{"id":"1d118e76-f55d-43be-9216-f46653943c32","arxiv_id":"1908.00703","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The two-user interference channel with limited cooperation under finite precision CSIT has sum-GDoF exactly equal to a min of four linear bounds, so each bit of cooperation buys 0, 1, 1/2, or 1/3 over-the-air bits.","lead":"This paper gives the exact generalized degrees of freedom (GDoF) of a two-user interference channel when transmitters share a limited number of bits under finite precision channel knowledge. Each bit of cooperation is shown to buy 0, 1, 1/2, or 1/3 bits over-the-air, with the 1/3 slope a new effect in strong interference.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Converse for the 1/3-slope bound rests on an unstated sum-set inequality from [16] applied to dependent partitions A, B, C; if that theorem requires independence, equation (93) does not follow.","rationale":"The paper's most significant and novel result is the exact strong-interference GDoF characterization with a 1/3 cooperation slope, which arises only under finite-precision CSIT. The proof of the matching upper bound rests on the sum-set inequality from the authors' unpublished preprint [16], invoked at (64), (78), and (82). The reader's weakest_assumption identifies exactly this dependency. My review of the surrounding argument found the internal algebra of sections 4 and 6 consistent: the entropy bounds (89)-(92) match the partition widths, the combination of (81) and (87) yields (88) correctly, and the achievable schemes in section 6 are explicit and parameterized. The decisive uncertainty is external: whether [16]'s inequality applies to the dependent random variables A, B, C. If it does, the theorem likely stands; if not, the 1/3 slope bound collapses. This is a legitimate condition for a CONDITIONAL verdict, not a demonstrated error, so no change in verdict is appropriate. I agree with the reader rather than proposing a new objection, and the concrete test above would settle the issue.","tokens_in":31204,"tokens_out":14943,"duration_ms":134381,"concrete_test":"Obtain Theorem 1 of [16], write out its hypotheses, and check them against A, B, C from (43)-(45) for all strong-interference subcases with alpha12 >= alpha21. In particular, determine whether the theorem tolerates dependence between the two summands: A and C are both functions of W02 when conditioned as in (64), and both are functions of W01 when conditioned as in (78). If the theorem requires independence, test the inequality H(Y1^n|W11,W01,G) >= H(A,C|W11,W01,G) on a concrete instance (e.g., alpha11=2, alpha22=1, alpha12=4, alpha21=3, pi chosen so the 1/3 bound is active) by computing or bounding the relevant entropies for the finite-precision deterministic model; if the inequality fails, equation (93) and hence Theorem 1 in that regime are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central new claim is the 1/3 slope in the strong-interference regime. Its converse is equation (93), obtained by combining (81), (87), and (88), and the derivation of those steps invokes the sum-set inequality of [16] three times: at (64), (78), and (82). The paper never restates the hypotheses of that inequality or proves them for the partitions A, B, C defined in (43)-(45). This matters because A and C are not independent: A is a power-level segment of X1 = f1(W11,W01,W02) and C is a segment of X2 = f2(W22,W01,W02), so conditioned on W11,W01 (as in (64)), both A and C depend on the common message W02; similarly at (78) they share dependence on W01. A generic sum-set entropy inequality H(sum) >= H(A,C) is false for dependent summands, so the applicability of [16] is the load-bearing step in the 1/3-slope upper bound. The note before (46) that the combined power levels of A and C are below alpha12 is necessary but not sufficient. Since [16] is an arXiv preprint by the same group, the self-containedness and correctness of this step cannot be checked from the present text. This is a rigor/dependency concern, not a demonstrated contradiction; the rest of the converse algebra appears internally consistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper characterizes the sum generalized degrees of freedom (GDoF) of the two-user interference channel with limited transmitter cooperation under finite-precision CSIT, for both half-duplex and full-duplex cooperation. Theorem 1 gives the half-duplex sum-GDoF as min(DΣ,IC+π, D2e+π/2, D3e+π/3, DΣ,BC) in the strong interference regime, with D2e and D3e defined in (13)-(14), and the simpler min(DΣ,IC+π, DΣ,BC) in the weak and mixed regimes. Theorem 2 gives the analogous full-duplex characterization, with