Sparsity-Aware Extended Kalman Filter

• Sparsity-Aware EKF is a family of nonlinear estimators that leverage sparse structures in models, measurements, and covariances to reduce computational complexity.
• Variants like LB-EKF, EKF-SINDy, EKF-STDF, GSP–EKF, and progressive EKF tailor sparsity using techniques such as graph relabeling, sparse regression, factor graphs, and ℓ1 regularization.
• Empirical studies demonstrate that these approaches achieve significant scalability improvements and maintain estimation accuracy across diverse applications.

Searching arXiv for recent and directly relevant papers on sparsity-aware EKF. arXiv_search(query="Sparsity-Aware Extended Kalman Filter", max_results=10, sort_by="relevance") Searching with broader queries to cover structural, covariance, and model sparsity variants of EKF. arXiv_search(query="extended Kalman filter sparsity localization graph SINDy", max_results=10, sort_by="relevance") Sparsity-aware Extended Kalman Filter (EKF) denotes a class of nonlinear recursive estimators that preserve the EKF prediction–update structure while exploiting sparse structure in the model, measurements, latent state, or error covariance. Across recent work, this label has been applied to at least five distinct mechanisms: graph-induced banded information matrices for large-scale wireless sensor network localization, sparse process models identified by SINDy, sparse spatio-temporal factor graphs for cooperative positioning, $\ell_1$-regularized latent edge-weight estimation for dynamic graph tracking, and sparse covariance approximations for very high-dimensional data assimilation (Khan et al., 2023, Rosafalco et al., 2024, Cao et al., 2023, Dabush et al., 14 Jul 2025, Kang et al., 2018). The shared objective is to mitigate the cubic-time and quadratic-memory burdens of dense EKF variants, or to improve identifiability and robustness when the underlying system is intrinsically sparse.

1. Scope and formal viewpoint

The standard EKF operates on a nonlinear state-space model

$$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$

with Gaussian process and measurement noise, local Jacobians $F_{k-1}$ and $H_k$, and the familiar gain

$$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$

Sparsity-aware variants leave this recursive skeleton intact but change how $f$, $h$, $H_k$, $P_k$, or the latent state are represented and updated. In the literature, sparsity is not a single mathematical object; it can mean a sparse Jacobian, a banded information matrix, a sparse coefficient matrix in a learned dynamics model, a sparse factor graph, a sparse latent support, or a sparse approximation to the error covariance.

Variant	Sparsity target	Representative paper
LB-EKF	Measurement information matrix made banded by graph relabeling	(Khan et al., 2023)
EKF-SINDy	Sparse coefficient matrix in the process model	(Rosafalco et al., 2024)
EKF-STDF	Sparse factor-graph coupling over local neighbors	(Cao et al., 2023)
GSP–EKF	Sparse latent edge-weight state with $\ell_1$-regularized update	(Dabush et al., 14 Jul 2025)
Progressive EKF	Sparse error covariance in high dimensions	(Kang et al., 2018)

A useful unifying interpretation is that sparsity-aware EKF methods replace dense global couplings by structured local ones. In some cases the structure is exact, as when pairwise measurements induce a block-sparse Jacobian aligned with a measurement graph. In other cases it is a modeling prior, as when an $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$0 penalty is imposed on a latent edge-weight vector or when a covariance is projected onto a prescribed sparsity pattern.

2. Graph-structured measurements and banded information filtering

A prominent instance arises in large-scale wireless sensor network localization. For $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$1 agents in $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$2 dimensions, the global state is

$$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$3

and the measurement graph $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$4 is sparse under a limited sensing radius. Pairwise distance and bearing measurements couple only the incident nodes, so each pairwise row of the EKF Jacobian $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$5 has nonzero blocks only at the columns associated with the two measured agents. This yields a block-sparse measurement structure whose information contribution

$$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$6

is “Laplacian-like” at the block level. More specifically, the sparsity of the pairwise component matches the sparsity of the graph Laplacian tensored with a $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$7 all-ones matrix, and the matrix bandwidth satisfies

$$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$8

for graph Laplacian $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$9 (Khan et al., 2023).

