{"id":"c0cd3549-9516-4e54-9949-bd878f586a30","arxiv_id":"2506.11899","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The proposed VSTD-based location-specific SCSI database and MMSE-based channel estimators achieve lower channel estimation error than existing benchmarks in simulated MU-MIMO systems.","lead":"This paper combines statistical channel information with DMRS reference signal processing to estimate uplink channels for many users at once. It builds a location-specific database of channel statistics from noisy base station measurements using tensor decomposition, then uses that database to improve channel estimation in simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Grid-size assumption d=2 m is unvalidated; if the common-SCSI representation error is not well below the noise floor, the reported NMSE advantage of SA-BCE/SA-WBCE is not secured.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and the weakest assumption identified is the grid spatial consistency assumption with d=2 m set without measurement-based validation. My stress-test pass agrees that this is the load-bearing point. The paper does provide some internal evidence via Fig. 4 that SCSI accuracy depends strongly on d, which is good, but it does not establish that d=2 m is far below the correlation distance of the simulated or real environment, nor does it quantify the representation error ΔH_v in Eq. (22). Without this, the claimed NMSE gains could be an artifact of an overly optimistic grid size. I do not see an internal mathematical contradiction sufficient to reject the paper: the MMSE derivation and the tensor decomposition steps are mostly consistent, and the use of QuaDRiGa with spatial consistency is a reasonable simulation methodology. The main deficiency is empirical validation of a key modeling assumption, which supports a conditional rather than unconditional acceptance. I would keep the reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT. The proposed concrete test directly measures the magnitude of the assumption's violation and its downstream effect, so it would settle whether the concern actually lands.","tokens_in":22160,"tokens_out":10942,"duration_ms":157018,"concrete_test":"Re-run the UMa simulation with the SCSI database built from W=10 sampling points per grid while measuring the per-grid representation error ||ΔH_v||_F^2 / ||H_v||_F^2 from Eq. (22) using the best-fit common SCSI, and compare it with the noise variance at SNR_SC=10 dB. Also sweep d from 1 m to 6 m and plot LSCSI and NMSE for SA-BCE using disjoint test-user locations from the training sampling points; if the representation error is within 3 dB of the noise floor or the NMSE advantage over the SOMP baseline vanishes for d ≤ 2 m, the reported gains are not robust to the grid consistency assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SA-BCE and SA-WBCE achieve about -21.5 dB NMSE at 15 dB SNR depends on the accuracy of the location-specific SCSI database, which in turn rests on the grid spatial consistency assumption in Eqs. (21)-(22): all users inside a d×d grid share identical delays, angles, and powers. The paper asserts that d must be much smaller than the channel correlation distance, but it never validates this for the UMa scenario. It fixes d=2 m while QuaDRiGa's spatial consistency procedure actually generates channels with continuously varying parameters, so the representation error ΔH_v in Eq. (22) is nonzero and is only absorbed into the noise model without being quantified. The paper's own Fig. 4 shows that SCSI accuracy degrades by roughly 5 dB when d increases from 2 m to 6 m, confirming that the choice of d is not benign. If the representation error at d=2 m is not far below the noise floor at SNR_SC=10 dB, the recovered SCSI is biased and the downstream NMSE gain over the SOMP-based database and other baselines would shrink. Because both the SCSI acquisition novelty and the channel-estimation gains depend on this assumption, it is the most load-bearing weakness in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses uplink DMRS-based channel estimation for MU-MIMO systems under the 3GPP Release 18 Type II OCC pattern. It proposes two SCSI-assisted Bayesian estimators, SA-BCE and its windowed low-complexity variant SA-WBCE, and a location-specific SCSI database constructed from noisy received signals using a Vandermonde-structured tensor decomposition (VSTD). The authors derive the MMSE estimators, prove the equivalence of the beam-delay domain form (Lemma 1), approximate the windowed correlations as band matrices (Eq. 18), and evaluate the method in a QuaDRiGa UMa scenario, reporting NMSE near -21.5 dB for SA-BCE at 15 dB SNR against baselines above -17 dB.","tokens_in":22431,"tokens_out":7546,"duration_ms":91175,"significance":"If the reported performance is robust, the paper would be a useful contribution: it targets a relevant 3GPP Release 18 configuration, combines SCSI database construction with noisy received signals rather than