Warren Smith of the Center for Range Voting has produced some new graphics extending zestyping's analysis here... He seems to have used either a higher spread, or fewer trials (thus less tendency to "revert to the mean"), since there seems to be more fuzziness on the edges of various "win regions". Also, if I'm correctly understanding the first row of four pictures in the 14-candidate example, this example makes a fairly compelling case that Range may be better than Approval for single-winner, depending on your strategy assumptions about voters. While I like AV's tendency to produce some types of compromise, and disfavor candidates who are hated by significant chunks of the population, in cases like the far upper-left area of the fourth image, it seems like even with the vast majority of voters being more "upper-leftist" than the upper-left-most candidate, the system is still opting for a compromise with the tiny minority of "lower-rightists". That definitely seems unfair. On the other hand, one can make other strategy assumptions -- even that voters place their Approval threshhold randomly -- and have AV come out looking a lot better. (Later on, he admits that strategy considerations impact RV as well... Hmm. Have to process this...)
Of course, the demonstration of IRV's tendency to squeeze out centrists in favor of fringe candidates, even when public opinion is centrist, is even more compelling.