search

aimalz / qp

Quantile Parametrization for probability distribution functions module
MIT License
3 stars 9 forks source link

Better quantile spacing #35

Open drphilmarshall opened 7 years ago

drphilmarshall commented 7 years ago

I guess we want to choose the quantiles such that we sample the peaks and wings well: uniform in CDF does not seem to be optimal (or even superior to a histogram). In fact; don't we want to preferentially place interpolation nodes where the second derivative of the function is high? How about the first derivative? There must be a whole literature on this.

aimalz commented 7 years ago

Just commenting to document that we want to develop a better way to choose quantile spacing. Let's leave this issue open for this purpose.

drphilmarshall commented 7 years ago

Good good. Feel free to punt it to the post-launch DESC Note milestone.

On Fri, Dec 9, 2016 at 11:54 AM, Alex Malz notifications@github.com wrote:

Just commenting to document that we want to develop a better way to choose quantile spacing. Let's leave this issue open for this purpose.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/aimalz/qp/issues/35#issuecomment-266107183, or mute the thread https://github.com/notifications/unsubscribe-auth/AArY98IEgS0f8G4elR_4OQOcVaB9aPWQks5rGbHpgaJpZM4LHJvn .

drphilmarshall commented 7 years ago

OK, I made a milestone for this analysis (and other enhancements): Mock Photo-z Test

aimalz commented 7 years ago

It seems like optimization over each chosen metric (presumably in survey mode once #43 is done) is probably the way to approach this one.