Pareto results

Just so it doesn't look like I've made no progress when I meet with my adviser tomorrow, I ran the algorithms with a Pareto distribution. The significance of using Pareto rather than Laplace is that Pareto has infinite variance (but finite mean). Laplace just has a really big variance. That appears to matter.

CISS is dealing with it fine (stripping off the tails brings us back to finite, but large, variance). The Bag of Little Bootstraps goes completely haywire. Bootstrap-Metropolis-Hastings is somewhere in the middle. Using a query that introduces correlated filtering doesn't change things much one way or the other.

So, I think this is a better illustration if we're only going to use three distributions. I suppose there's no downside to showing all four, other than all these graphs are making the paper pretty long without adding much real meat.