Hi @vincent,
A useful reference that has a brief section on partial signal subtraction is Prof. Scheres’s chapter in Methods in Enzymology, Processing of Structurally Heterogeneous Cryo-EM Data in RELION. It mentions the limitation in that signal subtraction makes the assumption that the current poses are correct, thus any inaccuracies in the poses will result in subtracting slightly mis-aligned references. The more accurate the alignments are, the better the subtraction will work, so it can be beneficial to do subtraction after a round of local refinement (Oli described this workflow nicely in this thread).
Independent of that issue, @zhenyu_tan’s comment on heterogeneity is the most likely explanation behind why you are seeing residual density. Each image contains a different micelle due to disorder, which breaks the assumption behind subtraction (i.e. that each image contains a discrete number of rigid bodies). For that reason, we typically don’t recommend using subtraction for micelles (see here for another thread on this topic).
Best,
Michael