Causal Resolvent Rough Lifts under Nondominated Martingale Laws: Gaussian-Tail Cores and Conditional Wong--Zakai Transfer

Bridging the Gap Between Smooth Approximations and Rough Reality

Modeling complex systems often involves dealing with extreme uncertainty. In these scenarios, the underlying rules of randomness are not fully known. Researchers use mathematical tools to approximate these unpredictable paths. However, a fundamental tension exists. If you smooth out the noise too much to simplify calculations, you lose the very characteristics that define the system's behavior.

In stochastic analysis, engineers typically rely on Wong–Zakai approximations. These replace jagged, random noise with smooth, differentiable functions to make differential equations solvable. While this works for a single known model, it breaks down under "nondominated" uncertainty. This occurs when you do not know which specific probability law governs the system. Current approaches struggle to maintain stability when you stop, shift, or concatenate these paths. This makes them unreliable for real-time control or decision-making under model ambiguity.

Zhao’s paper addresses this by constructing a "causal resolvent." This is a mathematical bridge. It allows us to lift smooth approximations into the "rough" world of true stochastic processes without losing track of the underlying uncertainty.

The breakdown of smooth approximations

The status quo relies on the idea of convergence. If you approximate a rough path with a smooth one, the resulting solutions should eventually match the true stochastic solution. This is the essence of the Wong–Zakai theorem. However, this works best with a single, fixed probability measure.

When moving into nondominated martingale laws, the smoothness of the approximation becomes a liability. The paper considers a class where volatility is constrained by a Loewner interval (a mathematical bound on the allowed volatility matrix). This interval acts like a speed limit for how much the system can fluctuate. Standard methods lack a "common" representative that works across all possible models in the class.

Existing tools fail to provide a unified, "raw-causal" framework. Such a framework must remain valid under temporal operations. These include stopping a process or shifting a time window. Without this, errors in the approximation can accumulate or vanish inconsistently during restarts or path stitching.

The causal resolvent mechanism

The author's core architectural choice is the causal exponential resolvent. Instead of forcing a smooth path to look like a rough one, the method uses a convolution-based regularization.

1. The Resolvent Construction: For a raw path $x$ and a scale parameter $\varepsilon$, the resolvent $Y^\varepsilon$ is an exponential convolution. It is defined as $Y^\varepsilon_t = \varepsilon^{-1} \int_0^t e^{-(t-r)/\varepsilon} x_r \, dr$. This creates an absolutely continuous path that acts as a smoothed version of the original.
2. Defect and Dissipation: The paper introduces a "defect" $D^\varepsilon_t = x_t - Y^\varepsilon_t$. The authors derive a pathwise identity linking this defect to a "bracket" $Q^\varepsilon_t$ (a selector for quadratic variation, which measures the cumulative variance of the path). This identity is expressed as: $Q^\varepsilon_t = (D^\varepsilon_t)^{\otimes 2} + \frac{2}{\varepsilon} \int_0^t (D^\varepsilon_r)^{\otimes 2} \, dr$. This identity tracks the "energy" of the approximation directly from the smoothing error.
3. Aggregation into Tensors: By aggregating these resolvents, the authors construct a single Borel causal Itô tensor (a consistent mathematical rule for performing stochastic integration) and a quadratic-variation selector. These objects serve as a universal upgrade kit for any path in the specified martingale class.
4. The Master Core: To handle the lack of uniform tightness in the raw path space, the authors build "Gaussian-tail capacity cores" ($C_R$). These are increasing families of compact sets that act as safe operating zones. Within these cores, the approximated tensors and lifts are guaranteed to be continuous and stable.

Converging at scale

The paper reports several key convergence metrics. The most critical result for an engineer is the uniform convergence of the tensor and bracket approximations. The authors demonstrate that the error in the finite-scale bracket and the Itô tensor converges in $L^q$ at a rate of $O(q\sqrt{\varepsilon})$. This tells us that as our smoothing scale $\varepsilon$ decreases, the approximation error shrinks predictably across all models in the class.

Regarding the "roughness" of the path, the paper finds that for $1/3 < \alpha < 1/2$, the canonical signatures (a way to represent the path's geometry) of the resolvent $Y^\varepsilon$ converge in the $\alpha$-Hölder rough-path topology. This convergence is toward a common Stratonovich lift. The rate is $O(q\varepsilon^\theta)$ for any $\theta < 1/2 - \alpha$.

Furthermore, the authors establish that the capacity of the complement of their compact cores decays exponentially. Specifically, $c_{\mathfrak{M}_\Lambda}(C_R^c) \leq Ce^{-cR^2}$. This "Gaussian tail" is significant. It quantifies the vanishingly small probability of encountering a path that is too wild for the core's guarantees.

Limits of the framework

There are clear boundaries to where this utility ends.

First, the construction requires the "local domination" of the martingale class. The volatility must satisfy $d[X]^P_t \preceq \Lambda I_d \, dt$. If your system exhibits unbounded volatility that violates this Loewner interval constraint, the Gaussian-tail guarantees collapse.

Second, the convergence rates are tied to the parameter $\varepsilon$. Achieving high precision requires a very small $\varepsilon$. However, the resolvent equation (Equation 8) includes terms scaled by $1/\varepsilon$. This creates a "stiffness" issue in ordinary differential equation (ODE) solvers. This presents a tradeoff between theoretical precision and computational cost. As $\varepsilon$ gets smaller, the equations become harder to solve numerically. This may require much smaller time steps to maintain stability.

Finally, the paper focuses on "maximal history-independent correspondences." While it touches on path-dependent cases, the strongest guarantees regarding Bellman continuity and optimal stopping stability are reserved for the simpler class of controllers.

The verdict

This is a foundational piece of theory for anyone building robust stochastic control systems. The ability to provide a single, consistent "lift" that is stable under stopping and concatenation is a massive leap.

If you are working on purely deterministic problems, this is overkill. However, if you are designing systems meant to operate under model uncertainty, this framework provides necessary mathematical safety rails. It moves the conversation from how we approximate a path to how we ensure our approximations are stable under the operations we actually perform in production.