{"id":"31be8264-af01-444a-9917-dd09ee5e4356","arxiv_id":"1908.05030","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A hierarchical scheme that mixes over-the-air computation with orthogonal transmission is analyzed for multi-layer wireless networks, with closed-form resource allocation.","lead":"This paper proposes a way to compute aggregate functions, like averages, across many wireless devices arranged in messy, multi-hop networks. It combines two known communication techniques, over-the-air computation and orthogonal transmission, and derives formulas for how fast the network can compute.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof of the group-function rate divides every subgroup count by H(f(bv)) instead of the subgroup function's own entropy; with unequal subgroup entropies Eq. (14)/(17) and all Section V optimizations do not follow.","rationale":"The central claim is Theorem 2's general computation rate and the closed-form optimizations built on it. For that claim to hold, the number of subgroup-function values computed at a relay must be measured consistently. The proof of Theorem 1 uses U^(c) = R^(c) |T^(c)| / H(f(bv)) for every subgroup, whereas Definition 2 applied to the subgroup function requires its own entropy in the denominator. When subgroup functions have unequal entropies, the min in Eq. (16) is taken over incomparable quantities, so Eq. (14) and hence Eq. (17) do not follow. The simple two-subgroup example with entropies 1 and 10 bits shows the discrepancy is concrete and not a normalization convention: the published formula misidentifies the bottleneck. Because Sections V and VI use Eq. (17) as the objective for time allocation and power control, the reported optimal closed forms and the claimed generalization of CoMAC and orthogonal schemes are unsupported as stated. I found no independent support in the manuscript: there is no formal verification, and the simulations merely evaluate the derived formula rather than test it. The reader's weakest assumption correctly identified the implicit entropy assumption in Eq. (16); I agree with that identification. A secondary defect is the invalid Jensen step in Eq. (21), where C+(x) = max(1/2 log x, 0) is not concave, but the entropy-normalization error is already sufficient to reject the central claim as stated.","tokens_in":19556,"tokens_out":9825,"duration_ms":103607,"concrete_test":"Re-derive Eq. (16) for a minimal two-subgroup network: L=2, K1=2, C=2 subgroups each containing one node, desired function f=(S1,S2) with H(S1)=1 and H(S2)=10 bits, identical symmetric fading and equal per-subgroup average rate R0, and beta1=beta2=1/2. Apply Definition 2 separately to each subgroup to write U^(c) in terms of H(f^(c)_{l,k}) rather than H(f(bv)): U^(1) = R0 * (n/2) * 1 and U^(2) = R0 * (n/2) * 10. Then compute the resulting number of desired-function values per channel use and compare it with min_c beta_c E[R^(c)] = 0.5R0. If the two rates differ, Theorem 1 is false as stated; repeat with H(S1)=H(S2) to confirm the published formula only holds in the equal-entropy special case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is in the proof of Theorem 1, Eq. (16). Definition 2 defines an achievable computation rate for a target function as R = lim T_d/(n H(target)); when the same definition is applied to a subgroup function f^(c)_{l,k}, the number of subgroup-function values obtained during |T^(c)_{l,k}| channel uses must be U^(c) = R^(c) |T^(c)| / H(f^(c)_{l,k}(...)). The paper instead writes U^(c) = R^(c) |T^(c)| / H(f(bv)) for every subgroup, implicitly assuming all subgroup functions have the same entropy as the desired function. That is false in general: a subgroup sums only a subset of sources, so its function has smaller entropy. With unequal subgroup entropies, the bottleneck 'min_c U^(c)' in Eq. (16) is not equivalent to 'min_c beta^(c)_{l,k} E[R^(c)_{l,k}]', and the group-function rate R_{l,k} in Theorem 1 does not follow. Since Theorem 2's general rate in Eq. (17) is obtained by chaining these group rates, and the closed-form allocations in Section V optimize that formula, the central claim is unsupported as stated. A minimal counterexample: L=2, K1=2, C=2 subgroups of size one, desired function f=(S1,S2) with H(S1)=1 bit, H(S2)=10 bits, equal symmetric rates R0 and beta=1/2; Eq. (14) gives 0.5R0, while counting with correct entropies gives a rate that is bottlenecked by the second subgroup and differs from 0.5R0. The gap is not a constant-factor artifact; it changes which subgroup is the bottleneck.