Statistical challenges in analyzing

LIGO gravitational wave data

TASSGW ICTS-SAMSI Workshop

Research Triangle Park, May 2017

LIGO DCC G1700785 1

Jess McIver for the LIGO Scientific Collaboration

2NASA

Gravitational wavesRipples in the fabric of spacetime

generated by the acceleration of matter

3

Gravitational wave

propagation

Spacetime strain h(t) measured as

LIGO DCC P1500072

Observing GWs with interferometry

4

Where are the LIGO detectors?

The matched filter analysis

arXiv 1606.04856

GW150914

GW151226

LVT151012

Unmodeled time-frequency GW

analyses

arXiv 0802.3232

arXiv 1410.3835

• Reversible- jump Markov-chain Monte Carlo algorithm that models signal

and transient noise events as Morlet-Gabor wavelets

• Estimates the posterior distribution of signal and noise wavelet parameters

8PRL 116. 061102

9

LIGO data is non-stationary!

Blip glitches• The biggest contributor to the

transient GW search backgrounds

• Seen in both LIGO detectors (non-

coincident)

• No known correlation with

instrument behavior or

environment.

60-200 Hz non-stationary noise

• Pollutes LIGO-Livingston data in a

critical frequency range (~50-500Hz)

• Longer duration (10s or 100s of

seconds)

• Major contributor to CBC and burst

backgrounds

10

Templates most susceptible to background

noise

B. P. Abbott et al., in preparation

Highest re-weighted SNR of LIGO-Livingston CBC triggers during O1

By template duration and peak frequencyBy effective spin and total mass

11

time

fre

qu

en

cy

time

fre

qu

en

cy

time

fre

qu

en

cy

time

fre

qu

en

cy

12

Applying machine learning to LIGO noise

gravityspy.org Zevin et al, 2017, CQG

Diagnosing noise: auxiliary

channels

CQG 28, 13 (2012)

We record over 200,000 channels per detector that monitor the

environment and detector behavior.

We can use these to help trace the instrumental causes of glitches

that pollute the search backgrounds.

Physical environment channels

CQG 28, 13 (2012)

SAMSI astro

working group III

(Multivariate and

Irregularly

Sampled Time

Series) now has

access to two

weeks of LIGO h(t)

and PEM channels

for a prior science

run (S6) from an

MOU with the

LIGO Scientific

Collaboration.

15

The tale of the noisy

refrigerator

Once a noise source that contributes to the background is identified, ideally it

is fixed in hardware or software.

If this is not possible, the noisy data is vetoed using auxiliary channel data.

16

• Measure non-Gaussianity of the h(t) gravitational wave

data

• Suggestion: use ARIMA

• Predict glitchy h(t) response based on the behavior of

the auxiliary channels (very hard)

• Change-point detection of interferometer behavior

based on the auxiliary channels

• Could start with detecting day/night traffic cycle

• Lock loss (loss of light resonance) diagnostics

• Inferring properties about noise background

distributions

• Change point detection based on trigger rate

• Downweight data of poor quality

• Dealing with outliers: fit background distributions using

mixture models (Gaussians, Weibull, Student-T.)

Open challenges

17

LIGO-Livingston h(t)

LIGO-Livingston transient noise

during the second observing run

18

LIGO-Livingston h(t)

LIGO-Livingston transient noise

during the second observing run

.

19

Sliding an injected signal in real Advanced

LIGO noise

Injected signal properties:

• GW151226-like in mass (8,14

M_sun) zero spin

• Injected into two-detector LIGO

network

• All injections identical in mass,

spin, sky position, orientation

with an SNR of 30

Noise properties:

• Time selected from the first

part of Advanced LIGO’s

second observing run

• Scatter in LIGO-Livingston

• Clean data in LIGO-Hanford

20

Impact of scattering on sky localization of

8+14M_sun BBH GW signal with SNR=30

Minimum 90% confidence sky area

(2 seconds before the scattering noise

feature): 300 sq. deg.

Maximum 90% confidence sky area:

(During the first 0.5 seconds of the

scattering noise): 540 sq. deg.

Parameter estimation produced

with the lalinference pipeline:

arXiv 1409.7215

21

Posteriograms

Made by TJ Massinger

Show evolution of 1D

posterior distribution

function over time.

Example: declination

Challenge: comparing distributions

22

Goal: to quantify the change in estimated posterior distributions

for target parameters (sky location, masses, spins) over time

relative to some reference pdf(s)

Reference pdf (“clean” data) pdf to be compared (“glitchy” data)

Comparing distributions

23

Current approach: Kullback–Leibler divergence

Open questions:

• How to average two posterior distributions (with

different evidences)

• Better approaches than KL-divergence?

Gaussian distributions p(x) and q(x) KL area to be integrated

wikipedia

WG III has a data-sharing MOU with the LIGO

Scientific Collaboration to work on these

problems (the first of its kind)

24

Actively looking for collaborators! Insights

welcome.

25

Extra slides

26

Comparing LIGO events

arXiv 1606.04856

SNR 23.7

SNR 9.7

SNR 13.0

27

Laser glitches

h(t) vs.

microphones

27

h(t)

PSL microphone