Derivative estimates for SDEs with singular and unbounded coefficients

1. Problem Statement

The paper addresses the fundamental challenge of establishing pointwise gradient and Hessian estimates for Markov semigroups associated with Stochastic Differential Equations (SDEs) of the form:
$$dX_t = b(X_t)dt + \sqrt{2}\sigma(X_t)dW_t$$
where the coefficients $b$ (drift) and $\sigma$ (diffusion) are allowed to be singular (locally irregular) and unbounded (growing at infinity).

Key Challenges:

• Lack of Global Smoothness: Classical methods (Malliavin calculus, Bismut formulas) typically require global smoothness or non-degeneracy, which fails when coefficients are singular or merely Hölder continuous.
• Unboundedness: Standard coupling methods and spectral techniques often rely on global Lipschitz conditions or uniform ellipticity, which are violated when coefficients exhibit superlinear growth or degenerate behavior at infinity.
• Quantitative Dependence: Existing literature often provides estimates with constants that depend on global norms of coefficients. The authors seek estimates where constants depend explicitly and locally on the coefficients (e.g., norms on a unit ball $B_1(x)$), allowing for localized analysis of irregular systems.

2. Methodology

The authors develop a unified PDE-probabilistic framework that bridges local PDE regularity theory with probabilistic representations of SDE solutions.

A. Local PDE Analysis (The Backbone)

Instead of relying on global probabilistic formulas, the authors first establish a priori local regularity estimates for the associated time-inhomogeneous Kolmogorov equation:
$$\partial_t u + \mathcal{L}_t u + f = 0, \quad \mathcal{L}_t = \text{tr}(a(t,x)\nabla^2_x) + b(t,x)\cdot\nabla_x$$

• Local Maximal Regularity: They derive local $L^q_t L^p_x$ maximal regularity estimates. Crucially, the constants in these estimates depend explicitly on the local $L^\infty$-norms of the coefficients and the local ellipticity ratio $\Lambda(x)/\lambda(x)$.
• Local Schauder Estimates: They obtain Hölder estimates for the Hessian $\nabla^2 u$ under local Hölder continuity assumptions on coefficients.
• Technique: The proofs utilize a "freezing coefficients" technique combined with parabolic scaling and anisotropic interpolation inequalities (Gagliardo-Nirenberg type). A novel scale selection procedure determines the intrinsic radius of the parabolic cylinder based on the local ellipticity and regularity norms, ensuring the estimates remain valid even when coefficients vary wildly.

B. Probabilistic Representation

Once the PDE estimates are established, the authors link them to the SDE semigroup $T_t \phi(x) = \mathbb{E}[\phi(X_t(x))]$ via the probabilistic representation of solutions.

• They prove that the semigroup solution satisfies the Kolmogorov equation in a weak sense (using local Sobolev spaces).
• By applying the PDE estimates to the probabilistic solution, they translate the local PDE bounds into pointwise gradient and Hessian bounds for the SDE semigroup.

C. Long-Time Behavior

For time-homogeneous SDEs, they combine the short-time regularization (derived from the PDE estimates) with ergodic decay assumptions (Lyapunov conditions). This allows them to derive long-time decay bounds for derivatives, linking short-time smoothing to long-time convergence to equilibrium.

3. Key Contributions

1. Explicit Local Dependence: The most significant contribution is that the constants in the gradient/Hessian estimates depend only on local norms of the coefficients (e.g., $\|b\|_{L^{p_b}(B_1(x))}$, $[\sigma]_{C^\alpha(B_1(x))}$) and the local ellipticity ratio. No global boundedness, smoothness, or uniform ellipticity is required.
2. Unified Framework for Singular/Unbounded Coefficients: The framework handles coefficients that are:
   • Singular: Drifts in negative-order Sobolev spaces or distributions.
   • Unbounded: Drifts with superlinear growth (e.g., polynomial growth).
   • Degenerate: Diffusion matrices that may degenerate at infinity, provided local ellipticity holds.
3. Sharp Short-Time and Long-Time Bounds:
   • Short-time: Establishes the optimal $t^{-1/2}$ singularity for gradients and $t^{-1}$ for Hessians, with constants explicitly tracking local coefficient behavior.
   • Long-time: Provides decay rates (exponential or polynomial) for derivatives, quantifying the stabilization of the system.
4. Extension to Distributional Drifts: The authors successfully apply their framework to SDEs with distributional drifts (e.g., $b \in H^{\alpha}_p$ with $\alpha < 0$) using a Zvonkin-type transform, yielding the first pointwise gradient estimates for such systems.

