{"id":"8e44cba2-5ba0-4619-968a-42c1c9dd550c","arxiv_id":"1908.02462","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A greedy relocation framework builds multi-dimensional spatially-coupled LDPC codes with fewer short cycles and substantially better bit-error-rate than comparable one-dimensional codes.","lead":"This paper connects several one-dimensional spatially-coupled error-correcting codes into a multi-dimensional code by relocating the groups of connections that cause short cycles, which reduces those harmful cycles. Simulations show large bit-error-rate gains over single-chain codes, and a windowed decoder keeps the added decoding latency low.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimization objective counts only active cycles-k of a single constituent SC code, leaving cross-segment cycles in H_MD unmodeled; this is the main correctness risk.","rationale":"The paper makes a clear and useful contribution: Theorems 1 and 2 are internally consistent, the construction is parameterized and systematic, the simulations show large BER gains, and the latency analysis of the MD windowed decoder is sound as far as it goes. The reader's weakest assumption is precisely the gap between the optimized proxy (active cycles-k of a single constituent code) and the claimed outcome (reduced total short-cycle population of the final MD-SC code). This gap is load-bearing because the trimming step of Algorithm 2 discards solutions based solely on active cycle counts, and the theorems do not address cycles that are created by combining relocated circulants from different constituent codes across multiple replicas. Such cross-chain cycles are not representable as instances of a single cycle of one H_SC, so they are outside the set Gamma entirely. The empirical results demonstrate that the proxy works well for the tested cases, but they do not establish it generally, and no code or data are provided to independently verify the reported cycle counts or BER curves. Thus the central claim is plausible and well-supported in the tested regime, but not fully established; a CONDITIONAL verdict is appropriate. I agree with the reader that this is the weakest assumption and the main reason the verdict should remain CONDITIONAL rather than ACCEPT.","tokens_in":22503,"tokens_out":2046,"duration_ms":23219,"concrete_test":"For at least one reported parameter set, e.g., SC-Code 1 with L2=5, d=2, T=18, and also L2=5, d=5, compute: (a) the number of active cycles-6 used by Algorithm 2; (b) the total number of cycles-6 and cycles-8 in the final H_MD, subdivided into cycles contained within one constituent chain versus cycles that traverse more than one chain. If cross-chain cycles are non-negligible, or if an alternative run of Algorithm 2 that also minimizes cross-chain cycles yields a different code with fewer total short cycles, then the proxy is insufficient and the central claim needs qualification. If the two rankings agree, the concern is largely resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the score-voting relocation framework produces MD-SC codes with notably fewer short cycles and dramatic BER gains. Theorems 1 and 2 correctly analyze the fate of the L2 instances of a single cycle of one constituent 1D-SC code under relocation. However, the construction's decision and trimming steps optimize a proxy: the number of active cycles-k, i.e., cycles of one constituent H_SC that visit the middle replica and for which the Ineffective Relocation Condition (IRC, Eq. (4)) holds. Algorithm 2 enumerates only the set Gamma of such cycles (steps 3, 7, 19-20) and scores relocation options only through them (Algorithm 1). The final H_MD of Eq. (3) can contain cycles that traverse several constituent SC codes through the coupling segments At, and that do not correspond to any single cycle of any one H_SC. Such cross-chain cycles are never enumerated, scored, or trimmed. Consequently, the link between minimizing active cycles-k and minimizing the total short-cycle population of H_MD is empirical, not structural. The simulations support the proxy for the tested parameters, which is real evidence, but the search is explicitly greedy and trimmed to the proxy, so the claimed connection could fail in other regimes or even for the reported codes if cross-chain cycles are numerous. The paper does not quantify cross-chain cycles or compare the final total cycle count against the proxy. This concern agrees with the reader's weakest_assumption and is the most load-bearing correctness risk in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for constructing multi-dimensional spatially-coupled (MD-SC) LDPC codes by relocating problematic circulants from constituent 1D-SC codes into auxiliary coupling matrices. The construction is defined by the MD coupling depth d, coupling length L2, and density T, and the relocation decisions are made by