State Estimation on Non-Euclidean Manifolds

37 concepts24 papers1 proposal

In classical state estimation, we treat the world as a flat, Euclidean vector space. We track positions, velocities, and errors by adding and subtracting vectors. However, many real-world systems do not live in flat spaces. A robot's orientation lives on a sphere or a rotation group ($SO(3)$); a camera's trajectory moves through a space of rigid-body transformations ($SE(3)$); and simple variables like a wind direction or a radar phase live on a circle ($S^1$).

When we force these curved, periodic, or constrained spaces into flat Kalman filters or particle filters, the mathematics breaks down. This article explores why standard filters fail on non-Euclidean manifolds and how geometry-aware state estimation fixes them by operating natively within the true geometry of the state space.

The Failure of Flat Assumptions

To understand why standard filters fail, consider tracking a vehicle's heading on a circle. If the vehicle is facing $359^\circ$ and turns slightly clockwise to $1^\circ$, a standard linear filter attempting to compute an average or a state update might calculate $\frac{359 + 1}{2} = 180^\circ$. Instead of a minor adjustment, the filter points the vehicle in the exact opposite direction.

As noted by Gerhard Kurz et al. in Recursive Bayesian filtering in circular state spaces, standard linear operations fail because high uncertainty causes probability mass to wrap around the circular manifold. Simply applying a modulo operation to a Euclidean filter does not fix the underlying statistical distortion.

This problem compounds when tracking 3D rotations or full 6-DoF poses. If you add a perturbation vector directly to a rotation matrix, the resulting matrix is no longer orthogonal—the state "drifts" off the valid manifold into unphysical space. Historically, engineers bypassed this by using local coordinate charts or constantly re-normalizing the state.

However, Benjamin Busam et al. demonstrate in Camera Pose Filtering with Local Regression Geodesics on the Riemannian Manifold of Dual Quaternions that linear operations like moving averages "cut corners" across curved spaces. This introduces coordinate-induced errors and instabilities, especially when tracking rapid or large rotations. Early foundational work by Stefano Soatto et al. in Motion Estimation on the Essential Manifold highlighted that avoiding these Euclidean coordinate models prevents the convergence issues and instabilities typical of standard tracking frameworks.

Linear Average (Cuts Corners)
      o--------x--------o
     /  \       |       /  \
    /     \     |      /     \
   /        \   v     /        \
  o          \       /          o
  True Manifold (Curved Surface)

Intrinsic Geometry and Lie Groups

Instead of treating a manifold as a subset of a larger flat space and constantly dragging the state back down, intrinsic filtering operates entirely within the manifold. For spaces like $SO(3)$ or $SE(3)$, this is achieved using Matrix Lie Groups. A Lie group is a smooth manifold that also possesses a group structure, meaning we can move along the surface using group multiplication rather than vector addition.

To perform calculus on these curved surfaces, we utilize the tangent space—a flat vector space that locally approximates the manifold at a specific point. The exponential map ($\exp$) takes a vector from this flat tangent space (such as a local velocity or an uncertainty perturbation) and wraps it perfectly along a straight path (a geodesic) on the curved manifold. Conversely, the logarithmic map ($\log$) pulls a point on the manifold back to the local tangent space.

                 Tangent Space (Flat Vector Space)
                       ^
                       |  Perturbation / Velocity
                       |
                       +----->
                      /
                     /  Exponential Map (exp)
                    /
                   v
              .---------.
             /   True    \
            /  Manifold   \
            '-------------'

By leveraging this structure, Frank Chongwoo Park et al. established in Particle filtering on the Euclidean group: framework and applications that incorporating Lie group geometry yields coordinate-invariant filters. These filters maintain mathematical validity without ad-hoc normalization.

Similarly, Chi Zhang et al. extended this principle to continuous particle filtering in Feedback particle filter on matrix lie groups, proving that respecting the Lie group structure prevents particles from drifting into invalid configurations, even under severe sensor noise.

