{"id":"83cddf5d-f7e9-4947-9db2-33d1aa3ac207","arxiv_id":"2505.01687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-phase absorption and adaptation framework that estimates the distribution of CSI errors via deconvolution and uses it for power allocation improves QoS recovery in C-V2X networks under unknown imperfect CSI.","lead":"This paper proposes a two-phase framework that lets a cellular vehicle-to-everything (C-V2X) network absorb and adapt to inaccurate channel state information by estimating the error distribution on the fly. The framework then reallocates transmit power using the estimated distribution, and simulations show it reduces severe vehicle-to-vehicle delay and improves vehicle-to-infrastructure throughput relative to model-based and data-driven benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulation fixes all transmit-power bounds to the same 10 dBm value, so the proposed power-allocation algorithms are degenerate and the reported gains cannot be attributed to the claimed power design.","rationale":"The reader's deconvolution model-mismatch concern is real but concerns the robustness of modeling assumptions that the paper explicitly commits to; the additive-error and Gauss-Markov structure is stated in Section II-B. The power-bound degeneracy is more load-bearing because it undermines the only empirical support for the headline claim regardless of model assumptions. If the power bounds are taken literally, the proposed power-allocation algorithms have no role in the reported gains. In addition, the abstract's 'without compromising the network's QoS in the absorption phase' is contradicted by the paper's own Fig. 7a, which shows roughly 70% delay-QoS satisfaction for lambdaV = 0.5 against the required P0 = 95%, a point the text acknowledges as a deliberate tradeoff. These issues do not necessarily invalidate the conceptual framework, but they require correction before the quantitative claims can be accepted. The proposed concrete test would settle whether the reported gains are attributable to the power-allocation design or to matching and PDF estimation alone.","tokens_in":23894,"tokens_out":10102,"duration_ms":98756,"concrete_test":"Rerun the Section VI experiments with nontrivial power intervals, e.g., pV_min = 10 dBm, pV_max = 23 dBm, pI_min = 10 dBm, pI_max = 23 dBm, and log the optimized p* values from Algorithm 2. Verify that p* actually varies across time slots and that the 35%/56% V2V delay reduction and 14%/16% V2I throughput improvements persist; if the optimized powers never leave 10 dBm or the gains disappear, the current evidence is an artifact of the degenerate power grid.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section VI sets pV_min = pV_max = pI_min = pI_max = 10 dBm. Under these bounds, every power feasible region collapses to a single point: in the absorption phase, the closed-form solution (19) returns the same (pI_n,a, pV_m,a) regardless of lambda; in the adaptation phase, the interval (31d) shrinks to a single ct value, so the bisection and one-dimensional search in Algorithm 2 cannot change any transmit power. The framework's central contribution is hazard-rate-constrained power design and real-time power-allocation optimization; with degenerate power bounds, all QoS differences between the proposed design and the benchmarks can only come from the matching and the estimated PDF, not from any power decision. The headline numbers in the abstract and Fig. 9 therefore do not demonstrate the claimed power-allocation advantage. This is not a modeling assumption; it is a missing degree of freedom in the only reported evaluation of the central claim. The authors should confirm whether this is a typographical error; as written, the simulations cannot support the quantitative conclusions.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-phase resilience framework for C-V2X networks operating under imperfect CSI with an unknown error distribution. In the absorption phase, the RSU estimates the PDF of the CSI error via a deconvolution-based estimator built from RSS measurements, and jointly optimizes RB matching and absorption powers under hazard-rate constraints. In the adaptation phase, the estimated PDF is used in a real-time power allocation problem that minimizes the MSE between estimated and true QoS satisfaction probabilities. Simulation results report 35% and 56% reductions in conditional V2V delay and 14% and 16% improvements in average V2I throughput over model-based and data-driven benchmarks, respectively.","tokens_in":24128,"tokens_out":10210,"duration_ms":91262,"significance":"If the results hold, the framework is a meaningful contribution to resilient V2X resource management: it combines a statistically grounded deconvolution estimator