a mixed-regime expression min(DΣ,IC+π/2, DΣ,BC) and a strong-regime expression involving min(α12,α21)+π/2 and D3e+π/3. The converse is developed through aligned-images bounds and sum-set inequalities, while the achievability is given by explicit Gaussian coding schemes organized into numerous cases. Two corollaries state the minimum cooperation GDoF needed to reach the broadcast-channel bound.","tokens_in":31493,"tokens_out":9736,"duration_ms":86141,"significance":"If the main theorems are correct, the paper closes the GDoF gap between the no-cooperation interference channel and the full-cooperation broadcast channel under finite-precision CSIT, and it identifies a genuinely new 1/3 slope that does not appear under perfect CSIT. The GDoF formulas are explicit and the theorem statements are clear. The paper also provides a detailed case-by-case achievability construction, which is a substantial amount of work. The main caveat is that the 1/3-slope converse relies on a sum-set inequality from the authors' own arXiv preprint [16] without restating its hypotheses; this is a correctness-risk point that needs to be resolved before the characterization can be considered fully established.","major_comments":[{"comment":"The converse for the D3e bound, leading to (93), depends on the sum-set inequality (Theorem 1 of [16]) being applicable to the partitions A, B, C defined in (43)-(45). The paper invokes this inequality at (64), (78), and (82) but never restates its hypotheses or proves them for the present setting. This is load-bearing because A is a segment of X1 = f1(W11,W01,W02) and C is a segment of X2 = f2(W22,W01,W02). Conditioned on W11,W01 as in (64), both A and C depend on the common message W02; similarly, at (78) they share dependence on W01. A generic sum-set entropy inequality H(Y) ≥ H(A,C) can fail for dependent summands, and the observation after (46) that the combined power levels of A and C are below α12 is necessary but not sufficient. Since [16] is an unreviewed arXiv preprint by the same group, the correctness of the 1/3-slope upper bound cannot be checked from the present text. Please provide a self-contained statement of the inequality and a detailed verification of its hypotheses for the dependent partitions A and C, or supply a proof of the specific instances used at (64), (78), and (82).","section":"§4, Eqs. (64), (78), (82), and (93)"},{"comment":"The achievability case analysis contains boundary inequalities that are internally inconsistent as printed, which prevents verification of the claimed active-bound transitions. In §6.1 Case 2, after the first bullet states \"π ≤ α12−α21\", the second bullet writes \"When α21−α12 ≤ π ≤ 2α11+2α22−α12−α21\"; the lower bound appears to be a sign error for \"α12−α21\". In §6.2 Case 2, the bullet \"When α11+α22−2α12+α21 ≤ π/2 ≤ π+/2\" appears to involve a different threshold than the condition \"α11+α22+α12−2α21 ≤ π/2\" used in the preceding bullet. Because the achievability proof proceeds by partitioning the parameter range into intervals defined by these inequalities, such inconsistencies block a complete check of the strong-interference achievability claim. Please correct these boundaries and re-verify the nonnegativity of all rate assignments over the stated intervals.","section":"§6.1 Case 2 and §6.2 Case 2"}],"minor_comments":[{"comment":"The axis labels in Figure 2 are garbled (e.g., \"0:2\", \"ı`!\", \"Sum-GDoF`!\"), and the figure does not render the intended mathematical notation; it should be regenerated with proper labels.","section":"Figure 2"},{"comment":"Equations (64) and (65) appear to be identical as printed; if one of them was intended to display a different intermediate bound, please correct the duplication.","section":"§4, Eqs. (64)-(65)"},{"comment":"The paragraph on the X channel is informal and states several conclusions without proof; since it is presented as an aside rather than a theorem, it should be clearly labeled as an observation or conjecture, or supported with a proof.","section":"§7"},{"comment":"Because the converse relies on the sum-set inequality of [16], the paper should either state the theorem in the main text or include its statement in an appendix, especially given that [16] is an arXiv preprint rather than a published article.