{"id":"e61eb52f-7c6f-4bd1-a7b8-91da088de1ad","arxiv_id":"2506.22943","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An alternating optimization of beamforming, semantic compression ratio, and fluid-antenna port selection is proposed for near-field rate maximization, with performance evaluated on a Jensen upper bound rather than the true ergodic rate.","lead":"This paper combines a movable fluid antenna at the receiver with semantic compression at the transmitter, and proposes an alternating algorithm that picks beamforming, compression level, and antenna port positions to maximize a proxy for transmission rate. The reported simulations show large rate gains over fixed-antenna and non-semantic baselines, but the proxy is a Jensen upper bound on the true rate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported rate gains are computed from the Jensen upper bound of Eq. (19), not the true ergodic rate of Eq. (16a), so the 71% improvement claim is unverified until the bound's tightness is checked.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the paper optimizes and evaluates the Jensen upper bound rather than the true ergodic rate, and no tightness check is provided. My independent reading confirms this is the most serious issue. The algebraic derivation of Eq. (18) is correct for the stated Gaussian model, and the subsequent determinant manipulations in the port-selection subproblem are also valid, so the concern is not about internal inconsistency but about whether the optimized surrogate reflects the true objective. The numerical section does not disclose whether the plotted 'equivalent rate' is the true expectation or the upper bound, and the convergence proof in Section III-C cites the power constraint (21) rather than the objective function, which further weakens the support for the alternating algorithm. However, both weaknesses are testable: a Monte-Carlo evaluation of the true expectation can settle the tightness question, and a standard stationary-point or monotone-objective argument can replace the flawed convergence remark. For this reason the appropriate verdict remains CONDITIONAL rather than REJECT; the reader's conditional acceptance is the right level of scrutiny. I do not see an additional, more fundamental flaw that would change the verdict to rejection: the system model is coherent, the optimization subproblems are mostly well posed, and the comparison baselines are reasonable in principle.","tokens_in":7985,"tokens_out":4174,"duration_ms":47959,"concrete_test":"Rerun the optimized configuration used in Fig. 2 at SNR=15 dB and at SNR=0 dB, and compute the true equivalent rate as R_true = (1/rho) * (1/T) * sum_{t=1}^T log det(I_ma + (1/sigma^2) G_t Q G_t^H), with T >= 10^4 independent draws of the random path-response matrix O and fixed optimized Q, rho, and r. Compare this with the upper bound Rbar from Eq. (19) using the same parameters. Then compute R_true for the three baseline schemes under identical channel draws and power settings. If the gap between Rbar and R_true is small (e.g., below 5%) and the 71% advantage over 'FAS-non-semantic' persists when all schemes are evaluated with R_true, the central claim stands. If the gap is large or the advantage shrinks materially, the reported gains are artifacts of optimizing and evaluating an untested Jensen upper bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the alternating algorithm maximizing R = R0/rho under the power constraint achieves up to 71% higher equivalent rate than a FAS without semantic compression. The optimization, however, is performed on the Jensen upper bound Rbar in Eqs. (17)-(19), not on the true ergodic rate R0 = E_O log det(I + (1/sigma^2) G Q G^H). Equation (18) is algebraically correct for i.i.d. complex Gaussian O, but Jensen's inequality is strict whenever the channel is random, and the paper never reports the size of the gap between Rbar and the true expectation. Section IV plots 'equivalent rate' without stating whether the curves are the true rate or the upper bound; if they are Rbar, then the reported gains are gains in an upper bound and may not reflect the actual rate achievable by the proposed Q, rho, and r. Moreover, optimizing a loose upper bound can yield different and suboptimal choices for the true objective, so the central claim that the proposed scheme improves the true equivalent rate is not established. A secondary but related weakness is the convergence argument in Section III-C, which states that 'the value of (21) increases monotonously', yet Eq. (21) is the power constraint Pc = Pmax - tr(Q), not the objective; thus monotone convergence of the alternating procedure is not rigorously