MadAnalysis 5: Collider Phenomenology Framework

• MadAnalysis 5 is an open-source framework for collider phenomenology that integrates simulation, recasting, and statistical analysis for LHC data.
• The framework couples a Python-based interactive front-end with a high-performance C++ kernel, enabling rapid prototyping and detailed event processing.
• It supports diverse detector simulation interfaces, jet reconstruction methods, and multi-region statistical tools to validate experimental searches.

MadAnalysis 5 (MA5) is an open-source framework for collider phenomenology and LHC reinterpretation that processes Monte Carlo event samples at parton level, hadron level, and reconstructed level. It was introduced as a user-friendly environment coupling a Python-based front-end to a C++ analysis kernel, and it subsequently developed into a broader recasting platform that includes detector-simulation interfaces, multi-region analysis management, a public database of validated ATLAS and CMS analyses, and multiple statistical back-ends for exclusion and sensitivity studies (Conte et al., 2012, Araz et al., 11 Jul 2025).

1. Foundational architecture

MA5 is organized around two principal layers. In normal mode, analyses are written through an interactive Python command-line interface using a dedicated metalanguage for importing datasets, defining particle and multiparticle labels, plotting observables, and applying selections. In expert mode, analyses are implemented directly in C++ inside the SampleAnalyzer core, which exposes event readers, unified internal data formats, physics services, and output writing (Conte et al., 2012).

The architectural split is central to the software’s identity. The Python layer automatically generates a C++ analysis, compiles and links it to the SampleAnalyzer library and optional tools, executes it, and stores the results in the XML-based SAF structure. The front-end then renders reports in HTML, PostScript/LaTeX, or PDF that include histograms and selection efficiencies. This design preserves an interactive workflow for rapid prototyping while retaining a compiled execution path for large event samples and more complex analyses (Conte et al., 2013).

From the outset, MA5 was positioned as a framework for professional analyses of simulated signals and Standard Model backgrounds, including event selections to isolate signals, kinematic distributions, cutflows, and user-defined signal-over-background ratios. A plausible implication is that MA5 was designed not merely as a plotting utility, but as an analysis environment intended to support both exploratory phenomenology and reproducible reinterpretation workflows (Conte et al., 2012).

2. Analysis construction and event-processing workflow

MA5 accepts event samples generated by Monte Carlo event generators in several standard formats. Across the documented releases, supported inputs include LHE, StdHep, HepMC, LHCO, Delphes Root, and ROOT-based reconstructed samples, with compressed input supported when zlib is available (Conte et al., 2013, Conte et al., 2014).

The normal-mode workflow follows a stable pattern: import samples, group them into datasets, define particle labels, plot observables, apply event- or object-level selections, submit the analysis, and open the resulting report. Representative commands include importing multiple files with wildcards, defining composite labels such as mu = mu+ mu-, plotting MET or PT(mu), and applying cuts such as reject (mu) PT < 20 or select 80 < M (mu+ mu-) < 100 (Conte et al., 2013).

The observables available to this workflow include standard collider kinematics such as transverse momentum,

$p_T = \sqrt{p_x^2 + p_y^2},$

pseudorapidity,

$\eta = -\ln \tan(\theta/2),$

and angular separation,

$$\Delta R = \sqrt{(\Delta \eta)^2 + (\Delta \phi)^2}.$$

MA5 also supports invariant masses, missing transverse energy, scalar transverse-energy sums, particle ordering by hardness, and operations on four-momenta such as sums and differences (Conte et al., 2013, Conte et al., 2012).

In expert mode, the workflow is recast into the Initialize, Execute, and Finalize structure of a C++ analysis class. This mode exposes direct access to event.mc() and event.rec() collections, TLorentzVector-based observables, physics helpers for boosts and event-level quantities, and custom histogram logic. The expert interface was later extended to handle multiple signal and control regions without duplicating branching cutflows, using a dedicated manager with methods such as AddRegionSelection, AddCut, ApplyCut, and FillHisto (Conte et al., 2014).

3. Reconstruction, detector treatment, and object-level modeling

In its earliest documented form, MA5 analyzed reconstructed objects produced by external fast detector tools such as PGS 4 and Delphes, while also supporting truth-level event analysis. Subsequent releases introduced increasingly direct control over reconstruction, first through a FastJet interface for jet clustering and later through internal detector-emulation layers (Conte et al., 2012, Conte et al., 2013).

