| Presentations |
Joe DePasquale (CfA)
8 Sep 2015
1pm EDT
SciCen 706 |
- The Art & Science of Image Processing
- Abstract:
The Chandra Communications & Public Engagement group at the
Center for Astrophysics plays a central role in representing NASA's
Chandra X-ray Observatory to the public. Bi-weekly Chandra press
releases provide high-quality peer-reviewed science content for
non-experts, and the astronomical imagery produced to support such
releases can be considered the public face of Chandra. Each public
image is unique and a considerable effort is put into creating imagery
that is not only aesthetically pleasing, but also accurately and
effectively communicates the science. After a brief overview of the
Chandra program and the unique challenges in processing X-ray data,
I will walk through the creation of an image for public release,
providing key insights into the processing involved. With an eye
towards the future, I'll briefly describe an ongoing project to
merge the end result "pretty pictures" with science analysis.
- Presentation slides:
[.key] (v6+)
[.pdf]
-
|
Hyungsuk Tak (Harvard)
15 Sep 2015
12:37pm EDT
SciCen 706 |
- Microlensing corrections in TDC
- Abstract: I briefly overview the time delay
estimation problem and introduce a microlensing effect that
arises when stars in the lensing galaxy introduce
independent flickering noises into the quasar images. If the
timescale of this effect is much larger than that of the
quasar variability, the light curves can have different
long-term trends. Because our curve-shifting assumption does
not hold with the different long-term trends, our model
produces (approximate) marginal posterior distribution of
the time delay that has several modes near margins of the
entire range of the time delay. These modes near margins
tend to overwhelm the density near the true time delay,
making it hard to estimate the true time delay. I suggest a
way to estimate the time delay, reducing the effect of
microlensing.
- Slides [.pdf]
-
|
Xiao-Li Meng (Harvard)
06 Oct 2015
12:15pm EDT
SciCen 706 |
- Seeking Effective Adjustments for Effective Areas
- Discussion of the cross-calibration shrinkage method.
Given I instruments and J sources, measurements
in a variety of combinations thereof, to determine a regression
correction to the underlying effective areas to bring them all
in sync.
- Slides [.pdf]
-
|
David Jones (Harvard)
13 Oct 2015
1:15pm EDT
SciCen 706 |
- Designing Test Information and Test Information in Design
- Abstract:
Since telescope time is limited, astronomers wishing to classify
lightcurves must carefully select future time points at which to observe
lightcurves of interest in order to maximize the information that will
be gained for classification. This work proposes a framework for
constructing measures of test / classification / model selection
information and explores how to use them in experimental design problems
such as lightcurve classification. Degroot (1962) developed a general
framework for constructing Bayesian measures of the expected information
that an experiment will provide for estimation, and our framework
analogously constructs frequentist and Bayesian measures of information
for hypothesis testing. In contrast to estimation information measures
that are typically used in experimental design for surface estimation,
test information measures are most useful in experimental design for
model selection and classification problems. Indeed, our framework
suggests a probability based measure of test information, which in
hypothesis test applications has more appealing properties than variance
based measures. We also extend a result of Nicolae et al. (2008) linking
test and Fisher information, and propose a fundamental coherence
requirement for test information measures.
- Slides [.pdf]
-
|
Jane Huang (HU)
Peter Blanchard (HU)
27 Oct 2015
1:07pm EDT
SciCen 706 |
- Searching for vibronic progressions in the diffuse interstellar bands [JH]
- Abstract:
The diffuse interstellar bands, observed at optical, infrared, and UV
wavelengths, are a series of hundreds of absorption bands observed in
stellar spectra due to molecules in intervening clouds. Though the first
DIBs were discovered about a century ago, so far only one of the
molecular carriers has been confirmed. Identifying more of these
carriers would add to our understanding of how complex molecules form in
the interstellar medium.To constrain the characteristics of candidate
carriers, Duley and Kuzmin (2010) suggested searching for low-energy
harmonic progressions among DIBs in order to identify bands that may
arise from torsional motion of large molecules. I will discuss the use
of agglomerative clustering methods to search for harmonic progressions
and examine the likelihood of such progressions arising due to chance
alignments.
