{"id":"125639ed-781c-4fde-95d9-1f68d720b5c4","arxiv_id":"2508.13839","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A movable-antenna cell-free ISAC system with a graph-neural-network beamformer is claimed to robustly mitigate uncertain power-amplifier distortion, but the worst-case rate bound rests on a false inequality.","lead":"Cell-free wireless networks that both communicate and sense can suffer from power amplifier distortion; the authors propose moving antennas and a graph neural network to robustly optimize beamforming under uncertainty about that distortion. The paper shows simulated gains over fixed-antenna baselines, but a key mathematical bound in the robustness proof is invalid, so the worst-case guarantee is not established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (69) upper bound is invalid: |A+B|^2 <= |A|^2+|B|^2 drops the positive cross term, so L(epsilon) in Theorem 1 is not a guaranteed lower bound.","rationale":"The reader's weakest assumption points to the same load-bearing flaw: Eq. (69) incorrectly bounds |A+B|^2 by |A|^2+|B|^2, ignoring the potentially positive cross term. This flaw directly undermines the derivation of L(epsilon), which is the foundation of the claimed worst-case robust optimization. The concern lands because the inequality is not merely loose; it is invalid as an upper bound, as shown by a simple real-valued counterexample. Additional issues in the CRLB derivation, the unequal-antenna baseline comparison, and the absence of convergence analysis further weaken the paper, but the Eq. (69) error is sufficient to reject the central claim of a certified robust design. The simulations may indicate a useful heuristic, but they do not rescue the theoretical robustness guarantee, so the reader's REJECT verdict is appropriate and unchanged.","tokens_in":21017,"tokens_out":3637,"duration_ms":38797,"concrete_test":"Recompute the robust rate lower bound for a single-tAP, single-UE, single-MA instance with beta1 = 1, epsilon = 0.2, beta3 = 0.2, h = w = 1 (all real). Evaluate the actual interference term |beta1 + 2 beta3|^2 and the claimed upper bound |beta1|^2 + (2 epsilon)^2. The claimed bound gives 1.16, while the actual value is 1.96, so Eq. (69) fails and L(epsilon) is not a guaranteed lower bound. If the authors instead use the valid triangle-inequality bound (|A|+|B|)^2, all subsequent expressions in Theorem 1 and Appendix B must be rederived and the robustness guarantee re-established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central robust rate guarantee in Theorem 1 depends on bounding the interference power by Eq. (69), which asserts |sum_b h^H Xi_b w_b,j|^2 <= |sum_b beta1 h^H w|^2 + epsilon^2 |sum_b 2 h^H diag{WW^H} w|^2. Writing A = sum_b beta1 h^H w and B = sum_b 2 beta3 h^H diag{WW^H} w, the exact identity is |A+B|^2 = |A|^2 + |B|^2 + 2 Re(A*B). The cross term 2 Re(A*B) can be positive; the triangle inequality only gives |A+B|^2 <= (|A|+|B|)^2, which is strictly larger than |A|^2+|B|^2. The claimed bound therefore fails as a universal upper bound. Since this inequality is used to upper-bound the interference contribution in delta, the derived L(epsilon) can exceed the actual achievable worst-case objective, so the max-min problem in Theorem 1 is not a conservative reformulation of problem (23). A concrete scalar counterexample: with beta1 = 1, beta3 = 0.2, h = w = 1, the actual left side is |1.4|^2 = 1.96, while the claimed upper bound is 1 + 0.16 = 1.16. The sensing constraint in Appendix A has a separate issue (the delay-phase derivative is omitted), but the rate-bound flaw alone invalidates the paper's central certified-robustness claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies a movable-antenna (MA)-aided cell-free integrated sensing and communication (CF-ISAC) system in which power amplifier nonlinearity is modeled by a third-order memoryless polynomial with bounded uncertain coefficients. It formulates a worst-case max-min problem that jointly optimizes beamforming and MA positions under transmit-power and sensing-accuracy constraints, and it proposes a graph-neural-network (SACGNN) solution trained in an unsupervised manner. The main theoretical contribution is Theorem 1, which claims a