As has already been noted: there are two reasons to cite a paper, an email, a letter,
or anything else:

(1) To give credit.

(2) To refer to a result you need.

If a citation is for reason (1), you should cite anything and everything that is appropriate.
If someone explained a result to you in an email, cite them, or at least acknowledge them in the paper. If you rely on the results of a preprint, or a preprint proves important results germaine to your own work, you should cite it (whether it appeared on the arxiv or not!).

As for reason (2), it is up to you. In general, you shouldn't be relying on results unless you are confident that they are true. Typically, it is up to you to determine your own threshhold of confidence, and if publication in a peer-reviewed journal increases your confidence, you can take that into account. But in this context you must still take into account
criterion (1): i.e. even if you avoid relying on an unpublished result out of a sense of caution, but that unpublished result is closely related to what you are doing, you
should still mention it. (A typical situation might be that you need a special case of
an unpublished result that it is easy for you to prove directly, and you prefer to do so
rather than rely on the unpublished result; then you can certainly do so, but you should
point out that what you have proved is a special case of the more general result, and cite
that more general result.)

One thing to note is that some journals may not accept your paper if it relies crucially on unpublished results (including unpublished results of yours!). Thus, if you rely on such results, you may want to include a sketch of the proof, so as to make your presentation somewhat self-contained. In this case, if you giving a sketch of the proof of someone elses result, be sure to include appropriate citations from category (1) above!

shouldor not, it's a fact that many mathematiciansdocite arxiv papers. $\endgroup$ – Robin Chapman Oct 5 '10 at 13:05