There are multiple reasons why it's better to link to abstract pages rather than directly to the PDF versions:
Mobile users and users with slow internet connection prefer an abstract page over a PDF file. (Also, downloading PDFs can redirect to other applications instead of staying in-browser; this should be an explicit choice on the part of the reader.) Often the abstract page is enough information, and there's no need to download the whole PDF file.
For publications, the PDF file is often behind a paywall while the abstract page is free.
For arXiv papers, the abstract page often contains useful information that is not present in the pdf file itself - such as whether and where the preprint was published, or if there is an updated version. The PDF doesn't usually contain a link back to the abstract page, so for the person clicking on the link it's more useful if it goes to the abstract page directly.
(The same is also true with published journal papers - the abstract page contains more metadata than the PDF, and it is often more cleanly collected and easier to find.)
When citing use DOI links (or some other form of permalink if there is no available DOI) unless the document you're linking to really doesn't have such a link. This is crucial to avoid link rot.
Linking to webpages that require log-in should be minimized.
Generally, it's always a good idea to include a full bibliographic reference, mentioning explicitly the author, title, and bibliographic details (journal, volume, page, year, etc.). If permalinks are unavailable this is essential, to make it possible to reconstruct the link in case of link rot.
Linking to private clouds, dropbox, etc, is for various reasons not acceptable on SE. (This can e.g. potentially be misused to update a SE post without propelling it to the front page. They are also extremely susceptible to link rot.)
For the Phys.SE policy for self-citations & self-linking, see this meta post.