idealized samples, and proposes a complexity-reducing banded windowed estimator. The algebraic derivations of SA-BCE and the beam-delay equivalence are correct, and the simulations include meaningful baselines (SOMP-based SCSI database, OMP, VSD, EM-AMP) that strengthen the claims. The main value depends on whether the grid spatial-consistency approximation and the band-matrix approximation are quantitatively justified, since both are load-bearing for the central NMSE improvements.","major_comments":[{"comment":"The central assumption that all users in a d by d grid share identical delays, angles, and powers is not validated. In Eq. (22), the representation error Delta H_v is introduced, but in Eq. (23) it is absorbed into the Gaussian noise term without any quantification or justification. The simulation fixes d=2 m while the QuaDRiGa channel parameters are explicitly said to vary with position according to a spatial consistency procedure, so Delta H_v is nonzero. Fig. 4 shows that the SCSI accuracy degrades by roughly 5 dB when d increases from 2 m to 6 m, indicating that the choice of d is not benign. The authors should quantify the empirical representation error E||Delta H_v||_F^2/(N_d M) as a function of d and compare it with the noise variance at SNR_SC=10 dB; if it is not well below sigma^2, the claimed NMSE advantage over SOMP and other baselines is not secured.","section":"Section IV-A, Eqs. (21)-(23); Section V-C, Fig. 4"},{"comment":"The band-matrix approximation underlying SA-WBCE is unquantified. Equation (18) simply sets entries beyond band B_phi to zero, and the text claims that 'Due to the characteristics of the window functions, both Xi_f and Xi_s are band matrices with narrow band sizes.' For a finite DFT of a Kaiser or Hamming window, the transformed matrices are not exactly banded; they only decay away from the diagonal. The paper provides no bound on the approximation error ||eR_phi - bR_phi||_2 or on the resulting MMSE loss, so the claim that SA-WBCE maintains performance while reducing complexity is not supported. The authors should provide a perturbation analysis or at least a numerical study showing that the chosen band sizes (e.g., B_tau=15, B_a=20 in Fig. 8) make the truncation error negligible relative to the noise and interference terms.","section":"Section III-B, Eq. (18)"},{"comment":"The relaxed uniqueness condition for the high-rank CPD is not established for the noisy finite-sample case. Lemma 3 is cited to prior work, and Eq. (34) reduces the rank conditions to a dimension inequality min((K_1-1)K_2K_3, L_1L_2L_3W) >= Lbar. This is a generic necessary dimension condition, not a sufficient guarantee in the presence of noise and finite samples, and it does not account for the representation error absorbed in Eq. (23). Since VSTD's ability to recover hundreds of sub-paths is the basis for the SCSI database accuracy, the authors should provide either a perturbation bound for the subspace estimates in Eqs. (43)-(49) or an empirical identifiability check, such as recovery success rate versus SNR_SC and versus W, for the configuration used in Section V.","section":"Section IV-C, Lemma 3 and Eq. (34)"}],"minor_comments":[{"comment":"The normalization in E_f uses 1/N_d, but R_f and tilde R_f are defined as N_c by N_c matrices; the trace normalization should be 1/N_c. The normalization factor only shifts all curves by a constant, but it should be corrected for consistency.","section":"Equation (52)"},{"comment":"The NMSE metric averages the per-sample dB values, i.e., 10 log10 of each ratio, rather than computing 10 log10 of the ratio of summed powers. Averaging in dB can bias the reported values; please clarify whether this is intentional and, if not, use the standard ratio-of-sums definition.","section":"Equation (53)"},{"comment":"The first row of the matricization X_[3] repeats [H_g]_{1,1,1,1} twice; the second entry should be [H_g]_{1,1,1,2} (and similarly for subsequent rows).","section":"Equation (35)"},{"comment":"The proof of Lemma 1 is stated as 'substitution directly yields' the equivalence; the algebra is correct but a two-line derivation showing the identity (A + sigma^2 I)^{-1} = F_N(D + sigma^2 I)^{-1} F_N^H would make the proof self-contained.","section":"Lemma 1"},{"comment":"There is a typo 'respectivelv' in the Notations paragraph; also, the y-axis of Fig. 3 and Fig. 4 is labelled 'Accuracy of SCSI (dB)' although the metric in Eq. (51) is an MSE; consider using 'MSE (dB)' or adding a note that lower values mean higher accuracy.","section":"Notation and figures"},{"comment":"The Kaiser window shape parameter is listed in Table I, but the Hamming and rectangular windows used in Fig. 8 are not parameterized; please specify their definitions or cite the chosen formulations.","section":"Table I and Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the core derivations appear correct, but the two main performance claims rest on unquantified approximations (grid representation error and band truncation). The