We study high-dimensional sparse linear regression: discover the exact sampling complexity when the data is intrinsically sparse, and establish a sufficient condition on sample size when sparsifying originally dense data.