form: http://blog.sina.com.cn/s/blog_60021d3e0100f87j.html
This parameter is deliberately left vague in the DX documentation, because it is mostly up to the driver to define what it means.
It is needed because even within a fixed number of samples, there can be different ways of implementing the details. For instance with 2x AA, do you just double the width of the backbuffer, or do you rotate the sample kernel? When downsizing, do you just average two samples for each output pixel, or perform some more sophisticated (and hence expensive) filter to get better results? For instance in the GeForce3/4 NVidia implemented Quincunx filtering, which renders with a rotated 2x AA kernel but then samples five surrounding values to compute each output pixel.
The API needs a way to control which option you want, but it doesn't really know what options the driver may have implemented, so MultiSampleQuality gives a way for the driver to report "ok, I support 3 different ways of doing 2x AA", and then the game can say "right, I want the second of those".
Of course, it's not easy to make a sensible choice here because you don't really know what the different options represent (unless you happen to recognise the card) so most often this value should be left alone. Plus, even if you do try to specify it a lot of drivers do crazy stuff like overriding what the game requests based on control panel settings made by the user, at which point the whole thing becomes irrelevant in any case!
AND It doesn't mean anything on the Xbox.
The 360 implements multisampling directly inside the special memory chip that holds your framebuffer data. This makes it obscenely fast, but not especially flexible: you can only have two samples or four samples, and don't get any control over the filtering algorithm.
http://social.msdn.microsoft.com/forums/en-US/xnaframework/thread/f211a69e-9c2d-4176-af67-3c7dfc1a18f7/
http://www.gamedev.net/community/forums/topic.asp?topic_id=432756