Establish equilibrium convergence guarantees

Establish equilibrium-convergence or approximate-Nash guarantees for the policies produced by the optimistic dual robust Bellman framework, beyond its current one-sided robust-value regret guarantee.

Background

The paper defines regret using a signed one-sided difference between the robust minimax value and the value of the jointly generated policy pair. It explicitly distinguishes this criterion from two-sided saddle exploitability and states that sublinear regret under the chosen comparator does not itself certify an approximate Nash equilibrium. Consequently, convergence of the learned policy pair to an equilibrium remains unresolved.