WEBVTT

00:00:00.000 --> 00:00:07.869
A finite answer becomes infinity. This happens in an Apple MLX CPU test with a simple two by two matrix.

00:00:07.869 --> 00:00:15.502
Scale the identity matrix by two to the power eighty. The correct log determinant is about one hundred ten point nine.

00:00:15.502 --> 00:00:25.069
The small matrix path computes a determinant before its logarithm. That intermediate overflows, even though the requested answer is representable.

00:00:25.069 --> 00:00:32.736
A clean C++ build reproduces one cause: one hundred twelve failed forward checks out of eight hundred twenty four.

00:00:32.736 --> 00:00:42.874
Widening single precision inputs to double precision passes all four hundred twenty eight single precision checks. It does not solve the double precision defect.

00:00:42.874 --> 00:00:54.853
Another candidate fixes forward values but breaks derivative checks. No complete compatible fix is claimed. Code and logs are published. GPU and model impact remain untested.