[Paper Review] Calculation of Iterated-Integral Signatures and Log Signatures

Jeremy Reizenstein|arXiv (Cornell University)|Dec 7, 2017

Advanced Database Systems and Queries7 references15 citations

TL;DR

This paper presents a self-contained computational framework for calculating iterated-integral signatures and their compressed log signature counterparts for piecewise-linear paths. It details the algebraic foundations, numerical algorithms, and software tools—particularly Python implementations—for efficiently computing and visualizing log signatures, enabling robust path representation in machine learning and stochastic analysis.

ABSTRACT

We explain the algebra needed to make sense of the log signature of a path, with plenty of examples. We show how the log signature can be calculated numerically, and explain some software tools which demonstrate it.

Motivation & Objective

• To provide a clear, self-contained explanation of the algebraic and computational machinery required to compute the log signature of a path.
• To address the challenge of representing complex paths in machine learning by offering a compact, informative signature representation that preserves path geometry.
• To demonstrate numerical computation of log signatures using explicit algorithms and code, with emphasis on practical implementation and stability.
• To enable visualization of the log signature's components and their relationship to path geometry through interactive Mathematica tools.
• To compare different basis systems (e.g., Coropa, Hall, Lyndon) for log signature computation and clarify their algebraic structure and computational impact.

Proposed method

• Uses the tensor algebra $ T^{ ext{underline}{m}}(\mathbb{R}^d) $ to represent signatures up to level $ m $, truncating higher-order terms for computational feasibility.
• Employs the exponential and logarithm series in the tensor algebra to define the log signature via $ \log(1 + x) = x - \frac{1}{2}x^2 + \frac{1}{3}x^3 - \cdots $, with truncation at level $ m $.
• Applies generalized Hall and Lyndon bases to express the log signature in a minimal, non-reduundant basis, with special handling for bracketed expressions and ordering conventions.
• Implements a Python-based computational pipeline that generates code for BCH formula evaluation and log signature computation, supporting both Coropa and classical Hall basis conventions.
• Integrates interactive Mathematica visualizations (via view.m and bch.m) to link path geometry with log signature components, allowing real-time manipulation and observation.
• Uses lexicographic ordering and recursive bracketing rules to define basis elements, with depth and alphabetical ordering used to ensure uniqueness and completeness.

Experimental results

Research questions

• RQ1 How can the log signature of a piecewise-linear path be computed numerically with high accuracy and efficiency?
• RQ2 What algebraic structures—such as generalized Hall or Lyndon bases—enable a minimal and stable representation of the log signature?
• RQ3 How do different basis conventions (e.g., Coropa vs. classical Hall) affect the structure and length of the resulting log signature computation?
• RQ4 To what extent can interactive visualization tools help in understanding the geometric and algebraic relationship between a path and its log signature?
• RQ5 What are the numerical and computational trade-offs in using the log signature over the full signature in machine learning applications?

Key findings

• The log signature provides a compressed, information-preserving representation of the full iterated-integral signature, reducing the number of components while retaining all essential path information.
• The log signature is invariant under reparametrization and translation of the path, and is stable under small perturbations unless approximate backtracking occurs.
• For $ d=2 $ and $ m=4 $, the log signature has eight components, which can be visualized interactively using the provided Mathematica tools to explore path-log signature duality.
• The Python codebase generates efficient, basis-specific implementations of the log signature computation, with support for both Coropa and classical Hall basis conventions.
• The visualisation tool demonstrates that level 1 and level 3 components of the log signature correspond to 2D vector controls, while level 2 controls the signed area enclosed by the path.
• The algorithm correctly handles the non-commutative nature of the tensor algebra and ensures that the log signature is well-defined and computable via truncated series expansions.