Towards Sharp Minimax Risk Bounds for Operator Learning