The FastJet integration added jet clustering in reconstructed-level mode with support for kt, cambridge, antikt, genkt, siscone, cdfmidpoint, cdfalgo, and gridjet. Users can configure the jet radius and transverse-momentum threshold, control whether leptons from the hard process feed the jet algorithm, and set matching-based b-tagging efficiencies and misidentification rates. Reconstructed events can then be exported to LHE or LHCO according to simplified conventions (Conte et al., 2013).

Later developments broadened the detector layer. MA5 added interfaces to Delphes 3 and to Delphes-MA5tune, a modified Delphes version that stores lepton and photon isolation variables in the output so that isolation can be applied at analysis level rather than hard-coded in the simulation. This was presented as essential for faithful recasting and flexible object definitions (Conte et al., 2014).

A more substantial internalization of detector modeling came with the simplified fast detector simulator (SFS), available from MA5 v1.8.51. SFS allows detector parametrization through user-defined smearing functions, efficiencies, and taggers written directly in the MA5 interpreter, with C++ code generated automatically at run time. The documented comparison with Delphes states that predictions generally agree to a level of about 10% or better, while SFS is often a factor ~2 faster within recasting and can produce output files up to ~100× smaller in default settings (Araz et al., 2020).

The SFS was subsequently extended to long-lived particles. The added features include charged-particle propagation in a homogeneous magnetic field along the detector $z$-axis, computation of the point of closest approach, transverse and longitudinal impact parameters $d_0$ and $d_z$, configurable isolation cones for tracks, leptons, and photons, and energy scaling for reconstructed objects. These additions were used to validate recasts of displaced-lepton, disappearing-track, and displaced-vertex searches (Araz et al., 2021).

A common misconception is that MA5 itself always provides a realistic detector simulation. The historical record is more qualified: early MA5 relied on external reconstructed inputs, later versions integrated Delphes and lightweight FastJet-based options, and SFS introduced a parametrized internal emulator. The papers consistently stress that fast simulation remains an approximation and that validation against official cutflows and distributions is necessary (Conte et al., 2012, Araz et al., 2020).

4. Recasting infrastructure and statistical interpretation

A decisive change in MA5’s role came with the extension of expert mode to multi-region analyses and with the creation of the Public Analysis Database (PAD). From v1.1.10 onward, MA5 supported analyses with multiple signal regions, control regions, and subanalyses, and validated implementations could be distributed through the PAD with accompanying detector cards, XML metadata, validation material, and DOI registration on INSPIRE (Conte et al., 2014, Dumont et al., 2014).

This recasting infrastructure formalized the mapping between a public analysis and MA5 outputs. The C++ analysis code produces cutflows and histograms in SAF, while a companion .info XML file provides luminosity, observed counts, background expectations, and uncertainties. In early recasting workflows, MA5 supplied a CLs-based script, exclusion_CLs.py, that used the expected signal yield

$$N_{\text{sig}} = \mathcal{L} \times \sigma \times A \times \epsilon$$

together with the published background information to derive exclusions (Dumont et al., 2014).

The statistical layer was later expanded in several steps. MA5 v1.8 introduced explicit handling of theory uncertainties on the signal cross section, user-defined systematic uncertainties on the signal yield, and extrapolation to higher luminosities with several prescriptions for background-error scaling. These features were integrated into the recasting workflow through dataset attributes such as scale_variation, pdf_variation, and main.recast.add.extrapolated_luminosity (Araz et al., 2019).

Signal-region combination was then added through two complementary mechanisms: an interface to pyhf for ATLAS JSON-serialized HistFactory workspaces, and a simplified-likelihood treatment using CMS covariance matrices. The corresponding likelihoods preserve either the published HistFactory nuisance-parameter structure or a Gaussian covariance model for statistically disjoint signal regions (Alguero et al., 2022).

Method	Input published by experiment	MA5 implementation
Full likelihood	ATLAS JSON HistFactory workspace	pyhf interface
Simplified likelihood	CMS covariance matrices	simplified likelihood calculation
Legacy single-region limit	SR yields and background counts	CLs-based recast workflow

The most recent statistical update described in the supplied corpus is MA5 v1.11, which moved to the Python package spey, optionally with a pyhf plugin for HistFactory models. In that version, independent regions, correlated regions, and full HistFactory workspaces are handled within a common asymptotic profile-likelihood framework, and a new analysis_only_mode allows efficiency-only runs without statistical post-processing (Araz et al., 11 Jul 2025).

These developments also address an important caveat in reinterpretation. The single “best signal region” strategy can lose sensitivity, can over-exclude in some cases, and can induce numerical instabilities when the most sensitive region changes across parameter space. This was one of the main motivations for implementing global likelihoods and covariance-aware combinations in MA5 (Alguero et al., 2022).

