🔢 mathematics

Convergence Rates of Continuous-Time Random Walks to Time-Fractional Diffusions with Unbounded Coefficients

This paper establishes uniform weak convergence rates for a probabilistic numerical scheme combining discrete Markov chains and heavy-tailed random walks to approximate backward time-fractional diffusion equations driven by diffusions with unbounded coefficients, utilizing Feller semigroup techniques and high-order sensitivity analysis to derive bounds under specific killing conditions.

Original authors: Artur Sidorenko, Vasilii Kolokoltsov

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Artur Sidorenko, Vasilii Kolokoltsov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the Unpredictable

Imagine you are trying to predict where a drunk person (a "random walker") will end up after walking for an hour. In the real world, this person doesn't just walk in a straight line; they stumble, change direction, and sometimes stop to tie their shoe.

In mathematics, this is called a diffusion process. Usually, we have very good tools to predict where they will be. But this paper deals with a much trickier scenario: Time-Fractional Diffusion.

Think of "Time-Fractional" as a world where time itself is broken or "glitchy." Instead of time flowing smoothly like a river, it moves in bursts, pauses, and jumps. The drunk walker might stand still for a long time, then suddenly take three steps at once. This happens because their "internal clock" is being driven by a stable subordinator—a fancy way of saying their time is being controlled by a chaotic, heavy-tailed random process.

The authors want to build a computer simulation (a Continuous-Time Random Walk, or CTRW) to approximate where this walker will be. The big question is: How accurate is our simulation compared to the real, messy math?

The Problem: The "Unbounded" Wildcard

Most previous studies assumed the walker was walking in a safe, bounded neighborhood (like a city block). But in this paper, the authors tackle the unbounded case.

Imagine the walker isn't just in a city; they are in an infinite desert. The further they walk, the faster they might run, or the more wildly they might spin. Their speed and direction aren't capped; they can grow infinitely large depending on where they are. This is like Geometric Brownian Motion (used in finance to model stock prices), where a stock price can theoretically go to infinity.

Simulating these "infinite desert" walkers is hard because standard computer methods often break down when things get too big. The authors had to invent a new way to measure the error that doesn't explode when the numbers get huge.

The Solution: A Two-Part Strategy

To solve this, the authors used a clever two-part strategy, like building a bridge across a canyon.

Part 1: The "Sensitivity" Map (Kunita Stochastic Flows)

Imagine you are trying to predict the path of the walker, but you are also worried about how a tiny change in their starting point affects the outcome. If they start one inch to the left, do they end up a mile away?

The authors used a mathematical tool called Kunita Stochastic Flows. Think of this as a "sensitivity map." They didn't just track the walker; they tracked how the entire landscape of possible paths bends and stretches. They proved that even if the walker runs wild (unbounded coefficients), the "shape" of the possible paths remains smooth and predictable enough to calculate.

They treated these paths like tensor fields (which are just multi-dimensional grids of numbers). By using a special "chain rule" (a mathematical recipe for combining changes), they showed that they could control the "jaggedness" of these paths, ensuring the simulation stays stable.

Part 2: The "Clock" and the "Step"

The simulation has two moving parts:

The Step: The walker taking a step (the diffusion).
The Clock: The chaotic time mechanism that decides when the next step happens (the subordinator).

The authors approximated the chaotic clock using a heavy-tailed random walk. Imagine a clock that usually ticks once a second, but occasionally skips a whole hour, or sometimes ticks ten times in a second. They proved that if you use enough "ticks" (a fine enough grid), this fake clock gets very close to the real chaotic clock.

The Results: How Fast Does the Simulation Catch Up?

The paper calculates the convergence rate. This is simply: How much do we need to zoom in (make the steps smaller) to get a specific level of accuracy?

They found two distinct regimes, depending on a "killing" parameter (think of this as a "tax" or "discount" applied to the walker's path over time):

The "Safe" Zone (Linear Convergence):
If the "tax" is high enough to overpower the walker's tendency to run away into the infinite desert, the simulation is very accurate. The error shrinks linearly with the step size. It's like walking on a treadmill; no matter how fast you try to run, the belt keeps you in place, and your simulation is spot on.
The "Logarithmic" Zone (Slower Convergence):
If the "tax" is too weak to fully stop the walker from running wild, the simulation is still accurate, but it gets there slower. The error shrinks, but it involves a logarithmic factor.
- Analogy: Imagine trying to catch a runaway train. If you have a strong brake (high tax), you stop it quickly. If your brake is weak, you can still stop it, but you have to apply it for a much longer time, and the math gets a bit "sluggish" (logarithmic).

Why This Matters (According to the Paper)

The authors didn't just say "it works." They provided rigorous bounds. They proved that even when the coefficients (the rules of the walk) are unbounded and the time is fractional (glitchy), their specific numerical method converges to the true answer.

They specifically highlighted that their method works for Geometric Brownian Motion (the math behind stock markets). This means their "sensitivity map" and "clock" techniques can handle the wild, unbounded growth of financial models without the math breaking down.

Summary in One Sentence

The authors built a robust mathematical "safety net" that allows computers to accurately simulate chaotic, time-glitchy random walks that can grow infinitely large, proving exactly how fast these simulations converge to the truth under different conditions.

