Working on the arXiv version of a published paper, how should I cite?

If at all possible you should cite the most recent official (published, peer reviewed) version.

You could consider contacting the author to ask if the arXiv version differs in any way that might matter in your context. A librarian might be able to help with that too.

You can note in your bibliography that a version of article is also available on arXiv.

This is a known tricky situation, which I've often seen addressed incorrectly (people citing the published version, but actually meaning the arXiv version).

My personal way of solving this is: I cite the published version in the bibliography, but when I actually reference it, I add a footnote saying that I'm referring to the arXiv version. This is honest and helps the reader (who likely has the same troubles finding the published version as you do). Some editors might not like such citations, but keep in mind that, just as you do, the readers of your paper will probably be reading its arXiv version, so whatever mess the editors make of your references will have a limited effect (and won't be your fault anyway).

Of course, you should handle arXiv preprints with the usual care: if at all possible, check their work and point out any errors that may significantly impede their use. Also, as @EthanBolker pointed out, it's a good idea to contact the author asking for the precise relation between the arXiv version and the published one (and subtly hinting to update the former).

In fields where arxiv usage is standard (e.g. high energy physics or astronomy), standard practice is to cite the arxiv identifier of the article along with any journal reference. (Journals tend to remove them upon publication, but that is very much their problem.) Some of the most commonly used abstract indexing services in those fields (e.g. inSPIRE or NASA ADS) in fact automatically merge the records of the arxiv and the journal version. Consequently, if you use the bibtex entries exported from those services, you will typically get bibliography entries that refer to both the journal version and the Bibtex version.

This, of course, only applies if you are reasonably sure the article on arxiv and in the journal are actually the same article (same title and authors may not always be enough to guarantee this). In many cases this is easy to tell because the arxiv entry will have a link to the published version (Many publishers in fact require this link as a condition for authors being allowed to also post the pre/post-print to arxiv.) The presence of a link to the journal version does not guarantee that the arxiv version is identical (contentwise) to the published version (although an increasing number of funding agencies now require this in order to satisfy their open access requirements).

In the specific case presented by the OP there is no link to the journal version in the arxiv version. Hence we cannot be exactly sure they are the same article (although identical abstracts and similar length make this extremely likely). Moreover, the arxiv version was updated after the journal version was published online, making it highly likely that it was updated to match the journal version.