Collider-Style Hypothesis Tests for Pulsar Radius-to-Frequency Mapping

Pulsars emit radio beams from plasma streaming along their magnetic field lines. The standard picture of this emission, called radius-to-frequency mapping, says that each observing frequency comes from a particular height above the star: high frequencies come from low heights where the plasma is dense, and low frequencies from greater heights. Because the bundle of field lines widens with height, the pulse should look narrower at high frequency and wider at low frequency. New ultra-wideband telescopes keep finding pulsars that break this rule. One pulsar's profile widens with frequency. Others emit from the same height at every frequency. In a third, the part of the pulse that flickers moves from one side of a component to the other as frequency changes. Each anomaly is currently judged by fitting a power law to widths in each frequency band and reading the result by eye. The competing explanations, such as a partly empty beam or outer components that disappear at low frequency, are stated only in words. Nobody has attached a likelihood to them, set an upper limit on the flux of the component that is supposed to be missing, or checked how the measurement procedure itself biases the widths.

B-factory experiments like BABAR faced the same statistical problem for decades and solved it with a specific toolkit. Tag-and-probe means fully reconstructing one clean half of an event and using it to define an unbiased sample for studying the other half. Missing-component inference means using conservation laws to reconstruct particles the detector cannot see, and then stating in numbers how much of them could be hiding. Profile-likelihood limits means treating systematic effects as fitted "nuisance" parameters and reporting how strongly the data allow or exclude a hypothesis. This proposal borrows that analysis logic from the BABAR rare-decay searches, not BABAR's data. It builds a joint, cross-frequency forward model of single-pulse data that turns "does radius-to-frequency mapping hold here, and if not, which alternative?" into a calibrated test with injected-signal checks.

Core Claim

Pulsar tests of radius-to-frequency mapping suffer from two biases that no current analysis controls, and correcting them will change the verdict for at least some of the reported anomalies.

The first is a signal-to-noise bias in width. Widths quoted at a fixed fraction of the peak shrink when the outer wings sink into noise. Low-frequency data from different instruments usually have lower signal-to-noise, so this effect alone can produce an apparent "widening with frequency."

The second is a band-dependent state-labelling bias. Mode-switching pulsars are classified pulse by pulse, independently in each sub-band. If the misclassification rate depends on frequency, each band's "State A" profile contains a different mixture of states, and the fitted width and height trends are contaminated.

The proposal fixes both problems with two collider ideas. For the first, a forward model fits single-pulse data across all bands at once, with explicit terms for a hidden outer component. For the second, tag-and-probe classification uses one high-signal-to-noise sub-band of simultaneous wideband data to label pulses, and all other sub-bands are then analysed with those fixed labels. The output for each pulsar is a likelihood ratio among three hypotheses (standard mapping, frequency-independent height, and mapping plus a frequency-suppressed hidden component) together with a 90% confidence upper limit on the hidden component's fluence.

In short, BABAR is the methodological source and pulsar data are the application. Section 2 of the research file confirms that no work applies B-factory tagging or limit-setting to this problem. The contribution is not "run a classifier on pulsars." It is to import the specific inference structure those searches used to make claims about unseen components, together with their discipline for handling systematics.

Research Program

Physical model. The core object is a per-pulsar generative model. The emission beam is a set of components: a core plus one or more cones. Under the mapping hypothesis, each component's angular half-width scales as $\rho_k(\nu) = \rho_{k,0}\,(\nu/\nu_0)^{-a_k}$, where $\nu_0$ is a reference frequency and $a_k$ a fitted index. This form comes from dipole field lines widening as the square root of height, combined with a power-law link between height and frequency. The competing "frequency-independent height" hypothesis fixes $a_k = 0$. The "hidden component" hypothesis adds an outer component whose amplitude is multiplied by a visibility factor that switches off below a cutoff frequency.

Geometry constraint. The components are projected onto pulse longitude using the rotating vector model, a fit of the polarization angle sweep that gives the magnetic inclination and the observer's closest approach to the magnetic axis. The fit also includes the aberration-retardation phase shift, the offset between intensity and polarization centres that measures absolute emission height. Doing this ties the width evolution to the independently measured emission heights instead of fitting widths in isolation.

Instrument and propagation model. Band-specific nuisance parameters enter the likelihood the way data-simulation discrepancies do in the BABAR $B\to K\nu\bar\nu$ search. They cover residual dispersion-measure misalignment, scattering broadening at low frequency, per-instrument flux calibration, and baseline noise level. The noise term is the key one, because it produces the signal-to-noise width bias automatically. The model predicts what a noisy profile would look like, not an idealised width.

Tag-and-probe step. This applies only to data where all frequencies are recorded at the same instant, from the Parkes ultra-wideband receiver (704–4032 MHz) and the FAST ultra-wideband receiver. Each single pulse is labelled once, using the sub-band with the highest signal-to-noise. The labels come from the existing $\Delta\chi^2$ template method or from the joint model's own posterior over states. The other sub-bands are then analysed with those labels frozen. This mirrors how the BABAR analyses in Search for the Decay B+ → K+ τ∓ μ± and Search for B̄ → Λc+ X ℓ- ν̄ℓ decays fully reconstruct one B meson so that the recoiling side is studied without the selection sculpting it.

Hidden-component limits. For J0250+5854 the question is whether a missing outer component is required. We compute a profile-likelihood 90% confidence upper limit on the outer component's fluence fraction at 57 MHz and 149 MHz, following the logic by which Search for the rare decay B → K ν ν̄ and the Belle Λ plus missing energy search bound invisible final states. This converts the qualitative "underfilled beam" story of A broad-band radio study of PSR J0250+5854 into a number that future NenuFAR data can confirm or exclude.

Validation by injection. Everything is validated by injecting simulated pulse stacks. These have known geometry, known state mixtures, and a known hidden component, and are added at the real noise levels of each instrument. This plays the role of the Monte Carlo signal injections that establish coverage in collider searches. The analysis choices will be fixed on the injections and on a control pulsar before the anomalous pulsars are fit, which is the collider practice of blinding.

Data and Roles of Prior Work

Positive control. PSR J1926-0652 shows ordinary conal contraction with frequency. The method must recover standard mapping there, or it is broken.

Motivating anomalies. Three datasets motivate the proposal:

  • J0250+5854, which widens with frequency with a power-law index $B = 1.9 \pm 0.4$ from 57 to 1250 MHz, using FAST, LOFAR, GBT and NenuFAR data that were not taken simultaneously.
  • PSR J0614+2229 with Parkes UWL, whose two states sit at about 543 km and 425 km at 879 MHz, with neither height changing with frequency.
  • PSR B1859+07 with FAST, with 22,769 pulses, emission-shift offsets scaling as $\Delta\theta \propto \nu^{-0.169}$, and "double-vision" pulses in which both states appear at once.

Constraining case. PSR J1948+3540, where the modulating region moves from $M_I$ to $M_{II}$ between 400–750 MHz and 1250 MHz, constrains the model. A beam made of symmetric components cannot produce that asymmetric flicker. The model must either add an asymmetric emission region or honestly fail on this pulsar. We report whichever happens.

Contrasting methodology. Dual-frequency single-pulse study of PSR B0950+08 interprets 55 MHz and 1.4 GHz single pulses through the mapping, but without a formal comparison of alternatives. We add that comparison. Aspen Open Jets supplies the argument that representations learned from real data transfer across regimes. Our datasets are thousands to tens of thousands of pulses, not hundreds of millions, so we deliberately do not use a foundation-model route. An explicit physical likelihood is more appropriate at this scale.

Baselines

Naive reference. The do-nothing baseline is the published analysis itself: a power-law fit of profile width at a fixed intensity fraction against frequency, reported with its quoted error. This is how J0250+5854's $B = 1.9 \pm 0.4$ was obtained. We rerun it on exactly the same profiles as our model. Beating it means showing, on injected data with known truth, that its index is biased and ours is not. On the real data, beating it means the corrected index moves by more than its quoted uncertainty, or that the likelihood ratio favours a hypothesis different from the one the naive fit suggests.

Strongest existing method. The strongest competitor is the per-sub-band $\Delta\chi^2$ template classification of the J0614+2229 study. Each pulse is assigned to whichever state's average profile it matches best, followed by aberration-retardation height fits in each band. It is the strongest baseline because it produced the frequency-independent height result that most directly challenges the mapping, and it is simple and well understood.

The reported figures are the state heights of about 543 km (State A) and 425 km (State B) at 879 MHz, with spectral indices of about −1.53 and −1.83. We found no reported classification accuracy or misclassification rate for the $\Delta\chi^2$ method on Lacuna. We will estimate that rate ourselves on injected data at each sub-band's signal-to-noise. Both methods get the same pulses, the same templates, and the same height estimator. The only difference is where the labels come from.

The proposal beats this baseline if either of two things happens. The first is that its per-band height estimates show smaller bias on injections with a known height slope. The second is that on real data its frequency slope of height differs from zero at a level the baseline cannot resolve. If neither happens, the frequency-independent height result gains credibility, which is also a useful outcome.

Ablation 1. Run the full joint model but with the noise-aware width term replaced by idealised widths. This isolates the contribution of the signal-to-noise bias correction.

Ablation 2. Run the tag-and-probe pipeline but label each band independently. This isolates the contribution of fixing labels from the cleanest band.

Compute and tuning. Each variant gets the same compute and the same number of injection trials used for tuning.

Why not just…?

