{"id":"b2c582bb-5876-4d36-8f9c-99fb15e73073","arxiv_id":"2501.15729","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":11,"one_line_summary":"A 5-tap tapped-delay-line channel model with a two-state Markov birth-death process is fitted to 2.16 GHz railway measurements and validated against the same dataset and 3GPP RMa.","lead":"This paper builds a five-path radio channel model for 5G railway networks from measurements taken at 2.16 GHz on a test track. It adds a Markov switch that turns paths on and off to reflect the changing environment, and compares the model's delay spread with measured data and the 3GPP standard.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Markov state interval is defined as coherence time but implemented as 100 ns delay resolution, so Table I transition probabilities are tied to an unreported snapshot rate and cannot support the claimed non-stationary model.","rationale":"The central claim requires a Markov TDL whose transition probabilities describe MPC birth/death at a physically meaningful rate. That rate is set by the state decision interval, so the interval is load-bearing. The paper conflates coherence time with delay resolution: it uses 100 ns in Eq. (3) for tap-count determination and then silently uses the same number as the Markov state interval. This is not merely a terminology issue; it changes the computed number of taps and the meaning of every transition probability in Table I. The sounder's snapshot interval is never reported, so the model cannot be reproduced or transferred to other speeds or snapshot rates. The validation via RMS DS PDFs is too coarse to test the birth-death dynamics. The reader's weakest assumption identified the same issue, and I agree. The measurement effort is genuine and could support a revised model, but as published the non-stationary Markov TDL is not validated. I would keep the REJECT verdict rather than downgrade to conditional, because the missing coherence-time estimate and snapshot-rate reporting are prerequisites for any meaningful parameter table.","tokens_in":8502,"tokens_out":4260,"duration_ms":40951,"concrete_test":"Estimate the true coherence time from the measured CIRs, e.g., the 50% threshold of the time-frequency correlation function, or from the measured Doppler spectrum; report the actual snapshot interval of the sounder. Re-estimate the Markov transition probabilities with the state decision interval set to the estimated coherence time rather than 100 ns, and regenerate Fig. 4 at that interval. If p00/p11/p1 change materially from Table I, or if the regenerated RMS DS PDF no longer matches the measured PDF, the model's switching dynamics are an artifact of the sampling interval and the validation claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing defect is the ill-defined Markov state decision interval. Section III-A defines this interval as the channel coherence time, but Section III-B substitutes Tc=100 ns—the delay resolution of the 10 MHz sounder—into Eq. (3). These are not the same quantity: at fmax=160 Hz a standard coherence-time estimate is about 1/(4fmax)≈1.56 ms, four orders of magnitude larger. Eq. (3) is therefore dimensionally inconsistent; if Tc were the true coherence time, L=max(στ)/Tc+1 would be 1, not 5, so the advertised tap count itself rests on the same mislabeling. Because Table I's p00, p11, and p1 are estimated from state sequences sampled at the sounder's snapshot interval, which is never stated, the transition probabilities are tied to that unknown interval. Any link-level simulation at a different snapshot rate or train speed cannot reuse Table I, and the model is not transferable as claimed. The validation in Section III-C compares only RMS DS PDFs; that aggregate statistic is insensitive to the time scale of birth-death switching, so it cannot detect this defect. The measurement campaign is real and valuable, but the published parameters do not yet constitute a reproducible non-stationary TDL model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a channel measurement campaign for a 5G railway private network at 2.16 GHz with 10 MHz bandwidth and uses the data to construct a 5-tap tapped-delay-line (TDL) channel model. The model augments each tap with a binary switching function driven by a first-order two-state Markov chain to represent multipath component birth and death. Parameters are reported for tap delays, powers, lognormal amplitude distributions, phase and Doppler distributions, transition probabilities, and a tap amplitude correlation matrix. The model is compared with measured RMS delay-spread distributions and with the 3GPP 38.901 RMa model, with the claimed result that the proposed model better captures the non-stationarity of the 5G-R channel.","tokens_in":8821,"tokens_out":4654,"duration_ms":40145,"significance":"If the parameter set were internally consistent and independently validated, the model would fill a real