Fittino, a program for determining MSSM parameters from collider observables using an iterative method

This is a software manual, and a good one. Fittino is a C++ framework that takes collider observables, calls an SLHA-compliant spectrum calculator (SPheno in the shipped version), and returns MSSM-24 or mSUGRA/GMSB/AMSB parameters with a full covariance matrix, correlations, and 2D contours. The genuinely useful parts are the plumbing: a flexible steering file covering masses, edges, cross-sections, BRs, xsbr products, brratio, mass limits, and low-energy observables; automatic pull distributions; and the observable-importance ranking via individual Δχ² contributions. The tree-level inversion formulas in §2.2.1 are written out explicitly, and the subsector-fit sequence is described concretely enough to reimplement. That is real infrastructure value for the LHC/ILC SUSY community, and the code is publicly referenced. Credit where due: this is the kind of paper that saves the next student six months. The soft spots are real but bounded. First, the load-bearing claim — stable global convergence in ~20 dimensions from tree-level start values — is not demonstrated here. §2.2.2 asserts it for SPS1a' and SPS7, but Fig. 3 shows trajectories for only two coordinates (tan β and X_bottom), and the quantitative 19-parameter fits are in [8], listed as 'to appear'. The pull and χ² closure tests in Figs. 5–6 are genuine and valuable, but §3 says CalcPullDist runs with MINOS off and the text has you enter the true values in fitParameter, so those validate local error propagation, not the global search. That is a fair thing to weigh, though it is also exactly what a manual would defer. Second, 'without any a priori knowledge' (abstract:s4) is a bit generous: the tree-level step consumes user-supplied cos2φL, cos2φR pseudo-observables and user-chosen sector assignments via nofit, and BoundsOnX/ScanX impose hard ranges on X_t, X_b, X_τ. Minor, but the abstract should not oversell. Third, there are no timing or success-rate numbers beyond 'order of several days on a PIII 1.3GHz' — a referee should ask for a table of SA convergence rate versus MaxCallsSimAnn. Who should read it: anyone building or comparing global SUSY fits, and anyone who needs a documented, scriptable front end to an SLHA calculator. It is not a physics result and does not pretend to be. I would send it out, with instructions to the referee to check reproducibility of the example and to press on the deferred validation and the a-priori-knowledge wording rather than on the manual's structure.

The paper's central claim is that Fittino reconstructs ~20 MSSM parameters from observables with no a priori knowledge, via tree-level estimates plus subsector fits or Simulated Annealing, followed by a global MINUIT fit. For that to hold, the global χ2 search must reliably reach the true minimum from tree-level start values in ~20 dimensions, and the quoted error matrix must be meaningful there. The in-manuscript evidence for the first half is thinner than the text implies. §2.2.2 asserts 'The Simulated Annealing algorithm has been tested on fits with 20 free parameters in the scenarios SPS1a' and SPS7. In each case, the convergence to the true minimum of the χ2 is good and very stable,' but the only figure shown (Fig. 3) plots trajectories for two of the 19 parameters (tan β and X_bottom). Convergence of two projections does not establish that all 19 coordinates landed at the global minimum; a run can look converged in tan β while a weakly constrained direction such as X_bottom is still drifting. The pull and χ2 closure tests (Figs. 5–6: pull mean 0.06, width 1.16σ; χ2 mean 69.2 for 71 d.o.f.) are genuine and valuable, but §3 states that for CalcPullDist 'Only MINIMIZE is used in the fit, MINOS is switched off,' and the text does not say the pull fits are started from tree-level estimates rather than from the true values. If they start at truth, they validate local error propagation and the parabolic error matrix, not the global search. So the load-bearing empirical claim — stable global convergence in ~20 dimensions — is supported here by a two-parameter projection plus a local closure test, with the quantitative demonstration in [8], which is listed as 'to appear' and is therefore fair to weigh. Note also that 'without any a priori knowledge' is slightly overstated: the tree-level step consumes user-supplied pseudo-observables cos2φL, cos2φR (§2.2.1) and user-chosen sector assignments via the nofit command, and BoundsOnX/ScanX impose hard ranges (−6000 < X < 200

The reader's core diagnosis is right: the empirical demonstration of stable global reconstruction is largely deferred to [8], which is listed as 'to appear' and so is legitimately in scope under the citation rule. But the reader overstates the gap by saying 'the only fit shown here is a two-parameter example.' That is not accurate. Fig. 3 shows Simulated Annealing trajectories from a 19-parameter SPS1a' fit, and Figs. 5–6 show pull and χ2 closure distributions from a 20-parameter SPS1a fit (pull mean 0.06, width 1.16σ; χ2 mean 69.2 against 71 d.o.f.). Those are real in-manuscript consistency checks, and the χ2 mean agreeing with the d.o.f. is meaningful evidence that the error treatment is not grossly wrong. What the reader missed is the sharper version of the same concern: the pull/χ2 tests, as described in §3, run with MINOS off and only MINIMIZE, and the text never states the start values used. If those fits begin at truth, they test the error matrix, not the global search, and the global-search claim then rests on two trajectory projections out of nineteen coordinates. That is the precise soft spot, and it is narrower and more technical than 'validation is deferred.' I would not change the verdict. This is a software manual for a documented, publicly available tool (Fittino 1.1.1, http://www-flc.desy.de/fittino), with a reimplementable algorithm, a specified SLHA interface, and explicit diagnostics (the 'Error Matrix accurate' flag, the quickstart's instruction to start from true values first to check whether the problem is constrained at all). Deferring the physics-results validation to a companion paper is normal practice for a program description, and the paper is candid about the cost and the failure modes of MINOS and GetContours. The 'without any a priori knowledge' phrasing is mildly overstated given the cos2φL/R pseudo-observables and the user-imposed X bounds, but §2.2.1 is transparent that these are start-value aids only and are not used in the fit. Conditional acceptance remains the right call; my read refines the reason rather than changing the outcome.

If the claims hold, the tool addresses a genuine and non-trivial problem: extracting Lagrangian parameters from a highly correlated, loop-corrected observable set in ~20 dimensions without seeding the fit at the truth. The engineering choices are sound and worth publishing as documentation: the SLHA-based calculator interface makes the fitter independent of any single spectrum code; the observable grammar covers the full range of quantities that LHC and ILC analyses actually measure, including edges used directly rather than as reconstructed masses, which is the right way to avoid importing correlations; and the built-in pull, chi-square, and observable-importance machinery is exactly the validation apparatus the field needs for such fits. The code is publicly available and the input/output formats are documented to a level that a user could reproduce a run. The weakness is that the paper documents the machinery thoroughly but demonstrates the load-bearing capability only in a two-parameter example started at the true values, deferring the multi-parameter validation to a companion listed as 'to appear'.

Suppose supersymmetry is real and the LHC and a future linear collider measure dozens of masses, cross-sections and branching fractions with high precision. The physics question then becomes an inverse problem: which Lagrangian parameters produced those numbers? This paper describes Fittino, a C++ program that answers that question by minimising a global $χ^2$ over up to 24 low-energy MSSM parameters plus 7 Standard Model parameters, or over the high-scale parameters of mSUGRA, GMSB or AMSB models. The point of the program is that it needs no prior knowledge of the parameter values: it bootstraps its own starting point from tree-level relations, refines it either with a sequence of fits on restricted parameter subsets or with a Simulated Annealing search, and then hands everything to MINUIT for the final global fit and error analysis. What a reader should care about is the deliverable: not just central values, but a full error matrix, correlations, two-dimensional confidence contours, pull distributions and a per-observable breakdown of which measurement pins down which parameter.

