{"id":"a2faf797-f2a9-42bd-b4fe-4aa26c9a4e5d","arxiv_id":"2607.24283","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Adaptive spatial discretization from Gaussian random-field theory plus greedy/SA selection of measurement sites reduces CGM reconstruction AMSE by roughly 20% versus uniform grids under known mean and covariance.","lead":"This paper shows how to pick where to measure wireless channel gain so a full coverage map can be filled in with lower average error. It combines Kriging prediction with an adaptive grid that puts more sample points in fast-varying regions, cutting error about 20% versus a uniform grid in simulation.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The ~20% AMSE gain is demonstrated only in a self-consistent simulation: the truth is drawn from the exact exponential covariance model that the adaptive allocation and Kriging estimator assume, with oracle second-order statistics.","rationale":"The paper is a clean, internally consistent contribution: the Kriging variance objective (13)–(19) is standard and correct, the spectral truncation analysis (32)–(38) is right (modulo an apparent typo in (36) where (2πL_cρ)² is printed as (2π(L_c)³ρ)² — the subsequent t-substitution only works with the former), and the closed-form allocation (42) follows correctly. The weak point is exactly where the reader located it — oracle mean/covariance — but I would sharpen one consequence the reader only implied: the 20% figure is not merely contingent on knowing the statistics, it is contingent on the truth belonging to the assumed model family, because the synthetic ground truth is generated from the same covariance used for design and estimation. Fig. 10's demonstration that semivariogram family materially changes AMSE makes this sensitivity concrete rather than hypothetical. This supports keeping CONDITIONAL rather than moving to REJECT (the framework and theory are sound and useful; nothing suggests the gain is an artifact of an error) and precludes ACCEPT (the practical headline number has no demonstrated robustness to covariance estimation error or model mismatch, and no code or field data are provided). Confidence in this reading is high; the proposed test is cheap since it reuses the authors' existing simulation pipeline with a swapped generative covariance and estimated parameters.","tokens_in":23003,"tokens_out":4065,"duration_ms":130771,"concrete_test":"Re-run the Fig. 12 headline configuration (U=16000, N=400, SA and greedy) with two modifications: (i) generate the residual field from a mismatched covariance (Matérn ν=3/2, or empirical residuals from the Wireless Insite map after detrending) while the design and Kriging still assume the exponential model; (ii) replace oracle σ_r², L_c^r with parameters estimated from a pilot sample of N_pilot=50–100 randomly placed measurements plus the running measurement set. If the adaptive-vs-uniform AMSE gap collapses well below 20% or reverses, the headline gain is in-model; if it holds above ~10%, the claim is robust to estimation and moderate mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — adaptive discretization beats uniform by ~20% AMSE at U=16000, N=400 (Fig. 12) — rests on an evaluation that is matched by construction to the design model. §VII states the residual S(x) is \"synthesized via Cholesky decomposition of the covariance matrix analytically constructed based on (30)\", i.e., the same exponential covariance whose parameters (σ_r², L_c^r) drive both the adaptive allocation rule in (42) and the Kriging variance objective (19). Under this setup the Kriging variance equals the true MSE exactly, the subregion partition exactly matches the generative heterogeneity, and the only thing being tested is whether better grid allocation helps when the model is perfectly known and perfectly true. That is the easiest possible regime for the method.\n\nThis matters quantitatively, not just philosophically, because the paper's own Fig. 10 shows the semivariogram family (exponential vs Gaussian vs spherical) changes both the optimal measurement pattern and the AMSE ranking across N by margins comparable to the headline 20%. So the method is demonstrably sensitive to covariance structure, yet the headline number is evaluated at zero model mismatch. Compounding this, §I motivates the work by noting real CGMs have abrupt LoS/NLoS gain transitions that \"undermine the smoothness of the channel gain function\" — precisely the structure a stationary exponential random field cannot represent; the K-means partition into R=10 subregions is an unvalidated patch for this. And as the reader flagged, the allocation rule (42) needs σ_r² and L_c^r per subregion a priori; in a real campaign these must be estimated from sparse data, and misspecified L_c^r misallocates grids through the (σ_r²/L_c^r)^{3/4} weighting. The internal math (Eqs. 13–19, 32–42) checks out — I verified the truncation integral (37), its asymptotic form (38), and the Lagrangian solution (42) — so the concern is not correctness of the derivation but that the demonstrated gain is in-model, ","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper studies where to place a budget of N channel measurements for data-based construction of a channel gain map (CGM). Under the explicit assumptions that the deterministic mean µ(x) and the full covariance C(x,x′) are known, the authors use the