Fast radio bursts (FRBs) are millisecond radio flashes from distant galaxies. A leading idea is that they come from young neutron stars born in supernovae. The obvious test is to look for an old supernova at the position of each burst. The trouble is that an old transient sitting at the right place, and exploding before the burst, is only a correlation. Deeper optical surveys will make such coincidences common, and a star-forming host galaxy can produce many supernovae and many bursts without any one supernova causing any one burst. There is also a physical wrinkle. The ionized debris of a young supernova absorbs low-frequency radio waves, so a burst engine inside it stays hidden until the debris thins. At CHIME's 600 MHz this takes an estimated 6–10 years (Searching for Historical Extragalactic Optical Transients Associated with Fast Radio Bursts).
We propose to turn that delay from a nuisance into the test. The absorption depends on radio frequency in a fixed way, so a real supernova–burst link predicts a specific signature: surveys at higher frequencies should find their associated supernovae systematically younger, by a ratio set by plasma physics and independent of each explosion's mass and speed. Chance alignments and host-galaxy confounding predict no such frequency dependence. The contribution is a frequency-contrast association test, validated on simulated catalogs with injected ground truth. It borrows the causal-inference habit of asking what pattern a common cause could not produce, rather than accepting "the transient came first."
Core Claim
The absorbing process is free-free absorption: free electrons in the ionized shell soak up low-frequency radio waves as they scatter off ions. Its optical depth scales as $\tau \propto n_e^2 T^{-1.35} \nu^{-2.1} L$, where $n_e$ is the electron density, $T$ the gas temperature, $\nu$ the observing frequency, and $L$ the path length. For a shell of mass $M$ expanding freely at speed $v$, we have $n_e \propto M/(v t)^3$ and $L \propto v t$. At fixed temperature this gives
$$\tau(\nu,t) \propto M^2 v^{-5}\, t^{-5}\, \nu^{-2.1}, \qquad t_{\rm trans}(\nu) \propto M^{0.4} v^{-1}\, \nu^{-\beta},\ \ \beta \approx 0.42,$$
where $t_{\rm trans}$ is the age at which $\tau$ falls to unity. The mass and velocity factors scatter the transparency age from one explosion to the next, but they cancel in the ratio between two bands. The median age of genuinely associated supernovae should therefore scale as $(\nu_1/\nu_2)^{-\beta}$ across surveys. With the 6–10 yr figure at 600 MHz, this predicts roughly 4–7 yr at 1.4 GHz and 3–5 yr at 3 GHz. These are our extrapolations under the stated scaling, not reported values.
Temperature evolution and a non-uniform shell will shift $\beta$, so we treat $\beta$ as a free parameter with a physically bounded prior rather than a fixed number. The claim is not that $\beta$ is known. The claim is that under chance alignment, including confounding by host galaxy, the age distribution of matched transients is flat in $\nu$ once survey footprint and localization are matched. A causal link instead produces a monotone, power-law shift. That contrast is a test that spatial and temporal coincidence alone cannot provide.
Relation to Prior Work
The closest work is the historical cross-match of 176 localized FRBs against about 142,000 optical transients (Searching for Historical Extragalactic Optical Transients Associated with Fast Radio Bursts). It combines positional and redshift matching with a Monte Carlo chance baseline built from 415,000 fake FRBs. It then applies the transparency delay as a hard 6–10 yr rejection window at a single frequency. It finds no significant associations beyond FRB 20180916B / AT 2020hur, and it finds that only about 7% of matched transients are old enough to be viable. This paper supplies our null-construction method and our real-data sample. We differ by replacing the single hard window with a frequency-dependent transmission model, $\exp(-\tau(\nu,t))$, and by using the frequency dependence as the discriminating statistic.
The forward population code frbpoppy (Synthesising the repeating FRB population using frbpoppy) shows that a single repeating population can reproduce CHIME statistics that others read as two populations. The luminosity-function analysis of CHIME Catalog 1 (Luminosity distribution of fast radio bursts from CHIME/FRB Catalog 1) reads the same kind of data as a real energetic gap. Together they show how easily one-dimensional summary statistics are fit by very different source models. This result constrains our design: a useful association test must be robust to a wrong luminosity function. The frequency ratio cancels per-event nuisances precisely to get that robustness.
