Much of the math in the prior answers here is quite correct. And it is very correct, very learned, analysis addressing the wrong problem.
Consider that the sole point of a password is to be unguessable (ie, not findable) by an attacker. A password consisting of a single letter is pretty poor since it will fall in, on average, 13 guesses (we assume a 26 character alphabet here, sorry non-English speakers). Regardless of how randomly (ie, high entropy) that single letter was chosen. So, IF we assume brute force is an attacker's best choice, a long password, chosen entirely randomly, is a good choice. For the reasons covered in the prior answers.
But no human (Hans the clever horse might have been able to do better) can remember random stuff very well, and long random stuff still less well. This leads to writing the long random password down, and to the sort of attack one saw in War Games. Don't use Post-it notes stuck in clever places either.
Ok, so choose 10 (or 50 or whatever) names from the great cryptographers list, and concatenate them (alternating one backwards, then one forwards, and so on, if you like) to produce a password. Not in any dictionary anywhere, really long, probably not maximally random but the entropy estimate algorithms would probably rate it highly, ..., so a pretty good password, right? Actually, if your attacker knows you're a cryptography fan, there's no need to assume concatenated names of famous knitters, which makes a guessing attack easier than it might have been. Don't help your attacker, keep his work factor as high as possible. In the middle ground, famous quotes (not from philosophers, as you are known to disdain them -- perhaps an EWD quote, though?) probably aren't as good, and all three of your daughters' names are still worse.
All this is to make the point that a password which cannot be guessed is good, and one that can be guessed is bad. All else is obiter dicta, though perhaps useful. Rules of thumb for entropy estimates are tricky, though good ones might be helpful (how to estimate goodness in this respect?).
Lastly, a horrid example. The digits of pi are (surely it's been proved by now?!) random. So a bunch of them ought to be really good as a password (not easily remembered, but we can't have everything). Until someone gets the idea of trying some of the digit sequences, in which case the randomness is worthless, for password purposes. Entropy matters, but only in the context of expectation by the attacker community. Theoretical entropy is only tangentially related.