Reduced-Complexity Kalman Gain Calculation

• Reduced-complexity Kalman gain calculation is a set of methods that accelerate filtering by shifting matrix inversions from high-dimensional measurement space to smaller state space, often via algebraic reformulations like the Woodbury identity.
• Techniques include sequential, low-rank, and iterative approaches, as well as neural network approximations, each balancing computational speed with estimation accuracy.
• Practical implementations leverage sparsity, block structure, and model reduction in applications such as LiDAR-inertial navigation, CFD, and large-scale data assimilation to achieve significant runtime savings.

Reduced-complexity Kalman gain calculation refers to algorithmic and statistical techniques for accelerating the computation of the Kalman gain—central to both classic and extended Kalman filtering—by exploiting algebraic, structural, or learning-based properties of the underlying models. Classical Kalman gain updates scale cubically with the measurement or state dimensions, which is prohibitive in high-rate, high-dimensional applications such as LiDAR-inertial navigation, fluid simulation, spatiotemporal field estimation, and large-scale data assimilation. Modern research has developed a variety of approaches that reduce asymptotic complexity while preserving—sometimes exactly, sometimes approximately—the optimality and stability properties of the full gain. These span exact reformulation using matrix inverse identities, low-rank and subspace methods, measurement or state partitioning, algebraic and gradient-based iterative solvers, and neural network approximations.

1. Classical Formulation and Complexity

In the standard linear Kalman filter and its nonlinear extensions, the measurement update step involves computation of the optimal gain

$K = P^- H^\top (H P^- H^\top + R)^{-1}$

for state-error covariance $P^-\in\mathbb R^{n\times n}$, measurement Jacobian $H\in\mathbb R^{m\times n}$, and observation noise $R\in\mathbb R^{m\times m}$. Direct computation entails forming the $m\times m$ innovation covariance $S=H P^- H^\top + R$ and inverting it at a dominant cost $O(m^3)$, with additional $O(n^2 m)$ and $O(n m^2)$ costs for the matrix products. For extended Kalman filters (EKF) or iterated EKF (iEKF), this cost is compounded by the repeated linearization and update cycles.

This scaling is impractical for scenarios with large $m$—as in multi-feature sensor fusion, dense data assimilation, or spatiotemporal field estimation—and motivates reductions whose complexity depends on $P^-\in\mathbb R^{n\times n}$0 (state dimension, often small) rather than $P^-\in\mathbb R^{n\times n}$1 (measurement dimension, often large) (Xu et al., 2020).

2. Exact Algebraic Reparameterizations

A seminal reduced-complexity approach is to algebraically reformulate the Kalman gain expression to shift the inversion from the usually large measurement space ($P^-\in\mathbb R^{n\times n}$2) to the typically smaller state space ($P^-\in\mathbb R^{n\times n}$3), or to exploit sparsity/block structure for sequential updating.

2.1 Woodbury Identity ("FAST-LIO" Formula)

Applying the Woodbury matrix inversion lemma: $P^-\in\mathbb R^{n\times n}$4 to the innovation covariance $P^-\in\mathbb R^{n\times n}$5, with $P^-\in\mathbb R^{n\times n}$6, yields the equivalent "small-matrix" Kalman gain: $P^-\in\mathbb R^{n\times n}$7 This form eliminates direct formation and inversion of the $P^-\in\mathbb R^{n\times n}$8 matrix $P^-\in\mathbb R^{n\times n}$9, replacing it with the inversion of an $H\in\mathbb R^{m\times n}$0 matrix $H\in\mathbb R^{m\times n}$1. The dominant cost shifts from $H\in\mathbb R^{m\times n}$2 to $H\in\mathbb R^{m\times n}$3, which in applications like LiDAR-inertial odometry, where $H\in\mathbb R^{m\times n}$4 and $H\in\mathbb R^{m\times n}$5–$H\in\mathbb R^{m\times n}$6, results in 2–3 orders of magnitude runtime savings (Xu et al., 2020).

2.2 Sequential and Block-structured Updates

If $H\in\mathbb R^{m\times n}$7 is block- or diagonal-structured—or if the overall system admits partitioning into weakly coupled subblocks—updates can be applied sequentially, in blocks, or band-wise, reducing the per-update cost to $H\in\mathbb R^{m\times n}$8 per block of size $H\in\mathbb R^{m\times n}$9, often with minimal loss in estimation accuracy. These strategies are common in large-scale geophysical data assimilation, turbulence modeling, and CFD (Raitoharju et al., 2015, Meldi et al., 2017).

