{"id":"ae8e93c9-db44-4f53-9560-eb7db67cb837","arxiv_id":"1908.07408","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A mixed-timescale stochastic optimization algorithm jointly designs beamforming and power splitting for massive MIMO SWIPT IoT networks under imperfect CSI and reports simulation gains over fixed-beamforming baselines.","lead":"This paper proposes a mixed-timescale beamforming and power-splitting scheme for massive MIMO wireless power transfer networks, adapting beamformers to imperfect channel estimates while adjusting splitters only to long-term statistics. It reports simulation gains over zero-forcing and maximum-ratio baselines, with convergence guarantees delegated to prior work.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Optimizing the rate bounds instead of the true ergodic rate is not an equivalent reformulation for the claimed general concave utilities; MO-SSCA's stationarity guarantee is for a surrogate objective, so the network-utility claim is not established.","rationale":"The reader's weakest assumption is the right one: the equivalence between optimizing the rate bounds and optimizing the true ergodic rate is asserted, not proven. I have sharpened it in two ways. First, the claimed equivalence fails even between upper and lower bounds for the paper's own general utility class when per-device constants C_k are present; g(x)=\\log x provides an immediate counterexample. Second, the true ergodic rate differs from the upper bound by a nonconstant penalty that depends on F and \\rho, so using C_k is not just a constant shift. Theorem 3 is also delegated to a self-cited prior paper, but even accepting that proof would only establish stationarity for the surrogate problem P. The simulations use the linear utility g(\\eta)=\\sum \\eta_k, for which the constant shift is harmless, so the numerical gain claim may survive; however the abstract and theorem claim a general-utility result. The conditional verdict is therefore appropriate; the concern does not move the verdict because the reader already conditioned on this gap.","tokens_in":10334,"tokens_out":13693,"duration_ms":149818,"concrete_test":"Recompute the Section IV comparison using the true ergodic rate \\hat r_k from Eq. (1), including the - (1/T)\\sum_m \\log_2(1 + T Var(h_k^H f_m)/(\\rho_k\\sigma_k^2+\\delta_k^2)) penalty, evaluated at the converged MJBP solution and at a fine grid over candidate \\rho (or a few offline-optimized \\rho). If the true-utility maximizer differs materially from the bound-utility maximizer, the surrogate substitution is not benign and the reported gain is only for P, not for the actual network utility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is the substitution of the true ergodic rate by its bounds in problem P. Proposition 1 gives \\hat r_k \\le r_k and \\hat r_k \\ge r_k - C_k, where C_k = (1/T)\\sum_m \\log_2(1 + T Pmax/\\delta_k^2 E\\|h_k\\|^2) is independent of F and \\rho. The text then claims that optimizing the lower and upper bounds provides the same optimal solution, and P is formulated with this bound. For the stated utility class (g concave, nondecreasing, with Lipschitz derivative), this equivalence is false: maximizing \\sum_k g(\\eta_k) versus \\sum_k g(\\eta_k - C_k) can have different argmaxes when the C_k differ across devices (e.g., g(x)=\\log x or sqrt). Moreover, the true gap between \\hat r_k and r_k is not the constant C_k: it equals (1/T)\\sum_m \\log_2(1 + T Var(h_k^H f_m)/(\\rho_k\\sigma_k^2+\\delta_k^2)), which depends on F and \\rho; C_k is only a uniform upper bound. Hence P is a surrogate problem, and Theorem 3 at most gives stationarity of that surrogate. Fig.2 shows pointwise tightness of the bounds, not optimizer equivalence or equivalence to the true rate. Because the abstract and Theorem 3 claim a solution for general utility functions, the central claim is not established for the stated problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a downlink massive MIMO SWIPT IoT network where the transmit beamformer adapts to imperfect CSI while receive power splitters adapt only to long-term channel statistics. It formulates a mixed-timescale network utility maximization problem, proposes an online stochastic successive convex approximation algorithm (MO-SSCA) with a short-term fractional-programming block-coordinate-descent inner loop and a long-term surrogate update, and claims the algorithm converges to stationary solutions. The simulation study reports gains over MRT and ZF baselines