Laboratory "analogue black holes" are built from intense laser pulses moving through glass. The pulse locally changes how fast light travels, which creates a boundary that light cannot cross. Such a boundary should emit faint light analogous to the Hawking radiation of real black holes, which is far too cold to see in space. Every published claim of this light rests on one judgment: is a weak ultraviolet peak in a spectrum really horizon emission, or a classical lookalike such as plasma glow, Raman scattering, or ordinary resonant radiation? Machine-learning work on visual grounding has met the same problem in another setting. Models asked to find "the kid in blue" confidently highlight the kid in purple. They do not fail on random empty images. They fail on near-miss distractors that match everything except one attribute. The field made progress only after it built benchmarks with explicitly labeled no-target cases, including deceptive ones, and scored models on whether they correctly said "nothing here."
We propose to carry that lesson into analogue gravity. We will build a benchmark of simulated spectral sweeps where horizon emission is present, absent, or absent but mimicked by a classical process sitting at the predicted wavelength. We will also build a detection test that asks whether a candidate peak moves the way horizon kinematics require when experimental knobs are turned. A peak that merely sits in the right place does not pass. The claim is that standard practice, a fixed spectral window plus a significance threshold or a template fit, gives a high false-discovery rate on deceptive nulls. A test that checks whether the peak is "horizon-locked" stays near its nominal error rate at comparable detection power.
Core Claim
Horizon emission in a dispersive medium is not a fixed line. It appears where the frequency in the frame moving with the pulse is conserved:
$$\omega' = \omega - v\,k(\omega),$$
where $\omega$ is the laboratory frequency, $k(\omega)$ is the medium's wavenumber, and $v$ is the pulse velocity. This condition, worked out in Dielectric black holes induced by a refractive index perturbation and the Hawking effect and Blackbody Emission from Light Interacting with an Effective Moving Dispersive Medium, predicts where emission appears and how that position shifts as $v$, the medium, or the index perturbation changes. Raman lines are locked to the pump frequency. Plasma recombination and fluorescence are locked to material levels. Each classical mimic therefore has its own covariance signature.
Our test statistic compares two quantities across a sweep of control settings $\theta_j$, such as Bessel cone angle, which sets $v$, pulse energy, and medium. The first is the fitted peak position $\hat\omega_j$. The second is the phase-matching prediction $\omega^\ast(\theta_j)$. A horizon-locked component has slope near one against the prediction, while a mimic has a slope set by its own physics. The decision is three-way: horizon-locked emission present, a peak present but not locked (a "deceptive null"), or no peak at all.
The vision analogue is that the phase-matching relation plays the role of the referring expression. The target is "whatever obeys this kinematic description," and the system must be allowed to answer "nothing obeys it."
Relation to Prior Work
The physics papers motivate the gap. Experimental evidence of analogue Hawking radiation from ultrashort laser pulse filaments reports emission near 350 nm in fused silica. It excludes Rayleigh, Raman, and plasma backgrounds through manual polarization and geometry checks, and shows that bandwidth scales linearly with pulse energy. That scaling is a covariance argument made informally on one dataset, with no calibrated error rate. Negative-Frequency Resonant Radiation shows a negative-frequency peak at 341 nm in CaF₂ beside positive resonant radiation at 620 nm, matching predictions to within 1% in fibers. It also shows that this coupling is orders of magnitude weaker than the positive branch, which is exactly the regime where a strong neighbor can masquerade as the weak target. The exchange in Belgiorno et al. Reply: responds to a comment questioning whether such emission is genuine Hawking radiation or classical resonant scattering. We do not settle that interpretive debate. We only ask whether a given dataset contains a component that follows horizon kinematics, which is a necessary condition that both sides accept.
