Pile-Up Effects in Detector & Collider Physics

• Pile-Up Effects are distortions in recorded data caused by unresolved overlapping interactions, leading to merged signals and biased measurements.
• The phenomenon occurs in systems like X-ray detectors and colliders, where simultaneous arrivals skew energy spectra, count rates, and jet kinematics.
• Mitigation uses methods such as Fourier deconvolution, simulation-based forward folding, and machine learning to reconstruct individual event signals accurately.

Searching arXiv for recent and foundational papers on pile-up effects across detector and collider contexts. arxiv_search(query="pile-up effects pileup review detector x-ray collider", max_results=10) Pile-up effects are distortions induced when multiple physical quanta or interactions are registered within a detector or reconstruction interval that is too coarse to resolve them individually. Across photon-counting detectors, X-ray CCDs, scintillator readout, proton–proton collisions, and heavy-ion experiments, the common mechanism is superposition: signals that should be distinct are merged in time, space, or event reconstruction, producing biased spectra, altered count rates, degraded resolution, and nontrivial nonlinearities. In solid-state X-ray detectors, $n$ photons arriving within one detector-response window appear as a single count whose recorded energy is the sum of the individual photon energies (Hernández et al., 2016). In hadron-collider reconstruction, multiple proton–proton interactions in the same or nearby bunch crossings overlay diffuse soft activity and sometimes hard jets on top of the hard scatter, biasing jet kinematics, missing transverse momentum, and correlation observables (Soyez, 2018). In heavy-ion fluctuation studies, even a small fraction of unresolved double collisions can strongly distort higher-order cumulants of multiplicity distributions (1705.01256). These manifestations differ in implementation but are unified by a statistical problem of unresolved overlap.

1. Physical origin and operational definitions

Pile-up in photon-counting systems originates from finite detector response or readout time. In solid-state detectors used with laser-driven pulsed X-ray sources, each detected photon produces an electrical pulse of duration $\tau_d$, while the source emits photons in bursts of duration $\tau_p$ with $\tau_p \ll \tau_d$; many photons from one laser pulse may therefore interact almost simultaneously and be recorded as one event with summed energy (Hernández et al., 2016). In X-ray CCDs, whose pixels typically integrate for seconds at a time, pile-up occurs when two or more photons arrive into the same or immediately adjacent pixels within one frame; the instrument may register a single event whose pulse height is the sum of the photon energies, may change the event grade, or may reject the event entirely (Tamba et al., 2021).

In photon-counting electronics with shaped pulses, pile-up is the overlap of pulses within the shaping and recovery window. For bipolar shaping, two complementary effects arise: peak pileup, which raises measured energy when pulses overlap on the positive lobe, and tail pileup, which lowers measured energy when a later pulse lands on the negative undershoot and may even push it below threshold (Chaplin et al., 2012). In MAPMT single-photoelectron counting with extended dead time, pile-up appears as saturation and quenching of the counting rate when two photoelectrons arrive within the double-pulse resolution of the ASIC (M'sihid et al., 28 Nov 2025).

In collider experiments, pileup refers to multiple proton–proton collisions occurring in the same bunch crossing as the hard collision of interest. The additional soft particles are overlaid on the hard event and bias reconstructed jet transverse momentum, jet mass, substructure, and missing transverse energy (Bertolini et al., 2014). The average number of interactions per bunch crossing is denoted $\langle \mu \rangle$; values around $60$ in Run 2 and around $200$ under HL-LHC conditions are explicitly discussed in the literature (Vaughan et al., 4 Mar 2025). In heavy-ion running, pile-up denotes unresolved superposition of two independent collisions within the detector time or vertex-resolution window, so that the recorded event is a convolutional mixture of single-collision multiplicity distributions (Zhang et al., 2021).

A useful generalization is that pile-up can occur either at the level of detector response or at the level of event reconstruction. This suggests two broad classes: signal-domain pile-up, where analog or digitized detector pulses merge before measurement, and event-domain pile-up, where separately produced particles or collisions are reconstructed as a single physics event. The distinction is operationally important because the former often requires forward modeling of detector response, whereas the latter often requires subtraction, unfolding, or object-level identification.

2. Statistical and mathematical structure

The basic stochastic structure of pile-up is usually Poissonian at the level of arrivals. In pulsed X-ray spectroscopy, if the mean number of photons per laser shot in the detector is $\lambda$, then the observed count rate $r$ at laser repetition rate $\nu$ satisfies

$\tau_d$0

so that

$\tau_d$1

If exactly $\tau_d$2 photons arrive in one detector window, their summed-energy distribution is the $\tau_d$3-fold convolution $\tau_d$4 of the true single-photon energy-deposition density $\tau_d$5, and the measured piled-up density is the Poisson-weighted mixture

$\tau_d$6

In Fourier space this becomes

$\tau_d$7

with inverse transform

$\tau_d$8

providing an explicit depile-up transform (Hernández et al., 2016).

