PaperlayerView exact revision on arXiv

Source-linked AI summary

The Towers Were Standing: A Cause Decomposition of Cellular Outages During Hurricane Helene

Oluseyi Olukola, Oare Danielle Addeh, Esther Abiodun Konan, Nick Rahimi

arXiv:2609.10944v1cs.NIcs.CY

TL;DR

To explain Helene’s record outage, the paper separates site damage, transport loss, and commercial-power loss in FCC reports. It reconciles independent county and state extractions, finding that sites were mostly intact while dominant causes varied by terrain.

  • Problem

    Helene’s record cell-site outage raises whether sites were damaged or outages reflected loss of transport or commercial power.

    Source: §1 · pp. 2–3

  • Method

    The authors analyze FCC DIRS cause attributions and check extraction by matching county sums to independently extracted state totals.

    Source: §1 · p. 3, §3.2 · p. 8

  • Results

    1.1% of attributed cell-site-days were site damage across six states, while in mountainous western North Carolina transport accounted for 52.2% versus 47.3% for power.

    Source: §7 · p. 30

  • Takeaways & Limitations

    Backup power addresses only outages caused by commercial-power loss, while backhaul loss is a distinct engineering problem with different mitigations and restoration timelines.

    Source: §1 · p. 3

    Because cause counts leave a residual, North Carolina’s transport-over-power ordering would reverse if more than 66% of that residual were power, which the authors cannot exclude.

    Source: §6.1 · p. 27

Abstract

from arXiv · show

Hurricane Helene produced the largest absolute cell-site outage in the public FCC record, peaking at 4562 sites. The conventional model is physical: towers destroyed. Helene did destroy over 1700 miles of fibre, but almost none of it was cell sites. We present the first cause-decomposed study of the FCC's Disaster Information Reporting System, reconstructing 80 state-days and 580 county-days from 24 daily filings by two reconciled independent extractions. Damage to cell sites is negligible: 1.1% of attributed cell-site-days across six states, at most 3.8% anywhere. The sites were standing. What took them out divides by terrain: pooled, power dominates at 63.2%, but in mountainous North Carolina severed transport (backhaul) reaches 52.2% against 47.3%, and in Tennessee 69.9%. North Carolina's transport share rises from 7.0% to 85.0% across the event (\r{ho} = 0.92). Seventeen days after landfall, on 15 October, 47 sites lost transport across six contiguous North Carolina counties with no rainfall, no power loss, no damage, and recovery by the next report. Independent active-probe measurement corroborates it: responsive /24s fall 1.02% for twelve hours while Tennessee stays flat. We release the dataset. Backup power is the standard resilience investment; here it addresses the smaller half of the problem.

1. Introduction

The paper challenges a tower-destruction framing by decomposing Helene’s cellular outages into reported causes: cell-site damage was negligible, while transport and power varied in importance by terrain. It uses DIRS cause columns and a reconstructed dataset to examine that pattern and its evolution.

1. Introduction

  • Across six states, cell-site damage accounts for 1.1% of attributed site-days and never exceeds 3.8% in any state.The paper distinguishes damage to cell sites from physical destruction elsewhere in the network, including severed fibre.Source: §1 · pp. 2–3
  • Transport overtakes power only in the mountainous inland states, while commercial power remains the larger cause pooled across all six states.The paper scopes transport dominance to constrained-route terrain, where the outage was largest and longest.Source: §1 · p. 3

1.1. Background: the measurement instrument

  • DIRS daily reports separate outages attributed to cell-site damage, transport, and commercial power, a cause breakdown prior work had not analyzed.The paper uses these cause columns as its measurement instrument.Source: §1.1 · p. 3

1.2. Contributions •

  • The study reconstructs a per-cause, per-county daily record of Helene’s cellular outages and reconciles independent state- and county-level extractions.The dataset contains 80 state-days and 580 county-days.Source: §1.2 · p. 4
  • North Carolina’s transport share rises near-monotonically across the event, reaching 80.9% of attributed outages in the recovery tail.The trend is tested against serial correlation and mid-event changes in the reporting area.Source: §1.2 · p. 4
  • On 15 October, transport failed across six contiguous counties; weather, power, damage, and rescoping were tested against primary sources, and timing was confirmed by active probes.Source: §1.2 · p. 4

2. Related Work

Prior outage research often measured reachability or physical survival without decomposing cellular outages by reported operating cause. This paper uses DIRS cause columns to separate transport from power in a public cellular outage record.

