Spontaneous subtle emotions are expressed through micro-expressions, which are tiny, sudden and short-lived dynamics of facial muscles; thus poses a great challenge for visual recognition. The abrupt but significant dynamics for the recognition task are temporally sparse while the rest, irrelevant dynamics, are temporally redundant. In this work, we analyze and enforce sparsity constrains to learn significant temporal and spectral structures while eliminate irrelevant facial dynamics of micro-expressions, which would ease the challenge in the visual recognition of spontaneous subtle emotions. The hypothesis is confirmed through experimental results of automatic spontaneous subtle emotion recognition with several sparsity levels on CASME II and SMIC, the only two publicly available spontaneous subtle emotion databases. The overall performances of the automatic subtle emotion recognition are boosted when only significant dynamics are preserved from the original sequences.