{"id":"2aee89ac-72d4-419a-98cb-9b109eb36dc2","arxiv_id":"1908.00833","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Low-complexity successive-convexification algorithms for joint power control and user association in NOMA full-duplex systems reach within 1-2% of brute-force optimal spectral efficiency in simulation.","lead":"This paper presents two algorithms that jointly pick which users share a channel and how much power each gets, in a small cell that transmits and receives at the same time using NOMA and full-duplex radios. The methods target higher spectral efficiency with a minimum data rate for every user, and simulations show gains over conventional full-duplex and half-duplex schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rounding step (33) can yield binary associations that violate QoS or association constraints, and the reported 1–2% gap to BFS does not quantify this infeasibility.","rationale":"The reader's weakest assumption identifies the rounding-to-binary step as the load-bearing premise, and I agree. The paper explicitly concedes that the continuous relaxation can produce non-binary values and that the rounded point can be infeasible, yet Algorithm 1 and Algorithm 2 round anyway and the reported 1–2% optimality gap is computed from those rounded points. This is not a mathematical contradiction, but it is an unquantified empirical gap: if the fraction of infeasible rounded associations is non-trivial, the central claim that the proposed algorithms achieve near-optimal SE while satisfying per-user QoS is unsupported. The proposed test is a direct audit of that gap, requiring only a post-processing feasibility check and a re-averaging over feasible realizations. The convergence-theorem concern is real but secondary; it affects the strength of the theoretical justification, not the direct numerical comparison. The paper is otherwise carefully assembled: the tensor formulation, the ICA convexification, and the BFS benchmark are all reasonable, and the simulations are extensive. The verdict of CONDITIONAL remains appropriate because the gap is addressable in revision, and no evidence here suggests the algorithms are inherently broken. I therefore recommend no change to the reader's verdict.","tokens_in":31144,"tokens_out":5447,"duration_ms":56444,"concrete_test":"Reproduce Fig. 3(a) with a post-rounding feasibility audit: for every channel realization, after Algorithm 1 or 2 returns (α*,β*) via (33), re-optimize (w,p) with these binary variables fixed and check whether constraints (9d)–(9e) hold. Record (i) the fraction of realizations where the rounded association is infeasible, (ii) the distribution of QoS violations, and (iii) the average SE gap to ICA-BFS computed only over feasible realizations. If the infeasibility fraction is non-negligible or the feasible-only gap exceeds the reported 1–2%, the headline claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest link is the rounding step in Section III.A. After solving the continuous relaxation (12), the paper states that 'some relaxed variables are non-binary' and that the rounded point 'can be infeasible to problem (11)', then rounds via (33) and computes the final SE from the rounded point (Algorithm 1, Steps 8–9). Rounding a fractional doubly stochastic α or β to {0,1} does not guarantee preservation of the row/column constraints (11e), the UL ordering constraints (9j)–(9k), or the QoS constraints (9d)–(9e); no bound is given on the resulting SE loss or on the probability and magnitude of QoS violation. The central numerical claim in Section VI.B, that ICA-CR and ICA-CR-PF 'deviate only 1% ~ 2% from the optimal SE', is an average over points that may be infeasible for the very problem (11) they aim to solve. If infeasible realizations are included in the average or silently discarded, the 1–2% gap is not a valid measure of feasible performance. The convergence proof (Theorem 3) is also delegated to Property A of [34] and to the KKT-invexity condition of [38] without verification for the specific approximations (14)–(31), but the rounding issue is the more directly falsifiable threat to the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies joint downlink user clustering and uplink decoding-order selection together with power and beamforming control in a full-duplex NOMA small-cell system. The authors formulate a mixed-integer nonconvex sum-spectral-efficiency maximization problem (9)/(11), relax the binary association variables, and develop two inner-convex-approximation (ICA) based algorithms: Algorithm 1 (ICA-CR) and Algorithm 2 (ICA-CR-PF), the latter adding a penalty function to drive the relaxed variables toward binary values. Algorithm 3 is a brute-force benchmark that enumerates K!×L! associations and applies ICA to each resulting nonconvex power-control subproblem. Numerical results report that the proposed algorithms come within 1–2% of the BFS benchmark spectral efficiency, outperform conventional FD, FD-NOMA with random user association, and HD-NOMA baselines, and exhibit convergence over iterations. The paper also includes an extended discussion of three-zone NOMA, SI/CCI effects, and convergence behavior.","tokens_in":31296,"tokens_out":8171,"duration_ms":86161,"significance":"If the performance claims are correct, the paper makes a useful contribution to full-duplex NOMA resource allocation. The tensor-based user-association model is a genuine structural idea that reduces the number of binary variables; the ICA construction yields disciplined subproblems containing only SOC and linear constraints; and the penalty variant is a reasonable mechanism for accelerating binary recovery. The numerical study is extensive, with many topologies and channel realizations, and the BFS benchmark provides an independent comparison point. The paper does not ship code, but the structural parts—Lemma 1, the SOC reformulations, and the penalty-method theorem—are plausible and are partly corroborated by the simulations. The main value of the paper would be in the low-complexity SOCP per iteration and the reported small gap to a systematic association search, provided the feasibility of the rounded output and the benchmark's optimality claims are properly qualified.","major_comments":[{"comment":"The rounding step is load-bearing and currently unsupported. The paper states immediately before (33) that the ICA-CR solution 'yields nearly binary values at convergence, but some relaxed variables are non-binary' and that the rounded point 'can be infeasible to problem (11)'; Algorithm 1 nevertheless rounds α and β to binary values and then computes the reported spectral efficiency. Rounding can violate the row/column constraints (11e), the UL ordering constraints (9j)–(9k), and the QoS constraints (9d)–(9e). No bound is given on the resulting SE loss or on the probability and magnitude of QoS violation, so the 1–2% gap reported in Section VI-B is not established for feasible outputs. The paper also acknowledges infeasibility in Fig. 8(c) but does not state how infeasible trials are treated in the averages of Figs. 3 and 4. The authors should report the feasibility rate of the rounded points, exclude or repair infeasible realizations before averaging, or provide a rounding guarantee with quantified SE and QoS loss.","section":"Section III-A, Eq. (33); Algorithm 1, Steps 8–9; Section VI-B"},{"comment":"The convergence claim is delegated rather than proved. The proof asserts that the approximate functions satisfy 'Property A' of [34] and that the feasible sets satisfy the 'connectedness condition for KKT invexity' of [38], but neither hypothesis is verified for the specific approximations (14)–(31) and the relaxed feasible region (12). Because Algorithm 1 may terminate with a rounded point and Algorithm 2 modifies the objective with penalty terms, the monotone-improvement and local-optimality statement needs a direct argument or an explicit verification of the cited hypotheses. Please either provide that verification or weaken the claim to numerically observed convergence.","section":"Section V-B, Theorem 3"},{"comment":"The BFS benchmark is not a global optimum as labeled. Algorithm 3 enumerates all K!×L! associations, but for each association it solves the nonconvex problem (41) using the ICA method, which only guarantees a local optimum of that subproblem. Therefore the phrase 'optimal SE' in Section VI-B and the 'optimal solution' claim for Algorithm 3 are not justified; the benchmark is the best among many locally optimized association-specific solutions. Please rephrase these claims as 'best BFS-local solution' and state the 1–2% gap relative to that benchmark, not to a proven global optimum.","section":"Section IV, Algorithm 3; Section VI-B"},{"comment":"The equivalence used to convert the max in (10b) into the reciprocal-min form in (20) is not established for the relaxed problem. For binary α with one nonzero entry per column the identity is plausible only with the ε regularization, but for fractional α in (12), max_k α_kj A_k and min_k A_k/(α_kj+ε) are different objects. Consequently, constraints (22a) and (23a) may not be valid inner approximations of the SINR constraint (10b) during the relaxation, and the corresponding ICA minorant is not guaranteed to be a lower bound. Please provide a formal derivation of (20) for the relaxed domain or state explicitly that the reciprocal-min form is a heuristic and quantify the resulting error in the reported rates.","section":"Section III-A, Eq. (20) and (22a)"}],"minor_comments":[{"comment":"There are several typos: 'Transact ions' and 'witho ut' in the header line, 'Algorithmsp' in Theorem 3, and '3:=' instead of '3:' in Algorithm 3's initialization.","section":"General"},{"comment":"The legend label 'ICA-CRP-PF' should read 'ICA-CR-PF'.","section":"Fig. 8(c)"},{"comment":"The text around (C.5) contains a stray 'B.' in 'ln(P max bs B.σ 2)', which should be removed or replaced with the intended notation.","section":"Appendix C, Eq. (C.5)"},{"comment":"Reference [31] lists pages '4483–4454', which appears to be a reversed or incorrect page range and should be checked against the original source.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the algorithmic architecture is plausible, but the headline '1–2% from optimal SE' claim depends on two unverified steps: the feasibility of rounded binary variables and the optimality of the BFS benchmark. I would not require a global optimality proof for the benchmark, but the terminology should be corrected and the post-rounding feasibility must be quantified. If the authors can re-run the simulations with infeasible points either removed or repaired, and provide the requested verifications, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading if you work on full-duplex/NOMA resource allocation. The genuinely new piece is the joint optimization of DL user clustering (tensor T) and UL decoding order (β), combined with power control under QoS constraints. The related-work review supports the novelty: prior FD-NOMA designs used random UL order or fixed DL clusters. The ICA machinery is competently assembled, and the brute-force search benchmark (Algorithm 3) is a real effort to get a reference optimum rather than just comparing against themselves.\n\nThe main soft spot is exactly what the stress-test flags: the rounding step (33). The authors state that some relaxed variables are non-binary and that the rounded point \"can be infeasible to problem (11)\", then round anyway and compute the final SE from the rounded point. No bound is given on SE loss or QoS violation probability. The reported 1–2% gap is an average over these rounded points; if some of them violate the rate constraints, that gap is optimistic for feasible performance. This is not fatal, but it needs addressing: report the fraction of realizations where rounding preserves feasibility, and either discard or repair infeasible rounds before computing the gap.\n\nThe convergence proof (Theorem 3) is delegated to Property A of [34] and the KKT-invexity condition of [38]. That is common in this area, but the specific approximations (14)–(31) are not verified for those conditions; a comment noting the gap would suffice. The penalty parameter search is heuristic and disclosed, which is acceptable.\n\nBottom line: the formulation and the algorithm design are solid enough to deserve referee time. This paper is for researchers working on FD-NOMA or user-association optimization; a careful reader will get useful techniques and a clear statement of the joint problem. With a modest revision that quantifies rounding feasibility, the paper would be publishable. I would send it to review.","headline":"A competent FD-NOMA resource-allocation paper with a genuinely new joint association formulation, but the headline 1–2% gap to brute-force rests on a rounding step the authors admit can violate QoS.","tokens_in":31993,"tokens_out":1711,"would_cite":false,"duration_ms":18465,"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":"For full-duplex NOMA small cells, optimizing user pairing and uplink decoding order jointly with power control gets within 1–2% of the exhaustive optimum at polynomial per-iteration cost.","keywords":["full-duplex radios","non-orthogonal multiple access","non-convex programming","self-interference","spectral efficiency","successive interference cancellation","user clustering","power control"],"falsifier":"Run Algorithms 1 and 2 over thousands of random small-cell channel realizations, record the largest non-binary entries of the relaxed association variables at convergence, then round via (33) and compare the rounded spectral efficiency and feasibility against the brute-force benchmark of Algorithm 3. If any rounded point violates a rate constraint or misses the 1–2% gap, the central performance claim is falsified.","tokens_in":30796,"feed_emoji":"📶","tokens_out":8756,"duration_ms":85851,"temperature":0.7,"pith_summary":"This paper tries to establish that user association—which downlink users are paired in the same NOMA cluster and in what order the base station decodes uplink users—can be optimized together with transmit powers and beamforming in a full-duplex NOMA small cell, and that the resulting mixed-integer non-convex problem is tractable in practice. The authors relax the binary association variables, apply inner convex