with a hazard-rate-based QoS metric, and it explicitly quantifies the absorption-adaptation tradeoff. The paper includes analytical upper bounds, closed-form absorption power solutions, and a low-complexity adaptation algorithm, and the simulation study covers 100,000 channel realizations. However, the experimental support for the central claim is weakened by the degenerate power bounds in Section VI, and several proof steps are outsourced or rely on unquantified truncation assumptions.","major_comments":[{"comment":"The simulation fixes pV_min = pV_max = pI_min = pI_max = 10 dBm. With these bounds, the absorption-phase solution in (19) and the adaptation-phase feasible interval in (31d) collapse to single points, so Algorithms 1 and 2 have no power-allocation degree of freedom. Consequently, the reported 35% and 56% conditional-delay reductions and 14% and 16% throughput gains in Figs. 8 and 9 and in the abstract cannot be attributed to the proposed power optimization; at most they can reflect matching and PDF estimation. The authors should either correct what appears to be a typographical error and rerun the simulations with non-degenerate transmit-power ranges, or explicitly reframe the paper's claims around matching and estimation rather than power allocation.","section":"Section VI, first paragraph"},{"comment":"The proof of Theorem 1 is entirely delegated to the conference version [1], with no argument in the present manuscript. Since Theorem 1 provides the MSE upper bound that defines the absorption-phase objective and the edge weights in (16)-(18), the journal manuscript is not self-contained and the correctness of the absorption-phase optimization cannot be verified from the submitted text. Please include the proof or a complete proof sketch.","section":"Section IV-B, Theorem 1"},{"comment":"The model assumes perfect knowledge of large-scale fading, perfect CSI at the RSU for the V2I uplink and V2V-to-V2I interference channels, and a Gauss-Markov/Jakes model with known coefficient delta_m,t for the V2V direct link; only the additive error e_nm has an unknown distribution. If any of these premises fails, the quantity in (9) mixes several unknown error sources and the deconvolution estimate is not the PDF needed for adaptation. The Abstract and Section I claim that the framework handles 'arbitrary unknown imperfect CSI,' which overstates the scope; the claims should be qualified to the unknown distribution of a single additive error component under the stated structural assumptions.","section":"Section II-B and Eq. (9)"},{"comment":"The claim that the adaptation-phase gains are achieved 'without compromising the network's QoS in the absorption phase' is contradicted by Fig. 7a, where for lambda_V = 0.5 the probability of meeting the V2V delay requirement is about 70%, far below the target P0 = 95%, and about 1% lower than the Gaussian benchmark. The text acknowledges this as a deliberate tradeoff, so the abstract and conclusion should describe the absorption-phase QoS sacrifice accurately.","section":"Abstract and Section VI, Fig. 7a"},{"comment":"The monotonicity of u(c_t) is established only under the unquantified condition (30), with the argument that K2 can be chosen 'large enough.' The simulations use K2 = 10, and the condition depends on lambda_Y, which is determined by the absorption powers; the paper does not verify (30) for the simulated parameters. The monotonicity of beta(c_t) in Corollary 2 is stated without proof. Because Algorithm 2 relies on both monotonicity properties for its bisection and one-dimensional search, the theoretical guarantee is incomplete in the simulated regime. In addition, the proof of Theorem 2 in Appendix B uses the truncation constant K in (43)-(46) while the statement uses K2; this inconsistency should be corrected.","section":"Appendix B and C, Propositions 1 and Corollary 2"}],"minor_comments":[{"comment":"The term 'probability distribution function' (Abstract, Sections I and IV-B) should be 'probability density function' for f_E,m.","section":"Throughout"},{"comment":"The expression for b_m,t appears to contain an index error: the term involving p^V_{n,t} L^V_{nm}|g^V_{nm}|^2 should likely be the V2I interference term p^I_{n,t} L^I_{nm}|g^I_{nm}|^2. Please verify and correct.","section":"Eq. (5)"},{"comment":"'Z is essentially a sequence of i.i.d. samples' should refer to the random variable Z = e_nm + Y; the sequence is {z_{m,k}}.","section":"Eq. (9) and following text"},{"comment":"The notation ∫_{w≥|Kπ|} should be written as ∫_{|w|≥Kπ} for clarity.","section":"Eqs. (12) and (26)"},{"comment":"The sentence 'The large-scale fading is assumed to vary every 1,000 time slots, i.e., T = 1,000' should be rephrased: T denotes the absorption-window length, and the large-scale fading is constant over this window.","section":"Section VI"},{"comment":"Table I gives the path-loss model for h^I_n but not the explicit formulas for the WINNER+B1 models used for h^V_m, h^I_nm, and h^V_mn; please provide the complete expressions.