Learning sparse systems at sub-Nyquist rates: A frequency-domain approach