{"id":"e4c71692-d1f7-40fd-b33c-4622a82f7167","arxiv_id":"1908.09594","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PAC codes, formed by preceding the polar transform with a convolutional code and decoding with a Fano sequential decoder, show near-dispersion FER in a single BIAWGN simulation, with capacity guaranteed only because they contain polar codes.","lead":"This Shannon Lecture note recounts how sequential decoding led Arikan to polar codes, then introduces PAC codes, which add a convolutional precoder before the polar transform and use sequential decoding. A one-example simulation shows PAC codes near the finite-blocklength performance limit at blocklength 128, but no rigorous analysis is given.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-dispersion FER claim for PAC codes rests on a single unreproduced simulation and an informal 'looks random' heuristic; the capacity-achieving claim via the polar-code special case is sound.","rationale":"The reader's conditional verdict already identifies the same load-bearing weakness: the PAC performance claim is empirical, the supporting randomness heuristic is informal, and the decoder complexity question is explicitly open. My stress-test confirms that this is the right concern. The paper's rigorous contents — the polar-code results, the MLC/MSD polarization development, and the capacity-achieving nature of PAC codes via the polar-code special case — are internally sound and properly referenced. The only serious risk is that the headline PAC FER result cannot be independently verified or may not generalize beyond the single configuration shown. This risk does not justify rejecting the paper, because the manuscript honestly labels the PAC analysis as heuristic and open. It does justify keeping the conditional verdict: the central finite-blocklength performance claim should be backed by reproducible simulation artifacts or a precise statement of what is being claimed. No additional technical flaw was found, so the reader's verdict should remain unchanged.","tokens_in":14104,"tokens_out":3402,"duration_ms":40851,"concrete_test":"Independently reproduce Fig. 12 from the Section VII specification: N=128, R=1/2, RM score function for A, c=(1,0,1,1,0,1,1), BIAWGN channel, and a Fano decoder whose metric and search policy are either author-released or explicitly pinned down. Compare the resulting FER curve to the dispersion approximation at SNRs between 1.5 and 3 dB. If the PAC curve is more than 0.25 dB away from the dispersion approximation at FER 1e-3 to 1e-2, the central empirical claim is not supported; if it reproduces within that margin, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only original quantitative claim — that a PAC code with N=128, R=1/2 approaches the BIAWGN dispersion approximation (Fig. 12) — is supported by one computer simulation with no released code, no decoder pseudocode, no error bars, and no measured complexity statistics. Section VII specifies the code parameters (RM score function for A, c=(1,0,1,1,0,1,1)), but the Fano decoder is described only as using a time-varying bias; the exact metric, thresholds, and search policy cannot be reproduced from the text alone. The authors explicitly state in Section VIII that the performance and complexity of PAC codes 'are yet to be studied rigorously,' and the Section VII explanation that G=TPn 'looks sufficiently random' is informal. Since the finite-blocklength FER is the main new contribution, its evidential basis is the weakest load-bearing point. The capacity statement in Section VIII is not in doubt: setting T to the identity and A to a polar data index set recovers a polar code, so PAC codes inherit capacity achievability. Thus the concern is not internal inconsistency but unsupported empirical evidence for the central performance claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a written version of the 2019 Shannon Lecture. It traces the conceptual path from sequential decoding, through Massey's cut-off-rate example and Pinsker's scheme, to multi-level coding and then to polar coding. It then introduces polarization-adjusted convolutional (PAC) codes, which place an outer convolutional transform T before the inner polar transform P_n, and reports by computer simulation that an N=128, R=1/2 PAC code has frame error rate close to the BIAWGN dispersion approximation (Fig. 12). The paper further argues that PAC codes contain polar codes as a special case and hence can achieve channel capacity. Theorems 1 and 2 are standard polarization results with proofs cited to the literature; the new PAC-code performance claims rest on a single simulation and an informal 'looks sufficiently random' heuristic, and Section VIII explicitly acknowledges that performance and complexity of PAC codes are not yet rigorously studied.","tokens_in":14344,"tokens_out":4687,"duration_ms":50199,"significance":"If substantiated, the finite-length PAC-code claim would be practically notable: a rate-1/2 code of blocklength 128 with FER near the finite-blocklength dispersion bound would markedly improve on the polar and CA-SCL curves shown in Fig. 12, while using sequential decoding. The capacity-achieving statement via the polar-code special case (T equal to the identity and A a polar data index set) is correct and does not depend on any simulation. The paper also provides a useful historical synthesis of the ideas that led to polar coding, with correct citations to the polarization literature. However, the central quantitative claim is currently supported only by one unreproduced simulation, and the paper itself states that PAC-code performance and complexity remain open. The significance is therefore conditional on the reproducibility and further validation of that simulation.","major_comments":[{"comment":"The paper's principal quantitative claim—that a PAC code with N=128, R=1/2, RM design rule, and c=(1,0,1,1,0,1,1) has FER near the BIAWGN dispersion approximation—is supported by exactly one simulation. The decoder description in Section VII is incomplete: only 'time-varying bias' is mentioned, with no specification of the metric, threshold update policy, search limits, or stopping rule, and Fig. 12 provides no confidence intervals, repetition counts, or measured complexity statistics. The claim is therefore not reproducible from the manuscript. Please either give full decoder pseudocode and complete simulation data, or explicitly label the curve as preliminary simulation evidence rather than a demonstrated performance result.","section":"Section VII, Fig. 12"},{"comment":"The explanation for the near-dispersion behavior—that the combined transform G = TP_n 'looks sufficiently random'—is an informal heuristic. No random-like property is defined or measured; for example, the paper does not report the weight enumerator, distance profile, or comparison with random-code ML decoding. Since this heuristic is the only principle offered to connect the simulation to the dispersion bound, it needs to be either formulated as a testable conjecture with supporting analysis or replaced by a quantitative evaluation of the claimed 'sufficiently random' property.","section":"Section VII, paragraph beginning 'Evidently'"},{"comment":"The manuscript explicitly states that 'the performance and complexity of PAC codes are yet to be studied rigorously' and that understanding the computational complexity of the sequential decoder is an open problem. This is in tension with the motivational use in Section VII of the rate-profile criterion—that staying below the polarized cutoff rate profile indicates low Fano-decoder complexity—as a practical design guide. Without complexity statistics from the reported simulation or an analysis of Fano search effort, the complexity side of the PAC-code proposal is unverified. The paper should clearly separate this open heuristic from the rigorous polar-code results in Theorem 2.","section":"Section VIII, second paragraph"}],"minor_comments":[{"comment":"The phrase 'original idea s' contains a stray space; please proofread for similar typographical errors.","section":"Abstract"},{"comment":"'an codeword u' should be 'a codeword u'.","section":"Section VII, paragraph after Fig. 13"},{"comment":"The claim that the RM design rule 'suggests that, unlike polar codes, PAC codes are robust against channel parameter variations' is not supported by the single simulation at one SNR setting; please rephrase as a conjecture or add supporting experiments across channel parameters.","section":"Section VII, last paragraph"},{"comment":"The dispersion approximation is described as 'an estimate of the average ML-decoding performance' of a random code ensemble; in the cited reference [19] the normal approximation is a rate approximation for the maximal achievable rate, not an ensemble average. The wording should be adjusted for precision.","section":"Section VI, Fig. 12 caption"},{"comment":"The observation that the Fano decoder 'ran significantly faster' under the polar design rule is reported without any measured complexity data; consider adding mean or median search effort, or a histogram of decoder complexity.","section":"Section VII, paragraph on design rules"}],"recommendation":"major_revision","confidential_remarks":"This is a lecture-note-style manuscript rather than a full technical paper. The historical narrative and the rigorous polar-code statements are solid, and the capacity-achieving inclusion argument is correct. The main concern is that the paper's new quantitative claim—near-dispersion FER for PAC codes—rests on a single unreproduced simulation and an informal heuristic, which the authors themselves label as open. I recommend major revision with the goal of either providing reproducible simulation details and quantified evidence or explicitly reframing the PAC-code finite-length claims as preliminary and conjectural. The self-citation pattern is mild and appropriate for this format."