---
title: Continuous Ensemble Kalman Filtering
url: https://www.emergentmind.com/topics/continuous-ensemble-kalman-filtering
type: topic
---

# Continuous Ensemble Kalman Filtering

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 [1702.07241][1901.05204][1912.08406].

## 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 [1702.07241].

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
\[
G_t=\bigl[g(X_t^{(1)})-\bar g_t,\dots,g(X_t^{(M)})-\bar g_t\bigr],\qquad
\bar g_t=\frac1M\sum_{j=1}^M g(X_t^{(j)}).
\]
With gain
\[
K_t=\frac1{M-1}\,P_t\,G_t^\top\,C^{-1},
\]
the EnKBF SDE takes the form
\[
dX_t^{(i)}
= f(X_t^{(i)})\,dt
+ Q\,dW_t^{(i)}
+ K_t\Bigl(dY_t + C\,dV_t^{(i)} - g(X_t^{(i)})\,dt\Bigr).
\]
A deterministic transform variant, the ETKBF, replaces perturbed observations by
\[
dX_t^{(i)}
= f(X_t^{(i)})\,dt
+ Q\,dW_t^{(i)}
+ K_t\bigl(dY_t - \bar g_t\,dt\bigr)
-\tfrac12\,K_t\bigl(g(X_t^{(i)})-\bar g_t\bigr)\,dt,
\]
and this transform form preserves exactly the empirical covariance [2012.07572].

The common structural feature is the familiar “gain \(\times\) innovation” update. In the linear-Gaussian case one may write the stochastic EnKBF as
\[
dX_t^i
= A\,X_t^i\,dt+\sigma\,dB_t^i
+\Sigma_t^{(N)}H^\top\bigl(dZ_t-HX_t^i\,dt+dW_t^i\bigr),
\]
or, in deterministic form,
\[
dX_t^i
= A X_t^i dt + \sigma\,dB_t^i
+\Sigma_t^{(N)}H^T\Bigl(dZ_t-\tfrac12(HX_t^i+H\hat X_t^{(N)})dt\Bigr),
\]
which makes explicit the continuous distribution of the update in time [1702.07241].

## 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 \(h\to0\), to an interacting SDE system in ensemble-mean-square. For the modified deterministic square-root EnKF, one obtains
\[
\mathbb E\Bigl[\sup_{0\le t\le T}\|X_t^{(i),h}-X_t^{(i)}\|_{\ell^2,M}^2\Bigr]
\le C'(T)\,h,
\]
that is, a uniform \(O(h)\) rate [1901.05204].

A complementary derivation accommodates nonlinear and unbounded model and observation operators. Under globally Lipschitz and growth assumptions on \(f\) and \(g\), and if initially
\[
\frac1M\sum_{i=1}^M \mathbb E\lvert X_0^{(i),f}-X_0^{(i)}\rvert^2=O(h),
\]
the piecewise-interpolated discrete EnKF converges strongly to the EnKBF solution:
\[
\frac1M\sum_{i=1}^M
\mathbb E\Bigl[\sup_{0\le t\le T}\lvert X_t^{(i),a}-X_t^{(i)}\rvert^2\Bigr]
=O(h),
\]
and an identical statement holds for ESRF \(\to\) ETKBF [2012.07572].

A different continuous limit arises in inverse problems with equality constraints. Starting from a constrained quadratic minimization step and rescaling \(\Gamma^{-1}\mapsto \Delta t\,\Gamma^{-1}\) together with \(\lambda^{j,n+1}\mapsto \Delta t^{-1}\lambda^{j,n+1}\), the \(\Delta t\to0\) limit yields the semi-explicit DAE
\[
\frac{d}{dt}u^j(t)
= C_{uw}(t)\,\Gamma\,(y-G(u^j))
- C_{uu}(t)\,J_{\!\mathcal A}(u^j)^\top\,\lambda^j(t),
\qquad
0=\mathcal A(u^j(t)).
\]
For linear forward maps this becomes
\[
\frac{d}{dt}u^j
= -\,C_{uu}\bigl[\nabla_u\Phi(u^j;y)+J_{\!\mathcal A}(u^j)^\top\lambda^j\bigr],
\qquad
0=\mathcal A(u^j),
\]
namely a preconditioned gradient flow with algebraic coupling to the constraints [1912.08406].

## 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
\[
d\bar X_t
= A\,\bar X_t\,dt
+Q^{1/2}\,d\bar W_t
+\mathcal P_tH^\top R^{-1}\bigl(dY_t-[H\bar X_t\,dt+R^{1/2}d\bar V_t]\bigr),
\]
whose conditional law is \(\mathcal N(\mathcal M_t,\mathcal P_t)\). The interacting EnKBF particle system
\[
dX_t^{i,N}
= A\,X_t^{i,N}\,dt
+Q^{1/2}\,dW_t^i
+K_t^N\bigl(dY_t-HX_t^{i,N}\,dt\bigr),
\qquad
K_t^N=P_t^N H^\top R^{-1},
\]
is then viewed as a particle approximation of this McKean-Vlasov dynamics [2011.04342].

