KKT Reformulations for Single Leader and Multi-Follower Games
Parin Chaipunya, Thirumulanathan D, Joydeep Dutta
TL;DR
This work analyzes a single-leader, multi-follower bilevel game where followers respond by solving a $u^{\mathtt{l}}$-dependent Nash equilibrium. It derives a KKT/MPCC reformulation to convert the bilevel problem into a single-level MPCC and proves local equivalence between the SLMFG and MPCC under convex follower problems and Slater's condition, while illustrating the necessity of these conditions with counterexamples. It then offers a CRCQ-based, practically verifiable criterion for local equivalence and discusses broader global equivalence under stronger convexity assumptions. Finally, it extends the analysis to generalized Nash settings with shared constraints (RGNEP) and to grouped follower structures, showing that RGNEP reduces to NEP under joint convexity and that CRCQ-based insights extend to these cases, with cautionary examples highlighting the limits of the approach when assumptions fail.
Abstract
We consider a bilevel optimization problem having a single leader and multiple followers. The followers choose their strategies simultaneously, and are assumed to converge to a Nash equilibrium strategy profile. We begin by providing a practical example of such a problem in an oligopoly setting. We then show the existence of a Nash equilibrium when the objective function of each follower is convex in its optimizing variable, and the feasible set is compact, convex, and nonempty. We then consider the KKT reformulation of the single leader multi-follower game (henceforth, SLMFG), and show using examples that the solutions of both the problems need not be the same, even when each of the followers' problem is convex. In particular, we show that the global minima of both the problems may differ if the follower's problem does not satisfy the Slater's condition. We then show that the local minima of the SLMFG and its KKT reformulation are the same if, in addition to convexity and Slater's constraints, the local minimum point remains a local minimum for every Lagrange multiplier in each of the followers' problem. Given that this condition is hard to verify in practice, we provide another condition for the local minima of the two problems to be the same using constant rank constraint qualification (CRCQ). We again show using examples that the local optima of the two problems may differ if the conditions are not satisfied.