{"id":"5a9bd420-4127-4bbf-a795-ac02e8ec628f","arxiv_id":"2502.00928","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A quantization-theoretic alternating optimization framework tunes antenna tilts, powers, cell partitions, and new-site locations to improve coverage and capacity for ground and UAV users.","lead":"The paper gives a formula-based method, built on quantization theory, to tune cell coverage for networks serving both ground users and drones, including where to put new base stations. A generalist might read it to see a tractable alternative to machine-learning approaches for 3D cellular planning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The γ2 partition update can converge to a non-max-RSS partition, making the capacity-per-region KPI in Table II an artifact of an association users would not follow.","rationale":"The reader's weakest assumption correctly targets Section IV-A, Lemma 2 and Algorithm 3: the capacity-per-region objective does not make max-RSS partitioning provably optimal, and the greedy rule's convergence proof is omitted. My read agrees with that localization but sharpens the issue: the problem is not merely that a heuristic may converge to a suboptimal local optimum; it is that the algorithm can converge to a non-max-RSS partition, and such a partition is not a legitimate operating point under the standard association rule used elsewhere in the paper. All gradients for Θ, ρ, site locations, and bearings assume V is fixed, so the algorithm can terminate with a V that is inconsistent with the actual RSS-maximizing user association. The reported KPI is then evaluated under a partition users would not follow, which means the Table II numbers do not necessarily represent real network capacity per region. This does not invalidate the coverage-capacity KPI γ1 results, where Lemma 1/P Proposition 1 establish that max-RSS is the optimal and realistic partition. For γ2, however, the 'additional BSs consistently outperform antenna-only tuning' claim is not established until the final configuration is evaluated under the max-RSS partition. The proposed test directly settles this by recomputing the KPI at convergence under the correct association. I therefore keep the reader's CONDITIONAL verdict unchanged: the mathematical framework and γ1 results are plausible, but the γ2 claims require this re-evaluation before acceptance.","tokens_in":20462,"tokens_out":8721,"duration_ms":93690,"concrete_test":"After running Algorithms 3 and 4 to convergence for both GUE distributions, compute the max-RSS partition V_RSS from (18) at the final Θ, ρ, bP, and bΦ. Then recompute KPI (16) using V_RSS instead of the algorithm's final V, with no further optimization. Report (i) whether the final V equals V_RSS; (ii) the relative change in KPI; (iii) whether Algorithm 4 still outperforms Algorithm 3. If the KPI drops materially or the ordering flips, the γ2 results in Table II cannot support the central claim and the Section IV-A greedy partition rule needs to be replaced or re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 2 (Section IV-A) proves that max-RSS partitioning maximizes only the auxiliary KPI (34), which omits the denominator in (16). For the actual KPI γ2, the denominator couples each cell's contribution to its own mass, so the max-RSS partition is not provably optimal every iteration. Algorithms 3 and 4 therefore apply the max-RSS assignment only 'if and only if it improves the KPI.' This is not a true maximization over V, and the omitted convergence proof (Proposition 10, stated 'similar to Proposition 4 and omitted') does not address optimality. Since all gradient updates in (37), (38), (40), and (41) then hold V fixed, the algorithm may stop at a fixed point whose partition is not the max-RSS partition at the converged parameters. At such a point, users on cell boundaries would associate with a different BS than V specifies, so the reported P_γ2 in (33) and (39) does not describe the network's actual capacity per region under the paper's own association model. This directly undermines the 'consistently outperforms' claim for KPI #2: the gains in Table II (178.89 to 184.02 and 176.33 to 190.22) may be computed under an unrealizable partitioning rather than under real max-RSS association.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantization-theoretic framework for jointly optimizing cell partitions, vertical antenna tilts, transmit powers, and, for newly deployed base stations, site locations and horizontal bearings. The framework is applied to two KPIs: a coverage-capacity trade-off (γ1) and a capacity-per-region metric (γ2), for a 3D user population comprising ground users and UAV corridors. The authors present alternating optimization algorithms (Algorithms 1–4), prove the optimality of max-RSS partitioning for γ1, derive gradient expressions for the remaining variables, and report case studies showing that deployment optimization outperforms antenna-only tuning and that joint GUE/UAV optimization improves UAV performance with limited GUE loss.","tokens_in":20775,"tokens_out":16101,"duration_ms":166528,"significance":"If the technical claims are correct, the framework is a useful extension of the authors' earlier work [45] to the optimization of new-site locations and bearings, and it addresses a practically relevant 3D deployment problem with GUEs and UAV corridors. The proof of Lemma 1 and the boundary-cancellation argument for γ1 are standard and sound, and the gradient formulas appear dimensionally consistent. However, the paper currently has two substantial weaknesses: several propositions that are load-bearing for the deployment algorithms are stated without proofs, and the capacity-per-region KPI uses a heuristic partition update whose relationship