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The novelty and scope of the result are strong, and I do not see a reason to doubt the overall approach. However, the central 1/3-slope converse depends on a theorem from an unpublished same-group preprint whose hypotheses are not verified in the manuscript. I would ask the editor to require a self-contained verification of that inequality for the dependent partitions A and C, and to have the achievability case boundaries corrected, before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper closes the two-user limited-cooperation gap between the known interference-channel and broadcast-channel extremes. The genuinely new item is the 1/3 slope, which appears only in strong interference and only under finite-precision CSIT; it is absent from Wang-Tse's perfect-CSIT result and from the two extremes. That alone makes the paper worth reading carefully.\n\nWhat the paper does well: the statements are clean and complete. Theorems 1 and 2 give the full piecewise-linear sum-GDoF for half-duplex and full-duplex cooperation, with the four possible slopes 0, 1, 1/2, 1/3 explicitly identified. The achievability sections are detailed, with tables and cases that lay out codeword powers and decoding orders. The converse is mostly standard aligned-images machinery, and the algebra after the key steps appears internally consistent. The corollaries on the minimum cooperation needed to reach the broadcast bound are also useful and non-obvious.\n\nThe soft spot is exactly where the stress-test note points. The 1/3-slope bound (93) is the central new converse, and it depends on the sum-set inequality of [16] at (64), (78), and (82). The paper never restates the hypotheses of that inequality, and the partitions A, B, C are not obviously independent in the way a generic sum-set bound might require. In particular, A and C can both depend on the shared cooperative messages, so independence assumptions, if present in [16], would not automatically hold here. The note that the combined power levels of A and C stay below alpha_12 is necessary but not sufficient. This does not mean the bound is false, but it means the paper's main novelty is not checkable from the text alone. The referee will need [16] open on the desk, and the gap is a real one.\n\nSmaller soft spots: several achievability subcases, especially in the full-duplex setting, are dismissed with 'similar' or 'can be checked.' That is probably fine, but those checks should be done, since the full-duplex case involves partially wasted cooperation directions and the rate assignments are not all obvious. The citation practice is fair — the prior same-group results are legitimate and the paper builds on them rather than restating them.\n\nBottom line: this deserves serious refereeing. I would not accept it as-is; I would ask the authors to state the [16] theorem, verify its hypotheses for A, B, C, and expand the 'similar' full-duplex subcases. If the sum-set tool checks out, the result is a solid, citable characterization. If it does not, the 1/3 slope is unsupported. Either way, the paper is worth engaging with, not desk-rejecting.","headline":"The paper gives the first complete GDoF characterization for the two-user interference channel with limited transmitter cooperation under finite-precision CSIT, and the new 1/3 slope is real, but the converse for that slope leans on an unstated sum-set inequality from the same group's arXiv preprint, so the result is strong but not yet self-contained.","tokens_in":32038,"tokens_out":1423,"would_cite":true,"duration_ms":16225,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A40","94A24"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under finite-precision CSIT, the two-user interference channel with limited cooperation admits only four cooperation efficiencies, 0, 1, 1/2, and 1/3, and the paper proves the sum-GDoF is exactly the minimum of the four corresponding…","keywords":["generalized degrees of freedom","finite precision CSIT","limited transmitter cooperation","interference channel","strong interference regime","aligned images","sum-set inequality","half-duplex and full-duplex cooperation"],"falsifier":"At a strong-interference point where the 1/3 bound is active, for instance the paper's example $\\alpha_{11}=1.2,\\alpha_{22}=1,\\alpha_{12}=2,\\alpha_{21}=1.8$, compute the entropy difference in (81) directly on the deterministic model instead of bounding it via (89)-(92); if the difference exceeds the claimed $D_{3e}$ bound by more than $o(\\log P)$, the converse fails. A cheaper check is to verify each of the three invocations of the sum-set inequality of [16] against that inequality's stated hypotheses for these A, B, C partitions.","tokens_in":31022,"feed_emoji":"📡","tokens_out":10450,"duration_ms":97972,"temperature":0.7,"pith_summary":"This paper claims to determine exactly how many useful over-the-air bits limited transmitter cooperation can buy in a two-user interference channel when the transmitters know the channel only to finite