A low-rank ensemble Kalman filter for elliptic observations

Mathieu Le Provost, Ricardo Baptista, Youssef Marzouk, Jeff D. Eldredge

Abstract

We propose a regularization method for ensemble Kalman filtering (EnKF) with elliptic observation operators. Commonly used EnKF regularization methods suppress state correlations at long distances. For observations described by elliptic partial differential equations, such as the pressure Poisson equation (PPE) in incompressible fluid flows, distance localization cannot be applied, as we cannot disentangle slowly decaying physical interactions from spurious long-range correlations. This is particularly true for the PPE, in which distant vortex elements couple nonlinearly to induce pressure. Instead, these inverse problems have a low effective dimension: low-dimensional projections of the observations strongly inform a low-dimensional subspace of the state space. We derive a low-rank factorization of the Kalman gain based on the spectrum of the Jacobian of the observation operator. The identified eigenvectors generalize the source and target modes of the multipole expansion, independently of the underlying spatial distribution of the problem. Given rapid spectral decay, inference can be performed in the low-dimensional subspace spanned by the dominant eigenvectors. This low-rank EnKF is assessed on dynamical systems with Poisson observation operators, where we seek to estimate the positions and strengths of point singularities over time from potential or pressure observations. We also comment on the broader applicability of this approach to elliptic inverse problems outside the context of filtering.

Introduction

In data assimilation, the Ensemble Kalman Filter (EnKF) is a powerful tool for estimating the state of high-dimensional systems by combining model forecasts with noisy observations. However, a persistent challenge arises when the ensemble size ($M$) is significantly smaller than the state dimension ($n$). In such cases, the filter is prone to "sampling errors"—spurious correlations between distant, physically unrelated points that can lead to filter divergence.

The traditional remedy is distance localization, which artificially dampens correlations between distant points. While effective for systems with local interactions, this approach fails for elliptic observations, such as pressure in incompressible fluid flows. In these systems, physics is governed by elliptic partial differential equations (PDEs) like the Poisson equation, where every part of the system is coupled to every other part. Because physical interactions decay slowly (algebraically), long-range correlations are often physically meaningful. Truncating them via distance localization removes vital information, rendering the filter ineffective.

Methodology: The Low-Rank EnKF (LREnKF)

This research introduces a regularization method based on dimension reduction rather than spatial truncation. The core insight is that while the state space may be high-dimensional, the observations typically only provide strong information about a small, low-dimensional subspace. The proposed Low-Rank EnKF (LREnKF) identifies and operates within this "informative" subspace.

1. Gramian Construction and Eigendecomposition

The method begins by defining state and observation "Gramians" based on the expected inner product of the observation operator's Jacobian. This mathematical framework identifies which directions in the state space most strongly influence the observations. By performing an eigendecomposition of these Gramians, the algorithm identifies the "leading modes" of the system.

2. Subspace Update

Mathematically, this approach acts as a data-driven version of the Fast Multipole Method (FMM), identifying generalized "source" and "target" clusters. The Kalman gain is factorized and truncated, ensuring that the analysis (update) step is performed only within the low-rank informative subspace. This prevents the introduction of noise from uninformative directions.

3. Adaptive Ranking

To maintain flexibility, the LREnKF can adaptively determine the number of modes to retain based on a "cumulative energy" threshold (e.g., retaining enough modes to explain 99% of the spectral energy), ensuring the filter captures the most relevant physical dynamics.

Experimental Evaluation

The LREnKF was evaluated using "twin experiments" across several dynamical systems:

  • Point Source Model: A static case demonstrating that LREnKF modes naturally recover the structure of classical multipole expansions.
  • Point Vortices near a Wall: Five point vortices advecting in a potential flow, with pressure sensors placed along a boundary.
  • Vortex Patch (Navier-Stokes): A complex scenario where the "truth" is a viscous simulation, but the filter uses a simplified inviscid model. This tested the filter's robustness to model error.

Key Results and Insights

Stability and Efficiency

The LREnKF demonstrated superior stability compared to the standard stochastic EnKF (sEnKF), particularly with small ensembles. In the vortex-along-a-wall experiment, the standard sEnKF became unstable for ensemble sizes $M < 40$. In contrast, the LREnKF remained stable and accurate with as few as $M = 10$ members. To achieve a target median RMSE of 0.07, the standard filter required 60 members, while the LREnKF required only 20 members.

Physics-Informed Regularization

Unlike distance localization, which is a geometric heuristic, LREnKF leverages the Jacobian of the observation operator to understand the underlying physics. By embracing the non-locality of elliptic operators, the method turns a perceived problem into a tool for effective dimension reduction.

Robustness to Model Error

In the vortex patch test, the LREnKF successfully tracked the center of a diffusing vortex patch despite the filter's internal model lacking viscosity. It significantly outperformed the standard EnKF, which tended to create spurious displacements of the vortices due to sampling noise.

Conclusion

The Low-Rank EnKF provides a robust framework for data assimilation in systems governed by elliptic PDEs. By identifying the low-dimensional subspace where observations are most informative, it enables stable state estimation with small ensembles where traditional localization fails. Future work may involve applying randomized numerical linear algebra to reduce the computational cost of forming the Jacobian and extending the method to complex 3D flows and other elliptic inverse problems in geophysics and electromagnetism.

Summary updated

Abstract

We propose a regularization method for ensemble Kalman filtering (EnKF) with elliptic observation operators. Commonly used EnKF regularization methods suppress state correlations at long distances. For observations described by elliptic partial differential equations, such as the pressure Poisson equation (PPE) in incompressible fluid flows, distance localization cannot be applied, as we cannot disentangle slowly decaying physical interactions from spurious long-range correlations. This is particularly true for the PPE, in which distant vortex elements couple nonlinearly to induce pressure. Instead, these inverse problems have a low effective dimension: low-dimensional projections of the observations strongly inform a low-dimensional subspace of the state space. We derive a low-rank factorization of the Kalman gain based on the spectrum of the Jacobian of the observation operator. The identified eigenvectors generalize the source and target modes of the multipole expansion, independently of the underlying spatial distribution of the problem. Given rapid spectral decay, inference can be performed in the low-dimensional subspace spanned by the dominant eigenvectors. This low-rank EnKF is assessed on dynamical systems with Poisson observation operators, where we seek to estimate the positions and strengths of point singularities over time from potential or pressure observations. We also comment on the broader applicability of this approach to elliptic inverse problems outside the context of filtering.
BibTeX
@inproceedings{provost2022low,
  author = {Mathieu Le Provost and Ricardo Baptista and Youssef Marzouk and Jeff D. Eldredge},
  title = {{A low-rank ensemble Kalman filter for elliptic observations}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2022},
  doi = {10.1098/rspa.2022.0182},
}
Figure