The vision papers supply the evaluation method and part of the model design. GRES: Generalized Referring Expression Segmentation and GREx redefine grounding to include one-to-zero cases. They add a no-target head (ReLA) and identify deceptive expressions as the dominant failure mode. We reuse their benchmark logic, with explicit null labels and a separate accuracy on nulls, and their idea of a dedicated null output. Rex-Thinker shows that base multimodal models almost never reject: its rejection score is 7.1 for base Qwen2.5-VL and 68.2 after verification training. That finding constrains our design. Abstention has to be built in and evaluated, not assumed to emerge from confidence scores. Robust Referring Video Object Segmentation with Cyclic Structural Consensus checks whether a query and the visual evidence agree under a reconstruction cycle. Our locking test is a physics-grounded version of a consistency check: it asks whether the evidence responds to interventions the way the description says it should. Unlike all four, we do not apply a vision-language model to spectra. We borrow the problem structure, not the network. Gravitational parameter estimation in a waveguide bounds estimation uncertainty for spacetime parameters but poses no present-versus-absent decision. Its Fisher-information approach will guide how we choose sweep points.
Research Program
The first deliverable is a simulator-backed benchmark, which we call DeceptiveNull-Horizon. We will propagate pulses with standard one- and two-dimensional nonlinear propagation models, using Sellmeier dispersion for fused silica and CaF₂. Configurations follow the published setups: filaments and Bessel beams producing about 350 nm emission, and 60 fs pulses at 800 nm in 2 cm CaF₂.
We will inject horizon emission at its phase-matched position, with flux spanning several orders of magnitude around the detection floor. Background components are a plasma recombination continuum, Raman lines, fluorescence, positive resonant radiation, and four-wave-mixing products. Detector noise follows a photon-counting model for a cooled CCD, with dark counts and readout noise.
Each benchmark item is a sweep over several settings of cone angle, pulse energy, or medium. Items fall into three classes. In "present" items, locked emission is injected. In "random null" items, there is no emission at the predicted position. In "deceptive null" items, there is no locked emission, but a mimic is tuned to sit at the predicted wavelength at the nominal setting and moves according to its own physics.
We will keep a held-out family of mimics, with line shapes and drift laws never seen in design or training. This set tests robustness to backgrounds nobody modeled. The one-dimensional angular models are known to miss three-dimensional structure Experimental evidence of analogue Hawking radiation from ultrashort laser pulse filaments, so a two-dimensional subset serves as a check for mismatch between simulator families.
The second deliverable is the locking test in two forms. The analytic form fits peaks at each setting, regresses them on $\omega^\ast(\theta_j)$, and calibrates a three-way decision by permutation over sweep labels. The learned form is a small set-prediction network. It outputs zero or more spectral components, each labeled "locked" or "not locked," plus an explicit no-target head in the style of ReLA, and is trained only on simulated sweeps. The learned form exists to handle overlapping components where peak fitting breaks down. The ablation below decides whether it is needed.
The third deliverable is a protocol recommendation: the minimum number and spacing of sweep points needed to reach a target false-claim rate. Sweep design will use Fisher information on the slope parameter.
Baselines
The naive reference point is current practice: a fixed window around the predicted wavelength and a claim when counts exceed background by $5\sigma$. The window is estimated from side bands at the nominal setting only. It gets the same simulated data, and its window width is tuned on the training split to maximize power.
The strongest existing competitor is a profile likelihood ratio test. It fits the spectrum with templates for every known background plus a signal template at the predicted position, and is applied jointly across the sweep with the same data. When the background templates are complete, the Neyman–Pearson lemma makes this test optimal, so it is the real threat. It receives the true templates for all in-distribution mimics and the same sweep points as our test. We tune nothing beyond what it needs.
The ablation that isolates our key ingredient is the same learned network trained on the same data but given only the nominal-setting spectrum, with no sweep. It can recognize line shapes but cannot observe covariance. A second ablation removes the no-target head and thresholds component confidence instead.
Beating the strongest baseline means meeting a fixed requirement on held-out mimics. At injected flux where the likelihood test reaches 90% power, our test must keep its false-claim rate on deceptive nulls within twice its nominal level, while the likelihood test's rate is at least five times nominal. Power on "present" items must not drop by more than 10 points.
