Hunting down systematics in modern galaxy surveys · 2017. 1. 24. · Mohammadjavad Vakili/...

Hunting down systematics in moderngalaxy surveys

Mohammadjavad VakiliCenter for Cosmology and Particle Physics

New York University

Berkeley / Cosmology Seminar2017 January

Mohammadjavad Vakili/ 2017-01-24

Outline

I Large-scale structure mocks for estimation ofgalaxy clustering covariance matrices

I Weak lensing systematicsI Point Spread FunctionI Photometric redshifts


Accurate galaxy mocks for estimation ofgalaxy clustering covariance matrices

I Based on works in collaboration with:Francisco-Shu Kitaura (IAC), Yu Feng(Berkeley), Gustavo Yepes (UAM), ChengZhao (Tsinzua), Chia-Hsun Chuang (Leibniz),ChangHoon Hahn (NYU)


Future of spectroscopic galaxy surveys

I Measurement of growth rate and expansionhistory with sub-percent precision

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

z

0.2

0.3

0.4

0.5

0.6

0.7

fσ

8

6DFGS

SDSS MGSSDSS LRG

BOSS LOWZBOSS CMASS

VIPERS

WiggleZ

0 0.5 1 1.5Redshift z

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Gro

wth

rate

= d

lnD

/ dl

na

DGP

f(R) k=0.02CDM

f(R) k=0.1

Right: Planck Collaboration XIII (2015), Left:DESI Collaboration (2016)


Future of spectroscopic galaxy surveys

I Measurement of growth rate and expansionhistory with sub-percent precision

DESI Collaboration (2016)


We need mocks for both precision andaccuracy!

I Estimation of uncertainties (covariancematrix)

I Need a large number of mocks(Nmock >> Ndata)

I Mocks need to be statistically consistent(1-point, 2-point, 3-point, ...) with the data!

I We live in the era of systematic-limitedmeasurements

I Need accurate end-to-end simulations ofgalaxy surveys to characterize sytematicuncertainties


How do we efficiently generate mocks forgalaxy surveys?

I Requirements:

I Need to simulate large volumes to sample theBAO signal

I Need to accurately model nonlinear clustering(current kmax ∼ 0.25 hMpc−1)

I Need to resolve low mass halos that host faintgalaxies

I Need to accurately describe two-point andhigher-order statistics

I N -body simulations are expensive!


How do we efficiently generate mocks forgalaxy surveys?

I Approximate Methods:I Approximate (DM-only) structure formation

model + Empirical sampling of galaxies/halosfrom the dark matter field


State-of-the-Art: SDSS III-BOSS

I QPM (White et al. 2014)

I Low resolution N -bodyI Sample halos by matching the mass function

and large scale bias

I ALPT-PATCHY (Kitaura et al. 2016)

I perturbation theoryI Sample halos by matching the n−point

functions


State-of-the-Art Approximate Methods:SDSS III-BOSS

Two-point statistics:

ξ0(s), ξ2(s)

-150

-100

-50

0

50

100

150

0 20 40 60 80 100 120

s2ξ(s)

[h−2M

pc2]

s[h−1Mpc]

0.55 < z < 0.75

QPMMD Patchy

BOSS-CMASS DR12


Percentage-level accuracy galaxy mocks

I Precision large-scale structure cosmologyrequires mocks with percentage-level accuracy!

I Main challenges:I Nonlinear ScalesI RSDI Higher order Statistics


Goal: Percent-level accuracy

I Main Challenges: (Quasi)Nonlinear Scales,RSD (Chuang et al. 2015):

100

200

300

400

500

600

k1.

5P

0(k

)

BigMD.SO

COLA+HOD

EZmock

HALOgen+HOD

LogNormal

PATCHY

PINOCCHIO+HOD

0.1 0.2 0.3 0.4 0.5

k [h Mpc−1 ]

0.80

0.85

0.90

0.95

1.00

1.05

1.10

1.15

rati

o

10

100

20

30

40

50

60708090

200

300

k1.

