Which Measurements Could Tell a Pole from a Triangle? Learned Metrics for Exotic Hadron Line Shapes

Many recently discovered particles, the "exotic hadrons" such as the $Z_c(3900)$, $X(1835)$ and $P_c(4312)$, can be explained equally well by physically opposite stories. The peak might be a genuine new particle, a loosely bound two-particle molecule, or a kinematic artifact produced when particles rescatter (a "triangle singularity"). The data do not decide. In the coupled-channel analysis of the $Z_c(3900)$, four such scenarios all fit with $\chi^2/\text{dof} \sim 1.3$, and their differences stay below $3\sigma$ (Amplitude analysis and the nature of the Z(3900)). In our reading, this is not mainly a fitting problem. It is a geometry problem. The spectra predicted by rival hypotheses sit close together once each hypothesis is allowed to adjust its free "nuisance" parameters (mass, couplings, background shape). The standard distance between spectra, the $\chi^2$, spends most of its weight on directions that those nuisance parameters can absorb.

We propose to learn a better distance. We learn a Mahalanobis metric, meaning a weighted squared distance $(x-y)^\top M (x-y)$ with $M$ positive semidefinite, over very high-dimensional binned observables: several decay channels, several collision energies, and two-dimensional Dalitz-plot cells, reaching $d \sim 10^4$ coordinates. The metric is trained so that the rival hypotheses are pushed apart specifically at their best-fitting, most confusable configurations. Training uses leading-eigenvector (rank-one) updates, which avoid a full $O(d^3)$ matrix decomposition at every step. The resulting low-rank metric gives two things. It is a fast, amortized hypothesis classifier. More importantly, it yields a short list of observable directions (combinations of channels, energies and bins) that carry the discriminating power. We test whether that list correctly predicts which additional measurements would actually raise the statistical separation between hypotheses. The expensive way to answer that question is a profile-likelihood study.

Core Claim

A large-margin metric trained on nuisance-matched triplets produces an approximately rank-$r$ matrix $M$. A nuisance-matched triplet pairs a simulated spectrum with its closest-fitting rival-hypothesis spectrum. The top eigenvectors of $M$ span the "degeneracy-breaking subspace" of the observable space. We claim three things about it. First, distances in this metric track the expected profile-likelihood significance between hypotheses. Second, restricting $M$ to a candidate subset of measurements predicts how much separation that subset buys, which gives a cheap ranking of future measurements. Third, none of this needs cubic-cost matrix operations at realistic $d$.

We do not claim to beat the profile likelihood ratio in statistical power on a fixed dataset. For composite hypotheses it is the reference test, and no classifier trained on the same information should exceed it. The claim is that the metric reproduces its conclusions at a small fraction of the cost, in a form that can be read directly as which bins matter.

Where the Pieces Come From

The method comes from the leading-eigenvector metric learning literature. Positive Semidefinite Metric Learning with Boosting (BoostMetric) builds $M$ as a sum of rank-one matrices $vv^\top$. At each step, $v$ is the leading eigenvector of a constraint-weighted matrix, so each step costs $O(d^2)$. Scalable large-margin Mahalanobis distance metric learning (SDPMetric) applies the same column-generation idea to a smooth large-margin loss under a trace constraint, and reports being nearly 1000× faster than the interior-point solvers SeDuMi and SDPT3 on a 100-dimensional problem. Low-Rank Similarity Metric Learning in High Dimensions (LRSML) proves that the optimal matrix lies in the span of the training data, which removes the dependence on $d$. Information-theoretic metric learning (ITML) supplies the idea of regularizing toward a prior metric $M_0$.

We reuse SDPMetric's update rule and LRSML's subspace reduction. From ITML we take the choice of prior, setting $M_0$ to the inverse Poisson covariance so that the untrained metric is exactly the $\chi^2$. The known weakness of these methods is that the number of triplet constraints grows cubically with sample size (Scalable large-margin Mahalanobis distance metric learning). We sidestep it by building far fewer triplets, and choosing them deliberately rather than at random.

