Model-independent searches for new physics at particle colliders ask a simple question: does the observed data distribution differ from the Standard Model simulation anywhere? The question is hard because the simulation is never quite right. One example is hadronization, the poorly understood process that turns quarks and gluons into observable particles, which shifts jet properties by small, energy-dependent amounts. A sensitive search will flag that mismodeling as a "discovery." The usual remedy is to add each known uncertainty as a free parameter and fit it away, a procedure called nuisance profiling. That works in five dimensions and becomes slow and fragile on the rich learned jet representations that modern taggers produce. We propose a cheaper and more transparent step that runs before the statistical test. We measure, directly in representation space, which directions each systematic deformation moves events along. We then choose a low-dimensional linear projection that keeps background structure while being nearly blind to those directions. This is a generalized form of principal component analysis (PCA, the standard method for finding the axes along which data varies most). The projected features go to an existing goodness-of-fit test.
The second half of the idea concerns which deformations to include. Field theory gives a specific shape to hadronization uncertainty: corrections that fall as $1/Q$ with the hard scale $Q$ and evolve logarithmically with energy. Lattice studies have also found a correction falling as $\Lambda^2/Q^2$ ($\Lambda$ is the intrinsic QCD scale) that the standard operator analysis forbids. We treat that unexpected term as a stress test. The test asks whether a search made robust only to the textbook $1/Q$ form produces false discoveries when the reference is deformed by a $1/Q^2$ term instead, and whether the projection method catches the difference.
Core Claim
Our claim is that systematic robustness can be achieved largely by geometry rather than by fitting. The method has three steps. Embed jets or events with a pretrained network. Estimate a per-family "nuisance response" matrix by applying physically parameterized deformations to the same simulated events and recording how their embeddings move. Then solve
$$\Sigma_B\, w = \lambda\,(N + \epsilon I)\, w,$$
where $\Sigma_B$ is the covariance of background embeddings, $N=\sum_k \sigma_k^2\, \mathbb{E}[\delta z_k\, \delta z_k^\top]$ is the second moment of the per-event embedding displacement $\delta z_k$ under a unit shift of nuisance $k$ with prior width $\sigma_k$, $\epsilon$ is a small regularizer, and $w$ are the projection directions. Keeping the top $d$ generalized eigenvectors yields features with large background variation per unit of systematic motion. A reduced set of residual nuisances is then profiled with the New Physics Learning Machine (NPLM) on those $d$ features. We expect three results. False-discovery rates should match nominal levels under deformed references. Sensitivity should match full profiling on standard PCA features. Compute cost should scale with $d$ rather than with the number of nuisance families.
Relation to Prior Work
Several papers motivate the gap. Learning new physics from an imperfect machine builds NPLM, a neural likelihood-ratio test that profiles detector-scale and normalization nuisances alongside network weights. It works on a 5D dimuon setup and reports an "agnostic tax" of about 2.7 relative to a dedicated search. It does not reach high-dimensional inputs, and it does not model theoretical shape systematics. Robust resonant anomaly detection with NPLM moves NPLM to kernel methods on summary observables. Anomaly-preserving contrastive neural embeddings shows that the choice of compression decides the outcome. Its 4D contrastive embeddings reach 12.1σ on a blind black box, where raw 57D inputs reach 0σ and a 6D variational autoencoder reaches 1.4σ. That embedding is learned non-linearly and never conditioned on systematics. We reuse NPLM and contrastive or pretrained encoders unchanged. Our contribution sits between them as the linear selection of directions that are robust to systematics.
Other work supplies the method. Simple and Effective Unsupervised Redundancy Elimination shows that unsupervised PCA beats supervised linear projections for compressing retrieval embeddings. It also shows that PCA degrades under distribution shift, which is exactly our concern. Raunak et al. remove dominant nuisance-like directions before and after PCA. Our generalized eigenproblem is a principled version of that purification step, with the directions to remove measured from physics deformations rather than taken as the mean vector.