5P

2(k

)

BigMD.SO

COLA+HOD

EZmock

HALOgen+HOD

PATCHY

PINOCCHIO+HOD

0.1 0.2 0.3 0.4 0.5

k [h Mpc−1 ]

0.80

0.85

0.90

0.95

1.00

1.05

1.10

1.15

rati

o


Goal: Percentage-level accuracyI Main Challenges: high-order statistics!

I BAO detection (Slepian et al. 2015)I Breaking the degeneracy betweenf , σ8 (Gill-Maŕın et al. 2014)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

B(θ

)

1e8

BigMD.SO

COLA+HOD

EZmock

HALOgen+HOD

LogNormal

PATCHY

PINOCCHIO+HOD

0.0 0.2 0.4 0.6 0.8 1.0θ12/π

0.8

0.9

1.0

1.1

1.2

rati

o


PATCHY : Nonlinear Stochastic Biasing

For a given dark matter density field ρm, halos/galaxies aregenerated from a nonlinear stochastic bias model: (1)Empirical nonlinear bias

〈ρg〉(ρm) = fg θ(ρm − ρth

)︸︷︷︸threshold bias

× ραm︸︷︷︸nonlinear bias

× exp(− (ρ/ρ�)�

)︸︷︷︸exponential cutoff

(2) stochastic bias (deviation from Poissoinity):

ρg ∼ NB(〈ρg〉;β

)


How can we improve Patchy?

I Limitations of PATCHY:

I Brute-force estimation of bias parametersI Limited accuracy of ALPT as a gravity solver

ALPT = LPT (on large Scales) + SC (onsmall scales)

I Solution:I Automatic estimation of bias parameters with

MCMCI Replacing the gravity solver with an

approximate N -body solver that yields abetter 1-halo term clustering


New gravity solver: FastPM

I FastPM (Feng et al. 2016) : approximateparticle mesh N -body solver

I Enforces large-scale linear growth

I Scales well with resolution, time step, forceresolution, ...


Strategy for generation of mocks

I Generation of a DM field with low resolutionN -body

I Constraining the patchy bias parameters byfitting P (k)

I Generation of galaxy/halo mocks

Method is currently being tested as part of theEuclid covariance project.


Comparison with the BigMultiDarksimulation

Can we reproduce the population of halos (andsubhalos) in the BigMultiDark N -body Simulation(N3p = 3840

3) with a low-resolutionFastPM-PATCHY (N3p = 960

3)?


Dark matter density field

1250 h−1Mpc

From left to right: BigMD, FastPM, ALPT.


Dark matter density ield

625 h−1Mpc



Dark matter density field

312.5 h−1Mpc



Bias parametersδth = 1. 07+0. 02−0. 05

0.2

0.4

0.6

0.8

1.0

α

α = 0. 27+0. 03−0. 04

0.2

0.4

0.6

0.8

1.0

β

β = 0. 73+0. 09−0. 07

0.2

0.4

0.6

0.8

1.0

ρ²

ρ² = 0. 15+0. 07−0. 08

0.8

0.0

0.8

1.6

δth

0.8

0.6

0.4

0.2

0.0

²

0.2

0.4

0.6

0.8

1.0

α

0.2

0.4

0.6

0.8

1.0

β

0.2

0.4

0.6

0.8

1.0

ρ²

0.8

0.6

0.4

0.2

0.0

²

² = −0. 24+0. 18−0. 24


Comparison with the BigMultiDarkSimulation

One-point PDF

100 101

100

101

102

103

104

105

106

107

Nce

lls

BigMD BDM halos

FastPM-patchy halos

ALPT-patchy halos

100 101

Nhalos

0.5

1.0

1.5

Nm

ock

cells/N

Big

MD

cells

Vakili et al. (2017)


Comparison with BigMultiDark Simulation

Real Space P(k)

k [Mpc−1h]