across SNR, number of users, and number of antennas.","tokens_in":10706,"tokens_out":10288,"duration_ms":106362,"significance":"If the theoretical claims were fully supported, the paper would provide a practical mixed-timescale design for massive MIMO SWIPT under imperfect CSI and a useful template for two-timescale stochastic optimization. The system model and algorithmic architecture are clearly described, the short-term steps build on standard tools (SAA, fractional programming, MM), and the simulations cover several operating regimes. However, the central substitution of the true ergodic rate by an upper bound is not an equivalent reformulation for general utilities, and the main convergence theorem is not proved in the paper. These gaps currently prevent the main claims from being accepted as stated.","major_comments":[{"comment":"The assertion 'From Proposition 1, optimizing the lower and upper bound provide the same optimal solution' is not established and is false for the stated utility class. The lower bound r_k - C_k and the upper bound r_k differ by the constant C_k, so for linear utilities their maximizers coincide, but for a general concave nondecreasing g, sum_k g(r_k - C_k + gamma_k e_k) and sum_k g(r_k + gamma_k e_k) can have different argmaxes. More importantly, the true ergodic rate is not equal to either bound; the actual gap between r_k and the true rate depends on F and rho through terms of the form (1/T) sum_m log_2(1 + T Var(h_k^H f_m)/(rho_k sigma_k^2 + delta_k^2)), so replacing the true rate with r_k changes the objective by a variable-dependent term. Figure 2 only demonstrates pointwise tightness of the bounds at a simulated operating point, not optimizer equivalence. Consequently, problem P is a surrogate problem, and Theorem 3, even if valid, establishes stationarity for the surrogate rather than for the original ergodic-rate utility.","section":"Section II, Proposition 1 and Eq. (2)"},{"comment":"The main convergence theorem is stated without proof. The sentence 'Theorem 3 can be proven by a similar approach in [7]' is not an adequate proof because [7] addresses a different system (hybrid compression for C-RAN) and does not include the nonlinear energy-harvesting model or the SAA short-term subproblem used here. Since Theorem 3 is the basis for the claim that Algorithm 1 converges to stationary solutions of P, a full proof, or a precise identification of a theorem in [7] from which it follows with stated modifications, is required.","section":"Section III-C, Theorem 3"},{"comment":"As printed, the objective in Eq. (4) contains a positive term (w^n_k)^H w^n_k (rho_k(Gamma^n_k + sigma^2_k) + delta^2_k). With q and w fixed, Gamma^n_k is convex quadratic in F, so this term is convex in F; maximizing it over the convex set Lambda is not a convex problem. The text then states that problem (8) is convex and can be solved by CVX, which contradicts the displayed equations. Please verify the sign of the quadratic penalty in Eq. (4) and either correct it or provide a convexity argument for (8). If the sign is a typographical error, the correction is straightforward, but as written the short-term convergence of Algorithm 2 is not justified.","section":"Section III-A, Eq. (4) and problem (8)"}],"minor_comments":[{"comment":"Equation (1) is malformed as printed: the penalty term 'log_2(1 + T rho_k sigma^2_k + delta_k Var(h_k^H f_m))' lacks a fraction bar and correct parentheses; please re-typeset the expression.","section":"Section II, Eq. (1)"},{"comment":"In Lemma 1, the phrase 'solves the problem in (1)' should refer to problem P3 or Eq. (3), since Eq. (1) is a rate expression, not an optimization problem.","section":"Section III-A, Lemma 1"},{"comment":"The relation between the coherence interval T = 400 and the stated 'coherence time of 2 ms and a coherence bandwidth of 200 kHz' should be specified explicitly, as the product of time and bandwidth does not by itself determine T.","section":"Section IV"},{"comment":"There are several typographical errors, including 'postive' after Eq. (9) and 'an second-order cone programming' in the introduction; please proofread the manuscript.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main convergence theorem is delegated to a reference that shares authors with this manuscript. I would ask the editor to require a detailed proof or supplementary material for Theorem 3 before acceptance, especially because the problem structure differs from [7]. The surrogate-objective issue in Section II is the more fundamental concern and should be addressed in revision by either proving the equivalence under additional assumptions or restating the paper's claims for the surrogate problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes a mixed-timescale joint beamforming and power splitting scheme for massive MIMO SWIPT IoT networks, claiming to maximize a general concave utility under imperfect CSI. The setting is genuinely new—nobody has mixed timescales this way in SWIPT—and the MO-SSCA design, with short-term FP-BCD and long-term stochastic SCA, is a plausible way to tackle it. The simulations show consistent gains over MRT and ZF baselines, and the paper is clearly written.\n\nThe load-bearing problem is the surrogate. Proposition 1 gives bounds \\hat r ≤ r ≤ \\hat r + C, where C is constant w.r.t. F and ρ. The authors then say optimizing the lower and upper bound gives the same optimal solution and proceed to optimize \\hat r. That is only true for the upper bound \\hat r + C, which is constant-shifted. But the actual objective in P is r, and r is not \\hat r + C; the gap r − \\hat r is itself a sum over beamformers and power splitters (from the variance terms in equation (1)), so it is not constant. For linear utility like sum rate, optimizing \\hat r might be a reasonable surrogate, but for the claimed general concave utilities it is not equivalent. The pointwise tightness of the bounds in Fig. 2 does not fix this. So Theorem 3's stationarity result applies to the surrogate, not to P as stated.\n\nTwo smaller issues: Theorem 3's proof is deferred to a self-cited reference ([7]), and the finite-N approximation error at N=200 is not quantified. The simulations also lack error bars and no code is released, so it is hard to tell how robust the 'significant gain' is.\n\nNone of this makes the algorithm worthless. As an engineering design with a linear utility, the surrogate may be fine, and the mixed-timescale formulation is a useful addition to the SWIPT literature. But as it stands, the general convergence claim overreaches. A referee should ask the authors to either prove the equivalence under additional conditions, or reframe the problem as optimizing the lower bound and state the guarantees for that objective.\n\nI would send this to review, but I would flag the surrogate issue as the main revision point. The paper is worth engaging with for anyone working on stochastic optimization for SWIPT or massive MIMO, but cite it for the formulation, not for the convergence guarantee.","headline":"A useful mixed-timescale SWIPT design whose general convergence claim rests on an unproven (and likely false) equivalence between a rate bound and the true ergodic rate.","tokens_in":11195,"tokens_out":5287,"would_cite":false,"duration_ms":48968,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C15","94A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a mixed-timescale joint beamforming and power splitting scheme for massive MIMO SWIPT IoT networks, solved by an online stochastic successive convex approximation algorithm that provably reaches stationary solutions…","keywords":["SWIPT","massive MIMO","power splitting","mixed-timescale optimization","beamforming","stochastic successive convex approximation","ergodic rate bounds","nonlinear energy harvesting"],"falsifier":"On a small instance where the ergodic rate and its bounds can be evaluated exactly, numerically maximize the true ergodic rate and each bound separately over a coarse grid of power splitter values. If the maximizing power splitter differs, the bound-optimizer equivalence fails. Alternatively, run MO-SSCA on a synthetic channel distribution and compare its limiting utility with a brute-force search over $\\rho$ on the true objective; a gap would refute the claim that the algorithm maximizes the intended utility.","tokens_in":10149,"feed_emoji":"📡","tokens_out":6822,"duration_ms":59457,"temperature":0.7,"pith_summary":"Simultaneous wireless information and power transfer (SWIPT) in massive MIMO systems usually assumes perfect channel state information, which is unrealistic when the base station has many antennas. This paper proposes a mixed-timescale joint beamforming and power splitting (MJBP) scheme: digital beamformers adapt to the fast, imperfect channel estimates, while the power splitters at the IoT devices change only on the slow timescale of channel statistics, to respect hardware limits and signaling overhead. The resulting network utility maximization is nonconvex and stochastic, and