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers computation of a desired function of distributed sources in a wireless network whose topology may be disorganized. The authors propose reorganizing such a network into a layered hierarchy of groups and subgroups, and then introduce multi-layer function computation (ML-FC) in which each subgroup function is computed by CoMAC and each group function is reconstructed by orthogonal communication over the subgroup outputs. The main theoretical claim is a closed-form computation rate for the desired function, expressed as a minimum over layers, groups, and subgroups of time-sharing factors times per-subgroup CoMAC rates (Theorems 1 and 2). The paper then formulates time-allocation and power-control problems and derives closed-form solutions, with simulation results comparing fixed and adaptive power control and optimal versus equal time allocation. The central claim is that this framework generalizes both classical CoMAC and orthogonal communication schemes to multi-hop settings.","tokens_in":20051,"tokens_out":12321,"duration_ms":131400,"significance":"If the rate formula and optimization results were correct, the paper would provide a useful design rule for multi-hop over-the-air computation and would extend CoMAC beyond single-hop relay-free networks. The problem is timely, and the hierarchical decomposition of the network into subgroups and groups is a natural and appealing conceptual step. The manuscript also gives closed-form resource-allocation expressions and simulations over Rayleigh fading channels, which would be of practical interest. However, the derivation of the main rate formula contains a load-bearing error in the conversion from computation rates to numbers of computed values, and the optimization section applies Jensen's inequality to a function that is not concave. These issues invalidate the central claims as stated, so the current contribution is not established.","major_comments":[{"comment":"The proof of Theorem 1 mishandles the relation between computation rate and the number of computed function values. Definition 2 states R = lim T_d/(n H(f)), so the number of values obtained in |T| channel uses is T_d = R |T| H(f), not R |T| / H(f) as written for U^(c)_l,k. The displayed equation for U^(c)_l,k is therefore dimensionally inconsistent, and the cancellation leading to the final expression in Eq. (16) is invalid. Independently, the normalization should use the entropy H(f^(c)_l,k) of the subgroup function, not H(f(bv)) of the desired function; these entropies differ in general, since a subgroup function typically depends on only a subset of sources. For example, with two one-source subgroups computing f=(S1,S2), H(S1)=1 bit, H(S2)=10 bits, equal per-subgroup rates R0 and beta=0.5, the paper's Eq. (14) gives 0.5 R0, while the entropy-corrected computation gives a rate bottlenecked by the first subgroup and equal to R0/22. The error changes which subgroup is the bottleneck, so it is not a constant-factor issue. Since Theorem 2 and all of Section V are built on Eq. (16), the claimed general rate and closed-form allocations are unsupported.","section":"Section IV-B, Eq. (16)"},{"comment":"The Jensen step in Eq. (21) is invalid because C+(x) = max{0, 0.5 log x} is not concave on the positive reals. A concrete counterexample is X=0.5 with probability 0.5 and X=1.5 with probability 0.5: E[X]=1, so C+(E[X])=0, while E[C+(X)] = 0.25 log(1.5) > 0. Thus the asserted inequality E[C+(1/K + min |h|^2 P)] <= C+(1/K + E[min |h|^2] P) is false. Consequently Problem 1 and Problem 2 do not optimize the actual computation rate, but rather a rate expression obtained through an invalid upper bound. The subsequent closed-form solutions and the simulation comparisons based on those solutions therefore do not establish the claimed performance of the proposed time allocation. The same non-concavity also casts doubt on the convexity argument for Problem 4.","section":"Section V-A, Eq. (21)"},{"comment":"Remark 3 asserts without proof that every disorganized network can be equivalently represented as a layered hierarchy in which each node has exactly one transmission destination. In a general wireless network a node's transmission may be received by multiple relays, and a node may have multiple intended destinations, but the reorganization in Fig. 3 does not preserve such connectivity. As stated, the result applies only to networks with a single-destination routing structure. The authors should either prove the equivalence under explicit conditions or restrict the claims about 'disorganized