Spectroscopy for Training Photometric Redshifts. Jeff Newman, U. Pittsburgh/PITT PACC

Transcription

1 Spectroscopy for Training Photometric Redshifts Jeff Newman, U. Pittsburgh/PITT PACC

2 Spectroscopy provides ideal redshid measurements but infeasible for large samples Imaging-based projects require redshid informajon to constrain dark energy At LSST gold sample (i<25.3) depths, ~10 hours on a 30m telescope to determine a redshid (75% of Jme) spectroscopically With a 500-fiber spectrograph on a 30m telescope, >10,000 telescope-years to measure redshids for this sample (4 billion galaxies)! AlternaJve: use broad spectral features to determine z : a photometric redshid Advantage: high muljplexing Disadvantages: lower precision, calibrajon uncertainjes Credit: ESO

3 Two spectroscopic needs for photo-z work: training and calibrajon Becer training / opjmizajon of algorithms using sets of objects with spectroscopic redshid measurements shrinks photo-z errors for individual objects Zhan 2006 Benitez et al Training datasets will contribute to calibrajon of photo-z's. ~Perfect training sets can solve calibrajon needs.

4 Two spectroscopic needs for photo-z work: training and calibrajon For weak lensing and supernovae, individualobject photo-z's do not need high precision, but the calibrajon must be accurate - i.e., bias and errors need to be extremely wellunderstood or dark energy constraints will be off Newman et al uncertainty in bias, σ(δ z )= σ(<z p z s >), and in scacer, σ(σ z )= σ(rms(z p z s )), must both be <~0.002(1+z) in each bin for Stage IV surveys. CalibraJon may be done via cross-correlajon methods using DESI/4MOST redshids (Newman 2008)

5 Improved photometric redshid training would greatly increase the science gains from future dark energy experiments E.g.: all LSST probes of dark energy will rely on measuring observables as a funcjon of photometric redshid Smaller photo-z errors from becer-trained algorithms can improve dark energy constraints, especially for BAO and clusters Zhan 2006 LSST system-limited photo-z accuracy is σ z ~ (1+z) (vs. σ z ~0.05(1+z) in similar samples today): difference is knowledge of templates/intrinsic galaxy spectra Perfect training set would increase LSST DETF FoM by at least 40%

6 Requirements for training for future ( Stage IV ) imaging dark energy experiments Need highly-secure spectroscopic redshids for 20k-30k galaxies sampling full range of galaxy colors, magnitudes, and redshids Newman et al. 2015, Spectroscopic Needs for Imaging Dark Energy Experiments, presents a baseline scenario: >30,000 galaxies down to LSST weak lensing limijng magnitude (i~25.3) 15 widely-separated fields at least 20 arcmin diameter to allow sample/cosmic variance to be mijgated & quanjfied Equal cosmic variance to Euclid C3R2 plan but ~20% as much sky area Long exposure Jmes are needed to ensure >75% redshid success rates: >100 hours at Keck to achieve DEEP2-like S/N at i=25.3 See Newman et al. 2015

7 Instrument requirements to address both DOE Cosmic Visions & Kavli/ NOAO/LSST report recommendations: based largely on photo-z needs High muljplexing - Required to get large numbers of spectra Coverage of full ground-based spectral window - Minimum: micron, microns preferred Significant resolujon (R=λ/Δλ>~5000) at red end - Allows secure redshids from [OII] 3727 Å line at z>1 amongst sky lines Field diameters > ~20 arcmin for efficient coverage - >1 degree preferred for much of Kavli science Southern site preferred to cover full LSST footprint

8 Fiber-fed spectrographs that maximize FoV would be best-suited to this work For photo-z training/calibrajon via ELTs, field of view and muljplexing are key: doubling the number of slits/fibers is roughly as useful as doubling the collecjng area ~40 LSST "gold sample" lensing objects per square arcmin: to get ~2500x muljplexing (so one poinjng per field), need at least 60 arcmin 2 FOV, but more is MUCH becer as can opjmize targets E.g.: MANIFEST fiber feed needed to make GMT/GMACS compejjve for this; fiber-fed versions of TMT/WFOS much more efficient than slits AO)

9 Summary of (some!) potential instruments for photo-z training Instrument / Telescope Collecting Area (sq. m) Field area (sq. deg.) Multiplex 4MOST ,400 Mayall 4m / DESI ,000 WHT / WEAVE ,000 Magellan LASSI ,000 Subaru / PFS ,400 VLT / MOONS Keck / DEIMOS FOBOS ESO SpecTel ,333 MSE ,249 GMT/MANIFEST + GMACS v. A GMT/MANIFEST + GMACS v. B TMT / WFOS Fiber WFOS-pessimistic ,000 Fiber WFOS-optimistic ,000 E-ELT / Mosaic Optical E-ELT / MOSAIC NIR Updated from Newman et al. 2015, Spectroscopic Needs for Imaging Dark Energy Experiments