2.3 Sparsity and Square-root/Cholesky Techniques

Further acceleration is possible by leveraging sparsity patterns and using square-root (Cholesky) factorizations in the update step. Exploiting the block-diagonal, banded, or sparse structure of $R\in\mathbb R^{m\times m}$0, $R\in\mathbb R^{m\times m}$1, and $R\in\mathbb R^{m\times m}$2 reduces both compute and storage, particularly for spatial discretizations or large distributed sensor systems (Raitoharju et al., 2015). Cholesky-based algorithms also improve numerical stability.

3. Low-rank, Subspace, and Model-reduction Methods

A particularly powerful class of approaches rests on low-rank or subspace representations of the state covariance and gain, leveraging the empirical observation that the effective posterior variance concentrates in a low-dimensional manifold.

3.1 Dynamical Low-rank Filters

Approaches such as dynamical low-rank approximation (DLR) for the Kalman-Bucy process and the discrete-time Oja flow-based filter propagate a factorization $R\in\mathbb R^{m\times m}$3, with $R\in\mathbb R^{m\times m}$4 and $R\in\mathbb R^{m\times m}$5, with $R\in\mathbb R^{m\times m}$6. Riccati/Lyapunov equations are projected onto this low-dimensional manifold: $R\in\mathbb R^{m\times m}$7 Gain and mean updates require only operations on $R\in\mathbb R^{m\times m}$8 matrices and $R\in\mathbb R^{m\times m}$9 factors. The measurement update uses SVDs or inverted projected covariances in $m\times m$0 dimensions (Nobile et al., 14 Sep 2025, Tsuzuki et al., 2024, Schmidt et al., 2023). These methods achieve $m\times m$1 complexity per step, with accuracy directly controlled by the neglect of directions outside the subspace.

3.2 A Priori Dimension Reduction

Offline identification of empirical dominant subspaces (via PCA, KL expansion, or Gaussian process fits) enables subspace-constrained prediction and update; the Kalman gain is then computed in the reduced basis ($m\times m$2), and lifted to full state space, replacing the dominant $m\times m$3 costs with $m\times m$4 or less (Solonen et al., 2015). This strategy is highly effective for spatiotemporal fields with intrinsic smoothness.

3.3 Application to CFD and Large-scale Data Assimilation

In computational fluid dynamics (CFD), “observer” and “filtered-P” strategies utilize local (diagonal) or sparse-band approximations for $m\times m$5 and local Kalman gain computations, reducing the per-grid-cell cost to $m\times m$6–$m\times m$7 and overall overhead to 10–15% above the original solver without material loss of assimilation accuracy (Meldi et al., 2017).

4. Data-selection and Learning-based Gain Approximation

Novel approaches include measurement selection (“sketching”), neural approximators, and stochastic approximation for regimes where even reduced algebraic complexity remains prohibitive.

4.1 Data Sketching and Measurement Subsampling

In large-scale inference, random projections (e.g., SRHT or Hadamard transform) reduce measurement size prior to gain computation, with accuracy guarantees; censoring strategies select only “informative” innovations based on statistical thresholds or KL-divergence in sequential updates. The resultant approximate gains achieve significant complexity reduction with provable bounds and can be tuned for tradeoffs between efficiency and estimation error (Berberidis et al., 2016).

4.2 Neural Kalman Gain and SNN/RNN-based Filters

Neural network surrogates, such as spiking neural networks (SNNs) and recurrent neural networks (RNNs), are trained in a teacher-forcing or reward-modulated regime to approximate the Kalman gain mapping from innovation or prediction errors directly, bypassing explicit matrix inversion. These architectures reduce per-step gain computation from $m\times m$8/$m\times m$9 to $S=H P^- H^\top + R$0 (SNN forward-passes or small RNNs), with effective performance on neuromorphic or embedded platforms. Error reduction of 15–65% and real-time performance (RTF=0.09) are reported for challenging acoustic echo cancellation and UAV tracking tasks (Xiao et al., 17 Apr 2025, Yang et al., 2022). Such neural gain formalisms are strictly data-driven, with trainability and robustness determined by representativeness of training data and capacity of the approximator; exact optimality is not generally guaranteed.

5. Iterative Solvers and Gradient-based Gain Computation

Gradient descent approaches minimize the innovation loss $S=H P^- H^\top + R$1 directly in filter gain space, using derivations of $S=H P^- H^\top + R$2 that decompose as products of the closed-loop observability Gramian and cross-covariance between the filter error and innovation. Convergence to the true Kalman gain is guaranteed under a strengthened observability condition for $S=H P^- H^\top + R$3, and per-iteration costs can be reduced to $S=H P^- H^\top + R$4 via truncated Gramian/Lyapunov solves when $S=H P^- H^\top + R$5, $S=H P^- H^\top + R$6 are sparse or for restricted time horizons. These approaches enable scalable, interpretable gain optimization in both steady-state and adaptive time-varying filtering (Belabbas et al., 18 Jul 2025).

