IACHEC 13, Apr , La Tenuta dei Ciclamini. CalStat WG Report. Vinay Kashyap (CXC/CfA)

Similar documents
MONTE CARLO METHODS FOR TREATING AND UNDERSTANDING HIGHLY-CORRELATED INSTRUMENT CALIBRATION UNCERTAINTIES

1 Statistics Aneta Siemiginowska a chapter for X-ray Astronomy Handbook October 2008

JEREMY DRAKE, PETE RATZLAFF, VINAY KASHYAP AND THE MC CALIBRATION UNCERTAINTIES TEAM MONTE CARLO CONSTRAINTS ON INSTRUMENT CALIBRATION

Analysis Challenges. Lynx Data: From Chandra to Lynx : 2017 Aug 8-10, Cambridge MA

Accounting for Calibration Uncertainty in Spectral Analysis. David A. van Dyk

Statistical Applications in the Astronomy Literature II Jogesh Babu. Center for Astrostatistics PennState University, USA

pyblocxs Bayesian Low-Counts X-ray Spectral Analysis in Sherpa

Accounting for Calibration Uncertainty in Spectral Analysis. David A. van Dyk

Spectral Analysis of High Resolution X-ray Binary Data

10 th IACHEC meeting: concluding remarks and agreements. Matteo Guainazzi (ASTRO-H ESAC SOC & JAXA SOT)

Contamination WG Summary

The High Energy Transmission Grating Spectrometer

Accounting for Calibration Uncertainty

XMM-Newton Chandra Blazar Flux Comparison. M.J.S. Smith (ESAC) & H. Marshall (MIT) 7 th IACHEC, Napa, March 2012

Neutron inverse kinetics via Gaussian Processes

Introduction to Bayesian Learning. Machine Learning Fall 2018

Challenges and Methods for Massive Astronomical Data

Working Group. Vadim Burwitz. for High Energy Calibration Apr , Frascati, Italy

X-ray spectral Analysis

Overlapping Astronomical Sources: Utilizing Spectral Information

Brandon C. Kelly (Harvard Smithsonian Center for Astrophysics)

Fully Bayesian Analysis of Calibration Uncertainty In High Energy Spectral Analysis

High Energy Astrophysics:

Machine learning comes from Bayesian decision theory in statistics. There we want to minimize the expected value of the loss function.

Fitting Narrow Emission Lines in X-ray Spectra

New Results of Fully Bayesian

Statistical Tools and Techniques for Solar Astronomers

CHAPTER 4 PROBABILITY AND COUNTING RULES UC DENVER

PoS(ICRC2017)765. Towards a 3D analysis in Cherenkov γ-ray astronomy

PILCO: A Model-Based and Data-Efficient Approach to Policy Search

Chandra Analysis of a Possible Cooling Core Galaxy Cluster at z = 1.03

Legacy IACHEC Working Group: the first steps

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008

CPSC 540: Machine Learning

Deep learning with differential Gaussian process flows

Data Analysis I. Dr Martin Hendry, Dept of Physics and Astronomy University of Glasgow, UK. 10 lectures, beginning October 2006

Fully Bayesian Analysis of Low-Count Astronomical Images

Statistics and Data Analysis

Statistical Methods for Astronomy

Can you predict the future..?

Statistical techniques for data analysis in Cosmology

XSPEC. Keith Arnaud. Center for Research Excellence in Space Science and Technology

Hierarchical Bayesian Modeling

Error Reporting Recommendations: A Report of the Standards and Criteria Committee

Hypothesis testing:power, test statistic CMS:

Empirical Bayes Moderation of Asymptotically Linear Parameters

Practical Statistics

EVIDENCE OF INTERMEDIATE-SCALE ENERGY SPECTRUM ANISOTROPY IN THE NORTHERN HEMISPHERE FROM TELESCOPE ARRAY

Effective Areas. Herman L. Marshall Nov. 17, 2014

X-ray data analysis. Andrea Marinucci. Università degli Studi Roma Tre

arxiv: v1 [stat.me] 24 May 2010

Chapter 4 Dynamic Bayesian Networks Fall Jin Gu, Michael Zhang

Organization. I MCMC discussion. I project talks. I Lecture.

NuSTAR observation of the Arches cluster: X-ray spectrum extraction from a 2D image

Improving the Relative Accuracy of the HETGS Effective Area

COMP 551 Applied Machine Learning Lecture 20: Gaussian processes

Parameter Estimation. William H. Jefferys University of Texas at Austin Parameter Estimation 7/26/05 1

A short introduction to INLA and R-INLA

Contents. Part I: Fundamentals of Bayesian Inference 1

Pattern Recognition and Machine Learning. Bishop Chapter 2: Probability Distributions

Nonparametric Bayesian Methods (Gaussian Processes)

DD Advanced Machine Learning

XMM-Newton Calibration

Hierarchical Bayesian Modeling

Statistics. Lecture 2 August 7, 2000 Frank Porter Caltech. The Fundamentals; Point Estimation. Maximum Likelihood, Least Squares and All That

