They don't know, they guess, based on some heuristic algorithm.
In Google's case, their algorithm informed my son that the name on his birth certificate is not a "real name", and forced him to pick another one (he kept his real name as his nickname, and chose his more popular middle name as his first name, keeping his real last name). According to http://www.babynamewizard.com/voyager, his real first name is the 280th most popular in the US. http://howmanyofme.com/search/ says there are nearly 10,000 people with his first name, and a bit less than 4,000 with his last name in the US. But Google says it's not his real name.
They won't necessarily know, but the consequences can be really, really shitty if you get flagged (correctly or incorrectly) for violating a real name policy.
Australian guy has the full, legal name "Stilgherrian." Google+ eventually flags his account has having a fake name, and suspends him.
Now, imagine you have an uncommon name, and you've been using Facebook Connect to log into a bunch of sites. One day, your account gets suspended. Not only have you lost access to your social connections, but you've also been locked out of those third party sites. Because Facebook didn't think your name was "real" enough. That sucks for you, and it sucks for the businesses you were Connecting with.
If it was your real name, you may be able to provide passport scans and other documents to recover your account. If it was a pseudonym, you may be completely locked out.
How could someone's real name always fall "within some expected value"? Someone who's real name is an outlier is going to be an outlier, they're not going to "fall within some expected value."
If they're a real person, they're going to have all the other meta data a real person would have (which can range from a ton, to almost none, depending on how they feel about sharing every detail of their life with Facebook). They're going to have that same range of meta-data whether they call themselves John Smith, or Sister Boom-Boom. So how does that make any difference?
I would guess that actual fake accounts (spam bots, etc.) use common everyday names, so the "real name" policy isn't going to have any impact on them.
It's not Facebook going out of their way to target the subjects of prejudice, it's Facebook not bothering to care about the potential for abuse of the "real name" enforcement system they set up. "I'm just handing a loaded gun to the bigots, I'm not shooting anybody!" Oh, well that's OK then.
