Many rare-event searches in particle physics rest on a handful of events. A typical analysis fits a smooth signal-plus-background model to the exact measured coordinates of every event and reports a yield and a significance. The CLEO II search for τ→μγ, for example, reported 1.8 fitted signal events at 1.0σ. That leaves a question analysts rarely answer exactly: how many individual events would have to be removed before the conclusion changes? Removing two events might erase a claimed 3σ excess. A detector glitch might be holding up an upper limit. This proposal computes that number, which we call the deletion-robustness number κ, exactly rather than estimating it. Where exact computation is too costly, it brackets κ with a guaranteed interval.
The tool comes from parameterized algorithms. There, iterative compression solves "delete at most $k$ items to reach property Π" by starting from a valid solution of size $k+1$ and trying to shrink it. The oversized solution makes the smaller search tractable. We adapt that step to a setting it was never built for: a continuous likelihood, where no event is ever cleanly "violating." The opening claim is specific. In rare-event regimes, an incumbent removal set of size $k+1$ yields a provable influence threshold. Any size-$k$ removal set must draw from a small pool of events that clear that threshold. This turns a fragility estimate into a certificate.
Core Claim
The setting is an unbinned extended maximum likelihood fit, meaning one that evaluates continuous signal and background densities $P_s, P_b$ at each event's coordinates $\vec{x}_i$ and multiplies by a Poisson term for the total count. We test it on the 2D ΔE–Δm fit of the CLEO τ→μγ search (Update of the search for the neutrinoless decay τ → μ γ). Let $q_0(A)$ be the profiled discovery test statistic computed on event set $A$. It is twice the log-likelihood ratio between the best signal-plus-background fit and the best background-only fit, with yields and shape nuisances $\vec\theta$ refitted. For a threshold $z$ (say 3σ), define
$$\kappa = \min\{|S| : S \subseteq \{1,\dots,N\},\; q_0(\{1,\dots,N\}\setminus S) < z^2\},$$
the smallest number of events whose removal pushes the significance below $z$. The analogous definition for an upper limit asks how many events must be removed to tighten it by a stated factor.
Computing κ is a combinatorial min–max problem: a minimum over removal sets of a maximum over continuous parameters. It cannot be read off any single fit. Our claim is that for small κ, which is exactly the regime where fragility matters, κ can be computed exactly at practical cost. The method combines two ingredients. The first is a relaxation, in which events receive weights $w_i\in[0,1]$ instead of being kept or dropped. This gives a sound lower bound on κ. The second is a compression step: given a removal set $S'$ of size $k+1$ that already succeeds, we either find a size-$k$ set or certify that none exists.
Why iterative compression does not transfer directly, and what replaces it
The classical method, laid out in the Iterative Compression monograph and used to settle directed feedback vertex set and almost-consistent linear systems, depends on two facts. First, the target property is inherited by sub-instances, so items can be added one at a time. Second, after the oversized solution is deleted, what remains is "clean" (acyclic, consistent). Every leftover violation must therefore pass through the guessed discarded part.
Neither fact holds for a likelihood. Significance is not inherited: a prefix of the events can be more or less significant than the whole sample. And no residual sample is ever clean, because every event contributes a nonzero log-density.
We therefore keep only the disjoint-compression step and drop the element-by-element induction. The procedure runs as follows. Fix an incumbent $S'$ with $|S'|=k+1$, found by a heuristic attack. Branch over the $2^{k+1}$ possible intersections $S_1 = S'\cap S^*$ with a hypothetical size-$k$ solution $S^*$. Within each branch, the events of $S_2 = S'\setminus S^*$ are forced to stay in the sample. The remaining $k-|S_1|$ removals must come from outside $S'$.
The replacement for "cleanness" is a quantitative gap. The weighted relaxation is convex in the yields when shapes are fixed, and it is handled by bounded box search over the few shape nuisances otherwise. In each branch it bounds how much $q_0$ must still fall. It also bounds how much any single event outside $S'$ can lower $q_0$; we call this the event's influence bound. By pigeonhole, at least one remaining removal must have an influence bound of at least (required drop)/(remaining budget). Only events above that threshold need to be branched on.
In rare-event searches, very few events sit where $P_s/P_b$ is large, so this candidate pool is small. In the CLEO fit only a handful of events lie near the signal region. That smallness is the source of tractability. The parameter that controls the cost is κ together with the size of the pool, not the total number of events $N$.
Research Program
The first deliverable is the influence-bound lemma and its proof for two cases. In the first, only the yields float and the shapes are fixed; here the relaxation is a convex program. In the second, a few shape parameters float, such as a resonance mass, a width, or a background slope; here we bound influence uniformly over a box by Lipschitz arguments and refine the box adaptively.
The second deliverable is the full certification loop. A cheap heuristic attack produces the incumbent. Options include greedy one-at-a-time removal of the highest-influence event, and trimmed-likelihood refitting in the style of ITMLE, which repeatedly drops the lowest-likelihood points and refits. The incumbent is then compressed downward until a branch-wide certificate shows that no smaller set works. The output is either exact κ or an interval $[\kappa_{\text{lb}}, \kappa_{\text{ub}}]$ when the time budget expires.
The third deliverable is an empirical study of what κ reveals. We test whether excesses created by detector glitches or by classifier-induced background sculpting, the failure mode documented by Das et al., have a systematically smaller κ than genuine injected signals of the same significance.
This work differs from three neighbouring lines.
- Compressed maximum likelihood (Compressed Maximum Likelihood) compresses the data representation to gain sample efficiency, and leaves its NP-hard optimization to EM heuristics. We compress a candidate removal set to gain an exactness guarantee, and we answer a robustness question rather than an estimation question.
