Lie-Group-Based Propagation

• Lie-group-based propagation is a mathematical framework that employs the structure of Lie groups and algebras to design state-update algorithms which preserve manifold constraints and physical symmetries.
• It utilizes integrators such as Runge-Kutta-Munthe-Kaas and commutator-free schemes to achieve high-order accuracy and structure preservation in both deterministic and stochastic contexts.
• In practical applications like robotics, mechanical systems, and neural computation, these methods enable effective uncertainty quantification, Bayesian fusion, and generative modeling on manifold-constrained state spaces.

Lie-group-based propagation refers to computational and statistical methods that exploit the structure of Lie groups for modeling, integration, uncertainty quantification, learning, control, and generative modeling of systems whose states, dynamics, or intermediate representations evolve on group manifolds. Lie-group-based propagation differs fundamentally from classical Euclidean approaches in that updates, uncertainty flows, and transformations are implemented via group actions, Lie algebra maps, and intrinsic geometric structure, ensuring that the evolving state or distribution remains consistent with manifold constraints and physical symmetries.

1. Mathematical Foundations of Lie-Group-Based Propagation

A Lie group $G$ is a smooth manifold endowed with compatible group structure. Its associated Lie algebra $\mathfrak{g}$ is the tangent space at the identity element, equipped with a bracket $[\cdot,\cdot]$. The exponential map $\exp: \mathfrak{g} \to G$ provides local coordinates for propagation on $G$, while the logarithm map $\log: G \supset U \to \mathfrak{g}$ inverts $\exp$ near the identity.

In computational propagation, group-valued states $g(t) \in G$ evolve according to differential or stochastic differential equations (SDEs) of the form

$g^{-1}dg = h(g,t)\,dt + H(g,t)\,dW_t,$

where $h$ and $\mathfrak{g}$0 are adapted vector fields in the algebra, and $\mathfrak{g}$1 is (possibly multidimensional) Brownian noise. Such equations generalize standard Euclidean SDEs to the group context and are naturally compatible with manifold constraints, group symmetries, and physical invariances (Ye et al., 2024, Ye et al., 2023, Celledoni et al., 2021).

The exponential and adjoint actions, together with left/right Jacobians ($\mathfrak{g}$2, $\mathfrak{g}$3), are central for converting between local (algebraic) coordinates and group-valued increments, enabling both deterministic and stochastic propagation schemes.

2. Lie-Group Integrators and Propagation Algorithms

Lie-group-based propagation leverages several classes of numerical integrators:

• Runge-Kutta-Munthe-Kaas (RKMK) Integrators: These methods pull back ODEs on $\mathfrak{g}$4 to $\mathfrak{g}$5 via exponential coordinates, apply classical Runge-Kutta steps in the algebra, and map the solution forward via $\mathfrak{g}$6. Each step automatically respects the manifold structure. The update takes the form $\mathfrak{g}$7, with $\mathfrak{g}$8 increments determined by algebraic equations involving dexp and its inverse (Celledoni et al., 2021, Bogfjellmo et al., 2013).
• Commutator-Free Integrators: These propagate via sequences of exponentials over frozen or linearly combined algebra elements, minimizing commutator computations and enabling higher-order accuracy (Celledoni et al., 2021).
• Symplectic and Variational Lie-Group Integrators: For Hamiltonian systems on $\mathfrak{g}$9, these schemes ensure symplecticity and conservation of first integrals by variational discretization, usually through structure-preserving updates of $[\cdot,\cdot]$0 (Bogfjellmo et al., 2013, Celledoni et al., 2021).

Classical Lie-group-based propagation maintains group-invariant constraints (e.g., orthogonality for $[\cdot,\cdot]$1), provides global coordinate-free updates, and permits high-order accuracy and structure preservation in both deterministic and stochastic contexts.

