IceCube Alert Tracks: Real-time Neutrino Alerts

• IceCube Alert Tracks are real-time neutrino alerts based on muon tracks that provide sub-degree directional precision for astrophysical source identification.
• They exploit charged-current νμ interactions to produce long-ranging muons, enabling advanced reconstruction techniques and refined error contours.
• The system underpins multi-messenger astronomy by triggering rapid follow-up across electromagnetic and gravitational-wave observatories.

IceCube Alert Tracks are the track-like realtime neutrino alerts issued by the IceCube Neutrino Observatory for neutrino candidates with good directional reconstruction and moderate-to-high probability of astrophysical origin. They arise primarily from charged-current $\nu_\mu$ interactions that produce long-ranging muons, whose elongated Cherenkov light patterns provide markedly better pointing than cascade-like events and therefore enable rapid, targeted multi-messenger follow-up. Since the public alert program began in 2016, these alerts have become the principal IceCube channel for source-directed neutrino astronomy, with the 2017 event IC-170922A and its association with the blazar TXS 0506+056 serving as the canonical demonstration of their scientific use (Williams, 2019, Blaufuss et al., 2019).

1. Detector context and physical basis

IceCube is a cubic-kilometer neutrino observatory at the geographic South Pole, instrumenting glacial ice between depths of 1450 m and 2450 m with 86 strings carrying 60 Digital Optical Modules each. The detector operates continuously with $>99\%$ uptime, probes neutrino energies from GeV to PeV, and is supplemented by DeepCore, which lowers the energy threshold to about 5 GeV, and by IceTop on the surface, which provides air-shower measurements (Williams, 2019).

Within this instrument, event topology is central to alert-track physics. “Tracks” arise primarily from $\nu_\mu$ charged-current interactions that produce long-ranging muons, while “cascades” are quasi-spherical light patterns from neutral-current events of all flavors and charged-current interactions of $\nu_e$ and most $\nu_\tau$. Because muon tracks extend over tens to hundreds of meters and carry rich timing information across many DOMs, they afford superior angular resolution compared to cascades. In the public alert context, this distinction is decisive: through-going and starting muon tracks can reach sub-degree directional uncertainty, whereas cascades typically have angular uncertainties of $10\text{–}15^\circ$ or more and are therefore not the primary source-pointed alert channel (Williams, 2019, Blaufuss et al., 2019).

Two track topologies underpin the alert program. Through-going tracks originate outside the instrumented volume and, in the Northern Hemisphere, exploit the Earth as a filter against downgoing atmospheric muons, yielding a relatively clean $\nu_\mu$ sample. Starting tracks begin within the detector volume and use containment to veto entering atmospheric muons, extending sensitivity over the whole sky and especially into the southern sky at lower energies than traditional through-going samples (Williams, 2019).

2. Realtime alert architecture and public dissemination

The realtime framework couples low-latency South Pole filtering to more sophisticated Northern Hemisphere reconstructions and public dissemination channels. In the original 2016 deployment, the DAQ recorded triggered DOM signals and passed events to roughly 400 online calibration and filtering clients, which calibrated waveforms, extracted photon arrival times, performed fast reconstructions, and identified candidate alerts under a processing budget of about 30 s per event. Candidate alerts were transmitted through Iridium RUDICS, with a measured median end-to-end message latency of 33 s from event trigger to receipt in Madison; later descriptions summarize public notice latency as typically about one minute from detection (Collaboration et al., 2016, Blaufuss et al., 2019).

The public alert stream started in April 2016. The initial machine-readable alert is distributed through AMON to NASA’s Gamma-ray Coordinates Network, and a refined human-readable GCN Circular usually follows a few hours later after more computationally expensive reconstructions. The Circular can update the best-fit direction and energy and can retract an alert if subsequent checks identify mis-reconstruction, mis-identification, or coincident cosmic-ray muons (Williams, 2019, Blaufuss et al., 2019).

A 2019 overhaul expanded the public system into Gold and Bronze channels. Gold alerts correspond to candidates at least 50% likely to be astrophysical, while Bronze alerts correspond to candidates with 30–50% astrophysical probability. In the 2019 expansion, Gold alerts were expected at a rate of about 10 per year and Bronze alerts at about 20 per year; after the upgrade described for the operational stream, Gold alerts were issued at an average rate of about one per month and Bronze alerts at about 1.3 per month (Blaufuss et al., 2019, Williams, 2019).

