jepa_science
Copyright 2026 jepa_science contributors

This product is licensed under the Apache License, Version 2.0 (see LICENSE).

------------------------------------------------------------------------------
Provenance and attribution
------------------------------------------------------------------------------

This package is an independent, clean-room implementation of the method
"Sketched Isotropic Gaussian Regularization" (SIGReg) described in:

    Randall Balestriero and Yann LeCun.
    "LeJEPA: Provable and Scalable Self-Supervised Learning Without the
    Heuristics." arXiv:2511.08544 (2025). https://arxiv.org/abs/2511.08544

The code in this repository was written from the mathematical description in
that paper (the SIGReg objective, the hyperspherical Cramer-Wold reduction, the
Epps-Pulley characteristic-function statistic, and trapezoidal quadrature).
Mathematical methods, algorithms, and statistical tests are not themselves
subject to copyright; this is an original expression of them.

We deliberately did NOT copy code from the authors' reference implementation at
https://github.com/rbalestr-lab/lejepa. At the time of writing, that repository
ships a LICENSE file under Creative Commons Attribution-NonCommercial 4.0
(CC BY-NC 4.0) -- a NonCommercial license -- even though its pyproject.toml
declares "MIT". Those signals conflict, and a NonCommercial term is incompatible
with this package's permissive Apache-2.0 license. To avoid any licensing
ambiguity, none of that repository's code was reproduced here; only the
publicly described method (from the paper) was implemented.

The paper itself is distributed under CC BY 4.0 / CC BY-SA 4.0 on arXiv; this
implementation reuses the method it describes, not its prose.

If you use this software, please cite the LeJEPA paper (see CITATION.cff).
