That’s the estimate from what looks to be a respectable academic study based on a survey with a sample size of 33,000.
Main problems: registration errors and absentee-ballot problems. Another 2-4 million were “discouraged” by long lines or ID requirments. This is why it’s so hard to take seriously the right-wing whining about “voter fraud,” a problem that must be at least two orders of magnitude* smaller.
Looks like the new technologies did substantially reduce the undervote: from 2% in 2000 to 1% in 2008.
The report was prepared for the Senate Rules Committee. So it seems likely we’re going to get some action on this.
Update A reader rebukes me for overstating the problem of false-name voting: it must be at least three orders of magnitude smaller than the vote-deprivation problem, given that (1) there are almost no actual cases documented, despite heavy searching (and strong pressure from the Bush White House to find and prosecute “voter fraud”) and (2) it would be a stupid, risky, non-cost-effective way to try to steal an election.