Fast Radio Burst 121102 Pulse Detection and Periodicity: A Machine Learning Approach

Yunfan Gerry Zhang; Vishal Gajjar; Griffin Foster; Andrew Siemion; James Cordes; Casey Law; Yu Wang

Source-linked AI summary

Fast Radio Burst 121102 Pulse Detection and Periodicity: A Machine Learning Approach

Yunfan Gerry Zhang, Vishal Gajjar, Griffin Foster, Andrew Siemion, James Cordes, Casey Law, Yu Wang

arXiv:1809.03043v1Sep 9, 2018astro-ph.HEastro-ph.IM

TL;DR

This paper addresses limited sensitivity and statistical power in detecting fast radio transients and periodicity from sparse FRB arrival times. It applies neural-network detection with dedispersion verification and a Rayleigh-based periodicity test, finding 72 new pulses and excluding arrival-time periods above 10 ms with 99% confidence.

Problem
Traditional dedispersion searches face noise and radio-frequency-interference false positives, while prior FRB periodicity studies lacked quantified statistical significance for null detections.
Method
The paper applies convolutional neural networks to raw spectrogram data, verifies candidates with standard dedispersion, and introduces a Rayleigh-based periodicity test incorporating time-stamp uncertainty.
Results
72 new FRB 121102 pulses were detected, bringing the observation total to 93, while arrival-time periodicity above 10 ms was excluded with 99% confidence.
Takeaways & Limitations
The expanded pulse sample enables analysis of fluence, pulse rate, and frequency structure, while the periodicity result constrains models predicting constant arrival-time periods.
Takeaways & Limitations
Unknown undetected pulses may be clustered because of observational bias, weak or out-of-band emission, or no emission, and intrinsic periodicity remains model-dependent.

Abstract

from arXiv · show

We report the detection of 72 new pulses from the repeating fast radio burst FRB 121102 in Breakthrough Listen C-band (4-8 GHz) observations at the Green Bank Telescope. The new pulses were found with a convolutional neural network in data taken on August 26, 2017, where 21 bursts have been previously detected. Our technique combines neural network detection with dedispersion verification. For the current application we demonstrate its advantage over a traditional brute-force dedis- persion algorithm in terms of higher sensitivity, lower false positive rates, and faster computational speed. Together with the 21 previously reported pulses, this observa- tion marks the highest number of FRB 121102 pulses from a single observation, total- ing 93 pulses in five hours, including 45 pulses within the first 30 minutes. The number of data points reveal trends in pulse fluence, pulse detection rate, and pulse frequency structure. We introduce a new periodicity search technique, based on the Rayleigh test, to analyze the time of arrivals, with which we exclude with 99% confidence pe- riodicity in time of arrivals with periods larger than 5.1 times the model-dependent time-stamp uncertainty. In particular, we rule out constant periods >10 ms in the barycentric arrival times, though intrinsic periodicity in the time of emission remains plausible.

1. INTRODUCTION

This paper applies deep learning to direct detection of FRB 121102 pulses and introduces a statistically sensitive periodicity search for their arrival times.

1. INTRODUCTION: The study applies convolutional neural networks to direct detection of fast radio transient signals in raw spectrogram data.It re-analyzes Breakthrough Listen C-band observations of FRB 121102.
1. INTRODUCTION: The paper introduces a new periodicity-search technique for repeating radio transient detections.The method is designed to remain sensitive with a limited number of pulses.
1. INTRODUCTION: The analysis identifies 72 new FRB 121102 pulses using deep learning.This is presented as the first successful application of deep learning to direct detection of fast radio transient signals.
1. INTRODUCTION: The resulting pulse sample enables investigation of trends in pulse structure from the largest single-observation set reported for FRB 121102.The paper examines pulse fluence, arrival times, and frequency structure.
1. INTRODUCTION: Arrival-time periods ≳10 ms are excluded with 99% confidence, while the paper reports a statistical limit on FRB 121102 pulse aperiodicity.This conclusion concerns detected arrival times rather than necessarily intrinsic emission periodicity.

2. OBSERVATION

The observations used six hours of FRB 121102 monitoring with the Green Bank Telescope’s 4–8 GHz C-band receiver, producing ten 30-minute scans and time-frequency filterbank data.

