{"id":"ba09d025-4de7-488f-bb26-6f8f934850e4","arxiv_id":"2507.19177","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For a two-relay diamond network with channel state known only at the relays, the paper derives computable upper bounds and simple achievable schemes whose simulated rates nearly match those bounds.","lead":"This paper studies a two-relay wireless network where the relays know the fading channel but the central processor does not. It provides computable upper bounds and simple relay schemes whose simulated rates are close to those bounds.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The TCI achievable-rate formula (32)-(34) is derived by plugging the average post-selection SNR into a fixed-state Gaussian rate expression, without proving achievability when the CP never learns the instantaneous fading coefficient; the cited [33] covers source-coding mismatch, not this…","rationale":"The reader's weakest-assumption analysis focused on the explicit Gaussian-input restriction. That restriction is a legitimate scoping caveat, but it is clearly stated in Section I-B and does not threaten the internal validity of the paper's bounds under that assumption. The more load-bearing issue is the TCI scheme's missing achievability proof: Section II-C2 evaluates fixed-state Gaussian capacity formulas at an average post-selection SNR even though the CP does not know the instantaneous fading coefficient. The paper's own text flags the non-Gaussianity and cites [33], but the cited result concerns mismatched source coding, not a channel whose effective noise variance is random and unobserved by the decoder. If the TCI rate is not actually achievable, then the scheme that most closely approaches Rub in the low-fronthaul, high-SNR regime (Figs. 5a and 6a) is not a valid lower bound, and the central 'close to the upper bound' claim is unsupported in that regime. A single numerical comparison of the true single-letter rate for the TCI test channel against the asserted formula would settle whether this gap is substantive. The rest of the paper, including QCI, MMSE, the DPP approximation, and the two-user FC scheme, appears internally consistent, and the reader's Gaussian-input caveat remains worth noting, so I would not reject the paper; acceptance should be conditional on providing the missing achievability proof or qualifying the TCI claim.","tokens_in":25376,"tokens_out":57365,"duration_ms":638040,"concrete_test":"For the single-relay marginal (state (1,0)), fix a threshold Sth and per-selected-symbol rate ec = (C - eH)/P. Let the relay use the TCI test channel Z = eX + Q with Q ~ CN(0, d), where eX = X + S^*/|S|^2 N and the selected distribution is |S| >= Sth. Numerically compute the actual single-letter achievable rate R_actual(d) = min( I(X; Z), ec - I(Y, S; Z | X) ) under the Rayleigh-selected joint distribution, then optimize over d >= 0 and compare with the paper's eR_{1,0} = max_r min( log(1 + eρ(1 - 2^{-r})), ec - r ) with eρ = 1 / E[|S|^{-2} sigma^2 | |S| >= Sth]. Repeat for representative parameters from Fig. 5, e.g., Sth = 0.5 and C = 5 bits. If R_actual(d) < eR_{1,0} for any such parameter point, the TCI curve overstates the achievable rate and the low-capacity closeness claim is weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that the proposed achievable schemes perform close to the upper bounds over a broad parameter range. In the single-user case, the truncated channel inversion (TCI) scheme is the one that comes closest to Rub when the fronthaul is scarce (Figs. 5a and 6a). The derivation of its rate in Section II-C2 has a gap. After selecting symbols with |S_k| >= Sth, the relay forms eX_k = X + (S_k^*/|S_k|^2) N_k and compresses it. The CP knows only the binary selection event, not the realized S_k. The paper then defines eρ_k = 1 / E[|S_k|^{-2} sigma^2 | |S_k| >= Sth] and evaluates eR_{1,1}, eR_{1,0}, eR_{0,1} exactly as if the channel from X to the relay output were an additive white Gaussian noise channel with SNR eρ_k over every selected symbol. But the actual per-symbol noise variance is random and unknown to the CP; the induced channel is not the Gaussian channel used in [21, Thm. 5]. The text acknowledges that the observation is non-Gaussian and appeals to [33] for the use of Gaussian codebooks, but [33] is about rate-distortion with mismatched codebooks and does not by itself prove that the decoder can achieve the fixed-state Gaussian rate without knowing the fading coefficient. No single-letter achievability proof is given showing that the min-T expression with the actual joint distribution of (X, Y_k, S_k, Z_k) is bounded below by (33). If this step fails, the TCI rates in Figs. 5-6 are not valid lower bounds, and the paper's main numerical demonstration of tightness in the low-capacity, high-SNR regime is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a two-relay Gaussian 'diamond' network (single-user and two-user versions) with Rayleigh fading, where each relay knows its own channel state but the central processor (CP) has no channel state