Nested Sampling =============== Formalism ~~~~~~~~~ The starting point of Nested Sampling is Bayes' theorem, which allows us to evaluate the probability distribution of a set of parameters :math:`\vec{c}` associated to a model :math:`\mathcal{M}(\vec{c})` given a set of experimental measurements :math:`\mathcal{D}` , .. math:: P\left(\vec{c}| \mathcal{D},\mathcal{M} \right) = \frac{P\left(\mathcal{D}|\mathcal{M},\vec{c} \right) P\left( \vec{c}|\mathcal{M} \right) }{P(\mathcal{D}|\mathcal{M})} \, . Here :math:`P\left(\vec{c}| \mathcal{D},\mathcal{M} \right)` represents the posterior probability of the model parameters given the assumed model and the observed experimental data, :math:`P\left(\mathcal{D}|\mathcal{M},\vec{c}\right) = \mathcal{L}\left(\vec{c} \right)` is the likelihood (conditional probability) of the experimental measurements given the model and a specific choice of parameters, and :math:`P\left( \vec{c}|\mathcal{M} \right) = \pi \left( \vec{c} \right)` is the prior distribution for the model parameters. The denominator in the equation above, :math:`P(\mathcal{D}|\mathcal{M}) = \mathcal{Z}` , is known as the Bayesian evidence and ensures the normalisation of the posterior distribution, .. math:: \mathcal{Z} = \int \mathcal{L}\left( \vec{c} \right)\pi \left( \vec{c} \right) d \vec{c} \, , where the integration is carried out over the domain of the model parameters :math:`\vec{c}` . The key ingredient of Nested Sampling utilises the ideas underlying Bayesian inference to map the :math:`n` -dimensional integral over the prior density in model parameter space :math:`\pi(\vec{c} )d\vec{c}` , where :math:`n` represents the dimensionality of :math:`\vec{c}` , into a one-dimensional function of the form .. math:: X(\lambda) = \int_{\{ \vec{c} : \mathcal{L}\left(\vec{c} \right) > \lambda \}}\pi(\vec{c} ) d\vec{c} \,. In this expression, the prior mass :math:`X(\lambda)` corresponds to the (normalised) volume of the prior density :math:`\pi(\vec{c} )d\vec{c}` associated with values of the model parameters that lead to a likelihood :math:`\mathcal{L}\left(\vec{c}\right) ` greater than the parameter :math:`\lambda` . Note that by construction, the prior mass :math:`X` decreases monotonically from the limiting value :math:`X=1` to :math:`X=0` as :math:`\lambda` is increased. The integration of :math:`X(\lambda)` extends over the regions in the model parameter space contained within the fixed-likelihood contour defined by the condition :math:`\mathcal{L}\left(\vec{c}\right) =\lambda`. This property allows the evidence to be expressed as, .. math:: \mathcal{Z} = \int_0^1 \mathcal{L}\left( X\right) dX \, , where :math:`\mathcal{L}\left( X\right)` is defined as the inverse function of :math:`X(\lambda)` , which always exists provided the likelihood is a continuous and smooth function of the model parameters. Therefore, the transformation from :math:`\vec{c}` to X achieves a mapping of the prior distribution into infinitesimal elements, sorted by their associated likelihood :math:`\mathcal{L}(\vec{c})` . The next step of the NS algorithm is to define a decreasing sequence of values in the prior volume, that is now parameterised by the prior mass :math:`X` . In other words, one slices the prior volume into a large number of small regions .. math:: 1 = X_0 > X_1 > \ldots X_{\infty} = 0 \, , and then evaluates the likelihood at each of these values, :math:`\mathcal{L}=\mathcal{L}(X_i)` . This way, all of the :math:`\mathcal{L}_i` values can be summed in order to evaluate the integral for the Bayesian evidence. Since in general the likelihood :math:`\mathcal{L}({\boldsymbol c})` exhibits a complex dependence on the model parameters :math:`\vec{c}` , the summation above must be evaluated numerically using {\it e.g.