SOSTAT 2021: Time Series Analysis

Time series analysis
Dr. Abbie Stevens
Michigan State University & University of Michigan
[email protected]
@abigailStev
github.com/abigailstev

View Slide

Outline
•Intro to the Fourier domain
•Evenly-spaced and irregular time series
•Power spectra/periodograms
•Coffee break and tutorial time
•Wavelets, Hilbert-Huang transform
•Spectrograms/dynamical power spectra
•Examples of astronomical signals
•More coffee and more tutorial time
IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich 2

View Slide

The Fourier series
Any function can be represented as a sum of sines and
cosines (with some coefficients, which may also be functions)
f(x) = ∑ An
cos(n 𝜋 x) + ∑ Bn
sin(n 𝜋 x)
n=0 n=1
∞ ∞
Slide adapted from J. McIver

View Slide

The Fourier series
Image credit:
J. Belk via
Wikimedia
Example: first four Fourier
approximation terms for a
square wave
The more terms you add, the
closer the approximation gets

View Slide

The Fourier transform
Take a periodic or well-bounded function (of time or
space) by projecting f(t) onto an orthogonal basis of
sines and cosines
F(𝜈) = ∫ f(t) e-i2𝜋𝜈t dt
f(t) ⇾ F(𝜈) [or f(𝜈)]
^
f(t) = ∫ F(𝜈) ei2𝜋𝜈t d𝜈
Fourier transform
Inverse Fourier
transform

View Slide

IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich GIF: L. Vieira via Wikimedia 6
f(t) ⇾ F(𝜈) [or f(𝜈)]
^
Think of it like
decomposing the time
series function into its
component frequencies

View Slide

IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich
Problem
solution
solve
(hard)
Transformed
problem
Transformed
solution
solve
(easy)
8
Fourier
transform
inverse Fourier
transform
The Fourier transform and inverse Fourier transform
make “Fourier pairs”

View Slide

Sampling effect: Nyquist frequency

View Slide

Sampling effect: Nyquist frequency
•In practice, cannot add terms forever!
•Highest frequency you can sample:
𝜈Nyquist
= 1/(2*dt) = d𝜈/2

View Slide

Positive and negative Fourier frequencies
• Mirrored about 0 Hz
• In an array, often arranged: (0, pos., Nyquist, neg.)
Image: J. VanderPlas 2018

View Slide

Evenly-sampled time series
• Signal period << observation length
• In X-rays and gamma-rays, we count photons. It’s very possible
to have zero counts -- “sparse” light curves are common
• Instead of saving light curves with lots of zeroes, we use
event lists – can select a dt multiple of the detector’s dt
• In optical, bright-enough sources mean you detect flux above
background in every time bin (e.g., every 30 seconds)
• Kepler, K2, TESS can give evenly sampled time series
1700 1702 1704 1706 1708 1710
2000 4000 6000 8000 10
4
1.2×10
4
Count/sec
Time (s)
Start Time 12339 7:28:14:566 Stop Time 12339 7:29:32:683
Bin time: 0.7812Eï02 s
12

View Slide

Applying Fourier transforms to data
1016 1018 1020 1022 1024
5000 104 1.5×104
Count/sec
Time (s)
Time domain
Light curve
Frequency/Fourier domain
Power density spectrum
FOURIER
TRANSFORM2
Light curve broken into equal-length chunks or segments,
take power spectrum of each chunk, average those together
x(t)→X(ν)
P(ν)=X(ν)X*(ν)
=|X(ν)|2
13

View Slide

Irregularly sampled time series
• Many (most?) astronomical time series will be irregularly sampled
Ø Signal period > observation length
Ø Observing cadence
Ø Weather
Ø Visibility
Reference: J. VanderPlas 2018, “Understanding the Lomb-Scargle Periodogram”

