Fast radio bursts are millisecond flashes of radio emission, mostly from other galaxies, whose origin is still unknown. Pinning a burst to its host galaxy requires several widely separated radio stations to record the same event at full time resolution. Each station can hold that raw data for only seconds: CHIME keeps a roughly 20-second voltage buffer (The CHIME Fast Radio Burst Project: System Overview). The links between stations are far too slow to stream raw data. Today, each station searches alone and reports only bursts that clear a fixed local signal-to-noise threshold. The network therefore never sees a burst that is weak at every station, even when the stations together would have enough evidence to detect it.
We propose that each station send a few bits per candidate. Those bits encode a locally calibrated tail probability, meaning how surprising the candidate is under that station's own measured noise. It would not send a raw signal-to-noise ratio. A coordinator adds the log-probabilities across stations and decides, within the buffer window, whether to trigger a network-wide data dump. The central idea is that calibration, not cleverer compression, is what makes low-bit fusion work on real radio data. Real receiver noise is correlated in time, which inflates raw signal-to-noise at wide pulse widths (Fast radio burst detection in the presence of coloured noise). A station reporting uncalibrated values in that regime acts like the "inept agent" of belief-fusion theory: confident but uninformative. It poisons any sum it enters.
Core Claim
Distributed hypothesis testing asks how several sensors, each seeing part of the data, can decide between "only noise" and "signal present" when they may send only a few bits. For the simplest version, where both hypotheses are fully known distributions, The Sample Complexity of Distributed Simple Binary Hypothesis Testing under Information Constraints proves three things. The cost of a $D$-bit limit per node is only a factor of about $\max(1, \log n^*/D)$ in sample size, where $n^*$ is the unconstrained sample size. Simple threshold quantizers are essentially optimal. Sequential interaction helps by at most a factor of 4. This result supplies our method, and it also constrains us. It says a thresholded report is the right shape of message, so we should not invent elaborate codes.
What the result does not cover is precisely the FRB setting. First, the test is composite: arrival time, dispersion measure (the frequency-dependent delay imposed by intervening plasma) and pulse width are unknown. Second, the noise is coloured, which means correlated in time. Its null distribution for a boxcar search of width $N_{box}$ is a $\chi^2$ scaled by a width-dependent factor $\eta$, and that factor differs from station to station and from hour to hour (Fast radio burst detection in the presence of coloured noise).
Our claim is therefore specific. The right threshold quantizer in this setting applies to the per-station tail probability under the station's own $\eta$-scaled null, not to raw signal-to-noise. Fusing these quantized log-probabilities by summation, which is log-linear pooling, gives a network test with three properties. It stays calibrated when one station's noise degrades. At a fixed bit budget and a fixed network false-trigger rate, it detects bursts whose per-station signal-to-noise lies well below the usual local thresholds.
Research Program
Each station runs its existing dedispersion and boxcar search, for example HEIMDALL-style or BoxDIT/IGRID (Single-pulse Detection Algorithms for Real-time Fast Radio Burst Searches Using GPUs). It also re-estimates $\eta$ for each boxcar width on a rolling window of recent data, following the scaled-$\chi^2$ derivation of the coloured-noise paper. For each search cell $c$ (a combination of time bin, dispersion measure and width), the station computes a tail probability $p_{s,c}$ under its own null. It reports the cell only if $p_{s,c}$ falls below a censoring level $p^{\ast}$, which is set by the bit budget. The report is the cell index plus $\log p_{s,c}$ quantized to $q$ bits.
The coordinator aligns cells across stations using the known dispersive and geometric delays. It then forms the statistic
$$T_c = \sum_{s=1}^{S} \min\left(-\log \hat p_{s,c},\ -\log p^{\ast}\right)\cdot \mathbb{1}[\text{reported}] ,$$
where $S$ is the number of stations and $\hat p_{s,c}$ is the dequantized probability. Unreported stations contribute zero, which is the correct censored term. The coordinator triggers a dump when $\max_c T_c$ exceeds a level chosen for a target false-trigger rate. That level comes from the known null law of a censored sum of exponentials, not from simulation. The statistic is a censored, quantized Fisher combination. Its connection to geometric belief pooling, which On the Arithmetic and Geometric Fusion of Beliefs for Distributed Inference shows is faster than arithmetic pooling and immune to inept agents, is direct. That paper, however, assumes full-precision belief exchange with no bit limit, and that is the piece we add.