What carries the argument. The load-bearing mechanism is the staged navigation of a high-dimensional $χ^2$ surface that is known to contain many local minima. Three pieces carry it. First, tree-level inversion: relations such as $|μ| = m_W (Σ + Δ(cos 2φ_L + cos 2φ_R))^{1/2}$ and $M_2 = m_W (Σ - Δ(cos 2φ_L + cos 2φ_R))^{1/2}$, built from chargino masses and mixing angles, give starting values for the gaugino-Higgs sector, and analogous relations give $X_t$, $X_b$, $X_τ$ and the squark and slepton mass parameters. Second, Simulated Annealing: the $χ^2$ is treated as a potential at temperature $t$, parameter steps that worsen the fit are accepted with probability $e^{-(χ^2_i - χ^2_{i-1})/t}$, the step size for each parameter is re-tuned so that roughly half the moves are accepted, and $t$ is multiplied by a fixed factor each round until the search settles. Third, the global MINUIT fit with HESSE and MINOS, which converts the located minimum into an error matrix, correlations and contours. The subsector fits are the cheaper alternative to annealing: fit only the parameters a given set of observables actually depends on, with everything else frozen, so that MINUIT sees a lower-dimensional and better-conditioned problem.

Core claim. The paper's claim is that a full, loop-corrected, unbiased determination of SUSY Lagrangian parameters from collider observables is practically achievable as a piece of software, and it presents the architecture that makes it work. Fittino treats every observable as a function of every parameter, so loop corrections and parameter correlations are kept rather than approximated away; the theoretical predictions come from an external spectrum calculator (SPheno in the shipped implementation) reached through the SUSY Les Houches Accord text-file interface, so any compliant code can be substituted. The initialisation problem is solved in stages: tree-level relations are inverted to estimate the gaugino, Higgs, squark and slepton sectors, with the estimate's spread obtained by repeating the inversion 10,000 times with observables smeared within their errors; then either a chain of fits on parameter subsets (sleptons, then squarks, then a joint fit of $tan β$ with the third-generation mixing parameters) or a Simulated Annealing search moves the starting point close enough to the global minimum for a conventional fit to converge. The output is the parameter set, its covariance and correlation matrices, optional MINOS asymmetric errors, two-dimensional $Δχ^2$ contours, and diagnostics that let a user cross-check the fit against itself.

The premise it rests on. The whole construction rests on the assumption that the global $χ^2$ surface of a realistic 19-to-24-parameter fit can actually be navigated from tree-level starting values to the true minimum by the subsector sequence or the annealing schedule, and the evidence for that is deferred to a companion paper listed as 'to appear', while the only fit worked through in this paper is a two-parameter example that is started at the true values.

Given a simulated set of LHC and ILC measurements, a user can extract the underlying MSSM parameters and, crucially, the full correlation matrix, so the precision on any one parameter is quoted with the others marginalised rather than fixed.

Because the theory side is reached through the SUSY Les Houches Accord, the fitting layer is decoupled from the spectrum calculator, so improved loop corrections can be adopted by swapping the external code rather than rewriting the fit.

The per-observable $Δχ^2$ breakdown turns the fit into an optimisation tool for the collider itself: it identifies which measurements actually constrain which parameters, and therefore which centre-of-mass energies, luminosities and beam polarisations are worth buying.

Automatically generated pull distributions over many pseudo-experiments give an internal check for bias in the fitted central values and for whether the quoted uncertainties match the actual scatter.

The same machinery applies to high-scale models, so mSUGRA, GMSB or AMSB parameters can be fitted directly to collider observables and the reconstructed high-scale pattern compared with the low-scale fit.

The architecture implies a reusable pattern for any beyond-Standard-Model inverse problem: invert the analytic relations for a starting point, then use restricted-observable fits or annealing to cross the barriers, then a conventional minimiser for the errors. The pattern is independent of supersymmetry as such.

The observable-importance output is arguably the more durable product than any single fit result, because it converts a parameter-determination study into a detector-design argument, and that argument survives even if the underlying model is wrong.

A natural stress test the authors do not run here is a fit in which a parameter is genuinely unconstrained by the available observables; the expected signature would be a non-positive-definite error matrix, which the program already reports, and that report could be turned into a systematic map of which regions of MSSM parameter space the LHC plus ILC combination cannot close.

The reliance on a text-file interface to an external calculator sets a practical ceiling on throughput, since every $χ^2$ evaluation is a subprocess call; the paper's own annealing budget of hundreds of thousands of calls suggests that surrogate or cached spectrum evaluation would be the obvious next engineering step.

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review