Covariance Intersection for Consistent Fusion

• Covariance Intersection is a method for consistently fusing unbiased estimates when cross-correlation structures are unknown, ensuring conservative error bounds.
• The approach optimizes a fusion weight to minimize scalar error measures like trace or determinant, balancing accuracy and conservatism.
• Extensions such as Weighted, Split, and Overlapping CI enhance performance in nonlinear settings and distributed networks with sensor heterogeneity.

Covariance Intersection is a foundational method for the consistent fusion of estimates when the cross-correlation structure between errors is unknown or untracked. It is central to distributed estimation, decentralized sensor fusion, and large-scale cooperative systems where direct management of all cross-covariances is infeasible or impossible.

1. Formal Definition and Classical Fusion Rule

Covariance Intersection (CI) fuses two or more unbiased estimates of a common state $x \in \mathbb{R}^n$, denoted as $\hat{x}_i$ with covariances $P_i$, under the condition that cross-covariances $\mathbb{E}[(\hat{x}_i-x)(\hat{x}_j-x)^\top]$ are unknown. The aim is to produce a fused estimate that is guaranteed to be conservative (that is, with an error covariance that is an upper bound under any admissible cross-correlation).

Given two estimates $(\hat{x}_1,P_1)$ and $(\hat{x}_2,P_2)$, the CI fusion law is

$$P(\omega) = \bigl[\,\omega P_1^{-1} + (1-\omega) P_2^{-1}\,\bigr]^{-1},\qquad \hat{x}(\omega) = P(\omega)\,\bigl[\,\omega P_1^{-1} \hat{x}_1 + (1-\omega)P_2^{-1}\hat{x}_2\,\bigr],$$

for a fusion weight $\omega \in [0,1]$ (Trumpf et al., 22 Jul 2025, Chang et al., 2021, Li et al., 2024, Hu et al., 8 Apr 2025).

For multiple estimates, the information form generalizes as

$$P_{\text{CI}}^{-1} = \sum_{i} \omega_i P_i^{-1}, \qquad P_{\text{CI}}^{-1}\,\hat{x}_{\text{CI}} = \sum_i \omega_i P_i^{-1} \hat{x}_i, \quad \boldsymbol{\omega} \in \Delta^n,$$

where $\sum_i \omega_i = 1, \; \omega_i \ge 0$ (Chang et al., 2021, Hu et al., 2021, Tamjidi et al., 2016).

CI solves a minimax problem: find a fused covariance that upper-bounds the true fusion error for any admissible cross-covariance, and which minimizes a strictly isotone scalar measure (trace, log determinant, etc). This is essential in distributed environments where dependencies induced by shared process/measurement noise or past communication are untracked and may vary dynamically.

2. Optimization Criteria and Theoretical Guarantees

The fusion weight $\hat{x}_i$0 is chosen by minimizing a scalar functional: $\hat{x}_i$1 Both criteria are convex in $\hat{x}_i$2, ensuring a unique minimizer (Trumpf et al., 22 Jul 2025, Chang et al., 2021).

Key theoretical properties:

• Consistency: For any choice of $\hat{x}_i$3, $\hat{x}_i$4 is a valid upper bound: $\hat{x}_i$5, the true fusion covariance, for all admissible cross-covariances.
• Unbiasedness: The fused estimator is unbiased with respect to the original state.
• Optimality: Among all linear, unbiased, and conservative fusion rules, CI yields the minimal upper bound with respect to any strictly isotone cost (i.e., any increasing function in the positive semidefinite ordering) (Trumpf et al., 22 Jul 2025).
• Extremal Values: $\hat{x}_i$6 or $\hat{x}_i$7 correspond to adopting one of the estimates unchanged. For intermediate weights, the solution achieves the trade-off that is provably least conservative given the available information (Trumpf et al., 22 Jul 2025, Hu et al., 2021).

This optimality extends to partial-state fusion, fusing unbiased estimators of projections $\hat{x}_i$8, and generalizes to semidefinite-program-based formulations (Trumpf et al., 22 Jul 2025, Pedroso et al., 20 Mar 2026).

3. Practical Implementation and Extensions

Weight Selection and Algorithmic Structure

CI reduces the fusion problem to a one-dimensional optimization (for two estimates), or simplex-constrained optimization (for multiple estimates). Efficient closed-form surrogates for the weights exist (e.g., via normalized information or trace-inverse weights) (Li et al., 2024, Hu et al., 2021). For higher dimensions, line search, grid search, or standard convex optimization can be applied.

Batch (simultaneous) and enhanced sequential (block-wise or streaming) CI strategies have been developed, with structure-independent analytic weighting criteria ensuring that fuse order and grouping do not affect the final result (Hu et al., 2021). This resolves practical concerns in networks with unpredictable data arrival.

Extensions to Nonlinear Manifolds and RFS

CI has been embedded in invariant and unscented Kalman filters on Lie groups (e.g., for pose estimation on $\hat{x}_i$9), with fusion performed in the associated tangent (Lie algebra) space (Li et al., 2024, Ruan et al., 2024). Generalized CI (GCI) extends the method to random-finite-set densities for multi-object tracking, fusing representations such as Poisson multi-Bernoulli densities (García-Fernández et al., 23 Jun 2025, Li et al., 2019).