The physics papers supply the deformations and also constrain them. Power corrections to event shapes with mass-dependent operators organizes $1/Q$ hadronization corrections through a transverse-velocity operator. It gives those corrections a logarithmic running with $Q$, and it shows that Pythia 8 and Herwig++ disagree on the low-energy parameters. Burgio et al. and Burgio et al. find a $\Lambda^2/Q^2$ remainder in the lattice plaquette after subtracting an eight-loop series, and they attribute it to power terms in the running coupling. Those papers also warn that finite-volume artifacts are not fully excluded. That evidence concerns the gluon condensate, not jets. We therefore do not claim jets carry a $1/Q^2$ term. We use it as a plausible, distinctly scaling deformation that a robust search should not mistake for new physics.
Research Program
The first component is a deformation module. Starting from simulated background events, we apply two parameterized soft-recoil deformations to jet constituents. The first is built so that the induced shifts in jet mass and angularities scale as $\Omega_1/p_T$. It carries the transverse-velocity weighting and is run in $Q$ according to the anomalous dimension of Mateu et al.. The second is built so that the shifts scale as $\Lambda^2/p_T^2$. Before either deformation is used, we validate it by checking that the induced observable shifts follow the prescribed scaling across transverse-momentum bins. We add standard detector nuisances (jet energy scale, resolution) as in NPLM. Generator differences, where alternative-generator samples are available, enter only as a held-out test, never as a template.
The second component is the projection module. We embed events with two kinds of frozen encoder. The first is the Particle Transformer trained on JetClass, using its class-token and hidden activations (Particle Transformer for Jet Tagging, 86.1% JetClass accuracy). The second is the contrastive encoders of Metzger et al. We then estimate $\Sigma_B$ and $N$ on a calibration split and solve the eigenproblem above. Because the deformations act per event, we compute $\delta z$ by finite differences on paired events, which avoids comparing unpaired samples. The result is a fixed linear map whose effect on any nuisance can be read off directly. That transparency is lost with non-linear bottlenecks.
The third component is the testing module. NPLM, with either neural or kernel back ends, runs on the $d$ projected features and profiles only the nuisances whose residual response, $w^\top N_k w$, exceeds a threshold.
The three components are separated by fixed interfaces, following the decoupled-evaluation designs of Universal Causal Evaluation Engine and the seeding and version binding of DeepDIVA. This separation is a methodological requirement, not a product. Every ablation swaps exactly one module while holding data splits, seeds, and test statistics fixed. DeepDIVA also notes that GPU non-determinism can survive seeding, so we report variance over repeated runs rather than single numbers.
Baselines
The naive reference point is standard variance-ranked PCA to the same $d$, fed to NPLM with no nuisance handling. This shows how badly hadronization mismodeling inflates false discoveries when nothing is done.
The strongest existing method is full NPLM profiling as in D'Agnolo et al.. It runs on the same standard-PCA features, with every nuisance family (detector, $1/Q$, $1/Q^2$) profiled explicitly. This is the most direct competitor because it handles systematics by fitting rather than projection and is statistically well understood. It gets the same encoder, the same $d$, the same pseudo-experiment count, and the same hyperparameter search budget over regularization. On LHCO the closest reported number is a median Z of 0.30 ± 0.04 for NPLM against 0.1 ± 0.1 for BDTs at $S/B = 8.2\times10^{-4}$ (Grosso et al.). On the black box, the best reported compression is 12.1σ for the supervised SimCLR Transformer (Metzger et al.). Neither number includes hadronization nuisances, so we rerun both baselines under our deformed references. We count our method as beating full profiling if it reaches equal or better median Z at matched false-discovery rate with lower wall-clock cost. Our method also wins outright if full profiling fails to keep calibration when nuisance count grows and ours does not.
The contrastive 4D embedding of Metzger et al., used without projection, is a third baseline. It tests whether a good encoder alone is already robust.
