{"id":"72cac348-f4f4-46ca-b935-a1f8f6b8478e","arxiv_id":"2505.04933","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new two-dimensional pilot design and tensor-based estimator let massive MIMO-OFDM base stations estimate channels for hundreds of users with lower error and lower complexity than previous approaches.","lead":"This paper proposes a way for mobile base stations with many antennas to estimate wireless channels using test signals shifted in both frequency and time. The method is most useful when a base station serves hundreds of users, where it reduces channel estimation error by about 8 decibels in simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof fails for the oversampled DFT beam dictionaries used in the simulations: condition (40) does not force the cross-covariance to vanish even with perfect Wu.","rationale":"The reader's weakest assumption concerned imperfect knowledge of Wu and off-grid leakage; those are legitimate practical risks. My stress-test found a more basic internal error: the proof of Theorem 1 treats the oversampled DFT beam dictionaries as if their columns become orthogonal as M,K,Np grow. They do not. Since the paper's central formal claim is the asymptotic MMSE optimality under condition (40), and this claim fails in the oversampled regime in which all simulation gains are reported, the theoretical foundation of the TFPSP design is unsupported. This does not by itself disprove the empirical >8 dB NMSE gain; the simulations may still be reproducible and useful. But the current manuscript presents the theorem as the justification for 'optimal' TFPSP design and scheduling, so the verdict should be stricter than CONDITIONAL: the paper should be rejected in its present form and reconsidered only if the theorem is corrected (e.g., restricted to F=1 or proven under a physical separation condition) and the simulations are rerun under the corrected claim. I therefore set agreement_with_reader to 'disagree': the load-bearing concern is not the one the reader identified, though both concern the triple-beam grid.","tokens_in":29241,"tokens_out":13564,"duration_ms":141786,"concrete_test":"Compute V_s^H V_s for the Section II-C construction with M=128, Ntheta=256 (F=2) and print max_{p!=q} |[V_s^H V_s]_{p,q}|. Then evaluate Eq. (65) with beta_T=1 for Wu and Wu' each having one unit entry at q=(0,0,0) and p=(1,1,1), respectively, plus the analogous K- and Np-dimensional kernels. If the resulting Qu,u' is nonzero at order MKNp while condition (40) holds, Theorem 1 is false for the F=2 regime of the paper's own simulations. A corrected proof must either restrict to F=1 (square DFT dictionaries) or add an explicit separation condition on physical angle/delay/Doppler and re-derive a finite-system bound.","verdict_should_be":"REJECT","load_bearing_attack":"The formal anchor of the paper is Theorem 1, whose proof (Appendix A) reduces the cross term Qu,u' in (62) to products of unnormalized Dirichlet kernels alpha_M((p1-q1)/Ftheta), alpha_K((p2-q2)/Ftau), alpha_Np((q3-p3)/Fnu) in (64)-(65). The proof then asserts alpha_A(x) -> delta(x) as A->infinity, so all off-diagonal index pairs vanish. This is incorrect: with Ntheta=Ftheta*M etc. as defined in Section II-C, for fixed nonzero x the unnormalized kernel is O(A) and does not tend to zero; only alpha_A(x)/A converges (to a periodic delta train) in distribution. Concretely, set F=2 and take two users with single nonzero TB entries at q=(0,0,0) and p=(1,1,1) after scheduling. Condition (40) holds because the supports are disjoint, yet the product alpha_M(1/2)*alpha_K(1/2)*alpha_Np(1/2) is O(MKNp), not zero, so Qu,u' and hence RSFT_pilot_{u,u'} do not vanish. The beam matrices are redundant frames, not orthonormal bases; adjacent oversampled beams have inner products of order M, so disjoint coefficient support does not imply disjoint physical channel support. Thus the claimed asymptotic optimality of TFPSPs is not established in the F>=2 regime used in Figs. 4-9, independently of the Wu-estimation and off-grid leakage concerns raised by the reader.