Random failures of sqmatrix_assert_hermitian due to hard-coded error threshold

[oskooi]

The check in sqmatrix_assert_hermitian to verify the Hermitian property of the Maxwell operator occasionally fails CI due to a larger than expected error of the hard-coded value of 1e-10 from this line:

mpb/src/matrices/sqmatrix.c

Line 54 in 718f3fe

CHECK(err < 1e-10, "sqmatrix_assert_hermitian failed");

Since it is difficult to find a reproducible test case, there are two possible workarounds:

1. Increase the threshold error value of 1e-10. Choosing an appropriate value may involve outputting the actual error which caused the test to fail otherwise we choose some arbitrary value.
2. Disable this test entirely since it is not that useful.

[stevengj]

Since it is a debugging assertion, it could be enabled only in debug mode, i.e. surround it with #ifdef DEBUG

[oskooi]

Since it is a debugging assertion, it could be enabled only in debug mode, i.e. surround it with #ifdef DEBUG

It turns out that this check is surrounded by #ifdef DEBUG. We should therefore probably just increase the error threshold since 1e-10 is too small.

mpb/src/matrices/sqmatrix.c

Lines 37 to 57 in 718f3fe

	#ifdef DEBUG
	double err = 0, maxsq = 0;
	int i, j, p = A.p;
	for (i = 0; i < p; ++i)
	for (j = 0; j < p; ++j)
	maxsq = max2(maxsq, SCALAR_NORMSQR(A.data[i*p + j]));
	for (i = 0; i < p; ++i) {
	err = max2(err, (SCALAR_IM(A.data[i*p + i]) *
	SCALAR_IM(A.data[i*p + i])) / maxsq);
	for (j = i + 1; j < p; ++j) {
	scalar x;
	ASSIGN_CONJ(x, A.data[i*p + j]);
	ACCUMULATE_DIFF(x, A.data[j*p + i]);
	err = max2(err, SCALAR_NORMSQR(x) / maxsq);
	}
	}
	CHECK(err < 1e-10, "sqmatrix_assert_hermitian failed");
	#else
	(void) A; /* unused */
	#endif