Document the DFT/NUFFT trade-off on the sparse-operator setup path

[Jammy2211]

Summary

Closes the last open item from the pynufft removal (#475): whether apply_sparse_operator is still incompatible with TransformerNUFFT.

It is not, and has not been since May. The NotImplementedError guard was removed by #329 (bd18a769, 2026-05-22); only a stale workspace comment still claimed otherwise. Verified directly: apply_sparse_operator runs with TransformerNUFFT at every scale tried, and its dirty image matches TransformerDFT to ~3e-13 relative.

Both paths are supported. What was undocumented is which one you should actually pick — and the answer isn't the visibility count.

The crossover is the product, not the visibility count

The DFT setup costs O(N_vis × N_pix); the NUFFT costs O((N_vis + N_pix) log N) plus a fixed ~2 s overhead. So the product decides it. Measured on CPU (use_jax=False), seven points agree on a crossover near 1e7:

N_vis × N_pix	DFT / NUFFT time
1.3e6	0.21×
2.6e6	0.27×
5.2e6	0.70×
1.0e7	1.16×
2.1e7	1.50×
2.3e7	1.41×
8.3e7	1.92×

At a typical 64×64 mask that's ~5,000 visibilities — but on a 32×32 mask the DFT still wins at 4,000. Quoting a visibility count alone would be wrong on half the grids.

Memory is the decisive axis

At fixed N_vis=4000, growing the grid:

grid	N_pix	DFT ΔMB	NUFFT ΔMB
32²	648	60	81
64²	2 608	239	0
96²	5 860	293	0
128²	10 428	446	0

The DFT allocates with N_vis × N_pix; the NUFFT allocates nothing measurable beyond its working buffers. At 10.7 bytes/element that extrapolates to ~109 GB at 1M visibilities — independently corroborating the ~123 GB figure bd18a769 recorded by a different route, and confirming the NUFFT is the only feasible ALMA-scale path.

Why an INFO and not a warning

Interferometer defaults to TransformerNUFFT (dataset.py:34). A "NUFFT below the crossover" warning would therefore fire on the default path for every small dataset — every tutorial, test and workspace example. That's warning fatigue for a ~2 s cost.

So the guidance is an INFO on the existing setup log, emitted only when a NUFFT-backed transformer is used below the crossover: an explicit, expensive, opt-in call, never fired when the advice would be moot. The genuinely dangerous direction is already a hard DatasetException above 10,000 visibilities with an opt-out, so nothing is left unguarded.

API Changes

None — internal changes only. One added INFO log and docstring text; no signature, symbol or behaviour changes.

Test Plan

• Advisory fires for small+NUFFT; stays silent for small+DFT and for large+NUFFT (all three checked explicitly)
• pytest test_autoarray — 1179 passed
• Diff is purely additive (29 insertions, 0 deletions) — the repo's black version wants to reformat an unrelated DatasetException block, and that churn was deliberately kept out

---

Generated by Claude Code