Each alert notice includes the event time and date, reconstructed right ascension and declination, angular radii of the 50% and 90% containment regions, and the signal probability; the alert schema also includes a false alarm rate field defined as the yearly rate of alerts with equal or higher signalness. In GCN Notices, the reported angular error is set to be at least $0.2^\circ$ to account for systematics (Blaufuss et al., 2019).

Behind the public notices, the internal infrastructure has been modernized around a message-based workflow. The redesigned “Skymap Scanner” performs hierarchical HEALPix sky scans, while SkyDriver provides a REST interface to parallelized reconstruction algorithms and SkyMist integrates the realtime queue, database bookkeeping, task scheduling, result collection, auto-drafting of GCN circulars, and standardized counterpart queries. RabbitMQ, ZeroMQ, containerized clients, CVMFS-based staging, and persistent provenance records are used to improve portability, reproducibility, and FAIR-aligned long-term cataloging (Lincetto et al., 2023).

3. Alert selections and stream ecology

The public alert system is not a single selection but an ecosystem of track-focused streams with different background regimes, sky coverages, and energy scales. In the 2019 public framework, three track-oriented selections dominated: GFU through-going tracks, HESE tracks, and EHE tracks. GFU and EHE provide most alerts, while a smaller fraction are starting tracks from the HESE channel; if an event passes multiple selections, only one alert is sent, with hierarchy $\mathrm{GFU} > \mathrm{EHE} > \mathrm{HESE}$ (Blaufuss et al., 2019).

Stream	Event class	Role in alert program
GFU	Through-going tracks	Main track-alert base; Gold and Bronze
HESE tracks	Starting tracks with outgoing muon	Full-sky starting events with track-like pointing
EHE	Through-going tracks	500 TeV–10 PeV; Gold only
ESTReS	Southern-sky starting tracks	5–100 TeV; average purity about 50%

GFU alerts are derived from through-going muon tracks similar to time-integrated point-source samples and use boosted decision trees to select well-reconstructed, astrophysical-like events. HESE track alerts apply a veto-based contained-vertex selection and additionally require an outgoing muon track, measured track length of at least 200 m, and a track-like hypothesis favored by the reconstruction likelihood. EHE alerts target the highest-energy tracks, requiring $\mathrm{NPE} \ge 4000$, a good fit-quality parameter, and a declination-dependent NPE threshold; they are exclusively Gold (Blaufuss et al., 2019).

Southern-sky starting tracks were developed because atmospheric muons and atmospheric neutrinos from downgoing directions severely limit through-going selections. A 2019 starting-track program introduced an event-specific veto built around the reconstructed track. For each DOM, the expected photoelectron mean is modeled as $>99\%$0, and the entering-muon hypothesis is rejected through the “missing-light” probability

$>99\%$1

The resulting offline sample achieved less than about one atmospheric muon per year in the final southern sample, reached high astrophysical neutrino purity above about 10 TeV at declinations less than $>99\%$2, and was identified as suitable for a realtime southern-sky alert stream (Mancina et al., 2019).

The later Enhanced Starting Track Realtime Stream, ESTReS, is the operational real-time extension of the offline ESTES sample. It is explicitly designed to deliver track-like alerts from the southern sky in the 5–100 TeV range, a region described as not well covered by GFU, HESE, or EHE. ESTReS applies fast South Pole cuts—at least 450 photoelectrons after the reconstructed vertex time, $>99\%$3, incoming-muon probability $>99\%$4, and in-detector track length $>99\%$5 m—then transfers about 100 events per day by satellite to Madison, where the full ESTES chain applies a dynamic Starting Track Veto, a BDT classifier, Millipede-based directional reconstruction, and a Random Forest energy proxy. The expected outcome is 10.3 alerts per year with average astrophysical purity of about 50%, comprising 5.1 astrophysical events per year and 5.2 background events per year, with no overlap seen in 10.3 years of archival data with GFU, HESE, or EHE alerts (Osborn et al., 2023).

