Finite-Horizon Strategy in Infinite-Horizon Linear-Quadratic Discrete-Time Dynamic Games

Abstract: This paper explores a finite-horizon strategy, ``watching $T$ steps into the future and moving one step now,'' in an $N$-person infinite-horizon discrete-time linear-quadratic dynamic game. The game involves linear input/output/state dynamics and quadratic cost functions with heterogeneous discount factors. For the finite-horizon version, which forms the basis of the infinite-horizon game, we analyze the structure of the coupled generalized discrete Riccati difference equations related to the feedback Nash equilibrium (FNE) and derive a sufficient condition for the uniqueness of the finite-horizon FNE. Under this condition, the FNE can be efficiently computed via the proposed algorithm. In the infinite-horizon game, assume all players adopt this finite-horizon strategy. If the iterations of the coupled equations related to the FNE converge, and the invertibility and stability conditions hold, we prove the convergence of each player's total cost under the finite-horizon strategy, even when players use individual prediction horizons. Furthermore, we provide an explicit upper bound on the cost difference between the finite-horizon strategy and the infinite-horizon FNE associated with the limiting matrices, expressed via the distance between their feedback strategy matrices. This bound vanishes as $T$ tends to infinity, implying convergence to the infinite-horizon FNE cost. A non-scalar numerical example illustrates the convergence behavior.