About
Intrinsic Green Learning (IGL) is a Python library implementing the task-conditioned intrinsic-dimensionality discovery framework introduced in Quemy, A. (2026). Intrinsic Green's Learning: Supervised Learning on Manifolds via Inverse PDE. ICLR 2026 Workshop on AI and PDE.
Author
Alexandre Quemy (@aquemy,
ORCID 0000-0002-5865-6403)
— research and implementation. The library is developed under
Hother and released under the MIT license.
Where IGL sits in the landscape
Intrinsic-dimensionality estimation has historically split into two camps:
- Purely geometric estimators (Levina–Bickel, TwoNN, MLE-based
approaches in
scikit-dimension) look at the input data alone and ignore what the user actually wants to predict. - Implicit bottlenecks in deep networks (autoencoders, information-bottleneck methods, variational approaches) use a low-dimensional latent but don't report its size without a separate analysis pass.
IGL takes a third position: solve the supervised task and read off
the effective dimension in the same training run. The discovered
d_eff is a property of (input, task), not of the input alone — the
same dataset resolves into different effective dimensions for a
classifier, a regressor, and an autoencoder, and the hierarchy
\(d_{\text{cls}} \le d_{\text{reg}} \le d_{\text{recon}}\) falls out
naturally.
Theoretical underpinnings
The library realises the inverse-PDE formulation from the paper above:
- The latent-space operator \(L\) is encoded by the Green's-function
kernel \(G(z, s)\), which can be expressed either as a local product
of per-dimension kernels (the default
GreenKernel) or via its spectral expansion \(G(z, s) = \sum_n \phi_n(z)\,\phi_n(s) / \lambda_n\) (theSpectralKernelplus closed-form Fourier / Chebyshev / Legendre / Hermite / Laguerre bases, or learned Laplace–Beltrami / user-supplied graph bases). - The non-trivial null space of \(L\) — the modes the Green's-
function expansion cannot reach (e.g. the constant function under a
Neumann Laplacian) — is fitted as un-regularised columns of the
design matrix via the kernel-agnostic
NullSpaceBasisprotocol. - Variable Projection with random Matryoshka truncation trains the
encoder while sampling the latent-dimension cutoff \(k\), so a single
trained model exposes a quality-vs-
kdimension curve that reads offd_effat deployment time.
Concrete derivations and worked examples live under Concepts; the per-symbol API reference under Reference.
Status
IGL is active research-grade software. The API follows strict
typing discipline (basedpyright --strict, ≥99% test coverage with
property-based fuzzing on the numerical core), but minor versions may
introduce breaking changes as the framework evolves. The 0.x line is
appropriate for research and experimentation; production users should
pin specific minor versions.
Citing
See the
Bibliography
section of the project README for the recommended citation, or the
CITATION.cff
file for the structured form (GitHub's "Cite this repository"
widget reads it automatically).