Fitting strategies

In this section we describe the fitting methodology that is used to map the EFT parameter space. SMEFiT implements two different stategies:

  • the first bases on Monte Carlo replica fitting MCfit method presented in [EMM+21, HMN+19],

  • the second one using the MultiNest and Nested Sampling algorithm [FHCP13, FH08], a robust sampling procedure that is completely orthogonal to the MCfit method and that is based on Bayesian inference.

A benchmark comparison for fits based on quadratic EFT calculations has been performed in [EMM+21] and demonstrates that results obtained with either NS or MCfit are statistically equivalent. The two methods are in excellent agreement, both in terms of best-fit values and of the corresponding uncertainties. This said, for specific coefficients one observes small differences, with MCfit in general tending to provide somewhat looser bounds. The reason for this behaviour is that optimisation-based methods such as MCfit can be distorted by fitting inefficiencies, such as when the optimiser finds a local, rather than global, minimum. MCfit distribution exhibits broader tails, implying that the bounds obtained this way might in some cases be slightly over-conservative.

NS is usually adopted as the baseline method, since its not affected by potential inefficiencies in the minimisation procedure and can produce global fits within a reasonable execution time.

In the following paragraphs we describe the log-likelihood function and the treatment of uncertainties, which are common for the two fitting methodology. We also discuss the case of indivdual fits where just one EFT operator is take into account.

\(\chi^2\) definition

The overall fit quality is quantified by the log-likelihood, or \(\chi^2\) function, defined as

\[\chi^2 \left ( {\textbf{c}} \right ) \equiv \frac{1}{n_{\rm dat}}\sum_{i,j=1}^{n_{\rm dat}} \left (\sigma^{(\rm th)}_i \left( {\textbf{c}} \right) -\sigma^{(\rm exp)}_i\right) ({\rm cov}^{-1})_{ij} \left ( \sigma^{(\rm th)}_j \left( {\textbf{c}}\right) -\sigma^{(\rm exp)}_j\right)\]

where \(\sigma_i^{\rm (exp)}\) and \(\sigma^{\rm (th)}_i \left(\textbf{c} \right )\) are the central experimental data and corresponding theoretical predictions for the i-th cross-section, respectively.

Uncertanties treatment

The total covariance matrix, \({\rm cov}_{ij}\), should contain all relevant sources of experimental and theoretical uncertainties. Assuming the latter are normally distributed, and that they are uncorrelated with the experimental uncertainties, this total covariance matrix can be expressed as a sum of the separate experimental and theoretical covariance

\[{\rm cov}_{ij} = {\rm cov}^{(\rm exp)}_{ij} + {\rm cov}^{(\rm th)}_{ij}\]

As usual, the experimental covariance matrix is constructed from all sources of statistical and systematic uncertainties that are made available by the experiments. Moreover, the correlated multiplicative uncertainties are treated via the t_0 prescription [B+10] in the fit, while the standard experimental definition is used to quote the resulting \(\chi^2\) values.

Concerning the theoretical covariance matrix, \({\rm cov}^{(\rm th)}\), its contributions depend on the specific type of processes considered.

Individual fits

Individual (one-parameter) fits correspond to varying a single EFT coefficient while keeping the rest fixed to their SM values. While such fits neglect the correlations between the different coefficients, they provide a useful baseline for the global analysis, since there the CL intervals will be by construction looser (or at best, similar) as compared to those of the one-parameters fits.

They are also computationally inexpensive, as they can be carried out analytically from a scan of the \(\chi^2\) profile without resorting to numerical methods.

Another benefit is that they facilitate the comparison between different EFT analyses, which may adopt different fitting bases but whose individual bounds should be similar provided they are based on comparable data sets and theoretical calculations.

In the scenario where a single EFT coefficient, \(c_j\), is allowed to vary while the rest are set to zero, the theoretical cross-section (for \(\Lambda=1\) TeV) given by simplifies to

\[\sigma_m^{\rm (th)}(c_j)= \sigma_m^{\rm (sm)} + c_j\sigma^{(\rm eft)}_{m,j} + c_j^2 \sigma^{(\rm eft)}_{m,jj}\]

which results in a quartic polynomial form for the \(\chi^2\) when inserted into the \(\chi2\) definition, namely:

\[\chi^2(c_j) = \sum_{k=0}^4 a_k \left(c_j\right)^k\]

Restricting the analysis to the linear order in the EFT expansion further simplifies to a parabolic form:

\[\chi^2(c_j) = \sum_{k=0}^2 a_k \left(c_j\right)^k = \chi^2_0 + b\left( c_j-c_{j,0} \right)^2\]

where \(c_{j,0}\) is the value of \(c_j\) at the minimum of the parabola, and in this case linear error propagation (Gaussian statistics) is applicable.

To determine the values of the quartic polynomial coefficients \(a_k\), it is sufficient to fit this functional form to a scan of the \(\chi^2\) profile obtained by varying the EFT coefficient \(c_j\) when all other coefficients are set to their SM value.

The associated 95 % CL interval to the coefficient \(c_j\) can then be determined by imposing the condition

\[\chi^2(c_j)-\chi^2(c_{j,0}) \equiv \Delta \chi^2 \le 5.91\]

We note that if the size of the quadratic \(\mathcal{O}\left(\Lambda^{-4}\right)\) corrections is sizable, there will be more than one solution for \(c_{j,0}\) and one might end up with pairwise disjoint CL intervals.