====== Challenge 3: Fit BP/RP and RVS simultaneously ====== 

===== Objectives =====

  * Compare performance of GSP-Phot to Andy's SICK code when fitting only BP/RP.
  * Investigate fitting BP/RP alone, RVS alone, and BP/RP+RVS simultaneously with Andy's SICK code.
  * Is this a route to beat APSIS?
  * Where can BP/RP help to estimate abundances from RVS?
  * Where can RVS help to improve on BP/RP?
  * This project employs synthetic MARCS spectra.

===== GSP-Phot vs. SICK =====

== Conclusions ==

  * GSP-Phot and SICK overall agree very nicely. No algorithm clearly outperforms the other.
  * logg: GSP-Phot is slightly better than SICK. SICK may suffer from some convergence problems (red dots below dwarfs).
  * [Fe/H]: SICK is slightly better than GSP-Phot. GSP-Phot appears to systematically underestimate [Fe/H] by ~1dex for some stars. These problematic stars are fainter than G~15, i.e., SICK estimates [Fe/H] more robustly at lower S/N than GSP-Phot.

== Test data ==

  * 2000 synthetic MARCS spectra in Gaia BP/RP. 
  * Mostly Main-Sequence dwarfs but also some giants.
  * Teff from 4000K to 8000K.
  * Test spectra created by GOG. (Not by our own forward models!)

== Examples of posterior samples ==

  * Posterior distributions mostly well behaved for BP/RP, i.e., fast convergence, unimodal.
  * Even in simple case the posteriors cannot be well approximated by a Gaussian.
  * Rare examples with multimodal posterior.

{{:tests:example-MCMC-unimodal.png|}}

{{:tests:example-MCMC-multimodal.png|}}

== Results for fitting BP/RP alone ==

^  ^ GSP-Phot bias ^ SICK bias ^ GSP-Phot sigma ^ SICK sigma ^
| Teff [K] | 14 | 18 | 100 | 100 |
| AV [mag] | 0.02 | 0.03 | 0.07 | 0.08|
| logg [dex] | 0.02 | 0.07 | 0.19 | 0.29 |
| [Fe/H] [dex] | 0.12 | 0.00 | 0.37 | 0.28 |


{{:tests:sick-gspphot-teff.png|}}
{{:tests:sick-gspphot-A.png|}}
{{:tests:sick-gspphot-logg.png|}}
{{:tests:sick-gspphot-feh.png|}}


== Future tests ==

  * Compare initial guesses of GSP-Phot and SICK. (Possibly SVM fails for noisy data, which may explain why SICK estimates [Fe/H] better at faint magnitudes.)


===== SICK: BP/RP alone vs. RVS alone =====

===== SICK: Fitting BP/RP and RVS simultaneously =====