justified. Both issues are addressable, but the unverified Jensen surrogate is the load-bearing one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a downlink fluid antenna system (FAS) with semantic communication, where a multi-antenna base station performs semantic compression and a user with a fluid antenna selects activated ports. The central claim is that jointly optimizing the transmit covariance matrix Q, the semantic compression ratio ρ, and the FAS port selection r maximizes the equivalent rate R = R0/ρ under a total power constraint. The authors replace the true ergodic expectation in the objective with a Jensen upper bound, decompose the resulting problem into a beamforming/compression subproblem and a port-selection subproblem, and solve them alternately using Dinkelbach's transform and an enumerate-like coordinate ascent. Simulation results are reported for an SNR sweep, showing faster growth and a 71% gain over a FAS without semantic compression at SNR = 15 dB.","tokens_in":8250,"tokens_out":6360,"duration_ms":69317,"significance":"The integration of semantic compression with fluid antenna systems is timely, and the system model with near-field channels and probability-graph semantic extraction is a reasonable formulation. The algebraic simplification in Eq. (18) is correct for i.i.d. complex Gaussian path responses, and the use of Sylvester's identity in Eq. (26) is sound. The simulation study compares against several baselines and includes convergence plots. If the central claim could be established for the true ergodic rate rather than an upper bound, the proposed alternating algorithm would be a useful design for FAS-assisted semantic networks. However, the current version does not verify the tightness of the Jensen surrogate, and the convergence proof contains a misreference to Eq. (21); these are load-bearing gaps that prevent the main claim from being accepted as stated.","major_comments":[{"comment":"The optimization objective in (16a) is the true ergodic equivalent rate R = E_O[(1/ρ) log det(I_ma + (1/σ^2)G(r)Q G^H(r))]. In Eqs. (17)-(19) this is replaced by the Jensen upper bound Rbar = (1/ρ) log det(I_ma + (α^2/σ^2) tr(AQA^H) B^H(r)B(r)). For a random O, Jensen's inequality is strict, so Rbar is an upper bound on the achievable rate, not the achievable rate itself. The paper then optimizes Rbar in Sections III-B and III-C, and in Section IV plots 'Equivalent Rate' without stating whether the curves are R or Rbar. If the curves are Rbar, then the reported 71% gain over FAS-non-semantic is a gain in an upper bound, and the optimized Q and r may be suboptimal for the true objective. The manuscript must either compute the true expectation by Monte Carlo in the simulations, provide a tightness analysis of the bound, or otherwise establish that maximizing Rbar is faithful to maximizing R. Without this, the central claim that the proposed scheme improves the true equivalent rate is not established.","section":"III-A, Eqs. (17)-(19) and Section IV"},{"comment":"The convergence argument states that 'the value of (21) increases monotonously during the optimization and as (21) is boundary, the overall algorithm is bound to converge.' Equation (21) is the equality constraint P_c = P_max - tr(Q), not an objective function, and there is no reason its value should be monotone during the iterations. A valid convergence proof must show that the objective of the alternating procedure, e.g., η(i)(Q,r) defined in Algorithm 2, is non-decreasing and bounded above, and that the Dinkelbach inner loop achieves the claimed accuracy. The current text does not provide this. This is a load-bearing gap because the algorithm's termination guarantee and the claim of convergence in Fig. 1 are not rigorously supported.","section":"III-C, convergence paragraph after Algorithm 2"},{"comment":"Under the Jensen surrogate, the objective in (20) depends on Q only through the scalar tr(A Q A^H) (see (19)). Consequently, the 'beamforming' optimization in Algorithm 1 does not actually optimize the spatial structure of Q; any Q with the same value of tr(AQA^H) yields the same surrogate value. In the true rate (11), the structure of Q matters through G(r)Q G^H(r). This reinforces that the surrogate is not a faithful proxy for the beamforming problem and should be explicitly acknowledged; the paper's language describing joint design of 'transmit beamforming' overstates what is optimized.","section":"III-B, Eq. (19) and Eq. (20)"}],"minor_comments":[{"comment":"The text twice refers to 'Talor's approximation'; this should be 'Taylor's approximation'.","section":"II-B, Eqs. (2)-(7)"},{"comment":"The abstract and contributions state that the optimal