information, and the relays are oblivious to the users' codebooks. For the single-user case, the authors derive an analytical cooperative informed-receiver upper bound R^ub, a lower bound to the informed-receiver bound, a DPP-based approximation, and three achievable schemes: quantized channel inversion (QCI), truncated channel inversion (TCI), and an MMSE-based scheme. For the two-user case, they propose a cooperative upper bound and an achievable scheme based on joint fronthaul compression of the received signal and the channel state (FC). The paper's central claim is that the achievable rates perform close to the respective upper bounds over a broad range of parameters, supported by numerical results.","tokens_in":25656,"tokens_out":18300,"duration_ms":201244,"significance":"If the derivations are correct, the paper would provide computationally tractable bounds and practical low-complexity relay strategies for a relevant C-RAN model with partial CSI at the CP. The QCI and MMSE schemes appear to be rigorously derived, as is the analytical upper-bound framework based on [27], and the DPP approximation follows standard stochastic optimization theory. The paper is also honest about the difficulty of exact ergodic capacity and about the Gaussian-input restriction. However, the TCI and FC schemes contain load-bearing gaps in their achievability arguments; without those, the numerical demonstration that the schemes are close to the upper bounds is not fully established. The paper's value is therefore conditional on fixing these proofs.","major_comments":[{"comment":"The TCI achievable rate is derived by substituting the average post-selection SNR eρ_k into the fixed-state Gaussian rate expression, but the CP never learns the instantaneous fading coefficient S_k. After selection, the normalized observation is eX_k = X + (S_k^*/|S_k|^2) N_k, and the noise variance |S_k|^{-2}σ^2 remains random and unknown to the CP. The resulting channel is a Gaussian mixture, not the AWGN channel used in [21, Thm. 5], and its capacity is not generally equal to the expression with eρ_k. The appeal to [33] does not close this gap: [33] concerns rate-distortion with mismatched Gaussian codebooks, not joint decoding of an information message at a CP that lacks the fading realizations. No single-letter achievability proof is provided for Eq. (33). Consequently, the TCI rates plotted in Figs. 5 and 6 are not established valid lower bounds, and the claim that TCI approaches R^ub in the limited-fronthaul regime is unsupported.","section":"Section II-C2, Eqs. (32)-(34)"},{"comment":"The FC scheme defines an auxiliary variable Z^{(2)}_{k,g} = Z^{(1)H}_k X + bN_{k,g} + bQ_k in Eq. (74) that depends explicitly on the users' signal X. Since the relays are oblivious to the users' codebooks, they cannot generate a quantization codebook according to this distribution. Moreover, the actual observation is Y_k = Z^{(1)H}_k X + E^H_k X + N_k, where E_k is a non-Gaussian, signal-dependent term. Replacing E^H_k X + N_k with an independent Gaussian noise of the same second moment changes the channel model; no argument is given that the communication rate for the actual channel is at least R_fc. Thus the FC rates in Figs. 8-10 are not established achievable rates. A valid derivation would need to specify a test channel P_{Z^{(2)}_k | Y_k, Z^{(1)}_k} and then evaluate the mutual information terms in Eq. (73) for the true non-Gaussian source.","section":"Section III-B, Eqs. (74)-(75)"},{"comment":"Both the TCI and FC schemes invoke [33] to justify the use of Gaussian codebooks for non-Gaussian sources. That reference addresses the rate-distortion performance of mismatched Gaussian codebooks for source coding, not the end-to-end reliability of a relay channel where the CP lacks the state information needed to form the correct likelihoods. The paper should either provide a proper joint-typicality achievability proof for the proposed test channels or clearly state that the rates are heuristic and not established lower bounds.","section":"Section II-C2 and Section III-B"}],"minor_comments":[{"comment":"The terms eP1(1−bP2)eR1,0 and (1−eP1)bP2eR0,1 in Eq. (32a), and the analogous terms in Eq. (34a), use the QCI probability bP2 where the TCI probability eP2 is clearly intended; this typo should be corrected.","section":"Eqs. (32a) and (34a)"},{"comment":"The rate allocation eck = (Ck − eHk)/ePk can become negative when Ck < eHk, making the scheme infeasible. The optimization should explicitly restrict to thresholds Sth for which eck ≥ 0 for all k.","section":"Section II-C2, Eq. (32b)"},{"comment":"The constant B is defined in Theorem 1 as (1/K) ∑_k max{C_k^2, (Cmax−C_k)^2}, while the derivation in Eq. (85) uses B = (1/2) ∑_k max{...