- Slides [.pptx]
-
- The Impact of Positional Uncertainty on Gamma-Ray Burst Environment Studies [PB]
- Abstract:
While it is now established that long-duration gamma-ray bursts (LGRBs)
are a rare outcome of the death of some massive stars, it remains
unclear what special conditions are required for the production of an
LGRB. Studies of the preferred locations of LGRBs within their host
galaxies can shed light on this open question. I use ground-based
detections of LGRB afterglows to locate the bursts within
high-resolution images of their faint host galaxies obtained using the
Hubble Space Telescope. I measure the distribution of LGRB offsets from
the centers of their host galaxies, and compare the brightness of the
burst position relative to the total host light distribution. The
dominant limiting factor in these studies comes from the uncertainty on
the position of the LGRB relative to its host galaxy. In this talk, I
will discuss the impact of positional uncertainty on host galaxy
identification, and offset and light distribution measurement. It is
important to address this issue to avoid biases in the full sample
distributions so that LGRB distributions can be understood and compared
to those for other types of stellar explosions. After a careful
consideration of uncertainties, I will present the results I obtain and
their implications for the progenitors of LGRBs.
- Slides [.pdf]
-
|
Yang Chen & Xufei Wang
10 Nov 2015
1:07pm EST
SciCen 706 |
- Calibration Concordance
- Follow-up to Xiao-Li Meng's talk on Oct 6
- Outline:
Part 1: Explanation of Multiplicative Model
Part 2: log-Normal Model (shrinkage estimator, variance estimator, additive model)
Part 3: Poisson Model (simplified version)
Part 4: Question for discussion
- Presentation Slides [.pdf]
-
|
Hyungsuk Tak
24 Nov 2015
1:07pm EST
SciCen 706 |
- Down-Up Metropolis-Hastings Algorithm for Multimodality
- Abstract:
I suggest a down-up Metropolis-Hastings (DUMH) algorithm that
expedites Markov chain's jumps between modes of a multi-modal
distribution in a simple and fast manner. This algorithm is
essentially a Metropolis-Hastings (MH) algorithm that generates a
proposal via two steps, a downhill step and an uphill step. Given
the current state, an intermediate proposal is generated by a
Metropolis algorithm with a reciprocal ratio of the target
densities in the acceptance probability so that this Metropolis
algorithm prefers a downward movement in density. For example, if
the density of the intermediate proposal is smaller than that of
the current state, then the intermediate proposal is accepted
with a probability one. Given the intermediate proposal, a final
proposal is generated by another Metropolis algorithm with a
typical acceptance probability that prefers an upward movement in
density. This down-up movement in density increases the chance
that the final proposal is at a different mode. The DUMH
algorithm accepts the final proposal with an MH acceptance
probability. Because this MH acceptance probability involves a
ratio of intractable integrations, the DUMH algorithm uses an
auxiliary variable to cancel out the intractable ratio.
Simulation results show that the DUMH algorithm explores some
high-dimensional and multi-modal distributions more effectively
than a random-walk Metropolis algorithm.
- Slides [.pdf]
-
|
Kaisey Mandel (CfA)
26 Jan 2016
1:07pm EST
SciCen 706 |
- Supernova Cosmology in the Near-Infrared with Hierarchical Bayesian
Light Curve Models
- Abstract:
The Nobel Prize-winning discovery of the accelerating expansion of the
Universe was made by astronomers using optical observations of the
brightness time series of faraway exploding stars (Type Ia supernova
light curves) to determine cosmological distances. Current and future
optical supernova surveys aim to determine the physical nature of the
mysterious dark energy driving the acceleration. However, these
efforts are now limited by systematic, rather than statistical errors.
The fortuitous properties of supernova light curves in the
near-infrared (NIR) offer a powerful strategy for improving distance
estimates and cosmological constraints. I have constructed a
hierarchical Bayesian model for optical and NIR supernova light curves
incorporating multiple random effects and uncertainties, including
host galaxy dust, measurement error, and intrinsic supernova
variations across time and wavelength, to determine precise and
accurate supernova distances. I will describe ongoing efforts to
trace the history of cosmic expansion by applying this statistical
model to analyze new, large datasets of ground-based NIR observations
of nearby supernovae, as well as recent Hubble Space Telescope NIR
observations of cosmologically distant supernovae discovered by the
Pan-STARRS and Dark Energy Survey supernova searches.