The Low-Bandwidth EKF (LB-EKF) exploits this observation by relabeling the graph vertices before the EKF update. If $F_{k-1}$0 is a permutation of the vertices, the permuted state $F_{k-1}$1 induces permuted matrices $F_{k-1}$2 and $F_{k-1}$3. The aim is to minimize graph bandwidth

$F_{k-1}$4

so that $F_{k-1}$5 becomes banded and the EKF information system can be solved by banded linear algebra. The paper studies classical orderings such as Cuthill–McKee, Reverse CM, BFS layering, and geometric orderings, and proposes a geometric “VR algorithm” that sorts agents by their $F_{k-1}$6-coordinate. For random geometric graphs in a square domain, the resulting bandwidth is shown to grow sublinearly with network size, approximately like $F_{k-1}$7 in expectation, which supports scalable localization.

Algorithmically, LB-EKF replaces dense matrix inversion by an $F_{k-1}$8-banded inverse approximation or, equivalently, by banded Cholesky or $F_{k-1}$9 factorization. Dense EKF updates require $H_k$0 time and $H_k$1 memory. After relabeling, the banded system scales as approximately $H_k$2 in time and $H_k$3 in memory, where $H_k$4 is the bandwidth. In the reported 2D experiment with 30 agents in a $H_k$5 domain, 8 anchors, sensing radius $H_k$6, and 5000 Monte Carlo trials, VR ordering reduced the graph bandwidth to 8 on the illustrated instance, and LB-EKF+VR with $H_k$7 achieved mean-squared error close to EKF in both transient and steady-state; covariance ellipses from LB-EKF+VR and EKF overlapped closely, whereas LB-EKF without VR performed worse (Khan et al., 2023).

3. Sparse process models and analytic Jacobians via SINDy

A different notion of sparsity appears in EKF-SINDy, where the EKF process model is not made sparse by graph topology or covariance truncation but by sparse identification of nonlinear dynamics. SINDy posits

$H_k$8

with a library $H_k$9 of candidate nonlinear functions and a sparse coefficient matrix $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$0. Under this construction, the EKF inherits a parsimonious process model that can be identified offline from data and used online for joint state–parameter estimation. Parameters are included by state augmentation,

$$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$1

with a random-walk prior for quasi-static parameters (Rosafalco et al., 2024).

The central technical advantage is that Jacobians become analytic and inexpensive: $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$2 This removes what the paper describes as a usually extremely involved derivation step for physics-based models. In the reported implementation, the EKF predictor uses Euler forward discretization, while sparse regression is performed offline by STLSQ or LASSO. For partially observed systems, time-delay embedding is used before SINDy so that the EKF runs in reconstructed coordinates rather than directly in the inaccessible physical state.

Two demonstrations make the idea concrete. In a 2-storey shear-building model excited by real STEAD seismograms, observations were corrupted with white noise at $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$3, signals were re-sampled to $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$4, and SINDy used a polynomial library up to second order with STLSQ threshold $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$5 and ridge parameter $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$6. The unknown inter-storey stiffness $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$7 was initialized 20% above the true value, yet the estimate converged to ground truth in approximately 20 seconds; the full 60-second assimilation ran in approximately 1.5 seconds on an Intel i7-2600 CPU with 16 GB RAM, about $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$8 faster than the physics-based process (Rosafalco et al., 2024).

In a partially observed nonlinear oscillator, only $$K_k = P_{k|k-1} H_k^\top \big(H_k P_{k|k-1} H_k^\top + R_k\big)^{-1}.$$9 was observed. Time-delay embedding with $f$0, $f$1, and truncated SVD retained 4 modes explaining $f$2 of the variance. EKF-SINDy estimated either $f$3 or $f$4, with convergence around $f$5 in the reported cases. One case used $f$6 with an initial guess outside the SINDy training range and underestimating the truth by about $f$7; another used $f$8, entirely outside the training range. In both settings, the reconstructed output remained accurate and the method exhibited strong noise tolerance. A plausible implication is that sparsity here functions primarily as a model-selection prior rather than as a matrix-structure prior.

