Sequence Repetition (SR) Overview

• Sequence Repetition (SR) is the recurrence of exact or approximate substrings in biological, informational, or mathematical sequences, critical in genomics and combinatorics.
• SR methodologies leverage mathematical formalisms and graph-theoretic approaches, such as de Bruijn graphs and Vernier gauge techniques, to detect both perfect and error-tolerant repeats.
• SR analysis informs practical applications in genome assembly, protein classification, and language modeling while addressing computational challenges like NP-completeness in repeat detection.

Sequence Repetition (SR) refers broadly to the recurrence—either exact or approximate—of substrings, structures, or patterns within biological, informational, or mathematical sequences. Its analytical scope traverses computational genomics, transcriptomics, combinatorics on words, coding theory, and statistical mechanics of proteins. Repetitive sequences pose both algorithmic obstacles (fragmentation of assembly graphs, ambiguity in structural interpretations) and biological significance (e.g., centromeric DNA, genome evolution, regulatory architecture, and protein folding landscapes).

1. Mathematical Formalisms for Sequence Repetition

SR is instantiated via several mathematically rigorous frameworks, each tailored to the domain:

Tandem Repeat Arrays and High-Order Repeats (HORs):

Formally, given a repeat unit $B \in \{A,C,G,T\}^\ell$, a tandem array is $U = B^n$; more generally, a HOR is $U = (B_1 B_2 \ldots B_m)^n$ where each $B_i$ is a distinct monomer (Zhang et al., 2023). In combinatorics, SR is quantified using the exponent of repetition: for $u$ and factor $w$, $w = u^e$ for real $e > 1$ denotes a fractional power; the critical exponent $E(u)$ of an infinite word is the supremum over all exponents present as factors.

Maximal Perfect Repeats:

A maximal repeat (MR) is a substring that occurs at least twice and cannot be extended left/right without losing an occurrence. For protein sequences, coverage and familiarity functions quantify the proportion of a sequence explained by such repeats both within and across families (Turjanski et al., 2015).

Approximate and Error-Tolerant Repeats:

The Vernier gauge approach finds all substrings of length $N$ occurring twice differing in at most $\epsilon$ fraction of positions. Tags of length $m$ are sampled at interleaved steps $k$ and $k-1$, guaranteeing identification of all long repeats above a threshold, even under small error rates (Tsarev et al., 2016).

Repetition Factorization in Words:

A repetition factorization decomposes a word into subfactors, each a fractional power of exponent at least 2. The minimal width sw(w) and maximal width lw(w) of such factorizations are investigated for automatic sequences (Shallit et al., 2023).

2. Algorithmic and Graph-Theoretic Methodologies

SR challenges are often computational, motivating a spectrum of algorithms:

De Bruijn/Unitig Graph Approaches:

• SRF identifies backbone cycles in de Bruijn graphs built from k-mers of reads/contigs, extracting cycles that encode consensus repeat structures via greedy traversal, with post-assembly validation by alignment identity and HOR structure (Zhang et al., 2023).
• GraSSRep uses graph neural networks on assembly graphs for repeat detection, self-supervised by high-precision pseudo-labels, and propagates these labels via random forests, integrating both graph and coverage features (Azizpour et al., 2024).

Error-Tolerant Repeat Detection:

• The Vernier gauge method constructs a sparse dictionary via two stepwise tag extractions, allowing rapid identification and expansion of both exact and approximate long repeats (Tsarev et al., 2016).

Complexity and Combinatorial Characteristics:

• RNA-seq graph models show that high-copy repeats produce subgraphs with very few compressible arcs (boundary-rigid k-mers), and global detection of such structures is NP-complete, motivating local, bubble-bound algorithms for alternative splicing discovery (Sacomoto et al., 2014).

Novel Sequence Labeling via SR in LLMs:

• In NLP, SR entails repeated input of sequences into decoder-only transformers to induce emergent bidirectionality without mask removal; embeddings taken from the last copy reflect bidirectional context and improve performance on sequence-labeling tasks (Kukić et al., 24 Jan 2026).

