Convergence rate of numerical scheme for SDEs with a distributional drift in Besov space

This paper establishes the strong $L^1$ convergence rate of an Euler-Maruyama numerical scheme for one-dimensional stochastic differential equations with distributional drifts in Hölder-Zygmund spaces, supported by theoretical bounds and numerical implementation.

The Gist

Imagine you are trying to predict the path of a drunk sailor walking on a ship deck. In the world of mathematics, this sailor is a Stochastic Differential Equation (SDE). His path is determined by two things:

1. The Wind (Brownian Motion): A random, jerky force that pushes him around unpredictably.
2. The Drift (The Drift): A force that tries to pull him in a specific direction, like a current or a slope.

Usually, mathematicians assume the "Drift" is a smooth, well-behaved slope. You can easily calculate where the sailor will go next if the slope is gentle and continuous.

The Problem: The "Ghost" Slope
In this paper, the authors tackle a much harder problem. They ask: What if the slope isn't smooth at all? What if it's a "distributional drift"?

Think of this drift not as a smooth hill, but as a chaotic, jagged landscape made of invisible spikes, dust, and mathematical "ghosts." It's so rough that if you tried to measure the slope at any single point, the number would be infinite or undefined. In math terms, the drift is a "generalized function" living in a Besov space (a fancy way of saying it's a very rough, negative-smoothness object).

Trying to simulate a sailor walking on a landscape of invisible, jagged spikes using standard computer methods is like trying to draw a picture of a storm using a ruler. The standard tools break down.

The Solution: The Two-Step "Blurring" Strategy
The authors designed a new numerical scheme (a recipe for a computer to solve the problem) that works in two clever steps:

Step 1: The "Heat Gun" (Smoothing)

Since the landscape is too jagged to walk on, the authors use a mathematical tool called the Heat Semigroup.

• The Analogy: Imagine taking a high-resolution photo of that jagged, chaotic landscape and running a "Gaussian Blur" filter over it. The sharp spikes turn into gentle, rolling hills.
• The Math: They apply the "Heat Kernel" to the rough drift. This turns the impossible-to-handle "ghost" drift into a smooth, friendly function that a computer can actually understand.
• The Catch: Blurring changes the landscape. If you blur it too much, you lose the true shape. If you don't blur it enough, the computer still crashes. The authors had to find the perfect amount of blur.

Step 2: The "Staircase" (Euler-Maruyama Scheme)

Once the landscape is smoothed out, they use a standard method called the Euler-Maruyama scheme.

• The Analogy: Imagine the sailor is walking on this now-smooth hill. Instead of calculating his path continuously (which is hard), the computer takes tiny steps. It asks: "If he walks for 1 second, where does he end up?" Then it updates the position and asks again for the next second.
• The Trick: Because the landscape was smoothed in Step 1, these steps are safe to take.

The Balancing Act
The genius of the paper lies in balancing the two errors:

1. The Blur Error: How much did we distort the original jagged landscape by smoothing it?
2. The Step Error: How much did we miss by taking big steps instead of walking continuously?

The authors proved that if you choose the "blur amount" and the "step size" just right, the errors cancel each other out in a specific way. They derived a Convergence Rate, which is essentially a speed limit for how fast their computer simulation gets closer to the "true" answer as you make the steps smaller.

The Results: How Fast is Fast?

• The Theory: They proved mathematically that their method works and gave a formula for the speed of convergence. It depends on how "rough" the original drift was. The rougher the drift, the slower the convergence, but it still works!
• The Experiment: They wrote code (in Python) to test this. They created a "rough drift" using a Fractional Brownian Motion (a mathematical model for very rough, self-similar noise, like a coastline).
• The Surprise: Their computer experiments showed that the method worked even better than their strict mathematical proof suggested. The empirical results hinted that the speed might be twice as fast as their theoretical lower bound predicted. It's like proving a car can go 60 mph, but in the real world, it consistently hits 120 mph.

Why Does This Matter?
This is a big deal because many real-world phenomena involve "rough" forces that aren't smooth.

• Finance: Market crashes or sudden jumps in stock prices.
• Physics: Turbulent fluids or particles moving through rough media.
• Biology: Movement of cells in complex, messy environments.

Before this paper, we didn't have a reliable way to simulate these "rough" systems on a computer. The authors built a bridge between the world of impossible, jagged mathematics and the world of practical, smooth computer simulations. They showed us how to walk on the jagged rocks by first turning them into a smooth path, and they proved exactly how fast we can get there.

Technical Summary

Here is a detailed technical summary of the paper "Convergence rate of numerical scheme for SDEs with a distributional drift in Besov space" by Chaparro J´aquez, Issoglio, and Palczewski.

1. Problem Statement

The paper addresses the numerical approximation of one-dimensional Stochastic Differential Equations (SDEs) of the form:
$$X_t = x + \int_0^t b(s, X_s) ds + W_t$$
where $W_t$ is a standard Brownian motion. The core challenge lies in the drift coefficient $b$, which is not a function but a generalized distribution (Schwartz distribution) belonging to a negative-order Hölder-Zygmund space (equivalent to a Besov space) $C^{-\gamma}$ with $-\gamma < 0$.

Specifically, the drift $b$ is assumed to be $1/2$-Hölder continuous in time with values in$C^{-\gamma}$(denoted$b \in C^{1/2}_T C^{-\gamma}$).

