nature of the chaos.
The Goal: Create a model that doesn't just guess the average future, but generates thousands of realistic possible futures, complete with the bumps, jumps, and sudden changes, while staying mathematically consistent with what we know so far.

2. The Core Idea: The "Signature" Map

To understand a path, the authors use something called a Signature.

The Analogy: Imagine you are trying to describe a dance to someone who can't see it. You could say "he moved left, then right." But that misses the flair, the speed, and the order of the moves.
The Signature: Think of the Signature as a "mathematical fingerprint" of the dance. It captures not just where the dancer went, but the order they went there (did they spin before jumping?), how fast they moved, and the complex interactions between moves. It turns a messy, jagged line into a structured, high-dimensional code that computers can understand perfectly.

3. The Secret Sauce: "Spectral Whitening" (AVNSG)

The world is noisy. Sometimes the market is calm; sometimes it's screaming. If you try to predict a storm while the wind is howling, you need to adjust your sensors.

The Analogy: Imagine trying to hear a whisper in a quiet room versus a whisper in a rock concert. In the rock concert, you need a special filter to turn down the loud drums so you can hear the whisper.
The Solution (AVNSG): The paper introduces a "dynamic filter" called AVNSG. It constantly adjusts the "volume" of the data. If the market is volatile (loud drums), it turns down the noise so the model doesn't panic. If the market is quiet, it turns up the sensitivity. This ensures the model stays stable even when the future looks terrifyingly unpredictable.

4. How It Generates the Future: The "Steepest Descent"

How does the computer actually draw the new path?

The Analogy: Imagine you are blindfolded on a hill, and you want to get to a specific valley (the target future). You feel the ground with your feet. If the ground slopes down toward the valley, you take a step that way.
The Process: The ANJD model treats the "difference" between its current guess and the desired future as a hill. It calculates the "steepest descent"—the fastest way to slide down that hill to match the target.
- Continuous Flow: It walks smoothly when things are calm.
- Jump Logic: If the target requires a sudden jump (like a stock crashing), the model doesn't walk; it teleports (jumps) to the right spot instantly, just like the real world does.

5. The "Time-Travel" Aspect

The model is "Anticipatory."

The Analogy: Most drivers look at the road right in front of their bumper. This model is like a driver who can see 100 miles down the road and knows a bridge is out, so they start turning the wheel now to prepare for the turn 100 miles later.
The Mechanism: It uses a "moving target." It doesn't just aim for a fixed point in the future; it aims for a target that moves and changes as time passes, ensuring the path stays consistent with the latest information.

6. Why This Matters (The "So What?")

For Banks: It can simulate "worst-case scenarios" (black swans) much better than before, helping them prepare for market crashes without being blindsided.
For Physics: It can model particles that bounce around randomly and suddenly get hit by other particles, capturing the true "texture" of the chaos.
Efficiency: The authors figured out a way to do this incredibly fast (using a trick called "Nyström approximation"), so it doesn't take a supercomputer a week to run. It can run in real-time.

Summary in One Sentence

This paper teaches a computer how to paint realistic, jagged, and chaotic future paths by using a special "fingerprint" of the past, a dynamic noise-canceling filter to stay calm during storms, and a smart "gravity" system that pulls the future into shape, ensuring it looks exactly like the messy, unpredictable real world.

1. Problem Statement

The paper addresses the challenge of synthesizing forward-looking, càdlàg (right-continuous with left limits) stochastic trajectories that are consistent with time-evolving path-law proxies. Existing generative models (e.g., GANs, VAEs, standard Neural SDEs) struggle with:

Discontinuous Dynamics: They often fail to capture discrete structural breaks, regime shifts, and jump-discontinuities inherent in financial and physical systems.
Path-Geometric Integrity: They lack mechanisms to preserve higher-order dependencies (e.g., leverage effects, volatility clusters) and non-commutative moments when the underlying law undergoes abrupt changes.
Inversion of Path-Laws: While signature-based filtering can estimate expected path dynamics, inverting these infinite-dimensional moments back into concrete synthetic realizations on restricted Skorokhod manifolds remains a formidable challenge.

The goal is to create a framework that generates an ensemble of paths whose collective law matches a moving target proxy (representing anticipated future dynamics) while rigorously handling jumps and non-stationarity.