requested additional analyses are feasible within the manuscript's scope, so I recommend major revision rather than rejection. The self-citations appear only as background and do not raise novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real contribution: the paper shows how to build a location-specific SCSI database directly from noisy uplink received signals using Vandermonde-structured tensor decomposition, and integrates it with MMSE-based DMRS channel estimation for Type-II OCC. The combination is not in the cited literature. Second, the headline gain—about -21.5 dB NMSE at 15 dB SNR, several dB better than baselines—is conditional on the grid spatial consistency assumption (d=2 m) and on an unquantified band-matrix approximation in the windowed estimator. If those do not hold in practice, the gain shrinks.\n\nWhat it does well: the derivations of SA-BCE and the beam-delay equivalence (Lemma 1) are algebraically correct. The complexity reduction from O(N^3+M^3) to band-matrix inversion is argued clearly. The simulation study is extensive: SNR, delay spread, grid size, window functions, number of subcarriers. The comparison to SOMP-based SCSI database is appropriate, and the paper honestly shows SA-WBCE is worse than SA-BCE in the high-SNR regime.\n\nThe soft spots, in order of importance. The grid spatial consistency assumption (Eqs. 21-22) is load-bearing. The paper asserts d must be much smaller than the correlation distance, then fixes d=2 m without validating it against the QuaDRiGa spatial consistency procedure. Fig. 4 shows SCSI accuracy degrades by about 5 dB when d goes from 2 to 6 m, so the choice is not benign. The representation error ΔH_v is absorbed into the noise covariance (Eq. 23) without justification. If that error is not well below the noise floor at SNR_SC=10 dB, the recovered SCSI is biased and the downstream gain over baselines narrows. A minor point: Eq. (18) band approximation is heuristic; the paper gives no error analysis for finite band size, only simulation. Also the relaxed uniqueness condition (34) is checked numerically, not proven for the noisy finite-sample case. Finally, the closely related CKM construction in [23] is cited but not benchmarked; a direct comparison would strengthen the paper.\n\nThe math is generally solid, the citations look appropriate, and the self-citations are background, not circular. The paper would benefit from releasing code and adding error bars, but that is not a blocker.\n\nWho this is for: people working on DMRS estimation, channel knowledge maps, or tensor methods for 5G-Advanced/6G. It deserves a serious referee. My recommendation: send it to peer review, ask for a quantitative treatment of the representation error and a sensitivity analysis of grid size, plus ideally a comparison to [23].","headline":"A genuinely new VSTD-based SCSI database construction combined with MMSE channel estimation, but the unvalidated grid-size assumption and unquantified approximations leave the headline gains conditional.","tokens_in":22972,"tokens_out":3932,"would_cite":true,"duration_ms":42189,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that statistical channel state information tied to user location can suppress DMRS pilot interference, with the proposed estimators reaching about -21.5 dB NMSE at 15 dB SNR where baselines stay above -17 dB.","keywords":["MU-MIMO","DMRS","statistical channel state information","channel estimation","tensor decomposition","Vandermonde structure","spatial consistency","3GPP Release 18"],"falsifier":"Run the same VSTD database construction and SA-BCE pipeline on a measured or simulated channel whose spatial correlation distance is known, then sweep the grid size $d$. If SCSI accuracy or NMSE degrades sharply while $d$ is still well below the correlation distance, or if the gain over the SOMP baseline vanishes once the database is built from noisy finite samples, the grid-consistency assumption is falsified. A second check is to compare high-SNR NMSE against the ideal-SCSI upper bound: the claim that SCSI-assisted estimators significantly outperform baselines would fail if they fall back to baseline level in that comparison.","tokens_in":21948,"feed_emoji":"📶","tokens_out":6934,"duration_ms":81977,"temperature":0.7,"pith_summary":"In a multi-user MIMO uplink, demodulation reference signals from different users are superposed on the same resource elements and separated by orthogonal cover codes; with frequency-selective channels that separation leaks, so user count growth comes with pilot interference. This paper claims that statistical channel state information tied to user location can remove most of that leakage. It proposes two MMSE-based estimators, SA-BCE and a windowed beam-delay variant SA-WBCE, that use delays, angles, and path powers as prior statistics, and it builds those statistics offline as a grid-based location-specific SCSI database recovered from noisy received signals by Vandermonde-structured tensor decomposition. In simulations of a 3GPP urban macro scenario, the database reaches roughly -23.5 dB SCSI accuracy at 180 subcarriers and 10 dB SNR, and the estimators reach about -21.5 dB NMSE at 15 dB SNR where the compared baselines remain above -17 dB. If this holds, practical uplink MU-MIMO can serve more users without sacrificing channel estimation quality, and without repeatedly estimating channel statistics in real time.","feed_headline":"Location-aware statistical CSI pushes MU-MIMO NMSE to -21.5 dB","feed_subtitle":"Database-derived spatial statistics beat OMP, VSD, and EM-AMP baselines.","key_machinery":"The load-bearing structure is the fourth-order tensor $\\mathcal{H}_g$ formed by concatenating $W$ least-squares channel estimates from sampling points inside one grid. Its canonical polyadic decomposition has factor matrices $B^{(1)}, B^{(2)}, B^{(3)}$ that are Vandermonde matrices with generators $z_{1,l}=e^{-\\bar{\\jmath}2\\pi\\Delta f\\bar{\\tau}_{g,l}}$, $z_{2,l}=e^{-\\bar{\\jmath}\\pi\\cos\\bar{\\theta}_{g,l}}$, and $z_{3,l}=e^{-\\bar{\\jmath}\\pi\\sin\\bar{\\theta}_{g,l}\\cos\\bar{\\varphi}_{g,l}}$, together with a factor $B^{(4)}$ holding path gains. Vandermonde-structured tensor decomposition (VSTD) uses spatial smoothing and SVD to build shift-invariance relations, reads the Vandermonde generators by eigenvalue decomposition, and reconstructs delays, angles, and powers from those generators. On the estimation side, the windowed beam-delay MMSE estimator (SA-WBCE) is the complexity-reduction mechanism: windowing concentrates channel energy so that the delay- and beam-domain correlation matrices become band matrices, lowering inversion complexity from $O(N^3+M^3)$ to $O(N B_f^2 + M B_s^2)$.","core_discovery":"The paper's central claim is that code-domain DMRS pilot interference in uplink MU-MIMO can be substantially undone once the base station knows the users' statistical channel parameters, and that those parameters can be learned from noisy received signals rather than assumed perfect. Two estimators carry the claim: SA-BCE applies a frequency-domain MMSE step to separate the frequency-domain OCC and an antenna-domain MMSE step to suppress noise, while SA-WBCE moves both steps to the beam-delay domain and uses windowing so that the correlation matrices become band matrices, cutting matrix inversion cost. The SCSI itself is delivered by a grid-based location-specific database: each grid is represented by one common set of delays, elevation and azimuth angles, and path powers, recovered by Vandermonde-structured tensor decomposition of least-squares channel estimates from W sampling points inside the grid. The simulation evidence shows the database reaching a SCSI accuracy around -23.5 dB at $N_d=180$ and 10 dB SNR, and the estimators reaching NMSE around -21.5 dB (SA-BCE) and -19.5 dB (SA-WBCE) at 15 dB SNR, with all compared baselines above -17 dB.","pith_inferences":["Editorial inference: the spatial-consistency grid is the scheme's fragile link; the paper fixes $d=2$ m in simulation without measurement-based validation that this is below the true correlation distance, so a natural extension is to size the grid from measured correlation distance or adapt it per cluster.","Editorial inference: the relaxed uniqueness condition (34) is checked numerically rather than proven for noisy finite-sample recovery; a deterministic noise tolerance bound for the shift-invariance eigenvalue step would harden the central claim.","Editorial inference: the same tensor machinery could be reused for database refresh as the environment changes, treating each refresh as a low-cost update rather than a full rebuild from scratch."],"forward_implications":["With the location-specific SCSI database, SA-BCE and SA-WBCE outperform OMP, VSD, EM-AMP, and an SOMP-database baseline across SNR; at 15 dB SNR they reach about -21.5 dB and -19.5 dB NMSE versus above -17 dB for the baselines.","The SCSI-assisted estimators are nearly immune to delay spread: increasing delay spread from 200 ns to 500 ns costs only about 0.5 dB for the proposed schemes, compared with about 6.5 dB degradation for VSD and EM-AMP.","SA-WBCE keeps most of the estimation gain while replacing cubic-complexity matrix inversions with band-matrix inversions, and the Kaiser window offers about 2 dB improvement over a rectangular window at the tested band sizes.","Grid size directly trades database storage against acquisition cost: holding SCSI accuracy near -23.4 dB needs $N_d=120$ subcarriers at grid size $d=2$ m but $N_d=240$ at $d=5$ m.","Because the environment is assumed quasi-static over a period $T$, one database construction serves DMRS-based estimation over many subsequent OFDM symbols, avoiding frequent real-time SCSI estimation."],"supporting_citations":[{"why":"Supplies the structured high-rank tensor approach and the VSD benchmark for decomposing Vandermonde-factored channel matrices with hundreds of sub-paths.","marker":"[8]"},{"why":"Defines the 3GPP Release 18 Type II DMRS OCC pattern and 24-port