---

name: numpy-scipy-primitives description: "Routes autograd.numpy and autograd.scipy wrapper behavior, supported and unsupported NumPy patterns, complex numbers, xarray interoperability, and SciPy-extra troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

NumPy and SciPy Primitives

Use this sub-skill when the user asks about autograd.numpy or autograd.scipy behavior rather than about core derivative operators or custom rules. It is the right place for wrapper support questions, xarray __array_ufunc__ interop, and public limitations around NumPy/SciPy patterns.

Route here for

• autograd.numpy wrapper behavior, supported NumPy calls, and common array-construction helpers.
• autograd.numpy.linalg and autograd.numpy.fft behavior, including complex-number support and known shape/axis limits.
• autograd.scipy.special, signal, linalg, integrate, and stats wrappers, including logsumexp.
• NumPy containers or array-like objects that implement __array_ufunc__, especially xarray.DataArray.
• Troubleshooting whether a failure is a missing optional dependency or an unsupported wrapper pattern.

Route elsewhere for

• grad, jacobian, value_and_grad, make_vjp, make_jvp, Hessians, and other core derivative operators: differentiation-core.
• New primitives, custom VJPs/JVPs, or wrapper authoring: extend-primitives.

Read first

• API reference
• Workflow recipes
• Troubleshooting

Safe helper

Run the bundled wrapper smoke helper from this sub-skill directory or any shell that can resolve the script path:

python scripts/wrappers_smoke.py
python scripts/wrappers_smoke.py --simulate-missing scipy
python scripts/wrappers_smoke.py --simulate-missing xarray

The helper uses tiny in-memory arrays only. It does not download data or require network access. When SciPy or xarray are absent, it prints a clear skip message with the suggested install step instead of failing immediately.

Core rules

1. Prefer autograd.numpy for user-facing array code. Keep the final result scalar before calling differentiation operators from the core sub-skill.
2. Use np.dot, np.matmul, or np.tensordot; do not rely on the A.dot(B) method form.
3. Do not mutate differentiable arrays in place. Rewrite assignments and += style updates as pure expressions.
4. When a primitive takes a list or tuple, build an explicit array first with np.array(...) unless the wrapper already documents a safe list path.
5. For container checks on boxed values, use the Autograd builtins versions of isinstance and tuple when needed.
6. xarray.DataArray and other __array_ufunc__ containers can carry boxed values through NumPy ufuncs. Keep the container during the ufunc step, then reduce to a plain scalar with .data or an equivalent extraction at the end.
7. autograd.scipy requires SciPy. If the import is missing, install autograd[scipy] or scipy itself.
8. Complex numbers are supported, but non-holomorphic questions and low-level derivative semantics belong in the differentiation-core route.
9. Some matrix norms, repeated-axis FFT gradients, and odd-length real FFT gradient paths are unsupported. Treat NotImplementedError as a limitation signal, not a bug in the wrapper layer.

What this sub-skill does not do

• It does not teach custom primitive authoring or gradient-rule registration.
• It does not explain the full derivative-operator API surface.