Bismut Formula for Intrinsic Derivative of DDSDEs with Singular Interactions

Abstract: In recent years, remarkable progress has been made for Distribution dependent stochastic equations (DDSDEs) with singular interactions, existing results include wellposedness, propagation of chaos, entropy cost inequality and ergodicity. As a continuation to the existing study, in this paper we establish Bismut type formulas for the intrinsic derivative of DDSDEs with singular interactions, which extends the existing formula established for the case with Lion's differentiable drifts.

• The paper extends the Bismut–Elworthy–Li formula to the intrinsic derivative of DDSDEs with singular interaction kernels, overcoming previous regularity constraints.
• It establishes precise moment and regularity estimates using maximal function techniques to balance singularity and integrability in the analysis.
• The results enable quantitative sensitivity analysis and chaos propagation in mean-field models, with broad impact on stochastic control and PDE studies.

Bismut Formula for the Intrinsic Derivative of DDSDEs with Singular Interactions

Background and Motivation

The study of distribution-dependent stochastic differential equations (DDSDEs), also referred to as McKean–Vlasov SDEs, has evolved significantly since McKean's seminal work connecting them to nonlinear Fokker–Planck equations. Recent research has extended foundational well-posedness, propagation of chaos, regularity, and ergodicity results to DDSDEs with singular interaction kernels—cases critical for modeling strongly interacting particle systems and nonlocal mean-field phenomena.

For the analysis of regularity and smoothing properties of Markov semigroups associated with SDEs, the Bismut–Elworthy–Li (BEL) formula and its variants are fundamental. Derivatives with respect to the initial law ("intrinsic/Lions derivatives") offer insight into the sensitivity analysis of functional observables on the Wasserstein space, which underpins modern probabilistic analysis and mean-field games. Until this work, BEL-type formulas for these intrinsic derivatives required restrictive regularity assumptions (Lions-differentiability, Dini/Hoelder continuity) on the drift with respect to the measure argument. Such conditions categorically exclude singular interactions—for example, interaction drifts of convolution type with a singular kernel, which occur in physically realistic mean-field models.

Objectives and Main Results

The key objective is to establish a Bismut-type formula for the intrinsic derivative of the law-dependent semigroup associated with singular DDSDEs:

$$dX_t = [b_t(X_t) + (h_t*\mathscr{L}_{X_t})(X_t)]dt + \sigma_t(X_t)dW_t$$

where the interaction kernel $h_t(\cdot)$ is singular in space and fails to be (intrinsically) differentiable in the Wasserstein sense.

The main contributions are:

• Extension of the BEL formula: The intrinsic Bismut formula is proven for DDSDEs with singular interaction kernels, relaxing previous Lions-differentiability or regularity assumptions.
• Precise moment and regularity estimates: The necessary a-priori bounds are established for associated linearized SDEs and the variational equations required for the intrinsic calculus.
• Sharp growth and regularity constraints on the interaction kernel: The analysis handles very rough (non-differentiable) kernels including power-law singularities, as often arise in Coulomb, Riesz, or other strongly-repulsive interactions.

Technical Approach

To address the singularity in the interaction, the study focuses on a function space framework characterized by $L^q(L^p)$-norms, leveraging maximal function techniques (as in [XXZZ]) and tailored regularity classes $K$ for the pair of indices $(p,q)$. The analysis proceeds in the following stages:

1. Well-Posedness and Linearization

Under explicit integrability and regularity assumptions on $b_t$, $\sigma_t$, and the kernel $h_t$ (see Assumption (H)), strong uniqueness and existence of solutions to the DDSDE and a characterization of the push-forward semigroup on the measure space are secured. The moment bounds for the singular interaction terms are provided, ensuring that all nonlinear functionals are well-defined.

The intrinsic derivative is constructed via linearization: Given a direction $\phi$ in the $L^p$ tangent space $h_t(\cdot)$0, the derivative in this direction corresponds to a solution of a variational SDE with singular terms involving the gradient (in measure) of the convolution.

2. Bismut–Elworthy–Li Formula Construction

A novel BEL-type (Bismut) formula is proven for the intrinsic derivative, leveraging stochastic calculus of variations (Malliavin calculus) and controlled Girsanov transformations adapted to the singular DDSDE context. The formula expresses the intrinsic derivative as the expectation of an ‘integration by parts’ functional, specifically:

$h_t(\cdot)$1

where the stochastic integrands are made explicit in terms of the variational flow, the inverse of the diffusion matrix, and the linearized effect of perturbing the initial measure $h_t(\cdot)$2 along $h_t(\cdot)$3.

This formula holds under minimal integrability conditions that balance the singularity exponent of $h_t(\cdot)$4 (via $h_t(\cdot)$5) with the Wasserstein differentiability and the moment index $h_t(\cdot)$6.

3. Quantitative Bounds and Applications

Strong $h_t(\cdot)$7-moment bounds for the variational processes are proven, universally in the initial direction $h_t(\cdot)$8 and the measure $h_t(\cdot)$9. The paper gives explicit estimates showing how the singularity of $L^q(L^p)$0 is ‘averaged out’ in the mean-field interaction, provided the kernel is sufficiently integrable ($L^q(L^p)$1).

A nontrivial example is provided where $L^q(L^p)$2 with $L^q(L^p)$3, and the necessary integrability regime is exhibited.

Theoretical and Practical Implications

The results extend the scope of probabilistic regularity analysis for McKean–Vlasov SDEs to previously inaccessible singular regimes. This has several direct theoretical implications:

• Sensitivity analysis and functional inequalities for empirical measures in singular mean-field models are now feasible, providing a foundation for entropy methods, transportation inequalities, and functional analysis on singular interacting particle systems.
• Propagation of chaos with singular potentials: The formula supports quantitative chaos estimates in numerous models, including those in statistical physics with Riesz or Coulomb-type interactions.
• Gradient and Log-Harnack estimates for the associated nonlocal/Fokker–Planck equations now become tractable, even in the absence of the Lions differentiability.

In practice, the results open the door to:

• Rigorous analysis of mean-field limits and stochastic control problems in singular domains (e.g., swarming models, opinion dynamics with singular influence).
• Stochastic calculus tools for numerical schemes approximating non-smooth mean-field games.

Limitations and Directions for Future Research

While the intrinsic Bismut formula is achieved for a general class of singularities, further weakening of the kernel assumptions (especially into the regime of distributions or $L^q(L^p)$4-type kernels) remains open. The extension to path-space functionals or DDSDEs on manifolds with singular geometry is a promising direction. Leveraging the formula to obtain new functional inequalities, gradient flows, or large deviations for singular McKean–Vlasov diffusions is an interesting challenge for subsequent work.

Conclusion

This work resolves a substantial technical barrier in the stochastic analysis of DDSDEs with singular interactions by providing a rigorous Bismut-type formula for the intrinsic/Lions derivative of the law-dependent semigroup, precisely characterizing the regularity regime necessary for validity. These results are foundational for further development in the theory of singular interacting particle systems, stochastic control, and the analysis of mean-field PDEs with rough or singular nonlocal couplings (2604.08899).