Syntropy is a Python library for multivariate information-theoretic analysis of discrete and continuous data. It provides efficient implementations of information measures ranging from basic quantities like entropy and mutual information to modern constructs like the partial information decomposition, O-information, and information rates for time series.
The goal of Syntropy is to be the "Networkx" of information theory. In the same that Networkx made network science available to scientists across disciplines, Syntropy aims to make information theory equally accessible.
- Multiple estimators: Discrete (with corrections), Gaussian, KNN (Kraskov), neural (normalizing flow), and mixed discrete/continuous estimators
- Pointwise measures: Access local/pointwise values, not just expected values
- Higher-order information: Total correlation, dual total correlation, O-information, S-information
- Information decomposition: Discrete and continuous partial information decomposition (PID), partial entropy decomposition (PED), generalized information decomposition (GID), and more. Multiple redundancy functions for each class.
- Time series: Information rates, Lempel-Ziv complexity, sample entropy.
- Consistent API: Same interface across all estimator types
pip install syntropyxThen import as:
import syntropyImportant
The package is syntropyx on PyPI because syntropy was already taken. The x is just a workaround—the actual library is called Syntropy.
For development:
git clone https://github.com/thosvarley/syntropy.git
cd syntropy
pip install -e ".[dev]"Discrete estimators work with probability distributions represented as dictionaries:
from syntropy.discrete import mutual_information, o_information
# XOR distribution: pure synergy
xor = {
(0, 0, 0): 0.25,
(0, 1, 1): 0.25,
(1, 0, 1): 0.25,
(1, 1, 0): 0.25,
}
# Mutual information between inputs (0,1) and output (2)
ptw, mi = mutual_information(idxs_x=(0, 1), idxs_y=(2,), joint_distribution=xor)
print(f"I(X0,X1 ; X2) = {mi:.3f} bits") # 1.0 bit
# O-information (negative = synergy-dominated)
ptw, omega = o_information(joint_distribution=xor)
print(f"Omega = {omega:.3f} bits") # -1.0 bitFor continuous data with approximately Gaussian distributions:
import numpy as np
from syntropy.gaussian import mutual_information, total_correlation
# Generate correlated Gaussian data
n = 10_000
x = np.random.randn(n)
y = 0.8 * x + 0.6 * np.random.randn(n)
z = 0.5 * x + 0.866 * np.random.randn(n)
data = np.vstack([x, y, z])
cov = np.cov(data)
mi = mutual_information(idxs_x=(0,), idxs_y=(1,), cov=cov)
tc = total_correlation(idxs=(0, 1, 2), cov=cov)
print(f"I(X ; Y) = {mi:.3f} nats")
print(f"TC(X, Y, Z) = {tc:.3f} nats")Non-parametric estimation for continuous data:
import numpy as np
from syntropy.knn import mutual_information
# Non-linear relationship
n = 5_000
x = np.random.randn(n)
y = x**2 + 0.5 * np.random.randn(n)
data = np.vstack([x, y])
ptw, mi = mutual_information(idxs_x=(0,), idxs_y=(1,), data=data, k=5)
print(f"I(X ; Y) = {mi:.3f} nats")For complex, high-dimensional distributions using normalizing flows:
import torch
from syntropy.neural import mutual_information
# Generate data (samples x features format)
n = 10_000
x = torch.randn(n)
y = 0.7 * x + 0.714 * torch.randn(n)
data = torch.stack([x, y], dim=1)
ptw, mi = mutual_information(idxs_x=(0,), idxs_y=(1,), data=data, verbose=True)
print(f"I(X ; Y) = {mi:.3f} nats")For mutual information between discrete and continuous variables:
import numpy as np
from syntropy.mixed import mutual_information
n = 10_000
continuous = np.random.randn(1, n)
discrete = (continuous > 0).astype(int)
ptw, mi = mutual_information(discrete_vars=discrete, continuous_vars=continuous)
print(f"I(discrete ; continuous) = {mi:.3f} nats")
# The continuous entropies can be estimated with a Gaussian (default) or a
# KNN estimator. The KNN option recovers the true mutual information when the
# continuous marginal is non-Gaussian:
ptw, mi = mutual_information(
discrete_vars=discrete,
continuous_vars=continuous,
continuous_estimator="knn",
k=5,
)| Measure | Discrete | Gaussian | KNN | Neural | Mixed |
|---|---|---|---|---|---|
| Entropy | x | x | x | x | x |
| Conditional Ent. | x | x | x | x | x |
| Mutual Information | x | x | x | x | x |
| Conditional MI | x | x | x | x | |
| KL Divergence | x | x | x | x | |
| Total Correlation | x | x | x | x | |
| Dual Total Correlation | x | x | x | x | |
| O-Information | x | x | x | x | |
| S-Information | x | x | x | x | |
| Co-Information | x | x | |||
| TSE Complexity | x | x | |||
| Partial Info. Decomp. | x | x | |||
| Partial Entropy Decomp. | x | x | |||
| Generalized Info. Decomp. | x | x | |||
| Integrated ( |
x | x | |||
| Information Rates | x | x | x | ||
| Connected Information | x | ||||
|
|
x | ||||
| I_dep Decomp. | x |
Syntropy also includes a number of optimization algorithms.
- Finding optimally-synergistic submatrices from a covariance matrix (as done by Varley, Pope et al., 2023).
- Finding the maximum-entropy discrete distribution consistent with k-order marginals (as done in the DIT package).
In the utils.py files, you can also find a number of utility functions for interacting with discrete and continuous probability distributions.
Full documentation is available at syntropy.readthedocs.io.
pytest tests/If you use Syntropy in your research, please cite:
@software{syntropy,
author = {Varley, Thomas F.},
title = {Syntropy: Multivariate Information Theory for Python},
url = {https://github.com/thosvarley/syntropy},
}MIT License. See LICENSE for details.