What’s new

v0.2.3 (October 2026)

Bug fixes

• antropy.perm_entropy() with several delays on a 2-D array now returns one value per row (averaged across delays) instead of a single value for the whole array. Invalid delay inputs, such as an empty list, now raise a clear error.

• antropy.perm_entropy() (order 3 or 4) now compares the original values directly, so very large integers or very close floats are no longer mixed up. Rows containing NaN now return NaN without affecting the other rows.

Tests

• Made tests that compare 2-D and 1-D results robust to tiny floating-point differences across platforms.

v0.2.2 (April 2026)

Performance

• antropy.perm_entropy() gains a fast path for order ∈ {3, 4}: ordinal patterns are encoded via pairwise comparison bit-keys and counted with np.bincount rather than argsort + np.unique, giving a major speed-up compared to the previous implementation.

• The general path (order > 4) is also improved: the embedded matrix is now built as a zero-copy as_strided view and the hash is computed via a single BLAS @ product instead of an element-wise multiply + sum.

New features

• antropy.perm_entropy() now accepts 2-D arrays of shape (n_channels, n_times) for order ∈ {3, 4}, computing entropy for every row in a single vectorised call.

Bug fixes

• antropy.perm_entropy() now always applies a small positional epsilon jitter before sorting, making results fully deterministic across platforms and dtypes. Previously, integer-typed signals (and some float signals with exact ties) could produce different outputs depending on the platform’s unstable-sort tie-breaking behaviour.

• Normalized antropy.perm_entropy() output is now clipped to [0, 1] to prevent returning -0.0 for perfectly regular signals (IEEE 754 artefact from -(1.0 * log2(1.0))).

Docs

• Fixed doctest failures in antropy.spectral_entropy(), antropy.num_zerocross(), and antropy.hjorth_params() caused by NumPy’s new scalar repr (np.float64(x) / np.int64(x) instead of bare literals in newer NumPy).

• Added a Big O complexity column to the performance tables in README.rst and docs/index.rst.

• Performance tables now show timings at both 1000 and 10000 samples.

• Added benchmarks/benchmark_all.py script to reproduce all table timings.

• Rewrote docs/index.rst to match README.rst in content and style.

Maintenance

• Clean up docs/conf.py (remove redundant sys.path manipulation) and tighten pyproject.toml (remove stale extras, fix editable-install metadata).

v0.2.1 (March 2026)

Bug fixes

• Fix off-by-one in antropy.higuchi_fd(): the inner loop summed \(N_m - 1\) differences per sub-series instead of the correct \(N_m\), causing a systematic underestimate of curve length. Verified against the NeuroKit2 reference implementation.

• antropy.sample_entropy() now returns np.nan (instead of 0) when no template of length m matches within the tolerance, since the entropy is mathematically undefined in that case.

• Fix silent epsilon = 10e-9 (= 1e-8) typo in utils.py; corrected to 1e-9. This constant guards against division-by-zero in the Numba linear regression used by antropy.detrended_fluctuation().

• Fix all = [...] → __all__ = [...] in all three source modules, which previously leaked imported names (np, jit, KDTree, …) into the public namespace on from antropy import *.

New features

• antropy.hjorth_params() gains an optional sf parameter. When provided, mobility is returned in Hz instead of samples⁻¹ (multiplied by sf); complexity is unaffected.

Validation

• User-facing assert statements replaced with proper ValueError / TypeError raises throughout (asserts are silently stripped by python -O): perm_entropy(delay=0), spectral_entropy(method=…), app_entropy/sample_entropy(tolerance=…), and lziv_complexity(sequence=…, normalize=…).

Docs

• Docstring formula for antropy.higuchi_fd() corrected to match the fix above (\(\sum_{j=1}^{N_m}\) not \(\sum_{j=1}^{N_m-1}\)).

• Added notes to antropy.lziv_complexity() (float arrays are truncated to uint32, not discretised — binarise continuous signals first), antropy.num_zerocross() (signbit(0) == False behaviour), antropy.spectral_entropy() (DC component included), antropy.detrended_fluctuation() (\(\alpha \approx 1\) is ambiguous), and antropy.hjorth_params() (mobility units).

v0.2.0 (March 2026)

Build & CI

• Drop Python 3.9 (EOL), add Python 3.13 support. Minimum is now Python 3.10.

• Switch from pip to uv in all GitHub Actions workflows.

• Add explicit minimum versions for core dependencies: numpy>=1.22.4, scipy>=1.8.0, scikit-learn>=1.2.0.

• Migrate [project.optional-dependencies] to PEP 735 [dependency-groups].

• Bump setuptools>=80.0.

• Split CI into three jobs: test-core (3 platforms × 4 Python versions), test-dependency-combinations (4 dep combos from minimum to latest), and coverage.

