{"id":"f1507b56-5a29-41a3-8a4b-51f8446507ba","arxiv_id":"2508.20060","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Field measurements of operational 4G/5G networks show inter-cell interference is pervasive, largely uncoordinated, and worsened by frequency-selective fading, PCI collisions, and low-frequency-first resource allocation.","lead":"This paper measures real-world 4G/5G networks and finds that base stations rarely coordinate to avoid interfering with each other, so most users suffer avoidable signal-quality loss. It is one of the first detailed field studies of inter-cell interference, with practical implications for network operators and future interference-management algorithms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RB-level SINR estimator assumes zero-mean interference, but the paper's own PCI-collision results violate this on the measurement REs; the up-to-40 dB RB-level SINR gaps may be partly artifacts and should be re-verified.","rationale":"The reader's weakest-assumption identification is correct: the RB-level estimator in §II is load-bearing for the quantitative frequency-selective SINR conclusions, and §IV-D's PCI collision results directly contradict the zero-mean assumption used to justify m=228. I do not see a more serious flaw in the prevalence measurement, which is based on decoding neighbor PCIs and is fairly direct, nor in the raw observation that low-index RBs are preferred. The concern is not that the paper is wrong about the existence of interference; it is that the 40 dB RB-level gaps and the associated claim that current frequency-agnostic allocation is substantially suboptimal may be inflated by pilot contamination from PCI-colliding cells. This is an internal consistency issue rather than a disagreement with community consensus. The paper has real strengths: it reports a large measurement campaign, uses physical-layer signal processing, and its prevalence and low-RB-utilization observations are plausibly robust. However, the dataset and code are not yet public, and the estimator issue is testable by re-analysis of the recorded IQ data and by controlled simulation. Conditional acceptance remains the right verdict: the paper should be published with the requirement that the authors either demonstrate the RB-level findings survive after excluding or modeling PCI-collision-affected measurements, or clearly weaken the frequency-selective SINR claims. My read therefore does not change the reader's verdict.","tokens_in":15941,"tokens_out":8939,"duration_ms":117601,"concrete_test":"Use the recorded 4G IQ traces to recompute the RB-level interference and SINR-gap distribution separately for (a) UE locations where any interfering cell has the same RS position index as the serving cell (PCI mod 3 collision) and (b) locations free of such collisions. Then run a controlled simulation with known serving and interfering channel gains, injecting one PCI-colliding and one non-colliding interferer, to measure the estimator bias as a function of collision status. If the >30 dB SINR-gap tail and the >0.64 low-utilization collision probability persist in the collision-free subset, the concern is refuted; if they vanish, the paper's frequency-selective SINR claims are dominated by pilot contamination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of absent coordination and unnecessary interference leans on RB-level SINR gaps up to 40 dB and on the claimed suboptimality of current handover/resource allocation. That evidence rests on the estimator in §II: Eq. (4) assumes interference and noise over the m=228 reference-signal REs average to a small residual, and Theorem 1 chooses m under a zero-mean assumption (Appendix A). Section IV-D, however, reports that PCI collisions affect roughly 60% of 4G UEs and 70% of 5G UEs; under a PCI collision, a neighbor transmits on the same RS REs as the serving cell. If the colliding cell shares the same RS position index, the contamination term h2·X2/X1 in Eq. (3) does not average to zero over the measurement REs, so the LS estimate is biased. Subtracting the biased channel estimate in Eq. (5) means the residual is not true inter-cell interference. The paper itself states that PCI collisions degrade channel estimation, so its own observations invalidate the estimator's central assumption in a large subset of the very locations where the RB-level SINR gaps are measured. The prevalence finding—neighbor PCIs decoded per location—and the qualitative observation that BSs allocate RBs from the lowest index upward are more robust, but the quantitative frequency-selective SINR distribution, and the resulting recommendation to change