Treating Detector Inefficiency as a Missing-Modality Systematic

Particle detectors are multi-sensor instruments. A B-factory detector such as BABAR combines a silicon and drift-chamber tracker, a Cherenkov particle-identification system, a crystal calorimeter, and a muon system. Any one of these can fail to record a particle that was really there, and the calorimeter can also record energy that no particle deposited. Machine-learning work on "missing modalities" (datasets where some input channels are absent for some samples) mostly asks how to fill the gaps. Precision physics has a different concern. What limits the final measurement is usually not the missing entry in a single event but the fact that simulation gets the rate of missing and spurious entries slightly wrong. BABAR's search for $B \to K\nu\bar\nu$ carried a 14% signal-efficiency systematic from tracking alone (Search for the rare decay B → Kνν̄). The $B^+ \to \tau^+\nu$ search was limited by mismodeling of $E_{\text{extra}}$, the leftover calorimeter energy that fake clusters and lost tracks corrupt (Search for B+ → τ+ν decays with hadronic B tags).

This note proposes treating the parameters of the missingness process as nuisance parameters, meaning quantities the physicist does not care about but must account for: per-subsystem track-loss probability, identification inefficiency, and the rate of fake neutral clusters. The event classifier would be built to be insensitive to those parameters, rather than to reconstruct what is missing. Concretely, the classifier reads each subsystem as its own set of detector objects, with absence represented natively and never imputed. During training the dropout and fake-insertion rates are resampled across their calibrated uncertainty band. An adversary then tries to recover those rates from the classifier's output. The measure of success is the physics number that matters: the expected upper limit on a rare branching fraction once the missingness systematics are included in the fit.

Core Claim

The mainstream missing-modality answer is imputation: reconstruct absent views from present ones. For detector data this answer is wrong, for two reasons. First, much absence is physics rather than failure. Neutral particles leave no tracks, and soft charged particles curl before reaching the calorimeter. Imputing those channels invents measurements. Second, and more important, imputation turns an efficiency mismodeling (a normalization error that is easy to profile) into a shape distortion of exactly the variables that define the signal region, such as $E_{\text{extra}}$ and missing mass. That kind of distortion is much harder to control.

We claim that a mask-native, set-structured classifier made pivotal with respect to missingness-rate nuisances will give a smaller total uncertainty on the measured branching fraction than three alternatives: the published BDT approach with standard profiling, imputation-based multi-view networks, and flat-vector adversarial pivoting. We further claim that the gain comes from reducing the classifier's response to the missingness mechanism, not from extra raw discrimination.

Relation to Prior Work

The method's invariance mechanism comes from Learning to Pivot with Adversarial Networks. That paper trains a classifier against an adversary that tries to infer a continuous nuisance from the classifier's output, so the output distribution becomes independent of the nuisance. It reports that giving up some nominal accuracy raised Approximate Median Significance (a standard discovery-sensitivity score) from about 7.0 to about 7.8 under systematic variation. Its nuisances shift a fully populated flat feature vector. Ours instead delete and insert elements of subsystem-level object sets. That changes both the architecture needed and whether a pivot is achievable at all. The same paper also constrains the design: it shows that no optimal pivot exists when the nuisance is intrinsically correlated with the label. Track loss correlates with signal topology, so we expect an explicit power–invariance trade-off and will measure it rather than assume it away.

Anomaly-preserving contrastive neural embeddings builds contrastive "views" of LHC events by artificially masking particle features. It reaches 12.1σ in a model-independent search where raw 57-dimensional inputs give 0σ. This motivates masking as a representation tool. Their masks, however, are augmentations with no physical rate and no systematic attached. Ours are tied to calibrated inefficiencies and enter the likelihood.

Three multi-view papers define the imputation paradigm we test against. TACVI-Net compresses views with an information bottleneck before cross-view decoding to avoid "hallucinated noise." It also reports that models are more sensitive to missing inputs than to missing labels. Biased Incomplete Multi-View Learning shows that missingness is often class-dependent and imputes by sampling Gaussians rather than point values. Both of these motivate the proposal, because they establish that missingness is informative and that deterministic filling produces overconfidence. We depart from both by refusing to impute. DC-SPAN separates shared from view-private latent factors and fuses them with a product of experts. It explicitly assumes complete views. We borrow its per-view encoder with product-of-experts fusion because that design drops absent views naturally, but we do not use its clustering objective.