2. OBSERVATION: The observation session began on August 26, 2017 and used the Green Bank Telescope’s 4–8 GHz C-band receiver.The session was divided into ten 30-minute scans of FRB 121102.
2. OBSERVATION: The Breakthrough Listen digital backend generated both time-domain voltage data and integrated spectral data.The analysis used spectral-temporal filterbank data.
2. OBSERVATION: The filterbank data had 350 µsec time resolution and 366 kHz frequency resolution.These resolutions characterize the data used for the pulse analysis.

3. DETECTION

The detection pipeline trains a convolutional neural network on simulated FRB pulses and combines its candidate selection with dedispersion-based verification. The model achieves strong test performance, reduces the computational burden relative to brute-force dedispersion, and uses morphology-focused verification to distinguish dispersed pulses from RFI.

3.2. Data Preparation: The network is trained with balanced simulated-pulse and noise/RFI frames because scarce real pulses cannot cover the range of dispersion, amplitude, width, scintillation, and location variations.The dataset contains around 400000 images, split between simulated pulses and non-pulse frames, with four and a half hours for training and half an hour for testing.
3.4. Model Evaluation: 88% recall and 98% precision were achieved on the independent test set, with only minimal over-fitting relative to training.Recall measures sensitivity, while precision measures robustness to false alarms.
3.4. Model Evaluation: The CNN detected verifiable pulses above 20 Jyµs with over 95% recall, while thinner and shallower models showed noticeable recall reductions.The selected architecture was considered sufficient for this analysis, although a larger model could provide higher sensitivity.
3.4. Model Evaluation: A roughly 2% false-detection rate reflects incomplete representation of underrepresented RFI types in the training set and motivates manual rejection before verification.The authors describe developing more robust training methods for such RFI but use manual rejection in this work.
3.5. Computational Performance: Inference processed data about 20 times faster than a brute-force dedispersion search using 1200 DM trials, although hardware differences complicate the comparison.The brute-force search processed 3 seconds of observation per second in the stated setup.
3.6. Pulse Verification: Verification prioritizes quadratic dispersion morphology over signal strength by searching manually selected sub-bands across dispersion measure and arrival time.Sub-bands of at least 1.5 GHz are chosen to make dispersion-measure fitting more distinct rather than maximizing sub-band S/N.

4. PULSE PROPERTIES

The paper defines how pulse subbands, fluence, flux density, and arrival times are measured from dedispersed observations. Fluence is estimated using the signal-to-noise dispersion measure, while arrival times use the structural dispersion measure and energy peaks.

The reported pulse properties include barycentric arrival times, TOAs, DMS/N, fluence, flux density, subbands, widths, and peak frequencies.Fluence and flux density are calibrated with the radiometer equation, while widths and subbands support fluence measurements.
Peak frequency is defined as the channel with the greatest power, with each channel spanning 11 MHz.Intrinsic widths are not reported because noise prevents reliable measurement for many weaker pulses.
Fluence is measured by dedispersing to DMS/N and searching frequency ranges for the widest subband around each pulse’s peak frequency.Using DMS/N improves fluence determination when full-band signal-to-noise is low.
Arrival times are obtained from the energy peak after dedispersion to DMstruct = 565 pc cm−3 using the selected subbands.Multiple sub-pulses make the strongest sub-pulse determine the recorded arrival time.

5.1. Pulse Structure

The detected pulses exhibit complex time–frequency structure, including closely spaced multiplets whose interpretation limits the precision of arrival-time and periodicity analyses. Some pulses contain substructures extending to approximately 2 ms.

5.1. Pulse Structure: Pulse profiles show complex structures in both frequency and time, with reported substructures extending to widths of approximately 2 ms.Frequency-dependent tilts in dynamic spectra reflect the pulses’ frequency ranges.
5.1. Pulse Structure: The pulse catalogue records barycentric times, TOAs, dispersion measures, fluences, flux densities, subbands, widths, and peak frequencies for the reported events.Table 2 reports uncertainty estimates for fluence and flux density and derives both by dedispersing to DMS/N.
5.1. Pulse Structure: Several pulse groups cluster within 10 ms, but their signals do not share a unifying dispersion measure and may represent separate pulses or substructures.For pulses 81 and 82, the possible separation is 2.56 ms despite overlapping frequency channels.
5.1. Pulse Structure: If the closely spaced signals are sub-pulses, their roughly 10 ms spacing sets a lower bound on arrival-time uncertainty; if independent, it constrains possible periodicity.The authors leave the physical interpretation of these multiplets unresolved.

5.2. Parameter Statistics

Pulse parameters show clustered peak frequencies, time-dependent detection and fluence patterns, and substantial variation in pulse strength. Several interpretations remain uncertain because detection bias and propagation assumptions cannot be excluded.

5.2. Parameter Statistics: DMS/N values scatter around an average of 575 pc cm−3, so the analysis retains DMstruct from earlier work because low-energy pulses have large scatter.Figure 4 compares arrival time, DMS/N, fluence, and peak frequency pairwise.
5.2.1. Peak Frequency: Five Gaussian-mixture components place peak-frequency clusters at 7.54, 7.05, 6.23, 5.62, and 4.91 GHz during the first hour.A pulsar observed in the same session lacks similar modulation, while the clusters correspond to sub-pulse peak frequencies.
5.2.1. Peak Frequency: The first three pulses occur eight seconds apart and peak near 7 GHz, suggesting but not establishing a non-stochastic time dependence in peak frequency.More data are needed to constrain any variability timescale.
5.2. Parameter Statistics: Pulse detections are denser near the observation’s beginning across fluences, indicating that high pulse rate coincides with higher detected pulse energy.The cause is left open to alternative explanations involving the intrinsic fluence distribution or weaker after-bursts.
5.2. Parameter Statistics: The low-fluence end of the histogram falls off, but the paper cannot determine whether this reflects intrinsic emission or detection bias.If low-energy pulses dominate intrinsically, many such pulses may have gone undetected during the later four hours.
5.2. Parameter Statistics: Pulses with similar frequency structure span more than an order of magnitude in flux density, although the interpretation assumes unconfirmed propagation behavior.Under that assumption, the variation may originate at the source.

5.3. Pulse Rate

The five-hour pulse interval distribution is more consistent with stationary Poisson statistics than earlier reports, although the detection rate is non-stationary and the first 30 minutes retain a long-tail excess. Missing clustered pulses may explain the residual deviation.

5.3. Pulse Rate: On five-hour timescales, observed pulse intervals are more consistent with Poisson statistics than previously reported, despite a non-stationary detection rate.Nearly half of detections occur during the first 30 minutes, and the comparison uses a Poisson expectation with r = 0.05 s−1.
5.3. Pulse Rate: The interval analysis compares all pulses, the first 30 minutes, and the 15 highest-fluence pulses using 20-second bins below 300 seconds.The distributions are plotted against Poissonian expectations.
5.3. Pulse Rate: The first 30 minutes still show a skew toward longer intervals relative to Poissonian expectations.The deviation persists despite this observation containing the largest single-observation sample of FRB 121102 pulses at the time.
5.3. Pulse Rate: The long tail may reflect undetected pulses clustered by non-stochastic changes in pulse energy or frequency structure, or periods without emission.These possibilities imply that missing detections are not necessarily randomly distributed.

5.4. Periodicity

The paper searches for periodicity in FRB 121102 pulse arrival times despite timestamp uncertainty and potentially missing pulses. It folds arrival times over trial periods and tests whether the resulting phases are unimodally distributed.

5.4. Periodicity: The periodicity search must accommodate timestamp uncertainty and an unknown number of missing pulses, both of which affect the inferred candidate period.Timestamp uncertainty is modeled as an effective σt, while missed pulses can arise from undetectable energy or frequencies outside the observing band.
5.4. Periodicity: Fourier, autocorrelation, and separation-histogram methods can lose sensitivity when many pulses are missing, motivating the adopted folded-phase procedure.The method follows an approach designed for incomplete observations.
5.4.1. Hypotheses of Periodicity: The analysis constructs periodic and aperiodic hypotheses for folded pulse phases, with periodicity represented by a unimodal phase distribution.The null hypothesis states that arrival times are not periodic with the trial period and phases are not unimodally distributed.
5.4.1. Hypotheses of Periodicity: Trial periods are sampled densely enough that period-resolution errors remain below the assumed arrival-time uncertainty over the observation.The search covers periods from the lower bound through approximately one second, although nearby trial periods are correlated.

