Pooling Thousands of Faint Bursts to Find Hidden Polarization Flips

Radio emission from pulsars and fast radio bursts (millisecond flashes of radio waves from distant galaxies) leaves the source's magnetized plasma in one of two "normal modes." These modes are polarized exactly 90° apart. When the dominant mode changes, the measured polarization angle jumps suddenly by about 90°. Such jumps are among the few direct probes of conditions near a neutron star. They are also rare in the data we have. Only 3 of more than 2,500 bursts from the repeater FRB 20201124A showed one, all in the brightest bursts, and nobody knows whether fainter bursts lack jumps or simply lack the signal-to-noise to show them (Sudden Polarization Angle Jumps of FRB 20201124A). This proposal treats that question as a scaling problem. Right now each burst is analyzed on its own, much like a lookup table in reinforcement learning (RL) that keeps one independent entry per state. The standard remedy in RL is a shared model that pools evidence across items. We propose to build one here, and to hard-wire the one piece of physics that is exactly known: the two modes differ by 90°.

In two sentences: we fit a single population model to every burst at once, in which each burst's polarization is the sum of two non-negative mode intensities whose angles are locked 90° apart. We then ask whether the pooled model finds statistically real mode flips in faint bursts that per-burst analysis cannot see, and how the flip rate depends on burst energy.

Core Claim

A two-mode model with a hard 90° constraint and non-negative mode amplitudes, fit hierarchically across a whole burst sample, can detect orthogonal jumps at lower signal-to-noise than independent per-burst inspection. This would turn "3 detections" into a measured jump rate as a function of energy, with calibrated uncertainty. The RL literature on function approximation is the methodological lens here, not the application. Three results from it shape specific design choices.

The first result concerns hard versus soft constraints. MOORE builds a Gram-Schmidt layer into its network, which forces the experts' features to be exactly perpendicular. It reports 88.7% success on MetaWorld MT10 against 85.4% for PaCo, and reports that soft similarity penalties needed delicate tuning. This motivates encoding the 90° relation exactly rather than as a prior. The second result concerns the type of factorization. Hierarchical Successor Representation replaces SVD (singular value decomposition, a factorization into perpendicular components that can take either sign) with non-negative matrix factorization. SVD's components smeared across the maze's natural boundaries, while the non-negative, localized parts did not. This motivates non-negative mode intensities, which is also what incoherent addition of the two modes physically requires. The third result concerns keeping updates from interfering. ODICE shows that projecting one gradient onto the plane perpendicular to another stops the two from cancelling. We reuse this idea narrowly: jump-time parameters and smooth angle-sweep parameters are updated so that they do not cancel each other.

None of these papers touches polarimetry. Our contribution is not a new RL architecture. It is the transfer of their lessons about pooling and structure to a concrete physics inference problem where they have not been tried.

Research Program

The forward model describes each burst's de-dispersed dynamic spectrum in the four Stokes parameters. These are the standard measured quantities: total intensity $I$, the two linear-polarization components $Q$ and $U$, and circular polarization $V$. Over time $t$ within a burst, the linear polarization is modelled as

$$Q(t) + iU(t) = \big(a_O(t) - a_X(t)\big)\, e^{2i\psi(t)}\, e^{2i\,\mathrm{RM}\,\lambda^2},$$

where $a_O, a_X \ge 0$ are the intensities of the two modes, $\psi(t)$ is the angle of the O mode, $\mathrm{RM}$ is the burst's Faraday rotation measure (how much the intervening magnetized plasma rotates the angle), and $\lambda$ is wavelength. This form comes from adding two incoherent components at angles $\psi$ and $\psi + 90°$. Because of it, an observed jump shows up as a sign change of $a_O - a_X$, with a dip in linear polarization at the crossing. That is the depolarization signature described in the FAST pulsar literature. We will fit a free circular term $V(t)$ and track it separately. This matters because the DSA-110 sample attributes strong circular polarization to Faraday conversion, a propagation effect in dense magnetized plasma that turns linear polarization into circular (Deep Synoptic Array Science: Polarimetry of 25 New FRBs). The model should not mistake that effect for a mode switch.

The population layer ties bursts together. Per-burst angle sweeps, RMs and jump times are drawn from shared distributions. The jump rate is a smooth function of burst energy and fluence. Inference is amortized. A neural encoder is trained on simulated bursts produced by this forward model, with noise drawn from real off-pulse data, and it outputs posteriors for each burst's parameters. We then fit the population layer over those posteriors. The encoder only exists to make the computation tractable at thousands of bursts, and later at millions of single pulsar pulses. We will show it matches a slow, exact sampler on a subset.

Burst #1 in the FRB 20201124A sample jumped 65.9° ± 1.1°, which is inconsistent with exactly 90°. This conflicts with a hard constraint, so the model must handle it. We will include a nested alternative in which the separation angle $\Delta$ between the modes is free. We then ask whether the non-90° case is explained by the two modes overlapping within a single time sample under the hard model, or whether it genuinely prefers $\Delta \neq 90°$. A genuine preference would itself be a physical result. It would point toward propagation effects such as generalized Faraday conversion.