4. Sparse factor graphs and distributed spatio-temporal fusion

In cooperative positioning for sparsely distributed high-mobility wireless networks, sparsity enters through the communication and measurement graph. Each mobile agent has state

$f$9

follows a constant-velocity model, and measures only a small subset of neighbors and anchors at each time slot. The resulting factor graph is low-degree: each variable $h$0 is connected only to its EKF prior, a temporal factor induced by the internally measured traveled distance, and a small set of spatial range factors. The proposed EKF-STDF algorithm is therefore “sparsity-aware” because both computation and communication scale with the local neighborhood rather than with total network size (Cao et al., 2023).

The method has three stages. First, each agent performs EKF prediction to obtain a coarse Gaussian prior over position. Second, a distributed spatio-temporal data fusion (STDF) stage runs belief propagation on the sparse factor graph. Nonlinear range likelihoods are approximated by second-order Taylor polynomials, but the approximation is carried far enough to keep all outgoing messages Gaussian and closed-form. The paper parameterizes these messages by coefficients $h$1, so that each one has the form

$h$2

Third, the fused mean $h$3 and covariance $h$4 are treated as a Gaussian pseudo-measurement in an EKF refinement step with measurement matrix $h$5.

This construction avoids particle approximations in the factor-graph stage. The paper reports per-agent complexity per time slot of

$h$6

where $h$7 is the number of local links and $h$8 is the iteration budget. Particle-based alternatives incur an additional $h$9 factor. In the reported simulations, the area was $H_k$0, with 13 anchors, 30–60 agents in an inner $H_k$1 area, communication radius $H_k$2, initial speed $H_k$3, Gaussian speed variation of standard deviation $H_k$4, and $H_k$5. Range noise variance was set to $H_k$6 for spatial measurements, and internal-distance noise variance to $H_k$7. Under these conditions, EKF-STDF outperformed SPA-EKF and STOC-EKF in RMSE, especially when neighbors were scarce, while NEBP could become slightly superior only at higher density and with significantly higher computational complexity.

5. $H_k$8-regularized EKF for sparse dynamic graph topology

A more explicit sparsity prior appears in dynamic graph tracking. Here the hidden state is not a physical coordinate vector but the weighted edge set of a graph. For an undirected weighted graph on $H_k$9 nodes, the complete-graph incidence matrix $P_k$0 with $P_k$1 possible edges defines the Laplacian

$P_k$2

where $P_k$3 is the nonnegative edge-weight vector. Observations are graph-filtered signals,

$P_k$4

so nonlinearity enters through the dependence of $P_k$5 on the sparse latent state $P_k$6 (Dabush et al., 14 Jul 2025).

The EKF prediction step is standard, but the update is replaced by a locally linearized one-step MAP problem with $P_k$7 regularization: $P_k$8 The paper solves this by ISTA, initializes the proximal iterations with the unregularized EKF estimate, and then projects the result onto the nonnegative orthant. Because the posterior is no longer strictly Gaussian under the $P_k$9 penalty, the usual EKF covariance update is retained only as an approximation.

The main computational obstacle is the measurement Jacobian. For polynomial graph filters, the exact Jacobian column associated with edge $\ell_1$0 is

$\ell_1$1

A dynamic-programming scheme reuses powers of $\ell_1$2 and reduces Jacobian construction from naive $\ell_1$3 to $\ell_1$4. The paper also notes a necessary observability condition for the linear special case: since the time-varying observability matrix has $\ell_1$5 rows and $\ell_1$6 columns, one must have $\ell_1$7. This motivates the sparsity prior, because $\ell_1$8-sparse edge states can be recovered from fewer effective measurements than dense ones.

Empirically, the sparsity-aware GSP–EKF outperformed both standard EKF and a change-detection batch baseline in MSE and edge identification error rate, and approached an oracle with known support. The gains persisted for polynomial measurement nonlinearities of orders 4–5, across multiple noise levels and change rates, and single-step ISTA was reported to be sufficient in practice. This suggests that sparsity regularization can be integrated into EKF updates without abandoning online operation.

6. Sparse covariance propagation in very high dimensions

The oldest strand in this literature targets a different bottleneck: the covariance itself. In very high-dimensional data assimilation, dense EKF covariance propagation is intractable even when model and observation operators are local. The progressive EKF therefore assumes that the true error covariance can be well approximated by a sparse, symmetric positive definite matrix with a fixed or adaptively maintained sparsity pattern. The model and observation operators are evaluated in a component-based localized fashion so that cost is proportional to the allowed nonzeros rather than to $\ell_1$9 or $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$00 (Kang et al., 2018).