networks' accordingly. This is a scope issue for the paper's main title and motivation, and it is not addressed in the system model.","section":"Section III-A, Remark 3"}],"minor_comments":[{"comment":"The step in Algorithm 1 that says 'The given channel uses for the group {N1,1,...,N1,5} belongs a set T2,1' should say 'belong to a set T2,1' and should specify whether the group includes all five nodes or only those in the depicted subgroup.","section":"Section IV-A, Algorithm 1"},{"comment":"The equality between the sample average (1/|T^(c)|) sum over m in T^(c) and an expectation E[.] is only justified as an ergodic limit as the block length grows. It would be clearer to state this limit explicitly rather than writing the equality for a finite set of channel uses.","section":"Eq. (14) and Eq. (17)"},{"comment":"The denominator E[ min_i |h|^2 / |h|^2 ] uses h as a representative coefficient without defining its distribution or its relation to the channels in the subgroup. The expression would be clearer if written with explicit indices, for example E[ |h_i->k|^2 / |h_j->k|^2 ] under the stated symmetry assumption.","section":"Section V-B, Eq. (34)"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this paper has a sensible framing but a load-bearing error in the proof of its main theorem. I do not think the rate formulas are valid as stated, and the resource allocation sections inherit the problem.\n\nWhat is genuinely new: the idea of reorganizing a \"disorganized\" network into a layered hierarchy of subgroups and groups, then running CoMAC within subgroups and orthogonal communication within groups, is a natural and useful extension of the relay-free CoMAC framework. The reductions to existing results (L=2, C=1 recovers [21]; orthogonal time-sharing is also recovered) are correct and show the authors know the prior art. The closed-form time/power allocation is a nice exercise and would be valuable if the underlying rate expression were correct. The paper is also clearly written and the simulations are matched to the formulas.\n\nThe soft spots are not minor. In the proof of Theorem 1, Eq. (16), the number of subgroup-function values obtained in |T| channel uses is written as U = R|T| / H(f(bv)) for every subgroup. But Definition 2 defines the rate of a function relative to that function's own entropy. For a subgroup function f^(c) that depends on a subset of sources, H(f^(c)) is generally smaller than H(f(bv)). Using the same denominator for all subgroups makes the bottleneck \"min_c U^(c)\" meaningless; a subgroup with high entropy but high rate can be the bottleneck, and the min in Eq. (14) does not follow. Theorem 2 chains these group rates, so the general rate formula is unsupported. I checked the stress-test counterexample (two subgroups with H=1 and H=10); it works. This is not a constant-factor artifact; it flips which subgroup is the bottleneck.\n\nThere is also a second, independent error in Eq. (21): Jensen's inequality is applied to C+(x)=max(0,0.5 log x), which is not concave on x>0. That step is used to move the expectation inside C+, so the fixed-power optimization and the claimed closed form are not justified.\n\nRemark 3's equivalence claim (any disorganized network can be represented as such a hierarchy) is asserted without proof and seems too strong for arbitrary topologies.\n\nIn short: the framework and the special cases are right, but the general rate theorem and the optimization results built on it do not stand. This needs a real revision before it can be trusted. If I were editing, I would send it to review rather than desk-reject, because the idea is relevant and the errors are specific enough to fix. But the current version should not be accepted.\n\nWho is this for? Researchers working on CoMAC and hierarchical data aggregation will find the model and the special-case checks useful, and future work might fix the entropy issue. I would not cite the rate formula as it stands.