{"id":"975866d4-f8ba-40e9-8c14-1f698d705a29","arxiv_id":"2411.18480","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Q-stem connected BD-RIS with a least-squares scattering matrix design can match fully connected RIS sum channel gain at modest Q in simulated multi-user MISO systems.","lead":"Researchers propose a way to wire a reconfigurable intelligent surface where a small set of stem elements links to every other element, and they show in simulation that this can reach the performance of the most flexible design with fewer connections. They also give two algorithms for tuning the surface and say the scheme creates a useful trade-off between performance and circuit complexity for future wireless networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's Eq. (16) silently fixes the SVD-gauge phase Φ to identity; without a stated gauge normalization the LS method solves a gauge-dependent stricter condition, so the Q=2M-1 parity claim lacks a reproducible algorithmic basis.","rationale":"The reader correctly identified Eq. (16) as a weak point, but the issue is not merely that Φ is dropped without explanation. A diagonal phase matrix can be absorbed by reparameterizing the SVD of E, so setting Φ = I is theoretically harmless if the gauge is normalized. The load-bearing problem is that Algorithm 1 does not specify or perform this gauge normalization, so for the arbitrary SVD output from a standard routine, the LS algorithm solves a different, stricter condition. This makes the numerical evidence for the central claim potentially dependent on the SVD implementation. I credit the paper with a plausible new architecture, a correct upper-bound analysis under the standard descending-singular-value convention, and a valid demonstration that the relaxed upper bound is not achievable by symmetric unitary scattering matrices in general. The proposed concrete test of gauge-invariance would settle whether the phase omission affects the reported Q = 2M - 1 parity. Because the concern is about algorithmic reproducibility rather than a fundamental impossibility, the reader's CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":8996,"tokens_out":29011,"duration_ms":260122,"concrete_test":"Run Algorithm 1 on a fixed channel realization (e.g., L=5, K=3, N=64) exactly as written, recording the sum channel gain and the residual ||V_M^H Θ P_M - I||_F of the LS output. Then repeat the same experiment with a gauge-equivalent SVD: replace P_M by P_M D and W by W D, where D is a randomly generated diagonal unitary matrix, and likewise apply a diagonal unitary to V_M and U for H. Re-run Algorithm 1 on these modified SVD factors. If the resulting sum channel gain or the optimized scattering matrix changes by more than the Monte-Carlo noise (100 realizations), the algorithm is gauge-dependent and Eq. (16) is not equivalent to Eq. (12) as claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a Q-stem connected BD-RIS can attain the sum channel gain of a fully connected RIS, and this is evidenced primarily by the proposed algorithms, especially the LS design in Algorithm 1. The derivation of the LS linear system Eq. (16) from the sufficient condition Eq. (12) is not valid for an arbitrary SVD output. In Eq. (12) the target is V_M^H Θ P_M = Φ, where Φ is an arbitrary diagonal unitary. Because Θ is unitary, this condition is actually equivalent to Θ P_M = V_M Φ, and substituting the Cayley form of Θ gives B C = D with C = jZ0(P_M + V_M Φ) and D = P_M - V_M Φ. The paper instead sets Φ = I without comment. A diagonal phase can be absorbed by reparameterizing the SVD of E (replacing P_M by P_M Φ^{-1} and W accordingly), so the target Φ = I is without loss of generality only if the algorithm explicitly performs that gauge normalization before forming C and D. Standard SVD routines return arbitrary column phases, and for the returned V_M and P_M the phase Φ is generally not identity. As written, Algorithm 1 therefore solves the stricter condition V_M^H Θ P_M = I in the original gauge, which is not equivalent to Eq. (12) for fixed V_M, P_M. This makes the LS solution dependent on an unspecified gauge choice, so the simulated parity with fully connected RIS and the observed threshold Q = 2M - 1 may be artifacts of the particular SVD routine used rather than properties of the Q-stem architecture. Because the LS solution also initializes the quasi-Newton method, the