The motivation comes from the hadron side. The $Z_c(3900)$ analysis shows that the degeneracy is real, and it names full multi-channel, Dalitz-level information as the way out. BESIII's $X(1835)$ analysis presents the same structure: a Flatté molecule-like fit (coupling ratio $g^2_{p\bar p}/g^2_0 = 2.31 \pm 0.37$) and a coherent sum with a narrow $X(1870)$ describe the threshold kink comparably well (Observation of an anomalous line shape of the η′π+π− mass spectrum near the pp̄ mass threshold).

Two machine-learning papers are the closest prior work, and both also act as constraints. Deep learning exotic hadrons trains a feedforward network on synthetic line shapes to classify the $P_c(4312)$ into four pole classes, labelled by Riemann sheet (bound or virtual state on sheet II or IV). It uses SHAP attributions to show that the region near the $D^0$ threshold is decisive. However, it excludes triangle singularities from the hypothesis space and offers no geometric account of which variations are nuisance and which are discriminating. Σ Resonances from a Neural Network-based Partial Wave Analysis on K−p Scattering shows that a convolutional classifier loses accuracy as resonances overlap: 98.5%, 94.8% and 82.5% on its 1-, 2- and 3-additional-resonance test sets. This is exactly the regime of confusable hypotheses we target. We differ from both in the object we learn: a convex, low-rank metric whose eigenvectors live in observable space. We also differ in how the training signal is built, from fitted nearest rivals rather than labeled samples alone.

Research Program

The simulation engine reimplements the four $Z_c(3900)$ scenarios as coupled-channel amplitudes: a sheet-III pole, a pole plus triangle, a sheet-IV virtual state plus triangle, and a pure triangle. It also implements the four $P_c(4312)$ pole classes of Ng et al. and the two $X(1835)$ line-shape models. For each hypothesis $h$ we sample nuisance parameters $\theta$ and produce binned observables $x(h,\theta)$ over channels ($J/\psi\pi\pi$, $D\bar D^*\pi$), energies (4.226 and 4.257 GeV, the BESIII points in Observation of a Neutral Structure near the DD̄* Mass Threshold) and Dalitz cells. We then apply resolution smearing, acceptance variation and Poisson fluctuation.

The key ingredient is nuisance-matched triplet mining. For an anchor spectrum $x_i = x(h,\theta_i)$, the positive $x_j$ is another draw from the same hypothesis. The negative $x_k$ is obtained by fitting every rival hypothesis $h'$ to $x_i$ and taking its best-fit prediction $x(h', \hat\theta')$. This is the configuration a $\chi^2$ fit would find indistinguishable. The metric is trained, starting from $M_0$ and under a trace budget, with squared-hinge constraints

$$ (x_i - x_k)^\top M (x_i - x_k) - (x_i - x_j)^\top M (x_i - x_j) \ge 1, \qquad M \succeq 0, $$

using SDPMetric-style rank-one steps. Each step computes one leading eigenvector by Lanczos iteration. Because every negative is a fitted rival, the loss concentrates on exactly the directions the profile likelihood uses. Random-negative triplets or class means would instead reward directions that the nuisance parameters absorb.

Two analyses follow training. First, we project real BESIII and LHCb spectra into the learned subspace. We compare their nearest-hypothesis assignments, with bootstrap uncertainty, against published fits and against the Ng et al. conclusion for $P_c(4312)$, which favors the virtual state on sheet IV. Second, for measurement design we define candidate additions: a new energy point, finer Dalitz binning, an extra channel, or doubled luminosity. We score each by the separation that the restricted metric predicts. We check that ranking against Asimov-dataset profile-likelihood significances computed by brute force. An Asimov dataset is the idealized, noise-free expected dataset used to estimate median significance.

Baselines

The naive reference point is the untrained metric $M_0$, the plain $\chi^2$ distance with a nearest-hypothesis rule. It reproduces the published degeneracy. On the $Z_c(3900)$ scenarios the paper reports $\chi^2/\text{dof} \sim 1.3$ for all four, with separations below $3\sigma$. Any learned metric must separate hypotheses more sharply than this, at matched nuisance, to be worth anything.