handover/resource allocation, is not independently supported until the collision-contaminated measurements are separated out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a measurement study of inter-cell interference in operational 4G/5G networks. Using a smartphone for RSRP/SINR collection and USRPs for OFDM symbol capture, the authors detect interfering cells by decoding PCIs, estimate UE-level interference from RSRP minus SINR, and estimate RB-level interference via LS channel estimation from reference-signal REs. They report prevalence of interference, its impact on SINR, intra-BS interference, and four root causes: deployment density, imbalanced channel assignment, low-frequency-first RB allocation, and PCI collisions. The central claim is that inter-cell interference is prevalent and largely uncoordinated, so substantial signal-quality gains are available.","tokens_in":16331,"tokens_out":6930,"duration_ms":81687,"significance":"If the qualitative findings hold, this is a valuable first look at a practically important but under-measured problem. The study's strengths include a large field dataset (1,389 locations, 20,840 measurement points), a blanket cell search, the combination of commodity and SDR instrumentation, and a plan to release artifacts. Several observations are directly decoded from the air interface and do not depend on the contested estimator: the prevalence of interfering neighbors, the low-index-first RB allocation pattern, and the occurrence of PCI collisions. These alone would be useful to the community. However, the quantitative RB-level SINR claims, including the up-to-40 dB gaps and the consequent handover/resource-allocation recommendations, rest on an estimator whose central assumption is violated by the paper's own PCI-collision findings. The significance of the paper is therefore conditional on repairing or re-scoping that estimator.","major_comments":[{"comment":"The RB-level estimator assumes that interference and noise on the reference-signal REs are zero-mean, so averaging m REs isolates h1. Theorem 1 and the choice m=228 are derived under this assumption. However, §IV-D reports that about 60% of 4G UEs and 70% of 5G UEs experience PCI collisions, meaning a neighbor transmits RSs on the same REs as the serving cell. In Eq. (3), the contamination term h2·X2/X1 is then not zero-mean over the averaged REs; the LS channel estimate is biased, and the residual in Eq. (5) is not purely inter-cell interference but contains estimation error. The paper itself states that PCI collisions degrade channel estimation, confirming that its own data invalidate the estimator's core assumption for a large subset of locations. The up-to-40 dB RB-level SINR gaps and the handover/resource-allocation conclusions are not independently supported until the analysis sepa","section":"§II, Eq. (3)-(5), Appendix A; §IV-D"},{"comment":"The 40 dB RB-level SINR gap and the frequency-selective-fading impact are measured only for 4G; the text says 'our measurement focuses on 4G, but the conclusions generalize to 5G' without providing RB-level SINR estimates for 5G. The abstract and conclusion nevertheless present these as findings for both 4G and 5G networks. Since 5G uses a different reference-signal structure (SSB rather than dense CRS) and the paper's own Appendix B does not supply 5G RB-level data, this generalization is unsupported. Please either provide 5G RB-level measurements or explicitly restrict the quantitative RB-level claims to 4G.","section":"§III-B and Abstract"},{"comment":"The claim that BSs 'consistently' allocate RBs from the lowest index upward, and the resulting high collision probability at low utilization, appear to be based on a single pair of 4G cells monitored with two USRPs. The paper does not report the number of cell pairs, total traces, or cell-pair-level variability behind Figs. 12 and 13. Appendix B extends the observation to urban and rural areas but gives no methodological details (number of cells, trace durations, selection criteria). Because this is a load-bearing part of the 'absence of coordination' conclusion, the sampling basis should be stated clearly and the wording should be matched to the evidence.","section":"§IV-C, Figs. 12-13, Appendix B"}],"minor_comments":[{"comment":"The proof writes 'Var(|~Δ_i|^2) = E[|~Δ_i|^2]', which is not an equality as written; the object used in Bennett's inequality needs to be defined precisely. Please correct the notation and clarify whether the bound is on the variance of the real/imaginary parts or on the second moment.","section":"Appendix A"},{"comment":"Equation (1) defines UE-level interference from RSRP