Reciprocal Adversarial Learning via Characteristic Functions