Moving Covariances via Parallel Transport

In a standard Extended Kalman Filter (EKF), when the state estimate is updated, the uncertainty covariance matrix remains in the same global coordinate frame. On a curved manifold, however, the local coordinate system changes as you move. If you update your position on a sphere, "North" and "East" point in entirely different directions than they did at your previous position. If you do not shift your covariance matrix along with your state, your description of uncertainty becomes misaligned.

To track uncertainty accurately, filters must use an affine connection to perform parallel transport. Parallel transport slides a vector or a covariance matrix along a geodesic curve without changing its intrinsic length or orientation relative to the manifold.

       Covariance at State A
            [===]
             /
            /  Parallel Transport along Geodesic
           v
         [===]  Covariance properly rotated to match State B

In A Note on the Extended Kalman Filter on a Manifold, Pieter van Goor et al. demonstrate that an affine connection is the minimal geometric structure required to implement a high-performance filter on a manifold. When the filter updates its estimate, the local coordinate system shifts. By using parallel transport to mathematically transport the uncertainty covariance to the new reference point, they drastically reduce linearization errors.

The authors expand on this in Equivariant Filter Design for Discrete-time systems, showing that treating the covariance reset step as a parallel transport operation preserves the integrity of the uncertainty description. This geometric consistency prevents the information loss typical of coordinate shifts, yielding faster convergence and higher steady-state accuracy.

Geometry-Aware Distributions and Inference

A standard Gaussian distribution extends infinitely in all directions. On a compact or periodic manifold, this assumption introduces immediate errors. To resolve this, researchers design filters around probability distributions native to curved spaces.

In Unscented Orientation Estimation Based on the Bingham Distribution, Igor Gilitschenski et al. replace Gaussian noise with the Bingham distribution. The Bingham distribution naturally models antipodal symmetry on hyperspheres (where a quaternion $q$ and $-q$ represent the exact same 3D orientation), eliminating the distortions that occur when mapping Euclidean transitions onto rotations. For planar rigid-body motions, the authors utilize dual quaternions on the $SE(2)$ manifold in A stochastic filter for planar rigid-body motions to achieve closed-form measurement updates without linearization errors.

This geometric awareness also applies to optimization and variational inference:

Real-World Impact and Tooling

Operating natively on manifolds provides significant performance gains in practical engineering applications. In legged robotics, Tzu-Yuan Lin et al. utilize an Invariant Extended Kalman Filter (InEKF) in Legged Robot State Estimation using Invariant Kalman Filtering and Learned Contact Events. By exploiting the Lie group structure of the robot's pose and velocity, the filter maintains geometric consistency and tracks the robot's state reliably in environments where visual SLAM fails entirely.

Application Domain Manifold Structure Key Geometric Mechanism Benefit
Legged Robotics (Tzu-Yuan Lin et al.) Lie Groups ($SE(3)$) Invariant EKF updates Geometric consistency, robust tracking under impact
Medical Imaging (Yuanxiang Wang et al.) Product Manifold (Shape + Orientation) Karcher means & manifold covariances Symmetric positive-definite tensor preservation
Continuous Tracking (Anna Kutschireiter et al.) Circular Manifolds ($S^1$) Projection filtering (circKF) Analytical tractability, $10\times$ faster than particle filters

To bridge the gap between abstract differential geometry and practical engineering, Gerhard Kurz et al. developed libDirectional, detailed in Directional Statistics and Filtering Using libDirectional. This framework provides native support for recursive estimation on circles, tori, and hyperspheres, allowing engineers to implement geometry-aware filters without deriving the underlying manifold calculus from scratch.

Key Foundational Work

Particle filtering on the Euclidean group: framework and applications (2007) — contributed 2 concept elements to this direction.
Motion Estimation on the Essential Manifold (1994) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_manifold_constrained_state_estimation_3d0e08ff,
  title = {manifold-constrained state estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manifold-constrained-state-estimation/rdr_3d0e08ff28b34efd9a40423658c29f38},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_manifold_constrained_state_estimation_3d0e08ff_v1,
  title = {manifold-constrained state estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manifold-constrained-state-estimation/rdr_3d0e08ff28b34efd9a40423658c29f38/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake