Study efficiency - purity tradeoffs

[dkirkby]

Define a boolean truth selector:

T(v1) =  { c | zfit - ztrue | / (1 + ztrue) < v1 }

and a boolean data selector, for example:

D(v2) = { c (z68hi - z68lo) / 2 / (1 + zbest) < v2 }

Then we can count the numbers of fits in four exclusive categories, which sum to N(total):

• Positively selected in data: N(D)
  • True positives: N(D & T)
  • False positives: N(D & ~T)
• Negatively selected in data: N(~D)
  • True negatives: N(~D & ~T)
  • False negatives: N(~D & T)

Focus on two metrics:

• efficiency = N(D & T) / N(total)
• purity = N(D & T) / N(D)

This issue is to:

• Write code to calculate curves of efficiency and purity versus v2 given a definition of D(v2) and a fixed v1.
• Find the optimum definition of D(v2) and choice of v2 given a minimum allowed efficiency and purity for some v1.

The optimum might be different for different target categories and can be used to set ZWARN = ~D(v2).

[fjaviersanchez]

The results for the different object types show a small improvement but we lose some efficiency, and at this point it might not be worth to make the cuts until we have more realistic simulations. The biggest improvement has been to include z_avg (average of the posterior) instead of z_best (maximum of posterior) as our redshift estimate.

ELGs

LRG

QSO

STAR