How Much Separation Survives Mass Decorrelation? A Certified Ceiling from Quotient Transport Metrics

Collider physicists who search for new particles as bumps in a mass spectrum face a recurring trap. A classifier trained to separate signal jets from ordinary quark and gluon jets tends to learn the jet mass itself. When the analysis then cuts on that classifier, the smooth background acquires a fake peak, a problem called "mass sculpting." The usual fix trains a second network, an adversary, that tries to predict mass from the classifier's output, and penalizes the classifier when it succeeds. This works, but it is a tuned tradeoff with no guarantee. Nobody can say how much discriminating power any mass-blind tagger could possibly retain, so nobody knows whether a given decorrelated tagger is close to the limit or far from it.

We propose to compute that limit. The plan has three steps. First, measure distance between jets with the Energy Mover's Distance (EMD), the cost of transporting one jet's energy pattern onto another's. Second, declare every jet equal to its mass-rescaled versions, which collapses the space along the mass direction. Third, compute the optimal-transport distance between the signal and background populations inside this collapsed space. By Kantorovich–Rubinstein duality, that number is exactly the largest mean score gap that any mass-invariant tagger can achieve if its output changes by at most $L$ when the input jet moves by EMD 1. The function that attains the gap is itself a tagger that is mass-invariant by construction. The same computation shows when decorrelated tagging is hopeless: if the collapsed distance is near zero, no invariant tagger can do much.

Core Claim

Let $X$ be particle positions in the rapidity–azimuth plane around the jet axis. A jet is an energy-weighted measure on $X$, and EMD is the 1-Wasserstein distance $W_1$ on such measures. Let $g_\lambda$ be the angular dilation of a jet about its axis by a factor $\lambda$. This move rescales the jet mass at fixed transverse momentum, so it is the natural "mass nuisance" move. The localized jet metric is the largest pseudometric that lies below EMD and assigns zero distance between $J$ and $g_\lambda J$. Concretely, it is the path (chain) metric built from $\inf_{\lambda,\lambda'} \mathrm{EMD}(g_\lambda J, g_{\lambda'} J')$. Call it $d_\sim$. Write $P_s$ and $P_b$ for the signal and background jet populations. The central quantity is

$$ \Delta_L \;=\; \sup_{f\ \text{invariant},\ \mathrm{Lip}_{d}(f)\le L} \big(\mathbb{E}_{P_s} f - \mathbb{E}_{P_b} f\big) \;=\; L \cdot W_1^{d_\sim}(P_s, P_b), $$

Here $f$ ranges over taggers that are mass-invariant and $L$-Lipschitz with respect to EMD, and $W_1^{d_\sim}$ is the transport distance between the two populations with $d_\sim$ as the ground metric.

The equality has two parts. First, invariant $L$-Lipschitz maps out of EMD-space correspond exactly to $L$-Lipschitz maps out of the quotient; this is the quotient's universal property. Second, Kantorovich–Rubinstein duality applies on the quotient. If $L$ is swept upward, the curve $\Delta_L$ traces a robustness–discrimination frontier for mass-blind tagging. Every existing tagger can be placed below this frontier once its empirical Lipschitz constant is estimated.

Two caveats need stating up front. Invariance under $g_\lambda$ is a design choice that implies decorrelation only when background jets at different masses are approximately related by dilation. Whether that holds is an empirical question we test directly. Also, the bound is on the mean score gap, not on area under the curve or significance. We report those separately, and we treat the bound's relevance to them as something to measure.

Where the Three Directions Enter

Quantitative equational logic supplies the tagger class. This field reasons with approximate equations $t =_\epsilon s$, meaning "$t$ and $s$ are within distance $\epsilon$." Free complete Wasserstein algebras proves that probability measures with the $W_p$ metric form the free algebra of the theory of convex mixtures with a Wasserstein condition. Quantitative Algebraic Reasoning gives the deductive calculus behind that theory. In practical terms, freeness says that any non-expansive feature map $\phi$ on single particles extends uniquely to a non-expansive map on jets, namely the energy-weighted sum $\sum_i z_i \phi(p_i)$. These are energy-flow-style observables. Their Lipschitz constant with respect to EMD is certified rather than estimated.

