Related packages

A few existing tools cover overlapping ground.

genriesz (Kato, 2026) is a single Python package implementing the Bregman-unified Riesz regression framework from arXiv:2601.07752. It exposes LinearFunctional and BregmanGenerator abstractions, analogous to this project’s Estimand and Loss. It ships several basis-function classes (polynomial, random Fourier features, Nyström, KNN catchments, random-forest leaves, PyTorch embeddings) inside the package itself. Third parties cannot publish their own learners against a stable protocol. Python only.

EconML (Microsoft) provides RieszNet, ForestRiesz, and an automatic_debiased_ml module. The forestriesz package in this family wraps EconML’s BaseGRF. EconML is monolithic, with no third-party learner protocol, and Python only.

DoubleML (Bach, Chernozhukov, Kurz, Spindler) is a mature DML library with parallel Python and R implementations. It expects the user to supply outcome and propensity nuisances using sklearn-compatible learners. Riesz regression is not the focal abstraction.

tlverse (van der Laan group) is an R-only family of packages (sl3, tmle3, lmtp, hal9001, …) organized around TMLE and SuperLearner. The meta-package + sibling-learners shape is the closest organizational match to this project.

What’s distinctive here:

• The Backend / MomentBackend split, exposed as a stable Protocol, lets a third-party learner package depend on rieszreg and ship as its own PyPI / CRAN release. New learners do not require a PR upstream.
• The split itself reflects two structurally different fitting strategies: augmentation-style (kernel ridge, gradient boosting via fit_augmented) vs moment-style (forests, neural nets via fit_rows).
• Cross-language Python + R coverage at the family level via R6 wrappers per package, not just bindings to a Python core.