Searches for new particles often look for a small bump in a mass spectrum. The background under that bump is estimated by taking a simulation of known physics and correcting it with real data from neighboring mass windows, called sidebands, where no signal is expected. The weak point is the correction step. A flexible correction can quietly absorb a real signal. It can also invent structure that was never in the data, a false peak. Physicists currently guard against both by attaching a single empirical uncertainty, for example the gap between two different simulators. That number has no principled link to how much the correction was actually allowed to move.
This note proposes to borrow the rule that logic-based reasoning systems use when they update beliefs. That rule, the minimal change principle, says to change prior beliefs only as much as the new evidence requires. A refinement of it, conservative revision, adds a second rule: when the evidence does not settle a question, do not pretend it does. Translated into a background estimate, the procedure is as follows. Among all corrections to the simulated background that fit the sideband data, keep only those that stay nearly as close to the simulation as the closest possible fit. Report the full range of background counts those corrections predict in the signal window, rather than a single number. Claim a signal only if the observed excess survives every correction in that set. The aim is a background interval whose width comes from the data and the simulation, and which holds its stated coverage when the simulation is wrong in ways no single simulator comparison anticipates.
Core Claim
The belief-revision literature has a precise diagnosis of what goes wrong when an update picks one "closest" state. Conservative Belief Revision shows that revising by weak or disjunctive evidence and selecting the single nearest world produces "ghost beliefs": commitments the evidence never supported. Their example is concluding that exactly one person is home when told that at least one is. Their fix is to take the most plausible world for every way the evidence could be realized, and to believe only what all of those worlds share. Sideband data is weak evidence of exactly this kind about the signal window: many corrections fit it equally well and disagree in the window.
A single best-fit correction, whether a fitted curve or one learned reweighting, is the ghost-belief move. It commits to one signal-window background the sidebands do not determine. Implementing Bounded Revision via Lexicographic Revision and C-revision supplies the second ingredient. It implements bounded revision ("accept the new information as far as a reference bound, and a little further") with ranking functions, which gives a natural way to say how far beyond the minimal fit a correction may move.
Both papers are purely propositional, and we do not claim to carry their theorems into continuous densities. We use them as a design specification. The probabilistic counterpart of minimal change is the minimum relative-entropy update of the simulated background. The counterpart of conservative revision is the set of corrections within a bounded excess divergence of that minimum, together with the envelope of their signal-window predictions.
The claim to test is this: the envelope gives background intervals with close to nominal coverage under simulator mismatch, at a sensitivity cost small enough that discovery reach on standard benchmarks is preserved.
Research Program
Let $p_{\mathrm{ref}}(x, m)$ be the simulated background over event features $x$ and mass $m$. A correction is a reweighting $w_f = e^{f(x,m)}$, where $f$ is restricted to vary smoothly in $m$. A low-order dependence on $m$ times a flexible function of $x$ is one option; a Lipschitz bound in $m$ is another. Smoothness in $m$ is the assumption that lets sideband information reach the signal window, and we make it explicit rather than hiding it inside a network.
Let $D(f)$ be the relative entropy between the reweighted and the original simulation. Let $\mathcal{F}_{\mathrm{SB}}$ be the corrections that fit the sideband data within statistical tolerance. Write $D^\ast = \min_{f \in \mathcal{F}_{\mathrm{SB}}} D(f)$ for the size of the smallest correction the sidebands force; this minimum is the minimal-change correction. The conservative set is
$$\mathcal{F}_\alpha = \{\, f \in \mathcal{F}_{\mathrm{SB}} : D(f) \le D^\ast + \alpha \,\},$$
where $\alpha \ge 0$ is the bounded-revision allowance. For any anomaly selection $S$, the background prediction is the interval
$$\big[\, \min_{f\in\mathcal{F}_\alpha} N_S(f),\ \max_{f\in\mathcal{F}_\alpha} N_S(f) \,\big],$$
where $N_S(f)$ is the expected background count passing $S$ in the signal window under correction $f$.
The construction does two things at once. Setting $\alpha=0$ recovers the single minimal-change correction, the ghost-belief case. The width of the interval then replaces the systematic shift parameter of Accurate and robust methods for direct background estimation in resonant anomaly detection. That paper estimates its shift from Pythia-versus-Herwig differences or from sideband calibration. It does not derive the shift from a constraint on how the template may deform, which is the gap we target. The discovery test statistic is the minimum significance over $\mathcal{F}_\alpha$, so a signal is claimed only if no admissible background explains it.
We will build three components. The first is a kernel-based fit of $f$, reusing the fast kernel solver behind Robust resonant anomaly detection with NPLM, which estimates log density ratios between data and a reference with kernel methods. The second computes the two extremes by adding $\pm\lambda N_S(f)$ to the constrained objective and sweeping $\lambda$, which makes each extreme a convex problem in the kernel setting. The third chooses $\alpha$ without touching the signal window. We will slide a pseudo-signal-window across the sidebands, treat its outer neighbors as sidebands, and pick the smallest $\alpha$ whose intervals cover the held-out pseudo-window counts at the nominal rate. This mirrors bounded revision's "a little further" with a data-derived bound.
NPLM's end-to-end mode needs an accurate, fixed reference model, and that paper names reference fragility as its main limitation. The conservative set offers a principled corrected reference to feed it.
