New paper links cost curvature to Frank-Wolfe stability in congestion games
A single curvature condition determines both monotonicity and learning stability in games.
Tobias Harks' new paper provides exact characterizations for universal monotonicity—strict, strong, and basic—in atomic splittable congestion games. The key is a curvature inequality that links the first two derivatives of resource cost functions with the number of players. This framework gives a complete picture of when the variational inequality operator is monotone.
Remarkably, the same curvature condition also governs the stability of Euclidean-regularized Frank-Wolfe dynamics. For any convex strategy space and cost classes closed under positive affine transformations, the condition ensures both local and global stability. For simplex-based games, an interior equilibrium remains locally exponentially stable even without the curvature condition, offering a robust fallback.
- Curvature inequality involves first & second derivatives of cost functions and player count.
- Same condition characterizes monotonicity AND universal stability of Frank-Wolfe dynamics.
- On simplices, interior equilibria are locally exponentially stable regardless of curvature.
Why It Matters
Unifies game-theoretic monotonicity and learning dynamics, enabling predictable behavior in distributed resource allocation systems.