}; for K=2 these coincide, but the general statement should be made consistent by defining B as the bound on the second moment term in Eq. (83).","section":"Section II-B2 and Appendix B"},{"comment":"The paper should make more prominent that all rates are for Gaussian channel inputs only; the upper bounds and achievable schemes do not bound the unrestricted capacity over arbitrary input distributions. The current wording is clear in Section I-B but could be repeated in the numerical section to avoid overstatement in the conclusion.","section":"Section I-B and Eq. (1)"},{"comment":"The achievable rate for D = 0.001 at low C is reported as zero because the channel-state encoding consumes all available capacity; this behavior is correct but deserves a brief explanation in the text, as it is visible only in the plots and not discussed.","section":"Section IV-B2, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The reader's report is more optimistic than I am. The TCI and FC achievability gaps are not merely presentation issues: they affect the validity of the central numerical claim for two of the five proposed schemes. That said, the gaps are in principle fixable by specifying valid test channels and evaluating the true mutual informations, so the paper should not be rejected outright if the authors can supply the missing proofs. I would also encourage the editor to ask for a careful revision of the use of [33], which appears to be cited beyond its scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — here's my take on arXiv:2507.19177. The genuinely new thing is the random-state model for the two-relay diamond with CSI at the relays but not the CP, and the authors give computable informed-receiver upper bounds (including a nice DPP approximation), plus several relay processing schemes. The convex reformulations and the MC evaluations are careful, and the single-user QCI and MMSE schemes are legitimate lower bounds as far as I can tell. The two-user cooperative upper bound using Wishart eigenvalues is also clean.\n\nThe soft spot is the achievability proofs of the TCI scheme (Section II-C2) and the FC scheme (Section III-B). In both cases the relay compresses a non-Gaussian observation and the CP never learns the instantaneous noise variance. The paper plugs an average SNR into the fixed-state Gaussian rate expression and then cites [33] for using Gaussian codebooks. That citation supports compression of a non-Gaussian source with Gaussian codebooks, not the claim that the end-to-end mutual information equals the fixed-state expression when the CP does not know the fading realization. The TCI rate formula (32)-(34) is therefore not proven as a lower bound; same issue appears in the FC rate in (75)-(76) when E_k is replaced by Gaussian noise of variance 2D. The stress-test note is right. This matters because the paper's headline 'schemes perform close to the upper bounds' leans on TCI in the low-capacity/high-SNR single-user plots and on FC in the two-user limited-capacity regime. If those rates are not valid lower bounds, that specific tightness claim is unsupported. There are also minor typos in (32a) and (34a) where bP2 should be eP2.\n\nI don't think the central framework is broken. The upper bounds, the DPP evaluation, and the QCI/MMSE schemes stand. The TCI/FC gap may be fixable by sending quantized noise variance (as QCI does) or by providing a proper mismatched-decoding analysis. The Gaussian-input restriction is explicit and fine as a modeling choice. The paper is worth a serious referee, but it needs a major revision before it can be trusted as stated.","headline":"Solid upper bounds and a useful framework, but the TCI and FC achievability proofs have a gap that undermines the tightness claim in the low-fronthaul regime.","tokens_in":26359,"tokens_out":6223,"would_cite":true,"duration_ms":67591,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A24"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives nearly tight achievable rates for a two-relay diamond network where the central processor has no channel state information, under Gaussian inputs.","keywords":["disaggregated radio access network","oblivious relay","Rayleigh fading","ergodic capacity","fronthaul compression","information bottleneck","drift-plus-penalty","cloud RAN"],"falsifier":"For the single-user model, compute the TCI achievable rate at SNR = 30 dB and fronthaul C = 2 bits per dimension and compare it with the cooperative upper bound $R^{\\mathrm{ub}}$ of (10): the paper's figures show near-coincidence at C = 5, so a large gap at C = 2 would delimit the 'broad range' claim. For the two-user model, the decisive test is whether any local joint-compression scheme at C = 5 bits and SNR = 30 dB can approach $R^{\\mathrm{ub}}$; the paper's Fig. 10 shows a visible gap there, so that regime is the paper's own boundary case.","tokens_in":25083,"feed_emoji":"📡","tokens_out":7454,"duration_ms":70472,"temperature":0.7,"pith_summary":"The paper tries to pin down how much data can flow through a 'diamond' relay network—one or two users, two relays, and a central processor connected by rate-limited fronthaul—when the relays know their own fading channels but the central processor does not. Because the exact ergodic capacity is intractable, the authors build an analytical upper bound by granting the central processor access to the