3. Stochastic and Uncertainty Propagation on Lie Groups

Group-valued SDEs and their associated propagation laws enable natural uncertainty quantification on manifolds:

• Parametric SDE Framework: The McKean–Gangolli injection method expresses the SDE on $[\cdot,\cdot]$2 in exponential coordinates, using right or left Jacobians to obtain Itô or Stratonovich forms for $[\cdot,\cdot]$3 where $[\cdot,\cdot]$4. This coordinate transformation, together with appropriate mean and covariance fits, enables closed-form moment propagation and Bayesian fusion for concentrated distributions (Ye et al., 2024, Ye et al., 2023).
• Fokker–Planck Approach: The evolution of probability densities $[\cdot,\cdot]$5 on $[\cdot,\cdot]$6 is governed by a Fokker–Planck PDE in bi-invariant or right/left-invariant form. Exact ODEs for the mean $[\cdot,\cdot]$7 and covariance $[\cdot,\cdot]$8 are derived via moment closure, recentered at the group mean, with approximations available via Gaussian quadrature (unscented transforms) or small-error (Taylor) expansions (Ye et al., 2023).
• Kalman-Type/Extended Bayesian Fusion: For Bayesian filtering and fusion, innovations and gains are computed in the exponential chart, and updates are retracted/projection-mapped through the exponential and left Jacobian back onto $[\cdot,\cdot]$9 (Ye et al., 2024).

These propagation schemes are essential for nonlinear filtering, sensor fusion, and uncertainty-aware control in mechanical systems, robotics, and state estimation where configurations naturally live in $\exp: \mathfrak{g} \to G$0, $\exp: \mathfrak{g} \to G$1, or more general $\exp: \mathfrak{g} \to G$2.

4. Generative Modeling and Score-Based Diffusions on Lie Groups

Lie-group-based propagation underpins advanced generative models, generalizing Euclidean score-based diffusion processes to arbitrary Lie groups:

• Generalized Score Matching: For a group $\exp: \mathfrak{g} \to G$3 acting smoothly on $\exp: \mathfrak{g} \to G$4, fundamental vector fields $\exp: \mathfrak{g} \to G$5 corresponding to Lie algebra generators define a Lie-group "gradient" operator,

$\exp: \mathfrak{g} \to G$6

which yields a generalized Fisher divergence as the loss for models $\exp: \mathfrak{g} \to G$7. The forward (noising) and reverse (sampling) SDEs on $\exp: \mathfrak{g} \to G$8 incorporate group curvature and drift via the Casimir operator and decompose updates along Lie-algebra directions (Bertolini et al., 4 Feb 2025).

• Exactly Solvable Propagation of Flow Coordinates: In the flow-coordinates (latent $\exp: \mathfrak{g} \to G$9), the forward SDE becomes tractable and Gaussian. The reverse SDE implements group-aware denoising Langevin dynamics, and the Euler-integrated updates are aligned with group actions via the exponential map.
• Dimension Reduction and Symmetry Adaptation: By choosing $G$0 to reflect the natural symmetries of the data (e.g., $G$1 for molecular conformers, $G$2 for rigid-body docking), the space of learned score-components is drastically reduced, sampling is more physically faithful (e.g., enforcing chemical constraints), and generative processes interpolate between complex data distributions in a manifold-coherent manner (Bertolini et al., 4 Feb 2025).

5. Learning and Neural Architectures on Lie Groups

Recent work generalizes propagation mechanisms to learning scenarios where parameters, features, or hidden states evolve on or are constrained by Lie groups:

• Lie-Group Bayesian Learning Rule (BLR): The BLR parameterizes posteriors through group actions on a base distribution and updates via group exponentials, enabling natural-gradient learning that remains fully on manifold. The update follows $G$3, sidestepping the need for exponential-family constraints and ensuring geometric consistency (Kıral et al., 2023).
• Neural Operators with Lie-Group Manifold Constraints: Neural operator backbones (e.g., FNO) are augmented with "manifold-constrained layers" that update latent features via low-rank Lie-algebra actions followed by group exponentiation (e.g., $G$4 or $G$5). This enforces near-isometric, structure-preserving evolution of features, improving long-range stability and accuracy in PDE simulation tasks (Zhang et al., 18 Feb 2026).
• Block-wise Lie-Group Embeddings in Neural Dynamics: Propagation of neural activations or weights incorporates adjoint group actions and block-wise constraints. Algorithmic updates are performed via manifold projection and exponential/retraction, enabling stable, learnable neural flows on general $G$6 (e.g., $G$7), with provable equilibrium and stability guarantees (Wang et al., 24 May 2026).
• Path Development and RNNs: Recurrent neural architectures leveraging path development on Lie groups propagate sequence features by left-multiplying by group elements computed via matrix exponentials of algebra-weighted inputs. These layers mitigate gradient explosion/vanishing and provide data-efficient, robust sequence modeling (Lou et al., 2022).