In X-ray CCDs the arrival of photons in each pixel or small cluster during one frame is also modeled as a Poisson process with mean

$\tau_d$9

The pile-up probability grows roughly as $\tau_p$0, which at $\tau_p$1 is approximately $\tau_p$2 (Tamba et al., 2021). However, because grade migration and event rejection are present, no closed-form linear response exists in general; the appropriate formalism is nonlinear forward folding with a simulator (Tamba et al., 2021).

For counts-per-frame in CCDs, one minimal parametric extension of the Poisson model introduces a parameter $\tau_p$3 that represents an effective single-photon pile-up probability per existing count. The resulting observed-count distribution is no longer Poisson, and a second model for waiting times introduces a parameter $\tau_p$4 representing the fraction of zero-length waits lost to pile-up. These distributions can be embedded directly in $\tau_p$5 or likelihood inference for background or signal analysis (Sevilla, 2013).

In heavy-ion event pile-up, if $\tau_p$6 is the single-collision multiplicity distribution and a fraction $\tau_p$7 of recorded events are double collisions, then

$\tau_p$8

with corresponding distortion of cumulants generated through

$\tau_p$9

This directly couples small pile-up fractions to large biases in higher-order cumulants through products of lower moments (Zhang et al., 2021).

At hadron colliders, the dominant soft pileup contamination is often represented through an event-by-event transverse-momentum density $\tau_p \ll \tau_d$0, estimated as the median of $\tau_p \ll \tau_d$1 over event patches or jets,

$\tau_p \ll \tau_d$2

and subtracted from each jet as

$\tau_p \ll \tau_d$3

This framework describes mean additive contamination, but it does not by itself identify hard jets originating from pileup collisions, which require additional methods (Hautmann, 2015).

3. Observable distortions and phenomenology

The most direct spectral consequence of pile-up in photon detectors is energy migration. In pulsed X-ray spectroscopy, low-energy photons are promoted to higher bins because multiple hits are recorded as one event with summed energy, which biases any quantitative estimate that assumes one count equals one photon (Hernández et al., 2016). Experimental application showed that apparent differences between spectra acquired with different apertures were purely pile-up and that depiling caused them to collapse onto nearly the same curve (Hernández et al., 2016).

In X-ray CCDs, three nonlinear signatures recur: spectral hardening, flux loss, and grade migration. Spectral hardening arises because low-energy photons, which are more numerous in a steep spectrum, coalesce into fewer high-energy counts; flux loss occurs because merged events exceed standard event-selection thresholds and are rejected; grade migration occurs because the pattern of split-pixel charge shifts toward more extended or invalid shapes (Tamba et al., 2021). A systematic Suzaku XIS study found that the relative spectral hardening, split-event branching ratio, and detached-event branching ratio all increase monotonically as the pileup fraction increases (Yamada et al., 2011).

For scintillator-based neutrino detectors, $\tau_p \ll \tau_d$4C pile-up in JUNO biases the positron energy reconstruction through additive visible energy,

$\tau_p \ll \tau_d$5

and worsens the target resolution. A residual $\tau_p \ll \tau_d$6C activity of $\tau_p \ll \tau_d$7 leads to a few-percent worsening, quantified as $\tau_p \ll \tau_d$8 at $\tau_p \ll \tau_d$9 (Fang et al., 2 Mar 2026). The problem is especially acute when the charge ratio $\langle \mu \rangle$0 is small but the time separation $\langle \mu \rangle$1 is short, making waveform-level discrimination difficult (Fang et al., 2 Mar 2026).

For shaped-pulse gamma-ray detectors, peak pileup generates false high-energy continuum, while tail pileup creates low-energy deficits and energy-dependent losses due to sub-threshold events (Chaplin et al., 2012). In microdosimetry, superposed ionization signals distort lineal-energy spectra and thereby bias derived quantities such as the frequency-mean lineal energy $\langle \mu \rangle$2 and dose-mean lineal energy $\langle \mu \rangle$3 (Pierobon et al., 4 Apr 2025).

