New Insight of Variance reduction in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions