In microarrays, expression is measured as fluorescence intensities from probes that hybridise to their target cDNA sequences. Virtually all probes will fluoresce and return values.
Genes that are not expressed may fall into the background noise (or be slightly above it), the level for which is then corrected during RMA normalisation, the typical normalisation strategy. Background noise can be inferred from, for example, the expression gained from negative control probes. In RNA-seq, conversely, genes that are not expressed will have 0 count values.
You should ideally not do any filtering prior to running limma. Only probes that have been shown to repeatedly fail should be removed. Probes covering genes that you definitively know to not be expressed in your tissue of study could also be removed. It can be inferred that, after normalisation, those genes with the lowest values are reflective of genes that have virtually nil expression.