Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Bayesian Spectral Line Fitting Dr. Michelle Lochner With Ian Harrison & Michael Brown https://arxiv.org/abs/1704.08278 The Problem with Spectra Radio spectra can be noisy, contaminated with RFI and containing very faint spectral lines SKA HI Galaxy Survey HI emission line is one of the few spectral features of radio galaxies HI is intrinsically faint High SNR requirement means very few high redshift galaxies can actually be detected with the SKA (Yahya et al. 2014) This limits the usefulness of HI galaxies for cosmology SKADS Simulations Continuum selected galaxies (not a blind HI data cube search) Obreschow et al. 2009 HI Line Profile Obreschow et al. 2009 Signal to Noise Ratio We use the SNR definition from Yahya et al. 2015 Line Fitting Approach We want to: Line Fitting Approach We want to: Be able to tell if there is a detectable HI line Line Fitting Approach We want to: Be able to tell if there is a detectable HI line Fit this with the HI line profile to estimate the parameters (including redshift) Line Fitting Approach We want to: Be able to tell if there is a detectable HI line Fit this with the HI line profile to estimate the parameters (including redshift) Get the full probability distribution for the parameters Bayesian Statistics to the Rescue! Introduction to Bayesian Statistics Bayes' theorem tells us: Introduction to Bayesian Statistics Bayes' theorem tells us: Posterior Likelihood Bayesian Evidence Prior Introduction to Bayesian Statistics Bayesian Inference: Hard to do for N dimensions Marginalisation requires N dimensional integrals Fortunately you can use numerical samplers like MCMC or Nested Sampling Example 1d marginalised posterior σ μ Model Selection with Bayesian Evidence Bayesian Evidence Likelihood Prior Model Selection with Bayesian Evidence Model Priors Bayes Factor Ratio of Posterior Odds Model Selection with Bayesian Evidence Bayes Factor Model Selection with Bayesian Evidence Jeffreys’ Scale Trotta 2008 Model Comparison vs. We use Bayesian model comparison to decide if the HI line profile (with 6 parameters) is a better fit than a completely flat line (consistent with pure noise) Results SNR ~ 11 Results SNR ~ 11 Results Black true, red maximum posterior fit P(z) Redshift estimates (band 1) Redshift estimates (band 1) B>6 evidence cut Redshift estimates (band 1) B>6 evidence cut Redshift estimates (band 2) ? B>6 evidence cut Multimodal P(z) P(z) N(z) Conclusions We’ve developed a promising, automated, general approach to spectral line fitting in the radio Allows more realistic number counts for forecasts We introduce statistical rigour into detection The Bayesian nature of the technique allows for correct uncertainty propagation and interesting extentions Go to https://arxiv.org/abs/1704.08278 to read the paper and get the code Email me at: [email protected]