a greedy tree-search algorithm driven by a score-voting scheme. Theorems 1 and 2 characterize how the L2 instances of a short cycle in a constituent code are transformed into cycles of length k, 2k, or L2k depending on the relocation mapping, with the Ineffective Relocation Condition (IRC) identifying when relocations fail to lengthen cycles. The paper also presents a multi-dimensional windowed decoder with latency analysis. Simulations for girth-6 and girth-8 constituent codes report large reductions in cycle counts and BER compared to 1D-SC counterparts and to random MD-SC constructions.","tokens_in":22745,"tokens_out":12758,"duration_ms":147388,"significance":"If the framework is understood as a heuristic design method, it is a useful and substantial contribution: it generalizes MD-SC construction beyond random or topology-specific designs, allows arbitrary L2 and d, and provides explicit mapping matrices for reproducibility. Theorems 1 and 2 are internally consistent and the cycle-count and BER simulations are detailed. The comparison with random relocation policies in Section VI.C is valuable and gives direct evidence that informed relocation matters. The main significance is therefore practical: a systematic, validated construction and a low-latency decoder for MD-SC codes. The theoretical guarantee, however, is per-base-cycle rather than a global minimization of the short-cycle population, and the greedy algorithm optimizes a proxy; this limits the strength of the theoretical claims but not the empirical contribution.","major_comments":[{"comment":"The search and trimming in Algorithm 2 optimize the number of active cycles-k (cycles of H_SC visiting the middle replica for which IRC holds), but the stated goal and the reported tables are about the total short-cycle population of H_MD. The paper does not establish that minimizing the active-cycle count also minimizes the total number of cycles in H_MD, nor does it quantify cycles not represented in Γ, such as base cycles of H_SC that do not visit the middle replica or cycles whose projection traverses several constituent codes. Please provide either a dominance argument or an empirical comparison of the proxy against total cycles-6 and cycles-8 at each trimming step; without this, the connection between the optimization objective and the claimed cycle enhancement remains a heuristic rather than a property of the framework.","section":"Section IV.B, Algorithm 2 (steps 7 and 19-20)"},{"comment":"Theorems 1 and 2 analyze only the L2 instances of a single fixed cycle O_k of one constituent code, but the final matrix (3) also contains cycles that are not such instances. I believe every cycle in H_MD projects to a closed walk in H_SC because circulant powers are unchanged by relocation, but the manuscript does not state or prove this, and it does not show that all projected base cycles are captured by the set Γ of middle-replica cycles. Please add a formal statement of the correspondence between cycles of H_MD and cycles of H_SC, and clarify the exact scope of the per-cycle guarantees; this would also sharpen the discussion of what Algorithm 2 can and cannot be expected to control.","section":"Section IV.A, Theorems 1-2"}],"minor_comments":[{"comment":"Equation (4) and the expressions for δH, δV, and Δ_Ok use equality where they mean congruence modulo L2. Please make the modular notation explicit, since the proof and the scores R(O_k,t) depend on this distinction.","section":"Section IV.A, Eq. (4)-(6)"},{"comment":"The condition \"if L2/x = 0\" is ambiguous; it should read \"if x divides L2\" or \"if (L2 mod x) = 0\".","section":"Algorithm 1, step 10"},{"comment":"There are several typos and draft artifacts: \"perfromance\" on page 26, \"blueto\" in Example 3, \"girthblues\" in Section VI, and stray \"blue\" and \"magenta\" color commands in Sections III and VI. These should be cleaned in the final version.","section":"Section VI"},{"comment":"The phrasing \"if two VNs do not to share CNs\" is grammatically incorrect and appears twice; please revise to \"do not share CNs\".","section":"Section V.A"},{"comment":"The BER comparisons use SNR in dB while the MD-SC codes have slightly lower rates than their 1D counterparts (0.74 vs. 0.76 and 0.81 vs. 0.83). Reporting Eb/N0 as well would remove any concern that the small rate difference contributes to the observed gains.","section":"Section VI, Figs. 7-9"},{"comment":"The printed MD mapping matrices M2-M8 are hard to verify without row and column labels; a short explanation of the correspondence between matrix entries and the constituent codes would aid reproducibility.","section":"Section VII, Appendix"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within scope and the empirical results are strong. My main uncertainty is whether the greedy proxy in Algorithm 2 is sufficiently tied to the stated objective of minimizing