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a complete uplink channel acquisition pipeline for massive MIMO-OFDM. Section II-C introduces a triple-beam (TB) tensor channel model that represents the space-frequency-time channel as beam matrices multiplying a sparse TB-domain channel tensor. Section III proposes time-frequency phase-shifted pilots (TFPSPs), derives an 'optimal' interference-free condition (Theorem 1, condition (40)) under which the MMSE lower bound (39) is claimed to be achieved, and gives a DSatur-based pilot scheduling algorithm (Algorithm 1). Section IV extends the authors' earlier information geometry approach (IGA) to tensors for estimating the TB channel, with an FFT-accelerated low-complexity implementation and channel prediction. Section V reports QuaDRiGa simulations showing large NMSE gains over APSP-based channel acquisition (more than 8 dB at 300 UTs and SNR = 20 dB), faster convergence than GAMP and EPV, and lower complexity.","tokens_in":29475,"tokens_out":28894,"duration_ms":278487,"significance":"The high-level idea is strong and timely: exploiting joint angular, delay, and Doppler sparsity through a two-dimensional pilot phase-shift schedule to serve many UTs with fixed pilot overhead is well motivated, and the reported gains are substantial. The paper is clearly written, the tensor machinery is developed carefully, and the simulation study is reasonably broad (48 and 300 UTs, two speeds, three estimation baselines, complexity and convergence comparisons). The FFT-based simplification of the tensor IGA is a genuine algorithmic contribution, and the use of a physically motivated QuaDRiGa channel model rather than synthetic sparse channels strengthens the empirical claims. The estimator is inherited from the authors' prior work [9], [25], [26] rather than fitted to the data in this paper, which removes any circularity concern. However, the theoretical anchor of the pilot design, Theorem 1, is not correct for the oversampled DFT dictionaries used throughout the simulations (see Major Comment 1), so the 'optimal' terminology overstates what is proven; the headline practical claims rest on the simulations alone.","major_comments":[{"comment":"The proof of sufficiency of condition (40) is incorrect for the oversampled beam dictionaries used in all the simulations, and the counterexample below shows that the claimed sufficiency fails in that regime. The proof states that α_A(x) → δ(x) as A → ∞ and concludes that whenever any of p1 ≠ q1, p2 ≠ q2, p3 ≠ q3 holds, at least one of α_M(·), α_K(·), α_Np(·) in (65) vanishes in the limit. With Nϑ = FϑM, Nτ = FτK, Nν = FνNp (Section II-C) and Fϑ, Fτ, Fν ≥ 1, this is false: α_A(x) = Σ_{a=0}^{A−1} exp(−j2πax/A) is an unnormalized Dirichlet kernel, and for fixed nonzero x one has |α_A(x)| = O(A), not o(1); only the normalized kernel α_A(x)/A converges to a periodic delta train in distribution. The orthogonality the proof requires holds only in the critically sampled case Fϑ = Fτ = Fν = 1. Explicit counterexample in the paper's F = 2 setting: let Fϑ = Fτ = Fν = 2, take two users with single non-zero TB entries Wu = δ_{(0,0,0)} and Wu′ = δ_{(1,1,1)} and identical (zero) phase shifts. Condition (40) holds because the supports are disjoint, yet the cross-covariance in (36)-(32) is the rank-one tensor with entries [V^p_T]_{a,(1,1,1)}[V^p_T]^∗_{b,(1,1,1)}, of unit magnitude and independent of M, K, Np; equivalently, the kernel product in (65) equals α_M(1/2)α_K(1/2)α_Np(1/2), whose magnitude is O(MKNp), not zero. Hence the assertion lim_{M,K,Np→∞} Q_{u,u′} = 0 in Appendix A fails, C_{u,all} ≠ C_u, and the claimed achievability of εMSE,min is not established. Because Algorithm 1 schedules against condition (40) and the abstract advertises an 'optimal TFPSP design', this is a load-bearing error. Since Figs. 4-9 use F = 2 and F = 4, and Fig. 4 shows that the F = 2 gain over F = 1 is the main source of the reported performance, the theorem cannot be rescued by falling back to F = 1 without giving up the paper's central performance claims. Please either restrict Theorem 1 to critically sampled dictionaries (where the claim holds exactly for all finite M, K, Np), or replace the sufficiency proof with a quantitative analysis of the residual