103

104

105

P(k

)[M

pc3h−

3]

BigMD BDM halos

FastPM-patchy halos

ALPT-patchy halos

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

k [Mpc−1h]

0.90

0.95

1.00

1.05

1.10

Pm

ock(k

)/P

ref(k)



Bispectrum Comparison

0.5

1.0

1.5

2.0

B(α

12)

[Mpc6h−

6]×

109

k1 =k2 =0.1 Mpc−1h

BigMD BDM halos

FastPM-patchy halos

ALPT-patchy halos

0.2 0.4 0.6 0.8 1.0

α12/π

0.8

1.0

1.2

Bm

ock(α

12)/B

Big

MD(α

12) 0.2

0.4

0.6

0.8

1.0

1.2

B(α

12)

[Mpc6h−

6]×

108

2k1 =k2 =0.3 Mpc−1h

0.2 0.4 0.6 0.8 1.0

α12/π

0.8

1.0

1.2

Bm

ock(α

12)/B

Big

MD(α

12)



Anisotropic RSD (Preliminary)Work in progress!

0.0 0.1 0.2 0.3 0.4 0.5

k hMpc−1

0.85

0.90

0.95

1.00

1.05

1.10

1.15

Pm

ock(k,µ

)/P

ref(k,µ

)

µ=0.1

µ=0.3

µ=0.5

µ=0.7

µ=0.9


Summary

I We have presented a new version of thePATCHY code with MCMC estimation of biasparameters and FastPM gravity solver.

I By testing our method with the halos in theBigMultiDark simulation, we recover P (k) at∼ 2% level to high k modes (k ∼ 0.4 hMpc−1),and the bispectrum at a ∼ 15− 20% level!

I Redshift space clustering results are not idealyet! But a different approach for treatment ofRSD is currently being developed.


Tackling PSF and photometric redshiftsystematics in imaging surveys

I Based on works in collaboration with:David Hogg (NYU, CCA), Alex Malz (NYU)


LSST and the next generation of imagingsurveys

0.26 0.30 0.34 0.38N = 280001 Bandwidth = 0.001184

Den

sity

0.80 0.85 0.90N = 280001 Bandwidth = 0.002538

Den

sity

0.60 0.64 0.68 0.72N = 280001 Bandwidth = 0.001319

Den

sity

−1.5 −1.0 −0.5N = 280001 Bandwidth = 0.0208

Den

sity

−1.5 −0.5 0.5 1.5N = 280001 Bandwidth = 0.1012

Den

sity

0.80

0.85

0.90

1

1

1

1

1

1

1

1

0.60

0.64

0.68

0.72

1

1

1

1

1

1

−1.

5−

1.0

−0.

5

1

1

1

1

0.26 0.30 0.34 0.38

−1.

5−

0.5

0.5

1.5

0.80 0.85 0.900.60 0.64 0.68 0.72 −1.5 −1.0 −0.5 −3 −2 −1 0 1 2 3N = 280001 Bandwidth = 0.1012

Den

sity

clusteringcosmic shearclusterN3x2pt3x2pt+clusterN+clusterWL

Ωm σ8 h w0 wa

wa

w0

hσ 8

Pro

b

Kraus & Eifler 2016


LSST and the next generation of imagingsurveys

Jain et al. 2015Mohammadjavad Vakili/ 2017-01-24

Weak lensing measurements

I Weak lensing measurements are the basis ofmany powerful probes:

I Cosmic ShearI Galaxy Cluster CosmologyI Cross-correlation with CMB and galaxies

Hildebrandt et al. 2016Mohammadjavad Vakili/ 2017-01-24

Weak lensing measurements

I Weak lensing measurements are the basis ofmany powerful probes:

I Cosmic ShearI Galaxy Cluster CosmologyI Cross-correlation with CMB and galaxies

Mantz et al. 2014Mohammadjavad Vakili/ 2017-01-24

Weak lensing is limited by systematics

I The problem of inferring the cosmic shearsignals from observations is far from idealized.Cosmic shear signal is dominated by:

I the PSFI shape noiseI Intrinsic alignmentsI and many more: Blending, noise bias, ...


