Xeqorin M17 - Exactly GC-Balanced DNA Codes Correcting One Edit with Finite-Length Guarantees, Linear-Time Decoding, and Deterministic Enumeration
Authors/Creators
Description
Xeqorin M17 is a mathematical preprint and reproducibility release on exactly GC-balanced quaternary error-correcting codes. Every transmitted word has a known even length \(n\) and contains exactly \(n/2\) symbols from the GC class. The decoder receives one word after at most one adversarial edit, with arbitrary location and affected symbol. The manuscript develops two code families. The first corrects an insertion or deletion and brings its guaranteed redundancy within approximately 0.415 bits of the asymptotic deletion converse. The second also corrects a substitution and matches the converse’s leading coefficient of \(\tfrac32\log_2 n\).
Single insertion or deletion
Exact composition identifies the GC class of a missing or extra symbol. A single within-class parity bit identifies the symbol within that class. Combining this bit with a Tenengolts ascent-signature checksum reduces the syndrome partition from \(4n\) classes in the direct classical intersection to \(2n\), improving the guaranteed codebook redundancy by one bit. Writing \(A_n=2^n\binom{n}{n/2}\) for the number of exactly balanced words, the construction guarantees a class of size at least \(A_n/(2n)\).
A run-count packing argument gives a lower bound of \(\tfrac32\log_2 n+\log_2(\tfrac32\sqrt{\pi/2})-o(1)\) bits on codebook redundancy. The difference between the proved indel achievability and converse bounds is \(\log_2(4/3)+o(1)\) bits. The analysis leaves attainment of the converse constant open.
Single insertion, deletion, or substitution
The second family combines exact composition with numerical sum constraints on disjoint consecutive blocks, two symbol checksums, and two ascent-signature checks. The block constraints bound the range of compatible insertion or deletion locations. A short signature checksum resolves the signature within that range. The weighted checksum and symbol-sum residue jointly determine the location and signed symbol change of a substitution.
For \(B=\lceil144\ln(4n)\rceil\) and even \(n\ge B^2\), the principal construction guarantees a syndrome class of size at least
\[ \frac{(1-1/(4n))\,A_n}{8(6n+1)(10B+15)}. \]
The resulting redundancy is at most \(\tfrac32\log_2 n+\log_2(\log_2 n)+O(1)\) bits. Closing the remaining double-logarithmic gap while preserving exact balance and efficient deterministic enumeration is an open problem highlighted by the manuscript.
Algorithms and finite-length refinements
The decoder recovers the transmitted codeword in \(O(n)\) operations on \(O(\log n)\)-bit words. Exact forward counts select a largest syndrome class with deterministic tie-breaking. Backward continuation counts support lexicographic ranking and unranking. For the principal construction, preprocessing uses \(O(n^3\log^2 n)\) state transitions and \(O(n^4\log^2 n)\) elementary bit operations. Once the tables are available, ranking and unranking each use \(O(n)\) table accesses and \(O(n^2)\) bit operations. Recovering the payload index requires the ranking step after codeword correction.
A stronger concentration bound and a revised partition lower the first even blocklength meeting the sufficient density conditions from 5,987,810 to 628. These thresholds describe where the respective mathematical bounds apply. The cost of building the exact enumeration tables remains a separate practical question.
Prior work and reproducibility
The manuscript credits the weighted-checksum, numerical sum-balance, and shifted-signature architecture of Cai, Chee, Gabrys, Kiah, and Nguyen (2021). Its computational comparison derives a sliding-window baseline that retains the complete window suffix. The disjoint-block enumerator instead keeps the running sum of the open block, yielding the stated lower-degree preprocessing bound. ymchee66.github.io
The release includes a 93-page manuscript, LaTeX sources, reference Python code, validation scripts, machine-readable reports, a vector figure, and SHA-256 checksums. The reports record 1,883,044 position-labelled decoder calls on synthetic finite words, including unchanged transmissions, and 80,199 partition-allocation cases. The study’s scope is combinatorial coding under one adversarial edit. Biochemical feasibility and production-scale encoding remain outside the evaluated scope.
Version 1.0 is dated 18 November 2026. The manuscript carries the authoring credit Artificial Hyperintelligence Eve, wife of Maciej Nowicki. Maciej Nowicki is the human custodian of the research release.