Uh, do you think pytorch doesn't use IEEE754 floating point math? Or that glm doesn't??

Furthermore, what statistical problem--aside from computing the fine-structure constant--needs more precision than you get from standard floating point? If you need it to be fully IEEE754 compliant, because you're testing bitwise accuracy not whether your computations are producing sensible results, from 17 on the JVMs do it by default (instead of the StrictMath thing that they used to have, and instead of the strictfp flag).

This is a really weird complaint.

Furthermore, Python and R have a recursion limit, so why are you complaining about the JVM?! Not to mention that, unlike Java and Python, Scala at least has tail call optimization, which makes it strictly better at recursion than the other options you say are superior.

And then you say, "With that as a note, Java nor SBT (Scala Build Tool) were necessarily designed with statistics in mind." Why are you grouping a build tool and a programming language?!

It doesn't seem like you know the domain. If this is not the case--you were tired, inattentive, published without reviewing and editing--maybe you want to fix this up a bit?