Big Data Inference. Combining Hierarchical Bayes and Machine Learning to Improve Photometric Redshifts. Josh Speagle 1,

Connections between score matching, contrastive divergence, and pseudolikelihood for continuous-valued variables. Revised submission to IEEE TNN

Hierarchical Bayesian Modeling of Planet Populations

Lecture : Probabilistic Machine Learning

Embedding Astronomical Computer Models into Principled Statistical Analyses. David A. van Dyk

Statistical Methods for Astronomy

ADVANCED MACHINE LEARNING ADVANCED MACHINE LEARNING. Non-linear regression techniques Part - II

PART I INTRODUCTION The meaning of probability Basic definitions for frequentist statistics and Bayesian inference Bayesian inference Combinatorics

How to build an automatic statistician

Statistics 360/601 Modern Bayesian Theory

Bayesian Estimation of log N log S

Calibration Activities the MOS perspective

Recurrent Latent Variable Networks for Session-Based Recommendation

Computational Challenges from Imaging X-ray Polarimetry. Herman L. Marshall (MIT) and the IXPE Team

Expectation Propagation for Approximate Bayesian Inference

BERTINORO 2 (JVW) Yet more probability Bayes' Theorem* Monte Carlo! *The Reverend Thomas Bayes

Bayesian Deep Learning

MASSACHUSETTS INSTITUTE OF TECHNOLOGY PHYSICS DEPARTMENT

BAYESIAN CLASSIFICATION OF HIGH DIMENSIONAL DATA WITH GAUSSIAN PROCESS USING DIFFERENT KERNELS

Generative Learning algorithms

Observational Astrophysics 2

STAT 518 Intro Student Presentation

Advanced Introduction to Machine Learning CMU-10715

Bayesian Methods for Machine Learning

Introduction to Gaussian Processes

STA414/2104 Statistical Methods for Machine Learning II

Gaussian Processes. Le Song. Machine Learning II: Advanced Topics CSE 8803ML, Spring 2012

XMM-Newton SOC Technical Note

Target Tracking and Classification using Collaborative Sensor Networks

An introduction to Bayesian statistics and model calibration and a host of related topics

Calculating K corrections using S Lang and Sherpa

15-388/688 - Practical Data Science: Basic probability. J. Zico Kolter Carnegie Mellon University Spring 2018

Local Polynomial Modelling and Its Applications

Midterm: CS 6375 Spring 2015 Solutions

Setting UBVRI Photometric Zero-Points Using Sloan Digital Sky Survey ugriz Magnitudes

Transcription:

IACHEC 13, Apr 9-12 2018, La Tenuta dei Ciclamini CalStat WG Report Vinay Kashyap (CXC/CfA)

CalStats Working Group A forum for the discussion of statistical, methodological, and algorithmic issues that affect the calibration of astronomical instruments and how they are used in data analysis and interpretation of analysis results.

CalStats Working Group @IACHEC13 III WG Session 1, Tutorial on Gaussian Processes, David Jones VII (plenary) Updates to Concordance, Herman Marshall IX (plenary) Optimal binning and fit statistics for X-ray spectra, Jelle Kaastra XI WG Session 2, Binning and cstat (contd.)

CalStats Working Group @IACHEC13 III WG Session 1, Tutorial on Gaussian Processes, David Jones VII (plenary) Updates to Concordance, Herman Marshall IX (plenary) Optimal binning and fit statistics for X-ray spectra, Jelle Kaastra XI WG Session 2, Binning and cstat (contd.) If you are not on the CalStat WG mailing list (iachec-calstat@cfa.harvard.edu),

CalStats Working Group @IACHEC13 III WG Session 1, Tutorial on Gaussian Processes, David Jones VII (plenary) Updates to Concordance, Herman Marshall IX (plenary) Optimal binning and fit statistics for X-ray spectra, Jelle Kaastra XI WG Session 2, Binning and cstat (contd.) If you are not on the CalStat WG mailing list (iachec-calstat@cfa.harvard.edu), and would like to be,

CalStats Working Group @IACHEC13 III WG Session 1, Tutorial on Gaussian Processes, David Jones VII (plenary) Updates to Concordance, Herman Marshall IX (plenary) Optimal binning and fit statistics for X-ray spectra, Jelle Kaastra XI WG Session 2, Binning and cstat (contd.) If you are not on the CalStat WG mailing list (iachec-calstat@cfa.harvard.edu), and would like to be, send me (vkashyap@cfa.harvard.edu) an email.

I would have definitely attended your session but I was too busy having fun.

