Does replacing an exact inner optimization by a learned approximation change the estimand?

This problem is visible to members.

Membership is open to holders of a gatech.edu address, approved individually. The statement, the solution claims, and the discussion are all behind that gate.