6. Quantum and Asymptotic Acceleration

Fault-tolerant quantum algorithms, using block encoding and quantum singular value transformation (QSVT), can compute the full Kalman gain with a complexity scaling polynomially in $S=H P^- H^\top + R$7 and the condition number $S=H P^- H^\top + R$8 of the matrix to be inverted, achieving exponential speedup over classical $S=H P^- H^\top + R$9. The entire gain calculation (including Hamiltonian block encoding of $O(m^3)$0, $O(m^3)$1, and $O(m^3)$2) is embedded in a quantum circuit, with final measurement yielding the amplitude-encoded gain (Shi et al., 2024). While promising for future large-scale control, real-world feasibility currently depends on QRAM/hardware development and moderate condition numbers.

7. Practical Implementation Considerations

Implementation details depend on the target application and computational environment:

• For real-time navigation and robotics, streaming accumulation of $O(m^3)$3 and $O(m^3)$4 in batch iEKF systems enables per-scan update times $O(m^3)$5 ms, far faster than classical batch inversion (Xu et al., 2020).
• In CFD and PDE-constrained filtering, low-rank or locally-structured gains exploit the physics of locality, diagonal dominance, and discretization error to minimize global computation and storage (Meldi et al., 2017, Solonen et al., 2015).
• For online neural surrogate gain schemes, network parameter size, time window, and input feature encoding determine both theoretical and practical accuracy/runtime, with typical models remaining deployable on low-power neuromorphic hardware (Xiao et al., 17 Apr 2025, Yang et al., 2022).
• All reduced-complexity methods must account for statistical fidelity; in most cases, error bounds are available—via either information-theoretic (sketching), eigenvalue separation (low-rank), or convergence rate (gradient) analyses—to guide practical rank or subsampling choices.

Summary Table: Major Reduced-complexity Kalman Gain Techniques

Technique	Key Formula / Approach	Complexity Order
Woodbury/FAST-LIO Algebraic Form	$O(m^3)$6	$O(m^3)$7
Sequential/Block Update	Blockwise gain/inverse per measurement or cell	$O(m^3)$8
Low-rank Factorization (DLR, RRKF)	$O(m^3)$9, gain via projected update	$O(n^2 m)$0
Subspace/pca-enKF	Gain, state in reduced basis $O(n^2 m)$1	$O(n^2 m)$2
Neural Approximator (SNN/RNN)	Learned $O(n^2 m)$3, no inverse	$O(n^2 m)$4
Data Sketching	Random projection or update selection	$O(n^2 m)$5 for $O(n^2 m)$6
Quantum Block-encoding	Gain as quantum amplitude encoding	$O(n^2 m)$7

Each scheme provides a different trade-off profile among optimality, runtime, implementation complexity, and hardware suitability. The choice of technique should be informed by the size and structure of $O(n^2 m)$8, $O(n^2 m)$9, noise models, and the degree of accuracy required for stability and estimation objectives.

References:

• "FAST-LIO: A Fast, Robust LiDAR-inertial Odometry Package by Tightly-Coupled Iterated Kalman Filter" (Xu et al., 2020)
• "On Computational Complexity Reduction Methods for Kalman Filter Extensions" (Raitoharju et al., 2015)
• "A reduced order Kalman Filter model for sequential Data Assimilation of turbulent flows" (Meldi et al., 2017)
• "Spike-Kal: A Spiking Neuron Network Assisted Kalman Filter" (Xiao et al., 17 Apr 2025)
• "Low-Complexity Acoustic Echo Cancellation with Neural Kalman Filtering" (Yang et al., 2022)
• "The Rank-Reduced Kalman Filter: Approximate Dynamical-Low-Rank Filtering In High Dimensions" (Schmidt et al., 2023)
• "Dynamical Low-Rank Approximations for Kalman Filtering" (Nobile et al., 14 Sep 2025)
• "Low-rank approximated Kalman filter using Oja's principal component flow for discrete-time linear systems" (Tsuzuki et al., 2024)
• "On dimension reduction in Gaussian filters" (Solonen et al., 2015)
• "Data Sketching for Large-Scale Kalman Filtering" (Berberidis et al., 2016)
• "A quantum algorithm for the Kalman filter using block encoding" (Shi et al., 2024)
• "Interpretable Gradient Descent for Kalman Gain" (Belabbas et al., 18 Jul 2025)