View Slide

Lomb-Scargle periodogram saves the day!
• Multiplies your Fourier signal X(ν) with the Fourier transform
of the sampling window ⇾ “convolution”
Reference: J. VanderPlas 2018, “Understanding the Lomb-Scargle Periodogram”
When interpreting:
beware many noisy
peaks and harmonics
due to the sampling
window convolved
with noise in the data

View Slide

X-ray variability: Hard to see by eye
1014 1016 1018 1020 1022
5000 104 1.5×104
Count/sec
Time (s)
CYGNUS_Xï1
1696 1698 1700 1702 1704
4000 5000 6000 7000
Count/sec
Time (s)
GRS1915+105
Light
curves
Power
density
spectra
Noise: Cygnus X-1 Signal: GRS 1915+105
16

View Slide

QPOs → Damped harmonic oscillators
y = cos(⍵t)
17
I will discuss a little QPO physics
in the next lecture later today

View Slide

y = cos(⍵t) x e-bt
b=0
b=0.08
18

View Slide

b=0
b=0.08
b=0.22
19

View Slide

b=0
b=0.08
b=0.22
b=0.5
20

View Slide

b=0
b=0.08
b=0.22
b=0.5
b=1.0
The stronger the damping,
the wider the peak
21

View Slide

Poisson noise (“white noise”)
Poisson noise from
counting photons;
power-law slope=0

View Slide

Poisson noise + Lorentzian QPO

View Slide

Sampling effect: Aliasing
Resonance
between signal
freq. and
sample freq.
gives false
feature
Video credit: Honda
Windowing: similar
false feature due to
segment (“window”)
length

View Slide

Tutorial time
•In the SOSTAT2021 Jupyter hub:
tutorials/abigail_timeseries/timeseries_workbook.ipynb
•On my GitHub:
https://github.com/abigailStev/timeseries-tutorial

View Slide

Wavelets
• Fourier products don’t have
an intrinsic way to tell time
resolution (i.e., when in the
light curve the signal is
present)
• Wavelets easily represent a
signal in the time domain
and in the frequency
domain
Resource: “A really friendly guide to wavelets”, C. Valens, 1999

View Slide

Wavelets aren’t great for everything
•Averaged power spectra (~50+ segments) follow a
chi-squared distribution with 2 degrees of freedom,
about the underlying true power spectrum
• Errors are statistically well-defined and well-understood (and
easy to compute!)
•Wavelets do not follow such a well-defined and well-
known distribution
•No clear, easy way to assess statistical significance of a
signal (which is one of the things we often want to do)

View Slide

Wavelets are fantastic for gravitational waves!
See also: “Q-transform”

View Slide

Time (s)
40 41 42 43 44 45 46 47 48 49 50
Frequency (Hz)
1
2
3
4
5
6
7
8
9
10
Gaussian smoothing amplitude
0
5
10
15
20
25
30
Hilbert-Huang transform
• Frequency-domain product designed for data that are non-
stationary and non-linear
• Like an instantaneous Fourier transform → gives (some) time
localization!
• Error from standard
deviation of (1000+)
simulations
IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich See Su+15 & refs therein for application to black hole data
0.01 0.1 1 10 100
10 100
Frequency(Hz)
Power (Leahy)
Hilbert spectrum from Su+15 of a 4Hz QPO
29

View Slide

Elapsed time (in 64 s segments)
Power (rms2/Hz)
Elapsed time (in 64 s segments; not continuous)
Spectrogram (dynamical power spectrum)
•Evolution of a power spectrum in time
•Instead of averaging together, plot in colormap

View Slide

Elapsed time (in 64 s segments)
Power (rms2/Hz)
Elapsed time (in 64 s segments; not continuous)
Good Time Intervals (GTIs)
• Jumps in the spectrogram below are cut out
• Looks like tophat windows in big light curve, but the segments we use for
the periodogram and averaging are much smaller than the window length

View Slide

Things to sometimes worry about
• Deadtime occurs with X-ray detectors if your (bright) source is
emitting photons faster than your detector can handle. Once some
chip of the detector has detected a photon, it cannot detect another
photon until it reads out its existing photon detection through the
electronics.
ØMeasurable as deviation from expected Poisson noise power-
law at high frequencies in the power spectrum
ØAccumulates over an observation; for a few ks observation,
could have several seconds of deadtime to adjust the exposure
time by
• Pile-up is a sibling of deadtime that aﬀects spectra

View Slide

Slide adapted from D. Huppenkothen; credit: NuSTAR Observatory Guide
‘Deadtime’ ⇾ the detector is effectively dead for the brief readout period.
NuSTAR

View Slide

Slide adapted from D. Huppenkothen; credit: Huppenkothen & Bachetti 2021 (in press)
Huppenkothen & Bachetti (under review)

View Slide

Poisson noise (a form of “white noise”)
• In power spectra: P~ν𝛼, 𝛼=0
• Not correlated on any timescale

View Slide

Flicker noise
• In power spectra: P~ν𝛼, 𝛼=-1
• Correlated on medium-short timescales (short “memory”)

View Slide

Red noise (random walk)
• In power spectra: P~ν𝛼, 𝛼=-2
• Correlated on long timescales (long “memory”)
• Ornstein-Uhlenbeck: ~red noise + friction: tends
towards a mean value over long time

View Slide

Beware of red noise!
• Cannot apply standard peak-finding algorithms, since those
assume white noise (see Vaughan & Uttley ‘06)
• Bigger issue for SMBHs
than stellar BHs
due to timescales
White/Poisson noise
Red noise
Smith+18b
Fourier frequency

View Slide

Red noise vs signals
−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(a)
−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(b)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(c)
PG 1302-102, CRTS data
Vaughan+16 (figure); Liu+18
Data looks periodic!

View Slide

−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(a)
−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(b)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(c)
Uneven sampling, gappy
data, only ~1.5 cycles

View Slide

−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(a)
−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(b)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(c)
Vaughan+16 (ﬁgure); Liu+18
Also including LINEAR data
(but it isn’t)
Sampling a random red
noise process in same
way can look like a
“periodic” signal
41

View Slide

−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(a)
−2 0 2 4 6 8 10 12
time (yr)
V mag
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
15.2 15.0 14.8
(b)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(c)
Sampling a random red
noise process in same
way can look like a
“periodic” signal
Also including LINEAR data
When in doubt, simulate!
Also: claimed periodicity in
J0045+41 disproven by
Barth & Stern ‘18
42

View Slide

Cross spectra
x(t)→X(ν) for a narrow energy band
y(t)→Y(ν) for a broad-energy reference band
As you average
segments together:
signal adds,
noise cancels
CXY
(ν)=X(ν)Y*(ν)
real
imaginary
43
real
imaginary

View Slide

Cross spectra
x(t)→X(ν) for a narrow energy band
y(t)→Y(ν) for a broad-energy reference band
Cospectrum: real part of the cross spectrum (see Bachetti+15,
Bachetti+Huppenkothen 18 and Huppenkothen+Bachetti 18 for statistical details)
Note: for X(ν)=Y(ν),
cospectrum = cross amplitude
= power spectrum
44
Also used:
amplitude of the
cross spectrum

View Slide

What other science cases might use
these timing analysis tools?

View Slide

Strohmayer 2001
0.1 1 10 100
2
4
6
8
Type-A
0.1 1 10 100
101
Leahy Power
Type-B
0.1 1 10 100
Frequency [Hz]
101
Type-C
Stevens
Mo1a+17a
QPOs in black holes and neutron stars
• High-frequency: 100’s Hz
• Hot Keplerian blobs in inner disk?
• Low-frequency: ~0.01-10’s Hz
• Precession of corona/hot flow?
Magnetic warps in disk?

