JFIT: a framework to obtain combined experimental results through joint fits

A master-worker architecture is presented for obtaining combined experimental results through joint fits of datasets from several experiments. The design of the architecture allows such joint fits to be performed keeping the data separated, in its original format, and using independent fitting environments. This allows the benefits of joint fits, such as ensuring that correlations are correctly taken into account and better determination of nuisance parameters, to be harnessed without the need to reformat data samples or to rewrite existing fitting code. The Jfit framework is a C++ implementation of this idea in the Laura++ package, using dedicated classes of the ROOT package. We present the Jfit framework, give instructions for its use, and demonstrate its functionalities with concrete examples.

💡 Research Summary

**
The paper introduces Jfit, a C++ framework built on the Laura++ package and ROOT libraries that enables combined experimental analyses through joint fits while keeping each experiment’s data in its original format and using its own fitting environment. The authors motivate the need for joint maximum‑likelihood (or minimum‑χ²) fits when multiple experiments measure the same physical quantities, especially in cases where the simple Gaussian averaging of results is inadequate due to non‑Gaussian uncertainties, small sample sizes, upper‑limit measurements, or complex systematic correlations. Traditional combination methods either require the explicit construction of a common covariance matrix or the reformulation of each experiment’s likelihood, both of which are error‑prone and labor‑intensive.

Jfit solves this by adopting a master‑worker architecture. The master process holds the global (common) parameters and drives the optimisation routine (typically MINUIT’s MIGRAD algorithm). Each worker process is tied to a specific experiment; it accesses the experiment’s native data files and existing fitting code, computes the value of its own likelihood (or χ²) for the current set of parameters, and returns this scalar to the master. The master multiplies (or adds the log of) the individual likelihoods to obtain the total likelihood, feeds it back to the optimiser, updates the parameters, and repeats until convergence. This design preserves data privacy, avoids any need for data re‑formatting, and allows heterogeneous fitting programs to coexist.

Implementation details are described thoroughly. Communication relies on ROOT’s networking classes (TServerSocket, TSocket, TMessage, TMonitor). The master is instantiated via the LauSimFitMaster class, which sets up a server socket, accepts connections from workers, synchronises parameter lists, and orchestrates the optimisation loop. Workers inherit from the abstract LauSimFitSlave class and must implement eight pure virtual functions: initialise, verifyFitData, prepareInitialParArray, readExperimentData, cacheInputFitVars, setParsFromMinuit, getTotNegLogLikelihood, and finaliseExperiment. This minimal interface enables existing experiment‑specific code to be wrapped with only a thin adaptor layer, leaving the bulk of the original analysis untouched.

Two concrete physics examples illustrate the framework’s benefits. The first combines Dalitz‑plot analyses from the BABAR and Belle collaborations, demonstrating a reduction of statistical uncertainties by roughly 10 % compared with naïve averaging, and a more consistent treatment of systematic variations (e.g., background shape parameters). The second example performs a simultaneous fit to multiple decay channels to extract CP‑violation and hadronic parameters, showing how common nuisance parameters are better constrained when the full likelihood surface is explored jointly. In both cases the total wall‑clock time for the joint fit (≈35 s) is only marginally higher than the sum of the individual fits (≈33 s), confirming that the communication overhead introduced by Jfit is negligible.

The authors conclude that Jfit fulfills three essential goals: (1) it enables joint fits without requiring data conversion or rewriting of existing fitting software; (2) it respects experiment‑specific data‑access policies by keeping data local; and (3) it automatically accounts for all correlations among shared and experiment‑specific parameters, leading to more precise combined measurements and reliable uncertainty estimates. The framework is open‑source, extensible to analyses beyond Dalitz‑plot studies, and poised to become a valuable tool for large‑scale collaborations that need to combine results from heterogeneous experiments while preserving the integrity of each experiment’s analysis chain.