Technical Summary: Convergence Rates of Continuous-Time Random Walks to Time-Fractional Diffusions with Unbounded Coefficients

Problem Statement
The paper investigates the uniform weak convergence rates of probabilistic numerical methods applied to backward time-fractional diffusion equations. Specifically, it addresses the Cauchy problem:
$\partial^\beta_t u(t, x) = \frac{1}{\Gamma(1-\beta)} (L u(t, x) - c u(t, x)), \quad u(0, x) = f(x),$
where $L$ is the generator of a diffusion process with potentially unbounded coefficients (e.g., Geometric Brownian Motion), and $\partial^\beta_t$ is the Caputo–Dzherbashian fractional derivative. The solution admits a probabilistic representation via a time-changed diffusion: $u(t, x) = \mathbb{E}_x [e^{-c\hat{\sigma}_t} f(X_{\hat{\sigma}_t})]$ , where $\hat{\sigma}_t$ is the inverse of a $\beta$ -stable subordinator.

While weak convergence rates for Continuous-Time Random Walk (CTRW) approximations were previously established for bounded diffusion coefficients (citing [3]), this work extends the analysis to the case of unbounded coefficients, a setting that includes affine models like Black–Scholes and Langevin equations. The primary challenge lies in controlling the growth of derivatives of the diffusion semigroup when the coefficients are not globally bounded.

Methodology
The authors employ a combination of Feller semigroup techniques, Kunita's stochastic flow theory, and high-order sensitivity analysis. The methodology proceeds in three main stages:

Stochastic Flows and Tensor Fields: The paper utilizes Kunita's theory to treat the derivatives of the diffusion flow $X_t(x)$ with respect to the initial condition $x$ as random tensor fields. By employing a tensorial chain rule (based on the Faà di Bruno formula), the authors organize higher-order derivatives into tensor fields, allowing for pathwise analysis.
Sensitivity Estimates and Quasi-Contraction: The authors derive uniform bounds for all orders of sensitivities ( $D^k F_t f$ $D^{k} F_{t} f$ ) of the diffusion semigroup $F_t$ $F_{t}$ . They establish a quasi-contraction property for the semigroup on spaces of smooth functions with bounded (weighted) derivatives. This involves:
- Defining weighted function spaces ( $C_{W, \infty}$ and $\mathcal{C}^r_{\alpha, \infty}$ ) to handle unbounded observables.
- Proving that the semigroup is strongly continuous on these spaces.
- Using renorming techniques (Proposition 22) to convert growth bounds into quasi-contraction estimates, which are essential for error propagation analysis.
CTRW Approximation and Subordinator Coupling: The fractional dynamics are approximated by combining:
- A discrete Markov chain scheme ( $U_h$ ) for the spatial diffusion component.
- A heavy-tailed random walk approximation for the inverse stable subordinator (the "clock").
  The convergence analysis separates the error into two components: the error from the spatial discretization (diffusion approximation) and the error from the temporal discretization (subordinator approximation).

Key Results
The paper derives explicit weak convergence rates for the combined CTRW scheme under specific "killing" conditions (where the discount parameter $c$ dominates the growth rates of the semigroup).

Convergence Rates for Unbounded Coefficients:
- Bounded Derivative Regime: For observables in a space with bounded derivatives ( $D_4$ ), the paper establishes a linear convergence rate $O(h)$ for the spatial approximation, provided $c \geq \mu_4$ (where $\mu_4$ is the growth rate of the semigroup on $D_4$ ).
- Weighted Regular Regime: For observables in weighted spaces ( $\mathcal{C}^r_{\alpha, \infty}$ ), similar rates are derived under Assumptions 1–5, which control the growth of coefficients and the stability of the random walk scheme.
Fractional Convergence Theorems (Theorem 12):
The main result provides the convergence rate for the full fractional approximation $u_h(T, \cdot)$ to the true solution $u(T, \cdot)$ in a Banach space $B$ :
- Supercritical Regime ( $c \geq \mu_D$ ): If the killing rate dominates the growth of the smoother space, the error is bounded by:
  $\|u_h - u\|_B \leq C (h + h^{\chi(\beta)}) \|f\|_D,$
  where $\chi(\beta) \in (0, 1)$ is a constant derived from the subordinator approximation.
- Logarithmic Regime ( $\bar{\mu}_B \leq c < \mu_D$ ): If the killing rate dominates the base-space growth but not the stronger smooth-space growth, the error bound degrades to a logarithmic regime:
  $\|u_h - u\|_B \leq C ((\log(1/h))^{-1/\beta} + h^{\chi(\beta)}) \|f\|_D.$
  This regime arises because the deterministic approximation error must be truncated in operational time to ensure integrability against the heavy tails of the inverse subordinator.
Specific Applications:
The results are applied to Geometric Brownian Motion (Black–Scholes model), demonstrating that the theory covers affine coefficients where derivatives vanish for orders $\geq 2$ .

Significance and Claims
The paper claims to provide the first quantitative weak convergence rates for CTRW approximations of time-fractional diffusions with unbounded coefficients. Its significance lies in:

Extension of Theory: Moving beyond the bounded coefficient assumption found in prior literature (e.g., [3]) to include economically and physically relevant models like Geometric Brownian Motion.
Methodological Framework: Developing a robust methodology for analyzing sensitivities of diffusion processes as random tensor fields, utilizing Kunita's stochastic flows and tensorial chain rules to derive growth estimates for unbounded systems.
Precise Error Characterization: Distinguishing between the supercritical regime (linear convergence) and the logarithmic regime, clarifying how the interplay between the killing parameter $c$ and the growth of the diffusion coefficients dictates the convergence behavior.

The authors explicitly state that the verification of the stability assumption (Assumption 5) for specific random walk schemes under unbounded coefficients is left open, and the paper focuses on deriving the convergence rates conditional on this assumption holding. The analysis is strictly theoretical, relying on semigroup techniques and probabilistic estimates rather than numerical experiments.