{"id":"f5969933-3b90-44ea-9172-f11c67839f9a","arxiv_id":"2504.14584","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Two max-min fairness algorithms, built from geometric programming and gradient descent/ascent, are proposed for SIM-assisted multi-user MISO downlinks, with a statistical-CSI upper bound that is tight at low SNR.","lead":"This paper designs two algorithms that allocate power and adjust stacked-intelligent-metasurface phase shifts to give all users in a multi-user wireless downlink a fair minimum data rate. One algorithm uses exact channel knowledge; the other uses only channel statistics to cut overhead.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (15b) omits the user's own power p_k from the GP constraint, so the printed power-allocation step does not solve (P2).","rationale":"The reader identified the GDA convergence assumption as the weakest point, and that concern is legitimate: Algorithm 2 has no convergence guarantee for the nonconvex max-min subproblem. However, the concrete algebraic error in (15b) is more immediately checkable and directly affects the power-allocation half of the alternating algorithm. If the GP constraint as printed were used, the power update would not maximize the minimum SINR, and the claimed order-of-magnitude gains would be unsupported regardless of GDA behavior. The likely explanation is a typographical omission of p_k and a subscript error, and the intended GP is standard; with that correction the central approach may survive. Therefore the reader's CONDITIONAL verdict remains appropriate: the manuscript should be revised to state the correct constraint, and the numerical implementation should be verified or released so the reported curves can be reproduced. This does not move the verdict, but it narrows the required revision and supplements the reader's GDA-based objection with a specific falsifiable check.","tokens_in":20570,"tokens_out":18986,"duration_ms":183301,"concrete_test":"Independently re-derive (P2.2) from (P2.1) and check whether the right-hand side of (15b) contains p_k. Then solve both forms for a small instance, e.g., K=2 with fixed G, using the printed constraint and the corrected constraint sum_{j!=k} s_{k,j} p_j + sigma_k^2 <= s_{k,k} p_k / t. If the two optimal t values differ, the printed GP is not solving (P2.1); the numerical implementation must be inspected to confirm whether the correct p_k-dependent constraint was used to generate Figures 3-9.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the instantaneous-CSI algorithm, the GP reformulation (P2.2) is the power-allocation step that underpins every numerical min-rate claim. Starting from (P2.1), the constraint t <= |h_k^H G w_k|^2 p_k / (sum_{j!=k} |h_k^H G w_j|^2 p_j + sigma_k^2) must be rewritten as sum_{j!=k} s_{k,j} p_j + sigma_k^2 <= s_{k,k} p_k / t, with s_{k,k}=|h_k^H G w_k|^2. The displayed (15b) instead has s_{k,k}/t with no p_k in the numerator, and the surrounding definition gives s_{k,k}=|h_k^H G w_j|^2, using w_j instead of w_k. As written, the user's own transmit power does not improve its SINR constraint, so the GP step is not equivalent to (P2.1); if implemented literally, Algorithm 1 is solving a different, meaningless power-allocation problem. The statistical-CSI GP step (P6.1) is not affected, but the instantaneous-CSI results and the claim that wave-based beamforming dominates power allocation rely on the corrected version of (15b).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies a downlink stacked intelligent metasurface (SIM)-assisted multiuser MISO system and develops two max-min fairness algorithms. For the instantaneous-CSI case, the problem (P1) is decomposed into a geometric-programming power-allocation step and a gradient descent-ascent (GDA) wave-based beamforming step, followed by discrete phase quantization. For the statistical-CSI case, the average minimum achievable rate is upper-bounded by replacing each user's SINR with its SNR, yielding a deterministic objective (P6) that is minimized via alternating GP and gradient descent; the resulting policy is evaluated against an exhaustive search. Numerical simulations report substantial gains over equal-power/random-phase benchmarks and a tight upper bound in the low-SNR regime.","tokens_in":20866,"tokens_out":9532,"duration_ms":91366,"significance":"The problem is timely, and the statistical-CSI max-min fairness formulation for SIM-assisted systems appears to be new. The analytical building blocks—the gradient derivations