gap: a measurement-based non-stationary TDL model for the 5G-R band, useful for link-level simulation. The campaign at the National Railway Track Test Center with a real 5G-R dedicated base station and a train at 80 km/h is valuable, and the reporting of a tap correlation matrix and transition probabilities goes beyond standard WSSUS TDL models. However, as presented, the model cannot be used or reproduced because of fundamental parameter inconsistencies: the tap-count formula conflates delay resolution with coherence time, the amplitude distribution is given two different values, and the validation data are the same data used for fitting. These problems are load-bearing for the central claim.","major_comments":[{"comment":"In Eq. (3), Tc is called the coherence time, but in the implementation it is set to 100 ns, which is the delay resolution of the 10 MHz sounder rather than a time-domain coherence interval. At 80 km/h and 2.16 GHz, fmax = 160 Hz, so a standard coherence-time estimate is about 1/(4fmax) ≈ 1.56 ms; substituting that value into Eq. (3) yields L = 1, not 5. The advertised 5-tap count therefore rests on the conflation of delay resolution with coherence time, and it is not a robustly derived model parameter.","section":"Section III-B, Eq. (3)"},{"comment":"The tap amplitude parameter is specified inconsistently: Table I reports αl ~ LN(-3.66, 1.08), while the simulation diagram in Fig. 3 states 'LN(-1.91, 0.67)'. Since Fig. 3 is the generator used for the validation in Section III-C, the reader cannot tell which distribution was actually simulated. This ambiguity makes the reported model parameters non-reproducible.","section":"Table I and Fig. 3"},{"comment":"The validation is circular. Section III-B states that the data from the two measurement cycles not used in [33] are used 'to establish Markov TDL model and then verify it,' and Section III-C validates against the same data. Because the amplitudes, transition probabilities, and correlations are all estimated from these laps, the RMS DS PDF comparison in Fig. 4 partly re-describes the fitting data and provides no independent evidence for the model's predictive accuracy.","section":"Section III-B and Section III-C"},{"comment":"The Markov state decision interval is defined as the channel coherence time, but the coherence time is never measured or reported. The transition probabilities p00, p11, and p1 in Table I are estimated from state sequences sampled at the sounder's snapshot interval, which is also not stated. These probabilities are therefore tied to an unknown time granularity, and the model cannot be transferred to a different snapshot rate or train speed as claimed.","section":"Section III-A and Table I"},{"comment":"The RMS delay spread is an aggregate statistic that is insensitive to the time scale of MPC birth-death switching; even a WSSUS model could produce a similar RMS DS distribution. The validation therefore does not exercise the Markov switching mechanism, which is the central novelty of the model. A validation of the non-stationary behavior would need to compare metrics such as tap lifetime distributions or state-switching dynamics, not only the RMS DS PDF.","section":"Section III-C, Fig. 4"}],"minor_comments":[{"comment":"There are typographical errors: 'Notely' in Section III-B and 'fallowing' in Fig. 3 should be corrected.","section":"Section III-B, Fig. 3"},{"comment":"The notation in Eq. (2) uses S for the steady-state vector, but S is typeset in a way that suggests a matrix; clarify that S = [p0, p1]^T is a vector and define its relationship to the transition matrix T.","section":"Eq. (2)"},{"comment":"The phase distribution is reported as U[0, π] for all taps; since phase is a circular quantity, clarify whether the support is meant to be [0, 2π) or whether the interval [0, π] is intentional.","section":"Section III-B, item 2"},{"comment":"Reference [28] (Jin and Wang, on MEMS capacitive microwave power sensors) appears unrelated to non-stationary TDL channel modeling and should be replaced or removed.","section":"References"},{"comment":"The goodness of fit in Fig. 4 is judged visually; report a quantitative test such as a two-sample Kolmogorov-Smirnov test on the RMS DS distributions to support the claimed agreement.","section":"Fig. 4"},{"comment":"The paper states that two cycles were previously used in [33] and two cycles are used here, but the total number of measurement laps and the snapshot interval are not reported; include this information for reproducibility.","section":"Section II"}],"recommendation":"reject","confidential_remarks":"The measurement campaign is real and the topic is timely, but the manuscript has internal inconsistencies in the tap-count derivation and in the amplitude parameters, and the validation is circular. These are not merely presentational; they affect the published parameter set's reproducibility. If the authors revise, they should re-derive