total short cycles; this is fixable by adding a formal projection argument or an empirical proxy-vs-total-cycle comparison. I recommend major revision rather than rejection because the central construction and simulations are sound, but the theoretical framing needs to be made precise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. It's a genuine advance in finite-length SC code design, not just another random-coupling paper. The construction connects any number of 1D-SC codes at arbitrary depth by relocating the circulants that participate in the most 'active' short cycles, and the score-voting scheme with tree search is a sensible way to choose relocations. Theorems 1 and 2 give a clean modular characterization of what happens to a single cycle under relocation: it either persists as L2 short cycles or merges into one long cycle, with the gcd condition controlling the outcome. That part is correct and genuinely useful.\n\nThe simulations support the claims. Tables I-III show large reductions in total cycles-6/8, not just the proxy, and the BER gains (orders of magnitude over the 1D counterparts) are the kind of result that makes people in storage coding pay attention. The MD windowed decoder with latency analysis preserves the low-latency property, which is a nice bonus.\n\nSoft spots, roughly in order:\n\n- The optimization objective only counts active cycles-k from one constituent SC code; cross-chain cycles in H_MD are not modeled, so the link between the score and the total short-cycle population is empirical rather than structural. The paper does mitigate this by reporting total cycle counts for the final codes, so the gap is theoretical, not a demonstrated failure. Still, a skeptic could construct a regime where the proxy misleads. Worth a comment in the paper, not a rejection.\n\n- Section VI-C has a name swap: the text says Code 7's relocations are random while Code 10's are informed, which contradicts the construction. Looks like a typo, but it should be fixed.\n\n- No code or data, BER curves without error bars. For a construction paper this is forgivable but worth asking about in review.\n\n- Baselines are 1D OO-CPO codes from the authors' own prior work. That's legitimate since OO-CPO is published, but an independent 1D baseline would strengthen the comparison.\n\nOverall this is a solid contribution that deserves a serious referee. I'd send it to review, conditional on the authors addressing the proxy question honestly and fixing the textual errors. The construction is reproducible from the appendix, and the cycle-enumeration numbers are concrete enough to verify.\n\nRecommendation: engage with it. Bring to reading group maybe.","headline":"A systematic, tunable MD-SC construction that delivers large measured cycle and BER gains; the main gap is that the optimization scores a proxy (active cycles of one constituent) rather than total short cycles, but the reported totals keep the claims honest.","tokens_in":23343,"tokens_out":2409,"would_cite":true,"duration_ms":24391,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B60"],"pacs":[],"model":"deepseek-v4-flash","headline":"A score-voting relocation scheme rewires the circulants that cause the most short cycles, converting them into longer cycles in multi-dimensional spatially-coupled codes.","keywords":["spatially-coupled codes","LDPC codes","circulant permutation matrices","multi-dimensional coupling","cycle properties","error floor","windowed decoding","score-voting optimization"],"falsifier":"Enumerate every cycle of length $k$ and length $2k$ in the final parity-check matrix $H_{\\mathrm{MD}}^{\\mathrm{SC}}$ for the codes studied in the paper, using the reported mapping matrices, and compare the total counts with the active-cycle counts used by the construction. A single case in which a code with fewer active cycles has more total short cycles, especially cycles that travel through more than one constituent chain, would show that the optimization target is not the quantity being improved.","tokens_in":22230,"feed_emoji":"🔗","tokens_out":7847,"duration_ms":76725,"temperature":0.7,"pith_summary":"This paper claims that the short cycles that limit finite-length performance of spatially-coupled (SC) error-correcting codes can be systematically lengthened by wiring several SC codes together. It introduces a score-voting construction that relocates the most cycle-heavy circulant blocks from the middle of each constituent SC code into the same positions of neighbouring SC codes, and it proves how the length of a cycle changes under such relocations. The resulting multi-dimensional SC codes are reported to have roughly 90 to 99 percent fewer short cycles in the examples, and correspondingly lower bit-error rates even though code length and rate are kept comparable. The design also preserves the low-latency windowed decoding property of