When a particle experiment sees a new bump in a mass spectrum near the energy where two known hadrons could be produced together, the bump alone often cannot tell physicists what the object is. It could be a tightly bound four- or five-quark state, a loose "molecule" of two ordinary hadrons, a virtual state (an attraction too weak to bind), or no particle at all: a cusp or triangle effect created by the kinematics of a nearby threshold. Different stories fit the same data. BESIII's $X(1835)$ line shape was fit equally well by a single molecular-like state and by two interfering resonances (Achasov et al.). This proposal treats the question "which observables would break this tie?" as a feature selection problem. We search over subsets of measurable quantities, such as energy windows in each decay channel, ratios between channels, and angular moments. The goal is the smallest set that lets a classifier name the correct structure, and that keeps doing so when the physics model used to train the classifier is wrong.
The ingredient that matters most is the scoring rule. A candidate subset of observables is scored by its worst-case classification accuracy across several different families of amplitude models: Flatté, K-matrix, coherent sums of resonances, and non-pole threshold cusps. It is not scored by accuracy under a single model. The closest existing work, Deep learning exotic hadrons, trains a deep network on $10^5$ line shapes from one two-channel amplitude model and applies it to the LHCb $P_c(4312)$ data, where it favors a virtual state on Riemann sheet IV. It feeds in the full binned spectrum, explains the result afterwards with per-bin attributions, and inherits that one model's assumptions. We instead select observables under deliberate model uncertainty, so the output is a short list of measurements whose discriminating power does not depend on which amplitude family is true.
Core Claim
A selection loop that uses the classifier itself as the judge (a "wrapper") can do two things when its objective is worst-case accuracy over amplitude families and it is made cheap enough to search a structured combinatorial space. First, it can find small sets of observables that classify the pole structure as accurately as the full spectrum on the families seen during training. Second, it can find sets that are clearly more robust than the full spectrum on a family held out entirely. Writing $S$ for a subset of candidate observables, $c(S)$ for its measurement cost (number of channels or windows an analysis must control), $B$ for a cost budget, and $A_g(S)$ for validation accuracy of a classifier restricted to $S$ on synthetic data from amplitude family $g$, the search solves
$$S^* = \arg\max_{S:\, c(S)\le B} \; \min_{g \in \mathcal{G}} A_g(S).$$
The max-min form addresses a failure mode the research file documents twice. Classifiers latch onto features specific to the training environment. Robust Neural Particle Identification Models shows large error increases when LHCb particle-identification networks meet unseen decay channels. Population-Robust Feature Selection via Generalized Welfare Optimization shows that a single feature subset serving several subpopulations should be chosen with a welfare-style objective rather than an average. We borrow that second idea and treat amplitude families as the "populations." We use 0-1 accuracy rather than mutual information as the score because Ahmed, Narasimhan and Agarwal prove that mutual-information filters can prefer subsets with higher classification error. The physics question is a discrete call (bound, virtual, or non-pole), so the operational loss should be the selection loss.
Research Program
The first component is a hypothesis-diverse simulator. For the $P_c(4312)$ system we start from the two-channel ($J/\psi p$, $\Sigma_c^+\bar D^0$) setup of Ng et al. and add three families that describe the same pole classes differently: a K-matrix parameterization, a Flatté parameterization, and an effective-range expansion. We also add a fourth label for non-pole structures produced by threshold cusps. This follows the open problem in the research file that kinematic singularities can mimic resonances. For the $X(1835)$ system we simulate the two hypotheses BESIII could not separate, a Flatté state with a large $p\bar p$ coupling and a coherent sum of two resonances. Each family is sampled over broad parameter ranges and smeared with detector-like resolution and Poisson noise at the event counts of the real experiments.
The second component is the candidate observable pool, kept structured rather than arbitrary. It contains contiguous energy windows in each channel at several widths, ratios of yields between coupled channels inside matched windows, low-order angular moments where the decay allows them, and the location and size of slope changes at each threshold. Structure matters for the combinatorics. Windows are intervals, so the search can move window edges rather than flip $2^N$ bits. Channels act as natural groups, so the cost $c(S)$ counts channels and windows the way an analysis team would.
The third component makes the wrapper affordable. Retraining a classifier for every candidate subset is the standard bottleneck of wrappers (Feature Selection for Classification). We instead train one network with random input masking, so a single model can score any subset by zeroing the unused inputs and supplying a mask indicator. Search uses successive halving in the style of OpenFE's first pruning stage (OpenFE): many candidate subsets are scored on small validation samples, and only the survivors earn larger ones. The final few subsets are retrained from scratch to confirm their scores. Wrappers also overfit the validation split they query repeatedly (Malarvizhi). To guard against this, we keep a separate test split per family and one entire amplitude family that the search never sees. Stability of the selected subsets is measured across bootstrap resamples, since unstable wrapper subsets would undercut any physical reading of them.
Baselines
The naive reference point is a classifier trained on the full binned spectrum from a single amplitude family, which is the Ng et al. setting. It tells us how much accuracy collapses when the truth comes from a different family. The file reports the qualitative outcome on LHCb data (a virtual state on sheet IV is overwhelmingly favored, with robustness to noise up to 5%). It gives no test-set accuracy number, so for this baseline there is no reported result found on Lacuna for synthetic classification accuracy, and we will measure it ourselves under our simulator.
