{"id":"cda4f240-06f4-4f08-bcac-cbc4505d5f25","arxiv_id":"2509.08238","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A joint optimization of training/data durations, bit allocation, and offloading ratio minimizes aerial vehicle energy in an ISAC-assisted SAGIN IoT system.","lead":"This paper proposes an algorithm that reduces an aerial vehicle's energy use when it offloads IoT data to a satellite for computing, while also sensing a target. It jointly sets training vs data phase lengths, bit allocation, and offloading fraction, and shows in simulation that this saves energy compared to simpler schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SCA surrogates (15) and (17) violate the tightness/majorization conditions of [15, Lemma 1], so the claimed convergence to stationary points is unsupported.","rationale":"The reader's weakest assumption identifies exactly the load-bearing problem: the SCA surrogates in (15) and (17) do not satisfy the conditions of [15, Lemma 1]. My analysis confirms this in detail: (15) is off by a factor of two at the current iterate and is not a majorant, and (17) applies the wrong surrogate to a bilinear term. This removes the theoretical basis for the central algorithmic claim of convergence to a stationary point. However, the flaw is local and fixable in revision by using a proper tangent surrogate for the product term and a bilinear surrogate for B_n; the simulations may still be approximately valid once the algorithm is corrected. Thus I do not escalate the verdict beyond CONDITIONAL; I agree with the reader's assessment. The zero-energy reporting for infeasible fixed-mode benchmarks is a secondary comparison artifact but not the most load-bearing issue.","tokens_in":9410,"tokens_out":8092,"duration_ms":100001,"concrete_test":"Independently re-derive the surrogates for E_A and B_n from [15, Lemma 1], requiring equality at z_n(v) and gradient matching, and compare with (15) and (17). Concretely, take a feasible point from Table I (e.g., N_p=0.5, N_d=0.5, ρ=0.5, l_d=100 kbit), compute both surrogates at that point, and then run Algorithm 1 while recording the true objective (13) and the constraint residuals of (14b)-(14e) at every iteration. If the surrogate values do not match the original functions at the current iterate, or if the true objective does not monotonically decrease to a point satisfying the KKT conditions of (14), then the claimed convergence guarantee fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing link in the central claim is the SCA convergence statement in Section III-A: Algorithm 1 'converges ... all its limit points are stationary,' justified by [15]. For that theorem to apply, the surrogates (15) and (17) must satisfy the tightness and majorization conditions of [15, Lemma 1]. They do not as written. At z_n(v), (15) evaluates to γ_U(C_A)^3/N_d^2 × [f1(ρ(v)) f2(l_d,n(v)) + f1(ρ(v)) f2(l_d,n(v))], i.e. twice the original E_A term, because the term -f1(ρ(v)) f2(l_d,n(v)) is missing; the surrogate is not equal to the objective at the current iterate. More importantly, the two-term expression f1(ρ) f2(l_d(v)) + f1(ρ(v)) f2(l_d) is not a global upper bound for f1(ρ) f2(l_d): for ρ=0, ρ(v)=0.5, l_d=2, l_d(v)=1 it equals 2 while the original term equals 8. Thus even after inserting the missing constant, the majorization property needed for monotone descent is absent. Equation (17) is a separate failure: B_n(ρ,l_d) is bilinear in ρ and l_d, but the surrogate reuses f1=(1-ρ)^3 and f2=l_d^3, so ar B_n(z_n(v);z_n(v)) = 2(1-ρ(v))^3 l_d(v)^3 T/(B N_d), which is not B_n(z_n(v)) = ρ(v) l_d(v) T/(B N_d); constraint (18b) is not a tangent approximation to (16). Consequently the cited theorem cannot establish stationarity, and the lower energy reported in Figs. 3-4 is not supported by the paper's convergence argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies an ISAC-assisted SAGIN-based IoT system in which an AAV performs uplink sensing of a target and, together with a LEO satellite, provides hybrid edge computing for an IoT device. The authors formulate a nonconvex optimization problem that minimizes the total AAV energy consumption by jointly optimizing the training/data phase durations, the data bit allocation, and the partial offloading ratio, subject to offloading rate constraints and a sensing SINR constraint. They propose an SCA-based double-loop algorithm (Algorithm 1) and empirically compare it with partial-optimization and fixed-mode benchmarks, reporting lower AAV energy consumption.","tokens_in":9921,"tokens_out":5590,"duration_ms":73757,"significance":"The problem setting is topical and the system model is described in a fairly self-contained way, with explicit formulas for channel estimation, achievable rates, sensing detection, and energy consumption. If the SCA convergence claim were valid, the paper would provide a useful optimization framework for energy-efficient ISAC-assisted SAGIN IoT. However, the central algorithmic claim rests on surrogate functions that do not satisfy the required tangent and majorization conditions of the cited SCA theory. Since the simulation advantage is not backed by a valid convergence argument, the main contribution is not yet established. The paper's strengths are its clear problem statement, the inclusion of ROC-based sensing verification, and the comparison against several benchmarks.","major_comments":[{"comment":"The surrogate for E_A^n is not tight at the current iterate. With f1(ρ)=(1−ρ)^3 and f2(ld,n)=ld,n^3, evaluating (15) at zn=zn(v) gives γU C_A^3/N_d^2 · [f1(ρ(v))f2(ld,n(v)) + f1(ρ(v))f2(ld,n(v))], i.e., twice the original term, because the constant −f1(ρ(v))f2(ld,n(v)) is omitted. Thus the tangent condition of [15, Lemma 1] is violated. Moreover, the two-term expression is not generally an upper bound of f1(ρ)f2(ld,n); e.g., at ρ=0, ρ(v)=0.5, ld,n=2, ld,n(v)=1 the surrogate equals 2 while the original product equals 8. The missing constant would make this worse, so the majorization property also fails.","section":"Sec. III-A, Eq. (15)"},{"comment":"Equation (17) applies the product surrogate built for (1−ρ)^3 ld,n^3 to the bilinear term B_n(ρ,ld,n)=ρ ld,n T/(B N_d). At the current iterate, (17) evaluates to 2(1−ρ(v))^3 ld,n(v)^3 T/(B N_d), which is not equal to B_n(z_n(v))=ρ(v)ld,n(v)T/(B N_d). Therefore (18b) is not a tangent approximation of (16), and the equivalence between the original communication energy and the slack-variable reformulation is broken. The surrogate for B_n should be constructed from f1(ρ)=ρ and f2(ld,n)=ld,n, and its tangent/majorization properties must be checked explicitly.","section":"Sec. III-A, Eq. (17) and constraint (18b)"},{"comment":"The claim that the update sequence {z(v)} converges to a stationary point is unsupported because it is based on [15, Theorem 2] whose assumptions are not satisfied by the surrogates in (15) and (17), as noted above. Without a valid convergence proof, the energy reductions reported in Figs. 3 and 4 cannot be attributed to convergence to a stationary point of (14).","section":"Sec. III-A, convergence statement after Algorithm 1"},{"comment":"The total energy E_t in (13) includes E_CH^n(Np) defined in (10), which depends on l_p,n, but the decision vector is z_n=(N_p,N_d,ρ,l_d,n) and l_p,n is not defined as a variable, fixed parameter, or function of N_p/N_d in (14). This makes the objective incomplete as stated: the channel-estimation energy term cannot be evaluated or optimized unless l_p,n is specified. The authors should either add l_p,n to the optimization or provide a rule linking it to the training duration and/or bit allocation.","section":"Sec. II and Problem (14), l_p,n"}],"minor_comments":[{"comment":"Grammar and copyediting: 'In this letters', 'a IoT device', 'AAV' vs 'AA V' inconsistency, and 'energy comsumption' in Fig. 4 should be corrected.","section":"Abstract and throughout"},{"comment":"The sentence 'if the algorithm does not terminate in finite steps, these stationary points are not local minima of Problem (14)' is misleading; limit points can be saddle points, but stationarity does not imply they cannot be local minima. Recast as 'limit points are stationary but not necessarily local minima.'","section":"Sec. III-A after Lemma 1"},{"comment":"The feasible set is referred to as 'X' in the initialization but never explicitly defined; define X as the feasible set of (14) or the convex inner approximation (18).","section":"Algorithm 1"},{"comment":"The rate expressions and the constraints are dimensionally ambiguous: R_I,A^n and R_A,L^n include Nd/T, and then are multiplied by B and compared with bit counts ld,n and ρld,n without an explicit frame-time factor. Clarify whether l_d,n is bits per frame or bits per data phase and adjust the inequalities accordingly.","section":"Eqs. (7)-(8), (14c)-(14d)"}],"recommendation":"major_revision","confidential_remarks":"The main technical flaw is localized to the SCA surrogate construction and the convergence theorem. I believe it is fixable: the authors should supply correct surrogates satisfying the tangent and majorization conditions, re-derive the inner approximation, and then re-run the simulations. However, until that is done, the paper's central algorithmic claim is not supported. The l_p,n issue also needs to be resolved. This is not a reject because the problem formulation and simulation methodology are otherwise reasonable and the scope is appropriate for a letters paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain take: the paper is a competent system-level optimization for a single-IoT-device ISAC SAGIN scenario, but the convergence proof of the proposed SCA algorithm has a load-bearing error: the surrogates in (15) and (17) are not tight at the current iterate, so the cited theorem does not apply.\n\nWhat's genuinely new: the joint treatment of training/data phase duration, per-frame bit allocation, and partial offloading ratio under a sensing SINR constraint and offloading constraints. The system model is laid out carefully, with reasonable energy expressions for channel estimation, computation, and transmission. The simulations cover sensible benchmarks and show plausible energy-savings trends, which supports the qualitative conclusion that joint optimization helps.\n\nSoft spots: In Section III-A, equation (15) uses the product surrogate f1(ρ)f2(l_v)+f1(ρ_v)f2(l) without the subtraction term -f1(ρ_v)f2(l_v), so at the current point the surrogate equals twice the original term. That violates the tightness condition needed for SCA convergence. Equation (17) is more severely wrong: it applies the same cubic-product surrogate to the bilinear term B_n=ρ l_d T/(B N_d), which is not of the form f1(ρ)f2(l_d) with f1=(1-ρ)^3. Thus as written, the convergence claim and the stationarity statement are unsupported. There is also a minor presentation issue: in Figure 4 the fixed mode reports zero energy after the sensing SINR becomes infeasible; that should be marked infeasible, not zero.\n\nThese are fixable. Adding the missing term and constructing a proper surrogate for the bilinear term (e.g., via the arithmetic-geometric mean inequality) would restore the convergence argument. The rest of the paper—the model, the problem formulation, and the simulation setup—holds up.\n\nWho would benefit: readers working on ISAC-assisted edge computing or SAGIN resource allocation will find the formulation useful as a case study. The paper deserves a serious referee, because the flaw is identifiable, local, and repairable. I'd send it to peer review with a recommendation of major revision, focusing the authors on fixing the surrogates and the infeasibility reporting.","headline":"A narrow but useful ISAC-SAGIN offloading formulation whose SCA convergence proof as written fails due to invalid surrogates.","tokens_in":10336,"tokens_out":5763,"would_cite":false,"duration_ms":55901,"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 tuning pilot time, data time, bit allocation, and offload ratio minimizes a flying edge server's energy in a satellite-and-drone IoT system.","keywords":["integrated sensing and communication","space-air-ground integrated network","computation offloading","energy minimization","successive convex approximation","Internet of Things","LEO satellite","autonomous aerial vehicle"],"falsifier":"Take any feasible (rho(v), ld,n(v)) with 0<rho<1 and evaluate (15) at rho=rho(v), ld,n=ld,n(v); the printed form equals twice the original computation energy in (11) rather than the same value, so one algebraic check decides whether the surrogate is locally exact. If it is not, the paper's own citation [15, Lemma 1] cannot support the convergence claim, and a numerical run of Algorithm 1 from different starting points would show whether the limit energies match the predicted stationary values.","tokens_in":9336,"feed_emoji":"📡","tokens_out":10547,"duration_ms":117646,"temperature":0.7,"pith_summary":"Integrated sensing and communication lets a flying edge server (an autonomous aerial vehicle, AAV) use the same uplink signal to sense a target and to receive IoT data, while excess computation is offloaded to a low-Earth-orbit satellite. The paper's central claim is that the AAV's total energy—channel-estimation, on-board computation, and satellite-uplink transmission—can be minimized by jointly choosing two time durations (training and data), the per-frame bit allocation, and the fraction of data offloaded, subject to data-offloading and sensing-accuracy constraints. Solving that non-convex problem is the hard part; the paper builds a successive convex approximation whose inner problem is strongly convex and claims convergence to a stationary point under a diminishing step size. Simulations show the joint choice beats benchmarks that optimize only one variable class (frame duration alone, offloading ratio alone, or neither). If the proof holds, the practical message is that an integrated sensing and communication setup can do double duty: one pilot signal serves both radar detection and channel estimation, and careful time/bit splitting materially lowers the aerial node's energy budget.","feed_headline":"Joint tuning of timing and offloading lowers drone IoT energy","feed_subtitle":"Jointly picking time split, bits, and offload share beats optimizing any one alone in drone-satellite IoT.","key_machinery":"The central mechanism is the surrogate decomposition of the AAV energy objective. Each cubic computation term is a product f1(rho) f2(ld,n) of two nonnegative convex functions, and the transmit term is an exponential of rho*ld,n; the paper replaces these with a strongly convex inner approximation that separates the two factors and adds proximal penalties, yielding a convex program with a unique solution. The outer loop over Np supplies the training/data split while the inner SCA loop updates rho and ld,n; convergence to a stationary point of the original problem is claimed to follow from the cited SCA theory when the surrogate satisfies the matching conditions.","core_discovery":"The paper formulates the AAV energy minimization as Problem (14): minimize the sum over N frames of channel-estimation energy, on-board computation energy, and AAV-to-LEO uplink transmission energy by choosing the training duration Np, data duration Nd, per-frame data bits ld,n, and partial offloading ratio rho, subject to a total offloaded-data requirement, per-frame achievable-rate constraints at the AAV and at the LEO satellite, a sensing SINR constraint, and rho in [0,1]. The non-convexity comes from cubic product terms in the computation energy and from an exponential transmit term in rho*ld,n. To handle it, the authors fix Np and Nd in an outer loop and, for each fixed pair, replace th","pith_inferences":["If the surrogate equality is repaired by adding the missing subtraction term so that the inner approximation matches the original functions at the current iterate, the same algorithm should still converge; a direct numerical check at one frame would settle whether the reported energy gains survive the correction.","The same f1*f2 surrogate trick should extend to multiple IoT devices or multiple targets, since those extensions add more product terms of the same shape; the convergence proof would need to be rebuilt term-by-term, not assumed.","Because flight energy and trajectory are fixed a priori, the reported total AAV energy excludes the largest physical cost of an AAV; treating trajectory as a decision variable could shift the optimal balance from offloading ratio toward where the AAV flies.","The claim that the offloading ratio matters more than frame duration is parameter-dependent; at much shorter AAV-to-satellite distances the exponential transmit term becomes cheaper, so frame-duration optimization might gain weight."],"forward_implications":["When the mission time is longer (more frames at fixed frame length), the per-frame energy denominators N_p^2 and N_d^2 shrink the energy needed per bit, so the jointly optimized schedule is most beneficial in long missions.","The offloading ratio is the dominant lever: optimizing only the frame duration while fixing rho=0.5 costs much more than optimizing rho alone.","Tight sensing SINR constraints push the optimizer to lengthen the pilot phase, which inflates computation and satellite-uplink energy; beyond about -1.5 dB a fixed pilot split cannot satisfy the sensing constraint at all.","The per-iteration cost of the inner SCA step is linear in the number of frames, so the full double-loop algorithm is affordable for frame counts in the hundreds."],"supporting_citations":[{"why":"Supplies the SAGIN hybrid edge computing setup, FDD equal-bandwidth assumption, LOS channel model, and many simulation parameter values used in the energy problem.","marker":"[2]"},{"why":"Motivates space/aerial-assisted offloading for IoT and provides the partial-offloading baseline (Opt-O) against which the joint optimization is compared.","marker":"[3]"},{"why":"Provides the ML channel estimate, clutter-plus-noise hypothesis model, whitening matrix, and the sensing SINR/detection machinery used in constraints (4)-(6) and (9).","marker":"[10]"},{"why":"Supplies the Neyman-Pearson detection rule used to set the sensing decision in (6).","marker":"[12]"},{"why":"Directly supplies the sensing SINR expression in (9) used as the sensing constraint (14e).","marker":"[13]"},{"why":"Gives the cubic channel-estimation energy formula (10), one of the three terms minimized in (13).","marker":"[14]"},{"why":"Provides the SCA Lemma 1 and Theorem 2 that the surrogate construction and convergence claim rely on; the central algorithmic justification.","marker":"[15]"},{"why":"Defines the MRT beamforming vector used in the AAV-to-LEO rate (8) and transmit energy (12).","marker":"[11]"},{"why":"Supplies the UAV-ISAC trajectory/beamforming benchmark (Opt-F) and the frame-optimization comparison.","marker":"[6]"},{"why":"Supplies the fixed mode benchmark (fixed Np=rho=0.5) used as a comparison baseline.","marker":"[16]"}],"fun_headline_variants":["Joint time-offload tuning cuts drone IoT energy","Energy-smart offloading: optimize timing and bits","Drone-satellite IoT joint optimization trims power","Optimize offload, timing, bits to save drone energy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The claimed convergence to a stationary point of Problem (14) rests on the surrogate expressions (15) and (17) satisfying the standard tangent condition—equal to the original non-convex terms at the current iterate—and the printed expressions do not obviously satisfy it; if that equality fails, the stationary-point guarantee from the cited SCA theory does not apply.","fun_headline_variants_meta":{"raw":{"variants":["Joint time-offload tuning cuts drone IoT energy","Energy-smart offloading: optimize timing and bits","Drone-satellite IoT joint optimization trims power","Optimize offload, timing, bits to save drone energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1402,"prompt_tokens":696,"completion_tokens":706,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":440,"tokens_out":706,"duration_ms":8311,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:58:30.257115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any feasible (rho(v), ld,n(v)) with 0<rho<1 and evaluate (15) at rho=rho(v), ld,n=ld,n(v); the printed form equals twice the original computation energy in (11) rather than the same value, so one algebraic check decides whether the surrogate is locally exact. If it is not, the paper's own citation [15, Lemma 1] cannot support the convergence claim, and a numerical run of Algorithm 1 from different starting points would show whether the limit energies match the predicted stationary values.","supporting_citations":[],"review_version":1}