4. Main Results

Theorem 1.2 (Short-Time Estimates)

Under local regularity assumptions ($\sigma \in C^\alpha_{loc}$, $b \in L^{p_b}_{loc}$) and a moment condition (Lyapunov function $\rho$):

• Gradient: $|\nabla_x T_t \phi(x)| \lesssim \frac{\Gamma_t(x)}{\sqrt{t} \wedge 1} \ell_0(2t) \|\phi\|_{B_{\rho_0}}$.
  • Here, $\Gamma_t(x)$ depends on local norms of $b$ and $\sigma$ on $B_1(x)$ and the local ellipticity ratio.
• Hessian: Similar bounds hold for $[\nabla^2_x T_t \phi]_{C^\beta}$, scaling as $(\sqrt{t} \wedge 1)^{-2-\beta}$.
• Significance: The leading-order constants do not involve global norms, only local ones.

Theorem 1.5 (Long-Time Estimates)

Assuming an ergodic condition (existence of an invariant measure $\mu$ and decay rate $\ell_1(t)$):

• Gradient: $|\nabla_x T_t \phi(x)| \lesssim \ell_1(t-1) \cdot (\text{Local Coefficients}) \cdot \|\phi\|_{B_{\rho_0}}$.
• Significance: The derivative decays at the same rate as the semigroup converges to equilibrium, quantitatively connecting short-time regularization with long-time stability.

Theorem 5.1 & 5.2 (Poisson Equations)

The authors establish the solvability and regularity of the Poisson equation $-\mathcal{L}u = f$ on $\mathbb{R}^d$ with singular/unbounded coefficients.

• They prove $u \in W^{2,p}_{loc}$ and provide quantitative Schauder-type bounds for the solution $u$ in terms of local coefficient norms. This is crucial for homogenization and averaging problems.

Theorem 6.3 (Distributional Drifts)

For SDEs with drift $b = b_1 + b_2$ where $b_1$ is a distribution (in a Bessel potential space) and $b_2$ has linear growth:

• They derive the gradient estimate: $|\nabla_x \mathbb{E}[\phi(X_t)]| \leq C(1 + |x| + t^{-1/2})\|\phi\|_\infty$.
• This is achieved by transforming the singular SDE into a regular one via a Zvonkin transform and applying the main theorems.

5. Significance and Applications

• Homogenization and Averaging: The explicit local dependence allows for rigorous error analysis in multi-scale problems where coefficients vary rapidly or are unbounded, without requiring global smoothness.
• Numerical Analysis: The derivative estimates are essential for analyzing the weak error of numerical schemes (e.g., Euler-Maruyama) for SDEs with rough or unbounded coefficients.
• Stochastic Control: The results provide the necessary regularity for value functions in control problems with rough coefficients, facilitating the verification of near-optimal controls.
• Rough Path Theory: The ability to handle distributional drifts extends the applicability of SDE theory to models arising in fluid dynamics and physics where noise or drifts are not functions but distributions.

In summary, this paper breaks the reliance on global smoothness assumptions in stochastic analysis by introducing a robust, localized PDE-probabilistic framework. It provides the first quantitative, pointwise derivative estimates for a broad class of SDEs with singular and unbounded coefficients, bridging a critical gap between theoretical regularity and practical applications in complex, irregular environments.