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an extended version of the authors' conference paper, and the missing proof of Theorem 1, the degenerate power bounds in the simulations, and the notation inconsistencies in Appendices B-C suggest the paper needs substantial revision before it can be considered further. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this one. The absorption/adaptation framing is genuinely new in the C-V2X imperfect-CSI literature, and the deconvolution-based PDF estimation is a sensible way to avoid assuming a known error model. But the simulations, as written, cannot support the power-allocation claims: every transmit-power bound is fixed at 10 dBm, so the power optimization is degenerate and the reported gains are not attributable to the proposed power design.\n\nThe idea is real: prior work either assumes a Gaussian/bounded error or builds conservative uncertainty sets from samples, and none of it treats the learning phase as a QoS-relevant period. This paper defines an absorption phase with a dedicated power scheme and RB matching, estimates the unknown error PDF from RSS measurements, then uses that estimate in adaptation. The hazard-rate constraint is a reasonable way to express a preference for small violations over large ones. The deconvolution machinery is standard statistics, cited fairly, and the upper-bound analysis yields a testable tradeoff between absorption QoS and estimation accuracy. Theorem 2 and the appendices are mostly standard manipulations; nothing in the math is outrageously wrong as far as I can see.\n\nSoft spots, in rough order of severity. First, the degenerate simulation setup: Section VI sets pV_min=pV_max=pI_min=pI_max=10 dBm. In absorption, the closed form (19) always returns (10,10); in adaptation, the feasible interval for ct in (31d) is a single point, so the bisection and one-dimensional search in Algorithm 2 cannot change any transmit power. That means the headline numbers—35% and 56% delay reduction, 14% and 16% throughput gain—can only come from the matching and the estimated PDF, not from power allocation. If the bounds are a typo, the authors need to rerun; if not, the central algorithmic contribution is unevaluated. Second, the abstract says 'without compromising QoS in the absorption phase,' but Fig. 7a for λV=0.5 shows only about 70% of V2V delays meeting the requirement, below the 95% target and below the Gaussian benchmark. The body later calls this a deliberate tradeoff, which is honest, but the headline is wrong. Third, Theorem 1's proof is outsourced to the conference version; a journal paper should prove it. Fourth, the σ2=0 and large-K approximations are unquantified. These are standard moves in this literature, but sensitivity analysis would help. Citation pattern is fine; the self-citation is to the conference paper where the proof lives.\n\nNet: this is a paper for researchers working on robust C-V2X resource allocation. It deserves a serious referee, and I would send it to review, but the revision should re-run the simulations with non-degenerate power bounds and correct the absorption-QoS claim. I would not cite the numerical results until that happens.","headline":"Genuinely new two-phase resilience formulation for C-V2X under unknown CSI error, but the reported power-allocation gains are unsubstantiated because the simulations fix every transmit power to the same value.","tokens_in":24625,"tokens_out":4781,"would_cite":false,"duration_ms":44725,"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 two-phase learning framework lets C-V2X networks recover QoS from unknown CSI-error distributions, cutting excess V2V delay by 35–56% and lifting V2I throughput by 14–16%.","keywords":["C-V2X","imperfect channel state information","resilience","deconvolution estimation","power allocation","hazard rate","quality of service","vehicular communications"],"falsifier":"Take a channel trace in which the large-scale fading on a V2V direct link drifts measurably within one absorption window (for instance, a vehicle accelerating from 10 to 20 m/s over the $T=1000$ slots), feed the recorded RSS samples into the estimator in Eq. (11), and compare the recovered PDF to the true additive-error distribution obtained by subtracting the known channel components; if the empirical MSE violates the upper bound of Theorem 1, or if the adaptation-phase delay CDF no longer separates from the benchmarks, the constant-large-scale assumption is the