one-dimensional SC codes. This matters for data-storage and communication systems that need strong, cheaply decodable codes under bursty or spatially non-uniform noise.","feed_headline":"Rewiring hot circulants cuts short cycles by up to 99 percent","feed_subtitle":"Score-voting relocation lengthens the cycles that hurt iterative decoding while keeping windowed decoding fast.","key_machinery":"The load-bearing object is the Ineffective Relocation Condition (IRC), the congruence $\\sum_{u=1}^k (-1)^u M(C_{i_u,j_u})\\equiv 0 \\pmod{L_2}$, which detects whether the relocated instances of a cycle still close within one segment. Around it the paper builds a score: for a cycle $O_k$ and a relocation option $t$, the score is $R(O_k,t)=L_2/\\gcd(L_2,\\Delta_{O_k})$, the length of the cycles that option produces. Algorithm 1 assigns these scores over all active cycles visiting a targeted circulant; Algorithm 2 grows a tree of relocation sequences, always moving the circulant visited by the most active cycles, expanding all surviving options, and trimming every branch whose active-cycle count is not minimal. This machinery carries the whole argument because it converts the design goal, fewer small cycles, into a tractable greedy search over circulant relocations.","core_discovery":"The central discovery is an exact rule for what happens to a short cycle when some of its circulants are moved between coupled SC chains. For a cycle $O_k$ of length $k$ in a constituent 1D-SC code, let $M(C_{i,j})$ be the auxiliary matrix index to which each visited circulant is relocated. The instances of the cycle across the $L_2$ constituent chains merge into $\\tau$ cycles of length $L_2 k/\\tau$, where $\\tau = \\gcd(L_2,\\Delta_{O_k})$ and $\\Delta_{O_k}$ is the alternating sum of the relocation indices modulo $L_2$; when the Ineffective Relocation Condition $\\sum_{u=1}^k (-1)^u M(C_{i_u,j_u})\\equiv 0 \\pmod{L_2}$ holds, $\\tau=L_2$ and the short cycles survive unchanged. The construction algorithm, by relocating the circulants that belong to the most active cycles so that this congruence fails, converts those short cycles into cycles of length up to $L_2 k$ while keeping the overall parity-check matrix diagonal and locally structured. Simulations then show the resulting MD-SC codes have far smaller counts of cycles-6 and cycles-8 than one-dimensional SC codes of the same length and rate, with BER gains reaching several orders of magnitude in the waterfall and error-floor region.","pith_inferences":["Because the Ineffective Relocation Condition mirrors the Fossorier cycle condition for quasi-cyclic lifting, the same relocation score could probably be adapted to target other graph objects, such as absorbing sets, trapping sets, or minimum-distance bounds, by scoring a different cycle length.","The paper's optimization counts only cycles that live within a single constituent SC code; an immediate testable extension is to add cross-chain cycles to the objective and see whether the active-cycle proxy or the BER estimates change.","The early termination seen when the relocation density $T$ grows suggests a trade-off curve between density and depth; choosing both jointly by tracking the marginal reduction in active cycles is a natural refinement that Algorithm 2 already exposes."],"forward_implications":["For the girth-6 constituent code, MD-SC codes with $L_2=5$ and depth $d=5$ cut the number of cycles-6 by about 99 percent relative to a one-dimensional SC code of the same length, which the paper reports as BER gains of several orders of magnitude near $3.85$ dB.","For the girth-8 constituent code, increasing the depth beyond $d=2$ gives smaller additional gains; depth $d=4$ cuts active cycles-8 by roughly 82 percent compared with the one-dimensional counterpart, suggesting modest depth is enough for error-floor improvement.","Because relocations always copy a circulant to the same position in another segment, the window structure of each constituent chain is preserved, so the MD windowed decoder has identical window configurations and its latency is bounded by $(W_D+m)/L$ times the block decoder latency.","The framework is parameterized by the number of constituent chains $L_2$, the coupling depth $d$, the density $T$ of relocated circulants, and the target cycle length $k$, so the same algorithm applies to other underlying SC codes and to channels that have a known problematic cycle length."],"supporting_citations":[{"why":"Supplies the OO-CPO-optimized 1D-SC codes used as the constituent codes in all simulations, giving the improved starting point needed for the reported BER comparisons.","marker":"[7]"},{"why":"Introduces connecting spatially coupled chains into multi-dimensional codes, defining the construction space this paper improves upon.","marker":"[10]"},{"why":"Provides the random multi-dimensional SC construction that the paper's deterministic