{"id":"38b29f7c-c152-42aa-a184-54c5284e23d8","arxiv_id":"2412.15965","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An architecture-independent ADMM framework optimizes beyond-diagonal RIS for arbitrary circuit topologies, matching or beating prior methods in sum-rate and transmit power at lower computational cost.","lead":"This paper introduces a single optimization framework that can tune any beyond-diagonal reconfigurable intelligent surface (BD-RIS), a wireless hardware technology that controls signal reflections, no matter how the surface's internal connections are arranged. Prior methods worked only for specific wiring patterns such as fully- or group-connected designs; this new approach handles any pattern and does so with lower computational cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's uniform lower bound on σ_min(G†U_t) is not merely unproven; it is incompatible with feasible problem instances in which two users share the same RIS-user channel, so the claimed universal convergence guarantee is not established for a legitimate class of BD-RIS systems.","rationale":"The reader's weakest assumption identifies the unproven boundedness and singular-value conditions in Theorems 1 and 2. My stress-test sharpens this to a concrete failure mode for the power-minimization algorithm. The paper's contribution is real: the admittance-matrix parametrization and the pp-ADMM construction are novel, and the derivations are largely coherent. The complexity analysis and simulation comparisons are plausible, and the extension discussions are reasonable. However, Theorem 2's assumption (b) is load-bearing: without it, Lemma 3's bound on μ_t collapses and the sufficient-decrease argument in Lemma 4 no longer yields convergence. The duplicate-user-channel counterexample is not exotic; it corresponds to two users at the same location or direction from the RIS, a standard scenario in wireless. In that case any limit point satisfying the bilinear constraint has σ_min(G†U) = 0, so the assumption cannot hold at a feasible limit. Consequently, the theorem's guarantee is vacuous for a feasible class of problems, and the paper's reliance on 'numerous simulations' to justify the assumption is an empirical remark on random Rician channels, not a property of the model. A simple computational test can determine whether the algorithm diverges, stagnates, or still converges on such an instance. If it still converges, the paper needs a convergence proof not relying on (b); if it diverges, the 'universal' claim is materially false for a feasible class. In either case, the current theorems overstate the guaranteed convergence, but the empirical framework may still be valuable; therefore I keep the reader's conditional verdict.","tokens_in":26434,"tokens_out":19841,"duration_ms":179347,"concrete_test":"Run the pp-ADMM power-minimization algorithm (26) on a feasible instance with M = 32, N = K = 4, setting h_1 = h_2 (two users at identical angle/distance from the RIS), drawing the other two user channels randomly, with SINR thresholds Γ_k = 0.1 and noise σ^2 = 1. Track σ_min(G†U_t), the bilinear residual ||(I - iZ_0B_t)U_t - (I + iZ_0B_t)H||_F, and the QoS-constraint satisfaction at every iteration. If the residual decreases toward zero while σ_min(G†U_t) also decays to zero, or if the iterates diverge, then assumption (b) of Theorem 2 fails on a feasible problem. Repeat with Γ = 0.01 and with Γ = 5 dB to probe feasible and infeasible regimes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a universally convergent, architecture-independent framework for the power-minimization problem rests on Theorem 2, whose assumption (b) requires σ_min(G†U_t) ≥ c > 0 uniformly in t. This condition is used in Lemma 3 to bound the dual multiplier μ_t through ((U^t)^†GG†U^t)^{-1} (Appendix C, Eq. (48)); without it, the sufficient-decrease argument in Lemma 4 collapses. The paper states after Theorem 1 and in Section III-C2 that such assumptions are 'imposed for technical reasons' and that simulations show they are 'always satisfied', but no proof is provided. The concern is stronger than 'unproven': there are feasible problem instances for which the condition must fail whenever the algorithm converges to feasibility. Take two users with identical RIS-user channels, h_1 = h_2 = h, with N = K = 4, and low SINR thresholds Γ < 1 (e.g., Γ = 0.1). Equal-power beamforming satisfies both QoS constraints, so the problem (5) is feasible. At any point satisfying the bilinear constraint (22c), U = Θ^†H has two equal columns, hence G†U has two equal columns and σ_min(G†U) = 0. If the pp-ADMM iterates converge to a feasible limit, the residual in (22c) tends to zero and