semantic compression ratio is obtained in closed form, but the algorithm enumerates the piecewise segments θ_s and uses Dinkelbach's transform; the phrase 'closed form' overstates the procedure.","section":"I, contributions"},{"comment":"The parameters of the semantic compression model in Eq. (13), namely A_s, B_s, D_s and p_0, are not given numerical values; without these values the simulation results cannot be reproduced.","section":"IV, Simulation results"},{"comment":"The binary variables θ_s are used in Eq. (22) before they are formally defined; please define θ_s and the range of the summation immediately before or after Eq. (22).","section":"III-B, Eq. (22)"},{"comment":"The inner loop resets the iteration index with 'Set iteration index i = 0' at the start of each outer iteration; this is confusing because the inner Dinkelbach loop should continue incrementing i until convergence. Please restructure the loop description.","section":"Algorithm 1"},{"comment":"The text says 'SNR is defined as Pmax/σ^2' while the average noise power is set to 3 dBm; please clarify the relationship and state whether σ^2 is in linear or dB scale.","section":"IV, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The main technical obstacle is the unverified Jensen upper bound; adding a Monte Carlo evaluation of the true expectation would substantially strengthen the paper. The reliance on the Jensen surrogate also makes the beamforming optimization essentially a scalar trace allocation, which may be a deeper limitation than the authors acknowledge. The convergence proof needs to refer to the actual objective rather than Eq. (21). The paper's use of the computation-load model from reference [8], which shares authors, is noted but is not by itself disqualifying."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this paper is a reasonable first pass at a genuinely new combination—fluid antenna port selection, semantic compression ratio, and transmit covariance optimization in a near-field mmWave downlink. The alternating algorithm (Dinkelbach for Q and rho, coordinate ascent for ports) is standard but competently assembled, and the closed-form handling of the piecewise compression cost is a nice touch. The math inside the surrogate model checks out; equation (18) is correct for i.i.d. complex Gaussian O.\n\nThe soft spot is the one the stress test flags: the objective actually optimized is the Jensen upper bound in (19), not the ergodic rate in (16a). The paper never computes the true expectation, never quantifies the gap, and then plots 'equivalent rate' in Section IV without saying which curve it is. If those curves are the bound, then the 71% improvement over FAS-non-semantic is a gain in the bound, and the optimized Q, rho, and r may be suboptimal for the true rate. That's a load-bearing issue, not cosmetics. It's fixable by simulating the true expectation and reporting the gap, but as submitted the central claim is not established.\n\nSecondary issues: the convergence argument in Section III-C says 'the value of (21) increases monotonously,' but (21) is the power constraint Pc = Pmax - tr(Q), not the objective. That's a misstatement and needs a proper monotonicity argument on Rbar. Also, the simulation parameters for the semantic compression load function (A_s, B_s, D_s) are not disclosed, so the experiments are not fully reproducible. Both are addressable, and the Jensen gap is the main one.\n\nWho is this for? Read it if you work on FAS or semantic resource allocation; it gives a concrete formulation and a workable algorithm skeleton, but treat the numerical gains as conditional. I wouldn't cite it as it stands, but I'd be willing to revisit after a revision.\n\nRecommendation: send it to peer review. The problem is meaningful, the approach is coherent, and the key flaw is verifiable rather than fatal. Ask the authors to check the bound tightness, fix the convergence statement, and disclose parameters. With those changes it could become a solid contribution.\n\nBest,\n[You]","headline":"The novelty is real but the reported 71% gain is computed from a Jensen upper bound, not the true ergodic rate, so the central quantitative claim remains unverified until the bound's tightness is checked.","tokens_in":8805,"tokens_out":2207,"would_cite":false,"duration_ms":25273,"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":"Jointly optimizing transmit covariance, semantic compression ratio, and fluid-antenna port selection maximizes equivalent rate under a combined power budget; the paper's alternating algorithm reports up to 71% gain over a non-semantic…","keywords":["fluid antenna system","semantic communication","rate maximization","beamforming optimization","port selection","near-field communication","Dinkelbach transform","equivalent rate"],"falsifier":"For the Section IV parameters, compute the true equivalent rate $\\frac{1}{\\rho}\\mathbb{E}_O[\\log\\det(I_{m_a}+\\frac{1}{\\sigma^2}G(r)QG^H(r))]$ by Monte Carlo averaging over many realizations of $O$ for the algorithm's output $(Q,\\rho,r)$, and compare it with the optimized bound (19); a relative gap of more than a few percent would show that the optimized variables are not maximizing the actual rate. For small $M$, exhaustively checking all legal port selections and comparing with Algorithm 2's choice would directly test the port-selection subproblem.","tokens_in":7776,"feed_emoji":"📡","tokens_out":15884,"duration_ms":135155,"temperature":0.7,"pith_summary":"The paper tries to establish that in a near-field downlink where the base station compresses data through semantic extraction and the user has a fluid antenna that can activate any small set of its many ports, the equivalent rate $R=R_0/\\rho$ is maximized by jointly choosing the transmit covariance $Q$, the semantic compression ratio $\\rho$ (the fraction of the original data size that remains), and the activated ports $r$, subject to a single power budget covering both transmission and computation. Because the exact ergodic rate is hard to optimize, the paper replaces it with a deterministic upper bound obtained by moving the expectation inside the logarithm, then solves the resulting problem with an alternating algorithm: Dinkelbach's transform handles $Q$ and $\\rho$, and a one-port-at-a-time enumeration handles $r$. The simulation study, with a 20-antenna base station, a 35-port fluid antenna, and 5 activated ports, shows the scheme converging quickly and exceeding a fluid-antenna system without semantic extraction by 43% at 0 dB and 71% at 15 dB SNR. The practical interest is that semantic compression and movable antennas are two separate proposals for stretching spectrum and energy, and this paper supplies one way to make them work together under a coupled power constraint, with a closed-form compression-ratio update at each step.","feed_headline":"Joint tuning of beamforming, compression, and ports lifts rate 71%","feed_subtitle":"It maximizes equivalent rate under a joint power budget, beating a non-semantic fluid-antenna baseline by up to 71%.","key_machinery":"Three pieces carry the argument. The first is the Jensen upper bound of Eqs. (17)-(19): replacing $\\mathbb{E}_O[OAQA^HO^H]$ with $\\mathrm{tr}(AQA^H)\\alpha^2 I_{V_r}$ removes the expectation from the log-determinant and leaves a deterministic ratio of a concave function of $Q$ to a piecewise-linear function of the compression rate. The second is Dinkelbach's transform, which converts the fractional program $\\max f(Q)/g(Q)$ into the iterated problem $\\max f(Q)-\\tau g(Q)$ with $\\tau=f(Q)/g(Q)$, making each subproblem convex. The third is the port-selection update, which splits the objective as $\\frac{1}{\\rho}\\log\\det(I_{V_r}+\\gamma B_mB_m^H)+\\frac{1}{\\rho}\\log(1+\\gamma b^H(r_m)(I_{V_r}+\\gamma B_mB_m^H)^{-1}b(r_m))$ with $\\gamma=\\frac{\\alpha^2}{\\sigma^2}\\mathrm{tr}(AQA^H)$, so that the combinatorial choice of $m_a$ ports is reduced to $m_a$ single-port searches over the $M$ available positions.","core_discovery":"The paper's central claim is that the equivalent rate of a FAS-assisted semantic downlink is maximized by the joint triple $(Q,\\rho,r)$, and that the proposed alternating algorithm attains this maximum on a tractable surrogate. The objective is $R=\\frac{1}{\\rho}\\mathbb{E}_O[\\log\\det(I_{m_a}+\\frac{1}{\\sigma^2}G(r)QG^H(r))]$, where $G(r)=B^H(r)OA$ is the near-field channel built from transmit and receive field response matrices and a random path-response matrix $O$. The authors use $\\mathbb{E}_O[OAQA^HO^H]=\\mathrm{tr}(AQA^H)\\alpha^2 I_{V_r}$ to move the expectation inside the logarithm, producing the deterministic upper bound $R\\le \\frac{1}{\\rho}\\log\\det(I_{m_a}+\\frac{\\alpha^2}{\\sigma^2}\\mathrm{tr}(AQA^H)B^H(r)B(r))$, and then optimize this bound by alternating between Dinkelbach-based updates of $(Q,\\rho)$ and coordinate-wise enumeration of $r$. On this basis, the paper reports that the proposed scheme outperforms a fluid-antenna system without semantic extraction by 43% at 0 dB and 71% at 15 dB SNR, with the semantic-compression gain amplifying as the channel improves.","pith_inferences":["Editorial inference: the same alternating structure, using a deterministic surrogate, Dinkelbach updates, and one-port-at-a-time enumeration, would extend to multiple semantic users or to jointly optimizing the number of activated ports $m_a$, which the paper keeps fixed.","Editorial inference: a Monte Carlo comparison between the optimized upper bound (19) and the true averaged rate (16a) would settle whether the reported gains are realized, and the paper does not include that check.","Editorial inference: comparing semantic compression on a fixed antenna against the proposed scheme would isolate the value of the fluid antenna itself; the paper's main comparison is against non-semantic FAS, so the two ingredients' separate contributions are not fully separated.","Editorial inference: because semantic extraction consumes power through $P_c=c(\\rho)p_0$, the optimal compression ratio encodes a computation-versus-transmission tradeoff, with more power spent on compression when the channel is good; this pattern is consistent with the paper's SNR-dependent results but is not stated as a separate result."],"forward_implications":["Under the paper's model, semantic compression is most valuable when the channel is strong: the reported gain over non-semantic FAS grows from 43% at 0 dB to 71% at 15 dB SNR, because less transmit power is needed and the compression step magnifies the rate.","Port selection remains important even with semantic compression: the random-port semantic scheme underperforms the optimized scheme, so choosing which ports to activate and choosing the compression ratio are not separable decisions.","The algorithm's fast convergence, with one optimization round already close to the final equivalent rate, means the alternating procedure can serve as a practical resource allocator rather than a one-shot offline design.","The worst-case complexity $O(N^{4.5}\\log(1/\\epsilon_1)M^{m_a}/\\epsilon_2)$ keeps the approach feasible for moderate port counts, as in the simulated $M=35$, $m_a=5$ case."],"supporting_citations":[{"why":"Supplies the near-field FAS channel model and the non-semantic FAS algorithm used as the main baseline.","marker":"[4]"},{"why":"Supplies the probability-graph semantic extraction model and the piecewise computation-power function that couples the compression rate to power.","marker":"[8]"},{"why":"Supplies the Dinkelbach transform used to solve the fractional subproblem for transmit covariance and compression ratio.","marker":"[10]"}],"fun_headline_variants":["Joint port, beam, and rate tuning lifts semantic downlink 71%","Fluid antenna + semantic coding: joint tuning yields 71% rate gain","Alternating algorithm maximizes FAS semantic rate, up to 71% gain","Semantic compression and port selection boost FAS rate by 71%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole optimization is run on an upper-bound formula for the average rate, and the paper never checks how close that formula is to the true average rate; if the bound is loose, the optimized settings and the reported gains may not be real.","fun_headline_variants_meta":{"raw":{"variants":["Joint port, beam, and rate tuning lifts semantic downlink 71%","Fluid antenna + semantic coding: joint tuning yields 71% rate gain","Alternating algorithm maximizes FAS semantic rate, up to 71% gain","Semantic compression and port selection boost FAS rate by 71%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000478,"raw_usage":{"total_tokens":2377,"prompt_tokens":965,"completion_tokens":1412,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1332}},"tokens_in":581,"tokens_out":1412,"duration_ms":15882,"temperature":1.0,"reasoning_tokens":1332,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:54:42.139933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the Section IV parameters, compute the true equivalent rate $\\frac{1}{\\rho}\\mathbb{E}_O[\\log\\det(I_{m_a}+\\frac{1}{\\sigma^2}G(r)QG^H(r))]$ by Monte Carlo averaging over many realizations of $O$ for the algorithm's output $(Q,\\rho,r)$, and compare it with the optimized bound (19); a relative gap of more than a few percent would show that the optimized variables are not maximizing the actual rate. For small $M$, exhaustively checking all legal port selections and comparing with Algorithm 2's choice would directly test the port-selection subproblem.","supporting_citations":[{"cited_title":"Joint beamforming and antenna design for near-field fluid antenna system,","cited_arxiv_id":null,"evidence_quote":"Supplies the near-field FAS channel model and the non-semantic FAS algorithm used as the main baseline."},{"cited_title":"Generative ai-driven semantic communication networks: Architecture, technologies, and applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the probability-graph semantic extraction model and the piecewise computation-power function that couples the compression rate to power."}],"review_version":1}