The causal-inference literature motivates the framing more than it supplies the method. Neuroscience examples show that a signal preceding an outcome does not establish that it drives the outcome. Motor cortex leading locomotion by 136 ms (Locomotion-Related Population Cortical Ca2+ Transients in Freely Behaving Mice) and pre-stimulus EEG phase predicting perception (The Phase of Ongoing EEG Oscillations Predicts Visual Perception) both fit this pattern. The audit of interpretability claims (Causality is Key for Interpretability Claims to Generalise) found 53.5% of claims asserting a higher causal level than their evidence supports. The supernova–FRB case is the same trap: every old transient trivially precedes the burst. Our test targets a mechanism-specific prediction instead.
We deliberately do not make general graph-learning algorithms the core, for reasons given below. Those algorithms are iterative causal discovery for hidden confounders (Iterative Causal Discovery in the Possible Presence of Latent Confounders and Selection Bias), rank constraints under selection (Latent Variable Causal Discovery under Selection Bias), and exact Bayesian network scoring in astrophysics (Causal Discovery in Astrophysics: Unraveling Supermassive Black Hole and Galaxy Coevolution).
Research Program
We will build a simulator in three layers.
The first layer generates FRB sources with frbpoppy, observed by surveys at roughly 600 MHz, 1.4 GHz, and 3 GHz. Each survey has its own sky coverage, localization errors, and redshift reach.
The second layer assigns a fraction $f$ of sources to a parent core-collapse supernova. Each parent gets an explosion age and draws of $M$, $v$, and $T$. A burst is emitted only with probability $\exp(-\tau(\nu,t))$. A confounding switch lets host star-formation rate raise both the local supernova density and the ejecta column.
The third layer places background optical transients using the sky and time distribution of the real transient catalogs. It degrades this background toward what an LSST-depth survey would find, so chance-match density can be varied.
On these catalogs we compute a pair-level statistic. For each candidate FRB–transient pair it combines the positional likelihood ratio with the transmission-weighted age likelihood under a shared $\beta$. We then compare the fitted band-dependence of matched ages to a null. The null is built exactly as in the historical cross-match: fake FRBs are redrawn per survey, keeping each survey's own footprint and localization, which removes survey-dependent confounding by construction. The outputs are a significance for "associations exist," an estimate of $f$ with an interval, and an estimate of $\beta$. The last is itself a physics measurement.
On the real 176-FRB sample we expect upper limits on $f$ per band, not detections. The historical cross-match already reports no significant associations. We will report those limits honestly and pair them with a power curve: how many localized FRBs, and at what optical depth, are needed before the frequency contrast can separate a causal fraction $f$ from zero.
Baselines
The naive reference is positional plus redshift coincidence counting, with no age or frequency information, calibrated against fake-FRB injections. It measures how much any physics-based weighting adds.
The strongest existing method is the procedure of Searching for Historical Extragalactic Optical Transients Associated with Fast Radio Bursts: 5σ spatial matching, redshift consistency, and a 6–10 yr minimum-age cut, with a Monte Carlo chance baseline. Its reported results are no significant associations within 5σ localizations apart from FRB 20180916B / AT 2020hur, about 7% of matched transients old enough to be viable, and an estimate that roughly 22,700 well-localized FRBs would be needed for one chance match. We found no reported detection-power or false-positive-rate figure on Lacuna for this procedure, so we will re-implement it in our simulator. We give it its most favorable form: the age window is retuned per band using the correct simulated $t_{\rm trans}$, which it does not know in practice, and it runs on the same catalogs and the same fake-FRB budget. Beating it means higher power at a matched false-positive rate of 1%. Critically, it also means keeping that false-positive rate under host-galaxy confounding, where we expect a window method to inflate.