approximation, and add a penalty term to push the relaxed variables toward binary. In simulation, both proposed algorithms land within 1–2% of the spectral efficiency of brute-force enumeration over all pairings and decoding orders, while random pairing and conventional full-duplex designs fall further behind. That matters because full-duplex NOMA inherits self-interference, co-channel interference, and multiuser interference; the claim is that explicitly choosing the association is what keeps those in check while still meeting per-user rate targets.","feed_headline":"User pairing algorithm lands full-duplex NOMA within 1–2% of optimal","feed_subtitle":"Relaxation plus a penalty term beats random pairing and standard full-duplex in simulated small cells.","key_machinery":"The central object is the third-order tensor $T$ of zone-to-cluster association matrices, with $C_1 = I_K$ and all $C_i$ permutation matrices; the pairing between any two zones is recovered by $T^{iz}=C_i^{\\mathsf{T}}C_z$, which cuts the number of association variables from $\\binom{Z}{2}K^2$ to $(Z-1)K^2$. The argument runs on an inner-convex-approximation loop: each non-concave log-rate is lower-bounded by a concave minorant (e.g., $\\ln(1+1/\\omega)\\ge A(\\omega^{(\\kappa)})+B(\\omega^{(\\kappa)})\\omega$), each interference term is upper-bounded by the convex majorant of Lemma 1, and the resulting constraints are SOC-representable. The second algorithm adds the penalty $f_p(\\alpha_n)=\\varrho_n(\\alpha_n^2-\\alpha_n)$ to force the relaxed binaries to 0 or 1, with the log-sum-exp trick making the distinct-decoding-order constraint convex.","core_discovery":"On the paper's own terms, the central discovery is that joint user association and power control for FD-NOMA can be attacked through a continuous relaxation that stays nearly binary at convergence, paired with a sequence of second-order cone programs. Downlink clustering is encoded as a tensor of permutation matrices, so any two-zone association matrix is the change-of-basis product $T^{iz} = C_i^{\\mathsf{T}}C_z$; uplink decoding order is encoded by a binary matrix $\\beta$ subject to order constraints. The inner convex approximation turns the non-concave log rates and interference terms into concave minorants and SOC constraints, and the penalty variant adds $\\varrho_n(\\alpha_n^2-\\alpha_n)$ to drive the association variables to 0 or 1. The paper reports that the two ICA-based algorithms \"deviate only 1% ~ 2% from the optimal SE\" relative to a brute-force benchmark, and that they improve the achievable SE at every iteration and converge, beating conventional FD, FD-NOMA with random UA, and HD-NOMA in the simulated scenarios.","pith_inferences":["The same relaxation-plus-penalty recipe should transfer to other combinatorial resource-allocation problems in wireless design, such as subcarrier assignment, antenna selection, or cluster-size optimization, wherever the integer variables enter through products with continuous powers.","The tensor/permutation-matrix view identifies user pairing with the group of permutation matrices, which suggests an exact search or a bound on rounding over doubly stochastic relaxations rather than heuristic rounding.","The 1–2% gap is demonstrated under perfect CSI in slow-fading small cells; a natural next test is whether the gap widens with imperfect channel knowledge, since all convexifications use the true channels."],"forward_implications":["Both proposed algorithms solve one convex SOC program per iteration, so their per-iteration cost is polynomial in the numbers of antennas and users, unlike the $K!\\times L!$ brute-force benchmark.","The tensor representation cuts the number of downlink association variables from $\\binom{Z}{2}K^2$ to $(Z-1)K^2$ without losing the ability to recover any zone-to-zone pairing.","In the simulated small-cell configurations, the penalty variant converges faster and reaches a higher spectral efficiency than the plain ICA-CR variant, and both stay within 1–2% of the brute-force optimum.","Optimizing either the downlink pairing or the uplink decoding order alone already gives large gains over random association, and joint optimization gives the best results.","Whether two-zone or three-zone NOMA is better depends on the cell: three-zone NOMA wins in macro-cells, while two-zone NOMA is preferred in small cells."],"supporting_citations":[{"why":"Defines the conventional full-duplex small-cell system without NOMA that serves as the baseline to beat.","marker":"[15]"},{"why":"Supplies the two-zone NOMA pairing model and the inner-approximation inequality used to convexify the downlink rates.","marker":"[32]"},{"why":"Provides the sequential parametric convex approximation method whose convergence property is invoked for the proposed