Continuous Ensemble Kalman Filtering

• Continuous ensemble Kalman filtering is a continuous-time formulation that assimilates observational data via a covariance-based gain to update state estimates.
• It leverages ensemble approximations, such as the EnKBF and ETKBF, to replace exact posterior covariances and ensure empirical accuracy.
• The methodology underpins advancements in stability analysis, multilevel acceleration, and parameter estimation, offering scalable solutions in nonlinear dynamics.

Continuous ensemble Kalman filtering denotes the continuous-time formulations of ensemble Kalman methods in which observational information is assimilated through an ongoing feedback correction rather than a purely discrete analysis step. In continuous-time filtering this family is usually expressed through ensemble Kalman-Bucy filters (EnKBFs), while related continuous limits also arise in constrained ensemble Kalman inversion and other control-to-observable settings. Across these formulations, the defining mechanism is a covariance-based gain multiplying an innovation or residual term, with the gain estimated from the ensemble itself rather than from an exact posterior covariance (Taghvaei et al., 2017, Lange et al., 2019, Herty et al., 2019).

1. Canonical formulation

The continuous-time filtering problem is commonly posed through a signal-observation system on $\mathbb R^d$,

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$

with the objective of computing the posterior law of $X_t$ conditional on the observation $\sigma$-algebra $\mathcal Z_t=\sigma\{Z_s:s\le t\}$. In the linear-Gaussian case $a(x)=Ax$, $h(x)=Hx$, the exact solution is the Kalman-Bucy filter; continuous ensemble methods replace the exact posterior covariance by an empirical covariance computed from particles (Taghvaei et al., 2017).

A standard continuous-time ensemble approximation maintains particles $\{X_t^{(i)}\}_{i=1}^M$, ensemble mean $\bar X_t=\frac1M\sum_i X_t^{(i)}$, empirical covariance

$$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$

and centered observation matrix

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$0

With gain

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$1

the EnKBF SDE takes the form

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$2

A deterministic transform variant, the ETKBF, replaces perturbed observations by

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$3

and this transform form preserves exactly the empirical covariance (Lange, 2020).

The common structural feature is the familiar “gain $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$4 innovation” update. In the linear-Gaussian case one may write the stochastic EnKBF as

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$5

or, in deterministic form,

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$6

which makes explicit the continuous distribution of the update in time (Taghvaei et al., 2017).

2. Continuous-time limits of discrete ensemble Kalman methods

A major strand of the literature derives continuous ensemble Kalman filtering as the small-step limit of discrete-time forecast-analysis recursions. In one rigorous construction, a discrete EnKF is applied to Euler-Maruyama discretizations of continuous signal and observation SDEs, and the resulting time-continuous interpolation converges, as the step size $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$7, to an interacting SDE system in ensemble-mean-square. For the modified deterministic square-root EnKF, one obtains

$$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$8

that is, a uniform $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$9 rate (Lange et al., 2019).

A complementary derivation accommodates nonlinear and unbounded model and observation operators. Under globally Lipschitz and growth assumptions on $X_t$0 and $X_t$1, and if initially

$X_t$2

the piecewise-interpolated discrete EnKF converges strongly to the EnKBF solution: $X_t$3 and an identical statement holds for ESRF $X_t$4 ETKBF (Lange, 2020).

A different continuous limit arises in inverse problems with equality constraints. Starting from a constrained quadratic minimization step and rescaling $X_t$5 together with $X_t$6, the $X_t$7 limit yields the semi-explicit DAE

$X_t$8

For linear forward maps this becomes

$X_t$9

namely a preconditioned gradient flow with algebraic coupling to the constraints (Herty et al., 2019).

3. Mean-field, McKean-Vlasov, and transport descriptions

In linear continuous-time filtering, EnKBFs admit a natural McKean-Vlasov interpretation. The Kalman-Bucy filter can be represented as the conditional law of a diffusion

$\sigma$0

whose conditional law is $\sigma$1. The interacting EnKBF particle system

$\sigma$2

is then viewed as a particle approximation of this McKean-Vlasov dynamics (Chada et al., 2020).