in Lemmas I and II and the bounding chain in Eqs. (38)-(46)—are detailed and mostly correct, and the paper gives a clear comparison with exhaustive search for the statistical case. If the issues below are fixed, the claimed order-of-magnitude improvements and the fairness comparisons would be useful for the SIM literature. However, the paper does not provide convergence guarantees for the nonconvex subproblems, and the printed GP step is not equivalent to the stated power-allocation problem, so the numerical claims are not yet fully grounded.","major_comments":[{"comment":"The GP reformulation (P2.2) is not equivalent to (P2.1). Starting from (14b), the constraint must read sum_{j≠k} |h_k^H G w_j|^2 p_j + sigma_k^2 ≤ |h_k^H G w_k|^2 p_k / t, that is, with s_{k,j}=|h_k^H G w_j|^2, sum_{j≠k} s_{k,j} p_j + sigma_k^2 ≤ s_{k,k} p_k / t. The printed (15b) has s_{k,k}/t on the right-hand side with no p_k, and the line defining s_{k,k} uses |h_k^H G w_j|^2 instead of |h_k^H G w_k|^2. As written, Algorithm 1's power-allocation step does not solve (P2.1), and because every instantaneous-CSI result in Section V-A relies on alternating this step with the GDA phase update, the numerical claims cannot be traced to the stated problem. This must be corrected, and the simulation implementation should be checked to confirm that it includes p_k in the numerator of the SINR constraint.","section":"Section III-B1, Eqs. (14)-(15)"},{"comment":"No convergence guarantee is provided for the GDA loop. The objective f(lambda, theta) in (16a) is nonconvex in theta and only linear in lambda, so the two-timescale step-size rule epsilon = tau*mu with tau=10, described after Eq. (25), has no established convergence guarantee for this problem class; the cited reference [39] requires structural conditions that are not verified here. Since the central numerical claim—an order-of-magnitude improvement over equal-power/random-phase benchmarks—depends on Algorithm 2 reaching a useful point, the authors should either supply a convergence analysis under explicit assumptions or clearly state that convergence is empirical and scope the claims accordingly. Figure 9 alone, for one configuration, is not a substitute for such a statement.","section":"Section III-B2, Algorithm 2 and Problem (P3.1)"},{"comment":"The displayed equality tilde_R(tilde_theta^*, tilde_p^*) = exp(zeta sigma^2) E_1(zeta sigma^2) drops the factor log_2 e that appears on the preceding line. If the statistical-CSI curves in Fig. 10 evaluate (46) literally, the upper-bound curve is scaled by 1/log_2 e, which would invalidate the claimed tightness against the exhaustive search of the actual average rate in bps. The factor should be restored, and the simulation should be checked to ensure it uses the corrected expression.","section":"Section IV, Eq. (46)"}],"minor_comments":[{"comment":"The simulation setup says 'The BS has an M-antenna array along the x-axis,' but the system model in Section II uses N transmit antennas, with N=K in the simulations; this should be N-antenna array for consistency.","section":"Section V, system setup"},{"comment":"The symbol epsilon is used both as the bisection accuracy in the lambda projection (after Eq. (24)) and as the GDA step size in (17a) and (25a), and the input list includes both epsilon and tau; renaming one of these quantities would remove a serious notational clash.","section":"Algorithm 2 and Section III-B2"},{"comment":"Algorithm 1 says 'Repeat steps 2 and 3 until convergence,' but no stopping criterion is specified; the convergence plots in Figs. 9 and 12 should state the actual criterion used (for example, relative change in the minimum rate).","section":"Algorithm 1 and Fig. 9"},{"comment":"The claim that g(eta) is monotonic in eta is stated without proof, and the displayed bounds on eta* are not derived; a short monotonicity argument and derivation of the bisection interval would make the projection step self-contained.","section":"Section III-B2, Eq. (24)"},{"comment":"The benchmark names are described in the text, but the legends in Figs. 3-8 are not all visible in the typeset version; please ensure every curve is labeled in the figure files so that the reported gains can be verified.","section":"Section V-A, Figs. 3-8"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, workmanlike extension of RIS max-min fairness to stacked intelligent metasurfaces, with two real contributions (a statistical-CSI upper bound and its GP/GD solution) and one load-bearing typo that must be fixed before anyone implements the instantaneous-CSI algorithm.