Is the EnKF the same as the extended Kalman filter?

No. The extended filter typically propagates uncertainty through local linearization, while the EnKF uses an ensemble to estimate relevant statistics.

Is it a particle filter?

Both use samples, but ordinary EnKF analysis uses Kalman-style transformations rather than general likelihood weighting and resampling.

Must observations be perturbed?

No. Deterministic square-root variants provide another approach.

AI Glossary / Statistics and probability / Intermediate

Ensemble Kalman filter

Definition

The ensemble Kalman filter, or EnKF, is a method for updating estimates of a changing system by combining an ensemble of model forecasts with incoming observations. It uses the ensemble's sample statistics to approximate uncertainty and calculate a Kalman-style correction. It is especially useful for large state-estimation problems, but it does not generally produce an exact Bayesian posterior for nonlinear, non-Gaussian systems.

What it means

An ensemble contains multiple plausible versions of the system's state. Each version evolves through the forecast model. When a measurement arrives, the filter adjusts the ensemble according to the disagreement between predicted and observed measurements and the estimated uncertainties.
The ensemble is not merely a collection of unrelated predictions whose average is reported. Its spread and cross-variable relationships help determine how observations change the state estimate, including variables that are not directly observed. Incorrect ensemble relationships can therefore cause incorrect updates.

How it works

Initialize ensemble members to represent plausible uncertainty about the starting state. Propagate each member through the dynamic model, including a justified treatment of model or process error. The resulting forecast ensemble provides a mean and sample covariance.
Map the forecast states into observation space using the observation operator. For linear observations y = Hx plus measurement error with covariance R, a basic gain is K = P_f H transpose (H P_f H transpose + R) inverse, where P_f is the ensemble forecast covariance. In implementation, stable linear solves are preferable to explicitly forming an inverse.
The mean update uses the innovation, the observed value minus its forecast: mean_a = mean_f + K(y - H mean_f). Updating the ensemble spread requires additional care. Stochastic EnKF variants use appropriately perturbed observations. Deterministic square-root variants transform ensemble anomalies to achieve the intended analysis covariance without those random perturbations.
A finite ensemble can estimate only limited covariance structure. With m members, its unlocalized sample covariance has rank at most m - 1. Practical methods often use localization to reduce implausible distant correlations and inflation to counter insufficient spread, with choices evaluated for the application.
Repeat forecast and analysis steps as observations arrive. Check innovation behavior, ensemble spread, observation quality, numerical stability, and physical constraints. Nonlinearity, biased forecasts, unmodeled measurement errors, and multimodal uncertainty can undermine the approximation. More ensemble members reduce some sampling error but do not automatically fix model misspecification.

Three practical examples

A weather-assimilation system updates many atmospheric state variables using new sensor observations and an ensemble of forecasts.
An engineering team combines a process model with intermittent measurements to estimate internal conditions that cannot all be measured directly.
A water-management model updates estimates of reservoir or catchment conditions as new observations become available.

Think of it this way

Think of several plausible maps of a moving system. New measurements help revise the maps, and the differences among them describe uncertainty. If all maps share the same mistaken assumption, their agreement can still be misleading; a narrow ensemble is not automatically an accurate one.

Worked example

Consider an original scalar forecast ensemble with members 8, 10, and 12. Its mean is ten and its sample variance, using denominator three minus one, is (4 + 0 + 4)/2 = 4.
A direct observation is eleven, with assumed independent measurement-error variance one. The scalar gain is K = 4/(4 + 1) = 0.8. The analysis mean becomes 10 + 0.8(11 - 10) = 10.8, and the target Kalman analysis variance is (1 - 0.8)4 = 0.8.
For a deterministic scalar square-root update, multiply the forecast anomalies -2, 0, and 2 by the square root of 0.2, then add 10.8. The new members are approximately 9.9056, 10.8, and 11.6944. Their sample mean is 10.8 and sample variance is 0.8, as intended.
If every member instead received the same naive correction x + 0.8(11 - x), the resulting ensemble variance would be only 0.16. This shows why updating spread correctly matters. The example demonstrates one analysis step with specified assumptions, not a complete operational forecasting system.

Illustrative case study

Fictional case: Alder Water uses a simulator to estimate conditions across a catchment. Its first EnKF implementation becomes very confident after repeatedly assimilating measurements, yet later forecasts miss observed changes.
The team investigates sensor-error assumptions, shared model bias, ensemble initialization, and whether updates shrink spread too aggressively. It evaluates suitable inflation and localization settings against withheld observations rather than selecting them merely to make plots look smooth. The revised workflow tracks both forecast errors and ensemble uncertainty, while retaining checks for physically impossible states.

Common misconceptions and mistakes

Mistake: treating EnKF as exact nonlinear Bayesian filtering. Its Kalman-style update generally approximates the distribution.
Mistake: correcting every member toward an identical observation without the appropriate anomaly or perturbation treatment. This can underestimate uncertainty.
Mistake: assuming a small ensemble estimates a huge covariance matrix accurately. Sampling noise and rank limitations matter.
Mistake: treating inflation as a universal remedy. It cannot repair every structural bias or incorrect observation model.
Mistake: assuming measurements are perfect or independent when their errors are correlated.

Business use

Businesses can use EnKF methods when decisions depend on a changing, partly observed system represented by a useful simulator. Applications require domain knowledge about process dynamics, sensors, and uncertainty. Evaluate decision-relevant accuracy, reliability of uncertainty estimates, runtime, and failure behavior. An ensemble approach can make large estimation problems practical, but operational value depends on credible models and continuing diagnostic checks.

Check your understanding

Why is the corrected mean in the example closer to eleven than to ten?

Answer: The forecast variance exceeds the observation-error variance, giving the observation substantial weight.

Why is matching the updated mean insufficient?

Answer: An incorrect ensemble spread misrepresents uncertainty and affects later updates.

Related terms

Kalman filter
Data assimilation
Covariance
Localization
state estimation
ensemble forecasting
observation operator
inflation
and particle filter

Learn next

derive a scalar Kalman update, compare stochastic and deterministic ensemble updates, and inspect how ensemble size and observation-error assumptions affect forecast reliability.

Sources

Sources: Evensen, The Ensemble Kalman Filter: theoretical formulation and practical implementation, Ocean Dynamics (2003), DOI https://doi.org/10.1007/s10236-003-0036-9 ; accessible author conference treatment: https://www.ecmwf.int/sites/default/files/elibrary/2003/74424-ensemble-kalman-filter-theoretical-formulation-and-practical-implementation_0.pdf . Mandel, A Brief Tutorial on the Ensemble Kalman Filter: https://arxiv.org/pdf/0901.3725 . The scalar square-root example is original.

Previous: Ensemble Next: Ensemble learning