and reported SINR. This is an operational definition and makes the algebraic relationship tautological. The paper should state more explicitly that this metric inherits any implementation-specific filtering in the device's reported SINR.","section":"§II"},{"comment":"The paper says PCI collisions are 'detected' at each UE location, but the detection procedure is not described. Please specify how a collision is identified from the measurements (e.g., by comparing decoded PCIs modulo the number of RS patterns or by detecting RS-power superposition).","section":"§IV-D"},{"comment":"The 5G n41 center-frequency entries '2524.95(2565)' and '2662.95(2644.80)' are confusing; please clarify the intended center frequencies and bandwidths.","section":"Table I"},{"comment":"The caption says 'RB-level interference (left) and RSRP (right)', but the axes are labeled 'Subcarrier' and 'Symbol' with powers. Clarify whether the displayed quantity is per-RE, per-RB, or per-symbol power.","section":"§III-B, Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the conflict between the RB-level estimator's zero-mean interference assumption and the paper's own PCI-collision results. If the authors can re-run the RB-level analysis excluding collision-affected measurements, or show quantitatively that the bias is negligible, the paper could become acceptable. Given the paper is a measurement study, the planned artifact release should also be made a condition of publication so that the statistics and detection procedures can be audited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the good news: this is the only field study I know that systematically documents inter-cell interference in live 4G/5G networks. The measurement campaign is serious: USRP capture plus smartphone logs across campus, urban, and rural, covering multiple ISPs and bands. The qualitative findings—low-frequency-first RB allocation, high collision probability between neighbors even at low load, and PCI collision prevalence (60% of 4G UEs, 70% of 5G)—are new and by themselves valuable. The paper reads well and the authors are appropriately cautious about 5G-specific limitations.\n\nThe soft spot is real and it is not minor. The RB-level interference estimator in Section II and Appendix A relies on the assumption that interference and noise on the averaged reference-signal REs are zero-mean. That is fine in principle. But Section IV-D shows that PCI collisions are common: when two cells share the same RS position, the interfering cell's signal is persistent on exactly those REs, so the contamination does not average out. The paper itself admits PCI collisions degrade channel estimation. That means the up-to-40 dB RB-level SINR gaps and the claims about suboptimal handover/resource allocation are not independently supported unless the authors re-estimate after excluding collision-contaminated locations or extend the estimator to model a biased interferer. This is a load-bearing flaw for the quantitative claims, not a cosmetic one.\n\nWhat survives the flaw: the prevalence figures, the RB-usage traces showing ascending allocation, and the collision-probability matrix are direct measurements and do not depend on the disputed estimator. Anyone working on interference coordination should know those results.\n\nThe dataset and code are promised but not yet available, so independent verification is currently impossible.\n\nMy bottom line: this deserves a serious referee. It should not be desk-rejected. But it needs a major revision before acceptance—separate the PCI-collision cases from the RB-level SINR analysis, or relax the estimator's assumption and bound the bias. For my own work I would not cite the 40 dB numbers until that is sorted, but I would cite the qualitative findings.\n\nWho is this for: network measurement researchers and anyone designing SON or resource allocation for 4G/5G. Bring it to a reading group—it will generate discussion about when measurement-derived models are trustworthy.