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis one is worth your time, but not for the reasons the abstract suggests. The historical narrative from sequential decoding to polar codes is a well-told review, and the new thing is the PAC code family: put a convolutional precoder before the polar transform and decode with sequential decoding over the induced irregular tree. That construction is genuinely new, and it unifies polar and Reed-Muller rate profiles in a way that has already sparked follow-up work.\n\nThe paper is honest about what is proved and what is not. The capacity-achieving statement is sound because setting the convolution to the identity recovers polar codes. The polar-code claims are properly cited. The authors explicitly say PAC performance and complexity \"are yet to be studied rigorously,\" and the open-problems section is candid about the sequential decoder complexity question.\n\nThe soft spot is the load-bearing finite-length claim. The near-dispersion FER at N=128, R=1/2 rests on a single simulation, Fig. 12, with no released code, no decoder pseudocode, no error bars, and no complexity statistics. The \"looks sufficiently random\" explanation for why G=TPn performs near dispersion is informal. That said, the paper does not oversell it; the language is careful and the open problems are real. The issue is not circularity or internal inconsistency, it is that the central empirical contribution is not yet reproducible from the text alone.\n\nThis is a paper for coding theorists and information theorists who want to understand where polar codes came from and where they might go. It deserves a serious referee: the construction is important, the exposition is clear, and the honesty about limitations is exemplary. The right outcome is probably acceptance with a request for a reproducible simulation appendix or a clear separation between the review material and the new empirical claims.\n\nI'd bring it to reading group and I'd cite it, mostly for the PAC construction and the rate-profile viewpoint.\n\nBest.","headline":"The PAC code construction is a genuinely new idea with a sound capacity claim, but the near-dispersion FER result is a single unreproduced simulation and is honestly labeled as unproven.","tokens_in":14850,"tokens_out":1591,"would_cite":true,"duration_ms":16179,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B10","94B35","94A24"],"pacs":[],"model":"deepseek-v4-flash","headline":"PAC codes, a convolution followed by the polar transform and sequential decoding, bring finite-blocklength performance close to the random-coding dispersion bound while retaining polar codes' capacity-achieving property.","keywords":["channel polarization","polar codes","PAC codes","sequential decoding","Fano decoder","cutoff rate","finite blocklength","BIAWGN channel"],"falsifier":"Simulate PAC codes with Reed-Muller rate profiling on the binary-input additive-white-Gaussian-noise channel at $N=256$ and $N=512$, rate $1/2$, and compare their frame error rates against the dispersion approximation at those lengths; if the gap widens as $N$ grows, or if per-block decoding complexity grows super-polynomially even though the rate profile stays below the polarized cutoff-rate profile, the \"sufficiently random\" heuristic and the complexity guide would be refuted.","tokens_in":13858,"feed_emoji":"📡","tokens_out":15077,"duration_ms":137598,"temperature":0.7,"pith_summary":"The paper retells the intellectual path from sequential decoding to channel polarization, then uses that path to introduce a new code family, polarization-adjusted convolutional (PAC) codes. The central claim is that placing an outer convolutional code before the polar transform and decoding the resulting irregular tree code with a Fano sequential decoder yields codes that still achieve channel capacity—they contain polar codes as the special case where the convolution is the identity—and that at practical blocklengths they perform far better than polar codes. The paper's headline simulation shows a PAC code with $N=128$, rate $1/2$, and a weight-ordered (Reed-Muller) rate profile achieving frame error rates close to the dispersion approximation for the binary-input additive-white-Gaussian-noise (BIAWGN) channel, the finite-blocklength benchmark for random codes. A sympathetic reader should care because this offers a concrete route to closing the gap between polar coding's asymptotic optimality and its short-blocklength