4. Reconstruction, localization, and statistical characterization

Track alerts are reconstructed by fitting the energy, position, time, and direction from the pattern of light deposition recorded by the DOMs. In the realtime system, directional localization is obtained from hierarchical sky scans in equatorial coordinates using HEALPix. A coarse pass at $>99\%$6 is followed by refinement at $>99\%$7, then a final pass at $>99\%$8 or 1024 around the best-fit region. At each pixel the direction is fixed and the likelihood is maximized over the remaining parameters; containment contours are then derived directly from the resulting $>99\%$9 map rather than from a fixed Gaussian model (Lincetto et al., 2023).

The generic reconstruction objective can be written as

$\nu_\mu$0

with the fit minimizing $\nu_\mu$1. In the 2024 realtime update, IceCube adopted a hybrid Revision 1 follow-up strategy that alternates between two reconstructions according to deposited energy and topology: SplineMPE with a likelihood scan for low-energy-deposition non-HESE events, and Millipede Wilks for high-energy-deposition events and all HESE alerts. The confidence regions are constructed from

$\nu_\mu$2

with

$\nu_\mu$3

and the two-dimensional quantiles $\nu_\mu$4 and $\nu_\mu$5. This change replaced the historical use of fixed Millipede thresholds $\nu_\mu$6 and 64.2 for 50% and 90% containment, respectively, and reduced the 50% and 90% contour areas by factors of 5 and 4 while improving statistical coverage (Sommani et al., 9 Jul 2025).

Signal characterization is based on event-by-event astrophysical probability. In the public Gold/Bronze system this quantity is computed differentially in declination and a selection-specific energy proxy; ESTReS provides the explicit form

$\nu_\mu$7

with a sigmoid interpolation in $\nu_\mu$8 used to assign per-event metadata. ESTReS reports a median space-angle resolution of about $\nu_\mu$9, a 90% containment angular error contour summarized by a median bounding rectangle of approximately $\nu_e$0 in declination and $\nu_e$1 in right ascension, and an energy resolution of approximately 25–30% over 1 TeV–10 PeV. The stream also notes under-coverage in Monte Carlo studies and ongoing refinement of the error-contour generation (Osborn et al., 2023).

These probabilistic metrics should not be conflated with certainty. Gold and Bronze encode average astrophysical probabilities under the assumed diffuse-spectrum model used to optimize the realtime system, and the false alarm rate field measures how often background alerts with equal or higher signalness are expected per year (Blaufuss et al., 2019).

5. Multi-messenger use and empirical record

The paradigmatic alert-track case is IC-170922A. On 22 September 2017, IceCube observed an upgoing through-going muon track with a most likely energy of 290 TeV and a 55.6% probability of being astrophysical. After the alert, Fermi-LAT detected enhanced gamma-ray emission from the blazar TXS 0506+056 inside the neutrino localization region, and MAGIC observed very-high-energy gamma rays up to 400 GeV within 12 hours of follow-up between 24 September and 4 October. The correlation between the high-energy neutrino and the gamma-ray flare was preferred over chance coincidence at the $\nu_e$2 level. Subsequent spectroscopy measured the redshift of TXS 0506+056 to be $\nu_e$3, and archival IceCube data revealed an excess from the same direction in a Gaussian time window centered on 13 December 2014 with duration $\nu_e$4 days and $\nu_e$5 events above atmospheric expectations, inconsistent with background-only at $\nu_e$6 (Williams, 2019).

The broader scientific role of alert tracks is to trigger rapid follow-up across the electromagnetic spectrum and other messengers. Alerts are disseminated to gamma-ray facilities such as Fermi-LAT, MAGIC, and HAWC, as well as optical, radio, and gravitational-wave partners. Beyond public notices, IceCube contributes high-quality single tracks to AMON at a rate of about 650 per day, while HAWC contributes daily source excesses at about 800 per day; in an initial two-year coincident analysis one coincidence was found with a false alarm rate of 1.1 per year, and the plan stated for public coincidences was $\nu_e$7 per year. During LVC’s O3 run, IceCube also performed maximum-likelihood and Bayesian searches for neutrino counterparts to gravitational-wave alerts using the GW skymap as a spatial prior and issued a GCN circular typically within about an hour of the LVC alert (Williams, 2019).