6. Consensus, Synchronization, and Distributed Control on Lie Groups

Multi-agent consensus and synchronization over Lie groups exploit bi-invariant metrics to generalize Laplacian flows and coupling laws to manifold-valued states:

• Geodesic Error Potentials: Pairwise geodesic distance on $G$8 is measured via $G$9, and the consensus potential is $\log: G \supset U \to \mathfrak{g}$0.
• Gradient Dynamics: Agent propagation law is given by group-invariant gradient descent,

$\log: G \supset U \to \mathfrak{g}$1

which coincides with vector Laplacian or Kuramoto oscillators as special cases. For second-order systems, algebraic damping is added.

• Convergence Properties: Lyapunov/LaSalle arguments extend to the group manifold, with the structure of the coupling graph and group topology determining synchronization rates and invariant sets. Almost-global convergence to consensus is typical on compact or bi-invariant $\log: G \supset U \to \mathfrak{g}$2 (Chandrasekharan et al., 2022).

7. Practical Applications and Implementation Considerations

Lie-group-based propagation methods have demonstrated significant impact in:

• Mechanical and Multibody Dynamics: Structural preservation, global constraint enforcement, and symplectic integration for complex systems with $\log: G \supset U \to \mathfrak{g}$3, $\log: G \supset U \to \mathfrak{g}$4, and related symmetry groups (Celledoni et al., 2021, Bogfjellmo et al., 2013).
• Molecular Modeling and Structural Biology: Efficient generative models for conformer generation and ligand docking with strict symmetry constraints and reduced latent space (Bertolini et al., 4 Feb 2025).
• Robotics and Control: State estimation, filter fusion, and controlled actuation on $\log: G \supset U \to \mathfrak{g}$5 manifolds, with closed-form uncertainty propagation and group-theoretic Bayesian updates (Ye et al., 2024, Ye et al., 2023, Wang et al., 24 May 2026).
• Neural Computation and Learning: Stable, symmetry-aligned neural operator and RNN architectures for scientific computing, time series, and physical systems (Lou et al., 2022, Zhang et al., 18 Feb 2026, Wang et al., 24 May 2026).

Implementation requires efficient computation of $\log: G \supset U \to \mathfrak{g}$6, $\log: G \supset U \to \mathfrak{g}$7, derivatives of the exponential map (dexp), adjoint actions, and Jacobians for the group at hand (e.g., Rodrigues’ formula for $\log: G \supset U \to \mathfrak{g}$8), as well as batch projection and retraction algorithms for manifold-constrained learning. For stochastic and uncertainty propagation, practical choices between sigma-point integration, small-error Taylor expansion, or full moment closure depend on accuracy and computational resources (Ye et al., 2023, Ye et al., 2024).

---

References:

• (Bertolini et al., 4 Feb 2025): Generative Modeling on Lie Groups via Euclidean Generalized Score Matching
• (Kıral et al., 2023): The Lie-Group Bayesian Learning Rule
• (Bogfjellmo et al., 2013): High order symplectic partitioned Lie group methods
• (Ye et al., 2024): Uncertainty Propagation and Bayesian Fusion on Unimodular Lie Groups from a Parametric Perspective
• (Ye et al., 2023): Uncertainty Propagation on Unimodular Matrix Lie Groups
• (Celledoni et al., 2021): Lie Group integrators for mechanical systems
• (Lou et al., 2022): Path Development Network with Finite-dimensional Lie Group Representation
• (Zhang et al., 18 Feb 2026): Geometric Neural Operators via Lie Group-Constrained Latent Dynamics
• (Wang et al., 24 May 2026): Planning Neural Dynamics with Lie Group Embedding through Supervised Projective Manifold Learning
• (Chandrasekharan et al., 2022): A Unified Framework for Consensus and Synchronization on Lie Groups admitting a Bi-Invariant Metric