Weighted Covariance Intersection (WCI)

In complex, high-dimensional or heterogeneous state spaces (notably in 3D cooperative localization with inertial/attitude components), classical CI can suffer from scale imbalance, correlation mismatch, and lack of task relevance. Weighted Covariance Intersection (WCI) introduces a positive-definite weighting matrix $P_i$0 and minimizes $P_i$1, focusing fusion on the state components most relevant for task performance (e.g., position drift in inertial navigation) (Tu et al., 17 Aug 2025). WCI resolves the scale-sensitivity of CI and restores task-relevant error guarantees.

Modified and Split CI

In monitoring networks where sensors have heterogeneous reliability (e.g., non-homogeneous detection probabilities), Modified CI incorporates per-node performance metrics (e.g., detection probability, miss-detection growth) into the weighting (Daeichian et al., 2019). Split CI (SCI) further reduces conservatism by explicitly decomposing estimator errors into correlated (unknown cross-correlation) and uncorrelated components (e.g., independent measurement noise), resulting in less conservative fused bounds and provably optimal fusion under additional structural assumptions (Cros et al., 2023, Cros et al., 2024, Cros et al., 14 Jan 2025).

Extensions include ESCI (exploiting known correlated components such as shared process noise) with optimality guarantees for the minimal conservative fused bound (Cros et al., 14 Jan 2025, Cros et al., 2024).

4. Distributed Fusion, Scalability, and Robustness

CI is particularly prominent in distributed Kalman filtering, multi-agent cooperative localization, and decentralized tracking when maintaining the full cross-covariance matrix is infeasible due to communication or computation constraints.

• Communication Topologies: CI guarantees boundedness and consistency in distributed estimation provided weak connectivity or observability through local measurement graphs (Chang et al., 2021, Hu et al., 8 Apr 2025).
• Scalability: Classical centralized filtering scales cubically in network size; CI-based distributed filtering requires only local exchanges and computations, offering scalability to large swarms (Tu et al., 17 Aug 2025, Chang et al., 2021).
• Robustness: CI-based systems are robust to communication dropouts and agent failure; estimation remains stable, and reintegrated nodes do not inject inconsistent information to the network (Tu et al., 17 Aug 2025, Chang et al., 2021).
• Time-varying and Periodic Systems: The boundedness and convergence of CI-based filters extend to time-varying and periodic systems, with fixed-point and periodic Riccati solutions characterizing the steady-state behavior (Hu et al., 8 Apr 2025).

5. Generalizations: Partial Information and Overlapping Bounds

Recent work has unified CI and its variants into a family-optimal framework via semidefinite programming (SDP). Overlapping Covariance Intersection (OCI) accommodates partial and overlapping structural information on cross-covariances, parameterized via matrix inequalities and solved efficiently via SDP (Pedroso et al., 17 Mar 2026, Pedroso et al., 20 Mar 2026).

The general OCI problem is: $P_i$2 where $P_i$3 is constrained by partial (possibly overlapping) bounds on its principal submatrices. This subsumes CI, SCI, and ESCI as special cases, and enables real-time consistent fusion in large-scale systems using convex optimization tools (Pedroso et al., 17 Mar 2026, Pedroso et al., 20 Mar 2026).

6. Applications and Empirical Performance

CI and its extensions are fundamental in:

• Distributed cooperative localization for ground, aerial, underwater, or space swarms, especially under limited communication or when facing highly uncertain environments (Tu et al., 17 Aug 2025, Chang et al., 2021).
• Decentralized state estimation in large sensor networks, enabling robust and scalable fusion under network failures (Tamjidi et al., 2016).
• Multi-object and multi-sensor tracking, with GCI providing a mathematically principled fusion rule for random finite set filters, including Poisson multi-Bernoulli and related densities (García-Fernández et al., 23 Jun 2025, Li et al., 2019).
• Biomedical signal processing, e.g., robust fusion of respiratory rate estimates from heterogeneous sensor features (Zhang et al., 2020).

Empirical results consistently demonstrate that CI-based fusion yields bounded, consistent estimates under partial knowledge, with the price of slight conservatism relative to an oracle centralized filter, but with major gains in robustness and scalability. WCI, SCI, and OCI systematically reduce conservatism by exploiting task structure and partial correlation knowledge (Tu et al., 17 Aug 2025, Cros et al., 2024, Pedroso et al., 17 Mar 2026).

7. Limitations, Theoretical Boundaries, and Open Challenges

While CI is minimax-optimal given complete ignorance of cross-correlation, it is conservative; in the presence of structural knowledge (independent noise, common process noise), SCI and OCI outperform CI by delivering strictly tighter bounds (Cros et al., 2023, Cros et al., 14 Jan 2025, Pedroso et al., 17 Mar 2026). For some hybrid situations, no closed-form minimal Loewner bound exists, and family-optimality can only be realized via parameterized SDPs (Pedroso et al., 17 Mar 2026).

Remaining challenges include:

• Scalable distributed solution of SDPs in very large swarms,
• Automatic identification and exploitation of independence/correlation structure,
• Extending CI frameworks to nonlinear and non-Gaussian estimation, and
• Systematic task-driven weighting (such as WCI) for domain-specific performance optimization.

CI and its generalizations remain the cornerstone of provably consistent, robust, and scalable fusion in uncertain, distributed, and information-limited estimation problems.