σ_min(G†U_t) → 0, violating assumption (b) for all large t. Thus the theorem's hypothesis is incompatible with convergence on a feasible, physically meaningful channel realization, and the paper's broad claim of a universal framework with a theoretical convergence guarantee is not backed by the stated theory. The boundedness assumptions in Theorem 2(a) are a related but secondary gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a universal, architecture-independent optimization framework for beyond-diagonal RIS (BD-RIS) in MU-MISO systems, covering sum-rate maximization and transmit power minimization. The key modeling step is to replace the scattering matrix Θ with the admittance matrix B through the Cayley transform, and to encode any BD-RIS architecture as a sparsity constraint on B. The resulting nonconvex problems are reformulated via auxiliary variables (U = Θ†H and Y = U†GW) and solved with custom partially proximal ADMM (pp-ADMM) algorithms, with each subproblem solved in closed form or via a simple one-dimensional search. The paper states convergence to stationary points under boundedness and rank assumptions, gives complexity estimates, reports simulations comparing with PDD and two-stage FP for fully- and group-connected RISs, and extends the framework to general utility functions and MU-MIMO.","tokens_in":26808,"tokens_out":9462,"duration_ms":92412,"significance":"If the convergence claims were fully established, the framework would be a valuable contribution: it unifies BD-RIS architectures under a single optimization algorithm, avoids expensive matrix inversions through low-dimensional bilinear constraints, provides closed-form subproblem updates, and demonstrates a superior performance-efficiency trade-off relative to existing algorithms in the simulated regimes. The appendices contain detailed proofs, and the numerical comparison isolates the algorithmic gain rather than fitting parameters to data. The main weakness is that the central convergence theorems rest on assumptions that are not proven and, for the power-minimization theorem, are actually incompatible with a feasible class of channel realizations; this limits the claimed universality of the theoretical guarantee.","major_comments":[{"comment":"The boundedness assumptions in Theorems 1 and 2 (boundedness of {U_t}, {λ_t}, and, in Theorem 2, of {W_t} and {μ_t}) are stated as \"imposed for technical reasons\" and supported only by simulation evidence. While such assumptions are standard in nonconvex ADMM analyses, they should be presented as explicit limitations and, if possible, replaced by verifiable sufficient conditions. More importantly, for the power-minimization algorithm the rank condition of Theorem 2 cannot be guaranteed a priori even in simple feasible scenarios, as detailed in the previous comment. The revision should either prove the required conditions under stated channel/parameter assumptions, or substantially weaken the convergence claim and explicitly describe the class of instances to which the theorem applies.","section":"Section III-C2 and IV-C2"}],"minor_comments":[{"comment":"There are several typographical errors: \"catogarized\" (Section I), \"trackable\" (Section III-A), \"vise versa\" (Section III-A), \"dependance\" (Section III-B3), and \"decomposing ... into an easy-to-projection constraint\" (Section IV-A) which should read \"easy-to-project constraint.\"","section":"Abstract and throughout"},{"comment":"The sentence \"The complexity and convergence analysis of the proposed pp-ADMM algorithm is given in Section IV-C\" should refer to Section III-C; Section IV-C contains the analysis for the power-minimization algorithm.","section":"Section III-C"},{"comment":"The proof references \"Eq. (50)\" but the displayed equations in Appendix B-B are numbered (46); the cross-reference should be corrected.","section":"Appendix B-B"},{"comment":"The definition of S_i = {i, i+1, …, M} \\ {j | j > i, B_{i,j}=0} depends on the current iterate B, making the construction of the vector x architecture-dependent only through the current values of B. It should be defined from the fixed sparsity pattern of the BD-RIS architecture, not from the instantiated B.","section":"Section III-B4, Eq. (18)"},{"comment":"In the expression for r*_{k,k} when a_{k,k}=0, the denominator is typeset ambiguously as \"(1 + Γ_k)^2 + Γ_k σ^2\"; the parentheses should be clarified so that it is clear whether Γ_kσ^2 is added inside or outside the fraction.","section":"Appendix A, Eq. (29)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically solid in its algorithmic derivation and numerical evaluation, and the rank-deficiency issue is specific to Theorem 2 rather than