Three ablations isolate the key ingredient. The first replaces $N$ with the identity, which reduces to plain PCA. The second keeps the smallest-variance components ("minor PCA"), which tests the worry that anomalies hide in low-variance tails. The third builds $N$ from the $1/Q$ family only, which isolates whether including the unexpected $1/Q^2$ family matters.
Why not just…?
Why not just profile every nuisance in NPLM and skip the projection? This is the real threat. If profiling a dozen nuisances on 16–64 PCA features stays calibrated and fast, our projection adds little. No reported result on Lacuna tests NPLM beyond five dimensions, and the file notes that gradient-based profiling becomes slow and prone to local minima as nuisance count grows. Our runtime-and-calibration scan over nuisance count settles this. The hybrid design also means that a win for profiling is not a loss for the program. In that case the projection simply chooses which nuisances need profiling.
Why not just decorrelate the embedding adversarially from the nuisances? Louppe et al. decorrelate a tagger from jet mass, raising significance from 2.82σ to 3.67σ under 50% background uncertainty. That approach requires retraining for each nuisance set, offers no closed-form residual sensitivity, and targets one observable. We include an adversarially decorrelated linear head as an extra comparison at matched compute. If it matches us, the simpler argument for our method reduces to transparency and zero training.
Why not just take the envelope of Pythia and Herwig? Mateu et al. show that the two generators disagree on the non-perturbative parameters, and that the difference shifts extracted $\alpha_s(m_Z)$ by about ±0.0005. A two-point envelope also has no scale dependence. We keep the generator difference as a held-out test. If our parameterized nuisances do not cover it, the deformation family is incomplete, and we will report that.
Why not just rely on the 4D contrastive embedding, which already works? Its 12.1σ was obtained with no systematic deformation of the reference. Whether its compact space happens to be orthogonal to hadronization shifts is untested. Running it under our deformed nulls answers this directly.
Proposed Experiments
We use the LHC Olympics 2020 benchmark (LHC Olympics 2020), the RODEM set, and the blind black-box dataset of Metzger et al., with injected-signal sweeps where the benchmark permits.
The first decisive check is null calibration. We draw thousands of signal-free pseudo-experiments in which the "data" is the reference deformed by the $1/Q$ family, the $1/Q^2$ family, or a held-out generator. For each pipeline we report the fraction of pseudo-experiments exceeding 3σ. The critical comparison is the $1/Q$-only ablation evaluated on $1/Q^2$-deformed nulls. If it stays calibrated, the two scalings are not separable at collider transverse momenta, which is a useful negative result. If it fails while the full projection stays calibrated, the unexpected power correction matters as a nuisance.
The second check is sensitivity. We compare median Z under signal injection across all baselines at a matched false-discovery rate. We also quantify how much signal is lost when signals lie partly along nuisance directions, by measuring $w^\top \delta z_{\text{signal}}$, the residual signal overlap after projection.
The third check is cost. We measure wall-clock time and fit stability as the number of nuisance families grows from 2 to about 20.
The note's thesis fails if full profiling matches our calibration and sensitivity at comparable cost. It also fails if minor-component PCA beats the generalized eigenvectors on sensitivity, since that would show variance-to-nuisance ratio is the wrong criterion.
References
- Learning new physics from an imperfect machine
- Robust resonant anomaly detection with NPLM
- Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC
- Simple and Effective Unsupervised Redundancy Elimination to Compress Dense Vectors for Passage Retrieval
- Effective Dimensionality Reduction for Word Embeddings
- Power corrections to event shapes with mass-dependent operators
- Λ2-contribution to the condensate in lattice gauge theory
- Power corrections and perturbative coupling from lattice gauge theories
- Particle Transformer for Jet Tagging
- Universal Causal Evaluation Engine: An API for empirically evaluating causal inference models
- DeepDIVA: A Highly-Functional Python Framework for Reproducible Experiments
- Decorrelated jet substructure tagging using adversarial neural networks
- The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics