{"id":"da560a1a-335b-4ed6-94b6-229521a9a887","arxiv_id":"1908.03119","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A scalable cell-free massive MIMO framework with dynamic user-centric cooperation clusters, a distributed access, pilot assignment, and cluster formation algorithm, and partial MMSE combining and precoding that nearly matches unscalable full-cooperation performance.","lead":"The paper designs a cell-free massive MIMO system where each access point serves only a small, fixed number of users, keeping per-access-point computation and front-haul traffic finite even as the total number of users grows without bound. It provides algorithms and signal-processing rules that, in simulations, nearly match the performance of unscalable schemes in which all access points cooperate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The scalability mechanism is sound, but the paper's claim of 'negligible performance loss' relies on uplink-only evidence; the downlink comparison lacks any unscalable baseline, leaving a headline assertion unsupported.","rationale":"I read the central claim as having two components: a per-AP complexity guarantee and a near-optimality claim. The complexity guarantee is rigorous under the stated Assumption 1, which is an explicit and standard Massive MIMO assumption; Definition 1 concerns complexity, not maintaining SE as K grows, so I do not treat the constancy of tau_p as the primary soft spot. The near-optimality claim, however, is only evidenced in the uplink. The downlink simulations intentionally focus on interference-limited operation and bound tightness, not on comparison with unscalable schemes, so the abstract's 'performs closely to the best unscalable alternatives' and Section VII's 'negligible performance loss' are unsupported for the downlink. This is a falsifiable gap in the evidence rather than an internal inconsistency, and it is addressable with a straightforward simulation using the released code. The reader's verdict is already CONDITIONAL, and this concern reinforces that conditionality without moving the verdict, so I recommend UNCHANGED. I credit the paper for releasing reproducible code, which makes the proposed test directly executable.","tokens_in":22429,"tokens_out":10572,"duration_ms":119784,"concrete_test":"Using the released Matlab code, rerun the Section VI downlink scenario (L=400, N=1, K=100, tau_p=10) with an unscalable centralized MMSE precoder: compute the collective MMSE precoding from (20)-(36) with all APs serving all UEs, apply the same network-wide equal power allocation rho_k=rho/tau_p used for P-MMSE, and plot its per-UE SE CDF alongside Fig. 6(a). If P-MMSE attains at least 90% of the unscalable baseline's average SE, the 'negligible loss' claim extends to the downlink; otherwise the conclusion must be restricted to the uplink or reworded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline has two parts: a complexity guarantee (Definition 1) and a performance claim ('performs closely to the best unscalable alternatives', abstract; 'negligible performance loss', Section VII). The first part is established by construction: under Assumption 1, |D_l| <= tau_p with tau_p fixed, and the Section V-A algorithm bounds per-AP estimation, combining/precoding, and fronthaul by tau_p and cluster size. The second part is only tested in the uplink. Figure 5 compares P-MMSE and LP-MMSE against unscalable '(All)' MMSE, L-MMSE, and MR benchmarks, and the losses are indeed small. Figure 6, the only downlink evaluation, compares three scalable precoders (P-MMSE, LP-MMSE, MR) with different power allocations from Section V-D; no unscalable downlink baseline appears. Thus the general statements in the abstract and conclusions go beyond the evidence: the downlink is exactly where the scalability constraints (at most tau_p UEs per AP, local CSI, distributed precoding) could cause the largest loss, and the centralized/distributed comparison is confounded by unequal power budgets (40x difference, Section VI-B). The scalability proof is unaffected, but the contribution's practical significance—that scalability costs little—is not demonstrated for the downlink direction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a framework for scalable Cell-Free Massive MIMO. The key idea is to bound the per-AP load by having each AP serve at most one UE per pilot sequence, with the pilot pool size τp independent of K. The authors define scalability (Definition 1), prove Lemma 1 that a constant number of served UEs per AP suffices, and develop a joint initial access, pilot assignment, and cluster formation algorithm (Section V-A). They derive spectral efficiency expressions for centralized and distributed uplink/downlink processing, propose partial MMSE (P-MMSE) and local partial MMSE (LP-MMSE) schemes, and prove an uplink-downlink