- Presentation slides [.pdf]
-
|
Xiyun Jiao (Imperial)
2 Feb 2016
6:07pm GMT
10:07am PST/1:07pm EST
Imperial |
- Next-generation Gibbs-type Samplers: Combining Strategies to Boost Efficiency
- Abstract: Although the Data Augmentation (DA)
algorithm and the Gibbs sampler are widely used tools for
obtaining a sample from the posterior distribution under complex
Bayesian models, they are sometimes notoriously slow to converge.
Thus, a number of strategies have been proposed to improve their
convergence properties. Among them, we focus on three methods:
the Marginal Data Augmentation (MDA), Ancillarity-Sufficiency
Interweaving Strategy (ASIS), and Partially Collapsed Gibbs (PCG)
sampling. To further improve the convergence of Gibbs-type
samplers, we propose strategies for combining MDA, ASIS, PCG
sometimes in conjunction with Metropolis-Hastings type updates.
We construct a general framework to combine these methods into
coherent samplers and guarantee that the combined samplers
maintain their target stationary distributions and can only
improve the convergence properties of their parent Gibbs-type
samplers. We demonstrate the implementation and efficiency of our
framework by implementing it to fit a factor analysis model and a
cosmological hierarchical model, which is the motivation of our
work.
- Presentation slides [.pdf]
-
|
Wang Xufei and Chen Yang (Harvard)
16 Feb 2016
1:07pm EST
SciCen 706 |
- Progress report of 'Calibration with Multiplicative Means but Additive Errors: A Log Normal Approach'
- Abstract:
Useful information to calibrate instruments used for
astrophysical measurements is usually obtained by observing
different sources with well-understood characteristics
simultaneously with different detectors. This requires a
careful modeling of the mean signals, the intrinsic source
variations, and measurement errors. Because our data are
typically large (>>30) photon counts, we propose an
approximate log-normal model, with the advantage of permitting
imperfection in the multiplicative mean model to be captured
by the residual variance. The calibration takes an
analytically tractable form of power shrinkage, with a
half-variance adjustment to ensure an unbiased multiplicative
mean model on the original scale.
In this presentation, we will talk about the current progress
regarding the properties and model fittings (frequentist and
Bayesian methods) of the log-normal regression model. We will
also discuss the potential issues and pitfalls of fitting a
Poisson regression model, which is more realistic but more
restrictive. Simulation studies and real data fittings will be
shown.
- Presentation slides [.pdf]
-
|
Group
1 Mar 2016
9:00am IST
29 Feb 7:30pm PST / 10:30pm EST
IUCAA |
- IACHEC Calibration Uncertainties Working Group Meeting
- Preliminary agenda:
- Vinay Kashyap, Introduction to pyBLoCXS
Slides [.pdf]
- Konrad Dennerl, Modeling RMFs
- Herman Marshall, Introduction to Calibration Concordance Project
- Keith Arnaud, Updates to XSPEC
- Xiao-Li Meng/Yang Chen/Xufei Wang, Calibration Concordance
-
|
Shijing Si (Imperial)
22 Mar 2016
6:07pm GMT
10:07am PST/1:07pm EST
Imperial |
- Bayesian Hierarchical Models for Stellar Evolution
- Abstract:
In astrophysics, we often aim to estimate a parameter for each of a number of
objects in a population. For example, we may want to estimate the age of each
of a sample of halo white dwarf (WD). The standard strategy is to separately
study each of the objects using case-by-case analyses, and then in a follow-up
analysis, study the distribution of the fitted parameters across the population. In
this research, we develop novel methods that allow us to take advantage of existing
software designed for such case-by-case analyses to simultaneously fit parameters
of both the individual objects and the parameters that quantify their distribution
across the population. Our methods are based on Bayesian hierarchical modelling
which is known to produce parameter estimators for the individual objects that
are on average closer to their true values than estimators based on case-by-case
analyses. We verify this in the context of estimating ages of Halo white dwarfs
via a series of simulation studies. We apply our techniques to three astrophysical problems.
The first one is essentially meta-analysis of the distance modulus to the large magenllanic
cloud (MLC) by combining some published results. The second project is to obtain more precise estimates of WD ages
by fitting their photometric data in a hierarchical model. The third problem is to study the initial-final
mass relationships (IFMR) via combining three stellar clusters.
- Presentation slides [.pdf]
-
|
Jeremy Drake (CfA)
19 Apr 2016
1:07pm EDT
SciCen 706 |
- Monte Carlo Methods for Treating and Understanding Highly-Correlated Instrument Calibration Uncertainties.
- Abstract:
Accounting for measurement uncertainty is a fundamental aspect of any
credible scientific experiment. In astrophysical observations, there
are two main sources of uncertainty: noise in the acquired data, and
the uncertainty in the calibration of the instrument used to obtain
them. Astronomers traditionally only consider the former and calibration
uncertainties are almost universally ignored. Modern X-ray
observatories, such as Chandra and XMM-Newton, frequently acquire data
for which calibration uncertainties are likely to be the dominant
source of error. However, calibration uncertainties are riddled with
complicated correlations that render them both technically challenging to
understand and to employ in data analysis. Here, we describe Monte
Carlo methods developed to include highly-correlated instrument
performance uncertainties in typical astrophysical model parameter
estimation studies. We will also describe how these methods can be used
in combination with observations of cosmic X-ray sources by one or
more observatories to refine the calibration uncertainties themselves.
- Slides [.pdf]
-
|
Sara Algeri (Imperial)
10 May 2016
6:07pm BST
10:07am PDT/1:07pm EDT
Imperial |
- Multiple hypothesis testing and testing one hypothesis multiple times: two sides of the same coin?
- Abstract:
In statistics, the problem of testing one hypothesis
multiple times can be formulated in terms of hypothesis
testing when a nuisance parameter is present only under the
alternative. Each possible value of the nuisance parameter
specifies a different alternative hypothesis and a unique
global p-value is provided to summarize the statistical
evidence in support (or against) the null hypothesis. From
a physics perspective, this scenario occurs quite often in
the searches for new signals over an energy or mass spectrum,
and in both the nested and non-nested frameworks.
An alternative way to search for new emissions is to refer to
the classical and widely known multiple hypothesis testing
approach. Separate tests of hypothesis are conducted at
different locations producing an ensemble of local p-values,
the smallest is reported as evidence for the new resonance,
once adequately adjusted to control the false detection rate
(type I error rate).
While multiple hypothesis testing procedures are both easy
and quick to implement, they may be overly conservative
in terms of the false detection rate. On the other hand,
testing one hypothesis multiple times methods are robust
with respect to false discovery rate and power, but may
require considerable computational effort when dealing with
complicated models.
The aim of this talk is to review both approaches, evaluate
their performance with respect to sample sizes, statistical
power, false detection rate and computational effort
required. Finally, a simple graphical tool is provided
to identify recurrent scenarios where a simple multiple
hypothesis testing procedure can be used to provide
valid inference with respect to stringent significance
requirements, without encountering the usual problem of
over-conservativeness.
- Presentation slides [pdf]
- arXiv:1701.06820 [url]
-
|
Nathan Stein (UPenn)
17 May 2016
1:07pm EDT
10:07am PDT/6:07pm BST
UPenn |
- Valid Statistical Comparisons Without Valid MCMC Output
- Abstract:
We consider model comparison when test statistics are
derived from the output of a Markov chain Monte Carlo
algorithm. We investigate this comparison from a classical
hypothesis testing perspective and find that the power of
such tests can exhibit surprising behavior. First, power is
not guaranteed to increase monotonically with the number of
MCMC iterations, so stopping the MCMC runs early can lead to
better performance. Second, power is not guaranteed to
increase when the MCMC chains are coupled in an effort to
reduce noise. We find that a promising direction for
developing more powerful tests relies on the fact that valid
Monte Carlo tests can be constructed from statistics that
preserve exchangeability but not independence. This problem
is motivated by the analysis of X-ray images of quasars, in
which the goal is to determine evidence for X-ray jets by
evaluating departures from a null model that includes a
quasar, modeled as a single point source, but no jet. In
this application, the test statistics are computed from the
output of an MCMC algorithm that samples from the posterior
distribution of a model that allows flexible, nonparametric
departures from the quasar-only null model.
-
|
Chunzhe Zhang (UC Davis)
24 May 2016
10:07am PDT
1:07pm EDT/6:07pm BST
Davis |
- Model Selection for Galaxy Shapes
- Abstract:
This project is about a novel image model selection procedure
for galaxy shapes. In recent astronomy literature, the method
used to perform model selection for galaxy images is based on
Chi^2 statistics, however, it is very often that different
kinds of fitted galaxy shape models have very similar or even
the same Chi^2 value. Our new procedure is based on BIC, but
with a novel initial parameter estimation method for
likelihood maximization. Extensive simulation studies of our
procedure have shown promising results.
