Upload
vucong
View
213
Download
0
Embed Size (px)
Citation preview
Simeon Bird
Johns Hopkins University (& UC Riverside)
Strong Absorbers in the Lyman-alpha Forest and Primordial Black Holes
Talk Outline
1) A Lyman alpha forest emulator
2) Detecting DLAs with machine learning
3) Effects of DLAs/LLS on flux power
4) Primordial Black Holes
Lyman alpha Forest
● Statistics of hydrogen absorption ● Indirectly measures small-scale matter power
Ed Wright
Interpolation
● Want >104 parameter samples● Interpolate; estimate output of simulations from
existing set without actually running them.● Need map from cosmology → observables
For DESI need < 1% level accuracy
Interpolation
Universalis Cosmographia, the Waldseemüller map dated 1507 (wikipedia)
● Early maps of the world had a similar problem
● We are mapping Panamato 1% without ever visiting!
Measurements
Quadratic Interpolation
● Current analysis● Vary one parameter at a time● Flux power interpolation quadratic polynomial
Gaussian Processes
● Bayesian function interpolation, which computes probability distribution of f(x) conditional on input set.
● Standard python implementation
Left: Prior on function Right: Posterior conditioned on samples
Rasmussen & Williams (GPML)
Gaussian Processes
● Magic in: Kernel function● Describes how correlation between function
values depends on parameter distance.Left: Prior on function Right: Posterior
conditioned on samples
Rasmussen & Williams (GPML)
Gaussian Processes
● Magic in: Kernel function● Describes how correlation between function
values depends on parameter distance.● Dot kernel (linear interpolation) plus squared
exponential:
● L is estimated from the samples, for every parameter
Rasmussen & Williams (GPML)
Gaussian Processes
● Gaussian Process also estimates interpolation error● Include this uncertainty in likelihood.● Interpolation error just increases uncertainty
Emulator Results
● Each line is a different redshift● Ratio of prediction to extra simulation● 30 simulations, parameters are σ8 and ns
GP+hypercube Quadratic polynomial
Emulator Results
● Much better interpolation!● Increase simulation density in high confidence
region
GP+hypercube Quadratic polynomial
Talk Outline
1) A Lyman alpha forest emulator
2) Detecting DLAs with machine learning
3) Effects of DLAs/LLS on flux power
4) Primordial Black Holes
What are DLAs?
DLAs: neutral hydrogen reservoirs at z=4-2
Neutral hydrogen density: red is DLA, blue is LLS NHI > 2x1020 cm-2
Prochaska+ 2008
Teaching Computers to Detect DLAs
Roman Garnett , Shirley Ho, Me.
1605.04460, 1610.01165
Strong absorbers with NHI > 2x1020 cm-2
Galaxies: from galaxy physics, not cosmology
Foregrounds for Lyman alpha forest
Finding DLAs in Spectra
Normally done by visual inspection of spectraLook for wide dips in the spectrum below:
Finding DLAs in Spectra
Currently done by visual inspection of spectraLook for wide dips in the spectrum below:
Finding DLAs in Spectra
So we taught a computer to do thisLearned GP kernel for continuum + absorption
Finding DLAs in Spectra
Learn Bayesian model for quasar without DLA
Learn and test with DR9 visual catalog!
Model for DLA added using Voigt profile
Emission model
Absorption model+noise
Finding DLAs in Spectra
Likelihood of learned model vs like. of DLA.
Get DLA probability for poor detections: use all data, even if SNR < 1
Emission model
Absorption model+noise
Conclusions
• This gives us posterior probability for DLA
• Completely automatic
• Can clean all data, even with SNR < 1
• Catalogue purity vs completeness adjustable
• Extendable to other quasar features
Talk Outline
1) A Lyman alpha forest emulator
2) Detecting DLAs with machine learning
3) Effects of DLAs/LLS on flux power
4) Primordial Black Holes
Lyman alpha Forest
● Undetected DLAs or Lyman Limit Systemsare a foreground and mess up the statistics
● Not in Lyman alpha forest simulations
Ed Wright
Characterising DLAs/LLS
Template for effect on flux power
with Keir Rogers, Peiris, Pontzen, Font-Ribera
State of the art: McDonald 05:– inserted DLAs
– computed power spectrum
– Total abundance free parameter
Characterising Residuals
1. It is easier to detect strong DLAs
Undetected population is not total population
Column density dependent template
2. DLAs in smaller halos at high redshif
Redshif dependent template
Characterising Residuals
Illustris simulation:
● 75 Mpc/h box● Galaxy physics
Contains DLAs: no need to insert themJust make spectra
Illustris Galaxy Model: Ok!
Abundances Metallicity (halo mass)
Has roughly correct DLA abundance and clustering
DLA Flux Power Effect
Flux power spectrum including dense gas
Compare flux P(k) without LLS or DLA spectra
Find multiplicative bias
Strong Absorber Flux Redshift Evolution
● Higher redshift → more flux power● We also made a fitting function:
Comparison to McDonald 05
Our total is similar to McDonald
Power masking strong DLAs very different !
Did LIGO detect dark matter?
Simeon Bird (JHU)
I. Cholis, J. Munoz, Y. Ali-Haimoud, M. Kamionkowski, E. Kovetz, A. Raccanelli, A. Riess
arXiv: 1603.00464 PRL 116 201301
Did LIGO Detect Dark Matter?
Are LIGO black holes from stellar evolution? Or primordial?
Form from inflationary perturbationsSmall scales, no theory constraints
(LIGO collaboration)
Primordial Black Hole Dark Matter
Black Holes are all the dark matter
Form halos
Gravitational wave emission forms binaries
Does the merger rate match LIGO?
Does the Merger Rate Match LIGO?
Make standard assumptions about dark matter density
NFW profile, Press-Schechter halo mass function, concentration-mass relation
Gravitational Wave Cross-Section
Mergers dominated by the smallest halos:relative velocity ~ halo virial velocity
Mergers are fast compared to binary formation
(Quinlan & Shapiro 1989)
Merger Rate
Total mergers:
This number could have been 10±10
INTERESTING
Did LIGO Detect Dark Matter?
???
Stellar Heating
Black holes would heat & expand star clusters:
Existence of star clusters constrains PBHs
Brandt 2016
Disruption of Star Clusters
Caveats:
– Central black holes?
– Structure constraints need simulations
Detecting Primordial Black Holes
• Micro-lensing: Black hole in front of star
• Star brighter
• Strong constraints on PBH and other MACHO
More Microlensing
• No micro-lensing constraints if lens is too rare to go past
• Non-detection of lensing (OGLE, HSC):
More Microlensing
Star crossed a DM caustic and was highly magnified
PBH dark matter would disrupt the smooth caustic and reduce the max magnification
This is only one star...
Mass Function
Two populations → mass function bump
Curiously, 7/10 LIGO BHs are ~ 30 solar
Kovetz+ 2017