Baselines
The naive reference point is the uncorrected simulation used directly as the signal-window background. This shows how much simulator mismatch alone costs in coverage and false discoveries.
The strongest direct competitor is the direct background estimation of Accurate and robust methods for direct background estimation in resonant anomaly detection. It counts events in the signal window against the background template that CATHODE or CWoLa produce internally, with a systematic shift taken from simulator differences or sidebands. The paper reports discovery significance over 5σ for CWoLa and CATHODE at background efficiencies of $10^{-3}$ to $10^{-4}$ on the LHC Olympics data, against 8σ for an idealized detector. It also reports correctly low significance on background-only tests, though the file gives no number for that check. We will rerun it on identical data splits with the same selection classifier and the same tuning budget, using both its Monte Carlo and its data-driven shift estimates.
The second strong competitor is the NPLM classifier with hyper-test of Robust resonant anomaly detection with NPLM. It scans selection thresholds, takes the minimum p-value, and calibrates with signal-free pseudo-experiments. It reports a median Z of 0.30 ± 0.04 at signal-to-background $8.2\times10^{-4}$, against 0.1 ± 0.1 for the standard boosted-decision-tree approach. We will run its hyper-test on top of each background estimate so the comparison isolates the background step.
The key ablation is $\alpha=0$, the single minimal-change correction with no envelope. If our method beats the baselines only because of the kernel reweighting, and not because of the conservative set, this ablation will show it. A second ablation removes the smoothness-in-$m$ restriction, to show how much signal an unconstrained correction absorbs.
Beating the baselines means two things together. First, our intervals must cover the true signal-window background at close to the nominal rate under simulator swap, where the shift-parameter baseline either under- or over-covers. Second, median injected-signal significance must stay within a modest fraction of the direct-estimation result at the same false-positive rate on background-only data.
Why not just…?
Why not just profile a few hand-picked nuisance parameters, as standard analyses already do? That is a real alternative, and our set is a nonparametric relative of it. The difference is what fixes its size. Hand-picked shape parameters cover only the mismodeling directions someone thought of. Here the admissible directions are every smooth correction the sidebands allow, and their size is tied to the simulation by the divergence bound. The simulator-swap experiment settles the question: if a two- or three-parameter profiled fit covers as well at equal width, the extra machinery is not justified, and we will report that.
Why not just use the Pythia-versus-Herwig difference as the uncertainty, as Das et al. do? One simulator pair gives one direction and one magnitude of mismodeling. It can understate errors the pair happens to agree on, and overstate errors where they disagree for reasons irrelevant to the data. We test this directly by calibrating on one direction of the swap and evaluating coverage on the other. The shift-parameter method should do well when the test mismatch resembles its calibration mismatch, and degrade otherwise.
Why not just take the spread of an ensemble of CATHODE flows? Ensemble spread measures training randomness, not the range of backgrounds consistent with the sidebands. Accurate and robust methods for direct background estimation in resonant anomaly detection reports that flows interpolate poorly in low-density tails, where every member can be wrong in the same way. We include a flow ensemble as an extra baseline, and expect it to undercover precisely at tight selections.
Why not just decorrelate the classifier from mass, as in multi-background representation learning? Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning uses an adversarial critic to reduce mass sculpting, reaching a Jensen–Shannon divergence of 0.118 against 0.255 for an autoencoder. Sculpting persists, and the approach needs predefined background classes. Decorrelation addresses a different failure, the selection distorting the spectrum, and it complements our method rather than replacing it. We will report its sculpting divergence with and without our estimate, using the hls4ml jet data as a side check.
Proposed Experiments
The main testbed is the LHC Olympics 2020 R&D dataset from The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics. We use Pythia background as the reference and Herwig as "data", and the reverse, with signal injected at a ladder of strengths that includes $8.2\times10^{-4}$ to match NPLM.
Three measurements decide the question. The first is coverage of the true signal-window background count across hundreds of background-only pseudo-experiments, compared between the envelope and the shift-parameter interval. The second is the false-discovery rate on Black Box 2, which is known to contain no signal. The third is median significance on injected signals and on Black Box 1. Black Box 3, which most methods missed, is reported honestly: we do not expect a background estimator to rescue a weak selection, but its false-positive behavior there is informative.
The decisive real-data check uses the curated CMS 2016 DoubleMu slice from Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data. There, CATHODE raised the Upsilon significance from a raw 1.6σ to 6.4σ, and a same-sign muon control correctly found nothing. Our method must keep the same-sign control null. Its envelope-minimum significance on the opposite-sign sample then shows what fraction of that 6.4σ survives once every sideband-consistent background is admitted.
The proposal is falsified in any of three cases: the $\alpha=0$ ablation covers as well as the full envelope, a low-dimensional profiled fit matches it, or the envelope makes the Upsilon fall far below discovery level while the shift-parameter baseline holds coverage.
References
- Conservative Belief Revision
- Implementing Bounded Revision via Lexicographic Revision and C-revision
- Accurate and robust methods for direct background estimation in resonant anomaly detection
- Robust resonant anomaly detection with NPLM
- Robust anomaly detection for particle physics using multi-background representation learning
- The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics
- Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data