5.5. Rayleigh’s Test

The paper uses a Rayleigh-test statistic based on the mean resultant radius to evaluate periodicity in folded arrival phases. Simulations account for multiple trial periods and show that periods above 5.1σt can be excluded with 99% confidence, while the constraint depends on timestamp uncertainty and physical perturbations.

5.5. Rayleigh’s Test: The mean resultant radius ¯R measures phase alignment from 0 to 1, reaching 1 only when all folded pulse phases align.Under uniformly distributed phases, ¯R follows a Rayleigh distribution.
5.5.1. Top Scoring Periods: Simulations of 1000 random 93-pulse arrival sets provide empirical p-values and account for multiple testing across trial periods.The analysis samples random trials from the empirical arrival-time distribution and compares their cumulative distributions with the observed statistic.
5.5.1. Top Scoring Periods: A top period rejecting H0 at 97.5% has only 79% confidence after correcting for multiple testing, so it is not evidence of a significant periodicity detection.The correction considers the probability of obtaining at least one apparently strong result among multiple tested periods.
5.5.2. Period Exclusion: Periods larger than 5.1σt are excluded with 99% confidence, based on the observed upper bound ¯R = 0.33 and simulated periodic-phase distributions.With σt = 3 ms, this excludes periods larger than 15 ms in barycentric arrival times.
5.5.3. Model Dependence: The exclusion applies to apparent arrival-time periodicity and remains model-dependent because propagation or source acceleration can smear intrinsic emission periodicity.Constraining specific physical models is beyond the paper’s scope; with 45 pulses from the first 30 minutes, periods above 6σt are excluded in the acceleration example.

6. CONCLUSION

The paper demonstrates neural-network detection of fast radio bursts and uses the resulting abundance of detections to study pulse properties and arrival-time periodicity. It also limits the scope of its claims by noting that a general real-time survey pipeline and comprehensive comparisons remain outside this work.

6. CONCLUSION: 72 new pulses demonstrate a neural network’s direct application to fast radio transient detection, with potential sensitivity and computational-speed advantages over dedispersion pipelines.The paper describes this as the first application of a neural network for direct detection in spectral-temporal data.
6. CONCLUSION: The abundant detections support analyses of pulse fluence, pulse rate, frequency structure, and multiplets spanning 10 ms to 20 ms.The reported multiplets show non-monotonic variations in frequency structure.
6. CONCLUSION: The periodicity method excludes with 99% confidence all arrival-time periods greater than 10 ms when timestamp uncertainty is 2 ms.The method is designed for cases where most pulses may be unobserved and quantifies the significance of a null detection.
6. CONCLUSION: A general real-time survey pipeline and comparison with existing pipelines remain outside the scope of this work.The authors note that applying deep learning to general real-time detection involves additional caveats.
6. CONCLUSION: The abundance of pulses and their spectral and energetic variation suggests previous surveys may have underestimated FRB abundance and the fraction of repeaters.The authors attribute the detected abundance to high-sensitivity detection combined with wide bandwidth observation.

A. EXTENDED FIGURES

The extended figures display pulse detections, dedispersion verification, and fluence measurements using time, frequency, dispersion measure, and flux-density axes. Together, they provide visual checks of pulse arrival structure, dispersion alignment, and fluence integration windows.

A. EXTENDED FIGURES: Figure 7 plots the first-30-minute pulses with time since observation start on the horizontal axis and frequency in GHz on the vertical axis.Asterisks identify previously reported pulses, while panel labels correspond to Table 2.
A. EXTENDED FIGURES: Figure 8 continues the pulse plots and emphasizes that visible pulse morphology depends on frequency and time resolution.The figure notes that pulse morphologies vary and visibility is subjective to the displayed resolution.
A. EXTENDED FIGURES: Figure 9 uses time of arrival and dispersion measure axes, with light central colors indicating dispersion measures and arrival times near expectation.Colors represent flux density after incoherent dedispersion and are scaled separately for each panel.
A. EXTENDED FIGURES: Figure 10 plots flux density against time for each pulse and marks the time ranges used to compute fluence with red vertical lines.The pulses are dedispersed at their signal-to-noise-maximizing dispersion measures and shown in selected sub-bands.