Population studies around alert directions have been notably conservative. A dedicated archival search over 10.5 years of through-going muon data around 81 alert directions found no significant steady source, with the most significant alert direction yielding a p-value of 0.83 and the stacked analysis across all alerts yielding $\nu_e$8. A later 11-year study of 122 alert and high-energy track directions likewise found no evidence for significant continuous emission and identified only the transient emission associated with TXS 0506+056 at local significance $\nu_e$9; after correcting for 122 test positions, the global p-value was 0.156. That analysis placed a 90% confidence upper limit on the total continuous flux from all 122 positions at 100 TeV of $\nu_\tau$0 for an $\nu_\tau$1 spectrum, corresponding to 4.5% of IceCube’s astrophysical diffuse flux (Karl, 2019, Abbasi et al., 2023).

6. Catalogs, uncertainties, and future development

The alert-track program has been progressively formalized in public catalogs. IceCat-1, released in 2023, compiled 275 track-like neutrino events between 13 May 2011 and 31 December 2020 from the GFU Gold, GFU Bronze, HESE Gold, HESE Bronze, and EHE Gold channels. For each event it provided reconstructed energy, direction, false alarm rate, probability of astrophysical origin, and likelihood contours describing the spatial uncertainty, together with FITS log-likelihood sky maps and topology metadata (Abbasi et al., 2023).

IceCat-2, presented in preliminary form in 2025, extends the archive through 2 January 2025 and reprocesses the catalog uniformly with the September 2024 reconstruction upgrade. The preliminary catalog contains 365 muon-neutrino track alerts spanning May 2011 through January 2025, with an average alert rate of $\nu_\tau$2. Relative to IceCat-1, the reprocessing reduces the typical 50% and 90% angular uncertainties by factors of approximately 5 and 4, yields median 50% and 90% containment areas of about 0.4 and 1.3 square degrees, and tightens the area distributions by factors between 7 and 9 in standard deviation. Among the 340 reprocessed IceCat-1 events retained after the IceTop cosmic-ray veto, the new-to-old directional differences have median $\nu_\tau$3 and 90% quantile $\nu_\tau$4. The catalog also shows that the tightened contours can materially revise source-association statements: TXS 0506+056 remains within the revised 90% region of IC-170922A but not within the 50% contour, two alerts near NGC 7469 remain compatible with that source, and several tidal-disruption-event associations move outside the refined containment regions (Zegarelli et al., 8 Jul 2025).

These catalogs also expose the main caveats of alert-track interpretation. First, real-time parameters and catalog values are not identical: reprocessed coordinates and uncertainties can differ substantially from the values circulated at the time of the alert. Second, Gold and Bronze probabilities depend on the assumed astrophysical spectrum. IceCat-1 and the real-time channel definitions used an $\nu_\tau$5 assumption for signalness optimization, while IceCat-2 explicitly notes that future releases plan to revisit these probabilities using a softer spectral index measured for muon tracks (Abbasi et al., 2023, Zegarelli et al., 8 Jul 2025).

Operationally, the dominant limitations remain atmospheric muons and atmospheric neutrinos, especially in the southern sky where strong vetoes suppress downgoing muons at the cost of low-energy efficiency. Systematic uncertainties in ice optical properties and detector calibration affect reconstruction quality and localization. Ongoing mitigation includes the IceTop cosmic-ray veto, new lower-energy down-going starting-track streams, the possible future addition of cascades to the public stream once directional reconstructions are sufficiently reliable, and the IceCube Upgrade, which will deploy seven new strings with enhanced sensors and an extensive calibration program. The Upgrade is expected to improve event reconstructions and low-energy performance and to deliver better than 10% sensitivity to $\nu_\tau$6 appearance normalization in one year, with broader benefits for alert-quality tracks through improved timing and light modeling (Williams, 2019).

In this mature form, IceCube Alert Tracks are best understood as a calibrated, low-latency track-based localization system rather than a mere list of isolated neutrino candidates. Their scientific value lies in the combination of topology-driven pointing precision, quantified astrophysical probability, iterative offline refinement, and a catalog structure that supports both immediate follow-up and retrospective population studies.