10 Dark Jme (with 1/3 losses for weather + overheads) required for each instrument Instrument / Telescope Total time (years), >75% complete LSST sample Total time (years), >90% complete LSST sample 4MOST Mayall 4m / DESI WHT / WEAVE Magellan LASSI Subaru/PFS VLT/MOONS Keck/Deimos Keck/FOBOS ESO SpecTel MSE GMT/MANIFEST + GMACS v. A GMT/MANIFEST + GMACS v. B TMT / WFOS Fiber WFOS-pessimistic Fiber WFOS-optimistic E-ELT / MOSAIC Optical E-ELT / MOSAIC NIR Updated from Newman et al. 2015, Spectroscopic Needs for Imaging Dark Energy Experiments

11 Amount of time required for each survey from the Kavli/NOAO/LSST report (sorted by telescope aperture; in dark-years). Leader for each column shown in red. Total time, Milky Way Local Total Photometric halo survey dwarfs (8000 sq. Redshift (8000 sq. and halo Galaxy Supernova deg. halo Instrument / Telescope Training (y) deg., y) streams evolution hosts survey, y) Total (20k sq. deg. halo survey, y) 4MOST Mayall 4m / DESI WHT / WEAVE Magellan LASSI Subaru / PFS VLT / MOONS Keck / DEIMOS Keck / FOBOS ESO SpecTel MSE GMT/MANIFEST + GMACS GMT/MANIFEST + GMACS TMT / WFOS Fiber WFOS-pessimistic Fiber WFOS-optimistic E-ELT / MOSAIC Optical E-ELT / MOSAIC NIR

12 Total time required for all surveys from the Kavli/NOAO/LSST report (sorted by telescope aperture; in dark-years). Leader for each column shown in red. Instrument / Telescope Total (no halo survey, dark-years) Total (8000 sq. deg. halo survey, dark-years) Total (20k sq. deg. halo survey, dark-years) 4MOST Mayall 4m / DESI WHT / WEAVE Magellan LASSI Subaru / PFS VLT / MOONS Keck / DEIMOS Keck / FOBOS ESO SpecTel MSE GMT/MANIFEST + GMACS GMT/MANIFEST + GMACS TMT / WFOS Fiber WFOS - pessimistic Fiber WFOS - optimistic E-ELT / MOSAIC optical E-ELT / MOSAIC NIR

13 Conclusions ELTs can make major contribujons to cosmology through photo-z training/calibrajon surveys Minimum photo-z training survey, >~75% complete: 15 widely-separated poinjngs, DEEP2 S/N at i=25.3 = dark nights on an ELT, 30,000 spectra total Sample objects over full range of galaxy SEDs, 0 < z < 4 This would be a VERY interesjng galaxy-evolujon survey Could trade out fibers as secure redshids are achieved to get more bright objects (or have becer S/N for galaxy evolujon studies). Fiber feeds for GMACS or WFOS would make them much more efficient than E-ELT/MOSAIC for this work See the Newman et al. Spectroscopic Needs white paper for more: hcp://adsabs.harvard.edu/abs/2015aph n

14 A few notes Basic assumpjon for exposure Jme calculajons: with comparable resolujon and greater wavelength coverage than DEEP2, redshid success at the same conjnuum S/N should be no lower than DEEP2's (parjally because [OII] EW distribujon shows licle evolujon, so [OII] S/N conjnuum S/N) Difficult to make guarantees about how high success rates will be: many failures are "unknown unknowns", especially for IR-selected samples (as WFIRST). DEEP2- based predicjons are for success vs. i magnitude (and hence opjcal S/N). Equivalent IAB from 100 hours@keck Newman et al Based on theory papers, need >99% completeness over the full color/magnitude range used for analyses to keep biases subdominant for Stage IV (it takes a large number of independent spectra to demonstrate that is reached!). If that is achieved, training samples provide calibrajon too.

15 PotenJal photo-z performance for LSST ugrizy Green: Requirements on actual performance; grey: requirements on performance with perfect template knowledge (as in these sims)

16 Spectroscopic training set requirements Goal: make δ z and σ(σ z ) so small that systemajcs are subdominant Many esjmates of training set requirements (Ma et al. 2006, Bernstein & Huterer 2009, Hearin et al. 2010, LSST Science Book, etc.) General consensus that roughly 20k-30k extremely faint galaxy spectra are required to characterize: Typical z spec -z phot error distribujon Accurate catastrophic failure rates for all objects with z phot < 2.5 Characterize all outlier islands in z spec -z phot plane via targeted campaign (core errors easier to determine) Those numbers of redshids are achievable with ELTs, if muljplexing is high enough