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu I. Tutorial on Gaussian Process New methodologies are coming up rapidly in astronomical analysis: Hierarchical Bayes, Nested Sampling, Deep Learning, Gaussian Processes, etc. Gaussian Process is probably the most immediately useful technique for calibration problems, as it allows a semiparametric, data-driven accounting of fudge factors that are impossible to get away from in real world calibration A tutorial was conducted on Apr 9 by statistician David Jones (Duke/SAMSI and TAMU)

David Jones Mention of Gaussian Process'' in SAO/NASA ADS Abstract 70 60 50 40 30 20 10 0 01/82 12/86 01/87 12/91 01/92 12/96 01/97 12/01 01/02 12/06 01/07 12/11 01/12 12/16 I

David Jones 2 y = f(x) 0 2 10 5 0 5 10 x

~y ~y David Jones 2 ~y ~y N ~ ~ =,! ~y ~y N ~y, y = f(x) 0 2 10 5 0 5 10 x

10 5 0 5 10 x David Jones 2 y 0 2

10 5 0 5 10 x I y,...,y m x,...,x m ~y N ~, + t I m David Jones 2 I ij = R (x i,x j ) ~y N ~f,t Im ~ f N ~, y 0 ~ f =(f (x ),...,f (xm )) T 2

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu I. Tutorial on Gaussian Process Some resources for further learning Gaussian Processes for Machine Learning, Rasmussen and Williams 2006, MIT Press http://www.gaussianprocess.org/gpml/chapters/ The Kernel Cookbook, David Duvenaud 2014 http://www.cs.toronto.edu/~duvenaud/cookbook/index.html Beyond Chi-Squared: An Introduction to Correlated Noise, Ben Montet 2014 https://astrobites.org/2014/07/01/beyond-chi-squared-an-introduction-to-correlated-noise/ Fitting Gaussian Process Models in Python, Chris Fonnesbeck 2017 https://blog.dominodatalab.com/fitting-gaussian-process-models-python/ Contact David Jones if you want assistance with a project

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu II. Update on Concordance On Apr 10, Herman Marshall (MIT) discussed the status of the Concordance project and discussed the process from an IACHEC perspective A paper describing the full statistical foundation of the method has been submitted to a statistics journal and is on the verge of resubmission after dealing with the referees reports Calibration Concordance for Astronomical Instruments via Multiplicative Shrinkage, Chen, Meng, Wang, van Dyk, Marshall, & Kashyap 2018, submitted to Journal of Applied Statistical Methods https://arxiv.org/abs/1711.09429 A trial run describing astronomically relevant analysis procedures will be presented at SPIE 2018 (Astronomical Instrumentation and Operations), with follow-on submission to JATIS An astronomy oriented paper which includes variable prior uncertainty factors and applications to other datasets (e.g., strong lines from Capella in Chandra gratings) is in preparation (Marshall et al. 201X)

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu II. Update on Concordance To complete analysis, we need: τ a measure of the prior uncertainty for detectors over different energy bands DONE datasets compilations of homogeneously estimated fluxes for simultaneously (or in some cases contemporaneously) observed sources. Perhaps easier to send Herman spectra and corresponding ARFs and RMFs (via Heritage WG).

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu III. A Practical Guide On Apr 11, Jelle Kaastra (SRON) gave a plenary talk on recent advances he has made 1.on optimal binning of RMFs a prescription for how to determine the most informative binning scheme given instrumental features (LSF, effective area) and counts in the spectrum, see Kaastra & Bleeker 2016 (A&A 587, A151) 2.on computing goodness of fit measures when using the cstat fit statistic calculating the expected contribution of a nominal model intensity in a bin to cstat (and its variance), averaged over all possible counts, and presenting it as a lookup table for ease of computation, see Kaastra 2017 (A&A 605, A51) and followed it up with a detailed discussion during the afternoon WG session. Keith Arnaud (GSFC) reports from afar: the optimal binning scheme is implemented in the new tool ftgrouppha (there s a bug for square RMFs) and an unreleased tool that bins the RMF along the energy-axis.

Monte Carlo results Jelle Kaastra (Data binning) Order 2 Order 1 Order 0 ( SPEX recommended ) ( classical approach)

Jelle Kaastra

Jelle Kaastra

Jelle Kaastra

Jelle Kaastra

Jelle Kaastra

Jelle Kaastra Estimating the contribution to cstat from each bin, given the expected model intensity in the bin; tabulated as a look-up table and interpolated therefrom 41

iachec-calstat@cfa.harvard.edu vkashyap@cfa.harvard.edu III. A Practical Guide A recommendation was made to write a practical guide to best practices in X-ray analysis [I will set up an overleaf document and ask CalStat WG members to contribute sections.] Wstat is an alternative statistic in XSPEC that computes maximum likelihood source and background intensities assuming Poisson regime and using observed counts in source and background. This can lead to bad fits when background counts fluctuate and/or model intensities are unstable. Jelle recommends that spectral fitting always be done by simultaneous fitting of background and source models, or computing a smooth background and adding it to the source. Apropos to that, would be a good idea for all missions to describe how to set up a model background spectrum Asked of Jelle that power of cstat parameterization be computed (how often is a bad fit accepted as good?); Monte Carlo uncertainties for cstat goodness-of-fit be calculated to show fluctuations in the asymptotic limits are as expected; simulations be carried out to see how a continuumdominated vs line-dominated spectrum affect goodness of fit estimates. Some questions require consultation with statisticians. E.g., what effect does determining binning from the data have on the process of the fit? Can we compute χ² and determine goodness from that? How robust is the K-S test (or equivalent) to measure goodness?

Next year s theme High resolution spectra