You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Feature request: a cone block type that PSLP passes through untouched #134
Hi Daniel, a feature request from trying PSLP in front of SCS.
Many conic problems mix linear rows with cones that PSLP doesn't model, such as PSD, power or dual exponential cones. To presolve the linear part of such a problem, PSLP has to leave those rows alone: never remove, merge or reorder them, and never use them to derive bounds. Changes of variables are still fine (fixing a variable, or substituting one out through a doubleton equation), since they keep A_B x - r_B in K for any cone K.
It turns out this already works today. src/core/Cones.c says "EXP blocks are passed through unchanged", and the LP reductions skip rows that belong to cone blocks. So we declared PSD rows as PSLP_CONE_EXP triples, padding with empty rows to a multiple of three, and in our tests at 89d6762 that worked well:
On Netlib LPs with small PSD blocks added, PSLP reduced the linear part (bandm 807 x 482 to 642 x 232) and left the PSD rows in place. The postsolved point passed the KKT check on the original problem.
On DIMACS copo14 and on four SDPLIB problems with a linear part, the round trip was exact.
The catch is that we're relying on an implementation detail, and it would change if you add reductions specific to exponential cones (which you may well want to). Would you consider an explicit block type for this, say PSLP_CONE_OPAQUE? It would take any dimension, pass the rows through as they are, and have postsolve return the duals on those rows as given. That would let any conic solver use PSLP on the linear part of its problems.
Thanks, and happy to try out anything along these lines.
Activity
dance858 commentedon Nov 16, 2026
Great and simple suggestion. Fixed in #143.