\n\nRecommendation: send to peer review with an explicit request to run the RB-level analysis both with and without PCI-collision cases and quantify the difference.","headline":"First useful field study of inter-cell interference, but its headline RB-level SINR numbers rest on an estimator that the paper's own PCI-collision results undermine.","tokens_in":16770,"tokens_out":3541,"would_cite":true,"duration_ms":38890,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A first measurement study of live 4G/5G networks finds inter-cell interference is ubiquitous and uncoordinated.","keywords":["inter-cell interference","4G/5G measurement","resource-block-level SINR","frequency-selective fading","PCI collision","resource allocation","network deployment","signal quality"],"falsifier":"Re-run the RB-level analysis excluding reference-signal resource elements whose positions collide with a detected neighbor's PCI (or use only PCI-collision-free cells); if the >30 dB SINR gaps and >64% low-load collision probabilities vanish, the central claims rest on the estimator's zero-mean assumption rather than on network behavior.","tokens_in":15870,"feed_emoji":"📶","tokens_out":7014,"duration_ms":72710,"temperature":0.7,"pith_summary":"This paper sets out to measure something theory has long predicted but no one has checked in the field: how much inter-cell interference actually occurs in working 4G and 5G networks, and whether operators do anything about it. Using phone-level signal reports and software-defined radios that decode the physical layer, the authors find that every measured cell has at least one interfering neighbor, that almost all users sit in interference, and that none of the four management levers—deployment, channel assignment, time-frequency scheduling, and cell-identity configuration—is used to avoid it. A central result is that base stations systematically schedule the lowest-frequency resource blocks first, so neighboring cells collide on the same resources even when spectrum is mostly idle. If this picture holds, large signal-quality gains are available from comparatively simple coordination, rather than from building more base stations.","feed_headline":"Nearly every 4G/5G user faces avoidable neighbor-cell interference","feed_subtitle":"Field measurements show base stations collide on the same spectrum even when idle—simple fixes would help.","key_machinery":"The key mechanism is an RB-level interference estimator built from reference-signal resource elements. The received signal on each reference element is modeled as desired channel times known symbol plus interference plus noise; averaging m=228 reference elements (chosen via a Bennett-inequality bound so residual interference and noise are small) isolates the desired channel, and subtracting it leaves an interference estimate per resource block. The argument also turns on two named objects: the collision probability |A∩B|/|A| for two cells' scheduled resource-block sets, and the PCI position index (PCI modulo X) that places reference signals on identical resource elements when cells collide.","core_discovery":"The paper claims that inter-cell interference is not a theoretical edge case but the normal operating condition of deployed 4G/5G networks. On a 2.5 km by 1.2 km campus and in follow-up urban and rural measurements, it identified 132 4G cells and 197 5G cells and found every cell has at least one interferer on its channel; 5G users see roughly twice as many interfering cells as 4G users. At resource-block granularity, frequency-selective fading makes SINR vary by up to 40 dB across the band at a single user location, so a handover decision based on a single wideband RSRP or on the narrow SSB window can pick a poor cell or frequency. The root causes are concrete: an imbalanced channel assignm","pith_inferences":["The zero-mean estimator assumption and the PCI-collision findings collide; excluding collided reference elements is a direct robustness check the paper does not report.","Extending RB-level estimation to 5G data regions would require scheduled DM-RS or CSI-RS, which the paper flags as future work but could be enabled by logging scheduler grants.","If the measured low-to-high RB allocation is a vendor default rather than deliberate policy, a configuration change could recover most of the lost SINR without new algorithms.","The same measurement design can be reused to quantify gains before and after an operator changes PCI plans or scheduler policies, turning the paper's static findings into a before-after evaluation."],"forward_implications":["Per-resource-block SINR becomes a practical scheduling and handover input; wideband or SSB-window RSRP misses 30–40 dB dips and should be replaced.","Schedulers that spread resource blocks across the band instead of filling lowest-indexed RBs first would sharply reduce inter-cell collisions, including at low load.","PCI planning that avoids sharing reference-signal positions with strong neighbors would recover roughly 4 dB SINR for the most-interfered 5G users and lift SIB decode success.","In 5G, densification's main benefit is capacity, not SINR; interference coordination is the lower-cost lever for signal quality.","Because the estimator runs on commodity phone-plus-USRP hardware, operators and researchers can audit interference