Composing several such sums through a nonlinear head reuses the input. A Complete V-Equational System for Graded lambda-Calculus handles exactly this case: its graded, Lipschitz-tracking modality bounds how much reusing an input can amplify distances. We use that bookkeeping to certify the Lipschitz constant of the whole tagger. Both papers stay inside semantics of probabilistic programs and never touch physics. Conversely, Particle Graph Autoencoders and Differentiable, Learned Energy Mover's Distance uses EMD for jets but approximates it with a neural network that carries no certificate. We reuse the theory and supply the missing application.

Localization supplies the quotient and also a warning. Adding the equations $J =_0 g_\lambda J$ to the quantitative theory is a localization: certain maps are inverted, and every invariant map then factors uniquely through the result. Brown representability does not come for free shows that localizations can fail to exist even in well-behaved-looking categories. Hypergraph co-optimal transport: metric and categorical properties shows that categories of measure-preserving maps lack basic limits. Both results constrain the design. We therefore work in pseudometric spaces with non-expansive maps, where quotients always exist, and we treat their failure mode as degeneration rather than non-existence. The open question is whether the quotient of the free Wasserstein algebra is again free over the quotient particle space, so that energy-flow taggers on normalized coordinates span all invariant Lipschitz taggers. If the answer is no, then some invariant taggers cannot be written in our class, and the ceiling may be unattainable within it. We will try to prove the result, or exhibit a counterexample, for dilations.

Signal-to-background discrimination supplies the problem and the benchmark. Decorrelated jet substructure tagging using adversarial neural networks is our main foil. It trains the classifier against a mass adversary and acknowledges "slightly worse signal-background separation power" in exchange for robustness. Our ceiling measures how large that loss is compared with the best possible invariant tagger.

Research Program

The theory work proves the duality identity on the quotient and characterizes when $d_\sim$ is a genuine metric on jet classes. It also states the freeness question above, and derives the certified Lipschitz constant of composed energy-flow taggers through graded typing.

The computational work estimates $W_1^{d_\sim}(P_s,P_b)$ in two independent ways. The primal route computes exact transport by linear programming on subsamples of a few thousand jets. Each ground cost is an exact jet-to-jet EMD minimized over a one-dimensional grid of $\lambda$, with the chain closure taken over the subsample. The dual route trains a certified-Lipschitz invariant energy-flow tagger that maximizes the mean gap. The two must agree within sampling error; agreement validates both the estimator and the claim that the tagger class can reach the ceiling. We bootstrap over subsamples and report how the estimate behaves as subsample size grows.

The empirical work places every baseline on the $(L, \text{gap})$ plane and measures sculpting and significance at matched signal efficiency.

Baselines

The naive reference is a single substructure variable, the N-subjettiness ratio $\tau_{21}$, which Decorrelated jet substructure tagging using adversarial neural networks reports the parameterized network beating across 20–300 GeV. A cut on $\tau_{21}$ shows what an untrained, roughly mass-aware observable achieves.

The traditional unconstrained neural network from the same paper is the sculpting-prone upper reference for raw separation.

The strongest competitor is the adversarially decorrelated network from that paper. It is the only method in the file built specifically for mass-decorrelated tagging. It reports a discovery significance of 3.67σ under 50% background uncertainty, against 2.82σ for the traditional network. We regenerate the paper's $Z'\to q\bar q$ plus photon versus γ+jets benchmark with the MadGraph5, Pythia8 and Delphes chain listed in the file. We retrain the adversary with a sweep over its tradeoff weight $\lambda$, using the same number of tuning trials and the same GPU budget we give our method. All methods use identical events and splits.

Beating the adversary means one of two things. Either the certified invariant tagger matches its significance at 50% background uncertainty with lower background-mass distortion, or it exceeds its significance at matched distortion. Separately, the ceiling succeeds as a diagnostic if the adversary's measured mean gap falls between 0 and $\Delta_{\hat L}$ at its estimated Lipschitz constant $\hat L$, and the gap to the ceiling predicts how much our tagger improves.

Two ablations isolate the key ingredient. One keeps the certified Lipschitz energy-flow tagger but drops the quotient. This separates the effect of Lipschitz control from the effect of invariance. The other keeps scale-normalized inputs but uses an unconstrained network. As a scale reference, we apply the Particle Transformer of Particle Transformer for Jet Tagging, which reports 86.1% accuracy on its 100-million-jet JETCLASS set, to the same benchmark. This tests whether a larger unconstrained model escapes the frontier only by giving up Lipschitz robustness.

Why not just…?

Why not just keep training adversaries and tune λ? Adversarial training already works, giving 3.67σ in Decorrelated jet substructure tagging using adversarial neural networks. Its weakness is that it cannot say what it gave up. The proposal does not need to beat it to be useful. The ceiling tells an analyst whether further tuning is worth attempting. If adversarial taggers already sit on the frontier, we will report that, and the remaining contribution is the certificate.

Why not just feed mass-normalized inputs to an ordinary network? This is the real threat. It is plausible that rescaling angles by the jet mass and training any classifier yields invariance as cheaply as our construction does. For that reason we do not rest the proposal on our tagger outperforming this baseline. It is an ablation, and the headline result is the frontier on which this tagger must also lie whenever it is Lipschitz. If this baseline matches our tagger, the method collapses into the diagnostic. We regard that diagnostic as the more durable contribution anyway.

Why not just constrain any network with spectral normalization? Spectral normalization bounds Lipschitz constants in the Euclidean geometry of input arrays, not in EMD. A small Euclidean constant says nothing about robustness to soft or collinear particle moves, and it gives no tie to the transport ceiling. Our Lipschitz-only ablation measures how much of any gain comes from Lipschitz control alone, as opposed to the quotient.

Why not just compute exact EMD by linear programming and skip the algebra? We do use exact linear programming on the primal side. The algebra is what tells us which taggers count as invariant and Lipschitz, and freeness is what lets the dual tagger be a cheap energy-weighted sum rather than a quadratic-cost transport solve per event.

Proposed Experiments and Decisive Checks

The first check is tightness. On the regenerated $Z'$ benchmark, at each resonance mass point, the primal linear-programming estimate of $W_1^{d_\sim}$ and the dual tagger's achieved gap at $L=1$ must agree within bootstrap error. If they disagree persistently, either the tagger class is too narrow (the freeness question has a negative answer) or the estimator is biased, and we diagnose which.

The second check is sculpting. We cut on each tagger at signal efficiencies of 30%, 50% and 70%. We then measure the Jensen–Shannon divergence between the background mass distribution before and after the cut, and we compute significance under 50% background uncertainty using the protocol of the adversarial paper. The 3.67σ adversarial figure is the target.

The third check tests the invariance assumption. We test whether background jets in neighboring mass windows match after dilation, measured by their quotient transport distance. If this distance is large, invariance is stronger than decorrelation requires. Then the ceiling understates what decorrelated taggers can reach, and an adversary exceeding the ceiling would falsify our modeling choice rather than the theorem.

The fourth check tests whether the collapse diagnostic predicts failure. We add further nuisance moves, such as a pile-up-like uniform soft-energy injection, and track how $\Delta_L$ shrinks. The prediction is that the significance of every invariant baseline falls in step with the ceiling.

A secondary transfer test uses the LHC Olympics R&D dataset from The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics. The reference there is the Particle Graph Autoencoder's reported area under the curve of 0.77. We use this test only to check whether the quotient distance between signal and background predicts which detectors avoid sculpting dijet mass. We claim no result on Black Box 3.