duality that motivates using uplink combiners as downlink precoders. Numerical results show that in the uplink, P-MMSE and LP-MMSE perform close to unscalable MMSE benchmarks while MR is far worse; downlink results compare three scalable precoders under different power allocations.","tokens_in":22576,"tokens_out":9582,"duration_ms":98047,"significance":"If the scalability guarantee and the numerical claims hold, this is an important step toward practical Cell-Free Massive MIMO: it provides a formal complexity bound, concrete algorithms, and reproducible simulation code. The paper's main strength is that the scalability proof is clean and does not rely on fitted parameters; the P-MMSE and LP-MMSE schemes are evaluated against standard benchmarks in the uplink, and the complexity tables are consistent with the stated assumptions. However, the downlink performance claim is not supported by the current simulations, because Fig. 6 has no unscalable baseline and the centralized/distributed comparison is confounded by different power allocation policies. Since the downlink is the direction where scalable local processing could cause the largest loss, the practical significance of the framework is currently only demonstrated for the uplink.","major_comments":[{"comment":"The abstract and Section VII claim that the proposed scalable schemes 'perform closely to the best unscalable alternatives' and that 'the scalability can be achieved with a negligible performance loss'. This claim is only tested in the uplink (Fig. 5), where P-MMSE and LP-MMSE are compared against unscalable MMSE, L-MMSE, and MR baselines. The downlink evaluation in Fig. 6 compares only three scalable precoders (P-MMSE, LP-MMSE, MR) and uses different power allocation policies for the centralized and distributed schemes, with the distributed schemes using 40 times more transmit power (Section VI-B). Consequently, the downlink is missing both the unscalable baseline and a matched power allocation, so the margin by which scalability costs performance in the downlink is not quantified. The authors should either add downlink simulations with unscalable MMSE or L-MMSE precoding under the same power allocation rules, or restrict the 'negligible loss' statements to the uplink.","section":"Section VI-B, Fig. 6"},{"comment":"The scalability proof relies on Lemma 1, which requires |D_l| <= tau_p for every AP. The proposed access algorithm assigns each AP at most one UE per pilot through Steps 2 and 3, but the Master AP role is not explicitly load-limited: footnote 5 states that an AP 'can only be the Master AP of up to tau_p UEs', yet the algorithm does not enforce this, and the footnote proceeds to allow multiple UEs on the same pilot via time/frequency multiplexing in the 'unlikely event' of overloading. Such multiplexing would increase |D_l| beyond tau_p and break the bounded-complexity guarantee of Lemma 1. The authors should modify the algorithm to include a reservation or handover mechanism that guarantees the Master AP load is at most tau_p, or they should provide a probabilistic argument that the overload event has vanishing probability as K,L tend to infinity under the proposed access rule.","section":"Section V-A, Assumption 1 and footnote 5"}],"minor_comments":[{"comment":"The sentence 'Since we have established that the proposed scalable Cell-Free Massive MIMO provides very competitive performance' refers to uplink results only; it would be clearer to say 'very competitive uplink performance'.","section":"Section VI-B, first sentence"},{"comment":"The caption should state the power allocation policy used for each curve, since the centralized and distributed schemes have different total transmit power, which is material to the comparison.","section":"Figure 6, caption"},{"comment":"The phrase 'performs closely' should be 'perform closely' to agree with the compound subject.","section":"Abstract"},{"comment":"Even if an overload probability bound is outside the scope of the paper, the text should state explicitly that Assumption 1 is a condition on the algorithm's inputs, not a guaranteed consequence of the access procedure.","section":"Section V-A, footnote 5"},{"comment":"The perfect-synchronization assumption is flagged as practically infeasible; a brief discussion of the expected impact of imperfect synchronization on the SE bounds would improve the paper's completeness.","section":"Remark 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong and likely to be influential, but the downlink performance gap should be addressed before acceptance. The self-citation rate is high, but the references appear relevant and the technical content is novel enough to justify the paper. The scalability framework itself is sound under the stated assumptions; the main deficiency is the mismatch between the broad performance