For deterministic continuous-time EnKBFs, the large-ensemble limit can also be formulated through a nonlinear McKean-Vlasov SDE and a corresponding nonlinear Fokker-Planck equation. In the fully observed setting analyzed by de Wiljes, Reich, and Stannat, the empirical law converges to the law $\sigma$3 of a single copy $\sigma$4 satisfying

$\sigma$5

with $\sigma$6, $\sigma$7, and $\sigma$8 (Wiljes et al., 2016).

In constrained inverse problems the mean-field limit takes a Vlasov-type transport form. If

$\sigma$9

then one formally obtains

$\mathcal Z_t=\sigma\{Z_s:s\le t\}$0

together with the moment constraint

$\mathcal Z_t=\sigma\{Z_s:s\le t\}$1

Dirac measures $\mathcal Z_t=\sigma\{Z_s:s\le t\}$2 are stationary precisely when $\mathcal Z_t=\sigma\{Z_s:s\le t\}$3 satisfies the KKT conditions of the constrained least-squares problem (Herty et al., 2019).

A recent mean-field analysis in the linear detectable setting derives the MFEnKF from a variational approximation of the Bayes update using a covariance-weighted optimal transport metric. The resulting McKean-Vlasov SDE is

$\mathcal Z_t=\sigma\{Z_s:s\le t\}$4

with $\mathcal Z_t=\sigma\{Z_s:s\le t\}$5. This formulation is coupled to a measure-level weak equation for $\mathcal Z_t=\sigma\{Z_s:s\le t\}$6 and is used to prove contraction toward the Gaussian manifold (Hoffmann et al., 5 Jun 2026).

4. Well-posedness, stability, and accuracy

The continuous-time EnKBF and ETKBF with nonlinear unbounded coefficients are well posed under global Lipschitz and linear-growth assumptions on $\mathcal Z_t=\sigma\{Z_s:s\le t\}$7 and $\mathcal Z_t=\sigma\{Z_s:s\le t\}$8. Introducing explosion times

$\mathcal Z_t=\sigma\{Z_s:s\le t\}$9

one has unique strong solutions on $a(x)=Ax$0 together with moment bounds

$a(x)=Ax$1

The same analysis yields mean-square accuracy estimates relative to a reference trajectory, uniform up to any fixed finite horizon $a(x)=Ax$2 (Lange, 2020).

For constrained continuous EnKF dynamics, standard DAE-index theory applies. Under smoothness of $a(x)=Ax$3, full rank of $a(x)=Ax$4, and bounded inverse of $a(x)=Ax$5, the constrained system has differentiation-index $a(x)=Ax$6 and perturbation-index $a(x)=Ax$7. A locally unique solution $a(x)=Ax$8 exists for short time, and small perturbations in the data produce $a(x)=Ax$9 changes only. A Lyapunov-type argument further shows that any steady state satisfies the KKT conditions of the constrained minimization problem; when the constraint set is convex, this guarantees optimality (Herty et al., 2019).

Long-time results are strongest in special regimes. For the deterministic EnKBF with a fully observed process $h(x)=Hx$0, full-rank diffusion $h(x)=Hx$1, and small measurement noise $h(x)=Hx$2, one obtains uniform spectral bounds

$h(x)=Hx$3

for all $h(x)=Hx$4, provided the initial covariance satisfies $h(x)=Hx$5. In the same setting, the estimation error $h(x)=Hx$6 satisfies

$h(x)=Hx$7

with $h(x)=Hx$8, so asymptotically $h(x)=Hx$9 uniformly in $\{X_t^{(i)}\}_{i=1}^M$0 (Wiljes et al., 2016).

