Qiskit / qiskit-fermions / 30530363564

98%

push

github

perf(operators): compute group weights natively for all operator types (#235)

Sampling weights for a grouped Hamiltonian were reduced in Python: pull
the full per-term coefficient array and the full per-term group-index
array across the FFI boundary, then collapse both down to one value per
group in NumPy, plus a second pass to normalize by each group's term
count. The data crossing into Python therefore scaled with the *ungrouped*
term count purely so it could be aggregated straight back down again.

`group_weights()` does that reduction in a single native pass and hands
back only `num_groups()` values. On the N2 Hamiltonian from
docs/guides/sqdrift.rst (251,468 terms in 75,616 groups) this cuts the
weight computation from ~30ms to ~4ms per call.

A group index carried by no term weighs 0.0 rather than dividing by zero.
Index ranges are dense by construction, but nothing enforces that, and a
single NaN would poison the whole probability vector downstream.

Refs #229

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

64 of 73 new or added lines in 9 files covered. (87.67%)

10431 of 10632 relevant lines covered (98.11%)

312.11 hits per line