Among the BABAR analyses, B → Kνν̄ (20 bagged decision trees over 38 variables) and B+ → τ+ν (Search for B + → τ + ν decays with hadronic B tags) supply the physics task, the systematic budget to beat, and the evidence that inefficiency-driven terms dominate. The ratio method in B̄ → Λc+Xℓν̄ reminds us that cancellation by construction is the incumbent systematic-control tool. Our method has to earn its place on top of that tool, not replace it.

Research Program

The input representation is one variable-length set of reconstructed objects per subsystem: tracks with vertex and $dE/dx$ information, Cherenkov identification hypotheses, calorimeter clusters, and muon-system hits. Each set also carries links to the objects it matches in the other sets. Each set passes through its own permutation-invariant encoder. The resulting per-subsystem posteriors are fused by a product of experts, so an empty or absent subsystem simply contributes no factor. The tag-side B reconstruction is kept as in the published analyses.

The missingness model has a nuisance vector $\theta$. It collects per-subsystem, momentum- and angle-binned loss probabilities for true objects, plus a Poisson rate for fake calorimeter clusters. The prior on $\theta$ is set from the control-sample calibrations the BABAR analyses already quote. Training minimizes

$$\mathcal{L}(f, r) = \mathbb{E}_{\theta \sim p(\theta)}\big[\ell_{\text{cls}}(f(M_\theta(x)), y)\big] - \lambda\, \mathbb{E}_{\theta}\big[\ell_{\text{adv}}(r(f(M_\theta(x))), \theta)\big],$$

where $x$ is a simulated event, $y$ its signal/background label, $M_\theta$ the stochastic operator that drops true objects and inserts fake clusters at rates $\theta$, $f$ the classifier, $r$ an adversary predicting $\theta$ from the classifier output, $\ell$ the corresponding log-losses, and $\lambda$ the invariance strength.

The downstream fit is a binned profile likelihood on the classifier output. Nuisances for $\theta$ are propagated by rerunning $M_\theta$, so every method is charged for its own sensitivity to missingness in the same way.

Access to the preserved BABAR data and full simulation depends on the collaboration's data-preservation arrangements, and the file gives no open resource for it. The primary testbed is therefore a generated $\Upsilon(4S) \to B\bar B$ sample with a parametric BABAR-like detector response, reproducing the $B \to K\nu\bar\nu$ selection. If collaboration access is obtained, the same pipeline is rerun on official simulation and on sidebands. A secondary, fully public check injects controlled train/test shifts in missingness rate into the multi-view benchmarks the file documents (Animal, Corel5k). This tests whether imputation-based methods degrade under missingness-rate shift in ordinary machine-learning data too.

Baselines

The naive reference is the published recipe: bagged decision trees on the flat high-level variables of B → Kνν̄, with a tracking-efficiency systematic assigned by varying $\theta$ and profiling. We rerun it on our simulation with the same variables so that it yields our own expected limit. That paper reports a limit of $\mathcal{B}(B^+ \to K^+\nu\bar\nu) < 1.3 \times 10^{-5}$ at 90% confidence and a 14% efficiency systematic on real data. Those are the scale targets, not directly comparable numbers. A modern gradient-boosted tree with native missing-value routing is run alongside it, with the same hyperparameter budget.

The strongest machine-learning competitor on the invariance question is flat-vector adversarial pivoting as in Learning to Pivot, with $\theta$ as the nuisance and zero-plus-indicator encoding for absent features. Its reported gain was an Approximate Median Significance of about 7.8 versus about 7.0 for the unpivoted classifier, on jet data. The strongest missing-modality competitor is TACVI-Net-style cross-view imputation feeding the same classifier. TACVI-Net reports an average precision of 0.447 on Corel5k at 50% missingness. For Gaussian-sampling imputation (RIML) the file reports only a roughly 9.7% accuracy gain on Animal, and no reported result was found on Lacuna for any of these methods on collider data.

