Event generation and detector simulation

A hard-process calculation describes short-distance scattering, often using parton distributions for incoming hadrons. A shower approximates additional soft and collinear radiation, while hadronization models turn partons into hadrons and unstable particles decay. These ingredients have different accuracy claims: a fixed-order matrix element, a logarithmic shower approximation, and a tuned nonperturbative model are not interchangeable descriptions of exact quantum evolution. Matching and merging aim to combine domains without counting the same radiation twice.

Detector transport then models interactions in material and magnetic fields. Digitization converts deposits into electronic signals, including noise and thresholds; reconstruction infers tracks, clusters, and event quantities from those signals. Pileup and changing operating conditions can modify the response. A generator-level particle record is consequently different from a reconstructed event record. Keeping those stages distinct lets an analyst locate a discrepancy rather than treating a full simulation as one indivisible black box.

For an unweighted sample drawn from the normalized signal model, the fraction passing a selection estimates its probability. Multiplying that fraction by cross section and integrated luminosity gives a predicted event count. Generating more events reduces numerical Monte Carlo error but does not increase the luminosity of the actual experiment. It also cannot correct a missing physical process, an inaccurate material description, or an inappropriate calibration. Those are uncertainties in the integrand or response rather than sampling noise.

Weighted generation changes the estimator. If weights have the declared relative normalization, selected weight divided by total weight estimates the selected fraction within that generated population. Negative weights can arise in higher-order calculations, so the ratio need not behave like an ordinary binomial efficiency and an effective sample size can be much smaller than the event count. For a Poissonized independent-event sample, the sum of squared accepted weights estimates the variance of its unnormalized weighted yield. Fixed-size sampling and ratios require their own normalization and covariance treatment.

Closure tests ask whether an analysis recovers an injected truth when pseudo-data are produced by a known model. Passing closure is necessary evidence about implementation, but using the same response for generation and inversion makes this a limited test of modeling. Alternative generators, distorted spectra, and independent control data probe different vulnerabilities. Agreement after tuning should be evaluated on observables or samples that were not all used to choose the tune.

A useful uncertainty description identifies which variations represent perturbative truncation, nonperturbative tuning, detector calibration, and finite simulated statistics. Arbitrarily adding more simulation does not turn these into one statistical error. Likewise, reusing one event sample under many parameter weights creates correlations between predictions. Recording those correlations, generator versions, truth definitions, and normalization conventions makes a later likelihood or cross-section result reproducible and allows a disagreement to be interpreted physically.