Existence of best $M$-term approximation