precision. Working in generalized degrees of freedom (a high-SNR measure in which a link of strength $\\alpha$ carries about $\\alpha\\log P$ bits), it gives closed-form sum-GDoF formulas for every channel-parameter regime, under both half-duplex and full-duplex cooperation. The headline result is a finite menu of cooperation efficiencies: each bit of cooperation capability buys 0, 1, 1/2, or 1/3 additional over-the-air bits, with the 1/3 slope occurring only under finite-precision CSIT and only in the strong-interference regime. If the theorems are right, the gap between the no-cooperation benchmark and the full-cooperation broadcast-channel benchmark is completely explained, and the smallest cooperation budget needed to hit the broadcast-channel bound is known.","feed_headline":"Shared bits buy only 0, 1, 1/2, or 1/3 over-the-air bits","feed_subtitle":"Finite-precision CSI caps cooperation efficiency at 1/3 in strong interference; all regimes are now characterized.","key_machinery":"The carrying mechanism is the aligned-images method for converse bounds, specialized to a deterministic model with power-level partitions $A,B,C$ chosen from the transmitter symbol spaces (definitions (43)-(45)). These partitions are engineered so that $A$ is the part of $X_1$ that Receiver 2 cannot hear, $C$ is the part of $X_2$ that Receiver 2 can hear, and $B$ fills out the signal heard by Receiver 1; the sum-set inequality of [16] is then applied to relate their entropies. The resulting entropy chain yields the key GDoF inequality $3d_{11}+3d_{22}+2d_{01}+2d_{02}\\le D_{3e}$, which is exactly the $1/3$ slope. Achievability uses Gaussian superposition codebooks with carefully chosen power levels and successive interference cancellation at the receivers, with schemes running through four strong-interference subcases.","core_discovery":"On the paper's own terms, the central discovery is that the sum-GDoF of the two-user interference channel with limited transmitter cooperation and finite-precision CSIT is exactly $\\min(D_{\\Sigma,\\mathrm{IC}}+\\pi,\\,D_{\\Sigma,\\mathrm{BC}})$ in the weak and mixed half-duplex regimes, and in the strong-interference half-duplex regime is $\\min(D_{\\Sigma,\\mathrm{IC}}+\\pi,\\,D_{2e}+\\pi/2,\\,D_{3e}+\\pi/3,\\,D_{\\Sigma,\\mathrm{BC}})$, where $D_{2e}=\\alpha_{12}+\\alpha_{21}$ and $D_{3e}=\\min(\\alpha_{21}-\\alpha_{22},\\alpha_{11})+2\\max(\\alpha_{21}-\\alpha_{11},\\alpha_{22})+\\alpha_{12}+\\max(\\alpha_{12}-\\alpha_{22},\\alpha_{11})$; Theorem 2 gives the analogous full-duplex formula, with $\\pi/2$ replacing $\\pi$ in the mixed regime and $\\min(\\alpha_{12},\\alpha_{21})+\\pi/2$ in strong interference. This characterization is tight in every regime, so the authors claim no parameter combinations are left open. The $1/3$-slope bound, hidden in $D_{3e}+\\pi/3$, is the genuinely new phenomenon: it appears only under finite-precision CSIT and strong interference, and the paper identifies its mechanism, namely that an $\\epsilon$ increase in sum-GDoF requires an $\\epsilon$ increase in each of the three cooperative sub-messages and an $\\epsilon$ decrease in each of the two noncooperative messages.","pith_inferences":["The 1/3 slope points to a possible general law: under finite-precision CSIT, robust cooperation may require simultaneously upgrading all shared sub-messages, so one might expect slopes of the form $1/(K+1)$ in $K$-user extensions; the paper does not make this conjecture and its proofs are specific to two users.","The same aligned-images-plus-sum-set recipe, applied to the power partitions of this paper, could be aimed at the K-user broadcast channel's strong-interference regime, which the paper notes remains open; whether the recipe carries over is my inference, not its claim.","The closed-form cooperation-budget formulas can be inverted to optimize channel parameters for a fixed $\\pi$, an operation the paper does not perform; the resulting comparison against perfect-CSIT benchmarks would be a direct numerical test of the formulas.","The paper's closing remark that the 2-user X channel with limited cooperation becomes straightforward suggests that the difficulty is specific to partial message overlap; my inference is that adding more messages, not more users, may be the next hard step."],"forward_implications":["The GDoF of the two-user limited-cooperation interference channel under finite-precision CSIT is fully characterized for all parameter regimes, with no gap left between achievability and converse.","Every unit of cooperation capability buys exactly 0, 1, 1/2, or 1/3 over-the-air bits; under perfect CSIT only 0, 1, and 1/2 are possible, so 1/3 is the finite-precision