- Presentation slides [.pdf]
-
|
Jeff Andrews (CfA)
28 Jun 2016
1:07pm EDT
10:07am PDT/6:07pm BST
SciCen 706 |
- Beyond Population Synthesis: Modeling X-ray binary populations
- Abstract:
Recent studies of high mass binaries have traditionally
employed population synthesis codes to link evolutionary
models with observed populations. However for certain types
of binaries, particularly those that are rare or
short-lived, population synthesis is a poor technique; only
a small fraction of simulated systems appear similar to
those observed. For more complex problems, the computational
expense can be prohibitive. Correlating high mass X-ray
binaries (HMXB) with regions of recent star-formation in
nearby galaxies is one such problem. I will demonstrate an
alternative to traditional population synthesis based on a
Markov-Chain Monte Carlo method which uses the spatially
resolved star formation history as a prior on the HMXBs'
birth location and age. We develop our model for HMXBs in
the Small Magellanic Clouds to quantitatively constraining
their evolutionary histories and map their likely birth
locations. With adaptations, this method can be applied to
populations of compact object binaries in other nearby
galaxies.
- Presentation slides [.pdf]
-
|
Saku Vrtilek (CfA)
12 Jul 2016
1:07pm EDT
10:07am PDT/6:07pm BST
SciCen 706 |
- A Hertszprung-Russell analog for X-ray binaries
- Abstract:
Color-color (CI) plots (which provide spectral information
over different energy ranges) and color-intensity (CI) plots
(which show brightness variations for a given color) are
common and easily-obtained measurements that have long been
used to classify accreting binary types. We have found that
when CC and CI diagrams are combined in a single
color-color-intensity (CCI) plot various types of X-ray
binaries separate into complex, but geometrically distinct
volumes.
CCI diagrams are in fact a three-dimensional analog for
accreting binaries of the classic Hertzprung-Russell diagram
which proved so fundamental to our understanding of single
star evolution. We wish extend the investigation of CCI
diagrams in several ways. First, to optimize the choice of
energy bands and test statistical methods to differentiate
among types of accreting binaries and the states of individual
objects. Second, to apply the technique study behavior of
accreting binaries in external galaxies. And finally to
identify the position in the CCI diagram with observable
parameters (temperatures of thermal components, spectral
indices of power law components, absorbing column and
inclination) and to more model dependent physical parameters
(accretion rate, disk inner radius, emitting area of neutron
star surface). Not only does our study address the decades old
question of how to easily and unambiguously separate black
hole from neutron systems, it will also provide important
clues about how the different classes of accreting binaries
are related to each other.
- Presentation slides [.pdf]
-
|
Irina Udaltsova (UC Davis) & Vasileios Stampoulis (Imperial)
26 July 2016
10:07am PDT/1:07 EDT
6:07pm BST
Davis/SciCen 706 |
- Bayesian Modeling of logN-logS
- Abstract:
[IU] The study of astrophysical source populations
is often conducted using the cumulative distribution of the
number of sources detected at a given sensitivity,
represented as a "log N - log S" curve. Direct estimation of
log N - log S relationship is complicated by unobserved
sources resulting from the detector-induced biases,
background contamination, and uncertainty on both the source
flux and the number of sources. Knowledge of the probability
of non-detection allows us to correct for the non-ignorable
missing data mechanism and to build a hierarchical Bayesian
model leads to inference for physical model parameters and
the log N - log S distribution. We present a procedure for
examining the Bayesian model goodness-of-fit and propose a
model selection approach of the number of source
populations.
[VS] We extend the logN-logS by incorporating
the spectral data from the observed sources in order to
account for the uncertainty in the count-to-flux conversion
factor gamma. We present the results of fitting the logN-logS
model for the Chandra Deep Field South (CDFS) dataset
for both the constant gamma case and the case with gamma uncertainty
and discuss how the posterior estimates are affected.
- Irina Udaltsova slides [.pdf]
- Vasileios Stampoulis slides [.pdf]
-
|
|
-
|