2. Methodology

The proposed framework, Anticipatory Neural Jump-Diffusion (ANJD), treats path synthesis as a sequential matching problem on restricted Skorokhod manifolds ( $D_s$ ). The methodology integrates three core components:

A. Theoretical Foundation: Marcus-Signature RKHS

Time-Extended Signature: The model uses the Marcus-sense signature of time-extended paths ( $\tilde{\gamma}_s = (s, \gamma_s)$ ). This embedding ensures injectivity and universality for càdlàg paths, uniquely characterizing the path-law via non-commutative moments.
Signature Bochner Integral: The target is defined as the expected signature (path-law proxy) $\hat{\Phi}_{s|t}$ , which serves as a sufficient statistic for the conditional path-measure.
Adaptive Variance-Normalised Signature Geometry (AVNSG): A time-evolving precision operator $Q_s = (\Omega_s + \lambda I)^{-1}$ is introduced. It performs dynamic spectral whitening on the signature manifold, decorrelating features and normalizing variance to ensure stability during volatile regime shifts and heavy-tailed shocks.

B. The Generative Mechanism: ANJD Flow

The synthesis is governed by a Jump-SDE driven by a Neural CDE:
$dX_s = f_\theta ds + g_\theta dW_s + h_\theta dN_s$

Drift ( $f_\theta$ ) and Jump Intensity ( $\lambda_\theta$ ): These are functionally coupled to the moving target proxy $\hat{\Phi}_{s|t}$ and the AVNSG geometry.
Marcus Integration: The continuous part is interpreted in the Marcus sense to ensure the solution remains on the appropriate manifold and the signature remains group-valued across discontinuities.
Sequential Matching: The flow is designed to be the infinitesimal steepest descent direction for the Maximum Mean Discrepancy (MMD) functional between the synthetic path-measure and the moving target proxy.

C. Numerical Implementation

Nyström-Compressed Score-Matching: To handle the infinite-dimensional signature space, the model uses a Nyström approximation to compress the score function.
Hybrid Euler-Maruyama-Marcus (EMM) Integration: A sampling scheme that handles both continuous diffusion and discrete jumps.
Dynamic Precision Updates: The precision matrix is updated via Sherman-Morrison rank-1 updates ( $O(m^2)$ complexity) rather than full re-inversion, allowing real-time adaptation to structural breaks and clock evolution.

3. Key Contributions

Sequential Anticipatory Flow Framework: Introduction of the ANJD architecture, which bridges recursive filtering and path synthesis, enabling the sequential matching of càdlàg trajectories against forecasted structural breaks.
Theoretical MMD-Gradient Flow: Rigorous proof (Theorem 4.1) that the generative drift and jump intensity constitute the infinitesimal steepest descent for the MMD functional relative to a moving target, linking the process generator to continuous path-law minimization.
AVNSG Geometry: Definition of a time-evolving precision operator that ensures contractivity and stability under forecasted volatility explosions and heteroskedastic shocks via spectral whitening.
Statistical Generalisation Bounds: Derivation of high-probability bounds for generalization error (Theorem 5.1) and analysis of Rademacher complexity (Proposition 5), showing that the model's expressive power is regularized by the spectral radius of the AVNSG operator, effectively capping the influence of "black swan" terms.
Scalable Implementation: A numerical scheme achieving $O(m^2)$ complexity via dynamic Nyström updates, enabling the propagation of infinite-dimensional geometry through both diffusion and jump-discontinuities.

4. Key Results

Convergence: The empirical average of time-extended signatures converges to the target proxy in the $Q_s$ -norm at a rate of $O(1/k)$ using a greedy herding rule.
Stability: The flow remains contractive in the Skorokhod topology even under spectral expansion (increasing uncertainty), provided the jump-induced energy does not exceed the infinitesimal dissipation rate of the MMD-gradient.
Expressive Power: The model successfully captures non-commutative moments and high-order stochastic texture, including heavy-tailed innovations and leverage effects, which standard pure-diffusion Neural SDEs miss.
Efficiency: The use of rank-1 updates allows the system to react to high-frequency structural breaks in real-time without the computational cost of $O(m^3)$ matrix inversions.

5. Significance

This work represents a significant advancement in quantitative finance and stochastic modeling by:

Bridging Theory and Practice: It connects abstract path-signature theory (Lyons et al.) with practical generative modeling (Schrödinger bridges, Neural SDEs) specifically for discontinuous processes.
Handling Non-Stationarity: Unlike static generative models, ANJD is "anticipatory," explicitly incorporating forecasted regime shifts and structural breaks into the generation process.
Robustness to Tail Risk: By utilizing the Marcus signature and AVNSG, the model provides a mathematically rigorous way to generate paths that respect the algebraic constraints of discontinuous dynamics, offering a superior tool for risk management, stress testing, and scenario generation in environments prone to "black swan" events.
Computational Feasibility: It demonstrates that infinite-dimensional path-law synthesis can be implemented efficiently, making it viable for real-time applications in high-frequency trading and dynamic risk systems.

In summary, the paper proposes a novel, theoretically grounded, and computationally efficient framework for generating realistic, discontinuous stochastic trajectories that adapt to evolving market regimes and structural breaks.