claims and the uplink-only evidence for the 'negligible loss' assertion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the scalability mechanism is real: under Assumption 1 (each AP serves at most one UE per pilot, with a fixed pilot pool tau_p), the per-AP complexity and fronthaul load are bounded independent of K. The Section V-A algorithm delivers on this, and the complexity tables check out. Second, the 'performs closely to the best unscalable alternatives' claim from the abstract is only demonstrated for the uplink. Figure 5 compares against full MMSE, L-MMSE, and MR baselines and the losses are small. Figure 6, the only downlink evaluation, compares three scalable precoders under different power allocations; there is no unscalable downlink baseline, and the centralized scheme is further handicapped by a 40x power difference. So the claim of near-optimal performance is not yet established in the direction where scalability constraints bite hardest.\n\nWhat is genuinely new: the joint initial access, pilot assignment, and cluster formation algorithm; the P-MMSE construction in (23) with the interference set P_k kept independent of K; the LP-MMSE scheme in (29); and the explicit accounting of what makes each task scalable. The SE expressions are standard bounds applied carefully. Code is released, and no free constants are fitted to make the numerics come out; the LP-MMSE advantage over MR is a Monte Carlo outcome. The authors are also honest about prior work, building on DCC and the textbook rather than pretending those did not exist.\n\nSoft spots, in rough order of severity. The downlink evidence gap is the main one; the abstract and conclusions should be toned down or the experiments extended. The power allocation comparison in Fig. 6 confounds precoding gain with power budget, so both scalable and unscalable schemes should run under the same policy. The duality proof in the appendix shows that Gamma - Sigma has the same eigenvalues as Gamma - Sigma^T, hence invertibility, but it never shows that the downlink power vector rho in (48) is nonnegative. This is probably fixable, but as written it is a gap. Assumption 1 is load-bearing: if tau_p has to scale with the number of users, the entire guarantee collapses. The authors flag it, and it is the standard Massive MIMO pilot assumption, but it deserves a sentence in the limitations. The perfect synchronization idealization is acknowledged in Remark 2, which is fair. Minor: the footnote in Section V-A about multiplexing when a Master AP is overloaded quietly allows Assumption 1 to be violated; not important for the architecture but worth tightening.\n\nWho is this for? Anyone working on distributed MIMO or cell-free massive MIMO as a 6G candidate. It is a serious framework with reproducible code and a clear complexity story. The central scalability result holds up; the overclaim is in the performance comparison, not the proof.\n\nSend it to peer review. I would ask for a revised version that adds unscalable downlink baselines, harmonizes the power allocation, and fixes the nonnegativity detail in the duality proof. That is a normal revision, not a rejection.","headline":"Genuinely sound scalability framework with real per-AP complexity bounds, but the 'negligible performance loss' headline is only supported in the uplink; worth serious review with revisions.","tokens_in":23196,"tokens_out":3191,"would_cite":true,"duration_ms":34203,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a scalable implementation of Cell-Free Massive MIMO in which each access point's computation, fronthaul, and channel-estimation load stay finite even as the number of users grows without bound.","keywords":["cell-free massive MIMO","scalability","dynamic cooperation clustering","pilot assignment","initial access","uplink-downlink duality","partial MMSE combining","fronthaul signaling"],"falsifier":"Measure or compute the per-AP fronthaul load and channel-estimation complexity in a setting where the pilot pool size $\\tau_p$ grows with the number of users $K$ (for instance, a fixed short coherence block with increasing $K$). If the per-AP load grows without bound under the proposed access and clustering algorithm, the central scalability claim is falsified. A more targeted check is to create pilot collisions in which two users sharing a pilot have comparable channel gains to the same AP, violating Assumption 1, and observe whether the AP's serving set and processing load grow with the number of such collisions.","tokens_in":22134,"feed_emoji":"📡","tokens_out":7233,"duration_ms":64008,"temperature":0.7,"pith_summary":"Cell-free massive MIMO lets many distributed access points cooperate to serve users without cell boundaries, but the original scheme requires every AP to process every user, so per-AP load grows with the number of users. This paper argues that scalability can be restored by using dynamic cooperation clustering: each AP serves a small cluster of users, with the cluster size capped by the pilot pool size rather than the total user count. It proposes a provably scalable three-step algorithm for joint initial access, pilot assignment, and cluster formation, and adapts channel estimation, combining, and precoding to the clustered setting. A new uplink-downlink duality lets downlink precoders be built from uplink combiners, and simulations show the scalable schemes approach the performance of the best unscalable alternatives while far outperforming maximum-ratio processing. If correct, the paper removes the main implementation objection to cell-free massive MIMO.","feed_headline":"Cell-free MIMO goes scalable: per-AP load stays finite","feed_subtitle":"New clustering bound ties each access point's work to pilot count, not user count; simulations show near-MMSE performance.","key_machinery":"The load-bearing mechanism is the dynamic cooperation clustering (DCC) framework, i.e., a set of diagonal matrices $D_{il}$ that decide which AP antennas may transmit to or decode signals from which UE, together with the scalability condition of Assumption 1: each AP serves at most one UE per pilot, so the serving set $D_l$ satisfies $|D_l| \\leq \\tau_p$ with $\\tau_p$ fixed as $K \\to \\infty$. This single bound caps channel estimation, combining/precoding, and fronthaul load per AP independently of $K$, turning Lemma 1 into a sufficient condition for Definition 1. The second load-bearing tool is the uplink-downlink duality of Proposition 4, which shows that for any uplink combining vectors and powers there exists a downlink power allocation with the same total power under which the downlink SINRs equal the uplink SINRs when the precoders are chosen as scaled combiners; this makes the scalability of uplink combining carry over to downlink precoding. The partial MMSE (P-MMSE) and local partial MMSE (LP-MMSE) schemes are the concrete designs that realize this, with per-AP complexity independent of $K$.","core_discovery":"The central claim is that Cell-Free Massive MIMO can be made scalable in a precise sense: under the paper's Definition 1, a network is scalable when each AP's channel estimation, signal processing, fronthaul signaling, and power-control complexity remain finite as the number of users $K$ goes to infinity. The paper first shows that the original all-users-to-all-APs form of Cell-Free Massive MIMO fails this definition on all four tasks, and then proposes a framework that satisfies it. The key move is to impose Assumption 1, that each AP serves at most one UE per pilot sequence with a pilot pool size $\\tau_p$ independent of $K$, which bounds the AP's serving set $|D_l| \\leq \\tau_p$ and makes all four tasks finite per AP. A three-step distributed access algorithm assigns each new UE a Master AP and a pilot, and lets neighboring APs decide whether to join the cluster, guaranteeing that every UE is served by at least one AP. The scalable combining and precoding schemes P-MMSE and LP-MMSE have complexity set by $|D_l|$ rather than $K$, and the new uplink-downlink duality (Proposition 4) shows that downlink precoders chosen as scaled uplink combiners preserve the SINR, transferring scalability from uplink to downlink. Simulations with 100 UEs show LP-MMSE achieving about 2.7 times the average spectral efficiency of maximum-ratio processing and centralized P-MMSE reaching 89% of the average spectral efficiency of the optimal unscalable MMSE combining.","pith_inferences":["A testable extension would let the pilot pool size adapt to the active-user density and map the tradeoff between pilot contamination and per-AP complexity, since the paper fixes $\\tau_p$ but the framework suggests a smooth degradation curve.","With high user mobility the steady-state load stays bounded, but the re-clustering overhead of repeatedly running the access algorithm may become the practical bottleneck; this is an implication the paper leaves implicit.","The same DCC-plus-pilot-bound design could be applied to other user-centric architectures, such as Fog Massive MIMO, where per-AP load would similarly be capped by a pilot-pool-sized serving set."],"forward_implications":["Per-AP fronthaul, computation, and channel-estimation load become bounded by the pilot pool size and cluster size, so adding users to a large network does not require upgrading APs or backhaul.","The three-step access algorithm guarantees every user at least one serving AP, avoiding the dropped-user problem of earlier user-centric clustering, and assigns pilots to minimize contamination at the Master AP.","Because downlink precoders can be derived from uplink combiners via the UL-DL duality, distributed APs can implement spatially selective transmission without