---
title: Inverse Covariance Intersection (ICI) Overview
url: https://www.emergentmind.com/topics/inverse-covariance-intersection-ici
type: topic
---

# Inverse Covariance Intersection (ICI) Overview

Inverse Covariance Intersection (ICI) is a data fusion methodology for combining incompatible or correlatively ambiguous Gaussian estimates, ensuring estimation consistency and providing less conservatism than standard Covariance Intersection (CI). ICI is central to distributed filtering and estimation, particularly in sensor networks for both vector spaces and matrix Lie groups, as well as in advanced settings such as partial knowledge regimes (e.g., common information or common noise scenarios). The method enables robust fusion of information under unknown or only partially known cross-correlation, extending standard fusion techniques to contexts where rigorous consistency guarantees are required [2409.17997], [2506.05955].

## 1. Consistent Fusion under Unknown Correlations

ICI addresses the core challenge of combining two or more Gaussian estimates $x_1 \sim \mathcal{N}(x_1, P_1)$ and $x_2 \sim \mathcal{N}(x_2, P_2)$ whose cross-covariances are unknown, which precludes direct optimal linear fusion. Naive consensus approaches may yield fused covariances that underestimate the true post-fusion uncertainty, violating consistency. ICI constructs a fused Gaussian $\mathcal{N}(x_f, P_f)$ such that $P_f$ upper-bounds the actual error covariance for any admissible correlation, and aims to be less conservative than standard CI while remaining architecturally straightforward for distributed deployment.

ICI constructs the fused covariance as
$$
P_f = [\omega P_1^{-1} + (1-\omega) P_2^{-1}]^{-1},
$$
where $\omega \in [0,1]$ is a fusion weight satisfying $\omega \geq 0$, $1-\omega \geq 0$, and $\omega + (1-\omega) = 1$. The corresponding fused mean is
$$
x_f = P_f [\omega P_1^{-1} x_1 + (1-\omega) P_2^{-1} x_2].
$$
Weight selection relies on convex scalarizations, typically minimizing either $\operatorname{tr} P_f$ or $\log\det P_f$, resulting in a convex one-dimensional program in $\omega$ [2409.17997].


## 2. Weight Selection and Optimization Criteria

The effectiveness of ICI in balancing conservativeness and informativeness hinges on optimal selection of $\omega$. Canonical criteria are:
- **Minimize trace:** $\omega^* = \arg\min_{\omega \in [0,1]} \operatorname{tr}\{[\omega P_1^{-1} + (1-\omega) P_2^{-1}]^{-1}\}$
- **Minimize log-determinant:** $\omega^* = \arg\min_{\omega \in [0,1]} \log\det\{[\omega P_1^{-1} + (1-\omega) P_2^{-1}]^{-1}\}$

Both criteria yield a convex optimization problem that can be solved efficiently. These scalarizations enable implementers to tune fusion conservativeness or informativeness to application demands while preserving rigorous error bounds.

## 3. Extension to Matrix Lie Groups

In distributed estimation over matrix Lie groups (e.g., $\mathrm{SE}(3)$), ICI generalizes to handle manifold-valued state spaces and associated local uncertainties defined on Lie algebras. Each agent $i$ stores an estimate $\hat X_i \in \mathcal{G}$ and associated tangent-space covariance $P_i \in \mathbb{R}^{n \times n}$. The right-invariant local error is represented as $\xi_i$ via $\exp(\xi_i) = X_{\mathrm{true}} \hat X_i^{-1}$, with $P_i = \operatorname{Cov}(\xi_i)$. 

For a neighborhood $\mathcal{N}_i$, agents construct error-difference vectors 
$$
\Delta_{ji} = \log(\hat X_j \hat X_i^{-1}),
$$
which, under Baker–Campbell–Hausdorff (BCH) linearization, satisfy $\Delta_{ji} \approx \xi_i - \xi_j$, serving as noisy observations of $\xi_i$ with noise covariance $P_j$. Multi-fusion ICI operates in the Lie algebra:
$$
\begin{align*}
P_{\Gamma} & = \sum_{j \in \mathcal{N}_i} \omega_{ij} P_j,\quad \sum \omega_{ij} = 1,\, \omega_{ij} \ge 0, \\
\hat P_i & = \Bigl[ \sum_{j \in \mathcal{N}_i} P_j^{-1} - (|\mathcal{N}_i|-1) P_\Gamma^{-1} \Bigr]^{-1},\\
\Gamma_i & = \sum_{j \in \mathcal{N}_i} [P_j^{-1} - (|\mathcal{N}_i|-1) \omega_{ij} P_\Gamma^{-1}] \Delta_{ji}, \\
\hat\xi_i & = \hat P_i \Gamma_i,\\
\hat X_i^+ & = \exp(\hat \xi_i) \hat X_i.
\end{align*}
$$
This Lie group ICI extension enables consistent uncertainty fusion in a broad class of geometric estimation problems and supports integration with invariant filtering architectures [2409.17997].