closed-form ordinary-Kriging variance (Eq. 19) as a surrogate for MSE and formulate measurement-location selection as a combinatorial problem (P1), solved by greedy search and simulated annealing with complexity analyses. Small-geometry stationary-point analyses (§IV) give intuition for how optimal patterns depend on the correlation range relative to the geometry. The main methodological contribution is an adaptive discretization strategy (§VI): the space is partitioned into R subregions via K-means on a per-subregion path-loss law, and grid budget U is allocated by a Lagrangian minimizing mean-squared spectral truncation loss of the 3D exponential-covariance field, yielding the closed-form rule n_r ∝ V_r(σ_r²/L_c^r)^{3/4} (Eq. 42). Simulations (§VII) show adaptive discretization beating uniform by \"20%\" in AMSE at U=16000, N=400 (Fig. 12), and explore semivariogram-family sensitivity (Fig. 10).","tokens_in":23364,"tokens_out":5450,"duration_ms":110820,"significance":"If the results hold, the paper gives the CKM/radio-map community a tractable, interpretable two-level framework (discretization allocation + subset selection) with closed-form ingredients: the Kriging-variance objective, the 3D PSD of the exponential covariance, and the allocation rule (42), all of which are cheap to compute and easy to reproduce. The semivariogram-sensitivity study (§VII-B) and the per-regime analysis in §IV are genuinely instructive. The derivations I checked (Eqs. 15–19, 21, 27, 32, 36–42) are internally consistent, and the oracle assumptions are stated explicitly rather than hidden — a credit to the manuscript. However, the empirical evidence is entirely self-consistent simulation: the residual field is synthesized via Cholesky decomposition of the same covariance (30) that drives both the allocation rule and the Kriging objective, so the headline 20% gain is demonstrated only at zero model mismatch, in the regime most favorable to the method. No code or data release is mentioned, which limits reproducibility of Figs. 10–12.","major_comments":[{"comment":"The headline claim (Abstract; Fig. 12: ~20% AMSE gain at U=16000, N=400) is evaluated with S(x) synthesized from exactly the exponential covariance (30) whose parameters drive both the allocation rule (42) and the Kriging objective (19). Under this setup Kriging variance equals true MSE by construction, so the simulation only verifies optimization efficacy, not robustness. This is load-bearing because the paper's own Fig. 10 shows the semivariogram family changes both the optimal pattern and the AMSE ranking by margins comparable to the headline 20%. A concrete fix: report the adaptive-vs-uniform gap (i) when L_c^r and σ_r² used in (42) are perturbed or estimated from a small pilot set, and (ii) when the truth is generated under a mismatched covariance (Gaussian/spherical) or from ray-tracing-driven residuals. If the gain survives plausible mismatch, the claim is much stronger.","section":"§VII-C, Fig. 12"},{"comment":"The '20%' gain appears to be computed on the dB-valued AMSE axis of Fig. 12 (values ~40–50 dB; the annotated 20.3% matches a ratio of dB numbers). Percentage differences of logarithmic quantities are not meaningful; on a linear MSE scale the corresponding ratio is roughly an order of magnitude. The Abstract's 'twenty-percent performance gain' should be recomputed and stated on the linear AMSE scale, and the axis convention ('AMSE/dB2' in Figs. 10–11 vs 'AMSE' in Fig. 12) clarified.","section":"Abstract; §VII-C, Fig. 12"},{"comment":"The two oracle assumptions (known µ(x) and C(x,x′) for all pairs) are honestly stated, but the paper does not explain where the Table II parameters come from: the L_c^r values (6.5–25 m) drive the allocation via (σ_r²/L_c^r)^{3/4}, yet no estimation procedure, fit quality, or per-subregion residual semivariogram is shown. Moreover §I motivates the work by LoS/NLoS gain discontinuities that a stationary per-subregion Gaussian field cannot represent; the K-means partition into R=10 subregions is asserted to 'validate the second-order stationarity assumption' without diagnostics. Please show within-subregion stationarity checks and quantify sensitivity of the allocation to mis-estimated L_c^r.","section":"§II; §VII-A, Table II"},{"comment":"The problem is classical in adjacent literatures — near-optimal sensor placement in Gaussian processes (Krause, Singh, Guestrin, JMLR 2008), Bayesian experimental design, and geostatistical spatial sampling design minimizing mean kriging variance (e.g., Müller, 'Collecting Spatial Data'; space-filling and model-based designs) — yet none is cited. The claim of 'establishing the theoretical framework of spatial measurement' (Abstract/Conclusion) needs to be positioned against this work, in particular whether the average-Kriging-variance objective is known to have structure (approximate submodularity) that would give the greedy algorithm a guarantee.","section":"§I; §V"},{"comment":"The Fourier-transform convention is inconsistent. The PSD in (32) with its (2πL_c||k||)² factor is the Hz-convention transform (kernel e^{-i2πk·τ}), and (34)/(52) use the corresponding Hz Nyquist condition — but the Appendix derivation (45) uses the radian kernel e^{-ik·τ}, which yields (1+(kL_c)²)² without the 2π factors. The final numbers appear self-consistent under the Hz convention, so this is fixable, but the convention must be stated and the Appendix aligned. Related: the high-t_r approximation behind (38) requires t_r=πL_c^r/Δ_r ≫ 1; from Table II the smallest value is ~6.8 (region 1). Please quantify the error of (38) in the actual operating regime.","section":"§VI, Eqs. (32)–(38); Appendix (45)"}],"minor_comments":[{"comment":"Fig. 12 x-axis is labeled 'universal set size U', but U is defined (Table I) as the reduced candidate set, distinct from the universal set D. Terminology should be consistent.","section":"§VII-C, Fig. 12"},{"comment":"Table II's n_r column sums to 4296, but the figure results use U up to 4×10^4; state which U the table corresponds to.","section":"§VII-A, Table II"},{"comment":"§V-B claims SA 'converges to the globally optimal measurement pattern'; this holds only asymptotically under conditions on the cooling schedule, not for the practical schedule of Algorithm 2. Soften the wording.","section":"§V-B"},{"comment":"Nugget inconsistency: semivariogram (20) uses C0=12 in Figs. 3/5/6, but covariance (30) implies zero nugget; §VII-C says the semivariogram is 'fixed as the exponential type to comply with (30)'. State whether a nugget is used in the headline simulations.","section":"§IV, Eq. (20); §VII-C"},{"comment":"§VI's complexity argument for reducing D to U says AMSE 'must be performed with respect to the global universal set D', which retains an O(|D|) factor per candidate evaluation; the claimed 5.36-fold complexity saving (Fig. 12 discussion) should reconcile this.","section":"§VI"},{"comment":"Typos/notation: 'tradeoffof' (twice), 'semivriogram' (§IV-C), 'genuine global minima is' (§IV-B), 'Co' for C0 in t'(u) (§IV-C), 'the it is exceedingly flat' (§VII-B). Fig. 3's axis annotations are hard to read.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is competent and the derivations are careful, but the evaluation is entirely synthetic and matched by construction to the design model, so the central empirical claim currently demonstrates only that the method works when its assumptions are exactly true. The requested mismatch/estimation-error experiments are standard and inexpensive within the authors' existing pipeline, and would determine whether the paper is a solid contribution or an exercise in self-consistency. The reference list leans heavily on the authors' own CKM line of work; the omission of the GP sensor-placement and spatial-design literature should be corrected for positioning."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is a clean pipeline: treat CGM measurement placement as combinatorial AMSE minimization with ordinary Kriging variance as the closed-form surrogate, then allocate the candidate grid U across K-means path-loss subregions by a Lagrangian that minimizes mean-squared spectral truncation loss of an exponential Gaussian field. That adaptive rule (Eqs. 32–42) is the actual new piece; the rest is standard building blocks assembled carefully.\n\nWhat they do well: the Kriging derivation in §III is correct and readable, the low-dimensional geometry in §IV gives genuine intuition (measure at the points themselves when correlation length is short; interior stationary points when it is long), and the complexity discussion of greedy vs SA is honest. The math checks out—I verified the 3D Fourier PSD, the truncation integral, the high-t_r approximation, and the closed-form n_r allocation. Simulations under matched generative assumptions show a clear adaptive-vs-uniform gap (~20% at U=16k, N=400) and the expected SA advantage at large N.\n\nSoft spots in proportion. The load-bearing assumption is stated up front in §II: µ(x) and the full covariance are known a priori. Under that premise everything is well-posed; once mean and L_c^r, σ_r² must be estimated from the same sparse campaign, both the objective and the allocation weights become circular. The stress-test note is right that the headline number is in-model: residuals are Cholesky-synthesized from exactly the exponential covariance the design assumes, so Kriging variance equals true MSE by construction. Fig. 10 already shows that switching the semivariogram family moves AMSE by amounts comparable to the claimed gain, yet the main comparison never leaves the matched exponential world. The LoS/NLoS discontinuities that motivate the paper in §I are also outside the stationary-field model; K-means into R=10 is an unvalidated patch. No code, no field data.\n\nThis is for people already working on radio maps / CKM drive-test design who want a principled candidate-set construction. It deserves a serious referee—formally grounded enough, contribution real inside its model—but the practical claim needs a follow-up with estimated second-order statistics and model mismatch. I would cite the adaptive allocation derivation; I would not treat the 20% as externally validated.","headline":"Solid packaging of Kriging site selection plus a clean adaptive-grid rule from spectral truncation; the ~20% gain is real inside the matched model but untested under covariance mismatch.","tokens_in":24002,"tokens_out":581,"would_cite":true,"duration_ms":12514,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Adaptive grids guided by local correlation distance cut channel-map error about twenty percent versus uniform sampling.","keywords":["channel knowledge map","channel gain map","Kriging interpolation","Gaussian random field","adaptive discretization","combinatorial optimization","simulated annealing","spatial measurement design"],"falsifier":"With known mean and covariance, build the adaptive and uniform candidate sets of identical size U, run the same optimizer for the same N, and check whether adaptive AMSE is still ~20% lower at U=16000, N=400; if the gap vanishes or reverses, the central performance claim fails.","tokens_in":23641,"feed_emoji":"📡","tokens_out":883,"duration_ms":24643,"temperature":0.7,"pith_summary":"Building a channel gain map from sparse measurements requires choosing where to measure so the rest of the map can be filled in with least average squared error. The paper treats channel gain as a Gaussian random field, uses ordinary Kriging variance as a closed-form stand-in for that error, and turns site selection into a combinatorial problem solved by greedy search or simulated annealing. Uniform fine grids explode the search cost; the authors instead partition space into locally stationary subregions and allocate a limited candidate-grid budget by a Lagrangian that minimizes mean-squared spectral truncation loss, giving denser grids where correlation distance is short. Simulations show the adaptive candidate set yields roughly a twenty-percent lower average mean-squared error than a uniform set of the same size, and that simulated annealing eventually overtakes pure greedy once many sites are allowed. The result supplies both a theoretical frame for spatial measurement design and a practical rule for spending a fixed measurement budget.","feed_headline":"Adaptive grids cut channel-map error ~20%","feed_subtitle":"Spend denser samples where correlation is short; Kriging then fills the rest with less average error.","key_machinery":"Adaptive discretization: partition into homogeneous subregions, then allocate nr ∝ Vr (σr² / Lrc)^{3/4} grid points so that the total mean-squared PSD truncation loss is minimized under a fixed candidate budget U; Kriging variance then serves as the tractable AMSE objective for combinatorial site selection.","core_discovery":"When candidate measurement locations are chosen from an adaptively discretized set whose local density is set by subregion volume and correlation distance to minimize mean-squared information loss from the continuous Gaussian field, greedy or simulated-annealing selection of a fixed number of sites reduces global average Kriging MSE by about twenty percent relative to the same algorithms run on a uniform grid of equal cardinality.","pith_inferences":["If mean and covariance must be learned from the same campaign the method designs, a two-stage or sequential design (pilot estimates of Lrc then adaptive allocation) is the natural next algorithm.","The same Lagrangian grid allocation could transfer to other spatial fields with known second-order structure—radio environment maps, temperature, or pollution—wherever MMSE interpolation is the end goal.","When correlation distances themselves drift over time, periodic re-clustering and re-allocation would turn the static combinatorial problem into an online measurement scheduler."],"forward_implications":["A fixed measurement budget should be spent denser in short-correlation subregions (e.g., LoS/NLoS boundaries) and sparser where the field is smooth.","Candidate-set size U can be cut by roughly 1.75× while matching a target AMSE, cutting combinatorial complexity by several times.","Semivariogram shape (exponential vs Gaussian vs spherical) systematically changes the optimal geometry—from local clusters to sphere packing to near-uniform repulsion.","Ordinary Kriging variance becomes a practical design metric for deciding where to drive test equipment or place fixed probes when building CGMs for 6G."],"fun_headline_variants":["Adaptive grids cut channel-map error ~20%","Denser samples where correlation is short cut CGM MSE 20%","Adaptive discretization beats uniform for CKM measurement","Random-field grids trim global Kriging error by ~20%","Select sites on adaptive grid to cut average CGM error 20%"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The deterministic mean and the full spatial covariance (or local correlation distances and variances) are already known everywhere before any measurement locations are chosen.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive grids cut channel-map error ~20%","Denser samples where correlation is short cut CGM MSE 20%","Adaptive discretization beats uniform for CKM measurement","Random-field grids trim global Kriging error by ~20%","Select sites on adaptive grid to cut average CGM error 20%"]},"model":"grok-4.5","effort":"low","cost_usd":0.00421,"raw_usage":{"total_tokens":1302,"prompt_tokens":846,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":42104000,"prompt_tokens_details":{"text_tokens":846,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":388,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":846,"tokens_out":68,"duration_ms":6590,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T19:11:15.501784+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"With known mean and covariance, build the adaptive and uniform candidate sets of identical size U, run the same optimizer for the same N, and check whether adaptive AMSE is still ~20% lower at U=16000, N=400; if the gap vanishes or reverses, the central performance claim fails.","supporting_citations":[],"review_version":1}