the tap count with the correct coherence-time definition, report the snapshot rate and a measured coherence-time value, reconcile the amplitude distributions, and validate on a held-out lap with a metric sensitive to birth-death dynamics. Given the scope of these changes, I would not recommend a quick revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward extension of existing Markov birth-death TDL work to the 5G-R downlink band at 2.16 GHz. That is a legitimate and useful gap: the prior viaduct and V2V models don't cover this band/bandwidth, and the authors did a real measurement campaign on a loop track. The basic TDL structure, with a two-state Markov switch per tap, is sound, and the correlation matrix plus transition probabilities, if correct, would be directly usable by link-level simulators.\n\nThe problems are in the details, and they are not minor. First, Eq. (3) conflates the delay resolution with the coherence time. They define the Markov state interval as the channel coherence time, then plug in Tc=100 ns from the 10 MHz bandwidth. That isn't the coherence time; at 80 km/h and 2.16 GHz you'd expect roughly 1.5 ms. The tap count formula gives floor(350/100)+1 = 4, not 5. More importantly, the transition probabilities in Table I are estimated from state sequences sampled at whatever snapshot rate the sounder used, and that rate is never stated. Those probabilities are tied to an unknown time interval and can't be transferred to another simulation without conversion.\n\nSecond, the validation is circular. The text says they use two laps for the model and then verify it on the same two laps. There is no independent validation data. The RMS DS PDF comparison with the same data is a fitting check, not a validation.\n\nThird, there are internal inconsistencies: Table I gives lognormal parameters LN(-3.66, 1.08) while Fig. 3 instructs simulation with LN(-1.91, 0.67). One is a typo, but the paper is unreproducible as published. The phase and Doppler are assumed uniform, not extracted, which also doesn't match the abstract's claim.\n\nNone of this kills the underlying measurement value. The campaign is real, and the approach is the right one for non-stationary railway channels. But the current version doesn't support the validation claim, and the parameter table can't be trusted until the coherence-time issue is fixed and the parameters are corrected.\n\nThis deserves a serious peer review: the topic matters, the data exists, and the flaws are identifiable and fixable. I'd send it out, but I'd expect major revision. For my own work, I wouldn't cite the parameter table as-is.\n\nBest","headline":"A real 5G-R measurement campaign, but the validation is circular and the parameter table has inconsistencies that make the current model unusable as published.","tokens_in":9438,"tokens_out":5053,"would_cite":false,"duration_ms":43983,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A five-tap tapped-delay-line model, with each tap switched on and off by a first-order two-state Markov chain, reproduces the non-stationary fading of a 5G-railway channel at 2.16 GHz, and is validated against independent measurements.","keywords":["5G-Railways","tapped delay line model","non-stationary channel","Markov chain","birth-death process","channel measurement","RMS delay spread","link-level simulation"],"falsifier":"Estimate the coherence time directly from the measured channel (for example from the Doppler spread or the autocorrelation of the frequency transfer function) and re-derive the Markov transition probabilities using that interval. If the resulting $p_{00}$, $p_{11}$, and $p_1$ values differ appreciably from Table I, or if the simulated RMS delay spread PDF no longer matches the independent measurement set, the paper's state-interval assumption would be falsified. A simpler check is to repeat the measurement at a different train speed and see whether the transition probabilities change as predicted.","tokens_in":8260,"feed_emoji":"🚄","tokens_out":11420,"duration_ms":88426,"temperature":0.7,"pith_summary":"The paper claims that a five-tap tapped-delay-line (TDL) model, whose taps are switched on and off by a first-order two-state Markov chain, captures the non-stationary behavior of a 5G-railway (5G-R) downlink channel measured at 2.16 GHz with 10 MHz bandwidth on a railway test track. From the measured data it extracts the tap delays, powers, amplitude, phase and Doppler distributions, the Markov state-transition probabilities, and correlations between tap amplitudes. It then validates the model on an independent set of measurements by comparing the probability density of root-mean-square delay spread against the measured data and against the standardized TDL model, finding a much closer match than the standard model. If the claim holds, link-level simulations of 5G-R systems at this band can use the provided parameters instead of assuming stationary fading. The