17 What qualijes do we desire in our training sets? SensiJve spectroscopy of faint objects (to i=25.3) - Need a combinajon of large aperture and long exposure Jmes; >20 Keck-nights (=4 GMT-nights) equivalent per target, minimum High muljplexing - Obtaining large numbers of spectra is infeasible without it

18 What qualijes do we desire in our training sets? Coverage of full groundbased window - Ideally, from below 4000 Å to ~1.5μm - Require muljple features for secure redshid Comparat et al. 2013, submitted

19 What qualijes do we desire in our training sets? Significant resolujon (R>~4000) at red end - Allows redshids from [OII] 3727 Å doublet alone, key at z>1 Percentage of [OII] doublets resolved Comparat et al. 2013

20 What qualijes do we desire in our training sets? Field diameters > ~20 arcmin - Need to span several correlajon lengths for accurate clustering measurements (key for galaxy evolujon science and crosscorrelajon techniques) - r 0 ~ 5 h -1 Mpc comoving corresponds to ~7.5 arcmin at z=1, 13 arcmin at z=0.5 Many fields - Minimizes impact of sample/ cosmic variance. - e.g., Cunha et al. (2012) esjmate that ~0.1 deg 2 fields are needed for DES for sample variance not to impact errors (unless we get clever) Cunha et al. 2012

21 Biggest obstacle: incompleteness in training sets In current deep redshid surveys (to i~22.5/r~24), 30-60% of targets fail to yield secure (>95% confidence) redshids Losses are worst at the faint end RedshiD success rate varies with galaxy color, redshid, etc. In DEEP2, best parts of BRI color space have ~90% redshid success 4 night GMT depth would yield >~75% completeness; achieving >90% would require ~25 nights/ poinjng on GMT Data from DEEP2 (Newman et al. 2013) and zcosmos (Lilly et al. 2009)

22 Note: even for 100% complete samples, current false-z rates would be a problem Only the highestconfidence redshids should be useful for precision calibrajon: lowers spectroscopic completeness further when restrict to only the best A major reason why spli ng [OII] is important Approx LSST Req't Based on simulated redshift distributions for ANNz-defined DES bins in mock catalog from Huan Lin, UCL & U Chicago, provided by Jim Annis

23 How much time would be required to complete surveys from the Najita et al. Kavli/ NOAO/LSST report on different platforms? This is an attempt to take the largest surveys proposed in the Kavli report and work out how long would be needed to do them Common set of assumptions: one-third loss to instrumental effects, weather and overheads; 4m = Mayall/DESI; 8m = Subaru/PFS; all instrumental efficiencies identical; equivalent # of photons will yield equal noise; ignoring differences in seeing/ image quality and fiber/slitlet size. Only mediumresolution fibers included. Assuming full spectral range can be covered simultaneously (likely not true for E-ELT). See report (available at ) for details of these surveys Will give time in years on each platform; note that this is generally dark time (very faint targets!) Costs based on TSIP + inflation: $1k/m 2 /night

24 Brief descriptions of the Kavli/NOAO/LSST surveys Photometric redshift training sample: Minimum of 30,000 galaxies total down to i=25.3 in 15 fields >20' diameter 100 hours/pointing on 10m To improve photo-z accuracy for LSST (and study galaxy SED evolution) Highly-complete survey would require ~6x greater exposure time than used here Milky Way halo survey: ~125 g<23 luminous red giants deg -2 over 8,000 (or preferably 20,000) square degrees of sky 2.5 hours/pointing with 8m Allows reconstruction of MW accretion history using stars to the outer limits of the stellar halo. Other objects could be targeted on remaining fibers.

25 Brief descriptions of the Kavli/NOAO/LSST surveys Local dwarfs and halo streams: Local dwarfs were estimated to require 3200 hours on an 8m to measure velocity dispersions of LSST-discovered dwarfs within 300 kpc Requires FoV 20 arcmin (1 deg preferred) and minimum slit/fiber spacing < 10 arcsec. Characterizing ~10 halo streams to test for gravitational perturbations by low-mass dark matter halos was estimated to require ~25% as much time on similar instrumentation.

26 Brief descriptions of the Kavli/NOAO/LSST surveys Galaxy evolution survey: Minimum of 130,000 galaxies total down to M=10 10 M Sun at 0.5 < z < 2 over a 4 sq. deg. field 18 hours per pointing on 8m To study relationship between galaxy properties and environment across cosmic time Supernova host survey: Annual spectroscopy of ~100 new galaxy hosts of supernovae deg -2 with r<24 over the ~5 LSST deep drilling fields (10 sq. deg. each) ~8 hours per pointing on 4m Provides redshifts for most of the ~50,000 best-characterized LSST SN Ia (other transients/hosts could be observed on remaining fibers)