Low-Rank Ensemble Kalman Filter (LREnKF)

• LREnKF is a filtering method that constrains updates to a low-dimensional subspace, reducing computational complexity in high-dimensional state estimation.
• It employs diverse parameterizations such as affine subspace representation, spectral reduction via Jacobian factorization, and deterministic low-rank approximations.
• Applications in reaction-diffusion, pressure sensing, and inverse scattering demonstrate LREnKF’s efficiency and accuracy with lower ensemble sizes and faster computations.

Low-Rank Ensemble Kalman Filter (LREnKF) denotes a family of ensemble-Kalman methods in which the filtering update is constrained to a low-dimensional structure, so that high-dimensional state estimation can be performed with reduced computational cost. In the cited literature, this low-rank structure appears in several mathematically distinct forms: as the intrinsic rank-$\leq N-1$ sample covariance of an ensemble, as a dynamically evolving subspace $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$, as a Jacobian-informed factorization of the Kalman gain, and as deterministic covariance factorizations that serve as idealized counterparts of ensemble filtering (Nobile et al., 6 Feb 2026, Provost et al., 2022, Schmidt et al., 2023).

1. Low-rank structure in ensemble Kalman filtering

At the most basic level, ensemble Kalman filtering is already a low-rank approximation. For an ensemble of size $M$, the forecast covariance in the continuous-time limit can be written as

$P_t=\frac{1}{M-1}E_tE_t^\top,$

where $E_t$ is the anomaly matrix, so $\operatorname{rank}(P_t)\le M-1$. In discrete time the same structure appears as

$P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$

This makes any EnKF a covariance-restricted filter whose effective uncertainty representation is confined to the ensemble span (Lange et al., 2019).

A closely related formulation writes the anomaly matrix

$$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$

so that $P\approx AA^\top$. In that formulation, an EnKF with $N$ ensemble members is conceptually a low-rank Kalman filter with rank $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$0 (Schmidt et al., 2023). This observation is central because it explains why EnKF remains tractable in state dimensions for which full $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$1 covariance propagation is infeasible, but it also identifies the core limitation: the filter can only update directions that lie in the ensemble-induced subspace.

The continuous-time deterministic EnKBF formulation makes this interpretation especially explicit. In the linear case, the ensemble mean and sample covariance satisfy the Kalman-Bucy mean equation and the Riccati equation within the ensemble subspace, so the filter can be read as a low-rank factorization of the Kalman-Bucy dynamics rather than merely a Monte Carlo surrogate (Lange et al., 2019). This suggests that many LREnKF constructions are best viewed as structured evolutions of a low-dimensional covariance manifold.

2. Principal low-rank parameterizations

One prominent parameterization constrains each state particle to a time-dependent affine subspace,

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$2

with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$3 orthonormal, $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$4 the mean, and $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$5 reduced coordinates satisfying $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$6. The corresponding state covariance has the factorized form

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$7

In the state-parameter formulation of the DLR-EnKF, only the state is reduced, while the parameter ensemble is kept full because $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$8 is small (Nobile et al., 6 Feb 2026). The computational implication given in that work is that evolving $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$9 yields EnKF-like updates at cost $M$0 and $M$1 instead of $M$2 (Nobile et al., 6 Feb 2026).

A second parameterization arises from observation-informed spectral reduction. For nonlinear observation operators $M$3, the LREnKF for elliptic observations defines the whitened Jacobian

$M$4

then builds state and observation Gramians

$M$5

If $M$6 and $M$7 contain leading eigendirections of $M$8 and $M$9, the Kalman gain is approximated in the lifted low-rank form

$P_t=\frac{1}{M-1}E_tE_t^\top,$0

with $P_t=\frac{1}{M-1}E_tE_t^\top,$1 estimated in the reduced coordinates (Provost et al., 2022). Here the low-rank object is not the state itself but the analysis map from innovations to state increments.