a flaw in the overall ADMM approach. The authors should be given a chance to either repair the convergence argument (e.g., by handling rank-deficient G†U through pseudo-inverses or by adding a regularizer that provably maintains the full-rank condition) or to explicitly restate the guarantee in a narrower form. The example with identical RIS-user channels is easy to construct and should be discussed directly in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the bilinear reformulation of the scattering-matrix constraint: instead of optimizing over Θ with its unitary and symmetric structure, the paper writes (I − iZ0B)U = (I + iZ0B)H and lets the admittance sparsity pattern define the architecture. That is clean, and it makes the algorithm template genuinely architecture-independent. The pp-ADMM updates are worked out carefully, each subproblem is either closed-form or a simple bisection, and the complexity comparison against PDD is honest. The paper also delivers something the subfield lacked: a direct performance comparison of tree-, group-, and fully-connected architectures in multiuser systems, which previously required separate bespoke algorithms. Credit where due: the appendices contain complete proofs, the FP and proximal tricks are standard but applied competently, and the simulation results support the efficiency claims for the sum-rate problem.\n\nThe soft spots are concentrated in the power-minimization half. Theorem 2 assumes a uniform lower bound σ_min(G†U_t) ≥ c > 0 to control the dual multiplier. The paper says this is technical and always satisfied in simulations. It is not always satisfied: take two users with identical RIS-user channels, h1 = h2 = h, with N = K = 4 and a low SINR threshold. The problem is feasible, and any feasible limit point has U = Θ†H with two identical columns, so G†U is singular and σ_min = 0. If the ADMM iterates converge to feasibility, the singular value must tend to zero, violating the theorem's hypothesis. The universal convergence guarantee for the transmit power problem is therefore not established as stated. The boundedness assumptions in both theorems are also unproven, though that is a more standard gap and less alarming. On the reproducibility side, there is no code or data, no error bars, and the simulation setup omits several implementation parameters; the comparisons are believable but not independently checkable. The extensions to MU-MIMO and general utilities are sketches, which is fine for a paper that already has two full algorithms, but readers should not treat those as validated.\n\nNet assessment: this is a solid optimization paper with a real contribution and one load-bearing gap that is fixable. The convergence theory needs either a proof that the singular-value condition holds on the actual solution path, a restriction to generic channel realizations, or an explicit treatment of degenerate cases. With that repaired, the architecture-independent claim would be defensible. As it stands, the paper deserves peer review, not desk rejection, but a referee should push on Theorem 2.\n\nRecommendation: send to review, and ask the authors to address the convergence assumptions and release code.","headline":"A genuinely useful architecture-independent ADMM framework for BD-RIS, with solid derivations, but the power-minimization convergence theorem rests on an assumption that fails on feasible equal-channel instances.","tokens_in":27346,"tokens_out":2651,"would_cite":true,"duration_ms":28156,"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":"This paper proposes a single ADMM-based optimization framework that works for every beyond-diagonal RIS architecture, not just fully- or group-connected ones.","keywords":["beyond diagonal RIS","reconfigurable intelligent surface","architecture-independent optimization","admittance matrix","partially proximal ADMM","sum-rate maximization","transmit power minimization","MU-MISO"],"falsifier":"Run the pp-ADMM updates on a wide range of random Rician or Rayleigh channel realizations with the paper's own setup and monitor whether $U_t$ and $\\lambda_t$ stay bounded and whether $G^\\dagger U_t$ keeps a uniformly positive smallest singular value; if any realization produces divergence or a iterates that fail to approach a stationary point, Theorems 1 and 2 are falsified as stated.","tokens_in":26185,"feed_emoji":"📡","tokens_out":4980,"duration_ms":44737,"temperature":0.7,"pith_summary":"The paper claims to close a gap in reconfigurable-intelligent-surface (RIS) research: every prior optimization algorithm for beyond-diagonal RIS (BD-RIS) was tied to specific circuit topologies, such as fully-connected or group-connected architectures. The