Next Contents Previous

3. BAYESIAN PARAMETER INFERENCE

3.1. The general problem and its solution

The general problem of Bayesian parameter inference can be specified as follows. We first choose a model containing a set of hypotheses in the form of a vector of parameters, theta. The parameters might describe any aspect of the model, but usually they will represent some physically meaningful quantity, such as for example the mass of an extra-solar planet or the abundance of dark matter in the Universe. Together with the model we must specify the priors for the parameters. Priors should summarize our state of knowledge about the parameters before we consider the new data, and for the parameter inference step the prior for a new observation might be taken to be the posterior from a previous measurement (for model comparison issues the prior is better understood in a different way, see section 4). The caveats about priors and prior specifications presented in the previous section will apply at this stage.

The central step is to construct the likelihood function for the measurement, which usually reflects the way the data are obtained. For example, a measurement with Gaussian noise will be represented by a Normal distribution, while gamma-ray counts on a detector will have a Poisson distribution for a likelihood. Nuisance parameters related to the measurement process might be present in the likelihood, e.g. the variance of the Gaussian might be unknown or the background rate in the absence of the source might be subject to uncertainty. This is no matter of concern for a Bayesian, as the general strategy is always to work out the joint posterior for all of the parameters in the problem and then marginalize over the ones we are not interested in. Assuming that we have a set of physically interesting parameters ϕ and a set of nuisance parameters psi, the joint posterior for theta = (ϕ, psi) is obtained through Bayes' Theorem:

Equation 12 (12)

where we have made explicit the choice of a model M by writing it on the righ-hand-side of the conditioning symbol. Recall that L(theta ) ident p(d | theta, M) denotes the likelihood and p(theta| M) the prior. The normalizing constant p(d | M) ("the Bayesian evidence") is irrelevant for parameter inference (but central to model comparison, see section 4), so we can write the marginal posterior on the parameter of interest as (marginalizing over the nuisance parameters)

Equation 13 (13)

The final inference on ϕ from the posterior can then be communicated either by some summary statistics (such as the mean, the median or the mode of the distribution, its standard deviation and the correlation matrix among the components) or more usefully (especially for cases where the posterior presents multiple peaks or heavy tails) by plotting one or two dimensional subsets of ϕ, with the other components marginalized over.

In real life there are only a few cases of interest for which the above procedure can be carried out analytically. Quite often, however, the simple case of a Gaussian prior and a Gaussian likelihood can offer useful guidance regarding the behaviour of more complex problems. An analytical model of a Poisson-distributed likelihood for estimating source counts in the presence of a background signal is worked out in [14]. In general, however, actual problems in cosmology and astrophysics are not analytically tractable and one must resort to numerical techniques to evaluate the likelihood and to draw samples from the posterior. Fortunately this is not a major hurdle thanks to the recent increase of cheap computational power. In particular, numerical inference often employs a technique called Markov Chain Monte Carlo, which allows to map out numerically the posterior distribution of Eq. (12) even in the most complicated situations, where the likelihood can only be obtained by numerical simulation, the parameter space can have hundreds of dimensions and the posterior has multiple peaks and a complicated structure.

For further reading about Bayesian parameter inference, see [29, 10, 30, 11]. For more advanced applications to problems in astrophysics and cosmology, see [31, 16].

3.2. Markov Chain Monte Carlo techniques for parameter inference

The general solution to any inference problem has been outlined in the section above: it remains to find a way to evaluate the posterior of Eq. (12) for the usual case where analytical solutions do not exist or are insufficiently accurate. Nowadays, Bayesian inference heavily relies on numerical simulation, in particular in the form of Markov Chain Monte Carlo (MCMC) techniques, which are discussed in this section.

The purpose of the Markov chain Monte Carlo algorithm is to construct a sequence of points in parameter space (called "a chain"), whose density is proportional to the posterior pdf of Eq. (12). Developing a full theory of Markov chains is beyond the scope of the present article (see e.g. [32, 33] instead). For our purposes it suffices to say that a Markov chain is defined as a sequence of random variables {X(0), X(1), ..., X(M-1)} such that the probability of the (t + 1)-th element in the chain only depends on the value of the t-th element. The crucial property of Markov chains is that they can be shown to converge to a stationary state (i.e., which does not change with t) where successive elements of the chain are samples from the target distribution, in our case the posterior p(theta|d). The generation of the elements of the chain is probabilistic in nature, and several algorithms are available to construct Markov chains. The choice of algorithm is highly dependent on the characteristics of the problem at hand, and "tayloring" the MCMC to the posterior one wants to explore often takes a lot of effort. Popular and effective algorithms include the Metropolis-Hastings algorithm [34, 35], Gibbs sampling (see e.g. [36]), Hamiltonian Monte Carlo (see e.g. [37] and importance sampling.

Once a Markov chain has been constructed, obtaining Monte Carlo estimates of expectations for any function of the parameters becomes a trivial task. For example, the posterior mean is given by (where < ⋅ > denotes the expectation value with respect to the posterior)

Equation 14 (14)

where the equality with the mean of the samples from the MCMC follows because the samples theta(t) are generated from the posterior by construction. In general, one can easily obtain the expectation value of any function of the parameters f(theta) as

Equation 15 (15)

It is usually interesting to summarize the results of the inference by giving the 1-dimensional marginal probability for the j-th element of theta, thetaj. Taking without loss of generality j = 1 and a parameter space of dimensionality n, the equivalent expression to Eq. (3) for the case of continuous variables is

Equation 16 (16)

where p(theta1 | d) is the marginal posterior for the parameter theta1. From the Markov chain it is trivial to obtain and plot the marginal posterior on the left-hand-side of Eq. (16): since the elements of the Markov chains are samples from the full posterior, p(theta|d), their density reflects the value of the full posterior pdf. It is then sufficient to divide the range of theta1 in a series of bins and count the number of samples falling within each bin, simply ignoring the coordinates values theta2, ..., thetan. A 2-dimensional posterior is defined in an analogous fashion.

There are several important practical issues in working with MCMC methods (for details see e.g. [32]). Especially for high-dimensional parameter spaces with multi-modal posteriors it is important not to use MCMC techniques as a black box, since poor exploration of the posterior can lead to serious mistakes in the final inference if it remains undetected. Considerable care is required to ensure as much as possible that the MCMC exploration has covered the relevant parameter space.

Next Contents Previous