In collider reconstruction, pileup raises jet $\langle \mu \rangle$4 and mass, alters substructure, degrades $\langle \mu \rangle$5 resolution, and can inject spurious jets. Jet transverse momentum acquires an average shift $\langle \mu \rangle$6 and a smearing of order $\langle \mu \rangle$7 (Soyez, 2018). Soft pileup also washes out dijet angular correlations and populates incorrect regions in jet–photon correlation planes (Hautmann, 2015). In heavy-ion fluctuation measurements, even residual pile-up fractions of order $\langle \mu \rangle$8 can have outsized effects on $\langle \mu \rangle$9, especially at low beam energies where proton multiplicities are large (1705.01256).

4. Correction and mitigation methodologies

The principal correction strategies divide into inverse transforms, forward simulation, subtraction methods, data-driven mixing, and learned discriminants or regressors.

In laser-driven pulsed X-ray spectroscopy, the Poisson-convolution formalism yields an explicit depile-up algorithm. After estimating $60$0 from $60$1, the measured histogram is normalized to a pdf, zero-padded, Fourier transformed, mapped through

$60$2

and inverse transformed to recover the estimated single-photon spectrum (Hernández et al., 2016). If non-pile-up noise is significant, a parametric form such as continuum Maxwellian plus Gaussian lines can instead be forward-piled via the transform and fit directly to the measured spectrum (Hernández et al., 2016).

In X-ray CCD spectral analysis, the preferred approach is simulation-based nonlinear forward folding. For each trial source spectrum $60$3, a detailed detector simulator generates a pile-up–affected count spectrum $60$4 including physical interactions, charge-cloud generation, charge diffusion, event grading, frame readout, and data reduction. Parameters are then optimized by minimizing $60$5 or $60$6-stat against the observed spectrum (Tamba et al., 2021). This framework naturally handles grade migration and rejected events, which are not captured by simple deconvolution (Tamba et al., 2021). A complementary operational mitigation for Suzaku XIS is core exclusion: masking the point-spread-function core within radii corresponding to $60$7 or $60$8 pileup fraction (Yamada et al., 2011).

For detector electronics dominated by dead-time saturation, inversion of the saturation curve is central. In the extended dead-time model used for Mini-EUSO,

$60$9

so the true flux is recovered through the Lambert-$200$0 inversion

$200$1

once the effective double-pulse resolution $200$2 has been determined (M'sihid et al., 28 Nov 2025).

At hadron colliders, the standard soft-contamination correction is the area–median method, in which $200$3 is estimated event-by-event and subtracted from each jet (Soyez, 2018). Variants include Constituent Subtraction, which removes local soft contamination at the level of particles or calorimeter cells, and SoftKiller, which sets a dynamical event-level soft threshold (Berta et al., 2023, Soyez, 2018). PUPPI instead assigns each particle a weight based on a local shape variable $200$4 that probes whether the neighborhood is collinear and jet-like or diffuse and pileup-like, and rescales four-momenta as $200$5 (Bertolini et al., 2014). For jet shapes, a distinct subtraction formalism computes the numerical susceptibility of a shape to infinitesimal uniform background and extrapolates back to zero pileup, optionally including hadron-mass effects through $200$6 (Soyez et al., 2012).

A separate collider strategy is data-driven jet mixing. One constructs a mixed distribution by overlaying zero-pileup hard-scatter events with minimum-bias events recorded at the same pile-up, runs the standard jet finding, forms the bin-by-bin ratio

$200$7

and reweights the zero-pileup signal or theory prediction accordingly (Hautmann, 2015). This is designed to treat correlation observables and hard jets from pileup statistically, including outside tracker acceptance (Hautmann, 2015).

Machine-learning methods introduce per-particle or per-jet pile-up identification or regression. PUMA regresses the hard energy fraction of each particle-flow candidate using sparse transformers on raw reconstructed PF inputs and outperforms CHS and PUPPI in missing transverse momentum and jet observables in realistic detector simulation (Maier et al., 2021). PUMiNet uses event-wide attention over jets and tracks to predict per-jet hard-scatter energy and mass fractions, enabling jet-level corrections under $200$8 (Vaughan et al., 4 Mar 2025). In JUNO, 1DCNN, 2DCNN, and transformer models are used to classify positron events affected by single $200$9C pile-up from waveform-derived inputs (Fang et al., 2 Mar 2026).

5. Quantitative performance and trade-offs

A recurrent theme is that complete avoidance of pile-up is often suboptimal because it discards too much useful data. In laser-driven pulsed X-ray spectroscopy, a strict single-hit condition enforced by severe attenuation would reduce count rate by approximately $\lambda$0, increasing statistical noise by approximately $\lambda$1, whereas applying pile-up correction at $\lambda$2 improves signal-to-noise ratio by roughly a factor of three over the single-hit approach (Hernández et al., 2016). Numerical tests at $\lambda$3 with added $\lambda$4-distributed noise showed that direct Fourier depiling removes about $\lambda$5 of repetition-peak artefacts, while a parametric Maxwellian-plus-Gaussian fit corrects them nearly completely while preserving continuum shape (Hernández et al., 2016).

