{"id":"11cf2408-10c4-43fd-87bb-0210a5086e18","arxiv_id":"2501.02985","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-timescale estimation scheme, 2TCE-TSP, uses piecewise beam training to estimate a ratio vector per block, cutting per-block pilot overhead from M to about M/N_RF.","lead":"This paper presents a channel estimation method for RIS-aided near-field wireless systems that reduces pilot overhead by learning only a low-dimensional channel ratio each time block. It is relevant to future mmWave and terahertz networks where large intelligent surfaces make training expensive.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is false as stated: pairwise-distinct V(v_b) subspaces do not guarantee full-rank G_q; a 4x4 Hadamard example satisfies all hypotheses yet Rank(G_q)=3.","rationale":"The reader's weakest_assumption correctly identifies Theorem 1 and the invalid Lemma 1 as the load-bearing point. My analysis strengthens this: even when the assumed condition V(v_b1)!=V(v_b2) holds, full rank can fail, so the theorem is false as stated. This does not by itself prove the simulated method fails: the random near-field channels may satisfy a stronger transversal-intersection property with high probability, and Fig. 4 suggests good conditioning in practice. The paper still needs a corrected theorem (e.g., a genericity or probabilistic statement) and a numerical rank check over the actual channel ensemble before the overhead-reduction claim can be accepted. That is a major-revision requirement, not a definitive rejection of the empirical contribution, so I retain the reader's CONDITIONAL verdict.","tokens_in":19517,"tokens_out":19985,"duration_ms":180364,"concrete_test":"Analytically compute Rank(G_q) for the explicit 4x4 counterexample using (34)/(39): v1=(1,1,1,1), v2=(1,-1,1,-1), u1=(1,0,0,0), u2=(0,1,1,0)/sqrt(2), u3=(0,1,-1,0)/sqrt(2), u4=(0,0,0,1), lambda1=lambda2=1. If Rank(G_q)=3, Theorem 1 is falsified. To assess practical impact, also rerun the authors' condition-number experiment in Fig. 4 with Msub=32, NRF=16, Q=16, DFT and Hadamard v_b over the Section V near-field channel realizations, recording the rank/minimum eigenvalue of G_q for all q; any rank-deficient trial would invalidate the claimed M/NRF uniqueness in the intended regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's overhead reduction (Section IV-C, Table I) depends on Theorem 1, which asserts that B>=Bmin makes G_q full rank under the condition V(v_b1)!=V(v_b2). The proof via Lemma 1 is invalid: two distinct r-dimensional subspaces can intersect nontrivially, so the 2r vectors in Lemma 1 need not be linearly independent. The proof's step that the lower block C([r+1,Msub],:) has no all-zero row only shows each new vector has some component outside V(v_b1); it does not show those components are mutually independent, which would require that block to have full column rank r. The theorem is not merely unproved; it is false. Explicit counterexample: Msub=4, r=2, B=2, v1=(1,1,1,1), v2=(1,-1,1,-1) (two Hadamard columns), and a unitary U with u1=e1, u2=(0,1,1,0)^T/sqrt(2), completed arbitrarily. Set C=lambda1 u1 u1^H + lambda2 u2 u2^H with lambda_i>0. Then V(v1)=span(e1,u2) and V(v2)=span(e1,(0,-1,1,0)^T/sqrt(2)) are distinct, but their intersection is span(e1), so the four spanning vectors have rank 3. Consequently G_q has rank 3<Msub=4 while all hypotheses of Theorem 1 hold with B=Bmin. The claimed uniqueness guarantee behind the M/NRF overhead is therefore not established. The method may still work generically for the random near-field channels in Section V, but the paper provides no proof and its stated theoretical verification is invalid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-timescale channel estimation scheme, 2TCE-TSP, for RIS-aided near-field mmWave systems with hybrid beamforming. The key idea is to exploit the asymmetry between the long coherence time of the RIS-BS channel and the short coherence time of the user-RIS channel: the effective channel at block t is written as H_t^eff = H_0^eff diag(d_t), where d_t = diag(h_UR^0)^{-1} h_UR^t is a vector of length M. The initial effective channel is estimated once with the PW-CLRA method, and subsequent blocks are estimated by recovering only d_t from a piecewise beam training design. Each time block uses B subframes of Q pilots each, leading to a multiple-least-squares problem. The paper claims Theorem 1, which states that the Gram matrix of this problem is full rank when B >= B_min = max_q ceil(M/(Q min(N_RF, r_hat_q))) under a pairwise distinct-subspace condition, giving a per-block pilot overhead of about M/N_RF. Simulations show 38-78% pilot overhead reduction and roughly 98% complexity reduction at M=512 compared with the 2TCE-PWCLRA benchmark.","tokens_in":19946,"tokens_out":11658,"duration_ms":119317,"significance":"If the rank guarantee were valid, the proposed method would give a practically attractive reduction in pilot overhead and complexity for high-rank near-field channels, and the numerical comparisons across sparse, near-field, and Rayleigh fading channels are a useful empirical contribution. The paper also provides explicit formulas for pilot overhead and complexity and studies the effect of initial estimation error. However, the central theoretical result is false as stated: Lemma 1 is invalid, and Theorem 1 does not follow from its hypotheses. The distinct-subspace condition is neither verified for the proposed DFT/Hadamard training matrices nor sufficient for full rank. Because the M/N_RF overhead claim rests on this theorem, the theoretical support for the main contribution is currently missing. The underlying time-scaling idea and multi-LS formulation may still be salvageable, but the paper needs a corrected analysis or a clearly weakened claim.","major_comments":[{"comment":"Lemma 1 is false. Two distinct r-dimensional subspaces of C^{Msub} can intersect nontrivially, so the 2r vectors in the lemma need not be linearly independent. The proof's observation that C([r+1:Msub],:) has no all-zero row shows only that each diag(v_b2*)u_i has some component outside V(v_b1*); it does not show that those components are mutually independent. Consequently Theorem 1 is false as stated. A concrete counterexample is Msub=4, r=2, B=2, v1=(1,1,1,1), v2=(1,-1,1,-1), and C = lambda1 e1 e1^H + lambda2 u2 u2^H with u2=(0,1,1,0)^T/sqrt(2) and lambda_i>0. Then V(v1*) = span(e1,u2) and V(v2*) = span(e1,(0,-1,1,0)^T/sqrt(2)) are distinct, yet the four spanning vectors have rank 3, so Rank(G_q)=3<Msub while all hypotheses of Theorem 1 hold with B=Bmin. Thus the unique-solution guarantee for the multi-LS problem in (37), and hence the M/N_RF overhead reduction in Table I and Section IV-C, is not established.","section":"Section IV-A, Lemma 1 and Theorem 1"},{"comment":"The paper does not prove that the proposed beam training matrices (DFT or Hadamard columns) satisfy the condition V(v_b1*) != V(v_b2*) for all distinct b1,b2. More importantly, the counterexample above shows that even when this pairwise condition holds, the Gram matrix need not be full rank. A correct analysis must establish a stronger property, for example that the B subspaces V(v_b*) collectively span C^{Msub}, or must provide a different sufficient condition tied to the specific choice of vb. As written, the theoretical verification of the beam training design is absent.","section":"Section IV-A, distinct-subspace condition for DFT/Hadamard"},{"comment":"The small-timescale channel is defined as d_t = diag(h_UR^0)^{-1} h_UR^t, which requires every entry of h_UR^0 to be nonzero. The visual-region model in Section II-B sets entries of f to zero with probability 1-p; if the LoS component is blocked and the NLoS component is absent at that element, then h_UR^0(m)=0 and the time-scaling property is undefined. The paper neither states a nonzero-entry assumption nor adapts the estimation to the support of h_UR^0. Since the entire reduction to estimating d_t depends on this inverse, this issue must be addressed explicitly.","section":"Section III, Eq. (13) and Eq. (22)-(23)"}],"minor_comments":[{"comment":"The text refers to 'the optimal analog combiner WRB'; this appears to be a typo for WRF.","section":"Section III-B1"},{"comment":"The 'time-scaling property' in (12) is a definitional identity: d_t is chosen so that the factorization holds. The wording 'derive' overstates the content; please state explicitly that (12) is the definition of d_t, subject to the invertibility condition in (13).","section":"Section III, Eq. (12)"},{"comment":"Equation (44) uses the notation kappa(G_q) before kappa is defined in (52). Please define the condition number, including the log scaling, before it is used in the noise-amplification bound.","section":"Section III-B2, Eq. (44)"},{"comment":"The text says the large condition number at B=Bmin 'supports Theorem 1', but a large condition number is compatible with full rank and does not by itself confirm the rank. The interpretation should be softened to say that the condition number predicts practical estimation accuracy.","section":"Section IV-A, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the invalid Lemma 1 and Theorem 1. Since the