At mean-field level, the linear detectable setting supports pathwise asymptotic results. There exists a continuous map

$\{X_t^{(i)}\}_{i=1}^M$1

and a uniform $\{X_t^{(i)}\}_{i=1}^M$2 such that

$\{X_t^{(i)}\}_{i=1}^M$3

which is a strict contraction onto the Gaussian subspace. In addition, for bounded uniformly continuous test functions, the MFEnKF and the optimal Bayesian filter have coincident large-time moments for almost every observation path (Hoffmann et al., 5 Jun 2026).

5. Major variants and generalized settings

The continuous-time ensemble Kalman literature contains several distinct but structurally related formulations (Taghvaei et al., 2017, Lange, 2020, Nüsken et al., 2019, Ertel et al., 2022, Herty et al., 2019).

Variant	Representative law	Distinguishing point
Stochastic EnKBF	$\{X_t^{(i)}\}_{i=1}^M$4	Perturbed observations
Deterministic / ETKBF	$\{X_t^{(i)}\}_{i=1}^M$5	No observation perturbation; exact empirical covariance
State-parameter EnKBF	$\{X_t^{(i)}\}_{i=1}^M$6 with gains from $\{X_t^{(i)}\}_{i=1}^M$7	Joint state and parameter estimation
Correlated-noise EnKBF	Gain uses $\{X_t^{(i)}\}_{i=1}^M$8 and $\{X_t^{(i)}\}_{i=1}^M$9	Shared signal/observation noise
Constrained continuous EnKF	DAE with $\bar X_t=\frac1M\sum_i X_t^{(i)}$0 and multipliers $\bar X_t=\frac1M\sum_i X_t^{(i)}$1	Equality constraints

A generalized continuous-time state-parameter EnKBF arises when noisy signal increments are observed and model and measurement errors are correlated. With state $\bar X_t=\frac1M\sum_i X_t^{(i)}$2, parameter $\bar X_t=\frac1M\sum_i X_t^{(i)}$3, and observation increments

$\bar X_t=\frac1M\sum_i X_t^{(i)}$4

the constant-gain approximation uses empirical covariances

$\bar X_t=\frac1M\sum_i X_t^{(i)}$5

and, to first order,

$\bar X_t=\frac1M\sum_i X_t^{(i)}$6

Writing parameters as zero-dynamics variables $\bar X_t=\frac1M\sum_i X_t^{(i)}$7 yields a continuous-time ensemble Kalman-Bucy scheme for simultaneous state and parameter estimation (Nüsken et al., 2019).

Correlated observation noise leads to a different extension. In that setting, the ensemble dynamics depend on the Moore-Penrose pseudoinverse of the empirical covariance matrix,

$\bar X_t=\frac1M\sum_i X_t^{(i)}$8

because the covariance adjustment contains the factor

$\bar X_t=\frac1M\sum_i X_t^{(i)}$9

The analysis provides existence and uniqueness for both the mean-field equation and the particle system, together with convergence to the mean-field limit (Ertel et al., 2022).

In constrained inverse problems, nonlinear equality constraints are enforced at the level of each ensemble member through Lagrange multipliers, whereas linear constraints are automatically preserved by the unconstrained continuous EnKF flow. This distinction is central to the DAE/KKT formulation of constrained continuous ensemble Kalman dynamics (Herty et al., 2019).

6. Computation, multilevel acceleration, and applications

Passing from discrete updates to continuous-time dynamics has direct algorithmic implications. In constrained inverse problems, the discrete Kalman-gain step is replaced by an ODE/DAE that can be solved by standard stiff integrators such as IMEX-RK and BDF, and the continuous-time form permits adaptive time stepping. The corresponding mean-field PDE suggests particle-system approximations, stochastic-particle splitting algorithms, and direct discretization in function space for very large ensembles (Herty et al., 2019).