Impact of the PSF (CFHTLenS)

Heyman et al. 2011


Impact of the PSF (DES)

100 101 102

θ (arcmin)

10−7

10−6

10−5

ξ +∆e(θ)

Jarvis et al. 2016


A closer look at the atmospheric PSFVariation of LSST atmospheric PSF ellipticitiesacross the FoV Simulations run by LSST PhotonSimulator (Peterson 2011)


A closer look at the atmospheric PSFIn practice, we can only empirically estimate thePSF at the positions of stars and predict its valueelsewhere


LSST Atmospheric turbulenceHow can we optimally interpolate the PSF?

Vakili et al. in preparation: Gaussian Processinterpolation method beats a more traditionalpolynomial interpolation. Atmosphere still causesconfusion in sub-arcminute scales!


Weak lensing is limited by systematics : theimpact of Photo-z’s

I Accurate redshift probabilities are needed fortomographic two-point function calculations,determination of redshift distributions,inference of cluster masses.

1

0

1

2

3

4

5

n(z

)

0.1

Common photo-z estimation methods

I Template fitting

I Machine Learning

I Cross-correlation with spectroscopic sample

0.0 0.5 1.0 1.5 2.0z

0.5

1.0

1.5

2.0

2.5

3.0

P(z

)

Galaxy's true redshiftRandom forestPhysical model


Combining different datasets : WFIRST andLSST

I Accuracy and precision of P (z) for individualgalaxies can be enhanced by combining thedata from overlapping surveys:


LSST filters


WFIRST filters


LSST and WFIRST

P (z|F̂, {SEDk}) =∫ ∏

k

dtkP (z, tk|F̂, {SEDk})

F̂ = {F̂LSST, F̂WFIRST}Template library {SEDk} from Brown et al. (2014)used in LSST DC1.


P (z|F̂) with single exposure LSST andWFIRST?



WFIRST photo-z is limited by distinguishinggalaxy SED’s at WFIRST wavelengths


n(z) with single exposure LSST andWFIRST?

How well can we recover the redshift distributions?p(N|{dk}) ∝p(N ) exp[−

∫N (z)dz]×∏k ∫ p(zk|dk)p(zk) dzk

n(z) = dNdz



How well can we recover the redshift distributions?

0.3 0.6 0.9 1.2 1.5 1.8redshift

0.00

0.05

0.10

0.15

0.20

0.25

0.30

n(z

)

WFIRST

LSST

LSST + WFIRST

True n(z)



How well can we recover the redshift distributions?

0.3 0.6 0.9 1.2 1.5 1.8redshift

0.5

1.0

1.5

n(z

)/n

tru

e(z)

WFIRST

LSST

LSST + WFIRST


How do we optimally combine differentdatasets

I Treat different datasets independently

I Simultaneously constrain photometry andshapes with both datasets:

P (F̂, e|dpixel)

wheredpixel

is the pixel-level data from all band-passes


How do we optimally combine differentdatasets

I Real World scenario:


Joint vs Independent modeling of bandpasses

I Joint modeling of all band-passes at the pixellevel could mitigate the biases in fluxestimates and hence the redshifts


Summary

I The impact of PSF residual systamtics can becontrolled if we use a more flexible GaussianProcess model for PSF interpolation.

I We have presented results showing thataccuracy and precision of photometricredshift probabilities can be enhanced bycombining datasets.

I Joint modeling of all bandpasses at the pixellevel leads to more robust photometricredshift estimation.


Date post:	04-Feb-2021
Category:	Documents
Upload:	others
View:	4 times
Download:	0 times

Hunting down systematics in modern galaxy surveys · 2017. 1. 24. · Mohammadjavad Vakili/...

Documents