Outage measurement from the outside

  • Active and passive Internet-measurement studies can attribute outages coarsely, but do not report operator-filed causes such as damage, transport, or power.The cited approaches have limited visibility into NAT-heavy mobile networks; DIRS instead covers declared disasters and participating providers.Source: §2 · pp. 4–5

Disasters and physical infrastructure

  • Prior disaster-infrastructure studies measure impacts or risks, but none decomposes a cellular outage record by reported cause.Transport and power appear separately in some work, but only qualitatively and for a single operator.Source: §2 · p. 5

Cellular resilience engineering

  • Cellular-resilience research models physical survival or backhaul resilience, while this paper reports observed cause shares from a real outage.The cited backhaul work relies on simulation or proprietary operational data unavailable to the community.Source: §2 · p. 5

FCC outage data outside networking

  • Earlier FCC outage studies used counts or prediction targets rather than this cellular transport-versus-power decomposition.Kuhn decomposed public-switched-telephone-network outages using a different cause taxonomy.Source: §2 · p. 6

Novelty

  • The authors make the narrow novelty claim that this is the first study to exploit DIRS cause decomposition, separating transport from power in a public cellular outage record.They do not claim to be the first to use DIRS.Source: §2 · p. 6

3. Methodology

The methodology reconstructs daily Helene outage data from FCC DIRS reports, independently extracts state and county tables, and checks their agreement. It defines cause shares and accounts for changes in the reporting area when analyzing trends.

3.1. Data acquisition

  • The authors retrieved 24 daily DIRS reports covering 26 September–19 October 2024, with cause data available for 23 days at a fixed 09:00 EDT reporting instant.The 26 September report predates state-level cause reporting and serves only for county baselines.Source: §3.1 · p. 7

3.2. Extraction and reconciliation

  • Because PDF text collapses adjacent numeric columns, the authors used separate extractions for state-level cause tables and county-level tables, then cross-validated them.A corrupt row was recovered only when solving against printed percentages and totals yielded a unique value; no values were interpolated.Source: §3.2 · pp. 7–8
  • Independent state- and county-table extractions reproduce the state totals exactly on all 45 reconcilable state-days.The check is a cross-table accounting identity, not two runs of the same parser.Source: Figure 1 · p. 7, §3.2 · p. 8

3.3. Dataset composition

  • The reconstructed dataset has unbalanced panels because states and counties enter and leave the FCC disaster area over time.Source: Table 2 · p. 9, §3.4 · p. 9

3.4. Definitions and derived quantities

  • Transport share uses attributed causes as its denominator, with the alternative share of all out-of-service site-days also reported.Cause counts leave 6.9% of North Carolina site-days and 7.6% pooled unattributed; the denominator choice does not change the North Carolina cause ordering.Source: §3.4 · p. 9
  • Trends use within-state shares rather than counts across reporting-area changes, though the authors also bound possible compositional effects from departing counties.Source: §3.4 · pp. 9–10

4. Experiment

The paper runs four analyses on the reconstructed dataset: cause decomposition, temporal behavior, investigation of the 15 October North Carolina anomaly, and external corroboration. Together, they test how causes vary across states and time and whether the anomaly appears in an independent measurement.

E1: cause decomposition

  • Cause decomposition integrates cell-site-days by cause over the full record, both per state and pooled, to test physical damage and terrain dependence.Source: §4 · p. 10

E2: temporal behaviour

  • Temporal analysis tests daily transport-share trends while correcting for serial correlation and changes in the reporting panel, then aggregates results into event phases.Source: §4 · p. 10

E3: the 15 October event

  • The 15 October analysis tests rescoping, weather, power, damage, and region-wide disruption as explanations before examining the residual’s county-level geography.Source: §4 · p. 10

E4: external corroboration

  • External corroboration checks whether IODA active probes detected the 15 October event, using Tennessee as a same-activation control measured the same way.Source: §4 · p. 10

5. Results

The results characterize how power and transport contributed to attributed outages during Helene, including their variation in North Carolina and Tennessee. They also examine the 15 October connectivity-only outage and its corroboration in statewide probe data.

5.1. E1: cause decomposition

  • Across Helene’s pooled record, power caused 63.2% of attributed site-days versus 35.7% for transport; damage was 1.1% on filed attributions and never exceeded 3.8% in any state.Assigning all unattributed site-days to damage raises its pooled share to at most 8.6%, yet leaves damage the smallest of the three causes.Source: §5.1 · pp. 11–12

5.2. E2: the temporal inversion

  • North Carolina’s transport share rose from 7.0% of attributed outages on 27 September to 85.0% on 19 October; all four states shifted toward transport, while starting levels and endpoints depended on terrain and record length.Source: §5.2 · pp. 13–15, Table 5 · p. 15

5.3. Does the rescoping produce the trend?

  • On 12 October, 94.8% of North Carolina’s out-of-service sites were in counties retained after rescoping, so the panel change could not explain the subsequent rise to 82.9%.The departing counties contained only 5.2% of the outage; the observed rise was from 71.4% on 12 October to 82.9% on 14 October.Source: §5.3 · p. 16