interference for redundant beam dictionaries.","section":"Appendix A, Eqs. (63)-(65); Section III-B (Theorem 1)"},{"comment":"The 'optimal' claim is conditional on ideal statistics and exact grid matching, but the paper never analyzes the gap between the ideal setting and the simulation setting. The analysis assumes that Wu of all UTs is available to the BS (end of Section II-C) and that every physical path is exactly on the discretized TB grid (the approximation in (12)); in the simulations, by contrast, Wu is estimated from data via the method of [46] and the QuaDRiGa channels have continuous angles, delays, and Doppler shifts, so measured power leaks off-grid and the empirical Wu is not sparse in the exact sense used to define Snz and condition (40). Under mismatch, condition (40) can be neither certified nor verified, and Theorem 1 provides no finite-system or mismatch-robust guarantee. The threshold γ in Algorithm 1 is also a free parameter with no recommended value or sensitivity study. I recommend adding a robustness study (imperfect Wu, off-grid channels) and softening the 'optimal' terminology in the abstract and Section III-B unless it is explicitly restricted to the idealized case.","section":"Section II-C (last paragraph); Sections III-C and IV-A"},{"comment":"The low-complexity claim is not fully substantiated. The O(C̄1) figure is stated for a single application of A∗3B or AH_3∗3C ((57)-(58)), but the total cost of Algorithm 2 depends on the number of iterations, which is governed by the unspecified damping factor α and is only reported empirically (300 iterations); no bound or convergence-rate analysis is given for the tensor recursion (56). The convergence results cited from [9], [25], [26] are for the (simplified) IGA of those papers, not for the coupled D/F update in (56), so the paper should either prove convergence of (56) or state explicitly that convergence is inherited only at the level of the prior analyses. In addition, the complexity comparison of Fig. 10 does not include the cost of Algorithm 1 or the acquisition of Wu, although both are part of the proposed acquisition pipeline, and the runtime figures are reported without platform details. Please provide a formal per-iteration and total complexity statement and declare the settings of α and γ used in the simulations.","section":"Section IV-B and Fig. 10"}],"minor_comments":[{"comment":"The first Dirichlet factor is written as α_M((p1−q1)/Fν), but consistency with Eq. (65) and with the exponent e^{−j2πr1(p1−q1)/Nϑ} requires Fϑ.","section":"Appendix A, Eq. (64)"},{"comment":"The symbol γ is used both for the integer coprime to Np in the ZC sequence (24) and for the scheduling overlap threshold in Algorithm 1; please rename one of them.","section":"Eq. (24) and Algorithm 1"},{"comment":"The assignment 'φUi′ = φ' presumably should read 'φUi = φ'.","section":"Algorithm 1, step 15"},{"comment":"The text 'as shown in Fig. V' should refer to Fig. 9, and the legend of Fig. 6 lists 'VEP' where 'EPV' is meant.","section":"Section V"},{"comment":"Typographical issues: 'implys' after Theorem 1; 'has an affect on' in Section III-B; 'acquition' in the Fig. 4 caption; 'normalized factor' and 'iteraction terms' in Section IV-A; 'seperation' and 'accomodated' in Section III.","section":"Throughout"},{"comment":"The paper would benefit from including a genie-aided reference curve corresponding to εMSE,min in (39), so that the reader can calibrate how close the heuristic scheduling (with its free threshold γ) comes to the claimed bound.","section":"Figs. 5 and 8"}],"recommendation":"major_revision","confidential_remarks":"The skeptic's concern about the Dirichlet kernels is valid, and it is the decisive issue: Theorem 1 is not correct for the oversampled dictionaries that generate the paper's headline gains (F = 2 in Figs. 4-9). I nevertheless recommend major revision rather than rejection because the empirical part is substantial and there is a clean repair path (restricting the theorem to F = 1, or adding a quantitative residual-interference analysis). Editors may also wish to check: (i) the convergence and fixed-point theorems for the simplified IGA are imported from the authors' own preprints [25], [26]; the journal version should prove the tensor-case recursion (56) or state the inheritance explicitly; (ii) the novelty relative to the GLOBECOM 2024 version [1] is not itemized; (iii) the scheduling threshold γ and damping factor α are free parameters with no guidance, which limits reproducibility of the figures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look: it extends the SF-beam IGA/APSP line to a triple-beam (space, frequency, time) tensor channel model, introduces time-frequency phase-shifted pilots, and couples them to a tensor-based IGA with an FFT low-complexity implementation. That is a genuine extension, and the QuaDRiGa results with 300 UTs showing 8+ dB NMSE gains over APSP-IGA are eye-catching. If those gains survive independent testing, the idea has real practical value for dense 6G scenarios.\n\nBut the theoretical anchor has a serious flaw. Theorem 1 claims that the MMSE lower bound is achieved when equivalent TB power distributions are non-overlapping, condition (40). The proof in Appendix A reduces the cross-covariance to products of unnormalized Dirichlet kernels, alpha_M((p-q)/F), and then asserts alpha_A(x) -> delta(x) as A -> infinity. That is not true for the oversampled DFT dictionaries actually used (F>=2). For nonzero integer (p-q), alpha_M((p-q)/F) is O(M) and does not tend to zero; only alpha_A(x)/A converges to a delta train. Adjacent oversampled beams have inner products of order M, so disjoint coefficient support in the oversampled frame does not imply vanishing cross-covariance. In the F=2 setting of Figures 4-9, condition (40) is not sufficient to reach the bound. This is a load-bearing flaw, because the text repeatedly calls the TFPSP design \"optimal.\"\n\nTwo secondary soft spots: the whole pipeline assumes perfect knowledge of every UT's power tensor W_u, and the simulations omit tuning parameters (gamma, alpha), Monte-Carlo counts, and code. The W_u assumption is a standard idealization in this literature, but here it is essential to the scheduling and the estimator, and the paper does not quantify sensitivity to estimation error or off-grid leakage. The missing simulation details make the 8 dB hard to verify independently.\n\nBottom line: the paper is a serious technical effort with an interesting idea, but as it stands the core theoretical claim is not established. I would send it to peer review rather than desk-reject, with a request to fix Theorem 1 (either prove it for normalized dictionaries or state it as a heuristic condition) and to release the detailed simulation settings and code. I would not cite the optimality result until that is done.","headline":"Interesting pilot design and plausible empirical gains, but Theorem 1's proof is wrong for the oversampled DFT dictionaries used in the simulations.","tokens_in":30125,"tokens_out":6683,"would_cite":false,"duration_ms":72334,"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":"Time-frequency phase-shifted pilots reduce massive MIMO channel estimation error by over 8 dB while serving more users with the same pilot overhead.","keywords":["massive MIMO-OFDM","channel estimation","triple-beam channel tensor","time-frequency phase-shifted pilots","pilot scheduling","information geometry","tensor estimation","inter-user interference"],"falsifier":"Give each UT a channel with off-grid delay and Doppler values that do not sit on the discretized triple-beam grid, estimate $\\mathcal{W}_u$ from finite samples instead of assuming it known, then run the TFPSP scheduler and tensor IGA at 20 dB SNR with 300 UTs. If the pairwise overlap measure $\\eta(\\cdot,\\cdot)$ of the scheduled equivalent power tensors is strictly positive for most pairs and the NMSE advantage over APSP-IGA drops below 8 dB, the central claim as stated fails in that regime.","tokens_in":28957,"feed_emoji":"📶","tokens_out":4479,"duration_ms":42924,"temperature":0.7,"pith_summary":"This paper proposes a channel acquisition method for massive MIMO-OFDM that lets many user terminals share the same pilot resource without a proportional loss in accuracy. The key move is to shift pilots in both frequency and time, not just frequency, and to model the channel as a sparse tensor over spatial, delay, and Doppler beams. The authors show that when each user's equivalent triple-beam power distribution does not overlap any other's, inter-user pilot interference disappears and the MMSE lower bound is reached. Simulations with 300 users report normalized mean-square error more than 8 dB lower than the adjustable phase-shifted pilot baseline, along with a lower-complexity estimator.","feed_headline":"Pilot shifts in time and frequency cut MIMO channel error by 8 dB","feed_subtitle":"Two-dimensional phase-shifted pilots let a base station serve far more users without extra pilot overhead.","key_machinery":"The triple-beam channel tensor model: the space-frequency-time channel is written as $\\mathbf{H}^{\\mathrm{SFT}}_{u,T} = \\mathcal{V}_T *_3 \\mathcal{H}^{\\mathrm{TB}}_u$, a product of DFT-structured beam matrices with a sparse tensor whose axes are spatial, delay, and Doppler beams. Phase-shifted pilots act as tensor cyclic shifts on $\\mathcal{H}^{\\mathrm{TB}}_u$, so scheduling pilots becomes the combinatorial task of choosing shifts that make the equivalent power tensors $\\mathcal{W}_{u,L_{\\varphi_u},L_{\\phi_u}}$ pairwise non-overlapping. The information geometry estimator projects the posterior onto a tractable Gaussian manifold, and the DFT structure of the beam matrices lets each projection step run through fast Fourier transforms.","core_discovery":"The central claim is that time-frequency phase-shifted pilots (TFPSPs) convert pilot separation into a tensor cyclic shift in the triple-beam domain, so two users stop interfering exactly when their shifted power tensors $\\mathcal{W}_{u,L_{\\varphi_u},L_{\\phi_u}}$ have disjoint supports. Theorem 1 states that as $M, K, N_p \\to \\infty$, this non-overlap condition achieves the MMSE lower bound for channel estimation without increasing pilot overhead. The paper also claims that the tensor-based information geometry estimator reaches near-MMSE accuracy with much lower complexity than direct tensor inversion, and that in high-user-count simulations the TFPSP pipeline outperforms the prior APSP-IGA approach by more than 8 dB in NMSE at 20 dB SNR.","pith_inferences":["If $\\mathcal{W}_u$ is estimated rather than known exactly, the non-overlap condition cannot be certified; a natural robustness test is to feed noisy covariance estimates into the scheduler and observe when the NMSE gain falls below 8 dB.","The same two-dimensional phase-shift idea may transfer to delay-Doppler grids such as OTFS, where fractional off-grid Doppler leakage would stress the grid-sparsity assumption.","The threshold $\\gamma$ in the DSatur grouping trades estimation performance for scheduling complexity, so an adaptive threshold based on the current UT population could improve pilot reuse without rerunning the full graph coloring.","Theorem 1 is asymptotic; finite arrays will retain residual interference from sinc-like leakage, so the practical reach of the claim depends on how fast the DFT factors $\\alpha_M(\\cdot)\\alpha_K(\\cdot)\\alpha_{N_p}(\\cdot)$ decay at finite $M$, $K$, and $N_p$."],"forward_implications":["A base station can serve many more UTs with the same pilot overhead, because each UT uses two phase-shift indices instead of one and only needs its sparse triple-beam support to avoid others.","The TFPSP scheduling criterion, pairwise zero Hadamard product of shifted power tensors, gives a direct and checkable design rule for pilot assignment.","The tensor-based IGA converges to the MMSE estimate at its fixed point, and its FFT-based implementation lowers complexity so a 300-UT system can be estimated in roughly a quarter of the runtime of GAMP or EPV.","The estimated triple-beam tensor directly supports channel prediction over the data segment of the slot, with growing benefit as UT speed increases from 30 to 60 km/h."],"supporting_citations":[{"why":"Supplies the adjustable phase-shifted pilot (APSP) baseline that TFPSPs are compared against.","marker":"[13]"},{"why":"Supplies the information geometry approach and the spatial-frequency beam model that the