gauge ambiguity propagates to the headline numerical result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new beyond-diagonal reconfigurable intelligent surface (BD-RIS) architecture, termed Q-stem connected RIS, which interpolates between single-connected, tree-connected, and fully-connected RIS by letting a set of Q ports connect to all other ports while remaining ports connect only to the first Q ports. The authors formulate a sum channel gain maximization problem for a multi-user MISO downlink, derive an SVD-based upper bound, and propose two algorithms: a closed-form least-squares (LS) design and an LS-initialized quasi-Newton method. Simulation results claim that with a specific Q (e.g., Q=2M-1 in one scenario) the Q-stem architecture attains the sum channel gain of fully connected RIS at reduced circuit complexity.","tokens_in":9346,"tokens_out":12746,"duration_ms":104608,"significance":"If the claims hold, the proposed architecture is a useful addition to the BD-RIS design space, offering a tunable complexity-performance trade-off. The circuit complexity expression QN + N - Q(Q+1)/2 is derived correctly and reduces to known values for Q=0, Q=1, and Q=N-1. The SVD-based upper-bound analysis is a standard and useful tool. The paper also shows the value of good initialization for quasi-Newton methods. However, the theoretical derivation of the LS algorithm contains a load-bearing algebraic error, so the significance of the algorithmic contribution is currently not established.","major_comments":[{"comment":"Equation (16) is not equivalent to Equation (12). Substituting the Cayley form (8b) into (12) with Phi = I gives V_M^H (I - jZ0B)(I + jZ0B)^{-1} P_M = I, which is not linear in B. In contrast, the equation B C = D with C = jZ0(V_M + P_M) and D = P_M - V_M is equivalent to (I - jZ0B) P_M = (I + jZ0B) V_M, i.e., to the stronger condition Theta P_M = V_M. The latter implies V_M^H Theta P_M = I but is not implied by it. The LS algorithm therefore solves a stricter alignment problem than the sufficient condition for the upper bound, so the stated derivation does not support Algorithm 1.","section":"III-B, Eq. (16)"},{"comment":"The diagonal phase matrix Phi in Eq. (12) is silently set to identity. For a general SVD output, the condition should be B(P_M + V_M Phi) = (P_M - V_M Phi)/(jZ0) (when expressed as the stronger linear form). Standard SVD routines return arbitrary column phases, so without an explicit gauge normalization of V_M and P_M, the LS solution depends on the SVD routine's choice of phases. The paper provides no such normalization, making the algorithm and the subsequent simulation results irreproducible with respect to the SVD gauge.","section":"III-B, Eq. (12)-(16)"},{"comment":"Proposition 1 states that fully connected RIS 'can not achieve' the upper bound when M > 1, but the proof only establishes that the symmetry condition Lambda = Lambda^T fails with probability 1 for generic channel realizations. There exist special channel matrices for which Lambda is symmetric, so the statement should be qualified as 'almost surely' or 'with probability 1'. This matters because the impossibility claim is used to motivate the LS approximation for all BD-RIS structures.","section":"Appendix, Proposition 1"},{"comment":"The least-squares solution b = (A^T A)^{-1} A^T z requires A^T A to be invertible. The paper does not prove that A has full column rank. The matrix A has size 2MN x (QN + N - Q(Q+1)/2), so for small Q the number of unknowns can exceed the number of equations, and for specific channel realizations rank deficiency can occur even when the dimensions are compatible. The authors should provide a rank analysis or replace the inverse with a regularized pseudo-inverse.","section":"Algorithm 1, Step 6"}],"minor_comments":[{"comment":"The heading reads 'least spare' and should be 'least squares'.","section":"Algorithm 1 heading"},{"comment":"The word 'connnected' is misspelled; it should be 'connected'.","section":"II-D"},{"comment":"The phrase 'a feasible susceptance matrice' should be 'a feasible susceptance matrix'.","section":"II-B"},{"comment":"The figures appear without axis labels or legends in the text; please add clear labels (e.g., 'Q' and 'sum channel gain') and a legend identifying each curve.","section":"Figures 3-5"},{"comment":"The DoF definition M = min(K,L,N) is presented as a limit; it would be clearer to state that M is the maximum rank of the effective channel F, which is at most min(K,L,N).","section":"III-A, Eq. (10)"},{"comment":"The phrase 'Without loss of generality, we set phi_m = 0' needs a justification, e.g., by the gauge freedom in the SVD of H^H and E. Without such a note, the reduction to phi_m = 0 appears unjustified.