A spinning black hole can drive a jet by twisting magnetic field lines threaded through the region around it. The jet carries off the hole's rotational energy, and in the process it builds strong, ordered magnetic fields at the jet's base. Those same fields are the medium in which a hypothetical light particle, the axion-like particle (ALP), would show itself. An ALP is a very light boson that a gamma-ray photon could turn into, and back from, while crossing a magnetic field. The conversion would leave irregular "wiggles" or extra transparency in the gamma-ray spectrum of an active galaxy. Current gamma-ray searches with ground-based Cherenkov telescopes report coupling bounds of roughly $10^{-11}$ to $10^{-12}$ GeV$^{-1}$ (Axion-like Particle Searches with IACTs). Those bounds rest on idealized, statistically random magnetic field models, and the field model is the dominant and least checked assumption.
Radio interferometry and polarization measurements now constrain the field geometry directly near the jet base. Examples are the Event Horizon Telescope (EHT) images of M87* and Sgr A* and the ordered jet-base fields in 3C 84. We propose to treat ALP conversion the way particle physicists treat a candidate new decay channel. It becomes a coherent amplitude sitting on top of a known background, fitted photon by photon with an unbinned likelihood. The magnetic field enters as a constrained nuisance distribution, built from polarimetry-calibrated simulations of magnetically dominated jets. It is not a free random field. The sharpest new handle is time. When a flaring jet's optical polarization angle rotates, the field orientation along the path rotates with it. Conversion in the jet should then change the spectral irregularity pattern in step with the rotation. Conversion in a static galaxy-cluster field cannot do this.
Core Claim
ALP limits from active galaxies depend on the magnetic field model. For jet-region conversion, that field is now partly measured rather than assumed. We claim three things. Conditioning the conversion model on polarimetry-calibrated field ensembles changes the inferred bounds by an amount comparable to the quoted statistical precision; it may tighten them or honestly loosen them, and either result matters. An event-level likelihood borrowed from rare-decay analyses is the right statistical tool for the few-hundred-photon flare samples where jet-region conversion matters most. A jet field that varies in time, traced by rotations of the polarization angle (the electric vector position angle, EVPA), gives a signature that separates jet-region conversion from cluster-scale conversion and from source-intrinsic spectral structure.
The physics behind the claim is simple. In a single coherent field domain, at weak mixing and above the critical energy, the photon-to-ALP conversion probability is
$$P_{\gamma\to a}\approx \left(\tfrac{1}{2}\,g_{a\gamma\gamma}\,B_T\,L\right)^2 ,$$
where $g_{a\gamma\gamma}$ is the photon-ALP coupling, $B_T$ is the field component transverse to the photon's direction, and $L$ is the domain length. Only the photon polarization component lying along $B_T$ mixes, so the field's orientation matters, not just its strength.
Polarimetry measures that orientation. It does not measure the field's sign. Under $B\to -B$ the conversion probability is unchanged. This means the ALP observable is insensitive to the ambiguity that most troubles horizon-scale interpretation. In First Sagittarius A* Event Horizon Telescope Results. VIII. Physical Interpretation of the Polarized Ring, the inferred sense of plasma rotation flips depending on whether the Faraday rotation (the twisting of polarization by magnetized plasma along the line of sight) occurs inside or outside the emission region. Our observable needs the orientation after de-rotation, which depends on the magnitude of the rotation measure (the size of that Faraday twist). It does not depend on the sense of plasma rotation.
Research Program
Field ensembles: the rotational-energy-extraction side. The EHT results establish that the relevant engines are magnetically arrested disks (MADs), accretion flows whose inner region is held up by accumulated magnetic flux. The M87 ring shows an azimuthal polarization swirl with mode amplitude $\beta_2\approx 0.04$–$0.07$ (First M87 Event Horizon Telescope Results. VII. Polarization of the Ring). Sgr A* is fit by a MAD model with spin $a_*=0.94$ (First Sagittarius A* Event Horizon Telescope Results. VII. Polarization of the Ring). 3C 84 shows high magnetic flux and ordered fields at the jet base (Ordered magnetic fields around the 3C 84 central black hole).
We will take general relativistic magnetohydrodynamic (GRMHD) jet simulations, which model magnetized plasma flowing around a spinning black hole. We will keep only the snapshots that pass the polarimetric scoring these papers use: net linear and circular polarization and the $\beta_m$ modes. We will then follow each surviving field along rays leaving candidate gamma-ray emission zones. For PKS 1510−089 we add the parsec-scale constraint from Multiwavelength observations of a VHE gamma-ray flare from PKS 1510−089 in 2015. That paper places a very-high-energy emission region of size about $2.8\times10^{16}$ cm outside the broad-line region, and its flare coincided with a roughly $100^\circ$ optical polarization-angle rotation and the ejection of a superluminal knot. Simulations supply the field beyond the emission zone. Polarimetry anchors the field's orientation and its rotation over time.