coordination continuously, not just once."],"supporting_citations":[{"why":"Supplies the smartphone measurement tool (CellularZ) that records the UE-level RSRP and SINR samples used throughout the study.","marker":"[7]"},{"why":"Documents the handover practice (UE-level RSRP over a limited frequency range) that the RB-level results argue is suboptimal.","marker":"[6]"},{"why":"Provides the lowest 5G MCS reference used to set the target RE-level SINR of -6.7 dB, fixing the averaging parameter m=228.","marker":"[10]"},{"why":"Bennett's inequality is the probabilistic bound that Theorem 1 uses to guarantee the channel estimator's error stays below threshold.","marker":"[33]"},{"why":"Shows RSRP and SINR are continuously available on smartphones, grounding the UE-level interference estimate in equation (1).","marker":"[9]"},{"why":"Explains RS-based interference-measurement principles in 5G NR, the acknowledged basis for the paper's measurement approach.","marker":"[19]"},{"why":"Surveys inter-cell interference coordination techniques, the decades of prior work whose real-world absence the paper documents.","marker":"[4]"}],"fun_headline_variants":["4G/5G cells collide on spectrum even when idle—fixes ignored","Every 4G/5G cell has an interferer; coordination absent","Real-world 4G/5G: inter-cell interference is the norm","Base stations reuse same resources despite idle spectrum"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The RB-level interference estimate assumes interference and noise on reference-signal tones are zero-mean random and average out, but the paper's own PCI-collision findings put persistent, non-random reference signals on those same tones—so the largest SINR gaps could be partly an artifact of the estimator.","fun_headline_variants_meta":{"raw":{"variants":["4G/5G cells collide on spectrum even when idle—fixes ignored","Every 4G/5G cell has an interferer; coordination absent","Real-world 4G/5G: inter-cell interference is the norm","Base stations reuse same resources despite idle spectrum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":2981,"prompt_tokens":712,"completion_tokens":2269,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":2201}},"tokens_in":456,"tokens_out":2269,"duration_ms":16111,"temperature":1.0,"reasoning_tokens":2201,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:13:30.633022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the RB-level analysis excluding reference-signal resource elements whose positions collide with a detected neighbor's PCI (or use only PCI-collision-free cells); if the >30 dB SINR gaps and >64% low-load collision probabilities vanish, the central claims rest on the estimator's zero-mean assumption rather than on network behavior.","supporting_citations":[{"cited_title":"Energy- efficient paging for duty-cycled lte backscatter,","cited_arxiv_id":null,"evidence_quote":"Supplies the smartphone measurement tool (CellularZ) that records the UE-level RSRP and SINR samples used throughout the study."},{"cited_title":"Leveraging context-triggered measurements to characterize LTE handover performance,","cited_arxiv_id":null,"evidence_quote":"Documents the handover practice (UE-level RSRP over a limited frequency range) that the RB-level results argue is suboptimal."},{"cited_title":"Method and apparatus for reporting channel state information for supporting 256qam in wireless access system,","cited_arxiv_id":null,"evidence_quote":"Provides the lowest 5G MCS reference used to set the target RE-level SINR of -6.7 dB, fixing the averaging parameter m=228."},{"cited_title":"Mobile access bandwidth in practice: measurement, analysis, and implications,","cited_arxiv_id":null,"evidence_quote":"Bennett's inequality is the probabilistic bound that Theorem 1 uses to guarantee the channel estimator's error stays below threshold."},{"cited_title":"Icellspeed: Increasing cellular data speed with device-assisted cell selection,","cited_arxiv_id":null,"evidence_quote":"Shows RSRP and SINR are continuously available on smartphones, grounding the UE-level interference estimate in equation (1)."},{"cited_title":"Interference measurement methods in 5G NR: Principles and performance,","cited_arxiv_id":null,"evidence_quote":"Explains RS-based interference-measurement principles in 5G NR, the acknowledged basis for the paper's measurement approach."},{"cited_title":"A survey on inter-cell interference coordination techniques in ofdma-based cellular networks,","cited_arxiv_id":null,"evidence_quote":"Surveys inter-cell interference coordination techniques, the decades of prior work whose real-world absence the paper documents."}],"review_version":1}