Kevin Smith has published "Setting the Record Straight about Elsevier" in Scholarly Communications @ Duke.
Here's an excerpt:
Each [article] version is a revision of the original, and the copyright is the same for all these derivatives. When copyright is transferred to a publisher, the rights in the entire set of versions, as derivatives of one another, are included in the transfer. Authors are not allowed to use their post-prints because the rights in that version are not covered in the transfer; they are allowed to use post-prints only because the right to do so, in specified situations, is licensed back to them as part of the publication agreement.
Once a copyright transfer has been signed, all of the rights that the author may still have are because of specific contractual terms, which are usually contained in the transfer document itself. In short, these agreements usually give all of the rights under copyright to the publisher and then license back very small, carefully defined slivers of those rights back to the author. One of those slivers is often, but not always, the right to use a submitted version, or post-print, in carefully limited ways.