algorithms.","marker":"[34]"},{"why":"Gives the circulator-based full-duplex radio prototype and self-interference parameters used in the system model and simulations.","marker":"[12]"},{"why":"Supplies the Rician self-interference channel model and the FD multiuser interference treatment that the paper extends to NOMA.","marker":"[19]"},{"why":"Provides the earlier FD-NOMA power and subcarrier allocation problem that this work generalizes with joint user association.","marker":"[22]"},{"why":"Provides the concave quadratic minorant used for the uplink rates in the brute-force subproblems.","marker":"[37]"},{"why":"Supplies the KKT-invexity connectedness condition invoked to conclude convergence to a local optimum.","marker":"[38]"},{"why":"Establishes the multi-zone NOMA clustering idea from which the tensor model and two-zone case study are drawn.","marker":"[9]"}],"fun_headline_variants":["Relaxation plus penalty puts full-duplex NOMA within 2% of optimal","Joint power-user association steers FD-NOMA to near-optimal rates","Near-optimal FD-NOMA via relaxed user association and power control","Tensor encoding and relaxation close gap to optimal FD-NOMA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the relaxed user-pairing and decoding-order variables end up almost exactly 0 or 1, so the rounding step that turns them into binary choices does not lose spectral efficiency or violate the users' rate requirements.","fun_headline_variants_meta":{"raw":{"variants":["Relaxation plus penalty puts full-duplex NOMA within 2% of optimal","Joint power-user association steers FD-NOMA to near-optimal rates","Near-optimal FD-NOMA via relaxed user association and power control","Tensor encoding and relaxation close gap to optimal FD-NOMA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3262,"prompt_tokens":1005,"completion_tokens":2257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2179}},"tokens_in":621,"tokens_out":2257,"duration_ms":15910,"temperature":1.0,"reasoning_tokens":2179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:31:49.524687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithms 1 and 2 over thousands of random small-cell channel realizations, record the largest non-binary entries of the relaxed association variables at convergence, then round via (33) and compare the rounded spectral efficiency and feasibility against the brute-force benchmark of Algorithm 3. If any rounded point violates a rate constraint or misses the 1–2% gap, the central performance claim is falsified.","supporting_citations":[{"cited_title":"On the spectral efficiency of full-duplex small cell wireless s ystems,","cited_arxiv_id":null,"evidence_quote":"Defines the conventional full-duplex small-cell system without NOMA that serves as the baseline to beat."},{"cited_title":"Precoder design for signal superposition in MIMO-N OMA multicell networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-zone NOMA pairing model and the inner-approximation inequality used to convexify the downlink rates."},{"cited_title":"A sequential parametric convex approximation method with applications to nonconvex truss topology design problems,","cited_arxiv_id":null,"evidence_quote":"Provides the sequential parametric convex approximation method whose convergence property is invoked for the proposed algorithms."},{"cited_title":"Full duplex radi os,","cited_arxiv_id":null,"evidence_quote":"Gives the circulator-based full-duplex radio prototype and self-interference parameters used in the system model and simulations."},{"cited_title":"A new design paradigm for secure full-duplex multiuser syst ems,","cited_arxiv_id":null,"evidence_quote":"Supplies the Rician self-interference channel model and the FD multiuser interference treatment that the paper extends to NOMA."},{"cited_title":"Optimal join t power and subcarrier allocation for full-duplex multicar rier non-orthogonal multiple access systems,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier FD-NOMA power and subcarrier allocation problem that this work generalizes with joint user association."},{"cited_title":"Spectral and energy efficiencies in full-duplex wir eless information and power transfer,","cited_arxiv_id":null,"evidence_quote":"Provides the concave quadratic minorant used for the uplink rates in the brute-force subproblems."},{"cited_title":"Invex Optimization Revisited","cited_arxiv_id":"1707.01554","evidence_quote":"Supplies the KKT-invexity connectedness condition invoked to conclude convergence to a local optimum."},{"cited_title":"Low complex ity beamforming and user selection schemes for 5G MIMO-NOMA systems,","cited_arxiv_id":null,"evidence_quote":"Establishes the multi-zone NOMA clustering idea from which the tensor model and two-zone case study are drawn."}],"review_version":1}