The output is a set of field ensembles. Each ensemble is weighted by its polarimetric likelihood and indexed by source and epoch, giving $B_T(s,t)$ along the line of sight as a function of path length $s$ and observation time $t$. The 3C 84 ensemble is the bridge, since its host NGC 1275 is one of the existing ALP targets, which allows a direct like-for-like comparison.
Amplitude-style likelihood: the decay-mode side. We propagate the three-state system (two photon polarizations plus the ALP) through each field realization. We then fit Cherenkov-telescope photon lists, meaning per-photon reconstructed energy and arrival time, with an unbinned extended maximum likelihood. This is the same estimator rare-decay analyses use when yields are small, as in the BESIII and BABAR measurements in the file (Amplitude analysis and branching fraction measurement of the decay D^+ \to K_S^0\pi^+\pi^0\pi^0; Observation and polarization measurement of B^0 \to a_1(1260)^+ a_1(1260)^- decay).
Concretely, the likelihood is
$$\mathcal{L}=e^{-\nu}\prod_{i=1}^{N}\int dB\;\pi(B)\,\rho(E_i,t_i\mid g_{a\gamma\gamma},m_a,\theta,B),$$
where $N$ is the number of observed photons, $\nu$ is the expected total count, $\pi(B)$ is the polarimetry-weighted field ensemble, $m_a$ is the ALP mass, and $\theta$ collects the intrinsic spectrum, extragalactic absorption, and instrument-response parameters. The predicted density $\rho$ folds the per-photon survival probability through the instrument's energy dispersion.
The analogy to isobar amplitude analysis is more than decorative. In both cases the signal is an interference pattern. Its shape depends on a geometric nuisance: resonance line shapes and helicity angles in one case, domain orientations in the other. Mis-modeling that nuisance creates fake signals. The $K_1(1270)$/$K_1(1400)$ degeneracy in the $D^+$ analysis is exactly this kind of model-dependence. A neural likelihood-ratio estimator trained on simulated photon lists is a secondary path, in case integrating over the field ensemble proves too expensive.
Time-correlation test. For flaring sources we fit a joint model in which the field orientation in the jet region tracks the measured EVPA curve, possibly with a lag. The cluster and intergalactic contributions are held static. A jet-region ALP signal predicts that the energies of spectral features shift and their depth changes coherently with EVPA rotation. Intrinsic spectral curvature and static cluster mixing predict no such correlation.
Baselines
The naive reference point is the no-ALP null. This is an intrinsic smooth spectrum (log-parabola or power law with cutoff) times standard extragalactic absorption, fitted to the same photon lists with the same instrument response. All test statistics are quoted relative to this null. Its goodness of fit, checked with photon lists simulated from the best-fit null, also tells us how much irregularity the data actually show.
The strongest existing method is the published Cherenkov-telescope ALP analysis summarized in Axion-like Particle Searches with IACTs. That analysis uses spatially integrated spectra and idealized random-cell turbulent fields, mainly cluster or galactic. It reports bounds on $g_{a\gamma\gamma}$ of about $10^{-11}$ to $10^{-12}$ GeV$^{-1}$ over sub-µeV mass ranges. It is the right competitor because it is what the field currently trusts. To compare fairly, we will rerun its field prescription inside our own pipeline, on the same NGC 1275 and PKS 2155−304 photon samples where accessible, or on high-level spectral data otherwise. Instrument response, absorption model, mass grid, and the number of field realizations will all match ours. Our method beats it if two things hold. First, on simulated data with a known injected coupling, our bound has correct frequentist coverage when theirs does not. Second, the expected bound at fixed coverage is narrower, or the observed bound moves by more than its own one-sigma statistical band. "Tighter on real data" alone does not count, because a wrong field model can also produce tighter bounds.
The ablation that isolates our key ingredient keeps everything identical but replaces the polarimetry-weighted ensemble $\pi(B)$ with the unweighted GRMHD ensemble. A further ablation replaces it with a single analytic helical jet field matched only to the mean EVPA. If the unweighted or helical versions recover our bounds and coverage, the polarimetric conditioning is not doing work.
A second ablation swaps the unbinned likelihood for the standard binned spectral fit. This isolates the decay-mode statistical contribution as a function of photon count.
Why not just…?