No reported result found on Lacuna for false-claim rates of any analogue-Hawking detection procedure. These baselines therefore have no published numbers, and all values will be measured here. The only reported abstention numbers in the file come from vision: ReLA's 63.60% gIoU on gRefCOCO validation, and Rex-Thinker's rejection score of 68.2 on HumanRef. They show the size of gains that explicit null handling has produced in that field. They are not comparable to our metrics.
Why not just…?
Why not just use a narrow bandpass filter and a significance threshold? The emission window coincides with plasma continuum and other classical light, as Experimental evidence of analogue Hawking radiation from ultrashort laser pulse filaments had to rule out by hand. A threshold measures excess light, not its origin. A deceptive null that places a mimic in the window is, by construction, invisible to this method, and the benchmark will quantify the resulting false-claim rate directly.
Why not just fit a profile likelihood with background templates? If every background is known and modeled correctly, this is the right answer, and on in-distribution mimics we expect it to match or beat us. Our claim concerns misspecification: an unmodeled mimic at the predicted wavelength gets absorbed into the signal template. The held-out-mimic split settles the question. If the likelihood test stays calibrated there, the locking test adds nothing, and we will report that. A hybrid that adds the locking slope as a parameter of the likelihood is a natural outcome and will be tested.
Why not just train a classifier on single spectra and threshold its confidence? The grounding literature shows that near-miss distractors receive high confidence, not low. GRES identifies deceptive expressions as the main failure, and Rex-Thinker needed explicit verification to raise rejection from 7.1. Our no-sweep ablation is exactly this alternative. If it matches the full model on deceptive nulls, then the covariance idea is wrong.
Why not just ask a multimodal model with chain-of-thought to inspect the plots? Such a pipeline has no access to dispersion physics beyond what it memorized, and costs about six times the latency of direct prediction (6.68 s versus 1.13 s per image in Rex-Thinker). We will run one such model zero-shot on rendered sweep plots as a reference point, expecting it to perform near the naive threshold.
Proposed Experiments
The central experiment produces detection-limit curves: power as a function of injected flux for each method. Each curve is paired with its false-claim rate on random nulls and on deceptive nulls, in both the in-distribution and held-out-mimic splits. The decisive check is the gap between the two null rates. A method that is truly robust shows little difference between random and deceptive nulls, while threshold and template methods should show a large one.
A second experiment varies the number of sweep points, from two to ten, and compares the observed false-claim rate with the Fisher-information prediction. The result is a concrete minimum sweep for experimenters.
A third experiment probes simulator mismatch. We train and calibrate on one-dimensional simulations and evaluate on two-dimensional ones. This tells us whether the test survives the dimensionality gap noted in the physics literature.
A fourth experiment applies the analytic test to published or shared raw spectra, wherever experimental groups can provide multi-setting data. It will report the locking slope with uncertainty, not a verdict on Hawking radiation.
The idea fails if the template likelihood stays calibrated on held-out mimics, or if the no-sweep ablation matches the full model. Either result would tell experimentalists that the current analysis is adequate, which is itself worth knowing before more claims are made near the detection floor.
References
- Dielectric black holes induced by a refractive index perturbation and the Hawking effect
- Blackbody Emission from Light Interacting with an Effective Moving Dispersive Medium
- Experimental evidence of analogue Hawking radiation from ultrashort laser pulse filaments
- Negative-Frequency Resonant Radiation
- Belgiornoet al.Reply:
- GRES: Generalized Referring Expression Segmentation
- GREx: Generalized Referring Expression Segmentation, Comprehension, and Generation
- Rex-Thinker: Grounded Object Referring via Chain-of-Thought Reasoning
- Robust Referring Video Object Segmentation with Cyclic Structural Consensus
- Gravitational parameter estimation in a waveguide