In cognitive radio (CR) networks, unlicensed (cognitive) users can exploit the licensed frequency bands by using spectrum sensing techniques to identify spectrum holes. This paper proposes a distributed compressive spectrum sensing scheme, in which the modulated wide-band converter can apply compressed sensing (CS) directly to analog signals at the sub-Nyquist rate and the central fusion receives signals from multiple CRs and exploits the multiple-measurements-vectors (MMV) subspace pursuit (M-SP) algorithm to jointly reconstruct the spectral support of the wide-band signal. This support is then used to detect whether the licensed bands are occupy or not. Finally, extensive simulation results show the advantages of the proposed scheme. Besides, we also compare the performance of M-SP with M-orthogonal matching pursuit (M-OMP) algorithms.