The day in conversation for 2026-11-02

discussion

Transcript

The discussion: Kolya: The claim about FOMTools achieving two orders of magnitude speedup over NumPy for sparse coding problems seems significant, but how does that comparison hold up when you consider the complexity involved in representing those iterations as computation graphs? Ada: It’s the efficiency gain on hardware that matters most, especially since they are generating specialized solver code for GPUs; I’m less concerned with the overhead of graph representation itself and more interested in whether this speedup scales linearly or degrades when we move from sparse linear systems to truly high-dimensional convex quadratic problems. Emmy: From a purely geometric perspective, I'm trying to reconcile that computational speed with the structural theorems being proven simultaneously. Take Alves da Silva’s work on the Sarkisov Program; proving it is automatically volume-preserving for elements of that specific decomposition group suggests a deep underlying rigidity in those cubic structures, regardless of how fast we can numerically solve them. Kolya: But that volume preservation result relies entirely on the assumptions about the decomposition group structure—specifically, what constraints must be placed on the underlying manifold to guarantee that property holds true when you apply it to elements derived from a nonsingular plane cubic. What if we relax those constraints slightly? Ada: Relaxing them might make the solver code generation less specialized, which would negate that speed advantage. My concern is whether the two orders

More episodes

← Home