- ITMLE gives minimax error rates for trimmed estimators in generalized linear models. It is a local alternating method with no certificate. We use it only as an incumbent generator and as a baseline.
- BASiL (D'Amico et al.) motivates the continuous-likelihood side of the work: it shows that single-event likelihoods beat hard cuts by 15–20% in precision. Its exact $2^N$ sum over signal/background assignments also constrains our design. Any exact method must avoid scaling with $N$, which is why our cost is tied to the candidate pool.
Baselines
Naive reference. We report the fitted significance alone, plus a leave-one-out jackknife. The jackknife removes each event in turn and records the largest drop in $q_0$, which gives κ = 1 or "greater than 1" and nothing more. No result is needed to beat it. It defines what analysts currently see.
Greedy influence removal. Repeatedly remove the event whose deletion lowers the refitted $q_0$ most. This is the obvious heuristic and gives an upper bound on κ. We beat it if, on instances where brute force gives ground truth, greedy overestimates κ on a non-trivial fraction of instances and our method recovers the true value.
Multi-start trimmed likelihood. This is the ITMLE procedure with many random restarts. It is the strongest published heuristic for removing contaminating points. On label-corrupted Gaussian regression, Awasthi et al. report error $O(\sigma\epsilon\log(1/\epsilon))$ as a minimax-matching rate; that is a rate, not a benchmark score. Our method receives the same wall-clock budget. Beating this baseline means certifying κ exactly where trimming either misses the optimum or, by construction, cannot prove optimality.
Generic branch-and-bound with the identical relaxation. We pose κ as a mixed-integer program over event weights and solve it by branch-and-bound using the same bounding relaxation. This baseline isolates the key ingredient. If it matches our runtime and certification rate, the compression step adds nothing beyond the relaxation.
Ablation. Our full method runs with and without the incumbent-derived influence threshold. Without it, the method enumerates the top-$M$ influence events for a fixed $M$ and has no exactness guarantee. This ablation measures how much the threshold contributes to certification.
Reference numbers. No reported result found on Lacuna for exact or approximate deletion robustness of unbinned likelihood fits. The physics anchors we reproduce on simulated analogues are CLEO's 1.8-event, 1.0σ fit with a 90% C.L. limit of 1.0×10⁻⁶ (CLEO), and the over-5σ significance reported for CWoLa and CATHODE on the LHC Olympics R&D set (Das et al.).
Why not just…?
Why not just brute-force the top few most influential events? This is the most serious alternative. In tiny samples it will often find the right answer. Its weakness is that it cannot know when it has found it, because the right pool size $M$ depends on the instance. Our influence threshold is exactly the rule that sets $M$ with a guarantee. The ablation above settles the question. If a fixed $M$ chosen in advance matches exact κ on nearly every instance, we will say so and reduce the contribution to the certificate itself.
Why not just use generic mixed-integer branch-and-bound? This is a real threat. With a good relaxation, a generic solver may do as well as compression. The head-to-head comparison at equal relaxation and equal compute decides it. We expect compression to help when the heuristic incumbent is already near-optimal, which is common in rare searches, because the incumbent fixes the pruning threshold before any search begins. We do not assume this. If generic branch-and-bound wins, the paper's claim becomes "the relaxation certifies κ," and the compression framing is dropped.
Why not just add a flat outlier component to the likelihood, or apply hard cuts? Both change the estimate; neither answers the robustness question. An added outlier term still yields a significance whose dependence on individual events is unknown. Hard cuts discard signal exposure, as BASiL shows. We will nonetheless compute κ for fits that include an outlier component, to test whether the component already makes glitch-driven excesses robust. That is a useful finding in either direction.
Why not just bin the data, where κ becomes a small integer program over bin counts? Binning makes events within a bin interchangeable, so exact κ is easy. However, the binned κ refers to a different and less sensitive fit, the kind that the CLEO-c Dalitz analysis avoids by fitting unbinned. We compute both and report how often binning misstates κ for narrow signals.
Proposed Experiments
Ground truth on CLEO-style toys. We generate toy samples matching the 2D ΔE–Δm setup of the CLEO search, with $N$ chosen so that exhaustive enumeration up to κ = 4 is feasible. We compare exact, greedy, trimmed, and branch-and-bound values of κ, along with runtime and certification rate.
Scaling on the LHC Olympics R&D set (LHCO 2020). We inject Z′ signals at controlled strengths and apply classifier selections at background efficiencies of $10^{-3}$ and $10^{-4}$, the rejection levels where Das et al. report that curve fits break down. We then fit the dijet mass unbinned and compute κ. The decisive check is whether excesses created purely by sculpting, in background-only selections and in the Black Box 2 analogue, have smaller κ than injected signals of equal significance. If they do not, κ is a calibration tool and not a discriminator, and we will report it as such.
Glitch injection. We plant 1–3 events where the background density is nearly zero, the failure mode the research file identifies for unbinned fits, and test whether κ flags them.
Stress test at higher event counts. We run on a Dalitz toy at the CLEO-c scale of 14,400 events and ask how many events must be removed to drop the 4.3% $f_0(1370)\pi^+$ component below 3σ. We expect κ to be large here and the candidate pool to grow. We will report where exact certification stops working and fall back to the interval bounds. The outcome shows how far the small-pool assumption, which drives the whole approach, actually holds.
References
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Iterative Compression
- Directed Feedback Vertex Set Problem is FPT
- Almost Consistent Systems of Linear Equations
- Trimmed Maximum Likelihood Estimation for Robust Generalized Linear Model
- Accurate and robust methods for direct background estimation in resonant anomaly detection
- Compressed Maximum Likelihood
- Signal estimation in on/off measurements including event-by-event variables
- Dalitz plot analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msubsup><mml:mi>D</mml:mi><mml:mi>s</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
- The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics