🔢 mathematics

Improved Representation of Matrix Lie Group Operations through Tensor Notation

This paper introduces a novel application of tensor and Einstein summation notation to matrix Lie groups, providing a clearer and more perspicuous mathematical framework for computing derivatives and operations in gradient-based estimation problems.

Original authors: Clark Taylor

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Clark Taylor

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Navigating a Curved World

Imagine you are trying to navigate a spaceship. You need to know exactly where you are and which way you are facing. In the world of robotics and space travel, "facing" (rotation) is tricky.

Usually, when we do math, we draw things on flat paper (a flat grid). But rotations don't live on flat paper; they live on a curved surface, like the skin of a basketball. If you try to do standard math (like drawing a straight line) on a basketball, you eventually run into problems. The math gets messy, breaks, or gives you the wrong answer.

For a long time, scientists have used a special toolkit called Lie Groups to handle this curved math. It works great, but the way they write it down is like reading a secret code. It's full of confusing symbols, and it's hard to see how the pieces fit together.

This paper introduces a new "language" (Tensor Notation) to write down these curved math problems. The author claims this new language doesn't invent new magic tricks; it just makes the existing magic tricks much easier to read, understand, and use.

The Problem: The "Flat Map" vs. The "Globe"

The paper explains that rotations are usually written as matrices (grids of numbers).

The Conflict: If you change a rotation matrix just a tiny bit to calculate a derivative (a rate of change), it often stops being a valid rotation. It's like trying to draw a straight line on a globe; the line eventually falls off the edge or distorts the map.
The Old Solution: Scientists use "Lie Algebras." Think of this as a flat map that represents a tiny patch of the curved globe. You do your math on the flat map (where it's easy), and then translate the result back to the globe.
The Messy Part: The old way of writing this translation involves complex, "opaque" formulas. It's like trying to follow a recipe written in a language you don't speak, where the ingredients are hidden inside nested boxes.

The Solution: The "Universal Translator" (Tensors & Einstein Notation)

The author suggests using Tensors and Einstein Summation Notation.

The Analogy: The LEGO Set
Imagine you are building a complex structure.

Old Way: You have a pile of instructions that say, "Take the red block, find the blue one that matches its left side, then glue it to the green one, but only if the green one is facing north." It's a long, confusing paragraph.
New Way (Tensors): You have a set of LEGO bricks with labels. The labels tell you exactly how they connect. If you see a brick labeled A with a 1 on the side, and another labeled B with a 1 on the side, you know they snap together. You don't need a paragraph of text; the shape and labels do the work for you.

What the Paper Claims:

Clarity: The new notation makes the "snapping together" of math operations obvious. It removes the guesswork about which numbers multiply with which.
Handling the "Curved" Math: It provides a clear way to calculate how things change on that curved surface (the Lie Group) without breaking the math.
The "Projection" Trick: One of the hardest parts of this math is converting a messy, curved result back into a flat vector. The paper introduces a specific "projection" tool (like a shadow caster) that automatically flattens the result correctly, even if the input is slightly messy.

Two Main Scenarios

The paper shows how this new language helps in two specific situations:

When the "State" is Curved:
- Scenario: You are trying to figure out the rotation of a camera (the "state").
- Old Way: Calculating how the error changes as you tweak the rotation was a headache of confusing symbols.
- New Way: The tensor notation writes the equation so clearly that you can see exactly how the rotation affects the error. It's like having a clear blueprint instead of a scribbled note.
When the "Measurement" is Curved:
- Scenario: Your sensors (like a star tracker) give you a rotation as a measurement.
- The Problem: You can't just subtract two rotations like normal numbers (e.g., $5 - 3 = 2$ ). You have to do a special "rotation subtraction."
- The New Way: The paper uses the new notation to define a "shadow" (projection) that takes the messy difference between two rotations and flattens it into a simple vector. This allows the computer to solve the puzzle much faster and more accurately.

The Results: Does it Work?

The author tested this new language on two problems:

Aligning Points: Trying to find the perfect rotation to match two sets of 3D dots. The new method found the answer very accurately.
Satellite Tracking: Simulating a satellite tracking its position using star maps and gyroscopes. The new method produced results that matched the theoretical "perfect" accuracy almost exactly.

The Bottom Line

The paper does not claim to have discovered a new way to fly a spaceship or a new type of sensor.

Instead, it claims to have found a better way to write the instructions for the math that already exists. By switching to Tensor Notation, the complex, confusing math of rotations becomes:

Clearer: You can see what is happening.
Simpler: The formulas are shorter and less prone to errors.
More Robust: It handles the "messy" parts of the math (like the Log function) in a way that is mathematically sound and easy to program.

Think of it as upgrading from a hand-drawn, messy map to a high-definition GPS interface. The destination (the solution) is the same, but getting there is much less confusing.