performance.","feed_headline":"Convolution-then-polar codes nearly hit the short-blocklength limit","feed_subtitle":"Adding a convolutional pre-code and sequential decoding lets PAC codes approach the dispersion limit at N=128 and still reach capacity.","key_machinery":"The central object is the PAC encoding transform $x = v T P_n$, where $v$ is a rate-profiled data carrier, $T$ is an upper-triangular Toeplitz matrix implementing convolution with impulse response $c$, and $P_n = [[1,0],[1,1]]^{\\otimes n}$ is the polar transform. Decoding uses a Fano sequential (depth-first tree-search) decoder over the irregular tree code generated by $T$ under the constraint that frozen coordinates are zero, with a time-varying bias metric computed recursively as in successive cancellation. This construction is an upper-lower decomposition of a generator matrix: it separates coding into a sparse convolution and a fast polar transform, and the paper argues that for good choices of the data index set $A$ and $c$, the combined matrix $G = T P_n$ looks sufficiently random to give near-dispersion performance while keeping encoding complexity $O(N \\log N)$.","core_discovery":"On the paper's own terms, the discovery is that the cutoff-rate boosting that motivated polar coding can be recovered at finite blocklengths by undoing the 0-1 rate simplification of polar codes: instead of freezing bit-channels, PAC codes run a convolutional code over the polarized bit-channels and decode the whole system as one irregular tree code. The paper reports that with the Reed-Muller design rule for the data index set and a suitably chosen convolution, the overall transform $G = T P_n$ behaves as if it were a random code, bringing the frame error rate at $N=128$, $R=1/2$ close to the binary-input additive-white-Gaussian-noise (BIAWGN) dispersion approximation for error rates above $10^{-3}$. It further claims that PAC codes achieve channel capacity in general because an identity convolution reduces them to polar codes.","pith_inferences":["If the \"sufficiently random\" explanation is correct, PAC codes should track the dispersion approximation across a range of rates and blocklengths; testing this at $N=256$ and $512$ and at rates away from $1/2$ would turn the heuristic into a measurable prediction.","The condition that the rate profile stay below the polarized cutoff-rate profile resembles a finite-length error-exponent comparison; making it precise could connect sequential-decoding complexity to coding error exponents rather than only to capacity.","The upper-lower-decomposition viewpoint suggests searching for other sparse factorizations of generator matrices: any fast transform paired with a compatible outer trellis might yield codes with near-dispersion behavior under an appropriate decoder.","If the Reed-Muller rate profile is shown to be universal across binary-input memoryless channels of a given capacity, PAC codes would become channel-agnostic finite-length codes, a stronger property than the channel-specific rate profiles used for polar codes."],"forward_implications":["PAC codes achieve channel capacity on symmetric binary-input memoryless channels, since taking the convolution to be the identity recovers polar codes.","At finite blocklengths PAC codes can outperform polar codes under both successive-cancellation and CRC-aided successive-cancellation list decoding, as the $N=128$, rate-$1/2$ BIAWGN simulation shows.","The rate-profile heuristic gives a design rule: a data index set whose cumulative rate stays below the polarized cutoff-rate profile should keep Fano decoding complexity manageable at that signal-to-noise ratio.","The best simulated performance came from a weight-ordered (Reed-Muller) rate profile, suggesting PAC codes may tolerate channel parameter variations better than polar codes; the paper leaves a rigorous universal-design statement open.","The main practical obstacle is the variable complexity of sequential decoding, and the paper points to fixed-complexity alternatives such as list Viterbi and beam search as future directions."],"supporting_citations":[{"why":"Provides the polar transform, successive-cancellation decoding, and the capacity-achieving theorem that PAC codes contain as a special case.","marker":"[13]"},{"why":"Establishes that the cutoff-rate barrier can be broken in principle by an inner block code with outer sequential decoding, motivating the design.","marker":"[8]"},{"why":"Introduces the channel-splitting example that boosts cutoff rate while conserving capacity.","marker":"[9]"},{"why":"Provides the multi-level-coding and multi-stage-decoding bit-channels whose polarization and cutoff-rate profiles underlie PAC rate design.","marker":"[12]"},{"why":"Gives the finite-blocklength dispersion approximation used as the benchmark in the