The key approximation is applied in the forecast covariance. Writing the model Jacobian as $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$01 for sufficiently small time steps, the exact propagation

$$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$02

is replaced by a first-order progressive approximation. A finite-difference matrix $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$03 is computed columnwise over the prescribed sparsity pattern, and the forecast covariance is projected back to that pattern: $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$04 After the analysis step, the covariance is again projected and thresholded. The paper recommends Joseph-form updating for numerical stability, symmetry restoration, and diagonal loading $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$05 if positive definiteness is lost. When the model time step is not sufficiently small, the interval can be subcycled into $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$06 smaller propagation steps.

This approach is explicitly contrasted with EnKF-style low-rank covariance approximations. The sparse covariance remains full rank; the algorithms provide updated error covariance for the next assimilation cycle; memory usage is reduced because only the sparse structure is stored; and the granularity of the sparse covariance can be adjusted to optimize parallelization. In the Lorenz-96 experiment with $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$07, $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$08, $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$09 measured states, $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$10, and initial $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$11, the EnKF baseline with $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$12, localization radius $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$13, and inflation $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$14 had median RMSE about $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$15. Progressive EKF reported median RMSE values of about $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$16 for $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$17, $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$18 for $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$19, $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$20 for $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$21, and $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$22 for $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$23. The improvement with increasing sparsity budget and subcycling is central: it shows that sparsity alone is not sufficient; the propagation approximation must also respect the local nonlinear dynamics.

7. Assumptions, limitations, and recurrent misconceptions

A recurrent misconception is that “sparsity-aware” refers to a single numerical technique. The literature instead uses the term for structurally different modifications of the EKF: banding an information matrix, learning sparse governing equations, restricting inference to a low-degree factor graph, adding an $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$24 prior to the latent state, or enforcing sparsity in the covariance. These approaches are compatible in spirit but not interchangeable.

Another frequent misunderstanding concerns “low bandwidth.” In LB-EKF, “Low-Bandwidth” refers to matrix bandwidth in numerical linear algebra, not network communication bandwidth. The method in that paper is centralized; its gain is computational efficiency and reduced memory bandwidth, not reduced inter-node messaging. The same paper also makes clear that performance depends strongly on the achieved ordering and on the choice of the band parameter $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$25: if $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$26 is too small relative to the true bandwidth, neglected long-range cross-covariances can bias the covariance and degrade performance or cause divergence (Khan et al., 2023).

Observability and initialization remain central despite sparsity. EKF-SINDy still requires sufficient persistency of excitation, and partial observation may require time-delay embedding; unknown inputs were not handled in the reported formulation. In the cooperative-positioning setting, extreme sparsity or long periods with very few neighbors limit observability, and second-order message approximations can deteriorate under strong nonlinearity or poor linearization points. In the high-dimensional covariance setting, long-range correlations or highly nonlocal observation operators force larger sparsity patterns, batching, or tempering, while large time steps make subcycling important for forecast accuracy (Rosafalco et al., 2024, Cao et al., 2023, Kang et al., 2018).

Finally, sparsity-aware updates need not preserve the exact probabilistic semantics of the Gaussian EKF. In the dynamic-graph formulation, the posterior covariance is not strictly Gaussian under $$x_k = f(x_{k-1}) + w_{k-1}, \qquad y_k = h(x_k) + v_k,$$27 regularization, so the EKF covariance recursion is retained as an approximation rather than as an exact posterior update (Dabush et al., 14 Jul 2025). This is not a defect specific to that method; it reflects a broader tradeoff in sparsity-aware filtering between strict Bayesian exactness and tractable online estimation.

Taken together, the literature shows that sparsity-aware EKF is best understood not as a single algorithm but as a design principle: identify the physically or statistically sparse structure in a nonlinear filtering problem, rewrite the EKF around that structure, and accept carefully controlled approximations where dense Gaussian filtering would otherwise be computationally or statistically untenable.