\n\nBest.","headline":"A promising hierarchical CoMAC framework undermined by a load-bearing entropy normalization error and an invalid Jensen step; the special cases are right but the general rate formula does not hold as stated.","tokens_in":20476,"tokens_out":3774,"would_cite":false,"duration_ms":33085,"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 shows that a disorganized wireless network can be reorganized into layers so that the fusion center's function-computation rate is the minimum of simple per-hop CoMAC rates, generalizing over-the-air computation and orthogonal…","keywords":["function computation","CoMAC","hierarchical networks","disorganized wireless networks","achievable computation rate","time allocation","power control","nested lattice codes"],"falsifier":"Simulate or build a three-layer network with four source nodes, two second-layer relays, and one fusion center under i.i.d. Rayleigh fading, and measure the maximum number of desired-function values per channel use with optimal time allocation; if the measured rate disagrees with Eq. (29), the paper's central claim is wrong.","tokens_in":19358,"feed_emoji":"📡","tokens_out":12107,"duration_ms":109992,"temperature":0.7,"pith_summary":"The paper is trying to prove that function computation in wireless data aggregation can be extended from single-hop networks to arbitrary multi-hop, disorganized topologies without losing the efficiency of over-the-air computation. It proposes multi-layer function computation, which reorganizes the topology into layers of subgroups and groups, computes subgroup functions by concurrent transmissions (CoMAC), combines them with orthogonal communication, and reconstructs the desired function hop by hop. The central result is a closed-form computation rate equal to the minimum over all hops of the per-subgroup rate, so the worst subgroup in the worst layer governs the whole network. If the result holds, network designers get a single formula for allocating time and power, and existing CoMAC and orthogonal communication schemes become special cases of one general scheme.","feed_headline":"Disorganized wireless networks can compute functions at one rate","feed_subtitle":"The rate is the worst per-hop over-the-air computation rate; ordinary CoMAC and time-sharing are special cases.","key_machinery":"The central object is the layered hierarchical reorganization: every layer's nodes are partitioned into groups, each group into subgroups; within a subgroup all nodes transmit simultaneously so CoMAC computes a subgroup function, and within a group the subgroup functions are forwarded over orthogonal time slots so the node reconstructs the group function. The load-bearing recurrence is Eq. (18), which states that the number of function values available at a node is the minimum of what its parents supply and what its own subgroups compute; Theorem 2 is the repeated minimum of these per-hop rates. The physical-layer engine that makes each per-subgroup rate achievable is a sequence of nested lattice codes.","core_discovery":"The paper establishes that any disorganized network whose nodes each have a single transmission destination can be reorganized into an $L$-layer hierarchy, and that in this hierarchy the fusion center can compute the desired function at rate $$R=\\min_{l\\in[2:L]}\\min_{k\\in\\mathcal{K}_l}\\alpha_{l,k}\\min_{c\\in\\mathcal{C}_{\\mathcal{N}_{l,k}}}\\$beta^{{(c)}}$_{l,k}\\mathbb{E}\\left[C_+\\left(\\frac{1}{$K^{{(c)}}$_{\\mathcal{N}_{l,k}}}+\\min_{i\\in\\mathcal{K}^{(c)}_{\\mathcal{N}_{l,k}}}|$h^{{i\\to k}}$_{l-1}|^$2P^{{i\\to k}}$_{l-1}\\right)\\right],$$ where $C_+(x)=\\max\\{\\frac{1}{2}\\log x,0\\}$, $\\alpha_{l,k}$ is the fraction of channel uses given to node $N_{l,k}$, and $\\beta^{(c)}_{l,k}$ is the fraction given to subgroup $c$. The proof is a recurrence: the number of desired-function values that survive to a node is the minimum of the numbers its parent groups supply and the numbers its own subgroups compute, so the overall rate is the worst per-hop subgroup rate. The paper then derives closed-form solutions for time allocation with fixed power and with adaptive power, and notes that with one subgroup per layer the rate reduces to CoMAC while with one node per subgroup it reduces to orthogonal time-sharing.","pith_inferences":["The paper does not spell out that its min-of-rates formula turns function computation into a tree resource-allocation problem, so standard scheduling and network-utility methods could be adapted to choose group sizes and routes that maximize the bottleneck rate.","If a subgroup function has a different entropy from the desired function, the rate normalization by $H(f(\\mathbf{b}_v))$ in Definition 2 would need to be layer-specific; testing this with non-symmetric functions would show where the formula needs correction.","The reorganization requires each node to have exactly one destination; extending the result to broadcasting or multi-destination nodes would likely require a min-cut-like bound over several per-node rates rather than a single minimum.","The simulations suggest that adding groups slows the rate loss as the number of sources grows; a natural next step is to prove a scaling