A third formulation is deterministic rather than ensemble-based. The rank-reduced Kalman filter approximates the covariance by

$P_t=\frac{1}{M-1}E_tE_t^\top,$2

and updates $P_t=\frac{1}{M-1}E_tE_t^\top,$3 by dynamical low-rank approximation and square-root conditioning formulas (Schmidt et al., 2023). Although this is not an EnKF, the paper explicitly frames it as a deterministic alternative to ensemble-based low-rank filtering and notes that the anomaly matrix in EnKF plays exactly the role of a covariance factor $P_t=\frac{1}{M-1}E_tE_t^\top,$4 (Schmidt et al., 2023).

3. Dynamical low-rank ensemble filtering and joint state-parameter estimation

The most explicit recent LREnKF construction is the Dynamical Low-Rank Ensemble Kalman Filter for joint state-parameter estimation. It considers an augmented variable

$P_t=\frac{1}{M-1}E_tE_t^\top,$5

with static parameter dynamics in the model and observations

$P_t=\frac{1}{M-1}E_tE_t^\top,$6

The method extends the earlier DLR-ENKF formulation for the Kalman-Bucy process to nonlinear drift and augmented state-parameter dynamics, while retaining the ansatz $P_t=\frac{1}{M-1}E_tE_t^\top,$7 for the state component (Nobile et al., 14 Sep 2025, Nobile et al., 6 Feb 2026).

A central feature is that the basis is not fixed offline. The basis itself evolves in time, and the forecast step is implemented by a Basis Update & Galerkin scheme combined with a forecast/analysis split. On one step $P_t=\frac{1}{M-1}E_tE_t^\top,$8, the method forms augmented coordinates and modes,

$P_t=\frac{1}{M-1}E_tE_t^\top,$9

then updates auxiliary matrices $E_t$0 and $E_t$1, orthonormalizes, performs a Galerkin projection, and finally truncates the zero-mean anomalies back to rank $E_t$2 by a generalized SVD (Nobile et al., 6 Feb 2026). For multi-variable states, truncation is applied variable-wise to avoid scale imbalance.

The analysis step is performed in reduced coordinates. With $E_t$3 and

$E_t$4

the reduced gain is

$E_t$5

and the full-space gain is $E_t$6. The state analysis update then has the EnKF form

$E_t$7

while the parameter update uses the usual augmented-state cross-covariance,

$E_t$8

with

$E_t$9

(Nobile et al., 6 Feb 2026).

For nonlinear drift evaluation, the same work introduces a DEIM-based CUR approximation. If $\operatorname{rank}(P_t)\le M-1$0, then

$\operatorname{rank}(P_t)\le M-1$1

with row and column indices selected by DEIM. The stated effect is to reduce the evaluation of reduced quantities such as $\operatorname{rank}(P_t)\le M-1$2 and $\operatorname{rank}(P_t)\le M-1$3 from $\operatorname{rank}(P_t)\le M-1$4 to $\operatorname{rank}(P_t)\le M-1$5, plus evaluation at selected indices (Nobile et al., 6 Feb 2026). When $\operatorname{rank}(P_t)\le M-1$6 and truncation is not applied, the low-rank scheme reduces to the full EnKF up to numerical roundoff (Nobile et al., 6 Feb 2026). This suggests that the method is a dynamically constrained reformulation of EnKF rather than a different estimator.

4. Spectral and operator-adapted LREnKF variants

For elliptic observation operators, the principal difficulty is not state propagation but analysis regularization. The LREnKF for elliptic observations argues that distance localization is inappropriate for operators such as Poisson maps because the Green’s function decays logarithmically in $\operatorname{rank}(P_t)\le M-1$7D or algebraically in $\operatorname{rank}(P_t)\le M-1$8D, so physical state-observation correlations are genuinely long-range. In the point-vortex pressure example, empirical cross-covariances decay like $\operatorname{rank}(P_t)\le M-1$9, remain significant at large distances, and exhibit different behaviors across state components, making a single localization radius inadequate (Provost et al., 2022). The proposed remedy is a low-rank Kalman gain built from Jacobian spectra, which the paper describes as a form of spectral localization.

The numerical evidence in that setting is explicit. For estimation of $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$0 point vortices from $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$1 wall-pressure sensors, with state dimension $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$2, the spectra of $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$3 and $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$4 decay sharply, and the informative subspaces are very low-dimensional. In the same example, sEnKF diverges for $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$5, whereas LREnKF remains accurate even for $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$6. To achieve median RMSE $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$7, the paper reports that sEnKF requires $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$8, while LREnKF with $P_k^f=\frac{1}{M-1}E_k^f(E_k^f)^\top.$9 spectral energy needs only $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$0. In the vortex-patch example with state dimension $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$1 and $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$2 pressure sensors, sEnKF diverges for $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$3, while LREnKF continues to give reasonable pressure estimates for $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$4; to achieve median MSE $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$5, sEnKF requires $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$6, whereas LREnKF with the tested thresholds needs only $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$7 (Provost et al., 2022).

A different operator-adapted construction appears in inverse medium scattering. There the low-rank space is defined by disk prolate spheroidal wave functions $$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$8, which satisfy

$$A=\frac{1}{\sqrt{N-1}}\,[x^{(1)}-\bar x,\dots,x^{(N)}-\bar x],$$9

for the Born forward operator and also diagonalize a Sturm-Liouville operator $P\approx AA^\top$0. The low-rank space is

$P\approx AA^\top$1

so its dimension is intrinsically determined by the wave number $P\approx AA^\top$2 through $P\approx AA^\top$3. In that basis the EnKF is run on low-rank coefficient vectors, initialized from an inverse Born estimate and a trace-class covariance

$P\approx AA^\top$4

whose spectrum is $P\approx AA^\top$5 (Meng, 7 Apr 2026).

That paper also proves low-rank stability statements. In the fully nonlinear case, if $P\approx AA^\top$6 lie in a convex compact subset of the DPSWF low-rank space, then

$P\approx AA^\top$7

In the Born region, the stability constant becomes explicit: $P\approx AA^\top$8 Numerically, for the “Cross 2D” scatterer at $P\approx AA^\top$9, $N$0 noise, and ensemble size $N$1, the inverse Born reconstruction is poor, whereas the low-rank-assisted EnKF improves amplitude by iteration $N$2 and adds finer details by iteration $N$3 (Meng, 7 Apr 2026). This suggests that in inverse problems with a known spectral basis for the forward operator, LREnKF can be tied directly to operator theory rather than to empirical covariance structure alone.

5. Deterministic low-rank counterparts and theoretical analogues

Several recent deterministic filters clarify what an LREnKF is approximating. The rank-reduced Kalman filter propagates low-rank covariance factors $N$4 by dynamical low-rank approximation on the manifold of rank-$N$5 matrices and uses the BUG integrator of Ceruti and Lubich for the Lyapunov prediction step. The update step is a deterministic square-root transformation in the column space of $N$6, and the method reproduces the exact Kalman filter as the low-rank dimension approaches the full state dimension or the true intrinsic rank (Schmidt et al., 2023). Under structural assumptions, one prediction-update step costs

$N$7

with worst-case $N$8 if those assumptions fail (Schmidt et al., 2023).

The same paper explicitly compares RRKF with standard EnKF and ETKF. In the linear advection example with $N$9 and true covariance rank $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$00, RRKF becomes exact once $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$01, while ensemble methods approach the Kalman filter more slowly. Across the reported experiments, RRKF is consistently closer to the exact Kalman filter than EnKF or ETKF at the same rank, but it can be overconfident for very small $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$02, and the paper notes that it currently lacks a principled analogue of covariance inflation (Schmidt et al., 2023). In this sense it functions as a deterministic benchmark for low-rank ensemble methods.

A more control-theoretic analogue is the discrete-time low-rank Kalman filter based on Oja’s principal component flow. There the covariance is represented as

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$03

with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$04 evolved by

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$05

The reduced gain is

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$06

and the reduced covariance obeys a rank-$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$07 Riccati recursion. The main theorem states that if $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$08 is the number of Hurwitz-unstable eigenvalues of $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$09, then the low-rank closed-loop filter is Schur stable if and only if $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$10 (Tsuzuki et al., 2024). This gives a precise lower bound on admissible rank and is directly suggestive for LREnKF design: the effective ensemble rank must cover the unstable subspace.

