PaperlayerView exact revision on arXiv

Source-linked AI summary

Bayes in the sky: Bayesian inference and model selection in cosmology

Roberto Trotta

arXiv:0803.4089v1astro-ph

TL;DR

Cosmology needs principled methods for inference and model comparison as data, models, and uncertainties become increasingly complex. This review develops Bayesian probability, parameter inference, numerical methods, and model comparison, concluding that Bayesian tools provide a consistent framework while highlighting important prior and methodological caveats.

  • Problem

    Cosmological analyses must handle growing data and model complexity, uncertainty, and questions ranging from parameter inference to model comparison.

  • Method

    The review synthesizes Bayesian probability theory, priors, parameter inference, MCMC, Bayesian evidence, model complexity, and cosmological applications.

  • Results

    Bayesian probability provides a consistent framework for uncertainty in parameter inference, model comparison, prediction, and optimization.

  • Takeaways & Limitations

    Bayesian methods offer a broad basis for updating knowledge and incorporating measurement, systematic, nuisance-parameter, and modelling uncertainty.

  • Takeaways & Limitations

    Bayesian model-selection results can depend strongly on prior volume, while information criteria are meaningful only under assumptions not always met in cosmology and astrophysics.

Abstract

from arXiv · show

The application of Bayesian methods in cosmology and astrophysics has flourished over the past decade, spurred by data sets of increasing size and complexity. In many respects, Bayesian methods have proven to be vastly superior to more traditional statistical tools, offering the advantage of higher efficiency and of a consistent conceptual basis for dealing with the problem of induction in the presence of uncertainty. This trend is likely to continue in the future, when the way we collect, manipulate and analyse observations and compare them with theoretical models will assume an even more central role in cosmology. This review is an introduction to Bayesian methods in cosmology and astrophysics and recent results in the field. I first present Bayesian probability theory and its conceptual underpinnings, Bayes' Theorem and the role of priors. I discuss the problem of parameter inference and its general solution, along with numerical techniques such as Monte Carlo Markov Chain methods. I then review the theory and application of Bayesian model comparison, discussing the notions of Bayesian evidence and effective model complexity, and how to compute and interpret those quantities. Recent developments in cosmological parameter extraction and Bayesian cosmological model building are summarized, highlighting the challenges that lie ahead.

1 Introduction

Bayesian methods have become increasingly important in cosmology as growing computational power and increasingly complex data and models make sophisticated inference feasible and necessary. This review introduces Bayesian methodology and its applications across cosmological analysis.

  • Motivation: Faster, cheaper computing made previously unsolvable Bayesian inference problems tractable and enabled extensive numerical simulation.The paper links this development to the exponential increase in computational power over recent decades.
  • Motivation: Cosmology adopted Bayesian methods mainly in response to the data explosion of the last decade and increasingly computationally intensive inference problems.
  • Motivation: More complex theories and observations require more refined statistical and data-analysis skills, constraining the scientific return of next-generation surveys.
  • Scope: The review presents Bayesian probability theory, parameter inference, model comparison, and illustrative cosmological applications for students and practitioners.It is intended to bridge textbook examples and current research while providing broader guidance to the literature.

2 Bayesian probability theory

Bayesian probability treats probability as degree of belief, allowing uncertainty to be addressed in repeatable and one-off situations. Bayes’ theorem updates prior beliefs using data, while prior specification remains important—especially for model comparison.

  • Probability concepts: Frequentist probability is criticized as circular, limited for unrepeatable situations, and exact only for infinitely repeated trials.These limitations motivate a broader treatment of uncertainty in cosmological settings with one observable Universe.
  • Probability concepts: Bayesian probability measures degree of belief and applies to both repeated experiments and one-off events, while incorporating uncertainty from noise, systematics, and imperfect knowledge.
  • Bayes’ theorem: Bayes’ theorem computes posterior belief by combining the likelihood of observed data under a hypothesis with its prior probability.The posterior represents belief after considering the data, and the normalizing quantity is central to model comparison.
  • Priors: Prior choice is a feature of Bayesian analysis, but its impact is substantially stronger in model selection because prior volume affects the penalty for complex models.Results should therefore be checked for stability under physically reasonable prior changes.
  • Priors: Different prior beliefs can converge to a common posterior when the likelihood is informative and has support where the likelihood is large.As more observations accumulate, the data progressively override the initial prior means.

3 Bayesian parameter inference

Bayesian parameter inference combines a model, priors, and a data-generating likelihood to obtain posterior distributions. Because analytical solutions are uncommon, cosmological inference relies heavily on numerical methods such as MCMC, whose exploration must be checked carefully.

  • Inference setup: Parameter inference begins with a model containing parameters, priors summarizing prior knowledge, and a likelihood representing how the data are obtained.Likelihoods can encode measurement noise and nuisance parameters such as unknown variance or background rates.
  • Posterior inference: The posterior is evaluated for parameters of interest by marginalizing over nuisance parameters, while Bayesian evidence is irrelevant for parameter inference but central to model comparison.
  • Posterior inference: Posterior results can be communicated through summary statistics or one- and two-dimensional marginalized distributions, especially for multimodal or heavy-tailed posteriors.
  • Numerical methods: Analytical solutions are rare in realistic cosmological and astrophysical problems, so numerical simulation—particularly Markov Chain Monte Carlo—is generally required.
  • Numerical methods: MCMC constructs a parameter-space chain whose point density is proportional to the posterior probability density.The resulting samples support Monte Carlo estimates of posterior quantities and marginalized distributions.
  • Practical cautions: Poor MCMC exploration can produce serious inference errors in high-dimensional or multimodal posteriors, so coverage of relevant parameter space requires careful assessment.

4 Bayesian model comparison

Bayesian model comparison evaluates competing models by balancing fit against complexity, with the evidence and Bayes factors quantifying this trade-off. Results depend on data informativeness, prior ranges, and the computational method used to evaluate the evidence.

  • 4.1 Shaving theories with Occam’s razor: Bayesian model comparison evaluates relative model probabilities using observed data and prior information, preferring added complexity only when supported by improved fit.This formalizes Occam’s razor through a comparison of predictive adequacy and model complexity.
  • 4.2 The Bayesian evidence: Bayesian evidence is the marginal likelihood integral that evaluates model performance, while the Bayes factor measures how data update relative odds between models.B01 > 1 increases support for model 0 over model 1; B01 < 1 decreases it.
  • 4.2 The Bayesian evidence: For λ ≫1, B01 ≪1 favors the more complex model, whereas for λ ≲1 and σ/Σ ≪1, B01 ≈Σ/σ favors the simpler model through the Occam penalty.When σ/Σ ≫1, the data are less informative than the prior and B01 →1.
  • 4.3 Computation and interpretation of the evidence: Evidence computation is numerically challenging because likelihoods may be sharply peaked, heavy-tailed, multimodal, or strongly degenerate across a multidimensional parameter space.Thermodynamic integration can require up to 10^7 likelihood evaluations, about two orders of magnitude more than MCMC-based parameter estimation.
  • 4.3 Computation and interpretation of the evidence: A more complex model gains evidence only when its fit improvement offsets the Occam penalty associated with unused parameter-space volume.The evidence combines best-fit likelihood, a volume factor, and a term suppressing models whose likelihood-maximizing parameters differ from posterior expectations.
  • 4.3 Computation and interpretation of the evidence: Prior width shifts model-comparison outcomes: overly restrictive priors can make results inconclusive or prior-dominated, while overly wide priors can unduly favor simpler models.For fixed prior width, better data move results toward greater information and larger significance when the more complex model is true.

5 Cosmological parameter inference

Cosmological parameter inference has become more demanding as observations have grown dramatically in size and precision. This section introduces the inference problem and the tools developed to address it.