model is intended for the 5G-R dedicated frequency band at 2.16 GHz.","feed_headline":"Five-tap Markov model matches measured 5G-railway channels","feed_subtitle":"Measurement-based birth-death taps reproduce non-stationary delay spread at 2.16 GHz better than the standard model.","key_machinery":"The central object is the per-tap switching function $z_l(t)\\in\\{0,1\\}$ in the TDL impulse response $h(\\tau,t)=\\sum_l z_l(t)\\alpha_l(t)e^{j[\\phi_l(t)+2\\pi f_{D,l}(t)]}\\delta(\\tau-\\tau_l(t))$. It is modeled as a first-order two-state Markov chain with transition matrix $T=[[p_{00},p_{01}],[p_{10},p_{11}]]$, where state 1 means the tap is present and state 0 means it is dead; the state is updated every channel coherence time. This mechanism converts a standard WSSUS TDL model into a non-stationary one by letting multipath components appear and disappear. The tap-existence states are extracted from measured power delay profiles using a 6 dB-above-noise-floor threshold, and the transition probabilities are estimated from those states.","core_discovery":"The central discovery is that the non-stationarity of the railway channel, understood as the 'birth and death' of resolvable multipath components, can be represented by a per-tap switching function $z_l(t)$ driven by a two-state Markov chain, and that all required parameters can be obtained from a measurement campaign. The resulting model has five taps at relative delays 0, 0.1, 0.2, 0.3, and 0.4 microseconds with average powers 0, -3.14, -17.02, -26.31, and -39.35 dB; the first tap always exists, while later taps have steady-state existence probabilities of 0.92, 0.77, 0.57, and 0.46. Each tap's amplitude follows a tabulated lognormal distribution, its phase is uniformly distributed on $[0,\\pi]$, and its Doppler shift is uniformly distributed on $[-160,160]$ Hz for the 80 km/h train speed. The paper also reports the tap-amplitude correlation coefficients, which are symmetric and reach 0.77 between taps 1 and 3. Validation compares the PDF of RMS delay spread from the simulated model, the measured data, and the standard TDL model; the proposed model's distribution is claimed to be much closer to the measurements.","pith_inferences":["We infer that the Markov transition probabilities are tied to the 100 ns state-update interval used in the paper; at a different train speed or snapshot rate the transition matrix would need re-estimation, since the coherence-time basis of the state interval is not independently measured.","A testable extension would be to estimate the channel coherence time from the Doppler spectrum or frequency-correlation function and re-derive the transition probabilities, checking whether the model's RMS delay spread fit is preserved.","We infer that the reported tap-amplitude correlations, especially the 0.77 value between taps 1 and 3, could be exploited in joint channel estimation for MIMO or multi-band 5G-R links, although the paper itself only tabulates them.","If the birth-death process is confirmed to track physical scatterer visibility, the model could be extended to predict handover or beam-management triggers in railway networks, but that connection is not made in the paper."],"forward_implications":["Link-level simulations for 5G-R at 2.16 GHz can directly use the tabulated five-tap parameters, Markov transition matrix, amplitude/phase/Doppler distributions, and tap correlation matrix to generate non-stationary channels.","Simulated RMS delay spread from the proposed model matches the measured PDF much better than the standardized TDL model, supporting its use for performance evaluation of 5G-R air interfaces.","Because the first tap always exists and the steady-state existence probability decreases with tap index, the model quantifies how often only two or three taps are resolvable in the measured rural scenario.","The dominance of $p_{00}$ and $p_{11}$ implies that multipath components persist over successive coherence intervals, so the channel's non-stationarity evolves on a time scale slower than small-scale fading but faster than large-scale effects.","The tap-amplitude correlation coefficients, not present in standard WSSUS TDL models, enable simulation of correlated fading across taps for receiver algorithms that exploit delay-domain structure."],"supporting_citations":[{"why":"Introduces the first-order Markov birth-death TDL approach for high-speed railway viaduct scenarios that this paper adapts to 5G-R.","marker":"[12]"},{"why":"Provides the Markov-chain state transition formulation for railway channels, directly reused as the two-state model here.","marker":"[24]"},{"why":"Supplies the measurement-based Markov modeling method and the 6 dB-above-noise-floor threshold used to decide tap presence from power delay profiles.","marker":"[32]"},{"why":"Gives the method for computing the number of TDL taps from maximum RMS delay spread and time resolution.","marker":"[25]"},{"why":"Shows