Baselines

The first reference point is the naive one: using no mode model and simply stacking bursts or smoothing position angles. We will align bursts on their peaks and fit a single angle curve. This shows how much is lost when a sample with varying geometry is pooled without structure.

The strongest existing method is per-burst time-resolved analysis: Stokes inspection, a fit of the position-angle curve and RM, and a jump flagged when the angle change is significant. This is the approach of the FRB 20201124A study. On that sample it reports 3 jump detections in more than 2,500 bursts, with jumps of 65.9° ± 1.1°, 97.7° ± 23.9° and 88.7° ± 30.0° on roughly 1 ms timescales (Sudden Polarization Angle Jumps of FRB 20201124A). We will re-implement it with DSPSR and PSRCHIVE, the pulsar processing and calibration packages used for the FAST dwarf-pulse work (Dwarf Pulses of 10 Pulsars). It will get the same calibrated data, the same RM search, and its threshold tuned to match our false-alarm rate on jump-free simulations.

Beating it means three things. We must recover all three published jumps. We must reach higher detection completeness at matched false-alarm rate in an injection-recovery test. And we must produce population jump rates in lower-energy bins whose uncertainty excludes zero, or excludes the bright-burst rate, which the baseline cannot do. For the DSA-110 sample the file reports polarization fractions but no per-burst jump analysis, so no reported result was found on Lacuna for DSA-110 jump detection.

Two ablations isolate the key ingredients. The first keeps the full hierarchical model but replaces the hard 90° link with a soft prior whose strength is tuned. The second keeps the hard two-mode model but fits each burst independently, with no pooling. If the pooled model wins only because of the orthogonality structure, or only because of the pooling, the ablations will show which.

Why not just…?

Why not just fit a classical hierarchical Bayesian model with MCMC and skip the neural encoder? This is the most serious alternative, and we expect it to be statistically equivalent where it can be run. The claim is that structure plus pooling does the work, not the network. We will run full sampling on a few hundred bursts and require the amortized posteriors to match it within calibration tolerance. The encoder is justified only by cost at the scale of the FAST samples and of single-pulse pulsar data. If exact sampling is affordable at full scale, we will simply use it.

Why not just lower the per-burst detection threshold? Lowering the threshold trades completeness for false alarms along a fixed curve. Pooling moves the curve. The injection-recovery experiment answers this directly by comparing completeness at matched false-alarm rate. If a lowered per-burst threshold performs as well, the pooling claim fails.

Why not use generic unsupervised decomposition, such as PCA or clustering on Stokes spectra? PCA components can take either sign and spread across the whole burst. This mirrors the smearing that Hierarchical Successor Representation found for SVD, though that evidence comes from gridworlds, not radio data. Those components also say nothing about RM or angles, so any "mode" they find needs a physical interpretation after the fact. We will include PCA and non-negative factorization without the 90° link in the injection test. If either matches the physical model, the physics constraint is not earning its keep.

Why not just a soft orthogonality penalty? MOORE found that soft penalties were fragile relative to hard Gram-Schmidt layers, but that was in robot control, not polarimetry. The soft-prior ablation tests it here. Burst #1's 65.9° jump also means a soft model could legitimately win, and we will report it if it does.

Proposed Experiments

We validate first on pulsars, where the answer is known. PSR B1944+17's 234 dwarf pulses lie mostly in the mode orthogonal to its normal pulses (Dwarf Pulses of 10 Pulsars). PSR B1859+07's Af mode sits in the orthogonal mode that matches the anomalous A mode (A New Emission Mode of PSR B1859+07). The model must assign these faint pulses to the correct mode without labels. We will also degrade bright pulses to dwarf-pulse signal-to-noise and check whether the mode assignment survives.

The central experiment is injection and recovery on FRB 20201124A. We will inject jumps with controlled size, timing and duration into real burst noise across the energy range, and measure detection completeness and false-alarm rate against signal-to-noise for all methods under the same compute budget. We will then run the model on the real sample and report the inferred jump rate against energy.

The decisive check has two possible outcomes. If faint bursts show a nonzero jump rate comparable to bright ones, the rarity of detections was a sensitivity effect. If the rate is significantly suppressed at low energy, then mode switching is tied to luminosity, which would constrain models of what triggers it. Either answer addresses the open question in the file.

As an out-of-sample test, we will apply the model to the DSA-110 non-repeaters with the dsa110-pol calibration pipeline (Deep Synoptic Array Science: Polarimetry of 25 New Fast Radio Bursts Provides Insights into their Origins Deep Synoptic Array Science: Polarimetry of 25 New Fast Radio Bursts Provides Insights into their Origins). There we will check whether the separate circular-polarization term absorbs the two high circular-fraction bursts without producing false jumps. Access to the raw FAST burst archives is the main practical risk. If full-resolution Stokes data are unavailable, the pulsar validation and DSA-110 analysis stand on their own.