Fix Boost tests

[jngrad]

BOOST_CHECK_CLOSE takes a percentage as tolerance. Epsilon values must be multiplied by 100.

In this example, the first line is incorrect:

espresso/src/utils/tests/int_pow_test.cpp

Lines 43 to 44 in d553bd5

	BOOST_CHECK_CLOSE(x * x, int_pow<2>(x), eps);
	BOOST_CHECK_CLOSE((x * x) * (x * x), int_pow<4>(x), 100. * eps);

In this example, the line is correct:

espresso/src/utils/tests/interpolation_test.cpp

Line 254 in d553bd5

BOOST_CHECK_CLOSE(interpolated_value, exact_value, 1.);

This affects both 4.1 and 4.2.

[fweik]

I don't understand the boost documentation.

[jngrad]

According to section tolerance-based comparisons, function BOOST_CHECK_CLOSE does a relative comparison and evaluates to assert(abs(x - y)/abs(x) <= tol && abs(x - y)/abs(y) <= tol) when both x and y are non-zero.

In the binary, ::boost::math::fpc::percent_tolerance(std::numeric_limits<double>::epsilon()) is used as the epsilon value, where percent_tolerance() divides by 100 while percent_tolerance_t() doesn't (according to section tolerance (decorator)). Here is a MWE:

#define BOOST_TEST_MODULE tolerance test
#define BOOST_TEST_DYN_LINK
#include <boost/test/unit_test.hpp>
/* compile with g++ test.cc -lboost_unit_test_framework */
BOOST_AUTO_TEST_CASE(tolerance)
{
  BOOST_CHECK_CLOSE(1, 1.01 - 1e-8, 1); // passes: slightly less than 1%
  BOOST_CHECK_CLOSE(1, 1.01 + 1e-8, 1); // fails: slightly more than 1%
}