chen-signatures 0.1.4

Fast rough path signatures - faster than iisignature, comparable to pysiglib

chen-signatures

Fast rough path signatures for Python, powered by Julia.

1.6× faster than iisignature with support for modern Python (3.9-3.13).

Why chen-signatures?

Feature	chen-signatures	iisignature
Speed	187 ms	299 ms
Python 3.10+	✅ Yes	❌ No (≤3.9 only)
Python 3.13	✅ Yes	❌ No
Autodiff	✅ Yes (ForwardDiff)	❌ No
Maintained	✅ Active	⚠️ Unmaintained

Benchmark: N=1000 points, d=10 dims, level=5

Installation

pip install chen-signatures

First import will automatically install Julia (via juliacall). Takes ~2 minutes once.

Quick Start

import chen
import numpy as np

# Your time series data
path = np.random.randn(1000, 10)  # 1000 timepoints, 10 dimensions

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

# Compute log-signature  
logsignature = chen.logsig(path, m=5)

API

sig(path, m)

Compute the truncated signature up to level m.

Parameters:

• path : (N, d) numpy array - Path data
• m : int - Truncation level

Returns:

• numpy.ndarray - Flattened signature coefficients

Example:

import chen
import numpy as np

path = np.array([[0., 0.], [1., 0.], [1., 1.]])
sig = chen.sig(path, m=3)
# Returns array of length d + d² + d³ = 2 + 4 + 8 = 14

logsig(path, m)

Compute the log-signature projected onto the Lyndon basis.

Parameters:

• path : (N, d) numpy array - Path data
• m : int - Truncation level

Returns:

• numpy.ndarray - Log-signature in Lyndon basis

Example:

logsig = chen.logsig(path, m=5)

Supported Types

• float32 and float64
• Any numpy array-like input

Performance

Production-scale benchmark (N=1000, d=10, m=5):

import chen
import numpy as np
import time

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

t0 = time.time()
sig = chen.sig(path, 5)
print(f"Time: {(time.time()-t0)*1000:.1f} ms")
# Output: Time: 187.4 ms

Compare to iisignature: 299.3 ms (1.6× slower)

Use Cases

• Financial time series: Extract signature features from price data
• Sensor data: Process multivariate sensor streams
• Neural CDEs: Differentiable features for neural networks
• Anomaly detection: Signature-based feature engineering

Limitations

• First import is slow (~2 min to install Julia environment, one-time)
• Not GPU-accelerated (CPU only, but very fast)
• Memory usage: Uses more RAM than iisignature (but negligible for most applications)

Requirements

• Python ≥3.9
• NumPy ≥1.20
• ~500MB disk space (for Julia installation)

Citation

If you use this in research, please cite:

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

License

MIT License - see LICENSE file.

Contributing

Issues and pull requests welcome at github.com/aleCombi/Chen.jl