Variational Data-Consistent Assimilation

TL;DR

This work develops DC-4D-Var and DC-WME 4D-Var by embedding data-consistent inversion into four-dimensional variational data assimilation, introducing a predictability-aware regularization that leverages a QoI map to stabilize estimation in nonlinear, partially observed systems. The authors derive the DC cost functions, Hessians (including Gauss-Newton approximations), and adjoint formulations, and prove existence and uniqueness of minimizers under a predictability framework. Numerical experiments on SWE and chaotic ODEs (Lorenz-63, Lorenz-96) show that DC-WME 4D-Var reduces RMSE and bias relative to standard 4D-Var and DC-4D-Var, particularly under high observation noise and short assimilation windows, with only modest computational overhead. The approach demonstrates robustness and scalability for high-dimensional data assimilation, providing a principled link between probabilistic DC updates and variational optimization, and suggesting practical impact for real-time, PDE-constrained forecasting such as coastal storm surge prediction.

Abstract

This work introduces a new class of four-dimensional variational data assimilation (4D-Var) methods grounded in data-consistent inversion (DCI) theory. The methods extend classical 4D-Var by incorporating a predictability-aware regularization term. The first method formulated is referred to as Data-Consistent 4D-Var (DC-4DVar), which is then enhanced using a Weighted Mean Error (WME) quantity-of-interest map to construct the DC-WME 4D-Var method. While the DC and DC-WME cost functions both involve a predictability-aware regularization term, the DC-WME function includes a modification to the model-data misfit, thereby improving estimation accuracy, robustness, and theoretical consistency in nonlinear and partially observed dynamical systems. Proofs are provided that establish the existence and uniqueness of the minimizer and analyze how a predictability assumption that is common within the DCI framework helps to promote solution stability. Numerical experiments are presented on benchmark dynamical systems (Lorenz-63 and Lorenz-96) as well as for the shallow water equations (SWE). In the benchmark dynamical systems, the DC-WME 4D-Var formulation is shown to consistently outperform standard 4D-Var in reducing both error and bias while maintaining robustness under high observation noise and short assimilation windows. Despite introducing modest computational overhead, DC-WME 4D-Var delivers improvements in estimation performance and forecast skill, demonstrating its potential practicality and scalability for high-dimensional data assimilation problems.

Figures (10)

• Figure 1: Spatial configurations (A) and (B) of observation stations (shown as red dots) used in the SWE experiments. Stations are distributed along transects perpendicular to the shoreline within a two-dimensional domain discretized into 144 triangular elements. These configurations are used as representative examples to demonstrate that all three assimilation methods perform equivalently in a well-conditioned, periodic setting, thereby serving as a baseline verification of correct implementation.
• Figure 2: Reconstruction errors in surface elevation on Day 3 of the SWE experiment for configurations (A) and (B). Each column corresponds to one of the three assimilation methods: standard 4D-Var, DC 4D-Var, and DC-WME 4D-Var. Error magnitudes and spatial patterns are nearly indistinguishable across methods, confirming that all implementations behave equivalently in this smooth, periodic regime. This representative snapshot serves as a baseline “sanity check” before applying the methods to more sensitive dynamical systems.
• Figure 3: Distribution of estimated lower bounds on the background variance $\sigma_b^2$ computed over 50 random trajectories of the Lorenz-63 system. For each trajectory, the bound is derived from the smallest eigenvalue of the forecast-observation Gram matrix, following the inequality \ref{['eq:eig_low_bound']}. The vertical dashed line denotes the average bound across all samples, which exceeds $4\sigma_{\mathrm{obs}}^2$, confirming the need for background inflation to maintain consistent and stable assimilation in nonlinear regimes.
• Figure 4: Time-averaged RMSE for 4D-Var, DC 4D-Var, and DC-WME 4D-Var applied to the Lorenz 63 system under varying observation noise levels. Each method uses a fixed inflation level of $\alpha = 4\sigma_{\mathrm{obs}}^2$, corresponding to the theoretically motivated lower bound on background variance. Results demonstrate the robustness of DC-WME 4D-Var to increasing observation noise, with the best-performing method at each noise level indicated by a blue star.
• Figure 5: Time-averaged Root Mean Squared Error (RMSE) over 1000 model time steps for 4D-Var, DC 4D-Var, and DC-WME 4D-Var applied to the Lorenz-63 system. RMSE is computed using the analysis state across 50 assimilation cycles. The results show that DC-WME 4D-Var consistently achieves the lowest error and exhibits greater stability over time, especially in the presence of nonlinear and chaotic dynamics.

Theorems & Definitions (9)

• Theorem A.1: Existence and Uniqueness
• Lemma B.1: Strict Convexity of $\mathcal{J}_{\mathrm{DC}}$
• proof
• Lemma B.2: Coercivity of the DC 4D-Var function
• proof
• Theorem B.1: Existence of Minimizer for DC 4D-Var
• proof
• Theorem B.2: Uniqueness of the DC 4D-Var Minimizer
• proof