\n\nThe gradient derivations in Lemmas I and II are algebraically sound as far as I checked. The statistical-CSI bound in (38)-(46) is genuinely new: dropping interference and using the independence of channel vectors to get a product of exponentials is standard but applied cleanly, and the exponential-integral identity in (46) is correct. The numerical claim that the bound is tight in the low-SNR regime is consistent with the fact that interference is small there. Credit also for giving complexity expressions and for testing against exhaustive search on the small case.\n\nSoft spots, in order of severity:\n\n1. Equation (15b) is wrong as printed. The constraint from (14b) should become sum_{j≠k} s_{k,j} p_j + σ_k^2 ≤ s_{k,k} p_k / t, with s_{k,k} = |h_k^H G w_k|^2. The text has s_{k,k}/t with no p_k in the numerator, and the definition of s_{k,k} uses w_j instead of w_k. This is likely a typo, but it is load-bearing: the GP step is what produces every instantaneous-CSI min-rate number in Figures 3-9. A literal implementation of (15b) solves a different problem where a user's own power never appears in its own SINR constraint. The authors need to correct this and ideally confirm the simulations used the corrected form.\n\n2. Algorithm 2 (GDA) has no convergence guarantee. For a nonconvex-nonconcave problem, the two-timescale step-size heuristic (ε > μ, τ=10) is reasonable but unsupported. The paper would be stronger with a critical-point convergence proof or at least a sensitivity study over step sizes.\n\n3. The novelty claim is overstated. The introduction says max-min fairness in SIM is 'unexplored' and 'largely unaddressed,' but reference [21] already does max-min spectral-efficiency power control for SIM cell-free massive MIMO. What is new here is the statistical-CSI bound and the specific GP+GDA combination for the MISO downlink, not the mere fact of fairness in SIM.\n\n4. Minor: no code or data, so the numerical results are not independently reproducible. The bound is validated only against an exhaustive search over the same simulation model.\n\nThe math core is honest and the derivations are mostly self-contained. This deserves a serious referee, but the referee should demand a fix to (15b) and a softening of the novelty claims. If the typo is corrected, I'd cite the statistical-CSI part.","headline":"Solid SIM max-min fairness paper with a correctable but load-bearing typo in the GP power-allocation step; the statistical-CSI bound is the real contribution.","tokens_in":21380,"tokens_out":2628,"would_cite":true,"duration_ms":22417,"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":"The paper claims that fairness in a SIM-assisted multi-user downlink can be optimized by alternating geometric programming for power allocation with gradient-based phase updates, giving order-of-magnitude minimum-rate gains over…","keywords":["stacked intelligent metasurfaces","max-min fairness","wave-based beamforming","geometric programming","gradient descent-ascent","statistical CSI","multi-user MISO","discrete phase shifts"],"falsifier":"Run the proposed algorithm on a small SIM configuration, for example L=2 and M=4, with many random phase initializations and compare against exhaustive search over the discrete phase set; if the gradient descent-ascent phase update yields final minimum rates that vary widely or sit far below the exhaustive optimum, the central claim that the alternating algorithm reaches the max-min operating point is not supported. Independently, evaluate the true average minimum rate for the statistical-CSI policy by Monte Carlo at mid and high SNR; if the gap between the proposed bound and the simulated truth grows large, the claim that the bound is a good stand-in fails.","tokens_in":20428,"feed_emoji":"📶","tokens_out":7767,"duration_ms":69360,"temperature":0.7,"pith_summary":"This paper targets the weakest user in a multi-user downlink where one base station serves several single-antenna users through a stacked intelligent metasurface (SIM), a stack of reconfigurable layers that shapes signals in the wave domain before they leave the antenna. The authors try to establish that the maximum of the minimum user rate, the max-min fair operating point, can be found by alternating two cheap steps: geometric programming for power allocation and gradient-based updates for the metasurface phase shifts. They handle two levels of channel knowledge: perfect instantaneous