Multiscale Nudging: From Macroscopic Observations to Microscopic Dynamics

Abstract: We introduce a measure-based nudging framework for assimilating macroscopic observations into microscopic mean-field particle dynamics. The central difficulty is a representation mismatch: the forecast is a labeled particle system, while the observations specify only a smoothed, permutation-invariant density. To address this mismatch, we define the forecast-observation discrepancy as a quadratic functional on probability measures after applying the same smoothing operator used by the observation process. The Wasserstein gradient of this functional induces a transport velocity on state space, which yields a particle-level correction without constructing particle-to-particle matching, linearizing the dynamics, or estimating ensemble covariances. For a fixed observation scale, we prove well-posedness of the assimilated McKean-Vlasov dynamics and propagation of chaos for the interacting particle approximation. Under exact smoothed observations and an observability condition at the kernel scale, we establish an $L<sup>2$-stability estimate showing exponential decay up to a bias floor controlled by model misspecification. Numerical experiments on linear, bimodal, chaotic, kinetic, and collective-motion systems demonstrate that the method can recover macroscopic structure from incomplete density-level observations.

• The paper develops a multiscale nudging approach that minimizes a smoothed measure-level misfit functional to align microscopic forecasts with macroscopic observations.
• It demonstrates robust performance across regimes—including linear, bimodal, and chaotic systems—with significant reductions in error, such as one order of magnitude lower L2 error.
• The framework avoids particle matching and covariance estimation by using permutation-invariant updates, providing strong theoretical stability and convergence guarantees.

Multiscale Nudging: Measure-Based Data Assimilation for Mean-Field Particle Dynamics

Problem Motivation and Representation Mismatch

The paper introduces a measure-theoretic data assimilation framework designed to correct microscopic particle forecasts using only macroscopic, permutation-invariant observations. Mean-field systems appear in diverse domains such as fluid mechanics, kinetic theory, collective behavior, and computational neuroscience. In these systems, the governing dynamics typically operate at the microscopic level (labeled particles, $N \gg 1$), but available data from physical measurements or experiments—the so-called observations—often manifest only at a coarse, macroscopic (aggregate, smoothed density) scale.

Classical assimilation techniques such as EnKF, particle filtering, and variational (3D/4D-Var) methods encounter conceptual and computational barriers in this setting. These approaches typically (i) act on labeled configurations or covariance structures that lack permutation-invariance, or (ii) require explicit particle-to-observation matching, which is infeasible when only marginal or smoothed densities are observed. This leads to representation mismatch and ill-posedness when assimilating macroscopic data into microscopic particle states.

Measure-Based Nudging and Theoretical Formulation

The proposed multiscale nudging method constructs a discrepancy functional $J_\mathrm{nud, h}(v, p_\mathrm{obs})$ by applying the same smoothing kernel $K_h$ to both the empirical forecast measure $v$ and the observation $p_\mathrm{obs}$. This yields a misfit functional defined purely at the measure level:

$$J_\mathrm{nud, h}(v, p_\mathrm{obs}) = \int |K_h * v(x) - K_h * p_\mathrm{obs}(x)|^2 dx,$$

which is then differentiated in the Wasserstein metric to produce a feedback vector field. The gradient flow (in $W_2$) induced by this functional yields a transport correction at the microscopic level:

$$\mathrm{d}Z_t = b_{\text{model}}(Z_t, v_t) dt - \lambda \nabla_x (K_h * (K_h * v_t - p_\mathrm{obs}))(Z_t) dt + \Sigma dW_t,$$

where $b_{\text{model}}$ is the possibly misspecified learned drift and $\Sigma$ governs diffusion. Feedback is thus produced at the macroscopic scale (coarse density) but imposed directly on individual particles, circumventing the need for particle-identity bookkeeping or matching.

The authors provide rigorous analysis of the well-posedness of the resulting McKean-Vlasov SDE, propagation of chaos for finite-particle implementations, strict positivity of the assimilated measure, and—under an observability condition—an explicit $J_\mathrm{nud, h}(v, p_\mathrm{obs})$0 contraction estimate. The latter demonstrates exponential error decay to a bias floor controlled by model misspecification and vanishing in the idealized (exact model) regime.

Theoretical stability and convergence guarantees are established without requiring ensemble covariance estimation, particle matching, or linearization. This is a notable departure from the requirements of standard filtering and variational methods.

Numerical Assessment Across Regimes

A comprehensive suite of experiments is conducted:

• Linear Gaussian Test: For biased interaction strengths, multiscale nudging robustly aligns forecast statistics (variance, density) with the reference law for increasing nudging intensity, outperforming unassimilated or weakly assimilated forecasts.
• Bimodal System: The scheme recovers correct modal structure in complex double-well systems, moving mass across metastable barriers, and reconstructing full distribution profiles from smoothed observations. Final-time and time-averaged Wasserstein-2 errors decrease monotonically with feedback strength.
• Mean-Field Lorenz Attractor: In strongly nonlinear and chaotic regimes, density-level feedback stabilizes the mean-trajectory on the correct attractor lobe, suppressing large trajectory errors.
• Vlasov-Poisson (Landau Damping, Two-Stream Instability): Particle forecasts corrected with measure-based nudging recover key qualitative features (electric field decay, filamentation, vortex merging) absent from the biased initializations.
• Experimental Collective Motion (Fish School): On real biological data lacking reliable particle identities, assimilated forecasts retain coherent structure (collective circulation, low-density regions) over significant timescales, quantitatively exhibiting one order of magnitude lower $J_\mathrm{nud, h}(v, p_\mathrm{obs})$1 error compared to uncorrected rollouts.

These experiments collectively demonstrate the generality of the method and its independence from the specifics of the underlying drift, kernel bandwidth, or observation operator beyond the basic compatibility assumptions.

Implications and Future Directions

The multiscale nudging paradigm advances measure-level data assimilation for large-scale particle systems by providing a mathematically principled, permutation-invariant update rule compatible with macroscopic information. This decoupling of correction and labeling is essential for practical deployment in real systems (e.g., turbulent fluids, large animal collectives, and experimental settings with incomplete data).

Practical implications:

• Assimilation at the measure level enables robust error control incorporating only observation-scale information, obviating the need for microscopic access or identity tracking.
• The algorithm can be efficiently implemented in high-dimensional settings without suffering from weight degeneracy (unlike SMC) or covariance degeneracy (unlike EnKF).

Theoretical implications:

• The convergence guarantees, positivity maintenance, and explicit error contraction to a model-dependent floor provide strong justification for the reliability of the approach.
• The presentation lays groundwork for extending measure-level feedback to partially observed, noisy, or non-Gaussian scenarios, as well as for adaptive multiscale schemes.

Future work should address nontrivial observation operators, phase-space marginal observations (particularly for kinetic equations), robustness to noise, and adaptive bandwidth or kernel selection. Expanding theoretical results to accommodate partial observability and noisy and nonlinear measurements is a critical direction. Integration with parameter and model learning frameworks may further advance the practical impact in scientific computation and model reduction.

Conclusion

This paper presents a rigorously analyzed and versatile framework—multiscale nudging—for assimilating macroscopic observations into the dynamics of interacting mean-field particle systems. By defining and minimizing a smoothed measure-level discrepancy, and correcting using the Wasserstein gradient, the method bridges the gap between permutation-invariant observations and microscopic dynamics. The strong theoretical results and comprehensive experimental assessment across a spectrum of particle models underscore its utility for data assimilation in large, high-dimensional stochastic dynamical systems where only aggregate, coarse-grained measurements are available (2606.06809).