Title: Improved Representation of Matrix Lie Group Operations through Tensor Notation

Problem Statement
Modern estimation frameworks in robotics, aerospace, and control systems (e.g., Kalman filters, factor graphs) rely heavily on gradient-based optimization. A fundamental challenge in these fields is the estimation of rotation, which is typically represented by matrix Lie groups (such as $SO(3)$ or $SE(3)$). While Lie algebras provide a mechanism to handle derivatives on these manifolds, existing mathematical notations for computing and representing these derivatives are often opaque, prone to error, or lack standardization.

Specific difficulties include:

Representational Complexity: Deriving equations involving matrix Lie groups often requires "vectorizing" matrices to compute Jacobians. This destroys the structural relationships between matrix elements, making complex equations difficult to interpret and manipulate.
Dimensional Mismatch: Local movements on a Lie manifold are represented by low-dimensional vectors (tangent space), while the group elements themselves are high-dimensional matrices. Standard Jacobian assumptions (vector-to-vector derivatives) break down when differentiating with respect to matrix elements.
Notational Ambiguity: Existing literature uses various notations (e.g., specific derivative matrices, circle-plus/minus operators) that can hide complexity or lead to communication errors regarding the proper derivatives, particularly for the logarithm map ( $\log$ ) and its derivatives.

Methodology
The paper proposes the use of tensors and Einstein summation notation as a unified framework for representing and computing operations on matrix Lie groups. The methodology involves:

Tensor Representation: Treating scalars, vectors, and matrices as special cases of $n$ -dimensional tensors. This allows for the explicit definition of operations like contraction (generalized dot products) and element-wise multiplication across arbitrary dimensions.
Linear Subspace Operations: Defining the "hat" ( $\wedge$ $\land$ ) and "vee" ( $\vee$ $\lor$ ) operators (mapping between Lie algebra vectors and matrices) as linear synthesis and projection operations using a basis set of matrices ( $B$ $B$ ).
- Synthesis: $v_{jk} = v_i B_{ijk}$ (mapping a vector to a matrix).
- Projection: Using a pseudo-inverse of the projection matrix to map a generic matrix back to the Lie algebra vector space.
Derivative Formulation: Applying the chain rule using tensor contraction. The paper introduces a specific technique for computing the derivative of the $\log$ function (which maps a matrix to a vector) by utilizing the pseudo-inverse of the derivative of the $\exp$ function. This avoids the need for closed-form expressions of $\log$ derivatives, which are often non-existent or complex for arbitrary rotations.
Optimization Framework: Integrating these tensor-based derivatives into a Gauss-Newton optimization loop. The Jacobian matrices required for the optimization are constructed directly from tensor contractions, preserving the structural clarity of the underlying equations.

Key Contributions

Novel Notation: The primary contribution is a new mathematical notation for matrix Lie group operations. The author explicitly states this is not a new capability but a more "perspicuous" (clear and transparent) way to represent existing capabilities.
Tensor-Based Hat/Vee Operators: Formalizing the hat and vee operators as linear subspace synthesis and projection operations, removing the need for algorithmic, case-by-case definitions.
Pseudo-Inverse Log Derivative: Introducing a method to compute the derivative of the $\log$ function using the pseudo-inverse of the $\exp$ derivative. This provides a mathematically sound way to handle perturbations in matrices that may not strictly lie within the Lie group, broadening the applicability of the derivative.
Clarity in Complex Equations: Demonstrating how tensor notation simplifies the derivation of Jacobians for complex equations (e.g., $x = AVBy$) where $V$ is a Lie group element, avoiding the "opacity" of vectorized matrix derivatives.

Results
The paper validates the proposed notation through two optimization-based estimation scenarios using the $SO(3)$ group:

State Estimation: Estimating a single $SO(3)$ rotation matrix to align two sets of 3D vectors. The tensor-based approach yielded highly accurate results, with the estimated rotation matrix closely matching the ground truth.
Factor Graph Optimization: A satellite attitude estimation problem involving multiple $SO(3)$ states, absolute rotation measurements (star trackers), and relative rotation measurements (gyroscopes).
- The results showed that the achieved accuracy closely matched the theoretically predicted performance ( $\sigma \approx \sigma_m / \sqrt{n}$ ).
- This demonstrated the efficacy of the tensor-based notation in correctly handling the derivatives of the $\log$ function within a complex factor graph structure.

Significance and Claims
The paper claims that the significance of this work lies in clarity and ease of understanding rather than new computational power.

Explicitness: The notation makes the operations and derivatives explicit, whereas previous methods often relied on implicit assumptions or complex, non-standard notations.
Computational Support: The notation aligns with existing computational tools (e.g., numpy.einsum in Python, Einsums in C++), facilitating implementation.
Modest Scope: The author emphasizes that the paper does not introduce a new estimation algorithm or capability. Instead, it offers a "more perspicuous mathematical notation" that clarifies the derivatives and operations necessary for working with matrix Lie groups in gradient-based estimation frameworks. The goal is to reduce the barrier to entry and potential for error in formulating these complex estimation problems.