cause.","tokens_in":23695,"feed_emoji":"🚗","tokens_out":22028,"duration_ms":176194,"temperature":0.7,"pith_summary":"This paper tries to establish that a cellular vehicle-to-everything (C-V2X) network can keep its quality-of-service targets when the statistical distribution of the channel-state-information (CSI) error is completely unknown, by splitting the problem into two deliberately designed phases. In the absorption phase the network spends a short window learning the error distribution from received-signal-strength samples; in the adaptation phase it uses the learned distribution to re-optimize transmit powers in real time. If the claim holds, it replaces the usual dilemma between model-based methods (which assume a distribution and fail when it is wrong) and data-driven methods (which ignore the transient QoS loss during learning) with a single framework in which the transient loss is controlled by a hazard-rate constraint and pays for a quantifiable improvement in eventual recovery.","feed_headline":"Two-phase learning cuts excess vehicle-to-vehicle delay by 56%","feed_subtitle":"The scheme learns the unknown channel-error distribution, then lifts vehicle-to-infrastructure throughput by 14–16%.","key_machinery":"The engine that carries the argument is the deconvolution estimator built on Eq. (9). Because the RSU knows the large-scale fading of the involved links and the Gauss-Markov coefficient $\\delta_{m,t}$ of the direct V2V channel, the normalized difference between measured and nominal received signal strength on a V2V receiver is a sum of the unknown additive error $e_{nm,t}$ and an independent exponential variable $Y$ with rate $\\lambda_Y = p^I_{n,a} L^I_{nm,a}/(p^V_{m,a} L^V_{m,a}(1-\\delta_m^2))$. Since the PDF of $Y$ is known, the target PDF is obtained in the Fourier domain as $\\mathcal{F}\\{f_{E,m}\\} = \\mathcal{F}\\{f_Z\\} / \\mathcal{F}\\{f_Y\\}$, with the numerator approximated empirically from the collected samples. Theorem 1 bounds the mean-square error of this estimator, and that bound becomes the absorption-phase objective, while the hazard rate of Eq. (13) — the conditional probability that an already-violated delay stays close to its requirement — constrains how much the learning phase may degrade QoS. In the adaptation phase, the estimated PDF is inserted into a Parseval-theorem evaluation of the delay-satisfaction probability, after which power control reduces to a one-dimensional search over $c_t = \\gamma_V p^I_{n,t} L^I_{nm,t}/(p^V_{m,t} L^V_{m,t}(1-\\delta_m^2))$.","core_discovery":"The central claim is that the unknown probability density of the additive error corrupting V2V-to-V2I interference links can be recovered from ordinary received-signal-strength measurements through the deconvolution identity $Z = e_{nm} + Y$, where the independent term $Y$ is exponential with a rate set by the ratio of the absorption-phase powers. With that density in hand, the probability that a V2V link meets its delay requirement becomes a computable quantity, and the paper proves an upper bound on the mean-square error of this computed probability (Theorem 2), then minimizes it by a one-dimensional power search. In the simulated Manhattan-mobility scenario, the resulting design reduces conditional V2V delay (delay values exceeding the requirement) by 35% against a Gaussian-error model-based design and by 56% against a high-probability-region data-driven design, while improving average V2I throughput by 14% and 16%.","pith_inferences":["The paper fixes the hazard-rate weights $\\lambda_m$ before absorption; an untested extension is to adapt them across links or time, e.g., raising $\\lambda_m$ for safety-critical V2V links and lowering it for throughput-oriented ones, so the network spends estimation effort where QoS is hardest to recover.","The deconvolution kernel requires exactly one known contamination distribution $Y$; the same estimator could be reused in other settings where a measurement is a sum of an unknown distribution and a known exponential one, such as residual co-channel interference estimation after subtracting a known serving-signal component.","Because the RB matching is decided once during absorption, the learned error PDF is not yet used to re-pair V2V and V2I links during adaptation; periodically re-solving the bipartite matching with the estimated PDF could yield further throughput gains beyond the reported 14–16%.","The claim that adaptation is achieved 'without compromising' absorption QoS holds under the hazard-rate interpretation, not under a strict satisfaction-probability guarantee: the paper's own Fig. 7a shows about 70% V2V delay satisfaction during absorption at $\\lambda_V=0.5$, below the 95% target, so an operator who treats $P_0$ as an absolute constraint must re-tune $\\lambda$."],"forward_implications":["A