to actual user association is not established. The paper would be considerably stronger with complete proofs for the location/bearing gradients and a clarification or fix of the γ2 partition update; as it stands, the central numerical claims for KPI #2 are not fully supported.","major_comments":[{"comment":"The max-RSS partition update is proved optimal only for the auxiliary objective in (34), which omits the load denominator in (16). For the actual KPI (33), the denominator couples each cell's contribution to its own mass, so the max-RSS partition need not be optimal at every iteration. Algorithms 3 and 4 therefore apply the max-RSS assignment only 'if and only if it improves the KPI,' which is a heuristic. The omitted convergence proof (Proposition 10) does not address optimality, and a fixed point of this rule can have a partition V that is not max-RSS at the converged parameters. If user association in the network follows max-RSS, as assumed for KPI #1 and in the SINR/rate CDFs, then the Pγ2 values reported in Table II (e.g., 176.33 to 190.22) are not the capacity-per-region realized under the paper's own association model. Please either evaluate the final KPI under the max-RSS partition at the converged parameters, explicitly model and justify a controllable (biased) association rule, or state clearly that the KPI #2 results are for a heuristic partition and remove the stronger claims.","section":"Section IV-A, Lemma 2; Algorithms 3-4; Table II"},{"comment":"The gradient formulas for site locations and bearings (Propositions 5, 6, 11, 12) are stated without proofs, and convergence of Algorithms 2, 3, and 4 is asserted with 'proof similar to Proposition 4' or simply omitted. These are not routine corollaries: the derivatives through SINR with respect to co-located site positions and bearings involve the three-sector coupling in (27)-(32), and Proposition 10 cannot follow from Proposition 4 because Lemma 2 does not establish optimality of the partition update for the true KPI #2. Since all deployment results in Tables I-II and Figures 2-5 rely on these algorithms, the authors should provide complete derivations and convergence arguments (e.g., in a supplementary appendix), or explicitly label the site-location/bearing updates and heuristic partition updates as heuristics supported only by the numerical study.","section":"Sections III-B and IV-B, Propositions 5-7, 10-12"}],"minor_comments":[{"comment":"The caption of Table II says 'Coverage-capacity performance comparison' but the section and KPI are for capacity per region; please correct the caption to match the content.","section":"Table II caption"},{"comment":"The algorithms do not specify step-size rules, convergence thresholds, initialization strategy, or the number of Monte Carlo samples used to evaluate the integrals in (21), (23), (26), (30), (37), (38), (40), and (41). Without these details, the reported four-decimal-place values in Tables I and II and the CDFs in Figures 2 and 5 cannot be reproduced.","section":"Algorithms 1-4 and Section V"},{"comment":"The sigmoid steepness κ in (14) is not given in the case-study setup in Section V-A; please specify its value together with T, β, and o_n.","section":"Section II-D, Eq. (14)"},{"comment":"Because the KPI γ2 in (16) depends on the partition V through the denominator, the performance function in (1) is not of the stated separable form with an integrand independent of V. Please generalize the framework definition or clarify that (1)-(2) apply only to KPIs whose integrand does not depend on V.","section":"Section II-A, Eq. (1) and Section II-D, Eq. (16)"},{"comment":"The statement that optimizing deployments 'consistently outperforms' antenna-only optimization is a monotonicity consequence of optimizing over a strictly larger feasible set, assuming the algorithms reach good optima. The numerical comparison is a useful sanity check on the local algorithms, but it should be presented as a validation of the algorithms rather than as an empirical discovery.","section":"Abstract and Section I-B"},{"comment":"There are several small typos: 'gradient ascend' appears in Algorithms 1, 3, and 4; 'UA Vs' appears with irregular spacing in the abstract; and Proposition 7 states that its proof is similar to Proposition 4 but no proof or reference to an appendix is provided.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the authors' own [45]; the main new technical content is the joint optimization of new-site locations and bearings. The most serious risk is the capacity-per-region partition heuristic, which currently makes the headline numbers in Table II not obviously comparable to realizable user association. If the authors can fix that issue and provide the missing proofs, the paper could become acceptable. The fit with cs.IT is appropriate, but the present version is not rigorous enough for publication as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this paper extends Karimi-Bidhendi et al.'s earlier quantization-theoretic work for UAV corridor optimization to jointly optimize new-site locations, bearings, tilts, and powers for mixed ground/UAV users. The coverage-capacity (γ1) half is largely sound. The capacity-per-region (γ2) half has a load-bearing gap: the max-RSS partition is not optimal for that KPI, and the greedy 'apply only if it improves' rule is not proven to converge to anything consistent with the association model. That undermines the γ2 performance claims in Table II.\n\nWhat's new: the joint optimization over new-site locations and bearings (Application #2) is new relative to [45], and the capacity-per-region KPI is a useful addition. The gradient expressions for tilts and powers are imported from [45] (properly cited), and the boundary-cancellation argument in Lemma 1 is standard and correct. The case study with 3GPP channel models, two GUE distributions, and UAV corridors is realistic and shows the framework's flexibility.\n\nSoft spots: several propositions (5, 6, 11, 12) give gradients without proofs. These look like routine chain-rule calculations from the RSS model, so they're likely correct, but in a paper whose selling point is mathematical rigor, 'omitted because of page limit' is unsatisfactory. Convergence of Algorithms 2-4 is likewise asserted without proof. More seriously, the γ2 partition update: for KPI #2 the objective divides by each cell's mass, so max-RSS does not maximize the true objective. Lemma 2 only proves max-RSS maximizes the auxiliary KPI without the denominator. The algorithm's 'if and only if it improves the KPI' step guarantees a non-decreasing sequence, but it does not ensure the final partition is the max-RSS association. At a fixed point, the reported KPI may correspond to a partition that users would not follow, which makes the Table II gains unreliable. The 'deployment beats antenna-only' result is a monotonicity consequence of the larger feasible set; it's true but framed as more than it is.\n\nWho's it for: researchers working on cell deployment or UAV corridor optimization will find the γ1 framework and the location/bearing gradient formulas useful. It deserves a serious referee, but the γ2 section needs either a genuine optimality proof for the partition update or a clear statement that the reported KPI uses the optimized partition rather than the association users would follow. I'd recommend a major revision.","headline":"A useful extension of the authors' prior quantization framework to site placement and bearings; the coverage-capacity half holds up, but the capacity-per-region KPI rests on a partition heuristic that is not proven and may not match real user association.","tokens_in":21294,"tokens_out":3313,"would_cite":true,"duration_ms":30317,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C30","90C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that jointly optimizing where new base stations sit and how their antennas point, using the alternating-update ideas of quantization theory, consistently outperforms adjusting only the tilts and powers of existing base…","keywords":["cell deployment optimization","quantization theory","UAV corridors","antenna tilt optimization","coverage-capacity trade-off","capacity per region","Voronoi tessellation","alternating optimization"],"falsifier":"Run Algorithm 4 on a small two-site network with a fixed user density and compare its converged capacity-per-region value against the optimum found by exhaustive enumeration of all cell partitions; a material gap would show that the greedy partition rule can stall at a poor local optimum and that the reported gains are not guaranteed.","tokens_in":20267,"feed_emoji":"📡","tokens_out":5025,"duration_ms":50358,"temperature":0.7,"pith_summary":"This paper tries to establish that cell deployment—where new base stations sit, which way their antennas point, and how loudly they transmit—can be optimized as a single mathematical problem by borrowing the alternating-update machinery of quantization theory. It claims that jointly optimizing the locations and bearings of additional base stations beats adjusting only the tilts and powers of existing ones, no matter how ground users are spread. It further claims that including drone corridors in the objective greatly improves aerial service while costing ground users very little. If these claims hold, operators could replace trial-and-error radio planning with a deterministic algorithm that works for arbitrary 3D user densities.","feed_headline":"New base stations beat antenna retuning for mixed drone and ground coverage","feed_subtitle":"A quantization-based algorithm jointly places and steers cells, lifting UAV SINR while barely hurting ground users.","key_machinery":"The load-bearing object is the generalized Voronoi tessellation: each user location $q$ is assigned to the base station giving the highest received signal strength, which the paper proves is the optimal cell partition whenever the KPI is a continuous increasing function of SINR. Around this partition, the algorithm alternates: recompute the max-RSS cells; run gradient ascent on vertical tilts $\\theta_n$ and powers $\\rho_n$; and, for new sites, run gradient ascent on site locations $\\hat{p}_m$ and reference bearings $\\hat{\\phi}_m$; then repeat. This is the alternating partition-and-update scheme of quantization theory transplanted to radio planning, and it is what turns an NP-hard joint problem into a sequence of tractable improvement steps.","core_discovery":"The paper's central claim is that a single performance function $P(\\mathcal{V},\\ldots)$, defined as the integral over each cell of a chosen key performance indicator $\\gamma^{(n)}(q)$ weighted by the user density $\\lambda(q)$, captures coverage, capacity, and load balancing, and that maximizing it by alternating between a generalized Voronoi partition and gradient ascent over all antenna and site parameters yields locally optimal configurations for realistic 3D networks. For the coverage-capacity KPI, the optimal partition is provably the strongest-signal partition, because both the log-rate and coverage terms are continuous increasing functions of SINR. For the capacity-per-region KPI, the paper adopts a greedy rule that applies the strongest-signal partition only when it improves the metric. The case studies report that deploying new base stations with optimized locations and bearings improves both KPIs over antenna-only tuning, and that optimizing for $r=0.5$, equal weight to ground users and UAVs, produces large UAV gains with small ground-user losses.","pith_inferences":["The same alternating structure could extend to other KPIs beyond log-rate and coverage indicator, as long as the per-cell contribution is a monotone function of SINR, in which case the max-RSS partition remains provably optimal.","The capacity-per-region greedy partition rule could be stress-tested against a small-network exhaustive search; if it stalls or converges to poor local optima, a swap-based partition update would be a natural remedy.","Real deployments with terrain shadowing would require substituting the idealized pathloss model with site-specific maps, which the integral formulation can accommodate but the case studies do not demonstrate.","The reported advantage of deploying new sites may diminish as the number of existing sites grows, since additional degrees of freedom have decreasing marginal value; the paper does not plot this scaling behavior."],"forward_implications":["Operators can tune tilts, powers, and new-site positions in one iterative procedure rather than through trial-and-error field measurements or separate optimization tools.","Deploying additional base stations with optimized locations and bearings yields larger KPI gains than exclusively retuning existing antennas, for both uniform and clustered ground-user distributions.","Including UAV corridors in the objective, for example with $r=0.5$, substantially raises UAV SINR and rate while producing only a small drop in ground-user performance.","The framework accepts any deterministic 3D user density $\\lambda(q)$, so it can model nonuniform ground hotspots and aerial corridors within the same optimization problem.","For the coverage-capacity KPI, the optimal cell partition is explicitly characterized by the maximum-RSS rule, giving a clear and simple association policy."],"supporting_citations":[{"why":"Supplies the quantization theory background and the alternating optimization perspective that the proposed framework adapts.","marker":"[21]"},{"why":"Establishes the RSS-versus-SINR equivalence and the gradient formulas that Propositions 1 through 3 build upon.","marker":"[45]"},{"why":"Provides the classic alternating partition-and-parameter update algorithm that inspires Algorithms 1 through 4.","marker":"[46]"},{"why":"Defines the standardized aerial-vehicle channel and deployment parameters used in the case study.","marker":"[6]"},{"why":"Supplies the directional antenna gain, pathloss, and channel model constants used in the simulations.","marker":"[7]"},{"why":"Justifies dropping the boundary integral terms in the gradient derivations for antenna tilts and transmission powers.","marker":"[30]"}],"fun_headline_variants":["Quantization-based cell placement lifts UAV coverage without hurting ground users","New base stations beat antenna-only tuning for mixed drone and ground coverage","Joint cell placement and steering improves UAV service with minimal ground loss","Optimized BS deployment outperforms antenna retuning in 3D heterogeneous networks","Quantization framework optimizes BS sites and antennas for UAV and ground users"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the capacity-per-region metric, the paper asserts without a full proof that the greedy strongest-signal partition update, applied only when it improves the metric, converges to a good solution, and this unproved convergence claim is the load-bearing step behind the capacity-per-region gains.","fun_headline_variants_meta":{"raw":{"variants":["Quantization-based cell placement lifts UAV coverage without hurting ground users","New base stations beat antenna-only tuning for mixed drone and ground coverage","Joint cell placement and steering improves UAV service with minimal ground loss","Optimized BS deployment outperforms antenna retuning in 3D heterogeneous networks","Quantization framework optimizes BS sites and antennas for UAV and ground users"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3176,"prompt_tokens":930,"completion_tokens":2246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":2154}},"tokens_in":546,"tokens_out":2246,"duration_ms":14615,"temperature":1.0,"reasoning_tokens":2154,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:12:19.884518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 4 on a small two-site network with a fixed user density and compare its converged capacity-per-region value against the optimum found by exhaustive enumeration of all cell partitions; a material gap would show that the greedy partition rule can stall at a poor local optimum and that the reported gains are not guaranteed.","supporting_citations":[{"cited_title":"Quantization,","cited_arxiv_id":null,"evidence_quote":"Supplies the quantization theory background and the alternating optimization perspective that the proposed framework adapts."},{"cited_title":"Optimizing cellular networks for UA V corridors via quantization theory,","cited_arxiv_id":null,"evidence_quote":"Establishes the RSS-versus-SINR equivalence and the gradient formulas that Propositions 1 through 3 build upon."},{"cited_title":"Study on enhanced LTE support for aerial vehicles (Release 15),","cited_arxiv_id":null,"evidence_quote":"Defines the standardized aerial-vehicle channel and deployment parameters used in the case study."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz (Release 16),","cited_arxiv_id":null,"evidence_quote":"Supplies the directional antenna gain, pathloss, and channel model constants used in the simulations."},{"cited_title":"Sensor deployment with limited communication range in homogeneous and heterogeneous wireless sensor networks,","cited_arxiv_id":null,"evidence_quote":"Justifies dropping the boundary integral terms in the gradient derivations for antenna tilts and transmission powers."}],"review_version":1}