PerturbDiff: Stability of Diffusive Systems

• PerturbDiff is a framework that defines explicit non-asymptotic bounds for the stability of diffusive stochastic systems under small parameter perturbations.
• It leverages parametrix expansions, convolution techniques, and normed distance measures to provide actionable insights into error propagation.
• The approach underpins practical applications such as uncertainty quantification, robust numerical schemes, and sensitivity analysis for both continuous diffusions and their Euler approximations.

PerturbDiff refers to a suite of rigorous, quantitative, and algorithmically constructive approaches for analyzing, predicting, and controlling the sensitivity (“stability”) of diffusive stochastic systems—continuous or discrete, single or multi-particle—to small perturbations of underlying model parameters, coefficients, or structures. The concept underlies several lines of research in stochastic analysis, statistical mechanics, and applied computation, with explicit non-asymptotic stability bounds, parametrix expansions, and practical procedures for uncertainty quantification. Here, we survey foundational results, methodologies, and ongoing developments in PerturbDiff, emphasizing the canonical framework for Markovian diffusions and Markov chains under coefficient misspecification (Konakov et al., 2015).

1. Mathematical Setup: Diffusion and Perturbation Framework

Consider a diffusion process on $[0,T]\times\mathbb{R}^d$ governed by the stochastic differential equation:

$$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$

where $b$ (drift) and $\sigma$ (diffusion matrix) are bounded, measurable functions, and $W_t$ is standard Brownian motion in $\mathbb{R}^d$. The process admits a transition density $p(s, t, x, y)$. A perturbed system is defined similarly with coefficients $(b_\varepsilon, \sigma_\varepsilon)$, yielding $X_t^{(\varepsilon)}$ and $p_\varepsilon(s, t, x, y)$.

Perturbations are quantified via:

• The $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$0 and $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$1 distances:

$$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$2

• Diffusion/Hölder distances:

$$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$3

• Cumulative perturbation norm:

$$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$4

The transition densities and their sensitivity to perturbations of $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$5 form the central object of analysis.

2. Key Assumptions and Structural Properties

PerturbDiff results rest on sharp regularity and nondegeneracy conditions (Konakov et al., 2015):

• Boundedness: $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$6 and $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$7.
• Uniform ellipticity: For all $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$8 and $$dX_t = b(t, X_t)\,dt + \sigma(t, X_t)\,dW_t,\qquad X_0 = x$$9, $b$0, and likewise $b$1, satisfy

$b$2

• Spatial Hölder continuity: For some $b$3 and constant $b$4,

$b$5

This framework accommodates general non-degenerate SDEs and their Markov chain (Euler–Maruyama) approximations.

3. Main Stability Theorems for Diffusions and Markov Chains

Continuous-Time Diffusions

The core result establishes first-order stability in the perturbation magnitude $b$6 (Konakov et al., 2015):

$b$7

where $b$8 and $b$9 depend on model parameters.

Integrating yields an $\sigma$0-stability bound: $\sigma$1

Markov Chain Euler Approximations

For discretizations of the same SDE (Euler-step Markov chains), with one-step increments driven by either Gaussian or high-decay density innovations, the transition density difference is controlled as

$\sigma$2

where $\sigma$3 is a proxy—Gaussian for Gaussian noise, stable-like for polynomial decay—mirroring transition tails.

The conclusion is that stability persists under time-discretization and non-Gaussian driving noise, with explicit control dictated by the normed difference of coefficients.

4. Parametrix Expansion and Analytical Techniques

The proof leverages parametrix representations of the transition density. This involves:

• Freezing coefficients: At the terminal point $\sigma$4 to define a Gaussian “proxy” process with density $\sigma$5.
• Difference kernel: $\sigma$6, where $\sigma$7 is the true generator, $\sigma$8 that of the frozen system.
• Iterated convolution: Yields the expansion

$\sigma$9

with $W_t$0, $W_t$1, and $W_t$2 denoting time–space convolution.

• Analytic bounds: Gaussian comparison, Hölder-continuity, and moment estimates ensure absolute convergence and furnish the sharp stability result.

The discrete case (Markov chain) adapts these constructions to finite steps, using discrete convolution and Edgeworth expansions.

5. Quantitative Bounds and Applications

The PerturbDiff result yields explicit, uniform, and non-asymptotic quantitative control:

• For $W_t$3, one has $W_t$4, i.e., the transition densities are linearly sensitive to small uniform or $W_t$5 coefficient errors.
• The same order applies to Euler-type Markov chain densities, including for innovations with only polynomial decay.
• This applies robustly to pricing functionals, forward/reverse Monte Carlo schemes, and scenarios where regularization or data-driven SDE inference introduces perturbations.

A summary of key formulas:

Concept	Formula	Notes
Diffusion stability	$W_t$6	$W_t$7: Gaussian proxy
Discrete chain stability	$W_t$8	$W_t$9: proxy density
Perturbation norm	$\mathbb{R}^d$0	Combines drift/diffusion

6. Broader Context and Related Sensitivity Frameworks

The PerturbDiff theory aligns with and informs parallel developments:

• Stability for degenerate diffusions: Extensions to Kolmogorov-type/hypoelliptic SDEs yield similar stability bounds, with two-scale Gaussian proxies and parametrix expansions (Kozhina, 2016).
• Many-body and Hamiltonian systems: In ergodic particle systems, perturbation spreading maps to a Lévy-walk–style single-particle propagator that predicts fronts and superdiffusive scaling—a distinct but conceptually related perturbation analysis (Zaburdaev et al., 2011).
• Numerical homogenization: Methods such as Petrov–Galerkin Localized Orthogonal Decomposition efficiently solve parameterized elliptic problems by recomputing only locally-affected multiscale corrections, guided by error indicators of the type $\mathbb{R}^d$1 (Hellman et al., 2019).
• Stochastic control and Riccati flows: Non-asymptotic expansions for matrix Riccati diffusions (covariance evolution in EnKF/Kalman filter) provide bias–variance control under finite-$\mathbb{R}^d$2 perturbations (Bishop et al., 2017).
• Fluctuating hydrodynamics and large deviations: Perturbative computation of quasi-potentials in non-equilibrium mean-field models employs linearization and transport-type equations for the rate function correction (Bouchet et al., 2015).

7. Significance and Implications

The PerturbDiff paradigm establishes that, under mild regularity and nondegeneracy assumptions, transition densities of diffusions and their chain approximations are uniformly stable under small (normed) perturbations of their coefficients. This enables:

• Rigorous uncertainty quantification for model misspecification or numerical discretization.
• Analytical and algorithmic sensitivity evaluation for parameter inference, filtering, and simulation.
• Direct transfer of finite-dimensional bounds to functionals and path-dependent observables, facilitating robust control and inference in applied settings.

The explicit dependence on normed differences $\mathbb{R}^d$3 quantifies the risk posed by modeling, numerical, or statistical inaccuracies in the input data or parameters, with the resulting $\mathbb{R}^d$4 and $\mathbb{R}^d$5 stability bounds precisely characterizing the propagation of errors through the stochastic system’s law (Konakov et al., 2015).