tractable conservative reformulation of the robust problem using a lower bound L(epsilon) on the communication objective and a conservative CRLB constraint g(epsilon, epsilon_0). The numerical section compares the proposed algorithm against DRL and fixed-position-antenna baselines.","tokens_in":21386,"tokens_out":5705,"duration_ms":60508,"significance":"If the robustness certificates were valid, the paper would address a timely practical problem—PA distortion uncertainty in distributed ISAC—and the combination of MAs with a distributed learning-based solver would be a useful engineering contribution. The paper also models hardware variability explicitly and compares against several baselines. However, the central theoretical guarantee is not established: the key inequality in Appendix B is mathematically incorrect, and the CRLB derivation omits a substantial dependence of the sensing channel on the target location. Consequently, the claimed conservative reformulation is not supported, and the numerical results remain a heuristic demonstration rather than a validated robust design. The paper provides no machine-checked proofs or reproducible code, so the burden falls entirely on the derivations, which fail at the load-bearing step.","major_comments":[{"comment":"The claimed bound |sum_b h_{b,k}^H Xi_b w_{b,j}|^2 <= |sum_b beta1 h_{b,k}^H w_{b,j}|^2 + epsilon^2 |sum_b 2 h_{b,k}^H diag{W_a W_a^H} w_{b,j}|^2 is not valid. Writing A = sum_b beta1 h^H w and B = sum_b 2 beta_{3,b} h^H diag{WW^H} w, the exact identity is |A+B|^2 = |A|^2 + |B|^2 + 2 Re(A^* B), and the cross term can be positive. Since this inequality is used to upper-bound the interference contribution in delta, the resulting L(epsilon) in Eq. (34) is not a guaranteed lower bound on the worst-case objective; Theorem 1 therefore does not provide the claimed conservative reformulation of problem (23). A concrete scalar counterexample with beta1 = 1, beta3 = 0.2, and h = w = 1 gives |1.4|^2 = 1.96 on the left-hand side, while the claimed bound evaluates to 1 + 0.16 = 1.16.","section":"Appendix B, Eq. (69)"},{"comment":"The FIM derivation omits the derivative of the round-trip delay phase e^{-j2 pi f_s tau_{a,b}} with respect to the target location e. The sensing channel in Eq. (9) depends on e through tau_{a,b}(e) = (l_{a,d}(e) + l_{b,d}(e))/c, where l_{a,d} and l_{b,d} are defined in Eqs. (11)-(12). This dependence contributes additional terms to dot H_{a,b} and ddot H_{a,b} that are not present in Eqs. (55)-(57). Without these terms, CRLB_b(e) in Eq. (22) and the robust sensing constraint in (23c)/(33b) are not correctly characterized, so the sensing-side robustness guarantee is also unsupported.","section":"Appendix A, Eqs. (55)-(57)"}],"minor_comments":[{"comment":"Eq. (60) writes E{F^b_{n1,n2}} while Eq. (58) defines F^b_{n1,n2}; the expectation and the expansion are not clearly derived, and the indexed quantities dot Upsilon, ddot Upsilon, and tilde Upsilon are introduced without defining all variants explicitly.","section":"Appendix A, Eq. (60)"},{"comment":"The statement that problem (33) is convex is not justified: L(epsilon) contains absolute values and products involving the optimization variables, and g(epsilon, epsilon_0) is a first-order Taylor expansion of a non-convex function. A proof or a more precise convexity argument is needed.","section":"Section IV.A, Eq. (33)"},{"comment":"The loss function in Eq. (39) uses a parameter beta > 0 that is not defined, and C1 is not explicitly connected to the constraint set of problem (36). Please define all symbols in the training objective.","section":"Section V, Eq. (39)"},{"comment":"Eq. (35) repeats the formulation already given in Eq. (32); if a different sensing-oriented problem is intended, the text should distinguish the two formulations clearly.","section":"Section IV.B, Eq. (35)"},{"comment":"The axis labels in Fig. 4 appear inconsistent with the text description: the horizontal axis is labeled 'Sum rate [bits/s]' while the text discusses 'number of episodes with high sum rate.' Please redraw the figure or revise the caption and description.","section":"Section VI, Fig. 4"}],"recommendation":"reject","confidential_remarks":"For the editor: the manuscript's central