arXiv's rule that catches people is the format one: with PDFLaTeX your figures must be .pdf, .jpg or .png; with classic (La)TeX they must be .ps or .eps, and arXiv explicitly does not convert figure files for you. That means the compression has to happen on the raster files before they go into the source archive, because a 12 MB PNG multiplied across fifteen figures is a submission archive that is slow to build and slow to fetch, and no amount of PDF optimisation downstream undoes it. The other thing worth knowing is that arXiv renders whatever resolution your raster carries — figures exported at screen size look soft in the HTML and PDF versions. The reliable order is: export each figure at the width it will occupy on the page, at around 300 dpi, compress the resulting PNG or JPEG to a sane weight, then let LaTeX embed it. Skwish compresses JPG, PNG and WebP in the browser with a per-file before/after readout, which makes it easy to find the two or three figures eating most of the archive. Nothing is uploaded, so pre-print material stays local.