Physicists who search for new particles without specifying in advance what they are looking for compare collision data against a reference sample. The comparison is either a two-sample test or a classifier that tries to tell data from reference. Any difference the classifier finds counts as evidence of something new. The weak point is that the detector is not a fixed ruler. Its energy scale, alignment and gains drift slowly over months of data-taking. A comparison that pools all the data then sees a smeared, shifted background and can report the drift itself as a discovery. Quantum communication faced the same problem from a different direction. There, two parties whose coordinate frames rotate unpredictably can still teleport a state if the protocol is arranged so that each party's unknown rotation cancels the other's. This works provided the frame is stable during each short local step.
We propose to carry that "cancel it locally, then combine" structure into data-driven new-physics searches. The method splits the data into time blocks short enough that the detector's calibration is effectively frozen within each one. Within each block it estimates the few parameters of the drift by matching a background-dominated control region to a fixed reference block, then undoes that drift before running a standard search. Because only a handful of group parameters are estimated per block, the leftover systematic error has a predictable size. That size sets an optimal block length. The design also keeps the absolute mass scale that a resonance search needs, which a fully drift-invariant representation would throw away.
Core Claim
Two ideas from frame-independent quantum teleportation transfer cleanly to drifting detectors.
The first is the condition that robustness holds only if the frame is stable during local correction. In Quantum teleportation with infinite reference-frame uncertainty, Verdon and Vicary show that teleportation can be made immune to arbitrary continuous frame rotations during transmission, so long as the frame is stable while the correction is applied.
The second is the trade-off between two of their schemes. The "tight" scheme becomes fully invariant to the frame and leaks nothing about it, but it degrades channel purity, for example to 0.65 under U(1) uncertainty. The "perfect" scheme recovers the state exactly by passing a small orientation token, at the cost of revealing that orientation.
Translated to detectors, a fully drift-invariant search corresponds to the tight scheme. One example is building features only from scale-free ratios, much as the reactor-ratio analysis in Model-independent ν̄e short-baseline oscillations from reactor spectral ratios cancels flux and efficiency uncertainties by dividing spectra measured at two baselines. This approach is clean but discards information. A dijet resonance search cannot become invariant to the jet energy scale without losing the resonance mass.
Our block-local estimate of the drift is the analogue of the perfect scheme's token. It is a low-dimensional summary of the frame, estimated from data. Removing it restores the information, and its estimation error becomes the explicit residual systematic.
The specific claim we test is quantitative. Suppose the drift parameters move at rate $v$ per unit time, events in the control region arrive at rate $r$, and blocks have duration $T$. The residual error then combines two terms. One is drift within the block, of order $vT$. The other is statistical error in estimating $k$ group parameters, of order $\sqrt{k/(rT)}$, from standard parametric estimation. Minimising the sum gives
$$T^\ast \propto \left(\frac{k}{r\,v^2}\right)^{1/3},$$
so the residual floor scales as $(k\,v/r)^{1/3}$. If this law holds empirically, it turns "residual calibration systematics" from a number assigned by hand into a quantity predicted from the drift rate and the data rate.
Relation to Prior Work
Several papers motivate the gap without closing it.
- A Lorentz-equivariant transformer for all of the LHC builds a network that respects spacetime symmetry. It handles detector geometry through fixed reference vectors, so it assumes the lab frame is static and perfectly calibrated. We instead treat the lab frame as a slowly moving, unknown group element.
- Symmetry Discovery with Deep Learning introduces the notion of an "inertial reference dataset" against which transformations are judged. It discovers symmetries in static data. We supply the method's anchor: a frozen reference block plays the inertial role. We do not discover the group; we posit a small one and estimate its element per block.
- The LT-RFI protocol in Study of practical state-preparation error tolerant reference-frame-independent quantum key distribution protocol combines frame drift with device flaws, but under quasi-static assumptions. It is the closest example of jointly budgeting drift and hardware systematics.
The statistical machinery comes from search methods we reuse rather than replace.
- The symmetrized two-sample test of Learning New Physics from Data -- a Symmetrized Approach fits both samples against a combined distribution. Its null distribution follows the textbook chi-square law without weight clipping. That property makes it a good detector of whether our correction has actually removed drift: a miscalibrated null should be visible.
- The contrastive embeddings of Anomaly-preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC supply the strongest existing search pipeline, assessed on static data.
Two lines of work constrain the design.
- Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference shows that methods ignoring nuisance shifts lose coverage badly: adaptive conformal sets fell to 40–60% coverage against a 90% target. Its remedy treats nuisances as distributional shifts and builds valid sets over them. It does not exploit the fact that a drift acts on every event in a block in the same way. That shared action is precisely what makes cancellation, rather than marginalisation, possible.
- Resonant anomaly detection with multiple reference datasets shows that combining several imperfect references beats any single one. It treats references as unstructured samples. We give the references a group structure: each block is a reference displaced by a known kind of transformation.
Research Program
The study runs in four stages.
- Drift simulator. We add time to the LHC Olympics dijet data (The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics). The events carry no timestamps, so we assign synthetic ones and apply a time-dependent calibration transformation from a small group. That group has three parts: a multiplicative jet energy scale, a separate multiplicative jet mass scale, and an azimuthal rotation. The rotation serves as a control case, because dijet observables are already nearly invariant to it and full invariance should be free there. Drift trajectories include slow linear drift, random walks and step changes, such as a recalibration between runs, with rates spanning roughly an order of magnitude.
- Block-local canonicalizer. For each block, we estimate the $k$ group parameters by minimising a classifier-based divergence between that block's mass sidebands and a frozen reference block. We then invert the transformation on all events in the block. Two choices make this safe:
- Estimation uses only sidebands, never the signal region.
- The fit has only $k\le 3$ parameters, so it cannot reshape the spectrum enough to absorb a localised bump.
The uncertainty on each block's estimate is propagated to the search as a constrained nuisance parameter in the symmetrized test.
- Search. We run the contrastive-embedding plus symmetrized-test pipeline on the corrected data.
- Analysis. We map the residual error and the false-discovery rate as functions of block length $T$ and drift rate $v$. We then check the predicted cube-root law against these measurements.
Baselines
Naive reference point. This is the pipeline of Anomaly-preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC, trained and evaluated on pooled drifted data with no correction. On static blind data it reports 12.1σ global discovery significance, against 0σ for raw 57-dimensional inputs and 1.4σ for a variational autoencoder. We run it with identical architecture, training budget and data under drift. We expect two failures: spurious significance on background-only data, and lost significance on real signals.
Strongest existing method: nuisance-aware training. This is the same pipeline, with the drift parameters treated as a nuisance in the spirit of Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference. Training is augmented over the full range of drift, and significance is computed with the drift profiled as a global nuisance. No reported result on LHC Olympics was found on Lacuna for this combination. The only reference figure is its 90% versus 40–60% coverage on single-cell batches. It receives the same tuning budget as our method.
Fully invariant ("tight") baseline. Features are built only from drift-invariant combinations, such as ratios of jet mass to dijet mass, and the search must work without an absolute mass scale.
Oracle. This is the true inverse drift applied per event, which sets the ceiling.
Ablation. One global drift estimate replaces per-block estimates, which isolates the contribution of time-locality.
Our method beats the field if it meets two conditions:
- Null calibration. It holds the background-only false-positive rate at the nominal level, with the null distribution consistent with chi-square, where the naive baseline does not.
- Recovered sensitivity. It recovers at least most of the oracle's significance on injected signals, where the invariant baseline and nuisance augmentation recover clearly less.
Why not just…?
Why not just augment training over drift and profile it as a nuisance? This is the real competitor, and we run it as the main baseline. Its weakness is structural. Augmentation teaches the network to tolerate any drift, which broadens its response and blurs exactly the narrow mass feature a resonance search relies on. A single global profile also cannot represent a drift that differs from block to block. If augmentation matches our method in signal recovery and null calibration at realistic drift rates, then block-local cancellation adds nothing. We would report that as the outcome.
Why not just use invariant features, as reactor experiments use spectral ratios? Ratios work when the quantity of interest survives the division, as with the oscillation pattern in Model-independent ν̄e short-baseline oscillations from reactor spectral ratios. Invariance to the jet energy scale removes the resonance mass, which is the analogue of the purity loss of the tight scheme in Quantum teleportation with infinite reference-frame uncertainty. The fully invariant baseline measures how large this loss is. For the azimuthal rotation, we expect invariance to be the right answer, and our method should simply match it.
Why not just recalibrate in situ with a standard candle, such as a known mass peak? Where a clean candle with enough rate exists, this is our method with a one-dimensional group. We do not claim to beat it. Our contribution covers the case without a candle, where several parameters must be estimated from generic sideband data. It also adds the block-length law, which tells an experiment how often recalibration must happen. We include a candle-style variant, which estimates only the energy scale from a single sideband peak. Including it shows when the extra parameters pay off.
Why not just feed time to an unconstrained network? A network conditioned on the timestamp can learn the drift. It can equally learn a time-localised signal as background, and nothing bounds its residual error. We test this variant on the background-only black box to see whether its null stays calibrated.
Decisive Checks
Three checks settle the proposal.
- Background-only null. On the background-only black box (BB2), under each drift trajectory, we compare the empirical false-positive rate at nominal α and the shape of the test-statistic distribution to chi-square across all methods. A correct method must stay at nominal.
- Signal recovery. With signal injected into the R&D set at several strengths, and on Black Box 1, we report discovery significance relative to the oracle.
- Scaling law. We sweep block length at three drift rates, measure the residual parameter error, and test whether the optimum moves as $v^{-2/3}$ and the floor as $v^{1/3}$. If the measured optimum does not track the prediction, the "stable during local correction" analogy fails, and the method reduces to ad hoc binning.
To check that the canonicalizer does not absorb signal, we compare the drift parameters it estimates with and without injected signal. A measurable shift would expose leakage of signal into the calibration.
References
- Quantum teleportation with infinite reference-frame uncertainty
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- A Lorentz-equivariant transformer for all of the LHC
- Symmetry Discovery with Deep Learning
- Study of practical state-preparation error tolerant reference-frame-independent quantum key distribution protocol
- Learning new physics from data: A symmetrized approach
- Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC
- Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference
- Resonant anomaly detection with multiple reference datasets
- The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics