Continuous-Discrete Ensemble Kalman Filtering

• Continuous-discrete EnKF is a sequential Bayesian state estimator that blends continuous-time particle propagation with discrete-time observational updates.
• It employs an ensemble of stochastic trajectories updated via empirical Kalman gains to reconcile model predictions with irregular measurement data.
• Rigorous convergence and stability results support its use in high-dimensional filtering, with extensions like multilevel and mollified variants enhancing efficiency.

Continuous-discrete ensemble Kalman filtering (EnKF) encompasses a class of algorithms for sequential Bayesian state estimation in partially observed dynamical systems, where the signal evolves in continuous time and observations are made at discrete or random (e.g., Poisson) times. In these methods, an ensemble of stochastic trajectories (“particles”) is propagated via the model dynamics in between observations and updated collectively using an empirical approximation of the Kalman gain upon the arrival of new measurement data. Continuous-discrete EnKF covers linear and nonlinear models, supports a range of discretization and assimilation schemes, and is a critical tool for high-dimensional filtering problems, with rigorously established convergence and stability properties under suitable conditions (Tanwani et al., 2024, Akyildiz et al., 2022, Hoel et al., 2015).

1. Mathematical Framework for Continuous-Discrete EnKF

The prototypical problem formulation consists of a stochastic differential equation (SDE) for the hidden signal $x_t$ on $[0,T]$,

$$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$

with $A \in \mathbb R^{n \times n}$, $Q \succcurlyeq 0$, and $W_t$ a standard $n$-dimensional Wiener process. Observations arrive at a sequence $\{t_k\}_{k\ge1}$ of discrete times—potentially random, such as the jump process of a Poisson process $N_t$ of rate $\lambda$, with

$[0,T]$0

where $[0,T]$1 and $[0,T]$2. The ensemble Kalman filter approximates the mean and covariance $[0,T]$3 of the filtering distribution, replacing the Riccati equation for $[0,T]$4 with an empirical estimate from a finite ensemble.

For general nonlinear $[0,T]$5 and discrete or continuous observations, the paradigm is similar, and the filter proceeds as a two-stage process:

1. Continuous-time propagation between observation times by integrating the SDE,
2. Discrete-time update at each $[0,T]$6 using empirical covariances and Kalman-type gain (Tanwani et al., 2024, Akyildiz et al., 2022, Bergemann et al., 2010).

2. Mean-Field Limit, Particle System, and Jump Updating

The infinite-ensemble (“mean-field”) limit is formulated as a McKean–Vlasov SDE with Poisson-driven jumps: $[0,T]$7 with mean-field covariance $[0,T]$8, mean $[0,T]$9, and gain

$$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$0

For a finite ensemble $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$1,

$$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$2

where $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$3 is computed from the empirical covariance.

The effect of the Poisson sampling is that updates occur only at jump times $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$4. Thus, between jumps, particles evolve according to the homogeneous SDE, and at each $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$5 are updated by

$$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$6

and the ensemble mean and covariance jump as

$$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$7

(Tanwani et al., 2024).

3. Discretization, Time-Stepping, and High-Dimensional Implementation

Discretization of the continuous propagation is crucial for high-dimensional or nonlinear systems. The predictor–corrector sequential Euler scheme allows for weak order 1.0 convergence and stable propagation with large step sizes:

• Predictor: $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$8
• Corrector: For each block $$\mathrm d x_t = A x_t \,\mathrm d t + Q^{1/2}\,\mathrm d W_t$$9, update iteratively with drift and noise (Akyildiz et al., 2022).

Ensemble propagation employs these schemes between observation times, followed by an analysis step at each $A \in \mathbb R^{n \times n}$0:

• Compute predicted mean/covariance,
• Calculate empirical cross-covariance and observation covariance,
• Compute Kalman gain and perform perturbed or transformed updates.

Numerical experiments in high-dimensional Lorenz-96 systems demonstrate that sequential schemes can use time steps $A \in \mathbb R^{n \times n}$1–$A \in \mathbb R^{n \times n}$2 larger than explicit Euler–Maruyama, yielding comparable or better estimation at lower computational cost by enabling smaller ensemble sizes or noisier systems (Akyildiz et al., 2022).

4. Theoretical Properties and Convergence Results

Under standard controllability and observability conditions and a lower bound on the observation sampling rate $A \in \mathbb R^{n \times n}$3, the empirical moments $A \in \mathbb R^{n \times n}$4 of the finite-ensemble system converge to those of the mean-field filter: $A \in \mathbb R^{n \times n}$5 with constants $A \in \mathbb R^{n \times n}$6 independent of $A \in \mathbb R^{n \times n}$7, uniformly on compact intervals in $A \in \mathbb R^{n \times n}$8 (Tanwani et al., 2024).