tensor IGA extends.","marker":"[9]"},{"why":"Supplies the simplified IGA fixed-point iteration that the tensor version adapts.","marker":"[25]"},{"why":"GAMP is used as a simulation baseline for TB domain channel estimation.","marker":"[23]"},{"why":"EPV is used as a simulation baseline for TB domain channel estimation.","marker":"[24]"},{"why":"Supplies the standardized urban macro non-line-of-sight channel model used to generate the simulation channels.","marker":"[45]"},{"why":"Supplies the method used to acquire the statistical CSI $\\mathcal{W}_u$ for each UT in simulations.","marker":"[46]"}],"fun_headline_variants":["Time-frequency pilot shifts deliver 8 dB gain in massive MIMO","Phase-shifted pilots in time and frequency cut MIMO error by 8 dB","Low-complexity tensor estimation with time-frequency shifted pilots for massive MIMO","8 dB NMSE gain: time-frequency phase-shifted pilots for massive MIMO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the base station knows every user's triple-beam power tensor $\\mathcal{W}_u$ exactly and that each channel is exactly sparse on the discretized angle-delay-Doppler grid, so the pairwise non-overlap condition can be certified; if the covariance is estimated with error or energy leaks off the grid, the interference suppression and the 8 dB gain can shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Time-frequency pilot shifts deliver 8 dB gain in massive MIMO","Phase-shifted pilots in time and frequency cut MIMO error by 8 dB","Low-complexity tensor estimation with time-frequency shifted pilots for massive MIMO","8 dB NMSE gain: time-frequency phase-shifted pilots for massive MIMO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2717,"prompt_tokens":948,"completion_tokens":1769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1686}},"tokens_in":564,"tokens_out":1769,"duration_ms":12376,"temperature":1.0,"reasoning_tokens":1686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:18:13.383348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give each UT a channel with off-grid delay and Doppler values that do not sit on the discretized triple-beam grid, estimate $\\mathcal{W}_u$ from finite samples instead of assuming it known, then run the TFPSP scheduler and tensor IGA at 20 dB SNR with 300 UTs. If the pairwise overlap measure $\\eta(\\cdot,\\cdot)$ of the scheduled equivalent power tensors is strictly positive for most pairs and the NMSE advantage over APSP-IGA drops below 8 dB, the central claim as stated fails in that regime.","supporting_citations":[{"cited_title":"Chan nel ac- quisition for massive MIMO-OFDM with adjustable phase shif t pilots,","cited_arxiv_id":null,"evidence_quote":"Supplies the adjustable phase-shifted pilot (APSP) baseline that TFPSPs are compared against."},{"cited_title":"Channel estimation for massive MIMO: An information geome try approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the information geometry approach and the spatial-frequency beam model that the tensor IGA extends."},{"cited_title":"Efficient Information Geometry Approach for Massive MIMO-OFDM Channel Estimation","cited_arxiv_id":"2401.02035","evidence_quote":"Supplies the simplified IGA fixed-point iteration that the tensor version adapts."},{"cited_title":"Generalized approximate message passing f or estimation with random linear mixing,","cited_arxiv_id":null,"evidence_quote":"GAMP is used as a simulation baseline for TB domain channel estimation."},{"cited_title":"Un ifying message passing algorithms under the framework of constrai ned bethe free energy minimization,","cited_arxiv_id":null,"evidence_quote":"EPV is used as a simulation baseline for TB domain channel estimation."},{"cited_title":"Q uaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,","cited_arxiv_id":null,"evidence_quote":"Supplies the standardized urban macro non-line-of-sight channel model used to generate the simulation channels."},{"cited_title":"A GAMP-bas ed low complexity sparse Bayesian learning algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the method used to acquire the statistical CSI $\\mathcal{W}_u$ for each UT in simulations."}],"review_version":1}