type ia sn light curve inference: hierarchical …hea-2011/01/25 · type ia sn light curve...
TRANSCRIPT
Kaisey S. MandelCfA Supernova GroupAstrostatistics Seminar
25 January 2011
Type Ia SN Light Curve Inference:Hierarchical Models for Nearby SN
in the Rest-Frame Near Infrared and Optical
1Tuesday, January 25, 2011
Related Papers
Mandel, K., G. Narayan, R.P. Kirshner. Type Ia Supernova Light Curve Inference: Hierarchical Models in the Optical and Near Infrared. 2011 submitted to ApJ, arXiv:1011.5910
Mandel, K., W.M. Wood-Vasey, A.S. Friedman, R.P. Kirshner. Type Ia Supernova Light Curve Inference: Hierarchical Bayesian Analysis in the Near Infrared. 2009, ApJ, 704, 629-651
Blondin, S., K. Mandel, R.P. Kirshner.Do Spectra improve distance measurements of SN Ia?2011, A&A 526, A81
Wood-Vasey, et al. Type Ia Supernovae are Good Standard Candles in the Near Infrared: Evidence from PAIRITEL.2008, ApJ, 689, 377-390
2Tuesday, January 25, 2011
Outline• Type Ia SN and Cosmology
• Statistical Inference with SN Ia Light curves
• Hierarchical Framework for Structured Probability Models for Observed Data
• Describing Populations & Individuals, Multiple Random Effects, Covariance Structure
• Statistical Computation with Hierarchical Models
• BayeSN (MCMC)
• Application & Results: Hierarchical Model for Nearby CfA NIR (PAIRITEL) and Optical (CfA3) SN Ia light curves
3
3Tuesday, January 25, 2011
Standard Candle Principle
1. Know or Estimate Luminosity L of a Class of Astronomical Objects
2. Measure the apparent brightness or flux F
3. Derive the distance D to Object using Inverse Square Law: F = L / (4π D2)
4. Optical Astronomer’s units: m = M +μ
4Tuesday, January 25, 2011
Type Ia Supernovae areAlmost Standard Candles
• Progenitor: C/O White Dwarf Star accreting mass leads to instability
• Thermonuclear Explosion: Deflagration/Detonation
• Nickel to Cobalt to Iron Decay + radiative transfer powers the light curve
• SNe Ia progenitors have nearly same mass, therefore energy
Credit: FLASH Center
5
5Tuesday, January 25, 2011
Type Ia Supernova ApparentLight Curve
!10 0 10 20 30 40 50 60
6
8
10
12
14
16
18
20
22
B + 2SN2005el (CfA3+PTEL)
V
R ! 2
I ! 4
J ! 7
H ! 9
Obs. Days Since Bmax
Ob
s.
Ma
g. !
kc !
mw
x
6Tuesday, January 25, 2011
Supernova Cosmology:Constraining Cosmological Parameters
using Luminosity Distance
vs. Redshift
AAS 215
!"##$%&! !""#$%&#'()(#&*'+,*'#!""#$%&#'()(#&*'+,*'#! -*)&(././0#1234#5.)6#-*)&(././0#1234#5.)6#
"(&0*#"7589#:*)#)7#;*))*&#"(&0*#"7589#:*)#)7#;*))*&#:*<(&()*#,7"7&#(/'#'+:)#:*<(&()*#,7"7&#(/'#'+:)#(/'#&*'+,*#:%:)*=().,:(/'#&*'+,*#:%:)*=().,:
! >?(=././0#@#;(/'#>?(=././0#@#;(/'#(/7=("%(/7=("%
! A&7<(0()*#<&7;(;.".)%#A&7<(0()*#<&7;(;.".)%#:+&B(,*#B7&#*(,6#4C#)7#:+&B(,*#B7&#*(,6#4C#)7#,7:=7"70%#D#;*))*&#,7:=7"70%#D#;*))*&#*:).=()*:#7B#:%:)*=().,#*:).=()*:#7B#:%:)*=().,#
! !"#$%&%'(")##,7/:)&(./):#,7/:)&(./):#7/#,7/:)(/)#57/#,7/:)(/)#5
*+,+-./01+*+,+-./01+EE+./.2+345657++./.2+345657+EE+./89+34:47+./89+34:47
• Nearby Hubble law is linear• High-z depends on cosmology• Host Galaxy Dust is a Major Confounding Factor
Credit: Gautham Narayan
7Tuesday, January 25, 2011
Cosmological Energy Content
Dark Energy Equation of state P = wρIs w + 1 = 0? Cosmological Constant
8Tuesday, January 25, 2011
Reading the Wattage of a SN Ia:Empirical Correlations
• Width-Luminosity Relation: an observed correlation (Phillips)
• Observe optical SN Ia Light Curve Shape to estimate the peak luminosity of SN Ia: ~0.2 mag
• Color-Luminosity Relation
• Methods:
•
• MLCS, Abs LC vs Dust
• SALT, App. Color single factor
Intrinsically Brighter SN Ia have broader light curves
and are slow decliners
9
!m15(B)
9Tuesday, January 25, 2011
Statistical inference with SN Ia
• SN Ia cosmology inference based on empirical relations
• Statistical models for SN Ia are learned from the data
• Several Sources of Randomness & Uncertainty
1. Photometric errors
2. “Intrinsic Variation” = Population Distribution of SN Ia
3. Random Peculiar Velocities in Hubble Flow
4. Host Galaxy Dust: extinction and reddening.
• Apparent Distributions are convolutions of these effects
• How to incorporate this all into a coherent statistical model? (How to de-convolve?)
10
10Tuesday, January 25, 2011
Advantages of Hierarchical Models• Incorporate multiple sources of randomness & uncertainty
• Express structured probability models adapted to data
• Hierarchically Model (Physical) Populations and Individuals simultaneously: e.g. SN Ia and Dust
• Intrinsic Covariance: Color/Luminosity/Light Curve Shape
• Dust Reddening/Extinction
• Full (non-gaussian) probability distribution = Global, coherent quantification of uncertainties
• Completely Explore & Marginalize Posterior trade-offs/degeneracies/joint distributions
• Deals with incomplete/missing data problems
• Make best inference/estimate for the observed data
• Modularity
11Tuesday, January 25, 2011
Directed Acyclic Graph for SN Ia Inferencewith Hierarchical Modeling
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
D1
z1
zN
DN
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
• Intrinsic Randomness• Dust Extinction & Reddening• Peculiar Velocities • Measurement Error
“Training” - Learn about Populations
12
Generative Model
Global Joint Posterior Probability
Density Conditional on all
SN Data
12Tuesday, January 25, 2011
Representing SN Ia Light curves:Differential Decline rates
• Gaussian Process over Decline Rates at different Wavelengths / Phases and Peak Luminosities
• Goal: Infer the Intrinsic Covariance Structure of SN Ia light curves over multiple wavelengths and phases
• Use to make “best” distance predictions
!10 0 10 20 30 40 50 60
!28
!26
!24
!22
!20
!18
!16
!14 B + 2
V
R ! 2
I ! 4
J ! 7
H ! 9
Obs. Days Since Bmax
Absolu
te M
agnitude
13Tuesday, January 25, 2011
Positive Dust only Dims and Reddens
0
50
100
SN
#
0 0.5 1 1.5 20
1
2
3
AV ! Expon(" = 0.37 ± 0.04)
Dust Extinction AV
Rela
tive F
requency
