Backend Support Matrix
May 24, 2026 ยท View on GitHub
This page is generated from tools/backend_support_matrix.py. It records a
smoke-test-backed support snapshot for selected public APIs. It is intentionally
conservative and does not prove full mathematical or numerical parity.
| Area | Capability | NumPy | PyTorch | JAX | Notes |
|---|---|---|---|---|---|
| backend.random | Seeded scalar/vector normal sampling | yes | yes | yes | JAX uses a process-global PRNG key unless explicit state is passed. |
| backend.random | Weighted choice with replacement | yes | yes | partial | JAX support depends on argument form and should be covered by focused tests. |
| backend.random | Weighted choice without replacement | yes | yes | partial | PyTorch support is smoke-tested with probability vectors via torch.multinomial. |
| distributions | GaussianDistribution.pdf / ln_pdf | yes | yes | yes | Smoke-tested with reference values. |
| filters | KalmanFilter.predict_linear / update_linear | yes | yes | yes | Backend-portable linear algebra path. |
| filters | UKFOnManifolds.predict / update | yes | yes | no | JAX is explicitly rejected by this API. |
| utilities | SciPy-heavy tracking/evaluation helpers | yes | partial | partial | Check NumPy behavior first for advanced workflows. |