• Fix test-dependency-combinations job: separate antropy install (--no-deps) from test-dependency install so that pytest’s own dependencies (e.g. pluggy) are always resolved.

• Fix Codecov upload to use ${{ secrets.CODECOV_TOKEN }} instead of a hardcoded token.

• Switch Ruff workflow from astral-sh/ruff-action@v1 to uvx ruff via astral-sh/setup-uv@v7.

• Extend Ruff rules: add W (pycodestyle warnings) and NPY (NumPy rules).

Tests

• Increase test coverage from ~54 % to 100 %.

• Add tests/test_utils.py covering all branches of the _embed helper (1-D and 2-D paths, all error conditions).

• Add edge-case tests: sample_entropy returning inf (m-length matches exist but no (m+1)-length matches); detrended_fluctuation returning NaN for a constant signal; spectral_entropy raising on an invalid method string.

• Set NUMBA_DISABLE_JIT=1 in the coverage CI job so coverage.py can instrument Numba JIT function bodies; other CI jobs still exercise real compiled code.

Bug fixes

• Fix antropy.higuchi_fd() returning a ValueError (math domain error) on constant or integer-typed input arrays: log(0) is now guarded to return -inf, matching Numba’s IEEE 754 behaviour.

Docs

• Switch documentation theme from sphinx_bootstrap_theme to pydata-sphinx-theme (dark/light toggle, GitHub icon, improved layout).

• Fix three broken intersphinx URLs: NumPy, SciPy, and MNE-Python.

• Add sphinx.ext.mathjax for LaTeX math rendering in docstrings.

• Add contributing.rst guide.

• Fix stale :py:func: cross-references in changelog entries v0.1.1–v0.1.3 (entropy.XXX → antropy.XXX).

• Fix typo in v0.1.6 changelog (“Fox for KDTree” → “Fix for KDTree”).

• Modernize README.rst and docs/index.rst: add PyPI, conda-forge, downloads, and Ruff badges; add uv installation instructions; fix broken links.

v0.1.9 (February 2025)

• Remove stochastic package from dependency to enable support for Numpy 2.x

v0.1.8 (December 2024)

• Switch to modern python packaging

• Use ruff instead of black/flake8

v0.1.7 (December 2024)

• Simplify Katz FD implementation. raphaelvallat/antropy#30

• Add tolerance parameter to the approximate and sample entropy. raphaelvallat/antropy#32

• Fix for scikit-learn ≥ 1.3 in approximate and sample entropy. raphaelvallat/antropy#36

v0.1.6 (July 2023)

This version requires numba >= 0.57.

1. Allow readonly arrays in numba jit signature. raphaelvallat/antropy#23

2. Improved sample entropy kernel. raphaelvallat/antropy#25

3. Fix for KDTree.valid_metrics which is method since sklearn 1.3. raphaelvallat/antropy#30

v0.1.5 (December 2022)

1. antropy.perm_entropy() will now return the average entropy across all delays if a list or range of delays is provided.

2. Handle the limit of p = 0 in functions that evaluate the product p * log2(p), to give 0 instead of nan (see PR3).

3. antropy.detrended_fluctuation() will now return alpha = 0 when the correlation coefficient of the fluctuations of an input signal is 0 (see PR21).

v0.1.4 (April 2021)

Important

The package has now been renamed AntroPy (previously EntroPy)!

1. Faster implementation of antropy.lziv_complexity() (see PR1). Among other improvements, strings are now mapped to UTF-8 integer representations.

v0.1.3 (March 2021)

1. Added the antropy.num_zerocross() function to calculate the (normalized) number of zero-crossings on N-D data.

2. Added the antropy.hjorth_params() function to calculate the mobility and complexity Hjorth parameters on N-D data.

3. Add support for N-D data in antropy.spectral_entropy(), antropy.petrosian_fd() and antropy.katz_fd().

4. Use the stochastic package to generate stochastic time-series.

v0.1.2 (May 2020)

1. antropy.lziv_complexity() now works with non-binary sequence (e.g. “12345” or “Hello World!”)

2. The average fluctuations in antropy.detrended_fluctuation() is now calculated using the root mean square instead of a simple arithmetic. For more details, please refer to this GitHub issue.

3. Updated flake8

v0.1.1 (November 2019)

1. Added Lempel-Ziv complexity (antropy.lziv_complexity()) for binary sequence.

v0.1.0 (October 2018)

Initial release.

1. Permutation entropy

2. Spectral entropy

3. Singular value decomposition entropy

4. Approximate entropy

5. Sample entropy

6. Petrosian Fractal Dimension

7. Katz Fractal Dimension

8. Higuchi Fractal Dimension

9. Detrended fluctuation analysis