theoretical claims are invalid as written, and the issues are not purely presentational. The invalid inequality in Eq. (69) and the incomplete CRLB derivation both bear directly on the paper's claimed robustness guarantee. I recommend rejection, though the authors could potentially revisit the problem with a corrected bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThis paper is worth a look, but the central robustness guarantee does not hold. The combination of movable antennas, PA distortion uncertainty, and a GNN-based distributed solver for cell-free ISAC is genuinely new, and the problem is well motivated. The system model is carefully laid out, and the simulation study is extensive, even though it lacks error bars and in Fig. 7 compares different antenna counts.\n\nThe soft spot is the core theorem. In Appendix B, Eq. (69) bounds |A+B|^2 by |A|^2+|B|^2 when |B| <= eps|C|. That is not an upper bound because the cross term 2 Re(A^*B) can be positive. A scalar example with A=1, B=0.4 gives |A+B|^2 = 1.96, larger than 1 + 0.16 = 1.16. Consequently L(eps) in Theorem 1 is not a guaranteed lower bound, and the max-min problem is not a conservative reformulation of (23). This is not a small typo; the bound is load-bearing for the paper's claim of certified robustness. The CRLB derivation in Appendix A also omits derivatives of the round-trip delay phase with respect to the target position, so the sensing constraint has the same kind of problem.\n\nCredit where it's due: the paper identifies a real hardware bottleneck, the problem formulation is plausible, and the SACGNN is a sensible way to reach distributed decisions. If the authors replace the invalid inequality with the correct triangle inequality and rework the analysis, the paper could be salvageable. The convergence claim should be proven or explicitly stated as empirical, and the simulations would be more convincing with multiple runs.\n\nThe paper should be read by researchers working on robust beamforming for ISAC or GNN-based physical-layer optimization; they will find the problem setup and architecture ideas useful, but they should not rely on the theoretical certificate. It deserves a serious referee because the topic is timely and the paper is substantial, but in current form the verdict should be reject. I recommend sending it to review with a note asking the authors to fix the inequality before any acceptance.","headline":"A fresh problem setup, but the key robustness bound is mathematically wrong: Eq. (69) drops a positive cross term, so Theorem 1 does not certify worst-case performance.","tokens_in":21910,"tokens_out":4307,"would_cite":false,"duration_ms":40356,"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":"This paper proposes a worst-case robust beamforming and antenna-positioning design that treats power-amplifier distortion as bounded uncertainty, and claims it improves the communication-sensing trade-off in cell-free ISAC systems.","keywords":["cell-free ISAC","movable antenna","power amplifier nonlinearity","third-order distortion coefficient","worst-case robust optimization","graph neural network","CRLB constraint"],"falsifier":"Compute $L(\\epsilon)$ and the true minimum over $\\beta_{3,a}$ of the objective in Eq. (17) for a fixed small network with two access points, two users, and one target; if the true worst-case sum rate falls below $L(\\epsilon)$ for any feasible beamforming and MA positions, the claimed guaranteed robustness is not established. The same check applies to the sensing constraint by comparing the exact Fisher information matrix, including the derivative of the round-trip delay phase, with the conservative bound $g(\\epsilon,\\epsilon_0)$.","tokens_in":20771,"feed_emoji":"📡","tokens_out":6630,"duration_ms":66488,"temperature":0.7,"pith_summary":"Integrated sensing and communication networks that share spectrum across many access points degrade when power amplifiers distort transmitted signals, and the distortion coefficients vary from one amplifier to the next. This paper asks whether movable antennas can make the system robust to that uncertainty, and answers by treating the third-order distortion coefficient $\\beta_{3,a}$ of each access point as lying in a bounded set $|\\beta_{3,a}|\\le\\epsilon$. It derives a conservative worst-case lower bound on the achievable sum rate and a conservative sensing-accuracy constraint, then solves the joint beamforming and antenna-positioning problem with successive convex approximation and a graph neural network. The reported simulations indicate that the resulting design improves the communication-sensing trade-off under distortion and outperforms fixed-position antenna baselines in both robustness and capacity. If correct, this would let cell-free ISAC systems operate under realistic amplifier impairments without knowing the exact distortion coefficients.","feed_headline":"Movable antennas can offset power-amplifier distortion","feed_subtitle":"A worst-case graph-network beamformer preserves rate and sensing accuracy when PA distortion is uncertain.","key_machinery":"The load-bearing machinery is the combination of a third-order memoryless polynomial PA model with the Bussgang decomposition, which expresses the distorted output as $\\mathbf{c}_a=\\boldsymbol{\\Xi}_a\\mathbf{s}_a+\\mathbf{d}_a$ with an uncorrelated distortion vector. The uncertainty set $|\\beta_{3,a}|\\le\\epsilon$ converts distortion into a bounded perturbation, and Theorem 1 supplies the conservative worst-case rate bound $L(\\epsilon)$ and the sensing constraint $g(\\epsilon,\\epsilon_0)$. Successive convex approximation handles the nonconvexity of the resulting max-min problem, and the SACGNN, a heterogeneous graph network with Transformer-style self-attention and convolutional layers, coordinates beamforming and antenna-position decisions across access points. The graph structure lets each access point aggregate channel, position, and distortion features from its neighbors before decoding transmit beamformers and MA positions.","core_discovery":"The paper's central claim is that PA-induced nonlinear distortion can be handled by a distributed worst-case design rather than by exact distortion compensation. Under a third-order memoryless polynomial model and the Bussgang decomposition, the amplified signal splits into a linear term and an uncorrelated distortion term whose covariance depends on the beamformers. The authors bound the unknown third-order coefficients as $|\\beta_{3,a}|\\le\\epsilon$ and derive a conservative lower bound $L(\\epsilon)$ for the worst-case sum rate together with a conservative CRLB constraint $g(\\epsilon,\\epsilon_0)$ on sensing accuracy. They then maximize this lower bound over beamforming matrices and movable-antenna positions. The claimed discovery is that this worst-case formulation, solved by the proposed SACGNN, substantially improves the communication-sensing trade-off under distortion and beats fixed-position antenna baselines in robustness and capacity.","pith_inferences":["Because the worst-case bound in Appendix B uses a triangle-inequality estimate that can be checked numerically, the design could be re-run with an exact worst-case evaluation over $\\beta_{3,a}$; a comparison of the true max-min rate with $L(\\epsilon)$ would show how much of the reported gain is real robustness versus conservatism of the bound.","The same bounded-uncertainty treatment could be applied to other amplifier imperfections, such as phase noise or I/Q imbalance, by adding their coefficients to the Bussgang distortion covariance.","In a deployment where each access point can estimate its own $\\beta_{3,a}$ online, the global fixed bound $\\epsilon$ could be replaced by per-AP adaptive uncertainty sets, potentially improving the trade-off further."],"forward_implications":["Under the bounded-3RDC assumption, the proposed max-min beamforming and MA-position policy yields higher worst-case sum rates than designs that ignore distortion uncertainty.","Movable antenna repositioning adds spatial degrees of freedom that can be steered away from distortion-dominated directions, with sum-rate gains growing as the number of antennas per access point increases.","The SACGNN solution is decentralized and has per-layer complexity linear in the number of graph edges, making it scalable to large cell-free deployments.","The conservative CRLB constraint gives a tunable link between sensing accuracy and communication rate, since relaxing $\\gamma_b$ frees beamforming degrees of freedom for communication."],"supporting_citations":[{"why":"Supplies the Bussgang decomposition used to split the PA output into a linear gain term and an uncorrelated distortion vector.","marker":"[17]"},{"why":"Supplies the distortion covariance model and the distributed distortion-aware beamforming context that this work extends to movable antennas.","marker":"[18]"},{"why":"Supplies the worst-case robust beamforming principle with nonconvex uncertainty sets that motivates the epsilon-bound and max-min formulation.","marker":"[27]"},{"why":"Supplies the convex solver used to solve the SCA subproblem derived from Theorem 1.","marker":"[28]"},{"why":"Supplies the Transformer self-attention mechanism used for heterogeneous message passing in the SACGNN.","marker":"[30]"},{"why":"Provides the GNN-based precoding and DRL baselines that the proposed method is compared against.","marker":"[31]"},{"why":"Provides the movable-antenna cooperative ISAC model and the Fisher information matrix derivative template used in the sensing analysis.","marker":"[32]"}],"fun_headline_variants":["Worst-case robust beamforming foils PA distortion","Distributed graph-network beamforming survives PA distortion","Movable antennas boost ISAC under uncertain distortion","Robust MA-aided beamforming beats fixed arrays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the validity of a triangle-inequality bound in Eq. (69), specifically that $|A+B|^2$ is no larger than $|A|^2+|B|^2$ for the linear and distortion terms; because the cross term $2\\operatorname{Re}(AB^*)$ can be positive, the derived $L(\\epsilon)$ may not be a guaranteed lower bound on the worst-case sum rate.","fun_headline_variants_meta":{"raw":{"variants":["Worst-case robust beamforming foils PA distortion","Distributed graph-network beamforming survives PA distortion","Movable antennas boost ISAC under uncertain distortion","Robust MA-aided beamforming beats fixed arrays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1331,"prompt_tokens":989,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":281}},"tokens_in":605,"tokens_out":342,"duration_ms":4039,"temperature":1.0,"reasoning_tokens":281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:12:33.205415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $L(\\epsilon)$ and the true minimum over $\\beta_{3,a}$ of the objective in Eq. (17) for a fixed small network with two access points, two users, and one target; if the true worst-case sum rate falls below $L(\\epsilon)$ for any feasible beamforming and MA positions, the claimed guaranteed robustness is not established. The same check applies to the sensing constraint by comparing the exact Fisher information matrix, including the derivative of the round-trip delay phase, with the conservative bound $g(\\epsilon,\\epsilon_0)$.","supporting_citations":[{"cited_title":"The bussgang decomposition of nonlinear systems: Basic theory and MIMO extensions [lecture notes],","cited_arxiv_id":null,"evidence_quote":"Supplies the Bussgang decomposition used to split the PA output into a linear gain term and an uncorrelated distortion vector."},{"cited_title":"Distributed distortion-aware beamforming designs for Cell-Free mMIMO systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the distortion covariance model and the distributed distortion-aware beamforming context that this work extends to movable antennas."},{"cited_title":"Robust adaptive beamforming via Worst-Case SINR maximization with nonconvex un- certainty sets,","cited_arxiv_id":null,"evidence_quote":"Supplies the worst-case robust beamforming principle with nonconvex uncertainty sets that motivates the epsilon-bound and max-min formulation."},{"cited_title":"CVX: Matlab software for disciplined convex programming, version 1.21,","cited_arxiv_id":null,"evidence_quote":"Supplies the convex solver used to solve the SCA subproblem derived from Theorem 1."},{"cited_title":"Toward energy-efficient massive mimo: Graph neural network precoding for mitigating non- linear PA distortion,","cited_arxiv_id":null,"evidence_quote":"Provides the GNN-based precoding and DRL baselines that the proposed method is compared against."}],"review_version":1}