Why not just train a boosted decision tree on the pulses, as BABAR did for $B\to K\nu\bar\nu$? Those classifiers were trained on simulated events whose true category was known. Pulsars have no labelled ground truth for their emission state, so a supervised classifier would only learn to copy whatever template labels it was given. The file offers no evidence that the classifier's power is the bottleneck. The labelling bias across bands is the more likely problem. We include a gradient-boosted state classifier trained on injected pulses as a secondary check. If it matches tag-and-probe on injections, that tells us the classifier choice does not matter and the label source does.

Why not just cluster single pulses without supervision, for example with a Gaussian mixture? This is a real alternative for finding states, and it handles double-vision pulses more gracefully than templates do. However, clustering each band on its own reintroduces the band-dependent labelling problem. Clustering all bands together is essentially our joint model without the physics, so it cannot say anything about emission height. We will run a joint-band mixture model as an extra baseline. If it matches our label accuracy on injections, we will adopt it for the labelling step and keep the physical model only for the hypothesis test.

Why not just refit J0250+5854 widths with a signal-to-noise correction and skip the whole machinery? For J0250+5854 alone, this could plausibly capture most of the effect, and it is the main threat to the claim that the joint model is needed there. The experiments settle it directly. If a simple correction, such as fitting widths on noise-matched smoothed profiles, reproduces the joint model's corrected index within errors, we will say so. The remaining unique contributions would then be the hidden-component fluence limit and the tag-and-probe result on simultaneous wideband data, and the simple correction cannot deliver either.

Proposed Experiments and Decisive Checks

Experiment 1: injection coverage. Simulated stacks cover each instrument's signal-to-noise and bandwidth, with the true width index ranging from −0.5 to +2 and a hidden-component fraction from 0 to 50%. The check is that the joint model's 90% intervals contain the truth about 90% of the time. We also measure how far the naive power-law index drifts from the truth as signal-to-noise falls.

The decisive check is simple: if at J0250-like noise levels the naive fit shows no bias, the signal-to-noise part of the claim is false.

Experiment 2: positive control. Fit J1926-0652. The method must favour standard mapping. We will fix all analysis choices before touching the anomalous pulsars.

Experiment 3: J0250+5854. Report three things: the corrected width index, the likelihood ratio among the three hypotheses, and the upper limit on hidden-component fluence at 57 and 149 MHz. These are compared against the published $B = 1.9 \pm 0.4$. The result is a success either way it falls. A corrected index consistent with standard mapping would dissolve the anomaly. A surviving positive index with a tight fluence limit would rule out the vanishing-outer-component explanation.

Experiment 4: tag-and-probe on simultaneous wideband data. The targets are Parkes UWL (J0614+2229, two epochs) and FAST UWB (B1859+07). We first measure how often per-band $\Delta\chi^2$ labels disagree with the tag-band labels, as a function of sub-band.

The decisive check is what that disagreement rate looks like and what it does to the heights. If it stays near zero in every sub-band, tag-and-probe offers no advantage, and we will report the frequency-independent heights as confirmed. If it grows toward the band edges, we recompute the heights with tag labels and test whether their frequency slope is still zero.

For B1859+07, the joint model gets a mixed-state term so that double-vision pulses are fitted as mixtures rather than forced into one state. We then ask whether the $\nu^{-0.169}$ offset law survives.

Experiment 5: stress test on J1948+3540. Fit the symmetric-component model to the uGMRT 400/750 MHz and FAST 1250 MHz data. A clear failure of goodness-of-fit is the expected outcome and will be reported as a limit of the method. Success of an asymmetric-component extension would be the more interesting finding.

References

  1. Search for the Decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>∓</mml:mo></mml:msup><mml:msup><mml:mi>μ</mml:mi><mml:mo>±</mml:mo></mml:msup></mml:math>
  2. Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:msubsup><mml:mi>Λ</mml:mi><mml:mi>c</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mi>X</mml:mi><mml:msup><mml:mi>ℓ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>ℓ</mml:mi></mml:msub></mml:math>decays in events with a fully reconstructed<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>meson
  3. Search for the rare decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:mi>K</mml:mi><mml:mi>ν</mml:mi><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>
  4. Search for <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> meson decays into <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="normal">Λ</mml:mi></mml:math> and missing energy with a hadronic tagging method at Belle
  5. A broad-band radio study of PSR J0250+5854: the slowest spinning radio pulsar known
  6. PSR J1926-0652: A Pulsar with Interesting Emission Properties Discovered at FAST
  7. The Study of Mode-switching Behavior of PSR J0614+2229 Using the Parkes Ultra–wide-bandwidth Receiver Observations
  8. The Study of Unusual Emission from PSR B1859+07 using FAST
  9. The Frequency-dependent Modulation Features of PSR J1948+3540
  10. Dual-frequency single-pulse study of PSR B0950+08
  11. Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics