chen-signatures 0.4.0

Fast path signatures

chen-signatures

Fast rough path signatures for Python, powered by a high-performance Julia backend.

chen-signatures brings the speed and numerical stability of Julia’s ChenSignatures.jl to Python via juliacall, offering a modern, actively maintained alternative to existing signature libraries.

It has been benchmarked against both iisignature and pysiglib, showing:

• Comparable performance to pysiglib, a modern C++/Python implementation
• Consistently faster performance than iisignature across typical configurations

Full benchmark notebooks and articles will be published separately.

---

Installation

pip install chen-signatures

On first import, juliacall will automatically install a lightweight Julia runtime. This happens once per environment.

---

Quick Start

import chen
import numpy as np

path = np.random.randn(1000, 10)

# Compute signature
signature = chen.sig(path, m=5)

# Compute log-signature (requires prepare step)
basis = chen.prepare_logsig(d=10, m=5)  # d must match path dimension
logsignature = chen.logsig(path, basis)

# Path augmentations (time/lead–lag) with convenience wrappers
augmented = chen.time_augment(path, Tspan=1.0)
leadlag = chen.lead_lag(path)
sig_time = chen.sig_time(path, m=5)
sig_leadlag = chen.sig_leadlag(path, m=5)

---

API

sig(path, m)

Compute a truncated signature up to level m.

signature = chen.sig(path, m=3)

Args:

• path: (N, d) array-like where N is path length, d is dimension
• m: truncation level (positive integer)

Returns: 1D numpy array of signature coefficients

---

prepare_logsig(d, m)

Precompute and cache the Lyndon basis for log-signature computation.

basis = chen.prepare_logsig(d=10, m=5)

Args:

• d: path dimension (positive integer)
• m: truncation level (positive integer)

Returns: Basis cache object (reusable across multiple paths)

---

logsig(path, basis)

Compute log-signatures using a precomputed Lyndon basis.

basis = chen.prepare_logsig(d=10, m=5)
logsig = chen.logsig(path, basis)

Args:

• path: (N, d) array-like where N is path length, d is dimension
• basis: Basis object from prepare_logsig(d, m)

Returns: 1D numpy array of log-signature coefficients (Lyndon basis projection)

Note: The basis object should match the dimension of your path. Prepare it once and reuse for multiple paths of the same dimension.

Path Augmentations

augmented = chen.time_augment(path, Tspan=1.0)   # add time as first coordinate
leadlag = chen.lead_lag(path)                    # standard lead–lag transform

sig_time = chen.sig_time(path, m=3)              # = chen.sig(time_augment(path), 3)
sig_leadlag = chen.sig_leadlag(path, m=3)        # = chen.sig(lead_lag(path), 3)

---

Supported Types

• float32, float64
• Any NumPy array-like input
• Contiguous arrays recommended (handled automatically)

---

Use Cases

• Financial time series and derivatives pricing
• Neural CDEs and machine learning
• Feature engineering and representation learning

---

Limitations

• First import is slow (Julia installation)
• CPU-only execution
• Uses more memory than minimal C++ libraries

---

Requirements

• Python ≥ 3.9
• NumPy ≥ 1.20
• ~500MB disk space for Julia runtime

---

Citation

@software{chen_signatures,
  author = {Combi, Alessandro},
  title = {chen-signatures: Fast rough path signatures for Python},
  year = {2025},
  url = {https://github.com/aleCombi/ChenSignatures.jl}
}

---

Contributing

Issues and feedback are welcome at https://github.com/aleCombi/ChenSignatures.jl/issues

However, pull requests are not being accepted at this time.