Two ablations isolate the key ingredient. One trains the mask-native set model with $\theta$-resampling augmentation but $\lambda = 0$, which removes the adversary. The other keeps the adversary but uses a fixed nominal $\theta$ during training.

For fairness, all methods see the same simulated events, the same tag reconstruction, the same fit, and the same nuisance priors. Each gets an equal compute budget and an equal hyperparameter search budget. Beating a baseline means a smaller expected 90% upper limit with systematics included, with the gap exceeding the spread across five training seeds. It also requires no worse bias under the stress test described below.

Why not just…?

Why not just impute the missing subsystem readings? Imputation assumes the absent value existed. For neutral particles in the tracker, or soft tracks at the calorimeter, it did not. Even for true inefficiencies, a decoder trained at nominal $\theta$ learns nominal-rate correlations, so a shift in $\theta$ moves the imputed $E_{\text{extra}}$ shape. The imputation baseline directly measures this: we record how far its classifier output shifts per unit change in $\theta$, and the same quantity for our model.

Why not just use gradient-boosted trees, which handle missing values natively? They handle absence at prediction time but have no pressure to be insensitive to the absence rate. That sensitivity is exactly what becomes the systematic. This is a real threat, because boosted trees plus profiling are strong and cheap. If they match our total-uncertainty limit, we will report it, and the conclusion becomes that the systematic is dominated by profiling rather than by representation.

Why not just augment with random dropout and skip the adversary? This is the most serious alternative. Training across the $\theta$ band may by itself flatten the classifier's response, and Louppe et al. show that adversarial invariance costs power when the nuisance correlates with the label. The $\lambda = 0$ ablation settles this question. If it matches the full model, the recommendation reduces to calibrated missingness augmentation, which would still be a practical result for archived-data reanalysis.

Why not just generate more simulation? More simulation shrinks the finite-sample term but not the uncertainty on $\theta$ itself, which comes from data control samples. We report the systematic budget split into these two components to show which one each method reduces.

Proposed Experiments

The decisive experiment compares expected 90% upper limits on the $B^+ \to K^+\nu\bar\nu$ analogue, with missingness systematics profiled, across all baselines and ablations. Each limit is reported alongside the statistical and systematic parts of the uncertainty.

The stress test generates pseudo-data at $\theta$ values up to twice the calibrated deviation from nominal and fits them with nominal templates. The resulting bias in the fitted branching fraction, and the coverage of the quoted interval, show whether a method's apparent robustness survives when the nuisance model is itself slightly wrong.

The trade-off curve sweeps $\lambda$. For each value it records nominal discrimination against the $\theta$-induced systematic, which locates the point where Louppe's no-pivot condition starts to bite. The same sweep is repeated with fake-cluster insertion only and with track loss only, to see which inefficiency the method actually tames.

The public-benchmark check trains on Animal and Corel5k at one missingness rate and tests at shifted rates. It asks whether imputation methods lose more performance under rate shift than a mask-native model, independent of any physics.

The proposal fails if the $\lambda = 0$ ablation or boosted trees with profiling match the full model on the systematics-included limit. In that case we report which simpler recipe suffices.

References

  1. Search for the rare decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:mi>K</mml:mi><mml:mi>ν</mml:mi><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>
  2. Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>τ</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>ν</mml:mi></mml:math>decays with hadronic<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>tags
  3. Learning to Pivot with Adversarial Networks
  4. Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC
  5. Task-augmented cross-view imputation network for partial multi-view incomplete multi-label classification
  6. Biased Incomplete Multi-View Learning
  7. DC-SPAN: A Dual Contrastive Attention Network for Multi-View Clustering
  8. Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:msubsup><mml:mi>Λ</mml:mi><mml:mi>c</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mi>X</mml:mi><mml:msup><mml:mi>ℓ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>ℓ</mml:mi></mml:msub></mml:math>decays in events with a fully reconstructed<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>meson