!20
!19
!18
!17
!16
MB vs B!H
MV vs V!H
!2 !1 0 1 2
!20
!19
!18
!17
!16
MR
vs R!H
!2 !1 0 1 2
MH
vs J!H
Intrinsic
Apparent
14Tuesday, January 25, 2011
Directed Acyclic Graph for SN Ia Inference:Distance Prediction
15
zs
Ds
µs
AppLCs
s = 1, . . . , NSN
AsV , Rs
V
AbsLCs
Training
PredictionApV , Rp
Vµp
DpAppLCpAbsLCp
DustPop
SN IaAbsLC
Pop
15Tuesday, January 25, 2011
Statistical Computation with Hierarchical SN Ia Models: The BayeSN Algorithm
• Strategy: Generate a Markov Chain to sample global parameter space (populations & all individuals) => seek a global solution
• Chain explores/samples trade-offs/degeneracies in global parameter space
Multiple chains globally converge from random
initial values
0
0.5
1
1.5
2
2.5
3
AV
SN2001az
SN2001ba
100
101
102
103
0
0.5
1
1.5
2
2.5
3
MCMC Sample
AV
SN2006cp
BayeSN MCMC Convergence
100
101
102
103
MCMC Sample
SN2007bz
16Tuesday, January 25, 2011
BayeSN
• Metropolis-Hastings within Gibbs Sampling Structure to exploit conditional structure
• Requires (almost) no tuning of jump sizes
• Generalized Conditional Sampling to speed up exploring trade-off between dust and distance: (Av, μ) → (Av, μ) + γ(1, -x)
• Run several (4-8) parallel chains and compute Gelman-Rubin ratio to diagnose convergence
17Tuesday, January 25, 2011
BayeSN in Graphs
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
D1
z1
zN
DN
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
D1
z1
zN
DN
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AbsLC
#N
AbsLC
Pop
AbsLC
#1
AbsLC
#N
AbsLC
Pop
AbsLC
#1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
Av, Rv
#N
Dust
Pop
Av, Rv
#1
Av, Rv
#N
Dust
Pop
Av, Rv
#1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AbsLC
Pop
D1
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
AbsLC
Pop
D1
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AbsLC
Pop
z1
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
AbsLC
Pop
z1
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
Dust
Pop
AbsLC
Pop
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1Dust
Pop
AbsLC
Pop
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AppLC
#NAbsLC
#N
Av, Rv
#NAbsLC
Pop
µN
DNAppLC
#NAbsLC
#N
Av, Rv
#NAbsLC
Pop
µN
DN
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AppLC
#NAbsLC
#N
Av, Rv
#NAbsLC
Pop
µN zN
AppLC
#NAbsLC
#N
Av, Rv
#NAbsLC
Pop
µN zN
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
18
18Tuesday, January 25, 2011
BayeSN in Graphs
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
D1
z1
zN
DN
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
AppLC
#NAbsLC
#N
Av, Rv
#N
Dust
Pop
AbsLC
Pop
µN
D1
z1
zN
DN
AbsLC
#1
AppLC
#1
Av, Rv
#1
µ1
18
18Tuesday, January 25, 2011
Practical Application of Hierarchical Model: NIR SN IaWhy are SN Ia in NIR interesting?
• Host Galaxy Dust presents a major systematic uncertainty in supernova cosmology inference
• Dust extinction has significantly reduced effect in NIR bands
• NIR SN Ia are good standard candles (Elias et al. 1985, Meikle 2000, Krisciunas et al. 2004+, Wood-Vasey et al. 2008, Mandel et al. 2009).
• Observe in NIR!: PAIRITEL /CfA
19
1989ApJ...345..245C
19Tuesday, January 25, 2011
Nearby SN Ia in the NIR: PAIRITEL
Credit: Michael Wood-Vasey, Andrew Friedman
20
Observed in NIRJ (λ=1.2 μm)H (λ=1.6 μm)Ks (λ=2.2 μm)
20Tuesday, January 25, 2011
Figure 1: 142 CfA Light curves from 2000-2004 (UBV RI) and 2004-2007 (UBV ri)
CfA3:183 Optical SN Ia
Light Curves(Hicken et al. 2009)
21Tuesday, January 25, 2011
Optical+NIR Hierarchical Model InferencePTEL+CfA3 Light-curves Marginal Posterior of Dust
!10 0 10 20 30 40 50 60
6
8
10
12
14
16
18
20
22B + 2SN2005eq (CfA3+PTEL)
V
R ! 2
I ! 4
J ! 7
H ! 9
Obs. Days Since Bmax
Ob
s. M
ag
. !