The analysis has three parts. The first is an accounting of the bit budget. The expected bits per station per second is roughly $p^{\ast}$ times the cell rate times $(\log_2 \#\text{cells} + q)$. The index cost $\log_2 \#\text{cells}$ dominates, echoing the $\log n^*/D$ term of the simple-binary result. The second is a derivation of the sensitivity loss from censoring and from quantization to $q$ bits, relative to an uncensored oracle sum. This gives a concrete, measurable counterpart to the $\chi^2$-contraction picture of rate-limited channels. The third is a robustness statement: under a per-station miscalibration of $\eta$, the network false-trigger rate stays controlled. Raw signal-to-noise summation has no such guarantee.
We treat the theoretical parts as supporting rather than headline. The headline is the empirical sensitivity-versus-bits curve on realistic noise.
Baselines
The naive reference point is single-station triggering with no network cooperation. Each station dumps whenever its own coloured-noise-corrected statistic crosses a threshold set for the same total false-trigger rate. This baseline shows how much cooperation buys at all.
The standard practice baseline is hard local thresholding plus coincidence. Each station applies a white-noise signal-to-noise cut of 7, as in the Nanshan search, and sends time, dispersion measure and signal-to-noise for each candidate. The coordinator then requires agreement from $k$ of $S$ stations. This mirrors the multi-station coincidence that GReX: An Instrument Overview and New Upper Limits on the Galactic FRB Population lists as future work. It also mirrors the local-threshold logic of the CHIME pipeline levels L1–L3. Under coloured noise, this baseline's bit rate is unbounded. The Nanshan search produced about 14,000 candidates above signal-to-noise 7 in 55 hours (Fast radio burst detection in the presence of coloured noise). We therefore run it both with its natural rate and with its threshold raised until it fits the shared bit budget.
The strongest competitor is coloured-noise-corrected hard coincidence. Each station computes the same $\eta$-calibrated tail probability as our method. It then sends only a one-bit "above threshold" report per cell, with a threshold jointly tuned together with $k$ to maximize completeness at equal bits and equal false-trigger rate. This baseline is the simple-binary theory's recommended quantizer applied with our calibration, so it isolates whether multi-level soft fusion adds anything beyond calibration. It receives the same noise data, the same injected bursts, and the same grid search over its two free parameters as our method receives over $p^{\ast}$ and $q$.
Three further arms isolate the key ingredient. The first is soft fusion of quantized raw signal-to-noise, summed arithmetically at the same bits. It keeps everything except calibration, so it tests the inept-agent hypothesis. The second is our method with a stale or white-noise $\eta$, which shows how much rolling calibration matters. The third is a public-coin hashed aggregation arm. Stations use their shared GPS-disciplined clocks as common randomness to agree on random groupings of cells, and each sends one parity-style bit per group, in the spirit of Inference Under Information Constraints II: Communication Constraints and Shared Randomness and Optimal high-dimensional and nonparametric distributed testing under communication constraints. Finally, a centralized oracle that sums unquantized, uncensored calibrated statistics gives the upper bound.
Our method beats the strongest baseline if it attains at least 90% detection completeness at a lower injected per-station signal-to-noise, at matched bits and false-trigger rate, across bit budgets. We will also report the 50% completeness point, where a gap of about one unit of per-station signal-to-noise would be astrophysically meaningful. No reported result was found on Lacuna for multi-station FRB coincidence completeness under a bit budget, so all baseline numbers here will be produced by us. The only external anchors are single-station ones. The Nanshan search reported a signal-to-noise 6.0 candidate at dispersion measure 1522.8 pc/cm³, with a 1% false-alarm probability over 55 hours after coloured-noise correction. GReX reports a detection threshold of about 50 kJy for a 1 ms burst.
Why not just…?
