ChenSignatures.jl

ChenSignatures.jl is a high-performance Julia library for computing path signatures and log-signatures—core tools in Rough Path Theory, stochastic analysis, and modern deep-learning architectures such as Neural CDEs.

It provides state-of-the-art speed, sparse algebra support, and a clean API for both research and production use.

---

📐 Overview

Path signatures are sequences of iterated integrals that encode the geometric structure of a path. They form a powerful, coordinate-free representation used in:

• financial time series and derivatives pricing
• machine learning (Neural CDEs, feature engineering)
• stochastic analysis and rough path theory

ChenSignatures.jl computes both signatures and log-signatures efficiently and accurately, with full support for dense and sparse tensor algebras.

---

✅ Key Features

Core Functionality

• Signatures: Compute truncated path signatures up to any depth.
• Log-Signatures: Lyndon-projected log-signatures for minimal bases.
• Tensor Algebra: Dense and sparse tensor algebra (SparseTensor).
• Algebraic Operations: Shuffle product, tensor exp/log, resolvent operators.

Performance

• Highly optimized kernels using:
  • StaticArrays.jl
  • Native Julia SIMD vectorization
• Designed to scale across:
  • long paths (large N)
  • high dimensions (d)
  • large truncation levels (m)

Validation

Benchmarked and numerically cross-validated against:

• iisignature
• pysiglib

Full benchmarking and correctness tooling is included (see below).

---

🐍 Python API: chen-signatures

ChenSignatures.jl powers the Python package:

📦 PyPI: https://pypi.org/project/chen-signatures/
📁 Source: python/

pip install chen-signatures

Features:

• Same API as Julia (sig, logsig)
• Works on Python 3.9–3.13
• Automatically installs a lightweight Julia runtime via juliacall
• Faster than iisignature and comparable to pysiglib

---

📊 Benchmarks

Cross-Language Benchmarks

For comprehensive performance comparisons against iisignature and pysiglib, see the dedicated benchmark repository:

👉 aleCombi/sig-benchmarks

This external suite provides:

• Isolated environment testing (Python vs Julia)
• Multiple library comparisons
• Detailed performance profiles and visualizations
• Methodologically fair benchmarking

Internal Julia Benchmarks

For quick development regression testing, see:

📁 benchmark/

Run Julia-only benchmarks:

julia --project=benchmark benchmark/benchmark.jl

---

📦 Dependencies

ChenSignatures.jl uses a minimal, high-performance stack:

• StaticArrays.jl
• LinearAlgebra
• ChainRulesCore.jl for AD support
• Enzyme.jl for automatic differentiation

---

🚀 Example

using ChenSignatures

path = randn(1000, 5)
sig_result = sig(path, 4)

# For logsig, need to prepare basis first
basis = prepare(5, 4)  # dimension=5, level=4
logsig_result = logsig(path, basis)

---

📖 Documentation

For detailed API documentation, mathematical background, and usage examples, see:

👉 ChenSignatures.jl Documentation

---

🧪 Testing

The repository includes:

• Full Julia test suite (run with julia --project -e 'using Pkg; Pkg.test()')
• Cross-language validation fixtures at test/validation/
• Automatic fixture generation via test/validation/generate_fixtures.py

---

📚 Citation

@software{chen_signatures,
  author = {Combi, Alessandro},
  title = {ChenSignatures.jl: High-performance signatures and log-signatures},
  year = {2025},
  url = {https://github.com/aleCombi/ChenSignatures.jl}
}

---

🤝 Contributing

Issues and feedback are welcome! However, pull requests are not being accepted at this time.

👉 https://github.com/aleCombi/ChenSignatures.jl/issues

Version Management

For Maintainers:

Version is managed in Project.toml only. Python package reads it automatically.

To release a new version:

1. Edit Project.toml: version = "0.3.0"
2. Commit: git commit -am "Bump version to 0.3.0"
3. Register with Julia General (comment @JuliaRegistrator register on PR/commit)
4. TagBot creates v0.3.0 tag automatically
5. Tag triggers PyPI publish automatically

Never edit version in python/pyproject.toml - it's computed dynamically.