law as both the number of sources and the number of groups grow."],"forward_implications":["In a two-layer network, setting the number of subgroups per group to one recovers the classical CoMAC rate, and setting it to the number of source nodes recovers the orthogonal time-sharing rate.","Because the overall rate is the minimum across subgroups and layers, the optimal time allocation equalizes the products $\\alpha_{l,k}\\beta^{(c)}_{l,k}$ for all subgroups, with the closed-form optimum given by the reciprocal-sum expression in Eq. (29).","Adaptive power control improves the achievable rate over fixed power control, and its optimal allocation is expressed through the Lambert W function.","Adding layers lowers the computation rate because each layer consumes channel uses, while increasing the number of groups can support a larger number of source nodes with a slower rate decline."],"supporting_citations":[{"why":"Supplies the single-hop CoMAC rate formula over fading MACs (Theorem 3) and the adaptive power-control rate (Theorem 5) that this paper generalizes.","marker":"[21]"},{"why":"Provides the nested-lattice-code framework that makes the per-subgroup CoMAC rates achievable in noise.","marker":"[20]"},{"why":"Defines the nomographic-function class that CoMAC computes, which motivates the subgroup and group function reconstruction.","marker":"[15]"},{"why":"Establishes the computation-over-multiple-access-channels problem that this paper extends to multi-hop hierarchical networks.","marker":"[18]"},{"why":"Motivates the substitution p = alpha*beta used to reformulate the bilinear time-allocation problem as a linear program.","marker":"[29]"},{"why":"Supplies the Lagrangian-duality and KKT conditions used to derive the closed-form optimal allocations.","marker":"[30]"}],"fun_headline_variants":["Disorganized networks compute at the worst per-hop rate","Hierarchy plus CoMAC computes functions in chaotic nets","Worst-hop rate governs function computing in disorganized nets","CoMAC+orthogonal scores a closed-form rate for messy nets","Reorganize chaotic links into layers to compute at one rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formula assumes that every disorganized network can be reorganized into a layered hierarchy in which each node has exactly one transmission destination and every intermediate subgroup and group function carries the same entropy as the desired function; if either condition fails, the overall rate is not simply the minimum over per-hop rates.","fun_headline_variants_meta":{"raw":{"variants":["Disorganized networks compute at the worst per-hop rate","Hierarchy plus CoMAC computes functions in chaotic nets","Worst-hop rate governs function computing in disorganized nets","CoMAC+orthogonal scores a closed-form rate for messy nets","Reorganize chaotic links into layers to compute at one rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3258,"prompt_tokens":1059,"completion_tokens":2199,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":2117}},"tokens_in":675,"tokens_out":2199,"duration_ms":16893,"temperature":1.0,"reasoning_tokens":2117,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:26:01.151474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or build a three-layer network with four source nodes, two second-layer relays, and one fusion center under i.i.d. Rayleigh fading, and measure the maximum number of desired-function values per channel use with optimal time allocation; if the measured rate disagrees with Eq. (29), the paper's central claim is wrong.","supporting_citations":[{"cited_title":"Computation over Gaussian networks with orthogonal components,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-hop CoMAC rate formula over fading MACs (Theorem 3) and the adaptive power-control rate (Theorem 5) that this paper generalizes."},{"cited_title":"Compute-and-forward: Harnessing interference through structured codes,","cited_arxiv_id":null,"evidence_quote":"Provides the nested-lattice-code framework that makes the per-subgroup CoMAC rates achievable in noise."},{"cited_title":"Nomographic functions: Efficient computation in clustered Gaussian sensor networks,","cited_arxiv_id":null,"evidence_quote":"Defines the nomographic-function class that CoMAC computes, which motivates the subgroup and group function reconstruction."},{"cited_title":"Mccormick-based relaxations of algorithms,","cited_arxiv_id":null,"evidence_quote":"Motivates the substitution p = alpha*beta used to reformulate the bilinear time-allocation problem as a linear program."}],"review_version":1}