signature.","Corollaries 1 and 2 give closed-form values for the minimum cooperation budget needed to reach the broadcast-channel bound, and in strong interference that budget strictly exceeds the simple shortfall $D_{\\Sigma,\\mathrm{BC}}-D_{\\Sigma,\\mathrm{IC}}$.","In the full-duplex model the mixed-interference bound uses $\\pi/2$ instead of $\\pi$, and in several strong-interference cases one cooperation direction is wasted, so half-duplex cooperation can beat full-duplex cooperation at the same total budget.","For symmetric channels the half-duplex and full-duplex formulas coincide, and for $2/3\\le\\alpha\\le1$ there is no cooperation gain, recovering and explaining the earlier broadcast-channel result."],"supporting_citations":[{"why":"introduces aligned image sets, the method on which all finite-precision converses in this paper rely.","marker":"[7]"},{"why":"supplies the broadcast-channel sum-GDoF that caps both theorems.","marker":"[10]"},{"why":"supplies the no-cooperation interference-channel sum-GDoF used as the baseline.","marker":"[11]"},{"why":"gives the perfect-CSIT limited-cooperation characterization whose slopes 0, 1, 1/2 are extended here.","marker":"[12]"},{"why":"shows two-user interference-channel GDoF are unchanged by finite-precision CSIT, justifying the baseline.","marker":"[15]"},{"why":"provides the sum-set inequalities invoked at (64), (78), and (82) to prove the 1/3-slope converse.","marker":"[16]"}],"fun_headline_variants":["Finite-precision CSIT reveals new 1/3 slope in strong interference","GDoF fully characterized for limited cooperation with finite-precision CSIT","1/3 slope emerges only under finite-precision CSIT with strong interference","Cooperation bits yield 0, 1, 1/2, or 1/3 over-the-air bits: finite-precision CSIT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 1/3-slope converse assumes that a known sum-set inequality applies to the particular power-level partitions A, B, and C constructed in (43)-(45); if that inequality does not cover these partitions, the 1/3 bound and with it the main theorems do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Finite-precision CSIT reveals new 1/3 slope in strong interference","GDoF fully characterized for limited cooperation with finite-precision CSIT","1/3 slope emerges only under finite-precision CSIT with strong interference","Cooperation bits yield 0, 1, 1/2, or 1/3 over-the-air bits: finite-precision CSIT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002192,"raw_usage":{"total_tokens":8553,"prompt_tokens":1073,"completion_tokens":7480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":7382}},"tokens_in":689,"tokens_out":7480,"duration_ms":45793,"temperature":1.0,"reasoning_tokens":7382,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:37:11.610238+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At a strong-interference point where the 1/3 bound is active, for instance the paper's example $\\alpha_{11}=1.2,\\alpha_{22}=1,\\alpha_{12}=2,\\alpha_{21}=1.8$, compute the entropy difference in (81) directly on the deterministic model instead of bounding it via (89)-(92); if the difference exceeds the claimed $D_{3e}$ bound by more than $o(\\log P)$, the converse fails. A cheaper check is to verify each of the three invocations of the sum-set inequality of [16] against that inequality's stated hypotheses for these A, B, C partitions.","supporting_citations":[{"cited_title":"Aligned image sets under channel uncertainty: Set- tling conjectures on the collapse of degrees of freedom under finite precision CSIT,","cited_arxiv_id":null,"evidence_quote":"introduces aligned image sets, the method on which all finite-precision converses in this paper rely."},{"cited_title":"Transmitter cooperation under finite precision CSIT: A GDoF perspective,","cited_arxiv_id":null,"evidence_quote":"supplies the broadcast-channel sum-GDoF that caps both theorems."},{"cited_title":"Interference mitigation through limited transmitter co- operation,","cited_arxiv_id":null,"evidence_quote":"gives the perfect-CSIT limited-cooperation characterization whose slopes 0, 1, 1/2 are extended here."},{"cited_title":"Generalized Degrees of Freedom of the Symmetric K-User Interference Channel under Finite Precision CSIT,","cited_arxiv_id":null,"evidence_quote":"shows two-user interference-channel GDoF are unchanged by finite-precision CSIT, justifying the baseline."},{"cited_title":"Sum-set Inequalities from Aligned Image Sets: Instruments for Robust GDoF Bounds","cited_arxiv_id":"1703.01168","evidence_quote":"provides the sum-set inequalities invoked at (64), (78), and (82) to prove the 1/3-slope converse."}],"review_version":1}