Why not just correct each station for coloured noise and keep one-bit coincidence? This is the real threat. The simple-binary result says threshold quantizers are near-optimal, so calibrated one-bit reporting could capture most of the gain. Our strongest baseline is exactly this, tuned to the same budget. If our multi-level fusion does not beat it by a meaningful margin, the correct conclusion is that calibration is the whole story. The paper would then be about calibrated one-bit coincidence, and we would report it that way. We expect a gap mainly for bursts that are marginal at every station. The composite search over many cells and the heterogeneous station sensitivities also fall outside the simple-binary theorem's assumptions.
Why not just stream compressed dynamic spectra to a central site? CHIME alone moves 142 Gb/s internally. The rationale for detection-oriented messages, covered in the overview of Limits of General Distributed Hypothesis Testing, is that codecs built to preserve waveform fidelity spend bits on detail that does not help separate noise from signal. We include one lossy-compression arm at matched bits, applying a standard transform codec to dedispersed time series, to confirm this empirically rather than assume it.
Why not add interaction, letting the coordinator ask stations for more detail on promising cells? The factor-of-4 bound in The Sample Complexity of Distributed Simple Binary Hypothesis Testing under Information Constraints suggests limited returns. In contrast, Distributed Hypothesis Testing: Cooperation and Concurrent Detection shows large gains from even one bit when detectors hold different side information. We test a single follow-up round within the latency budget: the coordinator requests finer quantization for its top cells. We keep this round only if it shifts the completeness curve beyond what spending the same bits non-interactively achieves.
Why not use shared randomness, since theory says public coins help testing? The public-coin gains are proven for flattening broad, high-dimensional distributions. An FRB is extremely sparse, occupying a handful of cells among millions. Random grouping then dilutes the signal, whereas censored tail reporting concentrates bits where the signal is. The hashed arm tests this directly. A win for it would itself be a notable finding.
Proposed Experiments
The testbed must use real coloured noise with truly independent station realizations. We build "virtual networks" from non-overlapping segments of the 55-hour Nanshan M82 data, if the authors release it, and from GReX station recordings. Each segment plays one station. We inject the same burst into all stations with station-specific gain, beam attenuation using the roughly 65° GReX beam, and a geometric delay. Bursts are drawn from frbpoppy populations (Synthesising the repeating FRB population using frbpoppy) so that widths and dispersion measures are realistic. We also include the instrumental smearing effects of Interpreting the distributions of FRB observables, since these create the narrow high-dispersion-measure bursts that are most likely to be marginal. A synthetic twin with controllable $\eta$ lets us vary the noise colour systematically.
The main measurement is completeness against per-station injected signal-to-noise at a fixed network false-trigger rate of one spurious dump per day. We sweep bit budgets per station from about 1 to 100 kb/s and network sizes of 3 to 8 stations. We also require end-to-end decision latency within the 20-second buffer.
Three checks decide the proposal. The first is the gap to the calibrated one-bit baseline at equal bits. The second is the false-trigger rate when one station's $\eta$ is deliberately corrupted. Our method should stay near nominal, while raw signal-to-noise fusion should blow up, which would be the inept-agent effect observed on radio data. The third is how close the method comes to the unconstrained oracle as the bit budget grows. The simple-binary theory predicts this gap closes roughly once bits per report exceed the log of the number of search cells. That prediction is directly checkable, and a clear failure would tell us the composite, coloured setting behaves differently from the theory.
References
- The CHIME Fast Radio Burst Project: System Overview
- Fast radio burst detection in the presence of coloured noise
- The Sample Complexity of Distributed Simple Binary Hypothesis Testing under Information Constraints
- Single-pulse Detection Algorithms for Real-time Fast Radio Burst Searches Using GPUs
- On the Arithmetic and Geometric Fusion of Beliefs for Distributed Inference
- GReX: An Instrument Overview and New Upper Limits on the Galactic FRB Population
- Inference under Information Constraints II: Communication Constraints and Shared Randomness
- Optimal high-dimensional and nonparametric distributed testing under communication constraints
- Distributed Hypothesis Testing: Cooperation and Concurrent Detection
- Synthesising the repeating FRB population using frbpoppy
- Interpreting the distributions of FRB observables