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the algebraic validity of Eq. (16). If the authors cannot provide a correct derivation of the LS algorithm, the paper's main algorithmic contribution is substantially weakened. The architecture itself is still of interest, and the quasi-Newton algorithm may remain useful even with a heuristic initialization, but the current manuscript overstates the theoretical support for the LS design. I would encourage the editor to ask for a rigorous re-derivation of the LS step and a clear statement of its relation to Eq. (12)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Q-stem connected RIS is a genuinely useful addition to the BD-RIS toolbox. The graph family cleanly interpolates between single, tree, and fully connected architectures, and the complexity formula is straightforwardly correct. The LS closed-form design is a reasonable way to target the SVD upper bound, and the idea of using it to initialize quasi-Newton is practical. The simulation figures show a smooth performance-complexity trade-off and a strikingly clean threshold at Q=2M-1 where Q-stem matches fully connected. That threshold is worth understanding, because right now it is purely empirical.\n\nThe main soft spot is the derivation of the LS equation. Eq. (16) claims that the condition V_M^H Θ P_M = Φ is equivalent to B C = D with C = jZ0(V_M+P_M), D = P_M-V_M. That is only true if Φ=I. You can absorb Φ by redefining P_M (a standard SVD gauge freedom), but the paper never says so. As written, Algorithm 1 solves the stricter condition V_M^H Θ P_M = I in whatever gauge the SVD routine happens to return. The resulting solution is gauge-dependent, and so is the parity claim. This is fixable with one sentence and a line of code, but it needs to be said.\n\nTwo smaller issues. Proposition 1 states that fully connected RIS cannot achieve the upper bound for M>1, but the proof shows only 'with probability 1' for a fixed gauge; the qualifier should be in the proposition. And the LS solution in (21) assumes A^T A is invertible without a rank argument.\n\nNone of this is fatal. The architecture is real, the complexity counts are right, and the empirical case is plausible. But without the phase fix and without code/data, the headline Q=2M-1 claim is an unverified empirical observation. I would send this to review, not desk reject; the authors can address these in a revision. The paper will be a useful reference for BD-RIS designers even if the threshold does not survive contact with a different SVD gauge.","headline":"Q-stem is a useful new BD-RIS topology, but the LS derivation drops the SVD phase and the headline parity claim is empirical.","tokens_in":9881,"tokens_out":8543,"would_cite":true,"duration_ms":74899,"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":"For a multi-user MISO downlink, a Q-stem connected BD-RIS reaches the sum channel gain of a fully connected RIS when Q = 2M − 1 (M the number of independent streams), using far fewer tunable admittances.","keywords":["beyond-diagonal reconfigurable intelligent surface","Q-stem connected RIS","sum channel gain maximization","scattering matrix design","least squares algorithm","quasi-Newton method","multi-user MISO","graph-based RIS architecture"],"falsifier":"Compute, for generic random channel realizations with M > 1, whether there exists any diagonal unitary Φ such that the Q-stem constraint set B_Q contains a symmetric B satisfying (I + jZ_0 B)^{-1}(I − jZ_0 B) = V_M Φ P_M^H + V_{N−M} X P_{N−M}^H; if the answer is no while the simulation still shows parity, then the observed gain parity is not explained by the stated derivation. A simpler direct test is to simulate N=128, L=K=8 with Q=2M−1 and compare the least-squares design against fully connected RIS over many channel realizations; any systematic gap refutes the parity claim.","tokens_in":8752,"feed_emoji":"📶","tokens_out":6811,"duration_ms":57579,"temperature":0.7,"pith_summary":"This paper proposes a new beyond-diagonal reconfigurable intelligent surface (BD-RIS) topology, the Q-stem