  • Cosmological data sets have increased rapidly in size, from ∼10^3 COBE CMB pixels to ∼10^6 WMAP pixels and projected ∼10^7 Planck pixels.
  • Larger and more precise data sets have driven the development of scalable map-making, component-separation, and parameter-inference methods.
  • The section surveys cosmological parameter inference and highlights innovative tools for tackling this broad problem.

5.1 The “vanilla” ΛCDM cosmological model

The vanilla ΛCDM model describes an expanding Universe containing a cosmological constant and cold dark matter. Its framework connects geometry, expansion, matter–energy content, and primordial fluctuations to observations.

  • Vanilla ΛCDM is a relatively simple cosmological scenario containing a cosmological constant and cold dark matter.
  • The model describes an expanding Universe whose geometry and scale factor determine the relation between redshift and comoving distance.
  • Its standard parameters include curvature, photons, neutrinos, baryons, cold dark matter, and the cosmological constant.
  • Inflation provides the accepted framework for generating primordial density fluctuations, stretching quantum fluctuations to cosmological scales.
  • Alternative gravity theories and other model formulations require model-comparison techniques to determine agreement with observations.

5.2 Cosmological observations

Cosmological observations probe different epochs and aspects of the Universe, including primordial radiation, galaxy clustering, supernova distances, and gravitational lensing. Their complementary information supports inference about cosmological parameters and structure formation.

  • Cosmic microwave background (CMB): CMB anisotropies provide a precise snapshot of the Universe at recombination, with temperature fluctuations of roughly ∆T/T ∼10^-5.
  • Cosmic microwave background (CMB): CMB temperature power-spectrum measurements combine data from WMAP 3-year, Boomerang 2003, and ACBAR with a best-fit ΛCDM model.
  • Large scale structures (LSS): Galaxy correlation measurements estimate the underlying dark-matter distribution, while their power spectrum depends on radiation-to-matter density, spectral index, and normalization.
  • Supernovae: Type Ia supernova observations extending to z ∼1.4 provided the first evidence in 1998 that the Universe’s expansion is accelerating.
  • Large scale structures (LSS): SDSS matter-power-spectrum data agree well with the best-fit CMB model even though those data were not used in fitting it.
  • Weak gravitational lensing: Weak gravitational lensing measures the small shape distortions induced in background galaxies by matter inhomogeneities along the line of sight.

5.3 Constraining cosmological parameters

Cosmological inference is specified by a model, parameter priors, and likelihood, then implemented through combined observations and posterior computation. Modern analyses use MCMC and find that a six-parameter ΛCDM description is currently sufficient for most available data.

  • The inference problem requires specifying which parameters vary, their prior distribution, and the likelihood for the considered data sets.
  • Cosmological parameters cover background dynamics, fluctuation initial conditions, nuisance quantities, and possible new-physics effects.
  • Independent-observation log-likelihoods add, and combining data sets can strengthen constraints by mutually breaking parameter degeneracies.
  • Hyperparameters can reweight discrepant data sets, with non-informative priors and Bayesian integration producing an effective chi-square.
  • Correlation studies are an exception to simply adding independent-observation likelihoods because they intentionally exploit correlations among observables.
  • Posterior probability densities are usually obtained numerically with MCMC, replacing earlier reliance on maximum-likelihood searches or parameter grids.
  • The six-parameter ΛCDM model appears sufficient for most presently available data, while extra parameters are better treated through model comparison.
  • Neural networks can reduce computational effort by learning likelihood evaluations from a user-provided training set and interpolating new points.

5.4 Caveats and common pitfalls

Bayesian inference is not a black-box procedure: prior specification, parameterization, and model-selection practice can materially affect cosmological conclusions. These choices require explicit scrutiny because apparent data constraints or significances may instead reflect methodological decisions.

