Connected Superlevel Set in (Deep) Reinforcement Learning and its Application to Minimax Theorems