connected RIS, in which the first Q ports connect to every other port while the remaining ports connect only to those Q ports. The authors argue this family interpolates between single-connected, tree-connected, and fully connected RIS, and that in an L-antenna base station serving K single-antenna users it can match the sum channel gain of a fully connected surface once Q = 2M − 1, where M is the number of independent data streams. To design the surface, they give a closed-form least-squares algorithm that approximately satisfies the channel-alignment condition V_M^H Θ P_M = Φ, plus a quasi-Newton refinement initialized from that least-squares solution. The practical claim is that near-full performance can be bought with far fewer tunable impedance components and much lower computational cost than full connectivity.","feed_headline":"Q-stem RIS matches fully connected gain at Q = 2M − 1","feed_subtitle":"A graph-based RIS topology with fewer tunable parts still reaches full-RIS sum channel gain in multi-user MISO.","key_machinery":"The load-bearing object is the Cayley-type scattering matrix Θ = (I_N + jZ_0 B)^{-1}(I_N − jZ_0 B), where the symmetric real susceptance matrix B encodes which ports are connected by tunable admittances. For Q-stem connected RIS the constraint is [B]_{n,m} = 0 whenever both indices exceed Q, so B is dense only in the first Q rows and columns. The design machinery is the independent vectorization veci(B) = b with a 0/1 assembly matrix R satisfying vec(B) = R b, which turns the channel-alignment condition B C = D into the linear system A b = z; the least-squares inverse (A^T A)^{-1} A^T z gives a closed-form scattering design, and the same b initializes the quasi-Newton method on the original non-convex problem.","core_discovery":"Working in a multiuser MISO downlink where the BS-RIS channel E and the RIS-user channels H are given, the paper shows that the sum channel gain ∥H^H Θ E∥$_F^{2}$ is upper-bounded by ∥S_M Σ_M∥$_F^{2}$ via the SVD of H^H and E, and that this bound would require the channel-alignment identity V_M^H Θ P_M = Φ with Φ a diagonal unitary matrix. It then proves that when the multiplexing gain M > 1, no reciprocal BD-RIS, including the fully connected one, can satisfy that identity exactly, so full connectivity is not the bound but the practical gold standard. The proposed Q-stem connected RIS—where the first Q ports connect to all ports and the last N − Q ports connect only to the hub—is shown in simulation to match the fully connected sum channel gain when Q = 2M − 1, using only QN + N − Q(Q+1)/2 tunable admittances instead of N(N+1)/2. The design is carried by a least-squares solution of the linearized identity B C = D and, when higher accuracy is wanted, a quasi-Newton refinement initialized from that least-squares point.","pith_inferences":["The Q = 2M − 1 rule, if it survives other settings, gives hardware designers a direct recipe: interconnect roughly twice as many hub ports as the number of spatial streams, independent of the total element count.","The same vectorization-and-least-squares machinery should transfer to any BD-RIS topology whose susceptance support is fixed, since only the 0/1 matrix R changes; multi-sector BD-RIS and STAR-RIS are natural next targets.","A clean proof that the phase matrix Φ can be absorbed into the SVD factors would upgrade the parity observation from simulation to theorem; without it, the reported Q=2M−1 law should be read as empirical.","The analysis assumes blocked direct links and Rayleigh fading; with strong direct paths or correlated channels the optimal Q may shift, so the threshold is worth testing beyond the current model."],"forward_implications":["When Q is fixed, the circuit complexity QN + N − Q(Q+1)/2 grows linearly with N, so the same surface can be scaled to many elements without quadratic hardware growth.","At the threshold Q = 2M − 1, the simulated sum channel gain of Q-stem connected RIS equals that of fully connected RIS, while the number of tunable admittances drops from N(N+1)/2 to roughly (2M−1)N.","The least-squares design is closed-form with O(Q^3 N^3) complexity, and using it to initialize quasi-Newton beats random initialization; for Q ≥ 7 the least-squares method alone matches the quasi-Newton result in the simulated N=64, L=K=4 setup.","Because Q=0, Q=1, and Q=N−1 reproduce single-connected, tree-connected, and fully connected