[Paper Review] Fitting and Comparison of Models of Radio Spectra

Bojan Nikolic|ArXiv.org|Dec 11, 2009

Gaussian Processes and Bayesian Inference6 references3 citations

TL;DR

This paper presents a Bayesian spectral fitting framework using nested sampling to objectively compare radio spectrum models, enabling full posterior inference and evidence-based model selection. Applied to nearby galaxies, it identifies the continuous-injection with synchrotron self-absorption model as strongly preferred for NGC 7331 due to high evidence, while simpler models suffice for others, demonstrating improved robustness over traditional maximum-likelihood fitting.

ABSTRACT

I describe an approach to fitting and comparison of radio spectra based on Bayesian analysis and realised using a new implementation of the nested sampling algorithm. Such an approach improves on the commonly used maximum-likelihood fitting of radio spectra by allowing objective model selection, calculation of the full probability distributions of the model parameters and provides a natural mechanism for including information other than the measured spectra through priors. In this paper I cover the theoretical background, the algorithms used and the implementation details of the computer code. I also briefly illustrate the method with some previously published data for three near-by galaxies. In forthcoming papers we will present the results of applying this analysis larger data sets, including some new observations, and the physical conclusions that can be made. The computer code as well as the overall approach described here may also be useful for analysis of other multi-chromatic broad-band observations and possibly also photometric redshift estimation. All of the code is publicly available, licensed under the GNU General Public License, at http://www.mrao.cam.ac.uk/~bn204/galevol/speca/index.html

Motivation & Objective

• To develop a statistically rigorous, objective method for fitting and comparing radio spectrum models in the presence of sparse, noisy, and non-Gaussian data.
• To overcome limitations of traditional maximum-likelihood fitting, such as lack of objective model selection and incomplete error characterization.
• To provide full posterior distributions of model parameters, including non-Gaussian and correlated uncertainties.
• To integrate physical priors naturally into the analysis, improving parameter estimation with external constraints.
• To create a publicly available, general-purpose tool applicable to multi-wavelength and photometric redshift analyses.

Proposed method

• Uses Bayesian inference via Bayes' theorem to compute the posterior distribution of model parameters given data and hypotheses.
• Employs the nested sampling algorithm (Skilling, 2006) to efficiently compute the Bayesian evidence Z and full posterior distributions.
• Models radio spectra using analytic expressions for synchrotron, thermal, and self-absorbed emission, with customizable priors.
• Calculates likelihoods by comparing predicted flux densities (from model and parameters) to observed fluxes at various frequencies.
• Visualizes model performance using fan-diagrams, showing predicted spectra across the posterior distribution.
• Implements the method in a publicly available, open-source code licensed under the GNU GPL for broader application in radio and multi-band astronomy.

Experimental results

Research questions

• RQ1 Which model of radio spectrum—power-law, continuous injection, or self-absorbed—best explains the observed flux densities in nearby galaxies?
• RQ2 How can model selection be made objective and statistically rigorous when data are sparse and errors are non-Gaussian?
• RQ3 To what extent do physical priors improve parameter estimation and reduce degeneracies in spectral fitting?
• RQ4 Can the Bayesian evidence quantitatively favor complex models like continuous-injection with synchrotron self-absorption over simpler alternatives?
• RQ5 How do fan-diagrams provide an intuitive and informative visualization of model fit and parameter uncertainty?

Key findings

• For NGC 7331, the continuous-injection with synchrotron self-absorption (CI+SSA) model has a Bayesian evidence value several orders of magnitude higher than other models, indicating strong preference.
• The marginal posterior distributions for the CI+SSA model parameters in NGC 7331 show significant non-Gaussianity, especially for the spectral index and break frequency.
• For NGC 3627, the continuous-injection model without absorption is preferred, with no evidence for low-frequency absorption in the data.
• The power-law model is the best-fitting model for the first galaxy studied, with the highest Bayesian evidence among the four tested models.
• The absorbed models (e.g., with low-frequency turnover at 30 MHz) perform worse due to prior constraints limiting the range of possible turnovers.
• Fan-diagrams effectively visualize the range of predicted spectra across the posterior, revealing where models fail to reproduce data features.