In simulation-based CCD analysis, validation spans negligible, moderate, and strong pile-up regimes. For PKS 2155–304, conventional linear and nonlinear simulation-based fits yielded identical photon indices and fluxes within less than $\lambda$6 differences (Tamba et al., 2021). For Aquila X-1, linear fitting underestimated both $\lambda$7 and flux by approximately $\lambda$8–$\lambda$9, while nonlinear analysis corrected $r$0 upward by about $r$1 and recovered about $r$2 more flux (Tamba et al., 2021). For the Crab Nebula, linear fits gave $r$3 with flux suppressed by about $r$4, whereas nonlinear simulation recovered $r$5 and flux within $r$6 of contemporaneous Suzaku HXD and NuSTAR measurements (Tamba et al., 2021).

Suzaku diagnostics identify practical tolerances. Relative spectral hardening is flat for $r$7 and then rises monotonically, reaching approximately $r$8 at $r$9 (Yamada et al., 2011). On that basis, events outside radii corresponding to a pileup fraction of $\nu$0 or $\nu$1 are recommended for spectral analysis (Yamada et al., 2011).

In LHC jet reconstruction, quantitative gains are typically expressed as reduced bias and improved resolution. ATLAS Run-1 found that for $\nu$2 dijets, the uncorrected slope $\nu$3 was about $\nu$4/vertex and was reduced to $\nu$5/vertex after $\nu$6 subtraction (Collaboration, 2015). JVT at $\nu$7 hard-scatter efficiency achieved a fake rate of about $\nu$8, compared with about $\nu$9 for a JVF-based selection (Collaboration, 2015). PUPPI, in a toy detector with $\tau_d$00, improved central-jet $\tau_d$01 resolution from about $\tau_d$02 for PFlow and about $\tau_d$03 for PFlow+CHS to about $\tau_d$04, and improved $\tau_d$05 resolution from about $\tau_d$06 for PFlow and about $\tau_d$07 for PFlow+CHS to about $\tau_d$08 (Bertolini et al., 2014).

New density estimators target reduced process dependence. The SignalFreeBackgroundEstimator reduced the sample-to-sample spread in $\tau_d$09 bias to at most $\tau_d$10 at $\tau_d$11 and at most $\tau_d$12 even at $\tau_d$13, compared with about $\tau_d$14 for the standard grid-median method; with charged-track seeds, the new $\tau_d$15 bias stayed below $\tau_d$16 for all $\tau_d$17 and all samples (Berta et al., 2023). Under these inputs, jet energy-scale bias for anti-$\tau_d$18, $\tau_d$19 jets in the $\tau_d$20–$\tau_d$21 bin was reduced from $\tau_d$22–$\tau_d$23 to at most $\tau_d$24 (Berta et al., 2023). PUMiNet, at $\tau_d$25, achieved $\tau_d$26 for hard-scatter energy fraction and $\tau_d$27 for hard-scatter mass fraction, and its predictions restored a narrow Higgs resonance near $\tau_d$28 in the reconstructed di-jet mass spectrum (Vaughan et al., 4 Mar 2025).

Waveform-classification performance in JUNO is reported through ROC AUC and time-separation dependence. At a working point with pure $\tau_d$29 acceptance of $\tau_d$30, 1DCNN and transformer models achieved AUC values of approximately $\tau_d$31, and in the critical window $\tau_d$32 achieved true-positive rates of about $\tau_d$33–$\tau_d$34, compared with about $\tau_d$35–$\tau_d$36 for 2DCNN (Fang et al., 2 Mar 2026).

6. Limitations, systematics, and unresolved issues

Pile-up corrections are intrinsically model-sensitive because superposition is nonlinear. In the Fourier depile-up formalism for pulsed X-ray sources, if $\tau_d$37 then effective resolution worsens substantially; for $\tau_d$38 and $\tau_d$39 the reported resolution degrades by more than a factor of two (Hernández et al., 2016). The discrete Fourier implementation assumes negligible spectral weight outside the selected energy window, so truncation induces Gibbs ringing; non-Poisson noise such as electronic baseline fluctuations or cosmic spikes violates the Poisson mixture model and requires background subtraction or parametric forward fitting (Hernández et al., 2016).