The strongest existing competitor is a full-spectrum network trained on the pooled mixture of all training families, with every bin and channel as input. It has the same architecture, training budget and hyperparameter search as our masked network. This is the real threat. If pooling alone yields robustness, selection only adds interpretability. Beating it means one of two outcomes on the held-out family. Either a selected subset at a cost of at most half the channels and windows matches its accuracy within bootstrap error, or a selected subset exceeds its accuracy by a margin larger than the bootstrap spread.
Two selection baselines run at equal subset size $k$ and an equal evaluation budget. The first takes the top-$k$ observables by SHAP attribution from the pooled network, as Ng et al. did for interpretation. The second is the 0-1 BayesGreedy filter of Ahmed et al., which relies on pairwise approximations. Its weakness here is predictable, because channel ratios only matter jointly with the windows they connect. We also run a classical baseline: a likelihood-ratio test between fitted Flatté and coherent-sum amplitudes on the full spectrum, which reproduces the BESIII ambiguity.
Two ablations isolate the key ingredient. The first swaps the worst-case objective for average accuracy across families while keeping everything else fixed. The second replaces masked-network scoring with full retraining on a subsample of candidates, which tells us whether the cheap scorer ranks subsets correctly.
For the secondary sanity check of the selection machinery we use the HIGGS benchmark. There, Baldi et al. report AUC 0.73 for shallow models on low-level features, 0.78 with engineered invariant masses, and 0.88 for a deep network on low-level features. A wrapper choosing among the 28 features with a shallow model should recover the engineered-feature level with few features. If it cannot, the search engine is suspect before any hadron conclusion is drawn.
Why not just…?
Why not just train the full-spectrum network on all amplitude families and skip selection? This may well match our accuracy on families seen in training, and we expect it to. The open question is behavior on an unseen family, where a full-input network can lean on bins whose meaning differs between models. The held-out-family experiment settles it directly. If pooled training matches the selected subsets there, the honest contribution shrinks to measurement economy: telling experiments which few channels and windows carry the decision. That result is still useful for planning analyses at BESIII or LHCb, but it is weaker, and we would report it as such.
Why not just read the important bins off SHAP values, as Ng et al. already do? Attribution tells you how one trained model spreads credit. It does not tell you which subset is best to measure. Neighbouring bins in a smooth line shape are strongly correlated, so attribution splits credit among them and can rank several redundant bins above a single informative ratio from another channel. The equal-$k$ comparison between SHAP-top-$k$ and wrapper subsets tests this. If SHAP subsets are as accurate and as robust, the wrapper is unnecessary, and we would say so.
Why not just use a fast filter such as mutual information or BayesGreedy? Ahmed et al. show that mutual-information filters can be misaligned with 0-1 error. Their own Bayes-error criterion relies on pairwise approximations, and the research file flags that these drop higher-order joint effects. Coupled-channel ratios are exactly such effects. The filter baseline runs at equal $k$, so the gap, or its absence, is measured rather than assumed.
Why not just fit amplitudes more carefully? Careful fits on the full $X(1835)$ spectrum already produced two equally acceptable hypotheses (Achasov et al.). Fitting does not say which additional measurement would separate them. Our selected observables could feed straight back into a conventional fit. We see the two approaches as complementary.
Proposed Experiments
The first decisive check is leave-one-family-out evaluation on the $P_c(4312)$ simulator, rotating which family is held out. We report worst-family accuracy, mean accuracy, and subset cost for every method above, with bootstrap intervals and subset-stability scores. The claim survives if the worst-case wrapper subsets beat SHAP-top-$k$ and BayesGreedy at equal $k$ on the held-out family. It also requires them to match or beat the pooled full-spectrum network at clearly lower cost. If the average-objective ablation does just as well, the worst-case objective is not doing the work.
The second check targets the BESIII ambiguity. We simulate both $X(1835)$ hypotheses at the published $1.09\times10^9$ $J/\psi$ statistics. We report whether any selected observable set, for example a matched $\eta'\pi\pi$ window together with a $p\bar p$ yield ratio, separates them where the full-spectrum likelihood ratio does not. We also estimate the event count each subset would need to separate them reliably. Either outcome is informative: a positive result names a measurement to make, and a negative one says the ambiguity is irreducible with these channels.
The third check reapplies the robust classifier, restricted to its selected subset, to the three LHCb $P_c(4312)$ datasets used by Ng et al. We ask whether the virtual-state reading survives once non-pole cusp shapes and alternative amplitude families are part of training. Agreement would strengthen that physical conclusion. Disagreement would show how much it depended on one parameterization, which we consider the more important finding.
The HIGGS sanity check runs first and gates the rest. Search cost is reported throughout as the number of subset evaluations and GPU hours, alongside the full-retraining validation of the masked scorer's rankings.
References
- Observation of an anomalous line shape of the $\eta^{\prime}\pi^{+}\pi^{-}$ mass spectrum near the $p\bar{p}$ mass threshold in $J/\psi\rightarrow\gamma\eta^{\prime}\pi^{+}\pi^{-}$
- Deep learning exotic hadrons
- Robust Neural Particle Identification Models
- Population-Robust Feature Selection via Generalized Welfare Optimization
- Bayes Optimal Feature Selection for Supervised Learning with General Performance Measures
- Feature Selection for Classification
- OpenFE: Automated Feature Generation with Expert-level Performance
- An Extensive Analysis of Feature Selection in Machine Learning
- Searching for exotic particles in high-energy physics with deep learning