5. Versioned extensions and advanced analysis capabilities

MA5’s development history is marked by the addition of specialized capabilities aimed at increasingly complex LHC analyses. The progression can be summarized through a small set of version milestones.

Version or branch	Extension	Source
v1.1.8	FastJet interface, jet reconstruction, DJR checks	(Conte et al., 2013)
v1.1.10	multi-region expert mode, Delphes-MA5tune recasting workflow	(Conte et al., 2014)
v1.8.51	simplified fast detector simulation	(Araz et al., 2020)
v2.0.4 and later in v2.0.X	HEPTopTagger integration and substructure tools	(Araz et al., 2023)
v1.11	Efficiency1D, rotations/boosts, RestFrames, spey/pyhf statistics	(Araz et al., 11 Jul 2025)

One line of extension concerns jet substructure and boosted-object reconstruction. In the MA5 2.0.X branch, the framework added a Substructure namespace interfacing FastJet and FastJet Contrib, together with a wrapper for HEPTopTagger v2. The documented interface allows multiple jet collections, reclustering, Variable-R reclustering, and direct invocation of the tagger in expert mode through Substructure::HTT. The study presenting this implementation emphasized that normal mode applies reconstruction efficiencies only to the primary jet collection, so expert mode is recommended for multi-collection substructure workflows (Araz et al., 2023).

Another line concerns frame-dependent observables and recursive jigsaw reconstruction. MA5 v1.11 introduced an Efficiency1D class for one-dimensional efficiency tables taken from HEPData, documented and extended Lorentz-rotation and boost utilities, and integrated the RestFrames library without ROOT by replacing vector, matrix, and minimization dependencies with MA5-native tools, Eigen, and a built-in Nelder–Mead algorithm. This infrastructure was used for soft-lepton electroweakino analyses employing jigsaw variables such as $M_T^S$ and $R_{\mathrm{ISR}}$ (Araz et al., 11 Jul 2025).

The advanced feature set has also sharpened the distinction between the two operating modes. Normal mode remains concise and interactive, but expert mode is repeatedly recommended when analyses require constituent-level smearing, multiple jet collections, full control over substructure, or bespoke region logic (Araz et al., 2023, Conte et al., 2014).

6. Scientific applications, ecosystem, and limitations

MA5 has been used as the computational basis for a wide range of reinterpretation studies. In one example, existing ATLAS and CMS Run-1 searches were recast through MA5 PAD modules to constrain the radiative decay $\tilde g \to g\,\widetilde{\chi}_1^0$, leading to the result that compressed spectra with $\eta = -\ln \tan(\theta/2),$0 are excluded up to about 750 GeV at 95% CL under the assumption $\eta = -\ln \tan(\theta/2),$1 (Chalons et al., 2015).

In another application, MA5 PAD recasts of Run-1 ATLAS dilepton plus missing-energy analyses were used to constrain the Inert Doublet Model. That study found exclusions reaching $\eta = -\ln \tan(\theta/2),$2 up to about 35–55 GeV, depending on $\eta = -\ln \tan(\theta/2),$3 and $\eta = -\ln \tan(\theta/2),$4, and $\eta = -\ln \tan(\theta/2),$5 up to about 135–140 GeV for very light $\eta = -\ln \tan(\theta/2),$6, thereby extending previous LEP limits in the $\eta = -\ln \tan(\theta/2),$7 regime (Belanger et al., 2015).

The public-analysis ecosystem around MA5 has expanded through repeated community efforts. The first and second MA5 recasting workshops documented the implementation and validation of numerous ATLAS and CMS analyses across monojet, monophoton, mono-Higgs, supersymmetry, long-lived particles, seesaw models, and four-top production, with most implementations released through the PAD and associated dataverses (Fuks et al., 2018, Fuks et al., 2021).

The framework’s scientific role is therefore twofold. First, it provides a reproducible environment for analysis design and phenomenological studies. Second, it acts as an interoperability layer between Monte Carlo generators, fast detector emulators, validated experimental recasts, and statistical tools. This suggests that MA5 has become a methodological bridge between phenomenological model building and the published output of the LHC experiments (Dumont et al., 2014, Araz et al., 11 Jul 2025).

Its limitations are explicit in the literature. Recast fidelity is constrained by the quality of the fast detector simulation, by the completeness of public experimental documentation, by object-calibration approximations, and by the availability of official likelihood information. Several papers note that fast simulation cannot replace full experimental reconstruction, that control-region handling may remain approximate, and that detector- or trigger-level details can dominate the residual discrepancies between MA5 and official cutflows (Conte et al., 2014, Araz et al., 2020).