kc !
mw
x
0.3 0.4 0.5 0.60
0.1
0.2
0.3
0.4
0.5
Dust Law Slope RV
!1
Ext
inct
ion (
mag)
AV
0.3 0.4 0.5 0.6R
V
!1
SN2005eq
AH
22Tuesday, January 25, 2011
Optical+NIR Hierarchical Model InferencePTEL+CfA3 Light-curves Marginal Posterior of Dust
!10 0 10 20 30 40 50 60
6
8
10
12
14
16
18
20
22
B + 2SN2006ax (CfA3+PTEL)
V
R ! 2
I ! 4
J ! 7
H ! 9
Obs. Days Since Bmax
Obs.
Mag. !
kc !
mw
x
0.2 0.3 0.4 0.5 0.60
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Dust Law Slope RV
!1
Ext
inct
ion (
mag)
AV
0.2 0.3 0.4 0.5 0.6R
V
!1
SN2006ax
AH
23Tuesday, January 25, 2011
SN NIR Population Inference: Peak Absolute Magnitudes
−18.35 −18.15
0.05
0.1
0.15
0.2
0.25
0.3
σ (
MX)
µ(MJ)
J
µ(MJ) = −18.26σ (MJ) = 0.17
−18.1 −17.9µ(MH)
H
µ(MH) = −18.00σ (MH) = 0.11
−18.35 −18.15µ(MKs)
Ks
µ(MKs) = −18.24σ (MKs) = 0.18
Marginal Distributions of SN Ia NIRAbsolute Magnitude Variances
Mandel et al. 2009 Kasen 200624Tuesday, January 25, 2011
Optical and Near InfraredLuminosity vs. Decline Rate
0.8 1 1.2 1.4 1.6
!20
!19.5
!19
!18.5
!18
!17.5
!17
!16.5
!16
! m15
(B)
Peak
Magnitu
de
0.8 1 1.2 1.4 1.6! m
15(B)
MB
B0!µ
MH
H0!µ
25Tuesday, January 25, 2011
Population Analysis
Intrinsic Correlation
Map for Abs Magnitudes
and Decline Rates
H-band provides nearly uncorrelated information on
luminosity distance
B V R I J H
B
V
R
I
J
H
! m
15(F
)
B V R I J H
B
V
R
I
J
H
! m15
(F)
B V R I J H
B
V
R
I
J
H
Peak MF
Peak
MF
!1
!0.5
0
0.5
1
!1
!0.5
0
0.5
1
!1
!0.5
0
0.5
1
26Tuesday, January 25, 2011
Host Galaxy Dust• Previous Analyses assumed all SN host
galaxy dust has same Rv
• Estimated Rv = 1-1.7 (Astier06, Conley07) if attribute all color variation to dust
• But Rayleigh Scattering: Rv = 1.2
• But Hicken09 found Rv = 1.7 with MLCS
• For individual high Av SN, Rv < 2
• But Rv may have a distribution, or depend on Av (e.g. grain growth)
27Tuesday, January 25, 2011
(Av, Rv) for Host Galaxy DustAssuming Linear Correlation
• Apparent Correlation of High Av / Low Rv• Low Av Rv ≈ 2.5 : High Av has Rv ≈ 1.7• Circumstellar dust at High Av ? • Multiple Scattering (Goobar 2008)
!0 (intercept R
V
!1 as AV " 0)
!1 (
slo
pe
)
P(RV
!1|AV)= N( R
V
!1 | !0+ !
