Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Suk Yee Yong
Supervisor: Professor Rachel Webster
Co-supervisor: Anthea King
Collaborators: Kathleen Labrie (Gemini);
Matthew O’Dowd (CUNY); Nick Bate (Cambridge)
Infer Structure of Quasar with Machine
Learning
ASA 2018
Overview
1 Quasar Population
2 MethodologySamplesApproach
3 ResultsMachine LearningStatistical TestsImplicationsProposed Disk-wind Model
4 Summary
Quasar Population
Broad absorption line quasars(BALQs)
Show BAL feature
∼ 20% of all quasars
Possible explanations:
I Evolutionary: BALQ →non-BALQ
I Orientation in a narrowwind context
Image credit: http://www.sdss3.
org/dr10/algorithms/qso_catalog.
php
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 1/11
Narrow Disk-wind Model
Invoke wind component toexplain emission andabsorption features of quasarspectra
BALs are viewed throughthe narrow wind
Expect differences inemission line propertiesdepending on theorientation:
I Equatorial narrow wind:broad line width, lessblueshifted, dominated byrotational motion
I Polar narrow wind: narrowline width, more blueshifted
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 2/11
Testing the Orientation Paradigm
VSBALQs non-BALQs
Are there any differences in their properties?
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 3/11
Testing the Orientation Paradigm
VSBALQs non-BALQs
Are there any differences in their properties?
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 3/11
Methodology
Samples
Data: 2773 quasars from Sloan Digital Sky Survey III(SDSS-III) Data Release 12 Quasar (DR12Q; Paris et al.2017) catalogue
C iv BALs: 313 quasars using traditional BAL definition ofWeymann et al. (1991)
Emission lines: C iv and Mg iv
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 4/11
Methodology
Approach
Supervised machine learning classification
I Decision tree, random forest, logistic regression, support vectormachine (SVM)
I Conduct grid and randomised search of parameter space
I Extract the feature importance and weighting
Additionally for comparison, conduct statistical tests:
I Anderson-Darling (A–D), Kolmogorov-Smirnov (K–S)
I Null hypothesis being samples from the same distribution
I p-value< 5% as significant, suggesting the two samples are notdrawn from the same distribution
Input features: Continuum and emission line properties
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 5/11
Methodology
List of Investigated Features
Property Feature Description
Continuum imag Absolute magnitude in i-band at z = 2z.pca PCA redshiftalphanu Spectral index
Emission fw(civ) FWHM of C ivciv ratioskew Asymmetry of C ivw(civ) EW of C ivfw(mgii) FWHM of Mg iimgii ratioskew Asymmetry of Mg iiw(mgii) EW of Mg iicivmgii diffv Velocity offsets of C iv and Mg iicivmgii ratiofwhm FWHM ratio of C iv and Mg iicivmgii ratioew EW ratio of C iv and Mg ii
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 6/11
Results
Machine Learning
alphanu
civ_ratio
skew
civmgii_diffv
civmgii_ratio
fwhm
civmgii_ratio
ewz.p
cafw
(civ)
fw(m
gii)
mgii_ratio
skeww(civ)
w(mgii)
imag
Feature
0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Imp
ort
an
ce
decision tree (grid)decision tree (rand)random forest (grid)random forest (rand)
Decision tree and random forest
alphanu
civ_ratio
skew
fw(m
gii)
civmgii_ratio
ew
mgii_ratio
skeww(civ)
fw(civ)
civmgii_diffv z.p
caim
ag
w(mgii)
civmgii_ratio
fwhm
Feature
0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Weig
ht
logistic regression (grid)logistic regression (rand)svm (grid)svm (rand)
Logistic regression and SVM
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 7/11
Results
Machine Learning: Comparison to predictions
Algorithm Decision Tree Random Forest Logistic Regression SVM
Class non-BALQ BALQ non-BALQ BALQ non-BALQ BALQ non-BALQ BALQ
Gri
dSea
rch
non-BALQ 73.40 48.57 79.59 41.43 66.60 31.43 70.52 34.29
BALQ 26.60 51.43 20.41 58.57 33.40 68.57 29.48 65.71
Ran
dom
ised
Sea
rch non-BALQ 74.64 45.71 79.38 41.43 66.60 31.43 70.52 34.29
BALQ 25.36 54.29 20.62 58.57 33.40 68.57 29.48 65.71
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 8/11
Results
Statistical Tests
3.0 2.5 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5
alphanu
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Fre
qu
en
cy
0.0
0.2
0.4
0.6
0.8
1.0
Cu
mu
lati
ve P
rob
ab
ilit
y
Non-BAL
BAL
EDF
Non-BAL BAL
balciv
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.5
1.0
1.5
alp
han
u
A–D α=0.04%; K–S p-value=3.05× 10−7%
0.0 0.5 1.0 1.5 2.0 2.5 3.0
civmgii_ratiofwhm
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Fre
qu
en
cy
0.0
0.2
0.4
0.6
0.8
1.0
Cu
mu
lati
ve P
rob
ab
ilit
y
Non-BAL
BAL
EDF
Non-BAL BAL
balciv
0.0
0.5
1.0
1.5
2.0
2.5
3.0
civm
gii
_rati
ofw
hm
A–D α=0.06%; K–S p-value=2.06%
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
civmgii_ratioew
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
Fre
qu
en
cy
0.0
0.2
0.4
0.6
0.8
1.0
Cu
mu
lati
ve P
rob
ab
ilit
y
Non-BAL
BAL
EDF
Non-BAL BAL
balciv
0
2
4
6
8
10
12
14
civm
gii
_rati
oew
A–D α=0.04%; K–S p-value=0.09%
0.0 0.5 1.0 1.5 2.0 2.5 3.0
civ_ratioskew
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Fre
qu
en
cy
0.0
0.2
0.4
0.6
0.8
1.0
Cu
mu
lati
ve P
rob
ab
ilit
y
Non-BAL
BAL
EDF
Non-BAL BAL
balciv
0
1
2
3
4
5
6
civ_
rati
osk
ew
A–D α=0.08%; K–S p-value=0.21%Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 9/11
Results
Implications
Difficult to distinguish BALQs from non-BALQs
The two populations generally exhibit similar continuum andemission line properties
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 10/11
So can we use this information to infer the structure of quasars?
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 10/11
Results
Proposed Disk-wind Model
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Results
Proposed Disk-wind Model
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Results
Proposed Disk-wind Model
Black holeDisk Torus
Ionisation cone
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Results
Proposed Disk-wind Model
Wide wind
Dense radial streams
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Results
Proposed Disk-wind Model
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Results
Proposed Disk-wind Model
Suk Yee Yong Infer Quasar Structure with ML [arXiv: 1806.07090] 11/11
Summary
Aim: Test orientation based on narrow disk-windexplanation for BAL phenomenon
Method: Using statistical tests and supervised machinelearning in an attempt to separate BALQs and non-BALQsbased on their continuum and emission line properties
Result: The two populations show similar properties
Proposed disk-wind model: Wide wind opening angle withmultiple radial streams of dense clumps
arXiv:1806.07090