Austin Conner

icon indicating a website link people.math.harvard.edu/~aconner icon indicating an email [email protected]

icon company Harvard University icon location Cambridge, MA icon location I study complexity and invariants of tensors and polynomials, particularly with applications to theoretical computer science.

Results 6 comments of Austin Conner

Implement Normal, a solver applying an inner solver to the normal equations

One question about the implementation of `allow_dependent_{rows,columns}`: Currently I implement them defensively with regard to the possibility that the solver state might be transposed, taking into account the possibility that...

Replace allow_dependent_rows and allow_dependent_columns with assume_full_rank

If you wanted to go further than this PR and did not mind a breaking change, you can make full_rank an operator tag. Then solvers requiring full_rank operators check this...

Replace allow_dependent_rows and allow_dependent_columns with assume_full_rank

> If Lineax defaulted to assuming singularity (as you seem to be suggesting here) and you left off the `full_rank` tag then you'd get silently expensive JVPs. Or conversely if...

Replace allow_dependent_rows and allow_dependent_columns with assume_full_rank

> Actually, there is one further problem here (and one that is a motivating reason for us not to do this ): if a user has what they believe to...

Replace allow_dependent_rows and allow_dependent_columns with assume_full_rank

I now understand your concern. At the very least the current allow_dependent_{rows,columns} are misnamed. Its really more like allow_nonconstant_{row,column}_space, as row span A(x) or column span A(x) is independent of...

Bug: NormalCG jvp can do CG on singular systems

One possibility to obviate both this issue and #156 is to make a lx.Normal wrapper taking a solver as input and returning another solver, and setting `lx.NormalCholesky = lx.Normal(lx.Cholesky())` and...