} Monte Carlo integration methods. In practice, one draws :math:`N_{\rm live}` points from the parameter prior volume :math:`\pi\left(\vec{c} \right)` , known as {\it live points}, and orders the likelihood values from smallest to largest, including the starting value of the prior mass at :math:`X_0=1`. As samples are drawn from the prior volume, the live point with the lowest likelihood :math:`\mathcal{L}_i` is removed from the set and replaced by another live point drawn from the same prior distribution but now under the constraint that its likelihood is larger than :math:`\mathcal{L}_i`. This sampling process is repeated until the entire hyper-volume :math:`\pi \left( \vec{c} \right)` of the prior parameter space has been covered, with ellipsoids of constrained likelihood being assigned to the live-points as the prior volume is scanned. While the end result of the NS procedure is the estimation of the Bayesian evidence :math:`\mathcal{Z}` , as a byproduct one also obtains a sampling of the posterior distribution associated to the EFT coefficients expressed as .. math:: \{ \vec{c}^{(k)} \}\, ,\qquad k=1,\dots,N_{\rm spl}\, , with :math:`N_{\rm spl}` indicating the number of samples drawn by the final NS iteration. One can then compute expectation values and variances of the model parameters by evaluating the MC sum over the these posterior samples together with their associated weights, in the same way as averages are done over the :math:`N_{\rm rep}` replicas in the MCfit method. Prior volume ~~~~~~~~~~~~ An important input for NS is the choice of prior volume :math:`\pi \left( \vec{c} \right)` in the model parameter space. In this analysis, we adopt flat priors defined by ranges in parameter space for the coefficients :math:`\vec{c}` . A suitable choice of prior volume where the sampling takes place is important to speed up the NS algorithm: a range too wide will make the optimisation less efficient, while a range too narrow might bias the results by cutting specific regions of the parameter space that are relevant. Furthermore, using a common range for all parameters should be avoided, since the range of intrinsic variation will be rather different for each of the EFT coefficients. Taking these considerations into account, we adopt here the following strategy. First, a single model parameter :math:`c_i` is allowed to vary while all others are set to their SM value, :math:`c_j=0` for :math:`j\ne i` . The :math:`\chi^2 \left( c_i \right)` is then scanned in this direction to determine the values :math:`c_i^{\rm (min)}` and :math:`c_i^{\rm (max)}` satisfying the condition :math:`\chi^2/n_{\rm dat}=4` . We then repeat this procedure for all parameters and end up with a hyper-volume defined by pairs of values :math:`\left( c_i^{\rm (min)},c_i^{\rm (max)} \right)` with :math:`i=1,\ldots, n_{\rm op}` which defines our initial prior volume. At this point, one performs an initial exploratory NS global analysis using this volume to study the posterior probability distribution for each EFT coefficient. Our final analysis is then obtained by manually adjusting the initial sampling ranges until the full posterior distributions are captured for the chosen prior volume. For parameters that are essentially unconstrained in the global fit, such as the four-heavy operators in the case of linear EFT calculations, a hard boundary of :math:`\left( -50 , 50 \right)` (for :math:`\Lambda=1` TeV) is imposed. Performance ~~~~~~~~~~~ To increase the efficiency of the posterior probability estimation by NS, we enable the *constant efficiency mode* in MultiNest, which adjusts the total volume of ellipsoids spanning the live points so that the sampling efficiency is close to its associated hyperparameter set by the user. With 24 cpu cores, we are able to achieve an accurate posterior for the linear EFT fits in :math:`\sim 30` minutes using 500 live points, a target efficiency of 0.05, and an evidence tolerance of 0.5, which results in :math:`N_{\rm spl}\simeq 5000` posterior samples. To ensure the stability of our final results, we chose 1000 live points and a target efficiency of 0.005, which yields :math:`\simeq 1.5\times 10^4` samples for the linear analysis and :math:`\simeq 10^4` samples for an analysis that includes also the quadratic EFT corrections. With these settings, our final global analyses containing the simultaneous determination of :math:`n_{\rm op}\simeq 36` coefficients take :math:`\sim 3.5` hours running in 24 cpu cores, with a similar performance for linear and quadratic EFT fits.