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 \(\hat\pi_t\) of a single copy \(\hat X_t\) satisfying
\[
d\hat X_t
= f(\hat X_t)\,dt
+ D\,\mathcal P_t^{-1}(\hat X_t-\bar x_t)\,dt
-\tfrac12\,\mathcal Q_t R^{-1}\bigl(h(\hat X_t)+\bar h_t-2\,dY_t\bigr),
\]
with \(\bar x_t=\hat\pi_t[x]\), \(\mathcal P_t=\mathrm{Cov}_{\hat\pi_t}(\hat X_t)\), and \(\mathcal Q_t=\mathrm{Cov}_{\hat\pi_t}(\hat X_t,h(\hat X_t))\) [1612.06065].

In constrained inverse problems the mean-field limit takes a Vlasov-type transport form. If
\[
f^J(t,u,\lambda)=\frac1J\sum_{j=1}^J \delta(u^j(t)-u)\,\delta(\lambda^j(t)-\lambda)
\longrightarrow f(t,u,\lambda),
\]
then one formally obtains
\[
\partial_t f
-\nabla_u\!\cdot \Bigl\{
C[f](t)\,\bigl[\nabla_u\Phi(u;y)+J_{\!\mathcal A}(u)^\top\lambda\bigr]\,f
\Bigr\}=0,
\]
together with the moment constraint
\[
\iint \mathcal A(u)\,f(t,u,\lambda)\,du\,d\lambda = 0.
\]
Dirac measures \(\delta(u-u^*)\delta(\lambda-\lambda^*)\) are stationary precisely when \((u^*,\lambda^*)\) satisfies the KKT conditions of the constrained least-squares problem [1912.08406].

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
\[
du_t = F\,u_t\,dt + \Sigma^{1/2}\,d\beta_t + K_t\bigl(dZ_t-Hu_t\,dt\bigr),
\qquad
K_t=P_tH^\top R^{-1},
\]
with \(P_t=\mathrm{Cov}_{\mu_t}(u)\). This formulation is coupled to a measure-level weak equation for \(\mu_t\) and is used to prove contraction toward the Gaussian manifold [2606.07729].

## 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 \(f\) and \(g\). Introducing explosion times
\[
\tau_n=\inf\{t\ge0:\;\mathrm{tr}(P_t)>n\},
\]
one has unique strong solutions on \([0,\tau_n)\) together with moment bounds
\[
\sup_{0\le t\le T\wedge\tau_n}\mathbb E[\mathrm{tr}(P_t)^p]<\infty,
\qquad
\lim_{n\to\infty}\mathbb P\{\tau_n\le T\}=0.
\]
The same analysis yields mean-square accuracy estimates relative to a reference trajectory, uniform up to any fixed finite horizon \(T\) [2012.07572].

For constrained continuous EnKF dynamics, standard DAE-index theory applies. Under smoothness of \(\mathcal A\), full rank of \(J_{\!\mathcal A}(u)\), and bounded inverse of \(\partial_\lambda \mathcal A\), the constrained system has differentiation-index \(1\) and perturbation-index \(1\). A locally unique solution \((u^j(t),\lambda^j(t))\) exists for short time, and small perturbations in the data produce \(O(1)\) 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 [1912.08406].

Long-time results are strongest in special regimes. For the deterministic EnKBF with a fully observed process \(h(x)=x\), full-rank diffusion \(D\), and small measurement noise \(R=\varepsilon I\), one obtains uniform spectral bounds
\[
C_2\sqrt\varepsilon
\le \lambda^{\min}(P_t^M)
\le \lambda^{\max}(P_t^M)
\le C_1\sqrt\varepsilon
\]
for all \(t\), provided the initial covariance satisfies \(\lambda^{\min}(P_0^M)\ge C_2\sqrt\varepsilon\). In the same setting, the estimation error \(e_t=X_t^{\rm ref}-\bar x_t^M\) satisfies
\[
\mathbb E[\|e_t\|^2]
\le e^{-\alpha t}\,\mathbb E[\|e_0\|^2]+\alpha^{-1}K,
\]
with \(\alpha^{-1}=O(\sqrt\varepsilon)\), so asymptotically \(\mathbb E[\|e_t\|^2]=O(\sqrt\varepsilon)\) uniformly in \(t\) [1612.06065].

At mean-field level, the linear detectable setting supports pathwise asymptotic results. There exists a continuous map
\[
\beta_c(P)=\lambda_{\min}\!\bigl(P^{1/2}H^\top R^{-1}H\,P^{1/2}+P^{-1}\bigr)
\]
and a uniform \(\beta_0>0\) such that
\[
W_{2}\bigl(\mu_t,G\mu_t;P_t\bigr)
\le
\exp\!\Bigl(-\tfrac12\!\int_0^t \beta_c(P_s)\,ds\Bigr)\,
W_{2}\bigl(\mu_0,G\mu_0;P_0\bigr)
\le
e^{-\tfrac12\beta_0 t}\,
W_{2}\bigl(\mu_0,G\mu_0;P_0\bigr),
\]
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 [2606.07729].