Why not just keep the random cell-turbulent cluster field and add more realizations? For NGC 1275 this is a real threat. The cluster field's product of strength and coherence length may dominate conversion over the jet's, in which case jet-field conditioning changes little there. Our first experiment is therefore a conversion budget split by region for each source, computed from the ensembles. If the jet region contributes negligibly to $P_{\gamma\to a}$ for NGC 1275, we report that and concentrate on sources without rich-cluster environments, such as PKS 1510−089 and PKS 2155−304. There the existing analyses must assume galactic or intergalactic fields, and the jet is the most strongly magnetized structure on the path. More realizations of the wrong field distribution cannot recover the time-correlated signature, which only a jet field produces.
Why not just use a simple helical jet field instead of GRMHD ensembles? An analytic helix fixes the domain structure by hand and ignores the turbulent component that MAD simulations show and that produces the depolarization used to disfavor low-magnetization models in the Sgr A* analysis. That component sets how coherent the conversion is. The helical ablation in our Baselines tests this directly. If it matches coverage and bounds, we adopt it as the cheaper model and the paper's contribution narrows to the statistical method and the timing test.
Why not just apply a deep classifier to the spectra, as in collider searches? Learned event-level classifiers do beat engineered summaries on simulated benchmarks. Deep Learning From Four Vectors reports AUC 0.880 against 0.733 for shallow networks and 0.73 for boosted decision trees on the Higgs benchmark. Searching for Higgs Boson Decay Modes with Deep Learning reports a smaller gain, 0.803 against 0.798. Those gains come from millions of training events in high-dimensional final states, whereas a gamma-ray flare offers two observables per photon and a few hundred photons. A classifier also does not itself deliver a calibrated coupling bound marginalized over the field. We include the neural likelihood-ratio estimator only as a computational substitute and judge it by coverage on simulated data. There is no reported result on Lacuna for any classifier on ALP gamma-ray spectra.
Why not just look for ALP effects in the millimeter polarization itself? At EHT frequencies and plausible couplings, conversion in jet fields is negligible by the scaling above. The mixing becomes strong only above a critical energy, so the polarimetry serves as a field measurement rather than a signal channel.
Proposed Experiments
The first experiment is the region-by-region conversion budget for NGC 1275/3C 84, PKS 2155−304, and PKS 1510−089. It decides which sources the program is about.
The second is an injection-recovery study on simulated photon lists. Lists are generated from field realizations held out of the fitting ensemble. On these we measure coverage and expected bounds for four approaches: the null, the published-style field prescription, the unweighted ensemble, the helical field, and the full method, each under binned and unbinned fits. This is the decisive check that the conditioned ensemble is calibrated rather than merely tighter. If coverage of the published-style prescription is already correct on jet-generated data, the motivation weakens and we say so.
The third experiment applies the full method to available photon data for NGC 1275 and PKS 2155−304 and reports the bound shift against the published range.
The fourth is the time-correlation test on the 2015 PKS 1510−089 flare, using its measured roughly $100^\circ$ EVPA rotation and the factor of about 5 flux increase. We quote the sensitivity the flare's photon count allows and treat the ±52-day uncertainty in the knot birth epoch as a lag nuisance. A null result here still yields a field-conditioned bound from the source's jet region. A positive correlation would be qualitatively new evidence, one no static-field model could mimic. It would require confirmation in a second flare before any claim.
References
- Axion-like Particle Searches with IACTs
- First Sagittarius A* Event Horizon Telescope Results. VIII. Physical Interpretation of the Polarized Ring
- First M87 Event Horizon Telescope Results. VII. Polarization of the Ring
- First Sagittarius A* Event Horizon Telescope Results. VII. Polarization of the Ring
- Ordered magnetic fields around the 3C 84 central black hole
- Multiwavelength observations of a VHE gamma-ray flare from PKS 1510−089 in 2015
- Amplitude analysis and branching fraction measurement of the decay $D^{+} \to K_S^0π^+π^0π^0$
- Observation and polarization measurement of<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:mo>→</mml:mo><mml:msub><mml:mi>a</mml:mi><mml:mn>1</mml:mn></mml:msub><mml:mo stretchy="false">(</mml:mo><mml:mn>1260</mml:mn><mml:msup><mml:mo stretchy="false">)</mml:mo><mml:mo>+</mml:mo></mml:msup><mml:msub><mml:mi>a</mml:mi><mml:mn>1</mml:mn></mml:msub><mml:mo stretchy="false">(</mml:mo><mml:mn>1260</mml:mn><mml:msup><mml:mo stretchy="false">)</mml:mo><mml:mo>−</mml:mo></mml:msup></mml:math>decay
- Deep Learning From Four Vectors
- Searching for Higgs Boson Decay Modes with Deep Learning