authors propose a framework in which any architecture, specified by the sparsity pattern of its admittance matrix, is optimized by the same partially proximal ADMM (pp-ADMM) algorithm. They apply it to sum-rate maximization and transmit-power minimization in multiuser MISO systems, and each update step either admits a closed form or is easy to solve. If the claim holds, wireless system designers could test new BD-RIS circuit layouts without designing a new optimizer for each one. The paper reports that the approach matches state-of-the-art performance at substantially lower CPU time.","feed_headline":"One optimizer now handles every beyond-diagonal RIS circuit","feed_subtitle":"Architecture becomes a sparsity mask on the admittance matrix, so the same ADMM solver serves sum-rate and power minimization.","key_machinery":"The central object is the susceptance matrix $B$ (the imaginary part of the admittance), whose sparsity pattern is the circuit architecture. The load-bearing mechanism is the exact reformulation of the scattering equation into the low-dimensional bilinear constraint $(I - iZ_0 B)U = (I + iZ_0 B)H$, combined with fractional programming for the sum-rate objective and an auxiliary variable $Y = U^\\dagger GW$ for the QoS constraints. These transformations turn a nonconvex problem with matrix-inversion constraints into a form ADMM can handle, and the proximal terms on $W$ and $B$ make those subproblems strongly convex, which is what the convergence proofs rely on.","core_discovery":"The central discovery is that the hard part of BD-RIS optimization, the non-diagonal scattering matrix $\\Theta$ with unitary and symmetry constraints, disappears once the problem is re-expressed through the admittance matrix $B$. Writing $\\Theta = (I + iZ_0 B)^{-1}(I - iZ_0 B)$ and constraining $B$ to be real symmetric with zeros wherever the circuit has no interconnection encodes every architecture as a sparsity pattern. The authors introduce $U = \\Theta^\\dagger H$ to replace the matrix-inversion constraint by the bilinear constraint $(I - iZ_0 B)U = (I + iZ_0 B)H$, cutting the constraint dimension from $M^2$ to $MK$ and making the problem tractable by ADMM. Their pp-ADMM algorithm, with proximal terms on $W$ and $B$, has each subproblem solvable in closed form or by a simple one-dimensional search, and Theorems 1 and 2 assert convergence to a stationary point under boundedness assumptions. This is claimed to be the first architecture-independent framework for BD-RIS optimization, and simulations show a better performance-versus-CPU-time trade-off than the PDD baseline [31] and the two-stage FP baseline [29].","pith_inferences":["Implicit in the paper but not pursued: because each architecture is just a binary mask on $B$, the framework could be used for architecture search, running the same pp-ADMM update over many masks and ranking topologies by the resulting sum-rate or transmit power.","A natural extension the authors leave open is adaptive or learned sparsity patterns, treating the interconnection mask as a discrete design variable that could be optimized jointly with the continuous susceptance values.","The CPU-time comparison against the PDD baseline could be sharpened by benchmarking against a dedicated solver for the PDD subproblems rather than a generic interior-point SOCP solver, which would test how much of the efficiency gain comes from the ADMM structure itself."],"forward_implications":["If the framework is right, any BD-RIS circuit topology, including tree- and forest-connected layouts whose scattering matrices have no simple structure, can be optimized with exactly the same algorithm by feeding in a sparsity mask.","The per-iteration complexity is $O(M^3 K + \\min\\{\\sum_i |S_i|, 2MK\\}^3)$, which for fully-connected RIS avoids the $O(M^6)$ per-iteration cost of the prior PDD approach when $M \\gg K$.","The paper's simulations show that tree-connected RIS, known to be optimal for single-user MISO, is no longer optimal in multiuser MISO, achieving performance comparable to group-connected RIS with group size 4 while using fewer impedances.","The same three-step recipe extends to max-min fairness, energy-efficiency maximization, and MU-MIMO systems."],"supporting_citations":[{"why":"Supplies the admittance-matrix characterization of BD-RIS circuit topologies and the tree-connected architecture model that the sparsity-mask formulation builds on.","marker":"[8]"},{"why":"Provides the universal multiport network analysis that justifies connecting the scattering matrix to the admittance matrix and the physical consistency constraints.","marker":"[32]"},{"why":"Introduces scattering-parameter modeling of BD-RIS and the fully-connected