For linear continuous-time filtering, multilevel Monte Carlo substantially changes the cost-accuracy trade-off. An ideal multilevel EnKBF achieves mean square error $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$0 with cost $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$1, whereas a single-level EnKBF requires cost $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$2 for the same $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$3 MSE. The theory is tested on a linear problem motivated by dimensions of order $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$4 and $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$5, and the numerical slopes in log-log MSE-versus-cost plots are approximately $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$6 for single-level and $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$7 up to logarithmic terms for multilevel (Chada et al., 2020).

Randomized multilevel constructions also yield unbiased estimators. In the linear-Gaussian setting, unbiased estimation for vanilla and deterministic EnKBFs combines randomization in the time discretization level and in the number of particles. Under standard strong and weak error rates, the resulting estimator has finite variance and, using $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$8 i.i.d. replicates, achieves MSE $$P_t=\frac1{M-1}\sum_{i=1}^M (X_t^{(i)}-\bar X_t)(X_t^{(i)}-\bar X_t)^\top,$$9 at cost $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$00 for any $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$01 (Alvarez et al., 2022).

Applications span state estimation, parameter identification, and control. Continuous-time state-parameter EnKBFs have been demonstrated on an Ornstein-Uhlenbeck model with unknown drift, averaging and homogenisation examples, a nonparametric drift reconstruction with dimension $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$02, and parameter estimation for a stochastic heat equation (Nüsken et al., 2019). Long-time deterministic EnKBF analysis has been verified numerically on the stochastic Lorenz-63 system, where the time-averaged root-mean-square error and the largest covariance eigenvalue scale like $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$03 over the tested range of measurement noise variances (Wiljes et al., 2016). In control, an EnKBF-based nonlinear model predictive control scheme propagates the forward state distribution with continuous-time EnKBF equations and approximates receding-horizon control laws; on an inverted pendulum with $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$04, $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$05, and $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$06, $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$07 Monte Carlo runs demonstrate stabilization around the unstable equilibrium $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$08 (Reich, 16 Mar 2025).

7. Scope, limitations, and related methodologies

A recurrent point in the literature is that continuous ensemble Kalman filtering is not, in general, an exact nonlinear Bayesian filter. The EnKBF is described as an approximate extension of the Kalman-Bucy filter to nonlinear problems, whereas the feedback particle filter is presented as a further extension that provides a consistent solution in the general nonlinear, non-Gaussian case (Taghvaei et al., 2017). In the nonlinear setting, the ensemble method is therefore often analyzed as a randomized state estimator or as a mean-field interacting particle system rather than as an exact posterior solver (Bishop et al., 2020).

Relative to sequential Monte Carlo methods, EnKBFs avoid importance weights and resampling. The update is implemented through feedback control, which potentially leads to smaller simulation variance and improved stability properties (Taghvaei et al., 2017). In the linear-Gaussian review literature this difference is sharpened by a negative result for the bootstrap particle filter: if $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$09 in an unstable discrete-time linear signal model, then for any finite $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$10 the particle-filter estimate loses the signal, whereas the ensemble Kalman filter and the optimal filter retain tracking capability. The same review also distinguishes heavy-tailed covariance behavior in the vanilla continuous-time EnKF from the lighter-tailed behavior of deterministic or square-root variants (Bishop et al., 2020).

The adjective “continuous” itself has two distinct uses in adjacent literatures. Most commonly in ensemble Kalman-Bucy work it means continuous time. In a different line of research, however, “continuous” refers to a continuous transition indexed by $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$11 between the ensemble and particle filter update: $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$12 In that construction, $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$13 recovers the particle filter and $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$14 recovers the EnKF, with $$dX_t = a(X_t)\,dt + \sigma(X_t)\,dB_t,\qquad dZ_t = h(X_t)\,dt + dW_t,$$15 chosen to balance bias against weight degeneracy (Frei et al., 2012). This terminological distinction is important because it separates continuous-time ensemble filtering from continuous interpolation between update mechanisms.