GUDHI

GUDHI is the widest general-purpose library in the field: simplex trees, Vietoris–Rips, alpha, witness and cubical complexes, persistence in several forms, and a representations layer that turns diagrams into vectors. A C++ core with a Python interface over it. Breadth is the point — it is what you reach for before you know which complex your data wants.

Its role in this corpus is less as a tool than as a floor. Several fixtures depend on it rather than reimplementing complex construction, and both clean-room featurizers written here — open-topoqa-featurizer and open-topodockq-featurizer — build their filtrations with it. Very little in the structure-QA path would exist without it.

The license is not one license

This note carries LicenseRef-gudhi-mixed rather than an SPDX id, and the reason is worth stating carefully because a reader skimming frontmatter will otherwise get it wrong.

GUDHI’s own code is MIT. But many of its modules depend on third-party libraries that are not — CGAL, Miniball, PyKeOps — and the project says so itself: its module list marks the affected entries “MIT (GPL v3)”, glossing that as GUDHI code is MIT, but there is a dependency on GPL code, so for practical purposes for a user it is as if this package was GPLv3. The conda-forge build declares the whole conjunction, MIT and BSD-3-Clause and MPL-2.0 and LGPL-3.0-or-later and GPL-3.0-or-later, because that is what is actually inside the artifact.

So no single id is true. MIT would describe the source and misdescribe everything anyone installs; any one copyleft id would misdescribe the parts that are permissive. A LicenseRef resolves deny-by-default, which is the correct posture for a package whose terms depend on a choice the consumer makes rather than on a field the publisher set.

That choice is not hypothetical here. Alpha complexes are a CGAL module, and alpha complexes are what both featurizers use for their H1 and H2 bars. Our own code stays MIT and nothing about that changes — but a locked environment or a container that ships GUDHI beside it is a distribution of the whole, and carries the terms of what is inside. Engineering observation, not legal advice, and the reason this note declines to simplify the field.

What it is not

Not a machine-learning pipeline and not a scorer. The representations layer does vectorize diagrams, but this corpus routes that question to persim instead, because separating the computation of persistence from its representation keeps two different failure modes apart.

In this corpus

gudhi-environment pins it straight from conda-forge with no recipe, which is why it is one of the few L3 fixtures here. The technique is persistent-homology; persistent-homology itself recommends it as the default when the right complex is not yet known.