channel states and letting the relays cooperate, then design three local-processing schemes for one user and a joint signal-and-channel compression scheme for two users. The message is that simple, locally computed relay strategies get close to this upper bound over a wide range of signal-to-noise ratios and fronthaul capacities, so a blind hub need not cost much rate. All rates are computed under Gaussian inputs, so the statement is about the Gaussian-input rate, not the unrestricted capacity.","feed_headline":"Blind central processor still gets near-optimal relay rates","feed_subtitle":"Upper bounds and three local relay schemes nearly agree across SNR and fronthaul range.","key_machinery":"The load-bearing device is the information-theoretic formulation of oblivious relay processing from [21] and [9], where each relay is represented by an auxiliary random variable $Z_k$ conditionally independent of everything else given its own observation $(Y_k,S_k)$, the communication rate is $I(X;Z_1,Z_2)$, and the fronthaul constraints appear as bottleneck inequalities. On top of this, the paper builds a computable upper bound by relaxing two restrictions at once: the central processor is given the full channel state ('informed receiver') and the relays are allowed to cooperate, which turns the network into a single two-antenna oblivious relay whose closed-form rate (10)/(63) follows from the eigenvalue distribution of the channel Gram matrix. The achievable schemes are then concrete choices of the auxiliaries: channel inversion with quantized noise levels, truncated inversion with a selection sequence, MMSE estimation followed by Gaussian compression, and—for two users—joint Gaussian compression of $(H_k,Y_k)$ with a distortion trade-off parameter $D$.","core_discovery":"On its own terms, the paper establishes that for the single-user two-relay Gaussian diamond with i.i.d. Rayleigh fading and channel state known only at the relays, the cooperative informed-receiver upper bound $R^{\\mathrm{ub}}$ in (10) is closely approached by three achievable schemes—quantized channel inversion ($R^{\\mathrm{qci}}$), truncated channel inversion ($R^{\\mathrm{tci}}$), and MMSE-based compression ($R^{\\mathrm{mmse}}$)—with the TCI scheme best at low fronthaul capacity and QCI best when capacity is ample. For the two-user case, the paper proves that the single-user local-processing ideas cannot be carried over, and instead shows that a fronthaul-compression scheme in which each relay lossily quantizes both its channel state and its received signal, then transmits them together, approaches the cooperative upper bound $R^{\\mathrm{ub}}$ in (63) whenever the fronthaul is not the bottleneck. The paper explicitly does not claim an exact capacity characterization, calling it 'surprisingly difficult.'","pith_inferences":["The same bounding strategy—relax to an informed cooperative receiver, then build local compression schemes—should extend to more than two relays; the single-antenna-per-relay channel inversion obstruction in the multi-user case suggests that for general $K$, joint $(H,Y)$ compression at each relay is the natural candidate.","The Gaussian-input restriction is the real ceiling on the reported numbers; a non-Gaussian ensemble designed to exploit the fading structure could in principle push the achievable rate above the Gaussian-input upper bound, so the 'near-tight' claim is conditioned on the input ensemble.","Because the channel states are i.i.d. per symbol, the setting is the $T=1$ extreme of a block-fading model; the paper's own discussion of block length $T$ suggests the gap between achievable rate and upper bound is likely to shrink further as $T$ grows, since the cost of conveying channel state to the CP falls.","The natural next test is to close the two-user low-fronthaul gap visible in the paper's Figure 10; if a better scheme succeeds there, the 'fronthaul is not the bottleneck' caveat can be dropped."],"forward_implications":["If the bounds are as tight as the simulations indicate, a system designer can predict the ergodic rate of a disaggregated RAN uplink without tracking channel state at the central processor, using only local relay processing and the statistics of the fading.","In the single-user case, the truncated-channel-inversion scheme is essentially optimal when the fronthaul is small: both it and the upper bound converge to $C_1+C_2$ as SNR grows (Lemma 1).","When the fronthaul is large, the quantized-channel-inversion scheme closes the gap to the upper bound, meaning the bottleneck moves from the relay processing to the fading channel itself, at rate $\\mathbb{E}[\\log(1+\\lambda/\\sigma^2)]$ in the limit.","For two users, the joint compression of channel state and signal is the first scheme that approaches the upper bound at all, and its distortion parameter $D$ trades channel-state accuracy against signal fidelity; no local filtering alone can work because a $1\\times 2$ channel vector is not invertible.","The drift-plus-penalty stochastic optimization provides an $O(1/V)$-accurate way