Bridging Kalman Filter

• Bridging Kalman Filter is a Monte Carlo filtering method that interpolates between the ensemble Kalman filter and particle filter using a scalar tempering parameter.
• It performs a two-stage update: a tempered EnKF-style update followed by a particle filter correction to balance computational feasibility and non-Gaussian representation.
• Adaptive tuning via effective sample size ensures improved accuracy and reduced errors in challenging high-dimensional, nonlinear data assimilation problems.

A Bridging Kalman Filter refers to a class of Monte Carlo filtering algorithms that interpolate continuously between the ensemble Kalman filter (EnKF) and the bootstrap particle filter (PF) by means of a scalar “bridging” parameter, typically denoted as $\gamma\in[0,1]$. The methodology, introduced by Frei and Künsch, is designed to address the dual challenges encountered in nonlinear, high-dimensional sequential data assimilation: the sample degeneracy intrinsic to particle filters and the Gaussian bias inherent to ensemble Kalman filters. By tempering the update step, such filters construct an adaptive transition—recovering classical EnKF for $\gamma=1$ and the PF for $\gamma=0$—thus leveraging the computational feasibility and sample-diversity of EnKF while partially retaining the non-Gaussian representational capacity of PFs (Frei et al., 2012).

1. Motivation: Limitations of Particle and Ensemble Kalman Filters

Monte Carlo filtering for nonlinear, high-dimensional state-space models is dominated by two approaches:

• Bootstrap Particle Filter (PF): Employs Bayes’ rule by weighting finite forecast samples by likelihoods and then resampling. In high dimensions or with “sharp” data, weight degeneracy arises—most weights collapse to zero except for a negligible fraction, degrading the effective sample size.
• Ensemble Kalman Filter (EnKF): Applies an affine, Gaussian-based update to each forecast particle, maintaining sample spread even for small ensembles. However, EnKF is only optimal with Gaussian priors and linear observation operators; under nonlinear or non-Gaussian conditions, it exhibits systematic bias, failing to represent multimodality or skewness.

A bridging strategy aims to combine the degeneracy-resistance of EnKFs with the non-Gaussian fidelity of PFs, balancing computational tractability with expressivity (Frei et al., 2012).

2. Classical Algorithms: PF and EnKF Update Steps

Both PF and EnKF operate on forecast ensembles $\{x_i^f\}_{i=1}^N$ for state $x$, given observation $y$ with likelihood $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$ (with $H$ linear and $R$ covariance):

• PF update: Compute normalized weights $\omega_i = \ell(x_i^f)/\sum_{j}\ell(x_j^f)$, form posterior empirical measure $\gamma=1$0; resample to yield $\gamma=1$1.
• EnKF update: Estimate forecast covariance $\gamma=1$2, calculate Kalman gain $\gamma=1$3. Each member is updated by $\gamma=1$4, $\gamma=1$5. This produces a balanced mixture of Gaussians centered at $\gamma=1$6 with covariance $\gamma=1$7.

EnKF’s balanced mixture can be interpreted as sampling from $\gamma=1$8, whereas the PF approximates the exact (potentially non-Gaussian) posterior (Frei et al., 2012).

3. The $\gamma=1$9-Indexed Bridging Filter: Algorithmic Structure

The Bridging Kalman Filter introduces a “tempering” parameter $\gamma=0$0 and performs a two-stage Bayesian update:

Posterior decomposition: $\gamma=0$1

• First stage: Apply an EnKF-style update using the tempered likelihood $\gamma=0$2. The effective observation noise becomes $\gamma=0$3.
  • Kalman gain: $\gamma=0$4
  • Updated intermediate means: $\gamma=0$5
  • Intermediate covariance: $\gamma=0$6
  • The ensemble after this stage is the mixture $\gamma=0$7
• Second stage: Apply a PF-style correction using the remaining likelihood power $\gamma=0$8 to each Gaussian in the mixture.
  • Weights:

  $\gamma=0$9 - Further update each mean via an EnKF step, with $\{x_i^f\}_{i=1}^N$0, where $\{x_i^f\}_{i=1}^N$1 - Common covariance: $\{x_i^f\}_{i=1}^N$2 - Resample indices $\{x_i^f\}_{i=1}^N$3 and draw $\{x_i^f\}_{i=1}^N$4

As $\{x_i^f\}_{i=1}^N$5, the procedure reproduces the EnKF; as $\{x_i^f\}_{i=1}^N$6, it is equivalent to the classical PF (Frei et al., 2012).

4. Adaptive Selection of the Bridging Parameter $\{x_i^f\}_{i=1}^N$7

Choosing $\{x_i^f\}_{i=1}^N$8 is crucial: small $\{x_i^f\}_{i=1}^N$9 incurs PF degeneracy, large $x$0 induces EnKF bias. The effective sample size (ESS)

$x$1

is used as a degeneracy metric. The standard criterion is to select the smallest $x$2 such that $x$3 for some threshold $x$4 (e.g., $x$5). Grid or binary search is typically used; one may optionally monitor ensemble spread to further tune $x$6 (Frei et al., 2012).

5. Computational Complexity and Scaling

The computational cost is dominated by linear system solves for the Kalman gain ($x$7 for dense matrices, $x$8 in ensemble form). Covariance localization or tapering is essential for high-dimensional problems, reducing to $x$9 with sparse structures. The Bridging Kalman Filter adds only a small overhead compared to the EnKF, arising from an additional gain computation and ESS evaluation. Unlike PF, the bridging filter is stable and maintains diversity for moderate $y$0 in high-dimensional settings (Frei et al., 2012).

6. Numerical Performance and Empirical Results

On the 40-dimensional Lorenz-96 model (with $y$1), the Bridging Kalman Filter attains a root-mean-square error (RMSE) of $y$2 (target ESS $y$3), versus $y$4 for the EnKF—a $y$5 improvement. The continuous ranked probability score (CRPS) on unobserved variables improves from $y$6 (EnKF) to $y$7 (bridged). The adaptively selected $y$8 typically lies in $y$9–$\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$0, preserving $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$1–$\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$2 ensemble diversity.

In the Korteweg–de Vries PDE problem ($\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$3 grid points, $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$4), the EnKF produces physically implausible samples under strong nonlinearity and small $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$5, while the bridging filter maintains dynamical consistency. Notably, even a PF with $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$6 is unable to match the bridging filter’s statistical realism at comparable computational cost (Frei et al., 2012).

7. The Bridging Filter in the Context of Hybrid and Homotopy Approaches

Related hybrid filters have been proposed using likelihood splitting and multi-stage updates with parameters analogous to $\ell(x) = p(y|x) \propto \varphi(y; Hx, R)$7 (e.g., in the SIR–ESRF hybrid filter) (Robinson et al., 2020); these variants differ in the order and nature of updates, but share the bridging philosophy—controlled interpolation between PF and EnKF extremities. Continuous-time dynamic homotopy-based filters (e.g., Schrödinger-bridge filters (Reich, 2022, Bao et al., 22 Dec 2025)) generalize the idea, constructing a path of distributions interpolating between the prior and posterior, with similar aims: controlling the trade-off between sample collapse and linear-Gaussian bias.

---

References

• "Bridging the ensemble Kalman and particle filter" (Frei et al., 2012)
• "A hybrid particle-ensemble Kalman filter for problems with medium nonlinearity" (Robinson et al., 2020)
• "Data assimilation: A dynamic homotopy-based coupling approach" (Reich, 2022)
• "The Ensemble Schrödinger Bridge filter for Nonlinear Data Assimilation" (Bao et al., 22 Dec 2025)