configuration used in the signal model.","marker":"[6]"},{"why":"Gives the MMSE two-port DMRS estimator that SA-BCE extends to multi-user OCC decomposition.","marker":"[10]"},{"why":"Provides the channel and fading model with uncorrelated path gains and the coherence-time arguments used in the system model.","marker":"[24]"},{"why":"Supplies the standard MMSE estimation equations used in the frequency and antenna domains.","marker":"[26]"},{"why":"Provides the Vandermonde-factor CPD spatial-smoothing uniqueness condition and recovery method behind Lemma 3 and the VSTD algorithm.","marker":"[40]"},{"why":"Provides the ray-tracing channel generation software used to produce simulation channels.","marker":"[43]"},{"why":"Defines the 3GPP 38.901 urban macro channel model and the spatial consistency procedure used in simulations.","marker":"[44]"},{"why":"Supplies the OMP algorithm used as a channel-estimation benchmark.","marker":"[45]"},{"why":"Supplies the SOMP algorithm used as the SCSI-database construction baseline.","marker":"[46]"}],"fun_headline_variants":["Location-aware CSI boosts MU-MIMO channel estimation","MU-MIMO channel estimation gains from location-specific SCSI","SCSI database from tensor decomposition improves MU-MIMO pilots","Bayesian estimators with spatial statistics cut MU-MIMO NMSE","Grid-based SCSI database lifts MU-MIMO uplink performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that every user inside the same $d\\times d$ grid sees the same multi-path delays, angles, and powers, so the grid size must be far smaller than the channel's spatial correlation distance; if that is not true in a deployment, the database stores wrong statistics and the estimation gain disappears.","fun_headline_variants_meta":{"raw":{"variants":["Location-aware CSI boosts MU-MIMO channel estimation","MU-MIMO channel estimation gains from location-specific SCSI","SCSI database from tensor decomposition improves MU-MIMO pilots","Bayesian estimators with spatial statistics cut MU-MIMO NMSE","Grid-based SCSI database lifts MU-MIMO uplink performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1510,"prompt_tokens":1118,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":311}},"tokens_in":734,"tokens_out":392,"duration_ms":3936,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T01:02:16.386657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same VSTD database construction and SA-BCE pipeline on a measured or simulated channel whose spatial correlation distance is known, then sweep the grid size $d$. If SCSI accuracy or NMSE degrades sharply while $d$ is still well below the correlation distance, or if the gain over the SOMP baseline vanishes once the database is built from noisy finite samples, the grid-consistency assumption is falsified. A second check is to compare high-SNR NMSE against the ideal-SCSI upper bound: the claim that SCSI-assisted estimators significantly outperform baselines would fail if they fall back to baseline level in that comparison.","supporting_citations":[{"cited_title":"Estimating channels with hundreds of sub-paths for MU-MIMO uplink: A structured high-rank tensor approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the structured high-rank tensor approach and the VSD benchmark for decomposing Vandermonde-factored channel matrices with hundreds of sub-paths."},{"cited_title":"Standard 3GPP T.R","cited_arxiv_id":null,"evidence_quote":"Defines the 3GPP Release 18 Type II DMRS OCC pattern and 24-port configuration used in the signal model."},{"cited_title":"MMSE Channel Estimation for Two-Port Demodulation Reference Signals in New Radio","cited_arxiv_id":"2007.14168","evidence_quote":"Gives the MMSE two-port DMRS estimator that SA-BCE extends to multi-user OCC decomposition."},{"cited_title":"Tse and P","cited_arxiv_id":null,"evidence_quote":"Provides the channel and fading model with uncorrelated path gains and the coherence-time arguments used in the system model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard MMSE estimation equations used in the frequency and antenna domains."},{"cited_title":"Blind signal separation via tensor decomposition with Vandermonde factor: Canonical polyadic decompo- sition,","cited_arxiv_id":null,"evidence_quote":"Provides the Vandermonde-factor CPD spatial-smoothing uniqueness condition and recovery method behind Lemma 3 and the VSTD algorithm."},{"cited_title":"QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,","cited_arxiv_id":null,"evidence_quote":"Provides the ray-tracing channel generation software used to produce simulation channels."},{"cited_title":"document 3GPP T.R","cited_arxiv_id":null,"evidence_quote":"Defines the 3GPP 38.901 urban macro channel model and the spatial consistency procedure used in simulations."},{"cited_title":"On the exact recovery condition of simultaneous orthogonal matching pursuit,","cited_arxiv_id":null,"evidence_quote":"Supplies the SOMP algorithm used as the SCSI-database construction baseline."}],"review_version":1}