the paper solves it with a mixed-timescale online stochastic successive convex approximation (MO-SSCA) algorithm. The authors prove that the algorithm converges to stationary solutions of the problem, and simulations show larger average utility than fixed MRT and ZF beamforming baselines across SNR, user count, and antenna count. The takeaway is that joint optimization across two timescales is a practical and beneficial route for massive MIMO SWIPT under imperfect CSI.","feed_headline":"Mixed-timescale design lifts SWIPT IoT utility under imperfect CSI","feed_subtitle":"Fast beamformers track imperfect CSI while slow power splitters track statistics, boosting rate-energy tradeoffs.","key_machinery":"The carrying mechanism is the two-timescale decomposition together with a chain of convex surrogates. At each short-term slot, a fractional programming block coordinate descent (FP-BCD) algorithm uses a Lagrangian dual transform and complex quadratic transformation to turn the rate expressions into a tractable form, and majorization-minimization with a first-order Taylor expansion handles the nonconvex nonlinear energy-harvesting constraint; a sample average approximation with $N$ channel-error samples makes the expectations finite. On the long-term timescale, a concave surrogate function with proximal regularization and a recursively updated weight vector $v^t$ drive a projected closed-form update of each power splitter $\\rho_k$. Theorem 2 supplies an exponential convergence guarantee for the short-term solver, and Theorem 3 chains that to the almost-sure stationarity of the mixed-timescale iterates.","core_discovery":"The central claim is that the mixed-timescale problem—choosing beamformers per imperfect channel realization and power splitters per channel statistics—can be solved to stationarity by the proposed MO-SSCA algorithm. Theorem 3 states that every limit point of the iterates satisfies the first-order stationary conditions for the utility maximization problem, up to an error that vanishes exponentially as the number of channel-error samples grows. The paper also claims that the MJBP scheme meaningfully outperforms MRT and ZF baselines in average sum utility, and that the gains widen as the number of users grows and remain consistent as the number of antennas grows.","pith_inferences":["The same two-timescale template could be applied to other slow variables in massive MIMO, such as user scheduling, hybrid precoding phases, or resource allocation, where fast beamformers track instantaneous CSI.","If the equivalence between optimizing the ergodic-rate bounds and the true ergodic rate fails, one could replace the surrogate by an unbiased stochastic gradient of the true rate and retrain the long-term update; this is a direct testable modification.","The simulation setup, with a small number of paths and 64 antennas, suggests the scheme could be prototyped on a testbed; a useful experiment would measure how the gain over ZF changes under measured, non-Laplacian angular spreads.","The convergence proof omits details by citing an existing approach, so a re-derivation in a journal version would be needed before relying on the guarantee in a system design."],"forward_implications":["If the convergence result holds, MO-SSCA gives a principled way to optimize general utility functions in massive MIMO SWIPT without perfect CSI, using only slow updates of power splitters.","The joint design should let operators reduce feedback overhead: power splitter settings are broadcast once per channel-statistics coherence interval rather than per channel block.","The performance gains over MRT and ZF imply that optimizing the beamformer and the splitter together matters more as inter-user interference grows with the number of users.","Because the short-term subproblem is convex after transformation, each online step can be solved with standard convex solvers, making the scheme implementable.","The exponential convergence in sample size suggests that a modest number of channel-error samples suffices, limiting computational cost."],"supporting_citations":[{"why":"Supplies the achievable ergodic rate expression and the rate lower bound used in Proposition 1.","marker":"[6]"},{"why":"Provides the two-timescale convergence framework that Theorem 3's proof extends to this setting, plus the geometric channel model.","marker":"[7]"},{"why":"Provides the practical nonlinear energy harvesting model used in the harvested-power objective.","marker":"[8]"},{"why":"Justifies the sample