IntelPython / dpnp / 39826719889

83%

push

github

Add a deterministic minres test for the eps stopping criterion (#3105)

The `istop = 3` branch of `minres` (`_iterative.py` L859, the `epsx >=
beta1` stopping test) is reached only by
`test_minres_x0_exact_solution`. There `x0` is the exact solution, so
the initial residual `beta1` is rounding noise (~1.5e-16) of the same
order as `epsx` (~2e-16), and whether the branch is hit depends on the
runner.

That is why Coveralls flips by one line currently on PRs with no source
change.

This PR adds `test_minres_x0_null_space`, which uses a singular `A =
[[1, 0], [0, 0]]`, a large `x0 = [0, 2**40]` in its null space and a
tiny `b = 2**-40`. All values are exact in fp64 and `epsx / beta1 =
2**27`, so the branch is always reached. The result matches SciPy's
`minres` for the same input.

1604 of 2990 branches covered (53.65%)

Branch coverage included in aggregate %.

29404 of 34540 relevant lines covered (85.13%)

19634.23 hits per line