5.4. E3: the 15 October event

  • North Carolina’s outages rose from 54 sites on 13 October to 123 on 15 October, as transport outages climbed from 33 to 115 while power fell and damage stayed at zero.The two successive increases were entirely transport; the analysis focuses on the larger 14–15 October step, which also has independent corroboration.Source: §5.4 · p. 17

5.4.1. Reconciling the two numbers

  • The 14–15 October increase was +43 total sites out but +47 transport-caused sites out; these figures differ because cause columns are not a strict partition of the total.The county table counts total sites out and cannot provide a county-level cause breakdown.Source: §5.4.1 · p. 17

5.4.2. Eliminating alternatives

  • Every station across nine counties recorded 0.00 inches on 13 October; the maximum was 0.19 inches on 14 October and 0.06 inches on 15 October, while power outages fell and damage remained zero.Cable subscriber outages also fell, weakening a region-wide disruption explanation.Source: §5.4.1 · p. 18, §5.4.2 · p. 19

5.4.3. Geographic signature

  • Six contiguous counties lost sites while three interior counties did not, a pattern consistent with a provider- or route-specific failure rather than a purely geographic cause.The county pattern supports an inference of a shared upstream path, not an observation of county-level causes or identification of a specific facility, provider, or sites.Source: §5.4.3 · pp. 19, 22, Figure 4 · p. 20

5.5. E4: external corroboration

  • North Carolina’s responsive /24 blocks fell 1.02%—270 blocks—for twelve hours on 15 October, while Tennessee stayed within 0.18% of baseline; the controlled dip was 4.0 standard deviations below October’s median.The probe measures fixed-line address space that can share backhaul fibre with cell sites, so it corroborates the transport event only partially.Source: §5.5 · p. 23
  • At the 09:00 EDT snapshot, IODA recorded 26,131 responsive blocks on 15 October, down 213 from 14 October and recovered to 26,314 on 16 October.IODA’s North Carolina-specific excursion coincided within hours with the otherwise unexplained DIRS increase in out-of-service sites; its BGP signal changed by only 3 of 79,007 prefixes across the twelve-hour window. The corroboration is statewide, not county-level, and does not identify the facility or affected counties.Source: §5.5 · pp. 25–26

5.5.1. What the 15 October analysis establishes

  • On 15 October, 47 intact and powered sites lost service across six counties in one reporting interval, with no rainfall, power-outage rise, or reported damage.The event occurred 17 days after landfall and establishes a failure mode distinct from physical destruction. The inference to one shared backhaul facility remains indirect: DIRS does not identify providers or sites, and IODA cannot localize below the state level.Source: §5.5.1 · p. 26

5.6. Implications

  • Backup power addresses 47.3% of the North Carolina event and 63.2% pooled across six states, while transport diversity addresses the larger, more persistent share in the mountains.A regional broadcaster stayed on air by substituting satellite connectivity for terrestrial backhaul; FCC transport-facility reporting harmonization is described as a precondition for observing this failure mode (Federal Communications Commission, 2026c).Source: §5.6 · p. 27
  • The 15 October outage occurred amid intense reconstruction, when debris removal and rebuilding can sever fibre and potentially cause re-failure.If this generalizes, peak reconstruction activity is also a period of elevated topological failure, a feedback loop current planning does not represent.Source: §5.6 · p. 27
  • Mainstream communication applications require paths to distant servers, so a region with surviving local infrastructure but severed backhaul can lose them all at once.Source: §5.6 · p. 27

6. Limitations and Future Work

The authors delimit their claims by the aggregate, self-reported record and the single-event analysis, then identify cross-event testing, finer-grained measurement, and architectural evaluation as future work.

6.1. Limitations of the source

  • The cause columns do not partition outages exactly; damage is at most 8.6% pooled, while North Carolina’s transport lead reverses if over 66% of its residual is power.The authors cannot exclude such a residual skew, so they advise treating the North Carolina ordering as unresolved in that case.Source: §6.1 · p. 27
  • DIRS reporting-area rescoping changes denominators discontinuously, so raw subscriber-count shifts can reflect counties leaving the reporting area rather than recovery.North Carolina’s denominator fell from 1452 to 901 sites between the 12 and 13 October reports.Source: §6.1 · p. 28
  • DIRS does not identify providers or failed sites, cause classifications are self-reported and unaudited, and county-level causes are not observed.The county-level attribution and shared-facility explanation are inferences from spatial patterns, not measurements.Source: §6.1 · p. 28

