stillwater-sc / universal / 30414562427

85%

master: 84%

push

github

fix(interval): round arithmetic outward to restore the containment property (#1246)

* fix(interval): round arithmetic outward to restore the containment property

interval<Scalar> performed its arithmetic under round-to-nearest, so the computed
enclosure could be narrower than the true result range -- the Fundamental Theorem of
Interval Arithmetic (the only reason to use an interval type) silently failed. The
classic symptom: interval(0.1) * interval(0.1) returned a ZERO-width interval that did
not contain the exact product.

Fix (Stage 1 of #1234): round every inexact endpoint OUTWARD -- lower bound toward
-inf, upper bound toward +inf -- via nextafter. nextafter is used rather than the FPU
rounding mode (fesetround) because it is Scalar-generic: the hardware rounding mode has
no effect on posit/cfloat/lns arithmetic, and interval<posit<...>> is a target use case.

Outward rounding applied to: operator+= -= *= /= (the reciprocal endpoints too), sqr(),
sqrt(), and rounded UP in width()/rad() so mid()+/-rad() covers. Constructors enclose a
value outward only when the conversion to Scalar is inexact (round-trip check), so
exactly-representable operands keep zero-width endpoints (interval<double>(2) stays [2,2])
while an inexact cross-type cast (interval<float>(0.1)) is enclosed. The value/range
constructors, rad() and width() are no longer constexpr (nextafter is a runtime op).
Free operators and scalar-mixed overloads inherit the fix (they delegate to members/ctors).
Unchanged (already exact): unary -, abs, intersect, hull, comparison/contains predicates.

Tests: the existing addition/subtraction/multiplication/division suites asserted EQUALITY
on small integer operands (vacuous -- no rounding ever occurred), so a widening fix makes
them fail by design; converted them to CONTAINMENT (expected result must be a subset of the
computed interval). New containment.cpp exercises inexact operands (0.1*0.1, 1/3, sqrt(2),
an 8-term sum, cross-type ctor), a rand... (continued)

35 of 35 new or added lines in 1 file covered. (100.0%)

41536 of 48741 relevant lines covered (85.22%)

7162173.23 hits per line