channel state information, and statistical knowledge given only the channel covariance. If their algorithms are right, the weakest user's rate is an order of magnitude above simple equal-power and random-phase baselines, and all users get nearly equal rates. They also show that with statistical CSI, an upper bound on the average minimum rate can be optimized instead of the intractable exact average, and that this bound is tight at low signal-to-noise ratio.","feed_headline":"Stacked metasurface tuning lifts the weakest user's rate 50-fold","feed_subtitle":"Two algorithms tune power and phases, with instant channels or statistics only, to serve every user fairly.","key_machinery":"The object carrying the argument is the SIM's wave-based beamforming matrix $G_\\vartheta = \\left(\\prod_{k=1}^{L-1}\\Theta_{L+1-k}W^{(L+1-k)}\\right)\\Theta_1$, the product of inter-layer diffraction matrices $W^{(\\ell)}$ and diagonal phase-shift matrices $\\Theta_\\ell$, with the phase entries as the optimization variables. The solution mechanism is alternating optimization: an epigraph variable converts the power-allocation problem into standard geometric programming, while the phase problem is handled by defining a weighted sum of SINRs and updating the weights and phases with gradient descent-ascent, with the discrete phase constraint enforced by final quantization. For the statistical-CSI variant, the key step is bounding the average minimum rate by an integral over products of exponential complementary CDFs, which turns the stochastic problem into the minimization of a sum of reciprocal terms whose gradients with respect to each phase are derived in closed form.","core_discovery":"The central claim is that a SIM-assisted multi-user MISO downlink can be operated fairly by maximizing the minimum SINR, and that the joint optimization splits cleanly: for fixed phase shifts, the power-allocation subproblem is a geometric program; for fixed powers, the phase-shift subproblem can be attacked by relaxing discrete phases to continuous values and applying gradient descent-ascent on a min-max surrogate. Numerical experiments then show that the resulting scheme gives roughly 45 to 60 times the minimum rate of equal power with random phase shifts, about 20 times the rate of optimal power with random phases, and about 1.2 to 1.3 times the rate of equal power with optimized phases; fairness indices stay near 1 up to 10 users. In the statistical-CSI variant, the exact average minimum rate is replaced by an upper bound obtained by replacing each user's SINR by its SNR, whose distribution is exponential; minimizing the sum of reciprocal SNR terms yields the power and phase policies, and this bound is tight in the low-SNR regime. The paper also reports that the wave-based beamforming optimization contributes far more to the gains than the power allocation, and that eight-bit phase quantization recovers essentially the continuous-phase performance.","pith_inferences":["A natural next test is whether the reported 45-to-60x gains survive imperfect or delayed channel estimation, since the instantaneous-CSI algorithm assumes error-free channels in every coherence interval; the statistical-CSI algorithm should degrade more gracefully because it only needs the covariance.","The same alternating geometric-programming-plus-gradient recipe likely transfers to SIM designs with amplitude control or inter-element coupling, because the power subproblem stays polynomial and the phase objective stays smooth.","The low-SNR tightness of the bound suggests an operating rule the authors do not state explicitly: use statistical-CSI max-min optimization in power-limited or feedback-limited regimes, and reserve instantaneous-CSI optimization for high-SNR periods.","Near-unit fairness indices up to ten users suggest the max-min policy does not merely rescue the worst user but equalizes the whole rate distribution; plotting full rate histograms would expose any middle-user sacrifice."],"forward_implications":["A SIM can be configured for near-equal user rates using only its passive phase layers plus a power split, so the fairness burden moves from digital precoding to the wave domain.","With instantaneous CSI the configuration must be recomputed per coherence interval, whereas with statistical CSI one configuration serves many coherence intervals and approximates the average rate well at low SNR.","Eight-bit phase quantization recovers essentially the continuous-phase performance, which matters for hardware