The main proof approaches involve:

• Propagation of chaos: empirical measure of the particle ensemble converges to the law of the limiting McKean–Vlasov process, with $A \in \mathbb R^{n \times n}$9 rates,
• Moment-closure and Lipschitz bounds: ensuring gains are globally Lipschitz and empirical/true covariances remain positive-definite under the Poisson rate lower bound,
• Gronwall-type estimates: controlling mean and covariance errors over time (Tanwani et al., 2024). A plausible implication is that, provided the ensemble size is sufficiently large and the data rate high enough, ensemble filters can approximate the optimal filter with error $Q \succcurlyeq 0$0 over arbitrary time horizons.

5. Extensions: Multilevel, Mollified, and Continuous-Time Limits

Multilevel EnKF

Multilevel Monte Carlo (MLMC) techniques introduce a hierarchy of time steps, reducing the total computational complexity for a given accuracy. The multilevel EnKF (MLEnKF) addresses bias due to time discretization and variance due to Monte Carlo error. By careful allocation of samples across levels and covariance matching (including eigen-truncation for positiveness), the MLEnKF can achieve RMSE $Q \succcurlyeq 0$1 in computational cost $Q \succcurlyeq 0$2, versus $Q \succcurlyeq 0$3 for classical EnKF (Hoel et al., 2015).

Mollified (Continuous) EnKF

To mitigate spurious high-frequency adjustment in sequential (instantaneous) EnKF analysis, mollified EnKF (MEnK) replaces Dirac deltas in the analysis step with compactly supported kernels, effectively spreading each update over a finite window. In stiff or multiscale dynamical systems, MEnK achieves lower RMS error and superior stability compared to standard or incremental analysis update (IAU) EnKF (Bergemann et al., 2010).

Continuous-Time Limits

For frequent and noisy observations, discrete-time EnKF updates converge in the limit to continuous-time Kalman–Bucy-type ensemble SDEs,

$Q \succcurlyeq 0$4

well-posedness and $Q \succcurlyeq 0$5 strong convergence (with $Q \succcurlyeq 0$6 the discrete assimilation step) are established under global Lipschitz and polynomial growth for model/observation nonlinearities (Lange, 2020, Kelly et al., 2013).

6. Practical Considerations and Implementation Insights

• For accurate covariance estimation, ensemble size must exceed the number of unstable modes; insufficient ensemble sizes risk rank-deficiency, corrected with variance inflation or localization (Kelly et al., 2013).
• Sequential predictor–corrector schemes enable stable ensemble propagation with much larger time steps ($Q \succcurlyeq 0$7) compared to Euler–Maruyama, delivering the same accuracy with fewer steps, especially in high-dimensional settings (Akyildiz et al., 2022).
• For random (e.g., Poisson) observation schedules, a lower bound on observation frequency is essential for filter stability and convergence (Tanwani et al., 2024).
• Mollified updates and continuous-time assimilations prevent the excitation of spurious modes in multiscale or stiff models without additional integrations (Bergemann et al., 2010).
• In MLMC-based filters, careful construction of coupled forecast steps and multilevel covariance estimation is needed to guarantee cost reduction without loss of accuracy (Hoel et al., 2015).

7. Summary Table: Continuous-Discrete EnKF Methodological Variants

Methodological Variant	Key Feature	Source arXiv id
Poisson-sampled ensemble KF	Mean-field SDE w/ Poisson update, $Q \succcurlyeq 0$8 convergence	(Tanwani et al., 2024)
Predictor-corrector SEnKF	Stable for large $Q \succcurlyeq 0$9; sequential block correction	(Akyildiz et al., 2022)
Multilevel EnKF (MLEnKF)	MLMC hierarchy for computational acceleration	(Hoel et al., 2015)
Mollified EnKF (MEnK)	Mollified (spread) analysis for balanced assimilation	(Bergemann et al., 2010)
Continuous-time EnKBF/ETKBF	Ensemble SDE limits, well-posed for GLipschitz nonlinearities	(Lange, 2020, Kelly et al., 2013)

Each variant addresses specific challenges in high-dimensional filtering, balancing accuracy, stability, cost, and dynamical consistency. The field’s rigorous convergence and stability results provide a foundation for applied and theoretical advances in real-time state estimation for stochastic dynamical systems.