Reading again, I don't think the distribution in TFA ends up logarithmically spaced. The "first zero bit" function is biased low, so elements don't have an even

Reading again, I don't think the distribution in TFA ends up logarithmically spaced. The "first zero bit" function is biased low, so elements don't have an even half-life: elements towards one end are preferentially deleted. Given uniformly-spaced elements that preference might initially move the distribution in the direction of log-spaced, but it won't trend closely to it. (I think...)