• Context: While existence and uniqueness of solutions (as "virtual" or weak solutions) for such SDEs have been established theoretically (e.g., via martingale problems), there is a lack of numerical schemes with proven convergence rates for distributional drifts.
• Gap: Existing numerical literature for SDEs with low-regularity coefficients typically assumes the drift is at least a measurable function (positive Sobolev or Hölder regularity). This paper tackles the more singular case where the drift is a distribution.

2. Methodology

The authors propose a two-step numerical scheme to handle the singularity of the drift:

Step 1: Regularization via Heat Semigroup

Since the drift $b$ is a distribution, it cannot be directly evaluated in a standard Euler-Maruyama scheme. The authors regularize $b$ by convolving it with the heat kernel (applying the heat semigroup $P_{1/N}$):
$$b_N = P_{1/N} b$$

• Effect: This transforms the distributional drift $b$ into a smooth function $b_N$ (specifically, $b_N \in C^1_b$).
• Justification: The heat semigroup provides smoothing properties quantified by Schauder estimates, allowing the authors to control the error introduced by this approximation.

Step 2: Euler-Maruyama Discretization

Once the drift is smoothed to $b_N$, the authors apply the standard Euler-Maruyama (EM) scheme to the regularized SDE:
$$dX^N_t = b_N(t, X^N_t)dt + dW_t$$
The discrete approximation $X^N_m$ uses a time step $\Delta t = T/m$.

Error Analysis Strategy

The total error is decomposed into two parts:

1. Smoothing Error: The difference between the true solution $X$ (with drift $b$) and the regularized solution $X^N$ (with drift $b_N$).
2. Discretization Error: The difference between the regularized solution $X^N$ and its EM approximation $X^N_m$.

Key Technical Tools:

• Virtual Solutions: The authors utilize the concept of "virtual solutions" derived from a Kolmogorov-type PDE. They define a transformation $\phi(t, x) = x + u(t, x)$, where $u$ solves a PDE involving the drift. This transforms the singular SDE into a standard SDE with a strong solution, facilitating the error analysis.
• Local Time Bounds: A critical technical hurdle is that standard $L^2$ error analysis (using Itô's formula on $(X^N - X)^2$) fails because the diffusion coefficient of the auxiliary process is only Hölder continuous, preventing the application of Gronwall's lemma. Instead, the authors analyze the $L^1$ strong error and rely on a delicate bound for the local time of the error process at zero, borrowing techniques from De Angelis et al. [9].
• Parameter Balancing: The convergence rate depends on the regularization parameter $N$ and the time steps $m$. The authors optimize the relationship $N(m) = m^\eta$ to balance the smoothing error (which decreases as $N$ increases) and the discretization error (which decreases as $m$ increases).

3. Key Contributions

1. First Convergence Rate for Distributional Drifts: The paper establishes the first theoretical strong convergence rate for an Euler-Maruyama scheme applied to SDEs with drifts in negative Besov spaces ($C^{-\gamma}$).
2. Theoretical Upper Bound: The authors prove that the strong $L^1$ convergence rate is bounded by:
   $$r(\hat{\beta}) = \left( \frac{\hat{\beta} + 1}{(1/2 - \hat{\beta})^2 + 2} \right)^{-1}$$
   where $\hat{\beta}$ characterizes the roughness of the drift (related to the negative index $-\gamma$).
3. Handling Non-Standard Regularity: The work extends numerical analysis techniques to cases where the drift is not a function, requiring novel estimates on local times and the use of $L^1$ norms rather than the standard $L^2$.
4. Numerical Implementation: The authors provide a practical implementation strategy for computing the regularized drift $b_N$ using fractional Brownian motion (fBm) paths as test cases, addressing the difficulty of discretizing distributions.

4. Results

• Theoretical Rate: The derived convergence rate $r(\hat{\beta})$ depends on the regularity of the drift.
  • As the drift becomes smoother ($\hat{\beta} \to 0$, approaching measurable functions), the rate approaches 1/6.
  • As the drift becomes rougher ($\hat{\beta} \to 1/2$), the rate approaches 0.
  • Note: The authors acknowledge that 1/6 is likely not optimal for measurable drifts (where rates of 1/2 are known), suggesting their bound is conservative due to the specific techniques used for distributions.
• Numerical Experiments:
  • The authors implemented the scheme in Python using fBm paths to generate rough drifts.
  • Empirical Observation: The empirical convergence rates observed in simulations were significantly higher than the theoretical lower bound. The data suggested a rate close to $1/2 - \hat{\beta}/2$.
  • This empirical rate aligns with the natural extension of results for positive regularity (where rate $\approx 1/2 + \gamma/2$) to negative regularity.
  • Conclusion: The theoretical bound is likely suboptimal, and the true rate may be $1/2 - \hat{\beta}/2$.

5. Significance

• Bridging Theory and Computation: This work is a significant step forward in making SDEs with distributional coefficients computationally tractable, which is relevant for modeling phenomena in physics and finance where coefficients may be singular (e.g., polymer measures, singular potentials).
• Methodological Innovation: The use of heat semigroup regularization combined with local time estimates for $L^1$ convergence provides a robust framework for analyzing singular SDEs.
• Future Directions: The discrepancy between the theoretical bound (1/6) and empirical results ($1/2 - \hat{\beta}/2$) highlights a gap in the current theory. The authors suggest that future research utilizing the Stochastic Sewing Lemma (as used in recent works by Butkovsky et al.) might be required to prove the optimal rate for distributional drifts.

In summary, the paper successfully constructs a numerical scheme for SDEs with distributional drifts, proves its convergence, and provides numerical evidence suggesting that the convergence rate is better than the proven theoretical bound, pointing toward a more optimal rate of $1/2 - \hat{\beta}/2$.