and group-connected architecture concepts that the paper generalizes.","marker":"[5]"},{"why":"The prior PDD framework for fully- and group-connected BD-RIS that serves as the main baseline and the only existing algorithm for transmit power minimization.","marker":"[31]"},{"why":"The heuristic two-stage fractional-programming baseline that the paper compares against for sum-rate maximization.","marker":"[29]"},{"why":"Supplies the fractional-programming transformation used to rewrite the sum-rate objective in a form suitable for ADMM.","marker":"[39]"},{"why":"The classical ADMM convergence framework that the paper adapts with proximal terms to handle the bilinear constraint.","marker":"[40]"},{"why":"The classical QoS-to-SOC reformulation that inspires the transformation of the power-minimization constraint.","marker":"[36]"}],"fun_headline_variants":["One ADMM solver handles every BD-RIS architecture","Admittance matrix turns BD-RIS into sparsity pattern","Universal BD-RIS optimizer: one solver, all circuits","Architecture becomes sparsity: universal BD-RIS solver","No more special-case BD-RIS: one ADMM fits all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence theorems assume that the algorithm's intermediate variables stay bounded (and, for power minimization, that $G^\\dagger U_t$ keeps a uniformly positive smallest singular value), but the paper offers no proof that these conditions always hold, only that simulations consistently satisfy them.","fun_headline_variants_meta":{"raw":{"variants":["One ADMM solver handles every BD-RIS architecture","Admittance matrix turns BD-RIS into sparsity pattern","Universal BD-RIS optimizer: one solver, all circuits","Architecture becomes sparsity: universal BD-RIS solver","No more special-case BD-RIS: one ADMM fits all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001117,"raw_usage":{"total_tokens":4709,"prompt_tokens":1061,"completion_tokens":3648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":3565}},"tokens_in":677,"tokens_out":3648,"duration_ms":22454,"temperature":1.0,"reasoning_tokens":3565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:54:44.948129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pp-ADMM updates on a wide range of random Rician or Rayleigh channel realizations with the paper's own setup and monitor whether $U_t$ and $\\lambda_t$ stay bounded and whether $G^\\dagger U_t$ keeps a uniformly positive smallest singular value; if any realization produces divergence or a iterates that fail to approach a stationary point, Theorems 1 and 2 are falsified as stated.","supporting_citations":[{"cited_title":"Beyond diagonal reconfig- urable intelligent surfaces utilizing graph theory: Modeling, architecture design, and optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the admittance-matrix characterization of BD-RIS circuit topologies and the tree-connected architecture model that the sparsity-mask formulation builds on."},{"cited_title":"A universal framework for multiport network analysis of reconfigurable intelligent surfaces,","cited_arxiv_id":null,"evidence_quote":"Provides the universal multiport network analysis that justifies connecting the scattering matrix to the admittance matrix and the physical consistency constraints."},{"cited_title":"Modeling and architecture design of reconfigurable intelligent surfaces using scattering parameter network analysis,","cited_arxiv_id":null,"evidence_quote":"Introduces scattering-parameter modeling of BD-RIS and the fully-connected and group-connected architecture concepts that the paper generalizes."},{"cited_title":"Optimizing power consumption, energy efficiency, and sum-rate using beyond diag- onal RIS—a unified approach,","cited_arxiv_id":null,"evidence_quote":"The prior PDD framework for fully- and group-connected BD-RIS that serves as the main baseline and the only existing algorithm for transmit power minimization."},{"cited_title":"A low-complexity beamforming design for beyond-diagonal RIS aided multi-user networks,","cited_arxiv_id":null,"evidence_quote":"The heuristic two-stage fractional-programming baseline that the paper compares against for sum-rate maximization."},{"cited_title":"Fractional programming for communication systems—part II: Uplink scheduling via matching,","cited_arxiv_id":null,"evidence_quote":"Supplies the fractional-programming transformation used to rewrite the sum-rate objective in a form suitable for ADMM."},{"cited_title":"Linear precoding via conic optimization for fixed MIMO receivers,","cited_arxiv_id":null,"evidence_quote":"The classical QoS-to-SOC reformulation that inspires the transformation of the power-minimization constraint."}],"review_version":1}