View Slide

• Short-timescale
variability changes
on long timescales
(spectral state-
dependent)
• Short-timescale
variability is energy-
dependent
Corona-dominated state
Mostly-corona-dominated state
Disk-dominated state
Few months
for full
spectral state
transition
Mostly-disk-dominated state
If you want to read more, see the power colours paper by Heil+15a 47

View Slide

• Short-timescale
variability changes
on long timescales
(spectral state-
dependent)
• Short-timescale
variability is energy-
dependent
A B
C D
2-60 keV
6.5-13.1 keV 13.1-60 keV
2-6.5
keV
Homan+01 48

View Slide

Pulsations in neutron stars
• Spin-down: decreasing spin frequency
(e.g., losing rotational energy to the
environment)
• Spin-up: increasing spin frequency (e.g.,
accreting material and thus increasing
angular momentum)
• Glitch: sudden change in spin frequency
(due to superfluid NS core?)
• Seen in residuals of frequency or pulse
timing
-400
-300
-200
-100
0
Timing residuals (ms)
(a)
-100 -50 0 50 100
Days from MJD = 53067.1
-3.750
-3.745
-3.740
-3.735
ν (10-10 Hz s-1)
•
(d)
-100 -50 0 50 100
0
1
2
3
4
5
6
Δν (μHz)
(c)
-100 -50 0 50 100
-200
-100
0
100
200
Timing residuals (ms)
(b)
Espinoza+11 49

View Slide

Stellar pulsations
• Cepheid variables, RR Lyrae stars, Delta Scuti variables,
Blahzko effect (long-period
modulation of the periodicity)
• Period-luminosity relation
makes them standard candles
used as “cosmic distance ladder”
• Slow enough (periods > hours)
that time-domain photometry
is often used
IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich Image via APOD, credit: R. Vanderbei, ESA/Gaia/DPAC 50

View Slide

Asteroseismology (“starquakes”)
• Understanding the internal structure of stars using their
brightness oscillations
• Convective zone excites oscillations
• Fourier analysis of light curves:
often see many harmonics
IAA-SOSTAT 2021 ☆ Abbie Stevens, MSU & UMich Info thanks to online slides by T. Bedding and refs therein
Aerts+19
51

View Slide

QPOs in active galactic nuclei (AGN)
• ~1 hr “periodicity”, 91ks observation
• RE J1034+396 is a narrow-line Seyfert 1
AGN
• Saw 16 ‘cycles’ (periods) in one
uninterrupted observation!
• Evenly-sampled time bins
• Signal attributed to high-freq. QPO
• If at innermost stable circular orbit,
MBH
~7x106-1x107 M☉
Gierlinski+08
Alston+14
52

View Slide

Smith+18b
• 44 day low-freq. QPO in KIC 9650712
• NLS1 in original Kepler field
• 30-minute cadence over 3.5 years: ~30 cycles
• Tested periodicity via simulations (Uttley+02) and Lomb-
Scargle periodogram
QPOs in active galactic nuclei (AGN)

View Slide

• Open-source timing and spectral-timing software
(Astropy affiliated package)
Øhttps://docs.stingray.science
• Stingray: Python library of analysis tools
• HENDRICS: shell scripting interface
• DAVE: graphical user interface
• Tutorials in Jupyter notebooks
• Huppenkothen, Bachetti, Stevens et al. 2019, ApJ & JOSS
• Google Summer of Code students in 2016-2021
Stingray
Please remember to cite
software in your papers!

View Slide

Time series resources
Evenly-sampled time series:
• Uttley et al. 2014, “X-ray Reverberation Around Accreting
Black Holes”
Irregularly-sampled time series:
• VanderPlas 2018, “Understanding the Lomb-Scargle
Periodogram”
Wavelets:
• C. Valens 1999, “A really friendly guide to wavelets”
Software tools:
• Stingray, Lightkurve, GWpy, astropy.timeseries

View Slide


View Slide

SOSTAT 2021: Time Series Analysis

SOSTAT 2021: Time Series Analysis

Dr. Abbie Stevens

More Decks by Dr. Abbie Stevens

Other Decks in Science

Featured

Transcript