Have you read the papers I linked in detail? Some of them, such as HyperLogLog, provide corrections to give better estimates for small sets, and although I can't follow the proof in its entirety, they claim to be more efficient than the alternatives, including the one you propose.