N=128, R=1/2 simulation.","marker":"[19]"},{"why":"Supplies the Fano sequential decoder used to search the irregular tree code.","marker":"[21]"},{"why":"Provides the CRC-aided successive-cancellation list decoder used as a baseline in the performance comparison.","marker":"[20]"},{"why":"Establishes the polar-code error exponent that motivates improving short-blocklength performance.","marker":"[17]"},{"why":"Underlies the weight-based (Reed-Muller) score function for choosing the data index set.","marker":"[22]"},{"why":"Together with the weight-based score function, defines the Reed-Muller rate profile used in the best simulation.","marker":"[23]"}],"fun_headline_variants":["PAC codes near the short-blocklength dispersion limit","Convolutional precoding sharpens polar codes at N=128","Undoing polar's rate simplification boosts short-block PAC codes","From polar to PAC: closing the short-blocklength gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that at the blocklengths of interest the combined transform $G = T P_n$ behaves statistically like a random code, so PAC codes inherit the near-maximum-likelihood performance predicted by the dispersion approximation; the paper presents this as an informal heuristic, not a proof.","fun_headline_variants_meta":{"raw":{"variants":["PAC codes near the short-blocklength dispersion limit","Convolutional precoding sharpens polar codes at N=128","Undoing polar's rate simplification boosts short-block PAC codes","From polar to PAC: closing the short-blocklength gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001448,"raw_usage":{"total_tokens":5732,"prompt_tokens":748,"completion_tokens":4984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":364,"completion_tokens_details":{"reasoning_tokens":4917}},"tokens_in":364,"tokens_out":4984,"duration_ms":33288,"temperature":1.0,"reasoning_tokens":4917,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:06:32.061892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate PAC codes with Reed-Muller rate profiling on the binary-input additive-white-Gaussian-noise channel at $N=256$ and $N=512$, rate $1/2$, and compare their frame error rates against the dispersion approximation at those lengths; if the gap widens as $N$ grows, or if per-block decoding complexity grows super-polynomially even though the rate profile stays below the polarized cutoff-rate profile, the \"sufficiently random\" heuristic and the complexity guide would be refuted.","supporting_citations":[{"cited_title":"Channel polarization: A method for constru cting capacity- achieving codes for symmetric binary-input memoryless cha nnels,","cited_arxiv_id":null,"evidence_quote":"Provides the polar transform, successive-cancellation decoding, and the capacity-achieving theorem that PAC codes contain as a special case."},{"cited_title":"On the complexity of decoding,","cited_arxiv_id":null,"evidence_quote":"Establishes that the cutoff-rate barrier can be broken in principle by an inner block code with outer sequential decoding, motivating the design."},{"cited_title":"Capacity, cutoff rate, and coding for a direc t-detection optical channel,","cited_arxiv_id":null,"evidence_quote":"Introduces the channel-splitting example that boosts cutoff rate while conserving capacity."},{"cited_title":"A new multilevel coding method using error- correcting codes,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-level-coding and multi-stage-decoding bit-channels whose polarization and cutoff-rate profiles underlie PAC rate design."},{"cited_title":"Channel coding r ate in the finite blocklength regime,","cited_arxiv_id":null,"evidence_quote":"Gives the finite-blocklength dispersion approximation used as the benchmark in the N=128, R=1/2 simulation."},{"cited_title":"A heuristic discussion of probabilistic deco ding,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fano sequential decoder used to search the irregular tree code."},{"cited_title":"List decoding of polar codes,","cited_arxiv_id":null,"evidence_quote":"Provides the CRC-aided successive-cancellation list decoder used as a baseline in the performance comparison."},{"cited_title":"On the rate of channel polariz ation,","cited_arxiv_id":null,"evidence_quote":"Establishes the polar-code error exponent that motivates improving short-blocklength performance."},{"cited_title":"A class of multiple-error-correcting codes a nd the decoding scheme,","cited_arxiv_id":null,"evidence_quote":"Underlies the weight-based (Reed-Muller) score function for choosing the data index set."},{"cited_title":"Application of Boolean algebra to switch ing circuit design and to error detection,","cited_arxiv_id":null,"evidence_quote":"Together with the weight-based score function, defines the Reed-Muller rate profile used in the best simulation."}],"review_version":1}