icalcc

Locally centered contrast functions for scikit-learn FastICA. Drop-in replacement for sklearn.decomposition.FastICA with bounded LCC-tanh and LCC-exp contrasts, plus polynomial LCC contrasts of order 4, 6, and 8.

from icalcc import ICALCC
ica = ICALCC(n_components=4, K='ltanh', random_state=0)
S_hat = ica.fit_transform(X)

Installation

pip install icalcc

Supported K Values

Usage

from icalcc import ICALCC
# Bounded LCC-tanh (recommended for heavy-tailed or skewed sources)
ica = ICALCC(n_components=4, K='ltanh', random_state=0)
S_hat = ica.fit_transform(X)
# Bounded LCC-exp (Rényi-2 entropy interpretation)
ica = ICALCC(n_components=4, K='lexp', random_state=0)
S_hat = ica.fit_transform(X)
# Polynomial LCC order 8 (near-Gaussian sources)
ica = ICALCC(n_components=4, K=8, random_state=0)
S_hat = ica.fit_transform(X)
# Classical FastICA baseline
ica = ICALCC(n_components=4, K='tanh', random_state=0)
S_hat = ica.fit_transform(X)

Convergence tracking

ica.fit(X)
print(ica.converged_)  # True if all components converged

Reconstruction MSE

ica = ICALCC(n_components=3, K='lexp', random_state=0)
ica.fit(X)
mse = ica.reconstruction_mse(X)
print(f"Reconstruction MSE: {mse:.6f}")

reconstruction_mse transforms X to sources and reconstructs via the mixing matrix in the original feature space. Meaningful only in the undercomplete case (n_components < n_features); returns 0.0 exactly when n_components == n_features. Inherited automatically by gpuicalcc.

See Also

• gpuicalcc — PyTorch GPU-accelerated extension (40–48× speedup for bounded contrasts)
• Experiment code

Requirements

• Python ≥ 3.9
• numpy ≥ 1.24
• scikit-learn ≥ 1.3

Citation

If you use this package, please cite both the software paper and the underlying LCC kernel paper:

@article{saito2026icalcc,
  author  = {Saito, Tetsuya},
  title   = {{ICALCC}: Locally Centered Contrast Functions for
             {FastICA} with {GPU} Acceleration},
  journal = {TechRxiv},
  year    = {2026},
  doi     = {10.36227/techrxiv.177203264.46969730}
}
@article{saito2026lcc,
  author  = {Saito, Tetsuya},
  title   = {Locally Centered Cyclic Kernels for Higher-Order
             Independent Component Analysis},
  journal = {TechRxiv},
  year    = {2026},
  doi     = {10.36227/techrxiv.177220376.62411390}
}

License

CC BY 4.0

Changelog

0.1.4

• Added reconstruction_mse(X) method to ICALCC: computes mean squared reconstruction error in the original feature space via inverse_transform. Inherited by GPUICALCC without changes.
• Renamed contrast 'skew' to 'cube' for clarity; 'skew' is retained as a legacy alias.

0.1.3

• Initial public release.