1A
V, #
r
2)
P( !1 < 0) < 6e!04
0.25 0.3 0.35 0.4 0.45 0.5 0.55!0.05
0
0.05
0.1
0.15
0.2
2
3
4
RV
0 0.5 1 1.5 2
2
3
4
Dust Extinction AV
RV
P(RV
!1|AV)= N( R
V
!1 | 0.36+0.14AV, !
r
2=0.042)
Marginal Estimate
28Tuesday, January 25, 2011
Bootstrap Cross-validationLow-z SN
data
Training Set{Ds, zs}
Bootstrap
Test Set{Dt}
Test Set{zt}
SN Ia
Model Predict{µt}
Hubble{µ(zt)}
Compare
• Test Sensitivity of Statistical Model to Finite Sample
• Avoid using data twice for training and distance prediction
• Prediction/Generalization Error
29Tuesday, January 25, 2011
Nearby Optical+NIR Hubble Diagram
(Opt Only) rms Distance Prediction Error = 0.15 mag(Opt+NIR) rms Distance Prediction Error = 0.11 mag
Aggregate Precision ~ (0.15/0.11)2 ≈ 2
104
31
32
33
34
35
36
37
38
µ(p
red
)
h = 0.72
127 BVRI(JH) SN Ia (CfA3+PTEL+CSP+lit)
3000 5000 7000 10000 15000!1
!0.5
0
0.5
1
Velocity [CMB+Virgo] (km/s)
Diff
ere
nce
CV Pred Err (All, cz > 3000 km/s) = 0.14 mag (0.134 ± 0.010 intr.)
CV Pred Err (Opt+NIR & cz > 3000 km/s) = 0.11 mag (0.113 ± 0.016 intr.)CV Pred Err (Opt only & cz > 3000 km/s) = 0.15 mag (0.142 ± 0.013 intr.)
OpticalOptical+NIR
30Tuesday, January 25, 2011
33 33.2 33.4 33.6 33.8 34 34.2 34.40
0.5
1
1.5
2
2.5
3
3.5
4
4.5
SN2005el
Distance Modulus µ
PD
F
P(µ | BV)
P(µ | BVRI)
P(µ | BVRIJH)
E(µ |z) ± (300 km/s)
0.2
0.4
0.6
BV
SN2005el
0.2
0.4
0.6E
xtin
ctio
n A
V
BVRI
33.3 33.4 33.5 33.6 33.7 33.8 33.9 34 34.1 34.2
0.2
0.4
0.6
Distance Modulus µ
BVRIJH
30.5 31 31.5 32 32.50
0.5
1
1.5
2
2.5
3
3.5
4
SN2002bo
Distance Modulus µ
PD
F
P(µ | BV)
P(µ | BVRI)
P(µ | BVRIJH)
E(µ |z) ± (300 km/s)
0.4
0.6
0.8
1
1.2
1.4
BV
SN2002bo
0.4
0.6
0.8
1
1.2
1.4
Extinction A
V
BVRI
31.2 31.4 31.6 31.8 32 32.2 32.4
0.4
0.6
0.8
1
1.2
1.4
Distance Modulus µ
BVRIJH
31Tuesday, January 25, 2011
Improved Distance Precision
for Individual Opt+NIR LCs
• Precision = 1/Variance
• On avg, 2.2x better BVRI vs BV
• 3.6x better BVRIJH vs BV
• 60% better BVRIJH vs BVRI
34 34.5 35 35.50
0.5
1
1.5
2
2.5
3
3.5
4
SN2006cp
Distance Modulus µ
P(µ | BV)P(µ | BVRI)P(µ | BVRIJH)E(µ |z) ± (300 km/s)
0.20.40.60.8
BV SN2006cp
0.20.40.60.8
Extin
ctio
n A V
BVRI
34.4 34.6 34.8 35 35.2 35.4
0.20.40.60.8
Distance Modulus µ
BVRIJH
32Tuesday, January 25, 2011
Summary• Hierarchical models are useful statistical methods for
discerning multiple random effects
• BayeSN: an efficient MCMC Sampler for computing inferences with SN hierarchical models
• Apparent differential trend of Rv vs Av (local dust at high Av?)