Another related line is the random-walk fast Kalman filter based on hierarchical matrices and a low-rank perturbative covariance representation

$X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$11

with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$12. In the random-walk setting the method achieves $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$13 memory and $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$14 computational cost (Saibaba et al., 2014). Although not ensemble-based, it shares the same structural premise as LREnKF: posterior covariance differs from a background covariance primarily in a small data-informed subspace.

A related ensemble-Kalman development appears in EnK-RML for Bayesian smoothing. There, balanced truncation is tailored to the smoothing problem through a generalized eigenproblem for an observability-like Gramian and the prior precision, and the reduced posterior covariance converges to an optimal low-rank update of the prior covariance in the appropriate limit. The authors present this as directly relevant to LREnKF design because the nontrivial updates occur in an information subspace $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$15, while the complementary subspace remains unchanged (Stavrinides et al., 3 Jul 2025).

6. Performance regimes, applications, and limitations

The recent DLR-EnKF state-parameter study provides concrete evidence on the rank-accuracy tradeoff. In the Fisher-KPP reaction-diffusion example, the PDE is discretized with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$16 degrees of freedom and $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$17 unknown parameters. With full observations and rank $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$18, DLR-EnKF tracks both state and parameters almost identically to full EnKF, and DLR-SEnKF with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$19 matches SenKF performance. Under partial observations, ranks $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$20 and $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$21 are insufficient to match EnKF behavior; with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$22, most parameters are still identified well, but some remain biased. In the reduced $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$23D blood-flow model with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$24, $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$25, and $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$26, DLR-SEnKF with $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$27 estimates $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$28 accurately but gives biased $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$29 and underestimated variance, whereas $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$30 improves $X_t^{(p)} = U_t^0 + U_t Y_t^{(p)\top}$31 and brings performance closer to full EnKF. In both benchmarks, DLR variants are faster than the full EnKF, with runtime increasing with rank (Nobile et al., 6 Feb 2026).

The broader literature identifies several recurring failure modes. If the retained rank is too small, state covariance is underestimated and state-parameter cross-covariances can be biased, leading to biased parameter estimates or divergence (Nobile et al., 6 Feb 2026). In deterministic low-rank filtering, very small rank can also produce overconfident posteriors, as documented by Z-score analyses in RRKF experiments (Schmidt et al., 2023). In dynamical low-rank Kalman-Bucy theory, approximation quality depends on the model noise being small or concentrated near the evolving low-rank subspace; when neglected noise components are not small, the DLR approximation deteriorates (Nobile et al., 14 Sep 2025). In elliptic observation problems, the difficulty is different: distance localization fails because long-range correlations are physical rather than spurious, so spectral reduction is preferable to spatial tapering (Provost et al., 2022). In inverse scattering, the low-rank dimension grows with wave number, so higher-frequency reconstructions improve resolution but enlarge the reduced coefficient space and increase ensemble cost (Meng, 7 Apr 2026).

A common misconception is that “low rank” denotes a single algorithm. The cited work instead supports a more differentiated view. Low rank may refer to the ensemble covariance rank, the state subspace, the gain factorization, a deterministic covariance square root, or a basis supplied by operator theory. Another common misconception is that low-rank filtering is synonymous with fixed reduced bases. The dynamical low-rank literature explicitly rejects that equivalence: DLR-EnKF uses a dynamically evolving basis with no offline POD stage, and this is presented as advantageous for moving features, waves, and shocks (Nobile et al., 6 Feb 2026). Conversely, the operator-adapted elliptic and scattering constructions show that fixed spectral bases can be preferable when the observation operator itself imposes a rapidly decaying informative spectrum (Provost et al., 2022, Meng, 7 Apr 2026).

Taken together, these works define LREnKF not as a single canonical filter but as a research program organized around one principle: the informative part of filtering dynamics often occupies a subspace of dimension far smaller than the ambient state dimension, and the success of a low-rank formulation depends on how that subspace is represented, evolved, and coupled to the observation operator.