RIS, any performance gain found for intermediate Q is a smooth interpolation among those existing architecture classes."],"supporting_citations":[{"why":"Provides the scattering-parameter network model and the formula Θ = (I_N + jZ_0 B)^{-1}(I_N − jZ_0 B) that the entire architecture and optimization build on.","marker":"[2]"},{"why":"Introduces graph-theoretic tree connected RIS and the style of susceptance-matrix constraints that Q-stem connected RIS generalizes.","marker":"[5]"},{"why":"Gives the M=1 special case where the SVD upper bound is reachable, which the appendix contrasts with the M>1 impossibility result.","marker":"[10]"},{"why":"Supplies an existing low-complexity beamforming design for multi-user fully connected BD-RIS that the proposed algorithms extend and compare against.","marker":"[11]"},{"why":"Provides the degree-of-freedom channel-shaping analysis and the upper bound ∥S_M Σ_M∥_F used as the theoretical target in (11).","marker":"[13]"},{"why":"Supplies the standard least-squares solution formula used in the closed-form scattering design (21).","marker":"[14]"},{"why":"Provides the quasi-Newton optimization approach and the simulation channel parameters used to evaluate the proposed LS-based initialization.","marker":"[15]"}],"fun_headline_variants":["Q-stem RIS matches full gain at Q=2M−1","Full sum gain at Q=2M−1 with fewer tunable parts","Q-stem connected RIS: full performance, lower complexity","Match full-RIS sum gain with 2M−1 port hub","BD-RIS trade-off solved: Q-stem hits full gain at Q=2M−1"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the phase matrix Φ in the sufficient condition V_M^H Θ P_M = Φ can be set to identity when the condition is rewritten as B C = D, so that the least-squares solution targets exactly the condition that guarantees the upper bound rather than a stricter one.","fun_headline_variants_meta":{"raw":{"variants":["Q-stem RIS matches full gain at Q=2M−1","Full sum gain at Q=2M−1 with fewer tunable parts","Q-stem connected RIS: full performance, lower complexity","Match full-RIS sum gain with 2M−1 port hub","BD-RIS trade-off solved: Q-stem hits full gain at Q=2M−1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000318,"raw_usage":{"total_tokens":1838,"prompt_tokens":1030,"completion_tokens":808,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":708}},"tokens_in":646,"tokens_out":808,"duration_ms":6764,"temperature":1.0,"reasoning_tokens":708,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:09:38.141790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for generic random channel realizations with M > 1, whether there exists any diagonal unitary Φ such that the Q-stem constraint set B_Q contains a symmetric B satisfying (I + jZ_0 B)^{-1}(I − jZ_0 B) = V_M Φ P_M^H + V_{N−M} X P_{N−M}^H; if the answer is no while the simulation still shows parity, then the observed gain parity is not explained by the stated derivation. A simpler direct test is to simulate N=128, L=K=8 with Q=2M−1 and compare the least-squares design against fully connected RIS over many channel realizations; any systematic gap refutes the parity claim.","supporting_citations":[{"cited_title":"Modeling and architecture design of reconfigurable intelligent surfaces using scattering parameter network analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the scattering-parameter network model and the formula Θ = (I_N + jZ_0 B)^{-1}(I_N − jZ_0 B) that the entire architecture and optimization build on."},{"cited_title":"SNR maximization in beyond diagonal RIS-assisted single and multiple antenna links,","cited_arxiv_id":null,"evidence_quote":"Gives the M=1 special case where the SVD upper bound is reachable, which the appendix contrasts with the M>1 impossibility result."},{"cited_title":"A low-complexity beamforming design for beyond-diagonal RIS aided multi-user networks,","cited_arxiv_id":null,"evidence_quote":"Supplies an existing low-complexity beamforming design for multi-user fully connected BD-RIS that the proposed algorithms extend and compare against."},{"cited_title":"Fully connected reconfigurable intelligent surface aided rate-splitting multiple access for multi-user multi-antenna transmission,","cited_arxiv_id":null,"evidence_quote":"Provides the quasi-Newton optimization approach and the simulation channel parameters used to evaluate the proposed LS-based initialization."}],"review_version":1}