roadside unit can meet QoS requirements without any prior statistical model of the CSI error: the only inputs are received-signal-strength feedback, large-scale fading estimates, and the known Jakes coefficient.","The hazard-rate constraint gives the operator a tunable knob $\\lambda_m$: raising it protects absorption-phase V2V delay at the cost of a less accurate error PDF, so link-criticality priorities can be encoded directly into the optimization.","The Theorem 2 bound means the quality of adaptation depends jointly on the accuracy of the learned error PDF and on the instantaneous CSI quality, so adding absorption samples or shortening the CSI feedback delay both improve the final QoS.","In the simulated Manhattan-mobility setting, the proposed design reduces the conditional V2V delay (delay above 15 ms) by 35% and 56% relative to the Gaussian-error and high-probability-region benchmarks, while raising average V2I throughput by 14% and 16%."],"supporting_citations":[{"why":"Supplies the proof of Theorem 1's upper bound on the deconvolution estimator's mean-square error.","marker":"[1]"},{"why":"Model-based benchmark assuming a known Gaussian CSI error; the paper's adaptation-phase and absorption-phase results are compared against it.","marker":"[9]"},{"why":"Motivates the first-order Gauss-Markov model for V2V direct links with imperfect CSI, used in Eq. (3).","marker":"[10]"},{"why":"Data-driven benchmark constructing a high-probability region from imperfect CSI samples; the main alternative the proposed design outperforms.","marker":"[12]"},{"why":"Provides the Jakes statistical model and the coefficient $\\delta_{m,t}=J_0(2\\pi f_D\\Delta t)$ used in Eq. (3).","marker":"[16]"},{"why":"Establishes the truncated Fourier-inversion deconvolution estimator used in Eq. (11).","marker":"[21]"},{"why":"Supplies the empirical-characteristic-function approximation used to replace the Fourier transform of $f_Z$ in Eq. (10).","marker":"[22]"},{"why":"Provides the truncated regularization used in Eq. (25) to make the adaptation-phase probability integral convergent.","marker":"[24]"}],"fun_headline_variants":["Learning channel errors cuts V2V delay by 56%","Adaptive scheme cuts V2V delay 56%, boosts V2I throughput 14-16%","Estimating channel-error distribution cuts V2V delay 56%","Learn error stats to reduce excess V2V delay 56%","Two-phase learning trims V2V delay 56%, lifts V2I throughput"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme's load-bearing premise is that the large-scale fading of every congested link is known and constant over the absorption window, the RSU has perfect knowledge of the V2I uplink and V2V-to-V2I interference channels, and the direct V2V link's small-scale fading follows a Gauss-Markov/Jakes model with a known coefficient, so the only unknown left is the additive interference error whose distribution is the deconvolution target.","fun_headline_variants_meta":{"raw":{"variants":["Learning channel errors cuts V2V delay by 56%","Adaptive scheme cuts V2V delay 56%, boosts V2I throughput 14-16%","Estimating channel-error distribution cuts V2V delay 56%","Learn error stats to reduce excess V2V delay 56%","Two-phase learning trims V2V delay 56%, lifts V2I throughput"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001397,"raw_usage":{"total_tokens":5698,"prompt_tokens":1042,"completion_tokens":4656,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":4554}},"tokens_in":658,"tokens_out":4656,"duration_ms":32902,"temperature":1.0,"reasoning_tokens":4554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:13:09.095180+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a channel trace in which the large-scale fading on a V2V direct link drifts measurably within one absorption window (for instance, a vehicle accelerating from 10 to 20 m/s over the $T=1000$ slots), feed the recorded RSS samples into the estimator in Eq. (11), and compare the recovered PDF to the true additive-error distribution obtained by subtracting the known channel components; if the empirical MSE violates the upper bound of Theorem 1, or if the adaptation-phase delay CDF no longer separates from the benchmarks, the constant-large-scale assumption is the cause.","supporting_citations":[{"cited_title":"Robust Resource Allocation for Vehicular Communications With Imperfect CSI,","cited_arxiv_id":null,"evidence_quote":"Model-based benchmark assuming a known Gaussian CSI error; the paper's adaptation-phase and absorption-phase results are compared against it."},{"cited_title":"Deconvolution with unknown error distribution,","cited_arxiv_id":null,"evidence_quote":"Provides the truncated regularization used in Eq. (25) to make the adaptation-phase probability integral convergent."}],"review_version":1}