because it limits feedback and control complexity.","The power-allocation step alone captures only a small part of the gain; most of the fairness improvement comes from optimizing the phase profile of the SIM.","The max-min formulation avoids exhaustive search over $2^{bML}$ phase combinations, making the problem tractable at realistic layer and element counts."],"supporting_citations":[{"why":"It supplies the Rayleigh-Sommerfeld inter-layer propagation model and the layered form of the SIM wave-based beamforming matrix.","marker":"[2]"},{"why":"It provides the prior statistical-CSI optimization approach for SIMs that motivates replacing per-coherence optimization with covariance-based policies.","marker":"[14]"},{"why":"It is the earlier SIM power-and-beamforming design that the instantaneous-CSI algorithm extends from sum rate to max-min fairness and whose projected-gradient update serves as a comparator.","marker":"[15]"},{"why":"It is the source of the min-max surrogate equivalence for max-min SNR and of the gradient descent-ascent treatment of discrete phase shifts.","marker":"[30]"},{"why":"It supplies the geometric-programming formulation and complexity estimate for max-min power control under an alternating optimization structure.","marker":"[31]"},{"why":"It is the standard reference for geometric programming, KKT-based projection, and backtracking line search used in the algorithms.","marker":"[38]"},{"why":"It provides the two-timescale gradient descent-ascent theory whose step-size rule, epsilon greater than mu and tau equal to 10, the phase update adopts.","marker":"[39]"}],"fun_headline_variants":["Stacked metasurface max-min fairness lifts worst user 50x","Max-min fairness in stacked metasurface MISO: 50x for weakest user","Stacked metasurface lifts worst user's rate 50x with max-min","Fair beamforming with stacked metasurface: 50x gain for worst user"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole performance claim depends on the phase-update loop reliably finding a good point of a difficult nonconvex problem; the paper gives no convergence guarantee and borrows its step-size rule from a heuristic, so a stalled or cycling run would erase the gains.","fun_headline_variants_meta":{"raw":{"variants":["Stacked metasurface max-min fairness lifts worst user 50x","Max-min fairness in stacked metasurface MISO: 50x for weakest user","Stacked metasurface lifts worst user's rate 50x with max-min","Fair beamforming with stacked metasurface: 50x gain for worst user"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2674,"prompt_tokens":1030,"completion_tokens":1644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1562}},"tokens_in":646,"tokens_out":1644,"duration_ms":11478,"temperature":1.0,"reasoning_tokens":1562,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:45:38.811044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed algorithm on a small SIM configuration, for example L=2 and M=4, with many random phase initializations and compare against exhaustive search over the discrete phase set; if the gradient descent-ascent phase update yields final minimum rates that vary widely or sit far below the exhaustive optimum, the central claim that the alternating algorithm reaches the max-min operating point is not supported. Independently, evaluate the true average minimum rate for the statistical-CSI policy by Monte Carlo at mid and high SNR; if the gap between the proposed bound and the simulated truth grows large, the claim that the bound is a good stand-in fails.","supporting_citations":[{"cited_title":"Passive refleccommuntion o ptimization for IRS-aided multicast beamforming with discrete phase sh ifts,","cited_arxiv_id":null,"evidence_quote":"It is the source of the min-max surrogate equivalence for max-min SNR and of the gradient descent-ascent treatment of discrete phase shifts."},{"cited_title":"Achievab le rate analysis and max-min SINR optimization in intelligent refle cting surface assisted cell-free MIMO uplink,","cited_arxiv_id":null,"evidence_quote":"It supplies the geometric-programming formulation and complexity estimate for max-min power control under an alternating optimization structure."},{"cited_title":"Two-timescale gradien t descent ascent algorithms for nonconvex minimax optimization,","cited_arxiv_id":null,"evidence_quote":"It provides the two-timescale gradient descent-ascent theory whose step-size rule, epsilon greater than mu and tau equal to 10, the phase update adopts."}],"review_version":1}