6.2. Limitations of the analysis

  • Site-weighted outage shares characterize infrastructure, not population experience, because rural sites serve fewer subscribers than urban ones.The estimate of roughly 350,000 residents in western North Carolina counties experiencing transport-caused outage is order-of-magnitude only.Source: §6.2 · p. 28
  • The results describe one hurricane, and six states without route-diversity data cannot test the terrain mechanism; the trend is descriptive, not replicated or causal.A transport-caused outage is an upper bound on what architectural mitigation could address, not service available but unused.Source: §6.2 · pp. 28–29

6.3. Future work

  • Future work should test the terrain hypothesis across disasters, seek finer-resolution external outage measurements, and evaluate partition-tolerant designs against this trace.The authors also propose aggregated NORS disclosure across four or more providers to test the shared-path hypothesis.Source: §6.3 · p. 29

7. Conclusions

Across six states, damage at cell sites was negligible, while the dominant outage cause varied with terrain. In mountainous North Carolina, backhaul loss exceeded power loss and rose sharply during recovery.

7. Conclusions

  • 1.1% of attributed cell-site-days were due to site damage across six states, and no state exceeded 3.8%.The sites were, for the most part, standing and powered.Source: §7 · p. 30
  • In western North Carolina, backhaul accounted for 52.2% of attributed cell-site-days against 47.3% for power, rising from 7.0% to 85.0% across the event.On 15 October, 47 sites across six contiguous counties lost backhaul without weather, power loss, or damage; independent probes recorded the same twelve hours.Source: §7 · p. 30
  • The towers were mostly standing and powered; what they had lost was the path to the network core.Source: §7 · p. 30

CRediT authorship contribution statement

Oluseyi Olukola and Nick Rahimi contributed to the paper’s conception and methods, while all four authors contributed to writing.

CRediT authorship contribution statement

Data availability

The supplementary materials include the reconstructed outage dataset and associated analysis materials, with public deposit planned on acceptance.

Data availability

  • The reconstructed dataset covers 80 state-days and 580 county-days with per-cause breakdown.Source: Data availability · p. 30
  • The supplementary materials include extraction and reconciliation code, statistical checks, and figure sources.Source: Data availability · pp. 30–31
  • The materials will be deposited in a public repository under a CC-BY-4.0 licence with a citable DOI on acceptance.Source documents are public FCC filings, and complete NCEI and IODA request URLs are provided for rerunning the external checks.Source: Data availability · p. 31

Appendix A. Full daily record, North Carolina

Table A.8 presents the complete per-day North Carolina record underlying Figures 2 and 3, with denominator changes marked and no values interpolated.

Appendix A. Full daily record, North Carolina

  • Table A.8 gives the complete per-day North Carolina record underlying Figures 2 and 3.Source: Appendix A · p. 36
  • The table marks denominator changes and does not interpolate values.Source: Appendix A · p. 36
  • The panel narrows between 12 and 13 October, when the served count falls from 1452 to 901.Source: Appendix A · p. 36

Appendix B. Trend statistics

Table B.9 reports corrections for serial correlation to test whether it accounts for North Carolina’s rising daily transport share. The appendix explains the Durbin–Watson discrepancy and notes reporting-panel changes in the daily record.

Appendix B. Trend statistics

  • The Durbin–Watson statistic (1.16) is lower than the 1.57 and 1.69 lag-1 approximations because of an end effect: endpoint residuals carry 53% of residual sum of squares, almost all from the 27 September onset day.Source: Appendix B · p. 36
  • The North Carolina daily record flags served-count changes on 29 September and 13 October alongside the two-day rise on 14–15 October analyzed in §5.4.The transport share is computed against the sum of the three cause columns.Source: Table A.8 · p. 37
  • Using the larger lag-1 estimate for Bartlett correction is conservative; the Yule–Walker estimate gives effective n = 16.9 and p = 2.2 × 10−7.The moving-block bootstrap uses 104 resamples of contiguous 4-day blocks, matching the recovery timescale.Source: Appendix B · p. 37
  • Table B.9 reports three trend statistics so readers can judge corrections for serial correlation; the statistics establish that serial correlation does not account for the shift, not the shift’s size.A monotone recovery makes consecutive days correlated by construction, so the naive p-value is anticonservative.Source: Appendix B · p. 36, Table B.9 · p. 38

Appendix C. External check queries

The external checks are reproducible from public APIs, with query parameters and retrieved series included in the artifact.

Appendix C. External check queries

  • Precipitation checks use NCEI daily-summaries PRCP data from 31 stations across nine counties, while active-probe checks use the IODA v2 ping-/24 signal for North Carolina and Tennessee.The artifact includes full request URLs and retrieved series for both checks.Source: Appendix C · p. 38
Loading 2609.10944v1…