to approximate the harder informed-receiver upper bound, so the tightest bound in the paper can actually be computed (Theorem 1)."],"supporting_citations":[{"why":"Supplies the oblivious-relay capacity expression (2) and the fixed-state Gaussian rate formula (5) on which all single-user bounds and schemes build.","marker":"[21]"},{"why":"Extends the oblivious-relay capacity region to K users; its Theorem 1 and Theorem 5 give the two-user sum-capacity expression (54) and the fixed-state Gaussian solution (57).","marker":"[9]"},{"why":"Provides the closed-form information-bottleneck rate for a vector oblivious relay with channel state, whose eigenvalue distribution is used to evaluate the cooperative informed-receiver upper bounds (10) and (63).","marker":"[27]"},{"why":"Gives the drift-plus-penalty stochastic optimization machinery used in Algorithm 1 and Theorem 1 to approximate the informed-receiver upper bound.","marker":"[30]"},{"why":"Supplies the drift-plus-penalty stochastic optimization convergence theory and the $O(1/V)$ bound used in the proof of Theorem 1.","marker":"[31]"},{"why":"Provides the rate-distortion function for Gaussian sources used to set the channel-state quantization rate $R_k(D)$ in the fronthaul-compression scheme and the entropy of quantized noise levels in QCI.","marker":"[32]"},{"why":"Justifies the use of Gaussian codebooks for non-Gaussian sources, which underlies the TCI and FC achievable-rate evaluations.","marker":"[33]"}],"fun_headline_variants":["No CSI at CP: relay schemes nearly achieve rate bound","Blind central processor: local CSI at relays recovers rates","Fronthaul compression with channel state at relays wins","Two-user relay diamond: compress channel+signal for near-optimal","Relay-side knowledge almost matches full-CSI rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that users are restricted to Gaussian random codebooks: if non-Gaussian input distributions were allowed, the true ergodic capacity could exceed every rate reported in the paper, so the near-tight bounds are bounds on the Gaussian-input rate, not on unrestricted capacity.","fun_headline_variants_meta":{"raw":{"variants":["No CSI at CP: relay schemes nearly achieve rate bound","Blind central processor: local CSI at relays recovers rates","Fronthaul compression with channel state at relays wins","Two-user relay diamond: compress channel+signal for near-optimal","Relay-side knowledge almost matches full-CSI rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1612,"prompt_tokens":1075,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":691,"completion_tokens_details":{"reasoning_tokens":454}},"tokens_in":691,"tokens_out":537,"duration_ms":5607,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:00:47.326572+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the single-user model, compute the TCI achievable rate at SNR = 30 dB and fronthaul C = 2 bits per dimension and compare it with the cooperative upper bound $R^{\\mathrm{ub}}$ of (10): the paper's figures show near-coincidence at C = 5, so a large gap at C = 2 would delimit the 'broad range' claim. For the two-user model, the decisive test is whether any local joint-compression scheme at C = 5 bits and SNR = 30 dB can approach $R^{\\mathrm{ub}}$; the paper's Fig. 10 shows a visible gap there, so that regime is the paper's own boundary case.","supporting_citations":[{"cited_title":"Communication via decentralized processing,","cited_arxiv_id":null,"evidence_quote":"Supplies the oblivious-relay capacity expression (2) and the fixed-state Gaussian rate formula (5) on which all single-user bounds and schemes build."},{"cited_title":"On the capacity of cloud radio access networks with oblivious relaying,","cited_arxiv_id":null,"evidence_quote":"Extends the oblivious-relay capacity region to K users; its Theorem 1 and Theorem 5 give the two-user sum-capacity expression (54) and the fixed-state Gaussian solution (57)."},{"cited_title":"Information bottleneck for an oblivious relay with channel state information: the vector case,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-form information-bottleneck rate for a vector oblivious relay with channel state, whose eigenvalue distribution is used to evaluate the cooperative informed-receiver upper bounds (10) and (63)."},{"cited_title":"Stochastic network optimization with application to communication and queueing systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the drift-plus-penalty stochastic optimization convergence theory and the $O(1/V)$ bound used in the proof of Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the rate-distortion function for Gaussian sources used to set the channel-state quantization rate $R_k(D)$ in the fronthaul-compression scheme and the entropy of quantized noise levels in QCI."},{"cited_title":"On the role of mismatch in rate distortion theory,","cited_arxiv_id":null,"evidence_quote":"Justifies the use of Gaussian codebooks for non-Gaussian sources, which underlies the TCI and FC achievable-rate evaluations."}],"review_version":1}