sharing network-wide CSI.","The scalable P-MMSE and LP-MMSE schemes beat maximum-ratio processing by a large margin and approach the spectral efficiency of unscalable MMSE alternatives, indicating that scalability costs little in performance."],"supporting_citations":[{"why":"Supplies the MMSE estimation, capacity lower bounds, and hardening bound used to derive all spectral-efficiency expressions in the paper.","marker":"[2]"},{"why":"Defines the original Cell-Free Massive MIMO scheme with all APs serving all UEs, which the paper shows is unscalable and uses as the main benchmark.","marker":"[5]"},{"why":"Another original Cell-Free Massive MIMO formulation with network-wide power optimization, used as a scalability counter-example and performance benchmark.","marker":"[6]"},{"why":"Provides the MMSE combining and centralized/decentralized implementations that the scalable P-MMSE and LP-MMSE schemes are compared against.","marker":"[7]"},{"why":"Proposes the dynamic cooperation clustering framework that the scalable implementation builds on.","marker":"[17]"},{"why":"Textbook describing DCC and the guidelines for distributed Network MIMO that inspire the scalable implementation.","marker":"[10]"},{"why":"Provides the scalable power allocation heuristic used for distributed precoding in the proposed framework.","marker":"[33]"}],"fun_headline_variants":["Per-AP work stays finite as cell-free MIMO user count grows","Cell-free MIMO: per-AP complexity capped by pilot pool size","Scalable cell-free MIMO: load bounded, performance near MMSE","Bounded per-AP cost makes cell-free MIMO scalable to any K"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the pilot pool size stays constant and that each AP serves at most one user per pilot, so if the number of pilots must grow with the number of users, for example because coherence blocks are short or many users are highly mobile, the per-AP workload grows with $K$ and the scalability guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["Per-AP work stays finite as cell-free MIMO user count grows","Cell-free MIMO: per-AP complexity capped by pilot pool size","Scalable cell-free MIMO: load bounded, performance near MMSE","Bounded per-AP cost makes cell-free MIMO scalable to any K"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3505,"prompt_tokens":1066,"completion_tokens":2439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":2359}},"tokens_in":682,"tokens_out":2439,"duration_ms":19883,"temperature":1.0,"reasoning_tokens":2359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:24:42.083970+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure or compute the per-AP fronthaul load and channel-estimation complexity in a setting where the pilot pool size $\\tau_p$ grows with the number of users $K$ (for instance, a fixed short coherence block with increasing $K$). If the per-AP load grows without bound under the proposed access and clustering algorithm, the central scalability claim is falsified. A more targeted check is to create pilot collisions in which two users sharing a pilot have comparable channel gains to the same AP, violating Assumption 1, and observe whether the AP's serving set and processing load grow with the number of such collisions.","supporting_citations":[{"cited_title":"Massive MIMO networks: Spectral, energy, and hardware efficiency,","cited_arxiv_id":null,"evidence_quote":"Supplies the MMSE estimation, capacity lower bounds, and hardening bound used to derive all spectral-efficiency expressions in the paper."},{"cited_title":"Pre- coding and power optimization in cell-free Massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Another original Cell-Free Massive MIMO formulation with network-wide power optimization, used as a scalability counter-example and performance benchmark."},{"cited_title":"Making cell-free massive MIMO competitive with MMSE processing and centralized implementation,","cited_arxiv_id":null,"evidence_quote":"Provides the MMSE combining and centralized/decentralized implementations that the scalable P-MMSE and LP-MMSE schemes are compared against."},{"cited_title":"Optimality properties, distributed strategies, and measurement-based evaluation of coordinated multicell OFDMA transmission,","cited_arxiv_id":null,"evidence_quote":"Proposes the dynamic cooperation clustering framework that the scalable implementation builds on."},{"cited_title":"Optimal resource allocation in coordi- nated multi-cell systems,","cited_arxiv_id":null,"evidence_quote":"Textbook describing DCC and the guidelines for distributed Network MIMO that inspire the scalable implementation."},{"cited_title":"Scalability aspects of cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Provides the scalable power allocation heuristic used for distributed precoding in the proposed framework."}],"review_version":1}