The main ablation is our statistic with the frequency contrast removed. The transmission-weighted age likelihood is kept, but $\beta$ is fixed to zero, so all bands share one age distribution. This isolates whether the chromatic signature, rather than soft age weighting, does the work.
A second ablation fits a full selection-corrected hierarchical Bayesian likelihood jointly over luminosity function, $f$, $\beta$, and the per-event $M$ and $v$. This is the reviewer's strongest alternative and is discussed below.
Finally, we run iterative causal discovery from the IntelLabs Causality-Lab on pair-level features (age, band, host star-formation rate, dispersion measure, detection). This checks whether generic structure learning recovers the association at realistic sample sizes.
Why not just…?
Why not just use forward population synthesis, as the field already does? We do use it; frbpoppy is our generator. The question is what statistic to compare against. Its own authors show that very different luminosity assumptions reproduce the same survey summaries, and the CHIME luminosity study reaches the opposite population conclusion from similar data. A statistic that cancels per-event mass, velocity, and luminosity should survive that degeneracy. The experiment tests this directly: we fit with a wrong luminosity function and check which method stays calibrated.
Why not just require the supernova to precede the burst by the transparency time? That is the hard-window baseline. Precedence is satisfied by every old transient in the field, just as motor-cortex activity preceding movement cannot exclude a shared upstream driver. A window also discards the band-to-band information that distinguishes causal from confounded matches. We expect it to lose power and to leak false positives under confounding, and the injected-confounder runs will show whether it does.
Why not just run a general causal discovery algorithm on the catalog? With at most one tentative association among 176 bursts, conditional-independence tests have almost no power. Exact Bayesian scoring also assumes no hidden confounders, which is exactly the problem here. Rank-constraint methods assume linear relations, while transmission is exponential in $t^{-5}$. We include iterative causal discovery as a baseline because its hidden-confounder handling is the right idea. We predict it will underperform a statistic that encodes the known physics, and we will report the result either way.
Why not just fit the full hierarchical Bayesian model? This is a real threat. A correctly specified joint likelihood contains our frequency contrast as a special case and could be more powerful. Our bet is that it is fragile when the luminosity function or the confounding structure is misspecified, while the ratio test is not. If the full model stays calibrated under misspecification, we adopt it. The contribution then narrows to identifying the frequency contrast as the component that carries the causal evidence, which we would show by comparing each method with and without that term.
Decisive Checks
The first check is calibration under the null. With $f = 0$ and host confounding switched on, our test must hold its nominal 1% false-positive rate across survey mixes, and we will report where the hard-window baseline fails to.
The second check is recovery. With $f > 0$ and $\beta$ drawn from its prior, the fitted $\beta$ must be unbiased, and power must exceed both the window baseline and the $\beta = 0$ ablation at LSST-like chance-match densities.
The third check is misspecification. With temperature evolution and asymmetric shells added so the true exponent drifts, the test should lose power gracefully rather than produce false detections.
On real data, the deliverable is per-band upper limits on the supernova-born fraction from the 176-FRB sample. We will also give the projected number of localized multi-band FRBs at which a nonzero $f$ becomes detectable. If that number is smaller than the roughly 22,700 bursts needed for one chance match, the chromatic test is worth building survey strategies around. If it is not, that is a useful negative result for the field.
References
- Searching for Historical Extragalactic Optical Transients Associated with Fast Radio Bursts
- Synthesising the repeating FRB population using frbpoppy
- Luminosity distribution of fast radio bursts from CHIME/FRB Catalog 1 by means of the updated Macquart relation
- Locomotion-Related Population Cortical Ca2+ Transients in Freely Behaving Mice
- The Phase of Ongoing EEG Oscillations Predicts Visual Perception
- Causality is Key for Interpretability Claims to Generalise
- Iterative Causal Discovery in the Possible Presence of Latent Confounders and Selection Bias
- Latent Variable Causal Discovery under Selection Bias
- Causal Discovery in Astrophysics: Unraveling Supermassive Black Hole and Galaxy Coevolution