Spectroscopic requirements for photometric surveys

Transcription

1 Spectroscopic requirements for photometric surveys Carlos Cunha Stanford University ngcfht workshop, March 28, 23

2 Basics of photo- zs Probe strong spectral features (4 Å break) Flux in each filter depends on galaxy s type and redshim. Terminology: magnitude = A log(flux) color = magnitude - magnitude

3 Basics of photo- z s Photo- zs are omen not very good. Three steps before gevng to the cosmology: Get photo- zs; Es;mate photo- z errors and cull outliers; Calibrate error distribu;on, e.g. P(z t z p ).

4 Basics of photo- z s Photo- zs are omen not very good. Three steps before gevng to the cosmology: Get photo- zs; spectra recommended Es;mate photo- z errors and cull outliers; spectra recommended Calibrate error distribu;on, e.g. P(z s z p ). spectra required

5 Need spectra, so what? Good spectroscopic samples are hard to come by. Issues Selec;on in observables: typically have many more bright samples than faint samples. Selec;on in non- observables: sample selected for a different purpose with different bands (e.g. DEEP2 survey). Shot- noise: samples are small. Sample variance: surveys are pencil- beam. Spectroscopic failures: Can t get spectra for certain galaxies. Wrong spectroscopic redshims.

6 Need spectra, so what? Good spectroscopic samples are hard to come by. Solu;ons Selec;on in observables: e.g. Weights (Lima, Cunha et al 28) Selec;on in non- observables: Don t do it. (Cunha et al 29) Shot- noise: need many galaxies Sample variance: need lots of area. Spectroscopic failures: Can t get spectra for certain galaxies. Wrong spectroscopic redshios. Cunha et al. 22a Cunha et al. 22b

7 Sample variance LSS in one deg 2 sample ΔP(z t z p ) = P(z t z p ) phot - P(z t z p ) train Difference between true P(z t z p ) and that of calibra;on sample generates biases in cosmology.

8 Survey Calculator Number of patches x 3 /4 deg 2 /8 deg 2 /32 deg 2 Magellan VLT σ 95 ( bias ) =. Assuming fiducial σ(w)=.35, and perfect spectroscopic selechon. gals/patch Cunha, Huterer, Busha & Wechsler MNRAS (22)

9 Survey Calculator Number of patches x 3 /4 deg 2 /8 deg 2 /32 deg 2 Magellan VLT σ 95 ( bias ) =. Assuming fiducial σ(w)=.35, and perfect spectroscopic selechon. gals/patch ngcfht Assuming:.5 deg 2 FoV, 2 fibers Cunha, Huterer, Busha & Wechsler MNRAS (22)

10 Spectroscopic failures

11 N- body + photometry Spectroscopic simula;ons Spectroscopy: Simulated spectra: K- correct templates + resoluhon + noise Spectroscopic redshios: IRAF- rvsao on - D simulated spectra

12 Completeness issues Deep spectroscopic samples are very incomplete or too small.8 Case study: Simulahons of DES photometry (i<24) + VVDS- like spec- z s True SSR h exposures 8- m telescope Redshift True SSR: frachon of galaxies with correct redshims.

13 Completeness issues z true z true Figure 4. Leakage matrices (P (z spec z true ))forthetrainingsetsselectedbythecu panel). The spectroscopic redshifts were calculated using 6,2 secs exposures with t SSR: Spectroscopic Success Rate True SSR True SSR: frachon of galaxies with correct redshims. r-i N(z) Observed SSR: Frachon of galaxies with redshim confidence above some threshold (R>6) R: Strength of correlahon between observed spectra and best- fivng spectrum in a template library i-mag Observed SSR (R > 6.).8 N(z).5.6 r-i i-mag Figure 2. Top panel: True spectroscopic success rate (SSR T ), defined as fraction of correct redshifts as a function of true redshift. Rightpanel: Observed SSR (SSR O ), defined as fraction of galaxies with correlation R 6.. Both results assume 62 secs of integration time with the 3 Fig area and reds sam 5.2

14 Completeness issues z true z true Figure 4. Leakage matrices (P (z spec z true ))forthetrainingsetsselectedbythecu panel). The spectroscopic redshifts were calculated using 6,2 secs exposures with t SSR: Spectroscopic Success Rate True SSR True SSR: frachon of galaxies with correct redshims. r-i N(z) Observed SSR: Frachon of galaxies with redshim confidence above some threshold (R>6) R: Strength of correlahon between observed spectra and best- fivng spectrum in a template library. Cannot use spectroscopic sample for calibrahon of error distribuhon of photometric sample if selechon is different. r-i i-mag Observed SSR (R > 6.) i-mag Figure 2. Top panel: True spectroscopic success rate (SSR T ), defined as fraction of correct redshifts as a function of true redshift. Rightpanel: Observed SSR (SSR O ), defined as fraction of galaxies with correlation R 6.. Both results assume 62 secs of integration time with the N(z) Fig area and reds sam 5.2

15 Completeness issues z true z true Figure 4. Leakage matrices (P (z spec z true ))forthetrainingsetsselectedbythecu panel). The spectroscopic redshifts were calculated using 6,2 secs exposures with t SSR: Spectroscopic Success Rate True SSR True SSR: frachon of galaxies with correct redshims. r-i N(z) Observed SSR: Frachon of galaxies with redshim confidence above some threshold (R>6) R: Strength of correlahon between observed spectra and best- fivng spectrum in a template library. Not a problem (in the simulahons, at least). Machine learning seems to effechvely be able to reproduce spectroscopic confidence selechon in terms of the survey colors. r-i i-mag Observed SSR (R > 6.) i-mag Figure 2. Top panel: True spectroscopic success rate (SSR T ), defined as fraction of correct redshifts as a function of true redshift. Rightpanel: Observed SSR (SSR O ), defined as fraction of galaxies with correlation R 6.. Both results assume 62 secs of integration time with the N(z) Fig area and reds sam 5.2

16 Spectroscopic failures (wrong redshios) Issues: When spec- z s are wrong, they re really wrong. A small speck of wrong redshims is enough to mess up cosmological constraints. z spec R > Sample used in the plot has 98.6% correct redshims and conshtutes 6% of total sample (bias in w: O(%)) z true -4 Case study: Simulahons of DES photometry + VVDS- like spec- z s R: cross- correlahon parameter (measures redshim confidence) Cunha, Huterer, Lin, Busha, Wechsler 22b

17 Lessons learned Incompleteness: Does not introduce cosmological biases if selechon matching is performed. Stahshcal constraints suffer with reduchon of sample size. Wrong redshios: Cause severe biases. Need beoer than 99% correct redshims. If 99% accuracy not possible, need to calibrate spectroscopic error distribuhon P(z true z spec ) with deeper sample/beoer instrument. Moral of the story: Focus of spectroscopic surveys has to be on accuracy of derived redshios, and characterizahon of redshim confidence.

18 The next genera;on of spectroscopic simula;ons SPOKES Brian Nord (FNAL, Filipe Abdalla (UCL), Adam Amara (ETH), Michael Busha (Zurich), Oliver Coles (UCL), Carlos Cunha (Stanford), Tom Diehl (FNAL), Brenna Flaugher (FNAL), Jaime Forero- Romero (UAC), Laurenz Gamper (ETH), Lukas Gamper (ETH), Ben Hambrecht (ETH), Stephanie Jouvel (Barcelona), Donnacha Kirk (UCL), Rich Kron (FNAL), Andrina Nicola (ETH), Alexandre Refregier (ETH), Will Saunders (AAO), SanJago Serrano (PAU)

19 What about ngcfht?

20 What about ngcfht? Sample variance and shot noise For a.5 deg 2 FoV, shot- noise = sample variance with roughly galaxies per poinhng or less. An instrument with 2 fibers is, thus, well balanced, from a galaxy stahshcs perspechve. Depends on survey depth, redshim, galaxy (halo) bias

21 What about ngcfht? Spectroscopic failures LSST Gold sample I < 25.3 Extending spectrograph window into IR is crihcal. Fraction in confident detections λ max =.3 µm λ max =. µm 3 hrs integrahon Redshift Determined using emission- line simulahons. Assuming: PFS- like resoluhon and noise, no Ly- α emission, λ min =.38 μm.

22 Bonus slides

23 Spectroscopic selec;on True SSR s 486s fiducial original comb2 fiducial s 486s s 486s Redshift i-mag R (redshift confidence)

24 An example: - Template photo- zs. - Calibrahon using one field with deg 2. - Weak Lensing shear- shear tomography. - Difference between true P(z s z p ) and that of calibrahon sample generates biases in cosmology. LSS in one deg 2 sample w- bias for fixed ΔP(z s z p )=. ΔP(z s z p ) = P(z s z p ) phot - P(z s z p ) train w- bias for ΔP(z s z p ) of Patch 37

25 Q est : redshim confidence eshmated with neural net. SSR T : Percentage of correct redshims in training sample. Shear- Shear constraints on w 62 secs bias(w) Selection Gal. Frac. SSR T (%) σ(w) z true z spec Q est > Q est > Q est > z true : bias due to selechon matching with neural networks: is negligible z spec : bias due to selechon matching + wrong redshios: is substan;al Cunha, Huterer, Lin, Busha, Wechsler et al, 22

26 Q est : redshim confidence eshmated with neural net. SSR T : Percentage of correct redshims in training sample. Shear- Shear constraints on w 62 secs bias(w) Selection Gal. Frac. SSR T (%) σ(w) z true z spec Q est > Q est > Q est > z true : bias due to selechon matching with neural networks: is negligible z spec : bias due to selechon matching + wrong redshios: is substan;al Cunha, Huterer, Lin, Busha, Wechsler et al, 22

27 IFU on WFIRST for LSST photo- z calibration Get supernovae spectra while performing the imaging survey.! Focal Plane But SN only takes up one pixel in IFU! Use rest of IFU to get spectra!!! For i lim = x 4 galaxies - 3 deg 2! Excellent for cross- calibrahon. We can dream of direct calibrahon. Success Fraction Success Frachon R= λ max =2.2 µm N pix =2, n read = ks. ks 2. ks R= λ max =2.2 µm N pix =2, n read = ks. ks 2. ks Redshift RedshiM Imager IFU In collaborahon w/ S. Perlmuoer & J. Newman

28 Technical Notes: Photo- z Calibration Tests All predictions are based on the revised COSMOS Mock Catalog (Zoubian et al., in prep.), but based on comparisons to DEEP2 and 3D-HST data we made a few changes:! Used velocity dispersion 65 km/sec for all objects (catalog values were unrealistic). DEEP2 95% range is 4- km/sec! Used a fixed [OII] doublet ratio of :.3 (catalog values were unrealistic)! For all Hα-based lines (all but [OII]), catalog line fluxes were divided by 3 (the factor by which mock catalog overpredicts z~-2 Hα EW-mass relation from Fumagilli et al. 22); dropped Ly" (predictions are very rough at this point)! Consider only galaxies with star-forming spectral types! Early-types form a negligible fraction of the catalog at z>, and should readily yield redshifts from spectral breaks +.6μm bump for WFIRST! Investigate completeness for emission-line redshifts only! Spectral break information can improve z success: predictions are conservative! >33% improvement in FoM based on Hearin et al. 2; estimates are based on a redshift distribution with fewer objects at z>2.3 than COSMOS mock catalog!

29 Technical Notes: PFS/ngCFHT Simulations Use resolution, pixel scales, read noise, dark current, throughput, etc. from ; PFS-provided sky background adjusted by.8x to match observed DEEP2 sky flux rescaled according to telescope / instrument / fiber characteristics! Assume all lines from COSMOS mock catalog are well-resolved (optionally including [OII] doublet); signal-to-noise is calculated for an optimal noise-weighted template-based feature detection. The S/N estimates may be overoptimistic (Ellis et al. claim S/N~8 for a 5x -7 erg/cm 2 /sec flux [OII] doublet in 5 minutes exposure; with greater collecting area & throughput, DEEP2 gets S/N~5 in hour at same flux).! Assume six hours of integration time per night after overheads (e.g. weather).! Use average of best-case (zenith) and worst-case (6 from zenith) atmospheric transmission.! ~5 widely-separated PFS pointings (~3k useful spectra) are required to match 2k IFU spectra, based on sample/cosmic variance calculations using QUICKCV code of Newman & Davis (22). Note: 5x2 nights = 3 nights ~ year! Require detection of 2 emission lines at 5#, OR 3 at 3# significance, OR at 5# plus 3 at 2.5#, for successful redshift determination!

30 Technical Notes: WFIRST Simulations R=5,.5 spectral pixels/resolution element,.5-2.4μm wavelength coverage! 5% throughput,.45 square sky pixels, object flux distributed over 2 sky pixels! 3 e - RMS read noise,.3 e - /s dark current, 2 readouts. Pixel scale and CCD characteristics not optimized for 2.5m.! Assume cannot resolve [OII] or [SII] doublets; add Paα/β/γ with Case B flux ratios vs. Hα to COSMOS mock catalog line set! Assume can always place an i<25.3 galaxy on IFU during imaging operations! For LSST gold sample, 4 galaxies/arcmin 2 = per 9 arcsec 2 ; for a 6 x6 IFU, need to be able to move pointing center 2 to average target per pointing! Assume 5 deg 2 survey,.25deg 2 FOV; 2k imaging pointings, 2s exposure time! Expect 5-hour and 2-hour SN spectra! Assume can pick up another galaxy as well as host >5% of time! Little effort made to optimize IFU parameters for photo-z calibration! Require significant detections of 2, 3, or 4 lines with net false redshift rate, combining all criteria, of.%!

31 Flux overprediction in COSMOS mock catalogs Revised COSMOS mock catalogs are being used for many spectroscopic survey predictions, but current version (Zoubian et al.) appears to have significant issues.! Catalog-derived restframe H" EW is 3x higher for star-forming galaxies than observed by 3D- HST (Fumagilli et al. 22). Perhaps due to calibration to match Geach et al. 2 (which overpredicts flux compared to most z~2 samples; see their Fig. )???! Fumagilli et al. 22 Zoubian et al. (in prep.) catalog (Figure from J. Newman)

32 Technical Notes: Success vs. I AB : predicted vs DEEP2 rescaled Can compare predicted 2- night PFS redshift success to extrapolation from DEEP2, accounting for differences in collecting area and throughput! Equivalent I AB for 2 nights on PFS Predicted PFS redshift success rate is somewhat higher than might expect extrapolating from DEEP2, but in the right ballpark! z>.4 galaxies (very difficult to get with DEIMOS or VIMOS) may account for difference!