Dual variable recovery

[Transurgeon]

Some canonicalizations eliminate constraints and add new ones. We will need to do dual variable recovery for those.
This is also necessary for the GSoC project 2026 for optimality conditions checks for DNLP.

[dhru189]

Hi @Transurgeon, I'm interested in working
on this issue as part of my GSoC 2026
application for the post-solver feasibility
and optimality checks project.

I have already contributed:
→ PR #6 to cvxpy/solver-benchmarks
→ PR #3204 to cvxpy/cvxpy (DQCP fix)

I am currently studying the canonicalization
pipeline and Aryaman's 2023 GSoC work to
understand dual variable recovery.

Could you point me to the relevant files
in the codebase to start working on this?

[dhru189]

Hi @WilliamZijieZhang @PTNobel!

I've been working on the dual variable recovery for PSD/NSD attribute constraints we discussed here. Just opened PR #3214 (or whatever number) which introduces Problem.get_attr_dual(var)— ,
The first API to expose dual variables for PSD/NSD constraints.

Example:
X = cp.Variable((3,3), PSD=True)
prob = cp.Problem(cp.Minimize(X[0,0] - X[1,1]), [cp.trace(X) == 1])
prob.solve(solver=cp.SCS)
print(prob.get_attr_dual(X))

[[ 2. 0.]

[ 0. ~0.]]

You mentioned there was no API yet — happy to discuss the design before it gets reviewed!