  • Automated Bayesian inference can hide problem-specific pitfalls, so each real-world analysis requires careful statistical consideration.The review emphasizes that inference remains both a craft and a science, despite increasingly automated procedures.
  • Flat priors are not generally uninformative: their informativeness depends on the parameter being estimated and can propagate to derived observables.A prior flat in σ differs from one flat in ln σ, and flat fundamental parameters need not produce flat observable distributions.
  • Nonlinear reparameterizations can make physically equivalent setups yield widely different inferences when no principled prior measure selects between them.The review identifies unknown-amplitude problems, including isocurvature modes, as examples where this dependence can matter.
  • Prior sensitivity is especially relevant to dark-energy reconstruction, initial power-spectrum reconstruction, and inflationary-potential parameter determination.
  • Assessing an observed deviation from a null parameter value is often a model-selection problem rather than merely a parameter-significance calculation.The review frames this issue as widespread across domains, including photon-count estimation with background.
  • Using data to guide model building or prior choice and then quantitatively test the same effect can drastically overestimate its significance.The review notes that this practice is difficult to avoid when observations reveal unexpected phenomena, such as large-scale cosmic microwave background anomalies.

6 Cosmological Bayesian model building

Bayesian evidence is used to assess extensions and alternatives to the ΛCDM concordance model, with conclusions shaped by prior volume and parameterization. Current comparisons generally support ΛCDM or remain undecided, while selected results constrain isocurvature and dark-energy alternatives.

  • Model-comparison framework: Bayesian evidence comparisons ask whether cosmological extensions or reductions of ΛCDM are supported by data, using prior volume to quantify the Occam penalty.The review emphasizes full evidence rather than information-criterion approximations because prior choices can control the comparison.
  • Evidence for ΛCDM: Most comparisons support ΛCDM or are undecided, so its extra parameters are generally not required by current data.A reported exception is support for a large-scale power-spectrum cut-off, driven by anomalies whose cosmological origin remains open.
  • Initial conditions: Isocurvature fractions are constrained below about 10% for one mode and below about 50% for general mixtures, favoring the purely adiabatic model.The strength of this preference depends strongly on the parameterization of the isocurvature sector through the Occam effect.
  • Primordial spectrum: A power-law primordial spectrum with ns < 1 is the current consensus, although inflationary model comparison depends on tensor-mode priors and a specified alternative.Flat-r and log-r priors can produce very different results, while the simpler ns = 1 and r = 0 model may lack a meaningful non-inflationary alternative.
  • Dark energy: The cosmological constant remains a sufficient description of dark energy, while the required effective accuracy for preferring it over alternatives ranges from σeff = 0.05 to σeff = 5 × 10−5.The thresholds depend on the class of alternative model: phantom, fluid-like, or small-departure.
  • Non-nested alternatives: For non-nested alternatives, full evidence requires priors for all model parameters, and current data do not require fundamental departures from the underlying theoretical model.The review cites examples including anisotropic templates, Lemaitre–Tolman–Bondi models, and fractal bubble scenarios.
  • Model averaging: Model-averaged dark-energy posteriors can become tighter around w = −1 than any evolving model alone because the preferred ΛCDM model receives substantial posterior weight.Model averaging marginalizes over model choice and can differ substantially from model-specific constraints unless one model overwhelmingly dominates.

7 Conclusions

The review presents Bayesian methods as a framework for uncertainty, inference, and model comparison in increasingly complex cosmological analyses. It highlights computational advances, clarifies the role of priors, and identifies continued challenges in applying the field’s full potential.

  • Bayesian probability provides a consistent framework for uncertainty across parameter inference, model comparison, prediction, and optimization.
  • Markov Chain Monte Carlo techniques are now standard tools for deriving parameter constraints across varied posterior distributions.
  • Priors are unavoidable assumptions whose influence should be quantified, although choosing a fair representation of belief can remain difficult.
  • Bayesian evidence and complexity quantitatively compare models, favoring simpler explanations when they adequately account for observations.
  • Multi-model inference combines model comparison with parameter inference by producing model-averaged parameter constraints.
  • As cosmological data sets and models become more complex, basic statistical analyses are increasingly inadequate and Bayesian methods are expected to grow in importance.
Loading 0803.4089v1…