## 4. Integration with Distributed Filtering and the Invariant UKF

ICI is naturally incorporated in distributed sensor network filters. In the diffusion-based distributed invariant Unscented Kalman Filter (DIUKF-ICI), each agent performs the following per timestep:
- **A:** Local UKF propagation and measurement update yielding $(\bar X_i, \bar P_i)$.
- **B:** Incremental update incorporating neighbor measurements $(\check X_i, \check P_i)$.
- **C:** Diffusion-fusion via extended ICI: exchange $(\check X_j, \check P_j)$; compute $\Delta_{ji}$; solve convex program for $\omega_{ij}$; compute $(\hat P_i, \hat \xi_i)$ via ICI formulas; update $\hat X_i = \exp(\hat \xi_i) \check X_i$, $\hat P_i = \hat P_i$.

This structure is robust to intermittent measurements and time-varying communication topologies. Agents omit non-participating neighbors or assign vanishing weights to estimates with infinite (or very large) covariances. All fusion and weight constraints are localized to the current neighborhood, ensuring consistency across dynamic networks [2409.17997].

## 5. Dual ICI and Partial Knowledge Models

Beyond total ignorance of inter-estimate correlation, ICI admits dual formulations that exploit partial prior knowledge, notably the **common-noise** regime. For two unbiased estimates $\hat y_1, \hat y_2$ of $y$ with marginal covariances $P_1$, $P_2$ and a joint covariance $P_{12} = \begin{pmatrix} P_1 & X \\ X & P_2 \end{pmatrix}$ where $X = X^T,\, 0 \preceq X \preceq P_i$, dual-ICI constructs an upper bound $M \succeq P_{12}$ and applies fusion with weights enforcing $W_1 + W_2 = I$.

The upper bound is parameterized as
$$
M = \operatorname{diag}(P_1, P_2) + \frac{\omega}{2} \begin{pmatrix} I \\ I \end{pmatrix} B (I, I),
$$
where $B = (\alpha P_1^{-1} + (1-\alpha) P_2^{-1})^{-1},\, \alpha \in [0,1]$. The fused covariance bound is
$$
P_F = [w P_1^{-1} + (1-w) P_2^{-1}]^{-1}
$$
with $w = (1 + 1/\omega)^{-1}$. For the optimal $\alpha = 1/2$, $P_F$ coincides with the standard CI bound. 

Significantly, although the union of dual-ICI bounds across admissible $(\alpha, \omega)$ strictly improves the union of CI bounds with suboptimal weights, no single dual-ICI bound outperforms the CI bound at its optimal $w$ [2506.05955]. A plausible implication is that partial knowledge models can be exploited to tighten covariance bounds in multi-criteria or adaptive settings, but the gain is not realized in standard scalarized optimal fusion.

## 6. Properties, Guarantees, and Practical Considerations

ICI ensures the fused covariance is always a valid upper bound: $P_f \succeq \operatorname{Cov}(x_f\,|\, \text{all data})$. For the distributed Lie group setting, Theorem 1 of [2409.17997] establishes that, under uniform bounds on local covariances and measurement noise, the sequences of local and fused covariances remain uniformly bounded, guaranteeing mean-square bounded error trajectories.

Algorithmic robustness to intermittent observations is maintained by omitting absent neighbors or assigning vanishing fusion weights, while time-varying network topologies are handled via dynamic neighborhoods $\mathcal{N}_i^t$ and properly adjusted weight constraints. The convexity of the weight selection criteria simplifies distributed implementations.

## 7. Applications and Limitations

ICI underlies robust distributed state estimation in sensor networks, particularly where states live on Lie groups (e.g., target tracking in 3D environments). Its deployment in the DIUKF-ICI achieves bounded estimation error under extensive Monte-Carlo testing, even with communication failures and sporadic sensing [2409.17997]. 

In the context of partial correlation knowledge, dual-ICI offers a family of tighter upper bounds for regimes such as common-noise, though improvements are realized only in the ensemble of possible bounds or for suboptimal fusion weights, not at the single optimal scalarized bound [2506.05955]. Extending dual-ICI to more agents, partial state measurements, and dynamic fusion remains an active research direction.

Source: https://www.emergentmind.com/topics/inverse-covariance-intersection-ici