framework is compared against.","marker":"[11]"},{"why":"Presents laterally connected SC chains and an earlier idea of multi-dimensional windowed decoding that this paper formalizes.","marker":"[14]"},{"why":"Fossorier's condition for cycles in circulant-permutation-based quasi-cyclic LDPC codes is the template for the Ineffective Relocation Condition.","marker":"[29]"},{"why":"The tree-based search for harmful objects in SC codes inspires the tree structure used in Algorithm 2.","marker":"[30]"},{"why":"Establishes windowed decoding for protograph-based SC codes, whose latency model the MD windowed decoder extends.","marker":"[31]"},{"why":"Analyzes windowed decoding of spatially coupled codes, supplying the latency bound that the multi-dimensional windowed decoder inherits.","marker":"[32]"}],"fun_headline_variants":["Relocating hot circulants cuts short cycles by up to 99%","MD-SC code design: lengthen cycles via informed moves","How to make short cycles vanish in coupled codes","Multi-D coupling boosts BER by cycle lengthening","Cycle-aware relocation improves SC code performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design rests on the assumption that the count of active short cycles, cycles of the constituent code that pass through the middle replica and survive relocation, faithfully represents how many short cycles the finished multi-dimensional code will have, even though cycles that cross between constituent chains are never counted.","fun_headline_variants_meta":{"raw":{"variants":["Relocating hot circulants cuts short cycles by up to 99%","MD-SC code design: lengthen cycles via informed moves","How to make short cycles vanish in coupled codes","Multi-D coupling boosts BER by cycle lengthening","Cycle-aware relocation improves SC code performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000583,"raw_usage":{"total_tokens":2794,"prompt_tokens":1050,"completion_tokens":1744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":1667}},"tokens_in":666,"tokens_out":1744,"duration_ms":14229,"temperature":1.0,"reasoning_tokens":1667,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:44:55.836812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate every cycle of length $k$ and length $2k$ in the final parity-check matrix $H_{\\mathrm{MD}}^{\\mathrm{SC}}$ for the codes studied in the paper, using the reported mapping matrices, and compare the total counts with the active-cycle counts used by the construction. A single case in which a code with fewer active cycles has more total short cycles, especially cycles that travel through more than one constituent chain, would show that the optimization target is not the quantity being improved.","supporting_citations":[{"cited_title":"Finite-Length Construction of High Performance Spatially-Coupled Codes via Optimized Partitioning and Lifting,","cited_arxiv_id":null,"evidence_quote":"Supplies the OO-CPO-optimized 1D-SC codes used as the constituent codes in all simulations, giving the improved starting point needed for the reported BER comparisons."},{"cited_title":"New codes on graphs constructed by connecting spatially coupled chains,","cited_arxiv_id":null,"evidence_quote":"Introduces connecting spatially coupled chains into multi-dimensional codes, defining the construction space this paper improves upon."},{"cited_title":"Multi-dimensional spatially-coupled codes,","cited_arxiv_id":null,"evidence_quote":"Provides the random multi-dimensional SC construction that the paper's deterministic framework is compared against."},{"cited_title":"Laterally connected spatially coupled code chains for transmission over unstable parallel channels,","cited_arxiv_id":null,"evidence_quote":"Presents laterally connected SC chains and an earlier idea of multi-dimensional windowed decoding that this paper formalizes."},{"cited_title":"Quasi-cyclic low-density parity-check codes from circulant permutation matrices,","cited_arxiv_id":null,"evidence_quote":"Fossorier's condition for cycles in circulant-permutation-based quasi-cyclic LDPC codes is the template for the Ineffective Relocation Condition."},{"cited_title":"Efficient Search and Elimination of Harmful Objects in Optimized QC SC-LDPC Codes","cited_arxiv_id":"1904.07158","evidence_quote":"The tree-based search for harmful objects in SC codes inspires the tree structure used in Algorithm 2."},{"cited_title":"Windowed decoding of protograph-based LDPC convolutional codes over erasure channels,","cited_arxiv_id":null,"evidence_quote":"Establishes windowed decoding for protograph-based SC codes, whose latency model the MD windowed decoder extends."},{"cited_title":"Windowed decoding of spatially coupled codes,","cited_arxiv_id":null,"evidence_quote":"Analyzes windowed decoding of spatially coupled codes, supplying the latency bound that the multi-dimensional windowed decoder inherits."}],"review_version":1}