Simulation-based CCD analysis depends on faithful instrument modeling. The simulator must reproduce depletion depth, field bias, noise coefficients, thresholds, and grade rules; verification against the official RMF to within $\tau_d$40 in the zero-pile-up limit is explicitly required for retuning to other CCDs (Tamba et al., 2021). This suggests that the accuracy of nonlinear correction is bounded by the fidelity of the detector microphysics and reduction-chain emulation.

Collider soft-subtraction methods have known failure modes when hard jets bias the pileup-density estimate. The standard median estimator is sensitive to the number of hard jets in the event, which motivated signal-region exclusion and weighted-clipped estimators (Berta et al., 2023). Track-based methods are limited outside tracker acceptance, motivating calorimeter-based mixing techniques and purely calorimetric density estimators (Hautmann, 2015). Learned methods remain dependent on simulation realism; PUMiNet explicitly notes that real-data systematics, detector effects, and pile-up modeling remain to be validated within experiment frameworks (Vaughan et al., 4 Mar 2025), while PUMA assumes perfect charged-vertex assignment and no non-pile-up underlying event (Maier et al., 2021).

In heavy-ion cumulant analysis, the difficulty is not only correcting pile-up but determining the correction parameters. A model-dependent Glauber-fit extraction of the single-collision multiplicity distribution can fail closure tests even when the correction formalism itself is correct (Zhang et al., 2021). The proposed response-matrix unfolding avoids this by inferring the single-collision distribution directly from the measured piled-up distribution (Zhang et al., 2021). More broadly, residual pile-up can mimic qualitative signatures sought in critical-point searches, especially upward excursions of $\tau_d$41 at low energies (1705.01256).

A final recurring limitation is that some forms of pile-up are not best corrected by inversion at all. In SDD spectroscopy at LCLS, the reported conclusion is that pile-up spectrum fitting is relatively simple and preferable to pile-up spectrum deconvolution, especially when measurements involve several discrete photon energies known a priori (Blaj et al., 2017). This indicates a methodological divide between inverse recovery of a continuous underlying distribution and probabilistic decomposition of discrete mixtures.

7. Cross-domain synthesis and research directions

Pile-up effects admit a unifying interpretation as overlap-induced nonlinearity with three coupled consequences: additive bias, variance inflation or count loss, and structural misclassification. In X-ray spectroscopy these appear as spectral hardening, flux suppression, and grade migration (Tamba et al., 2021); in collider reconstruction as jet energy bias, spurious jets, and decorrelation of observables (Hautmann, 2015); in heavy-ion fluctuation analysis as convolutional distortion of multiplicity tails and cumulants (Zhang et al., 2021). The common statistical primitives are Poisson arrival processes, finite response windows, convolutional mixture laws, and inverse or forward mappings conditioned on detector response.

The methodological landscape is similarly convergent. Fourier-domain inversion, Monte Carlo forward folding, likelihood-based parametric extensions, per-object subtraction, event mixing, and attention-based regression all implement the same abstract program: estimate latent single-interaction structure from superposed observations. This suggests that transfer of ideas across subfields is plausible. For example, the distinction between soft diffuse contamination and hard contaminating objects that motivated jet mixing and signal-region exclusion at the LHC has an analogue in photon detectors, where non-pile-up electronic background must be separated from true pile-up before inversion (Hautmann, 2015, Hernández et al., 2016). Likewise, detector-accurate forward simulation in X-ray CCD analysis resembles response-matrix approaches to pile-up unfolding in heavy-ion cumulants (Tamba et al., 2021, Zhang et al., 2021).

Current research directions reflect increasing reliance on contextual inference. Attention-based models exploit event-wide information for pile-up mitigation at HL-LHC conditions (Vaughan et al., 4 Mar 2025, Maier et al., 2021), while JUNO applies CNNs and transformers to waveform-level discrimination of $\tau_d$42C overlaps (Fang et al., 2 Mar 2026). A plausible implication is that future pile-up mitigation will increasingly combine physics-based priors with learned representations rather than replacing one with the other. The existing literature already points toward hybrid workflows: simulation-tuned inference in CCDs (Tamba et al., 2021), data-driven mixing with classical soft removal at colliders (Hautmann, 2015), and ML-estimated double-pulse resolution followed by analytic Lambert-$\tau_d$43 correction in MAPMT counting (M'sihid et al., 28 Nov 2025).

Pile-up effects therefore remain less a single phenomenon than a family of overlap problems whose practical treatment depends on which degrees of freedom are unresolved: time, energy, position, vertex, or event identity. The underlying challenge is stable inference under superposition. Across the cited domains, progress has come from making that superposition explicit rather than treating it as a nuisance perturbation.