## 5. Major variants and generalized settings

The continuous-time ensemble Kalman literature contains several distinct but structurally related formulations [1702.07241][2012.07572][1903.10717][2205.14253][1912.08406].

| Variant | Representative law | Distinguishing point |
|---|---|---|
| Stochastic EnKBF | \(K_t(dY_t + C\,dV_t^{(i)} - g(X_t^{(i)})dt)\) | Perturbed observations |
| Deterministic / ETKBF | \(K_t(dY_t-\bar g_tdt)-\tfrac12K_t(g(X_t^{(i)})-\bar g_t)dt\) | No observation perturbation; exact empirical covariance |
| State-parameter EnKBF | \(da_t=0\) with gains from \(P_t^{xh},P_t^{ah}\) | Joint state and parameter estimation |
| Correlated-noise EnKBF | Gain uses \(P_t^M H_t^\top+\tilde C_tT_t^\top\) and \((P_t^M)^+\) | Shared signal/observation noise |
| Constrained continuous EnKF | DAE with \(0=\mathcal A(u^j)\) and multipliers \(\lambda^j\) | 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 \(X_t\), parameter \(a_t\), and observation increments
\[
dY_t = H\,f(X_t,a)\,dt + H\,G\,dW_t + R^{1/2}\,dV_t,
\]
the constant-gain approximation uses empirical covariances
\[
P_t^{xh},\qquad P_t^{ah},\qquad P_t^{hh},
\]
and, to first order,
\[
K_t \simeq \bigl(P_t^{xh}+QH^\top,\;P_t^{ah}\bigr)C^{-1}.
\]
Writing parameters as zero-dynamics variables \(da_t=0\) yields a continuous-time ensemble Kalman-Bucy scheme for simultaneous state and parameter estimation [1903.10717].

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,
\[
(P_t^M)^+,
\]
because the covariance adjustment contains the factor
\[
-\,(P_t^M H_t^\top+\tilde C_tT_t^\top)\,R_t^{-1}\,(H_tP_t^M+T_tC_t)\,(P_t^M)^+.
\]
The analysis provides existence and uniqueness for both the mean-field equation and the particle system, together with convergence to the mean-field limit [2205.14253].

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 [1912.08406].

## 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 [1912.08406].

For linear continuous-time filtering, multilevel Monte Carlo substantially changes the cost-accuracy trade-off. An ideal multilevel EnKBF achieves mean square error \(O(\varepsilon^2)\) with cost \(O(\varepsilon^{-2}\log(\varepsilon)^2)\), whereas a single-level EnKBF requires cost \(O(\varepsilon^{-3})\) for the same \(O(\varepsilon^2)\) MSE. The theory is tested on a linear problem motivated by dimensions of order \(\sim O(10^4)\) and \(\sim O(10^5)\), and the numerical slopes in log-log MSE-versus-cost plots are approximately \(-3\) for single-level and \(-2\) up to logarithmic terms for multilevel [2011.04342].

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 \(M=O(\varepsilon^{-2})\) i.i.d. replicates, achieves MSE \(O(\varepsilon^2)\) at cost \(O(\varepsilon^{-2}|\log\varepsilon|^{2+\delta})\) for any \(\delta>0\) [2208.03947].

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 \(200\), and parameter estimation for a stochastic heat equation [1903.10717]. 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 \(O(\varepsilon^{1/2})\) over the tested range of measurement noise variances [1612.06065]. 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 \(T=2\), \(\Delta t=10^{-3}\), and \(M=K=50\), \(5000\) Monte Carlo runs demonstrate stabilization around the unstable equilibrium \((0,0)\) [2503.12474].

## 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 [1702.07241]. 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 [2006.08843].

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 [1702.07241]. In the linear-Gaussian review literature this difference is sharpened by a negative result for the bootstrap particle filter: if \(|A|>1\) in an unstable discrete-time linear signal model, then for any finite \(N\) 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 [2006.08843].

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 \(\gamma\in[0,1]\) between the ensemble and particle filter update:
\[
\pi^{u,\gamma}(dx)\propto \pi^p(dx)\,\ell(x\mid y)^\gamma.
\]
In that construction, \(\gamma\to0\) recovers the particle filter and \(\gamma\to1\) recovers the EnKF, with \(\gamma\) chosen to balance bias against weight degeneracy [1208.0463]. This terminological distinction is important because it separates continuous-time ensemble filtering from continuous interpolation between update mechanisms.

Source: https://www.emergentmind.com/topics/continuous-ensemble-kalman-filtering