These Bootstraps Are Made For Walking

Today I ran Nmix (100000 sweeps, burn-in of 100000) on a bunch of bootstrapped samples of the Willman I velocity data, this time only including the 40 with lower metallicities.

I'll summarize the results of the 12 trials that I just recorded:

1) This one had almost no discernible structure (~90% for more than 2 components)
2) Again, no discernible structure (~90% for more than 2 components)
3) No structure (~96% for more than 2 components)
4) About 10% confidence for 1 component, 75% for more than 2
5) ABSOLUTELY ZERO STRUCTURE
6) Lotta structure. 21% for 1 component, 55% more than 2
7) Reasonably high confidence for 1 component: 46%, ~30% more than 2
8) No structure
9) Around 10% for 1 component, about 75% more than 2
10) 13% for 1 component, 20% for 2, about 65% for more than 2
11) No chance for 1 component, but about 30% for 2, another 20% for 3.
12) Nothing

Summary
1) 6 out of 12 of these runs, 50%, had no discernible structure.
2) 5 out of 12, 42%%, had above 10% confidence for a single component
3) Only 2 of 12, 17%, supported a single component at more than 20% confidence
4) 4 of 12, 33%, supported two components at more than 20% confidence

With the High-Metallicity Stars:
I then repeated the above with the extra 5 stars with the following results:
1) Nothing
2) Nothing
3) 15% for 1 component, 18% for 2
4) 5% for 1, but 27% for 2 and 20% for 3
5) Nothing
6) Nothing
7) 10% for 1, and a whopping 30% for 2
8) Nothing of interest, but slightly more probable than those other nothings--5% for 1
9) 33% for 1, 26% for 2
10) Holy crap, 50% for 1, 24% for 2
11) Absolutely nothing for 1, 28% for 2
12) Nothing

Summary
1) Again, 50% had no discernable structure
2) Only 33% detected a single component at more than 10% confidence
3) Only 17% at more than 20%
4) 5 out of 12, 42%, had a double component composition at above 20% confidence

Bootstrapping Hercules
I then devised a code that would automate the Nmixing, reading the files, and then writing the first line of the .k file (posterior probability of each k) to a new file.

I used this to bootstrap and Nmix Hercules 20 times, with rather surprising results.
Remember that Hercules only has 30 stars, but an Nmix on the actual data gives a 49% probability of a single component.

What was surprising was that only 20% (4 of 20) of the bootstrapped samples had higher than a 10% probability of a single component. 3 out of 4 of those positive results were above 35%, but it was still rather discouraging. This means that 80% of the bootstrapped samples had practically no discernible overriding structure.

I'm not sure if this means that there just aren't enough stars (30) for proper bootstrapping or that Nmix just sucks. I need to test it on Segue I, with its higher (70%) single component confidence level.


TO DO
1) Do the same stuff to Segue1, remembering to exclude the last stars that don't belong
2) Try messing around with Bootes II, making sure that the right stars are selected from the data (Member is 1)
3) KNOW MY BENS!