Quadrature-Based Filtering

• Quadrature-based filtering is a technique that leverages high-order quadrature rules to approximate integrals in Bayesian state-space inference, enabling accurate estimation in nonlinear systems.
• It employs multi-point weighted sums using methods like Gauss–Hermite, Unscented Transform, and cubature rules to achieve polynomial exactness and manage numerical instability.
• This approach is widely applied in navigation, sensor fusion, quantum filtering, and probabilistic ODE/PDE solvers, with ongoing advancements in adaptivity and error quantification.

Quadrature-based filtering refers to a suite of deterministic, high-order numerical filtering methodologies that employ accurate multi-point weighted summation rules—quadrature formulas—for evaluating otherwise intractable integrals in Bayesian state-space inference, moment transforms, and signal estimation. These methods are foundational in both classical and quantum filtering, and their theoretical and algorithmic implications span nonlinear state-space models, probabilistic numerics, spectral methods, and quantum measurement theory.

1. Mathematical Foundations and Core Quadrature Rules

Quadrature-based filtering is predicated on the necessity to approximate integrals of the form

$I[f] = \int f(x) p(x) dx,$

where $p(x)$ is often a Gaussian, Student-t, or other known measure, and $f$ is the transformation or observation function. In practice, such integrals arise in the recursive prediction and update steps of nonlinear Bayesian filters:

• For prediction, the mean and covariance of the state after propagating through a nonlinear transition.
• For update, the marginalization over latent variables in the presence of nonlinear observations.

Key quadrature rules include:

• Gauss–Hermite Quadrature: Exact for polynomials up to degree $2m-1$ in one dimension using $m$ nodes. Multivariate extension via tensor-product incurs exponential growth in point count (the "curse of dimensionality") (Kokkala et al., 2015).
• Unscented Transform (UT): A third-order (and higher order generalizations) rule utilizing so-called sigma-points, widely used in Unscented Kalman Filters and their variants (Kokkala et al., 2015).
• Fully Symmetric Cubature: Higher-order symmetric rules (orders 5, 7, 9) enabling polynomial-degree control at polynomial-in-dimension point counts.
• Student-t Quadrature: Extends to heavy-tailed marginals via specially constructed sigma-points and kernel-weighted sums (Prüher et al., 2017).
• Gaussian Process Quadrature (GPQ): Treats the integrand as a GP, yielding both an integral estimate and a principled estimate of the quadrature error (Prüher et al., 2017).
• Moment-Based Quadrature: Uses orthonormal polynomials derived from raw moments to deterministically construct nodes and weights (Zhao et al., 2023).

The universal paradigm is to replace the integral by a weighted sum over function evaluations at deterministic points: $\int f(x) p(x) dx \approx \sum_{i=1}^N w_i f(x_i),$ where both the $x_i$ ("nodes" or "sigma-points") and weights $w_i$ are designed to guarantee exactness for classes of functions and to control numerical instability.

2. Algorithms and Recursion in Nonlinear State-Space Models

Quadrature-based filtering methods can be characterized by the filter recursion equations where closed-form map propagation is replaced by quadrature-based expectations: $$\begin{aligned} m_{k|k-1} &= \sum_{i=1}^N w_i f(m_{k-1|k-1} + L_{k-1|k-1} \xi_i), \ P_{k|k-1} &= \sum_{i=1}^N w_i [f(\cdot) - m_{k|k-1}][f(\cdot) - m_{k|k-1}]^T + Q, \end{aligned}$$ with analogous forms for the measurement update. Here, $\{ \xi_i, w_i \}$ are canonical sigma-points and weights, $L$ is a square root of the covariance, and $Q$ is the process noise.

Algorithmic choices dictate the numerical and statistical performance:

• Gauss–Hermite-based Filters (QKF): Provide optimal polynomial exactness but with computational complexity $O(m^n)$ (Kokkala et al., 2015, Singh et al., 2023).
• Sigma-Point Filters (e.g., UKF): Efficient linear or quadratic scaling, with controlled polynomial degree; extensions to higher orders exist (Kokkala et al., 2015).
• Cubature Kalman Filter (CKF): Spherical-radial rules providing quadratic point growth and improved accuracy over UT (Singh et al., 2023).
• Gaussian Sum Quadrature Filters: For models with multimodal posteriors (e.g., Wiener systems), Gauss-Legendre quadrature combined with Gaussian mixture updating enables non-Gaussian state propagation (Cedeño et al., 13 May 2025).

Moment-based quadrature filters recursively propagate and update low-order moments, reconstructing deterministic quadrature rules for each step (Zhao et al., 2023). For heavy-tailed or non-Gaussian noise, Student-t process quadrature propagates higher moments and assesses numerical integration error via an explicit variance term (Prüher et al., 2017).

3. Error Characterization and Bayesian Quadrature Extensions

Classical quadrature rules afford no built-in mechanism to quantify the approximation error. Bayesian quadrature (BQ) and its specializations (GPQ, TPQ) model the integrand as a random process—classically a GP or Student-t process—and propagate both the mean and variance of the resulting integral estimate: $\operatorname{Var}[I \mid y] = c - w^T K w,$ where $K$ is the GP/TP covariance matrix over nodes, $w$ is computed from the prior kernel and measure, and $c$ is an integrated kernel mean (Prüher et al., 2017).

The GPQ and TPQ methodologies permit adaptive quantification of quadrature uncertainty, leading to more conservative and credible estimation, notably outperforming traditional quadrature-based moment transforms and Kalman filters in RMSE and negative log-likelihood metrics (Prüher et al., 2017, Prüher et al., 2017).

4. Extensions: Stable Filtering, ODE Solvers, and Spectral Methods

Discontinuous Galerkin and Spectral Filtering

Quadrature-based filtering underpins stabilization and modal control in polynomial-based (e.g., DG) discretizations for PDEs. The filter operator is constructed via high-order accurate quadrature (e.g., Legendre–Gauss–Lobatto), leading to high-frequency attenuation, contractivity (semi-discrete energy bounds), and provable long-time stability (via modal cutoff and transmission problem contractivity) (Nordström et al., 2020).

Probabilistic Numerics and ODE Filters

Filtering-based probabilistic ODE solvers utilize quadrature-based prediction steps to propagate parameter and numerical solver uncertainty. Marginalization over uncertain parameters is handled via quadrature rules; discretization error is accounted for through additive process noise, leading to robust and calibrated uncertainty quantification in the state and parameters of the dynamical system (Yao et al., 6 Mar 2025).

5. Quantum and Stochastic Quadrature Filtering

Continuous-time Quantum Filtering

In continuous quantum measurement, quadrature-based filters arise in the Belavkin–Stochastic master equation, where the system evolution is conditioned on the homodyne record (field quadrature). Advanced constructions accommodate thermal and squeezed inputs via Araki–Woods representations, Bogoliubov transformations, and Tomita–Takesaki theory, ensuring the filter equations are representation-independent and model both phase-insensitive and phase-sensitive (squeezed) noise contributions (Gough, 6 Oct 2025, Gough et al., 18 Jan 2026).

Single-photon and Multi-channel Quantum Filters

Single-photon filtering with diffusive quadrature measurements (Q–P, Q–Q forms) exploits multiple output channels, each corresponding to distinct quadratures. The associated stochastic master equations yield optimal excitation performance—maximally energizing a two-level system—by exploiting the full measurement record, with back-action effects quantifiable and controllable (Dong et al., 2018).

6. Practical Applications, Limitations, and Empirical Comparisons

Quadrature-based filtering methods are routinely utilized in diverse domains:

• Nonlinear target tracking, navigation, and sensor fusion (Kokkala et al., 2015).
• State and parameter estimation in control systems, block-oriented nonlinear models, and model-based reinforcement learning (Cedeño et al., 13 May 2025).
• Model uncertainty propagation in scientific computing ODE and PDE solvers (Yao et al., 6 Mar 2025, Occorsio et al., 2021).
• Quantum state estimation, quantum feedback control, and squeezing-based measurement in quantum optics (Gough, 6 Oct 2025, Gough et al., 18 Jan 2026, Horrom et al., 2012).

A summary comparison (see Table):

Filter	Complexity	Integrals Exactly	Polynomial Order (n-dim)	Posterior Type
EKF	O(n³)	1st (linear)	1	Single Gaussian
UKF3	O(n)	≤3	3	Single Gaussian
CKF (order 3)	O(n)	≤3	3	Single Gaussian
GH Quadrature	O(mⁿ)	≤2m–1	2m–1	Single Gaussian
GPQ/TPQ	O(N³)	(data-adaptive)	- (GP-based)	Gaussian + variance
Moment-Quad	O(N³) (uni)	≤2N–1	2N–1	Deterministic approx
Gaussian Sum	O(KM)	(mixture)	(via mixture)	Gaussian sum
Particle Filter	O(M)	As M→∞	∞ (MC convergence)	Arbitrary

Key practical guidance:

• For low-dimensional problems requiring high accuracy, use Gauss–Hermite or higher-order symmetric cubature (Kokkala et al., 2015, Singh et al., 2023).
• For moderate dimensions or real-time constraints, fully symmetric (UKF5, UKF7) or cubic rules are computationally efficient and relatively accurate.
• For heavy-tailed or multimodal posteriors, GPQ, TPQ, and moment-based deterministic quadrature methods provide convergence guarantees and enhanced error quantification (Zhao et al., 2023, Prüher et al., 2017).
• In quantum systems, representation-independent quadrature filters account for both phase-insensitive and phase-sensitive (squeezing) noise (Gough et al., 18 Jan 2026).

Limitations include computational scaling (tensor-product rules scale exponentially), control of numerical instabilities (ill-conditioned moment systems), and the restriction to a class of posteriors (filters relying on Gaussian or low-order moment closure can fail for non-Gaussian, multimodal, or strongly nonlinear evolutions).

7. Recent Developments and Future Directions

Current frontiers include:

• Adaptive and Bayesian quadrature under active node selection, hyperparameter learning, and robust error bounds (Prüher et al., 2017, Prüher et al., 2017).
• Spline, polynomial, or kernel-based moment closure to overcome the curse of dimensionality and ill-conditioning (Zhao et al., 2023).
• Integration with probabilistic programming and variational inference for parameter and structural uncertainty.
• Quantum filters capable of leveraging squeezed state resources and exploiting measurement-induced back-action for enhanced quantum control (Gough et al., 18 Jan 2026, Gough, 6 Oct 2025, Horrom et al., 2012).
• Deep integration of quadrature methods in stable high-order ODE and PDE solvers, including hybrid particle-quadrature methods for scientific and engineering simulation (Yao et al., 6 Mar 2025, Occorsio et al., 2021).

The systematic adoption of quadrature-based filtering, in both classical and quantum domains, continues to provide a bridge between efficient, scalable algorithms and near-optimal estimation accuracy across a wide variety of high-dimensional and strongly nonlinear inference tasks.