parameter extraction for non-stationary TDL models, including lognormal amplitude fitting and tap-amplitude correlation coefficients.","marker":"[26]"},{"why":"Reports the same measurement campaign's delay-spread characterization, from which the maximum RMS delay spread (over 350 ns) used in the tap count is taken.","marker":"[33]"},{"why":"Defines the standardized TDL channel model used as the baseline whose RMS delay spread PDF is compared in validation.","marker":"[14]"},{"why":"Supports the choice of lognormal amplitudes and uniform phase and Doppler distributions for non-stationary TDL models.","marker":"[35]"}],"fun_headline_variants":["Measurement-based 5-tap Markov model beats standard for 5G-rail","Tracked: Markov TDL model captures non-stationary 5G-rail channels","Five-tap Markov model from real 5G-rail measurements","Non-stationary 5G-rail channel model built on measured taps","Markov tap model reproduces 5G-rail non-stationarity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's behavior is set by the Markov state decision interval, which the paper identifies with the channel coherence time; however, the coherence time is never estimated from the data, and the text instead uses the 100 ns delay resolution of the sounder as this interval, so the transition probabilities in Table I are only calibrated to that particular snapshot cadence and train speed.","fun_headline_variants_meta":{"raw":{"variants":["Measurement-based 5-tap Markov model beats standard for 5G-rail","Tracked: Markov TDL model captures non-stationary 5G-rail channels","Five-tap Markov model from real 5G-rail measurements","Non-stationary 5G-rail channel model built on measured taps","Markov tap model reproduces 5G-rail non-stationarity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3066,"prompt_tokens":1052,"completion_tokens":2014,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":1911}},"tokens_in":668,"tokens_out":2014,"duration_ms":11926,"temperature":1.0,"reasoning_tokens":1911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:59:37.841332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the coherence time directly from the measured channel (for example from the Doppler spread or the autocorrelation of the frequency transfer function) and re-derive the Markov transition probabilities using that interval. If the resulting $p_{00}$, $p_{11}$, and $p_1$ values differ appreciably from Table I, or if the simulated RMS delay spread PDF no longer matches the independent measurement set, the paper's state-interval assumption would be falsified. A simpler check is to repeat the measurement at a different train speed and see whether the transition probabilities change as predicted.","supporting_citations":[{"cited_title":"The dynamic evolution of multipath components in high-speed railway in viaduct scenarios: From the birth-death process point of view,","cited_arxiv_id":null,"evidence_quote":"Introduces the first-order Markov birth-death TDL approach for high-speed railway viaduct scenarios that this paper adapts to 5G-R."},{"cited_title":"Markov chain based channel characterization for high speed railway in viaduct scenarios,","cited_arxiv_id":null,"evidence_quote":"Provides the Markov-chain state transition formulation for railway channels, directly reused as the two-state model here."},{"cited_title":"Measurement-based markov modeling for multi-link channels in railway communication systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the measurement-based Markov modeling method and the 6 dB-above-noise-floor threshold used to decide tap presence from power delay profiles."},{"cited_title":"Vehicle–vehicle channel models for the 5- GHz band,","cited_arxiv_id":null,"evidence_quote":"Gives the method for computing the number of TDL taps from maximum RMS delay spread and time resolution."},{"cited_title":"Measurement-based determination of parameters for non- stationary TDL models with reduced number of taps,","cited_arxiv_id":null,"evidence_quote":"Shows parameter extraction for non-stationary TDL models, including lognormal amplitude fitting and tap-amplitude correlation coefficients."},{"cited_title":"Measurement-based channel characterization and modeling for 5G-Railways at 2.16 GHz,","cited_arxiv_id":null,"evidence_quote":"Reports the same measurement campaign's delay-spread characterization, from which the maximum RMS delay spread (over 350 ns) used in the tap count is taken."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz (release 18),","cited_arxiv_id":null,"evidence_quote":"Defines the standardized TDL channel model used as the baseline whose RMS delay spread PDF is compared in validation."},{"cited_title":"Analyzing non-stationary TDL channel models based on measurement data,","cited_arxiv_id":null,"evidence_quote":"Supports the choice of lognormal amplitudes and uniform phase and Doppler distributions for non-stationary TDL models."}],"review_version":1}