• NIR Light Curves have low correlation with optical, provide independent information on distance
• SN Ia Optical with NIR: Better dust and distance estimates than with Optical alone
33
33Tuesday, January 25, 2011
Future Work & Problems
• Application to Larger Sample of Opt+NIR SN Ia
• Application to high-z SN Ia & Cosmological Inference
• Accounting for Selection Effects
• Using Auxiliary Information
• Host Galaxy Information (e.g. P. Kelly, et al. 2010)
• Spectral? Blondin, Mandel, & Kirshner 2011
• Foley & Kasen 2011 (Color / Ejecta Velocity)
34Tuesday, January 25, 2011
Spectral Info correlate with SN Ia luminosity and light curves?
Blondin, Mandel, Kirshner 2011
Multiple Comparisons Problem
35Tuesday, January 25, 2011
Correlating Spectral Ratios with Luminosity
A&A 526, A81 (2011)
WRMS residuals [mag] Absolute Pearson correlation
R
(x1,R)
(c,Rc)
(x1, c,Rc)
Fig. 7. Results from 10-fold cross-validation on maximum-light spectra. From top to bottom: R only; (x1,R); (c,Rc); (x1, c,Rc). The left columnis color-coded according to the weighted rms of prediction Hubble residuals, while the right column corresponds to the absolute Pearson cross-correlation coe!cient of the correction terms with uncorrected Hubble residuals.
results in a Hubble diagram with lower scatter when comparedto the standard (x1, c) model. Using a flux ratio alone, Baileyet al. (2009) find R(6420/4430) as their most highly-rankedratio, while we find R(6630/4400) (see Table 2). The wave-length bins are almost identical, and in any caseR(6420/4430) is
amongst our top 5 ratios. For this ratio we find ! = !3.40±0.10,in agreement with ! = !3.5± 0.2 found by Bailey et al. (2009)3.3 In fact Bailey et al. (2009) find ! = +3.5 ± 0.2, but this is due to atypo in their equation for the distance modulus: !R really appears as anegative term in their paper (S. Bailey 2010, priv. comm.).
A81, page 10 of 24
A&A 526, A81 (2011)
WRMS residuals [mag] Absolute Pearson correlation
R
(x1,R)
(c,Rc)
(x1, c,Rc)
Fig. 7. Results from 10-fold cross-validation on maximum-light spectra. From top to bottom: R only; (x1,R); (c,Rc); (x1, c,Rc). The left columnis color-coded according to the weighted rms of prediction Hubble residuals, while the right column corresponds to the absolute Pearson cross-correlation coe!cient of the correction terms with uncorrected Hubble residuals.
results in a Hubble diagram with lower scatter when comparedto the standard (x1, c) model. Using a flux ratio alone, Baileyet al. (2009) find R(6420/4430) as their most highly-rankedratio, while we find R(6630/4400) (see Table 2). The wave-length bins are almost identical, and in any caseR(6420/4430) is
amongst our top 5 ratios. For this ratio we find ! = !3.40±0.10,in agreement with ! = !3.5± 0.2 found by Bailey et al. (2009)3.3 In fact Bailey et al. (2009) find ! = +3.5 ± 0.2, but this is due to atypo in their equation for the distance modulus: !R really appears as anegative term in their paper (S. Bailey 2010, priv. comm.).
A81, page 10 of 24
K-fold Cross-ValidationMultiple Comparisons
Blondin, Mandel, Kirshner 2011
36Tuesday, January 25, 2011
Open Problems• Photometric Classification of SN Light Curves
• Classification of SN by Spectra
37Tuesday, January 25, 2011