The strongest existing competitor for classification is the deep network of Deep learning exotic hadrons, with SHAP used for bin importance. That paper reports robustness to simulated noise up to 5%, and the file contains no reported class-probability percentages for the $P_c(4312)$. We retrain it on our identical simulated data, extended to include triangle-singularity classes, with the same number of simulator calls and the same hyperparameter search budget as the metric. It is the strongest competitor because it is amortized, nonlinear, and already has an attribution method. Beating it means one of two things. Either the metric matches its accuracy on held-out, nuisance-shifted test sets while its measurement ranking agrees better with the Asimov ground truth (higher Spearman correlation), or it is more stable to acceptance shifts outside the training range.

The gold standard for statistical power is the profile likelihood ratio itself: fitting every hypothesis to every test spectrum. It is not something we expect to beat. We report the fraction of its separation that the metric recovers, and the wall-clock ratio. Success is recovering at least roughly 90% of its expected significance at one to two orders of magnitude less cost per design configuration.

Linear discriminant analysis and PCA-then-$k$-nearest-neighbours are the classical linear baselines, trained on the same features. ITML with the same $M_0$ and random pairs is the closest metric-learning baseline.

Two ablations isolate our ingredients. Replacing nuisance-matched negatives with random rival-hypothesis draws tests whether the mining is what matters. Solving the same objective with projected gradient descent and full eigendecomposition checks that rank-one updates lose no accuracy while delivering the claimed speedup at $d \sim 10^4$. For reference, LRSML's reported 1-nearest-neighbour error on Reuters-28 (d=18,933) is about 7.29% against a Euclidean 16.56%. That result is a feasibility precedent for this dimensionality, not a hadron result.

Why not just…?

Why not just use Fisher's linear discriminant? For two Gaussian classes sharing a covariance, the Fisher discriminant is optimal. It too is a leading generalized eigenvector, so it is a real threat. Its weakness here is that it separates class means. Our hypotheses are curved families of spectra over nuisance parameters, and their means can differ along directions that nuisance variation covers anyway. If this nonlinearity is mild, LDA will match us. In that case we will say so, and the contribution reduces to nuisance-matched LDA. The random-versus-matched ablation, run with LDA as well, settles which ingredient carries the result.

Why not just run profile-likelihood Asimov scans for every candidate measurement? That is the correct answer when there are few candidates, and it is our ground truth. The obstacle is cost. Each design times each hypothesis pair times each nuisance configuration needs multi-channel coupled fits, which Σ Resonances from a Neural Network-based Partial Wave Analysis describes as numerically unstable with overlapping structures. The proposal is worthwhile only if the metric's ranking agrees with the scans. We measure that agreement directly on a design grid small enough to scan exhaustively.

Why not just use the deep classifier with SHAP? The network may classify as well or better. SHAP attributions, however, are local to individual inputs and are not a global statement about which observable directions separate hypotheses. A low-rank PSD metric is a global statement by construction. The ranking-correlation test compares the two directly. If SHAP-aggregated importances predict Asimov gains as well as the eigenvectors do, the metric's interpretability advantage is not demonstrated.

Why not just apply PCA and a simple classifier? PCA keeps the directions of largest variance, and here those are dominated by nuisance: mass shifts and background normalization. LRSML notes that heuristic bases like PCA often fail to align with the task. The PCA baseline tests this directly; we expect it to perform near the $\chi^2$ reference.

Decisive Checks

The idea stands or falls on three measurements. The first is the Spearman correlation between metric-predicted and Asimov-computed significance gains across the candidate-measurement grid for the $Z_c(3900)$ scenarios. We need clear superiority over SHAP and LDA, not a tie. The second is the fraction of profile-likelihood separation recovered on held-out spectra with shifted acceptance. The third is agreement with published conclusions on real data: the $P_c(4312)$ sheet-IV virtual-state preference and the unresolved status of $X(1835)$. Reproducing the $X(1835)$ ambiguity honestly, rather than manufacturing a confident answer, counts as a pass. If matched mining gives no gain over random negatives, or if LDA equals the full method, we will report the simpler method as the answer.