counterexample uses Hadamard columns, which are explicitly proposed in the paper, the rank guarantee cannot be repaired by a minor wording change. I would ask the authors to either provide a correct sufficient condition for full rank and verify it for their training matrices, or explicitly downgrade the rank guarantee to an empirical observation and adjust the overhead claims accordingly. The time-scaling decomposition itself is definitional rather than a derived physical property, but the small-timescale estimation idea is still a reasonable contribution if the theoretical claims are corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a genuinely useful idea for slashing per-block pilot overhead in RIS-aided near-field systems, but the central rank guarantee (Theorem 1) is false as stated, and the proof of Lemma 1 is invalid. I would send it to review, but the authors need to fix the theory.\n\nThe core trick is the time-scaling property: within the RIS-BS coherence block, each effective channel factorizes as H_eff^t = H_eff^0 diag(d_t). Estimating the M-dimensional d_t instead of the full effective channel cuts the per-block pilot cost from M to about M/N_RF when combined with their piecewise beam training and a multi-LS solve. That is a real practical contribution for near-field mmWave/THz systems, and the simulation results back up the claimed 38-78% pilot overhead reduction and ~98% complexity reduction at M=512.\n\nThe soft spot is the theoretical verification. Theorem 1 asserts that if the subspaces V(v_b) are pairwise distinct and B>=Bmin, the Gram matrix G_q is full rank. That is not true. Two r-dimensional subspaces of an Msub-dimensional space can be distinct yet intersect non-trivially, in which case the union of their bases does not span Msub dimensions. The counterexample is small: Msub=4, r=2, B=2, v1=(1,1,1,1), v2=(1,-1,1,-1), and U with u1=e1, u2=(0,1,1,0)/sqrt(2). The subspaces are distinct but share e1, so the four spanning vectors have rank 3. Lemma 1's independence claim is therefore false; the correct condition for the 2r vectors to be linearly independent is V(v_b1)∩V(v_b2)={0}, which is stronger than distinctness. This means the paper's uniqueness guarantee for the multi-LS solution is unsubstantiated. The method may still work generically for the near-field channels in the simulations, but the paper does not prove it, and the claim of theoretical verification in the abstract is overstated.\n\nA smaller issue: the definition of d_t divides by h_UR^0 entrywise, which the VR model can zero out. The NLoS component may save you, but it deserves a line of discussion.\n\nNet: this is a promising algorithm with strong simulations and a broken proof. The right outcome is a major revision, not a desk reject and not acceptance in the current form. I would tag it for a careful referee who actually checks the linear algebra.","headline":"Promising overhead-reduction scheme for RIS near-field channel estimation, but the central full-rank theorem is false as stated and needs major revision.","tokens_in":20411,"tokens_out":7914,"would_cite":false,"duration_ms":61858,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"After one large-timescale estimate, a time-scaling identity lets every later RIS effective channel be recovered from roughly M/N_RF pilots per block, cutting pilot overhead 38–78% versus current methods.","keywords":["reconfigurable intelligent surface","near-field communications","channel estimation","two-timescale estimation","piecewise beam training","multiple least squares","hybrid beamforming","pilot overhead"],"falsifier":"One concrete test: with the paper's own parameters (M=512, Q=16, N_RF=16, r_q=16), form G_q from (39) using DFT or Hadamard beam vectors with B=B_min=2; Theorem 1 predicts rank(G_q)=32. If numerical rank is below 32 for any q, the theorem fails and the M/N_RF overhead claim is not guaranteed. The paper's own Fig. 4 shows exactly this failure for the sparse channel, where the condition number stays high near B_min.","tokens_in":19356,"feed_emoji":"📡","tokens_out":8603,"duration_ms":79583,"temperature":0.7,"pith_summary":"RIS-aided near-field systems have a quasi-static RIS-to-BS channel and a rapidly varying user-to-RIS channel, giving two very different coherence times. The paper's central claim is that once the initial effective channel is estimated on the slow timescale, every later effective channel can be written as that initial channel times a diagonal matrix built from an M×1 vector d_t, so only d_t needs to be estimated. It proposes piecewise beam training in which the RIS phases are structured so the observations become a multiple-least-squares problem in d_t, and proves (Theorem 1) a full-rank Gram matrix exists when the number of beam subframes B exceeds a computable threshold B_min, bringing per-block pilot overhead down to about M/N_RF. This is why a reader should care: the numbers translate to a 38–78% pilot-overhead reduction and about 98% lower computational complexity at M=512 relative to the baseline, and it works with half-duplex hybrid-array base stations rather than full-duplex operation.","feed_headline":"RIS channel estimation cuts pilot overhead 38–78%","feed_subtitle":"A time-scaling identity plus piecewise beams recovers fast user-RIS channels from M/N_RF pilots per block.","key_machinery":"The load-bearing object is the time-scaling identity H_t^eff = H_0^eff diag(d_t), combined with piecewise beam training. The beam training splits the RIS into Q pieces and assigns each piece a beam vector v_b drawn from a unitary matrix; after a unitary transform at the receiver, each subframe gives a measurement whose sensing matrix is W_RF H_{q,0}^eff diag(v_b). Summing the normal equations over B subframes produces the Gram matrix G_q, whose rank is analyzed through the identity G_q = (V_B V_B^H) ∘ ((W_RF H_{q,0}^eff)^H W_RF H_{q,0}^eff) and the Hadamard-product rank inequality. The theorem that G_q is full rank for B ≥ B_min is what turns the multi-LS solution from underdetermined to unique and justifies the M/N_RF pilot count.","core_discovery":"The discovery is a structural identity: for t in the RIS-BS coherence window, H_t^eff = H_0^eff diag(d_t), with d_t = diag($h_UR^{0}$)^{-1} h_UR^t, so the entire fast-timescale channel variation is contained in one vector per block. The paper builds a channel estimator around it. The RIS reflection vector is split into Q pieces, each using a column of a unitary matrix as the piecewise beam; the received signals over B subframes yield B linear equations z_{[b,q,t]} = W_RF H_{q,0}^eff diag(v_b) d_{[q,t]} + noise, and the paper solves them jointly as a multi-LS problem. Its Gram matrix is G_q = (V_B V_B^H) ∘ ((W_RF H_{q,0}^eff)^H W_RF H_{q,0}^eff), and Theorem 1 asserts G_q is full rank whenever B ≥ B_min = max_q ceil(M/(Q min(N_RF, r̂_q))) and the beam subspaces from distinct subframes are distinct. With that, the per-block pilot overhead is about M/N_RF, independent of the number of pieces once each piece has rank at least N_RF, and simulations show the resulting NMSE beats the two-timescale low-rank baseline at reduced overhead in near-field and Rayleigh channels.","pith_inferences":["Because the M/N_RF figure does not depend on Q once each piecewise rank reaches N_RF, the paper's own equations imply a clean operational tradeoff: pick Q for complexity, not for pilot count, and use the freed pilots for accuracy via B = 2B_min or 3B_min.","The same time-scaling-plus-piecewise-LS pattern should extend to any scenario where a slowly varying full matrix multiplies a fast-varying diagonal factor, such as a fixed reflector and a moving user; the paper does not explore this generalization.","The Gram-matrix ill-conditioning shown for sparse channels suggests the method is essentially tuned to high-rank near-field channels; a natural testable extension would make B adaptive to a condition-number threshold instead of a fixed multiple of B_min.","Using random phase vectors rather than deterministic DFT or Hadamard columns would make the subspace-distinctness condition in Theorem 1 hold with high probability, sidestepping the paper's unproven deterministic condition."],"forward_implications":["Per-block pilot overhead drops from M to about M/N_RF, so the saving grows with the number of RF chains and with RIS size.","The two-timescale strategy no longer requires full-duplex base stations; it runs on a hybrid-array half-duplex BS, which is the standard architecture.","At fixed transmit power, higher SNR buys a directly measurable pilot reduction: the simulations report NMSE 10^-2 at about 128 pilots for 20 dB and below 64 pilots for 30 dB.","Small-timescale estimation complexity is O(M^3/Q^2), so splitting the RIS into more pieces trades a little extra initial estimation work for a large complexity cut, around 98% at M=512.","Estimation accuracy in later blocks is limited by the initial large-timescale estimate; Fig. 8 shows performance degrades when that first estimate is imperfect, so the first-block estimator must be accurate to realize the paper's gains."],"supporting_citations":[{"why":"Supplies the PW-CLRA large-timescale estimator and the piecewise low-rank approximation used to obtain the initial estimated effective channel, and defines the baseline with per-block overhead M.","marker":"[22]"},{"why":"Introduces the two-timescale estimation strategy and the 2TCE-FD benchmark that this paper extends from full-duplex to half-duplex hybrid-array operation.","marker":"[23]"},{"why":"Provides the CLRA low-rank baseline whose complexity and overhead are compared in Table II and in the simulations.","marker":"[21]"},{"why":"Defines MIMO-ARD and MIMO-RD, the near-field distance thresholds and polar-domain sparsity that justify the high-rank effective channel model.","marker":"[14]"},{"why":"Provides the visual-region (VR) blockage model for the RIS-BS and User-RIS channels used in the system model.","marker":"[24]"},{"why":"Supplies the rank inequality for Hadamard products that bounds the Gram matrix rank and leads to the B_min condition.","marker":"[31]"}],"fun_headline_variants":["Piecewise beam training cuts RIS pilot overhead","Two-timescale identity speeds RIS channel estimation","Near-field RIS: one vector per block estimates fast channels","Time-scaling property enables low-cost RIS channel recovery","Beam splitting reduces pilots in RIS near-field systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire overhead saving rests on the claim that, for every pair of distinct beam-training subframes, the subspaces spanned by the r dominant rows of the measurement matrix are different, so their union spans the full M dimensions; the paper asserts this condition but does not prove it for the DFT or Hadamard matrices it actually uses, and the lemma supporting the proof does not establish linear independence of the combined vectors.","fun_headline_variants_meta":{"raw":{"variants":["Piecewise beam training cuts RIS pilot overhead","Two-timescale identity speeds RIS channel estimation","Near-field RIS: one vector per block estimates fast channels","Time-scaling property enables low-cost RIS channel recovery","Beam splitting reduces pilots in RIS near-field systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000299,"raw_usage":{"total_tokens":1763,"prompt_tokens":1017,"completion_tokens":746,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":673}},"tokens_in":633,"tokens_out":746,"duration_ms":6946,"temperature":1.0,"reasoning_tokens":673,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:59:16.569273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete test: with the paper's own parameters (M=512, Q=16, N_RF=16, r_q=16), form G_q from (39) using DFT or Hadamard beam vectors with B=B_min=2; Theorem 1 predicts rank(G_q)=32. If numerical rank is below 32 for any q, the theorem fails and the M/N_RF overhead claim is not guaranteed. The paper's own Fig. 4 shows exactly this failure for the sparse channel, where the condition number stays high near B_min.","supporting_citations":[{"cited_title":"Near-field los/nlos channel estimation for ris-aided mu-mimo systems: Piece-wise low-rank approximation approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the PW-CLRA large-timescale estimator and the piecewise low-rank approximation used to obtain the initial estimated effective channel, and defines the baseline with per-block overhead M."},{"cited_title":"Two-timescale channel estimation for reconfigurable intelligent surface aided wireless communications,","cited_arxiv_id":null,"evidence_quote":"Introduces the two-timescale estimation strategy and the 2TCE-FD benchmark that this paper extends from full-duplex to half-duplex hybrid-array operation."},{"cited_title":"Near-field channel estimation for xl-ris assisted multi-user xl-mimo systems: Hybrid beam- forming architectures,","cited_arxiv_id":null,"evidence_quote":"Provides the CLRA low-rank baseline whose complexity and overhead are compared in Table II and in the simulations."},{"cited_title":"Near-field channel estimation in mixed los/nlos envi- ronments for extremely large-scale mimo systems,","cited_arxiv_id":null,"evidence_quote":"Defines MIMO-ARD and MIMO-RD, the near-field distance thresholds and polar-domain sparsity that justify the high-rank effective channel model."},{"cited_title":"Channel estimation for xl-ris-aided millimeter-wave systems,","cited_arxiv_id":null,"evidence_quote":"Provides the visual-region (VR) blockage model for the RIS-BS and User-RIS channels used in the system model."},{"cited_title":"The singular values of a hadamard product: A basic inequality,","cited_arxiv_id":null,"evidence_quote":"Supplies the rank inequality for Hadamard products that bounds the Gram matrix rank and leads to the B_min condition."}],"review_version":1}