average approximation that converts the expectation-based short-term problem into a finite-sum problem.","marker":"[9]"},{"why":"Supplies the Lagrangian dual transform and complex quadratic transformation that make the rate and energy terms tractable.","marker":"[10]"},{"why":"Gives the majorization-minimization convergence theorem used for the short-term FP-BCD algorithm.","marker":"[11]"},{"why":"Provides the MM framework and related convergence argument cited in the proof of Theorem 2.","marker":"[13]"},{"why":"Supplies the uniform exponential convergence result for sample average approximation used to bound the short-term error.","marker":"[14]"}],"fun_headline_variants":["Mixed-timescale design lifts SWIPT IoT utility","Fast beamformers, slow splitters: better SWIPT IoT","MO-SSCA solves mixed-timescale SWIPT optimization","Imperfect CSI tamed by mixed-timescale SWIPT","Mixed-timescale beamforming splits SWIPT gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the lower and upper bounds on the ergodic rate derived in Proposition 1 have the same optimal power-splitting and beamforming solution as the true ergodic rate; the paper asserts this without proof, so the objective being optimized may be a surrogate that differs from the intended network utility.","fun_headline_variants_meta":{"raw":{"variants":["Mixed-timescale design lifts SWIPT IoT utility","Fast beamformers, slow splitters: better SWIPT IoT","MO-SSCA solves mixed-timescale SWIPT optimization","Imperfect CSI tamed by mixed-timescale SWIPT","Mixed-timescale beamforming splits SWIPT gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001198,"raw_usage":{"total_tokens":4870,"prompt_tokens":809,"completion_tokens":4061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":3982}},"tokens_in":425,"tokens_out":4061,"duration_ms":29323,"temperature":1.0,"reasoning_tokens":3982,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:50.697809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small instance where the ergodic rate and its bounds can be evaluated exactly, numerically maximize the true ergodic rate and each bound separately over a coarse grid of power splitter values. If the maximizing power splitter differs, the bound-optimizer equivalence fails. Alternatively, run MO-SSCA on a synthetic channel distribution and compare its limiting utility with a brute-force search over $\\rho$ on the true objective; a gap would refute the claim that the algorithm maximizes the intended utility.","supporting_citations":[{"cited_title":"On the ergodic rate lower bounds with applicat ions to masssive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the achievable ergodic rate expression and the rate lower bound used in Proposition 1."},{"cited_title":"Two-timescale hybrid compression and forward for massive MIMO aided C-RAN,","cited_arxiv_id":null,"evidence_quote":"Provides the two-timescale convergence framework that Theorem 3's proof extends to this setting, plus the geometric channel model."},{"cited_title":"Pract ical non- linear energy harvesting model and resource allocation for SWIPT systems,","cited_arxiv_id":null,"evidence_quote":"Provides the practical nonlinear energy harvesting model used in the harvested-power objective."},{"cited_title":"Shapiro, D","cited_arxiv_id":null,"evidence_quote":"Justifies the sample average approximation that converts the expectation-based short-term problem into a finite-sum problem."},{"cited_title":"Fractional programming for communic ation systems – Part II: Uplink scheduling via matching,","cited_arxiv_id":null,"evidence_quote":"Supplies the Lagrangian dual transform and complex quadratic transformation that make the rate and energy terms tractable."},{"cited_title":"An expanded theoretic al treatment of iteration-dependent majorize-minimize algorithms,","cited_arxiv_id":null,"evidence_quote":"Gives the majorization-minimization convergence theorem used for the short-term FP-BCD algorithm."},{"cited_title":"Optimization of MIMO Device-to-Device Networks via Matrix Fractional Programming: A Minorization-Maximization Approach","cited_arxiv_id":"1808.05678","evidence_quote":"Provides the MM framework and related convergence argument cited in the proof of Theorem 2."},{"cited_title":"A note on uniform exponential converge nce of sample average approximation of random functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the uniform exponential convergence result for sample average approximation used to bound the short-term error."}],"review_version":1}