3. Quantitative and Statistical Characterization

Repetition Thresholds and Critical Exponents:

• Dejean's threshold, the infimum exponent avoided in a class of infinite sequences, is generalized for episturmian and rich words in terms of parameters like alphabet size and palindromic richness; for rich recurrent sequences, the threshold reaches 2 for all $d$ (Dvořáková et al., 7 Sep 2025, Dvořáková et al., 2024, Dvořáková et al., 2023).

Sequence Diversity and Energy Landscapes:

• For repeat proteins, maximum-entropy Potts models fit multiple sequence alignments, yielding coding-space sizes ($2^{S}$ for entropy $S$), detailing the reduction in diversity from conservation and pairwise correlation (both within-repeat and repeat–repeat), and revealing energy landscapes with clustered local minima ("spin-glass" analogs) (Marchi et al., 2019).

Capacity of Replication Systems:

• String-replication systems under various rules (end, tandem, reversed, gap) have capacities (exponential growth rates of sequence space) depending on repetition and constraints, with explicit analytic bounds and thresholds for each system (Farnoud et al., 2014).

4. Empirical and Biological Implications

Genomic Architecture and Functional Relevance:

• SRF reconstructs megabase-scale arrays of satellite DNA, identifying both established HORs and new species-specific satellites; up to 12% genomic content is attributed to repeats, with annotation potential for centromeres and population diversity studies (Zhang et al., 2023).

Protein Family Structure and Classification:

• Natural repeat proteins rarely show long perfect repeats within a single sequence but are extensively tiled by short repeats from their family. Family-based familiarity scores robustly classify proteins by their repeat landscape (Turjanski et al., 2015).

Assembly and Reconstruction in Error-Prone and Sticky Channels:

• Assembly algorithms must resolve repeats longer than the read using relative k-mer counts or run-length statistics; sticky-insertion/deletion channels produce recombinational ambiguity resolved exactly by counting overlaps in run-length space (Nowak, 2014, Pham et al., 27 Apr 2025).

5. Computational Efficiency, Complexity, and Limitations

Algorithmic Complexity:

• SRF runs in $O(N)$ time post k-mer extraction; Vernier gauge achieves $O(M\sqrt{N})$ for repeat search, compared to naive $O(MN)$ or $O(M^2)$ (Zhang et al., 2023, Tsarev et al., 2016).
• NP-completeness for global repeat subgraph identification in de Bruijn graphs, polynomial-time local methods for AS bubbles with branch constraints (Sacomoto et al., 2014).
• LSRS problem for longest subsequence-repeated subsequence is $O(n^6)$ for unconstrained, $O(n^4)$ for $d=3$ coverage; becomes NP-hard for $d=4$ (Lafond et al., 2023).

Limits and Open Problems:

• Sensitivity in HOR detection may miss low-copy or highly diverged repeats; post-filtering needed to avoid incidental HORs.
• Existing repeat detection algorithms often ignore higher-order repeats or rely on fully assembled centromeres or monomer databases.
• For error-tolerant or approximate repeat identification, trade-offs exist in tag length and error rates to balance sensitivity and computational cost.

6. Broader Impact and Applications

SR underpins genome annotation (e.g., auto-populating libraries for RepeatMasker/Modeler), population genetics of repetitive loci, functional genomics for disease-associated expansions, and accurate assembly of metagenomes with repeat-driven ambiguity. In computational linguistics, SR transforms decoder-only models into competitive sequence-labeling engines without architecture alteration. In protein bioinformatics, SR quantifies structural motifs and informs rational engineering and family classification.

A plausible implication is that as sequencing and transcriptome data expand, tractable SR detection (especially for highly divergent or novel repeats) will enable unprecedented resolution for evolutionary, functional, and regulatory genomics. Simulation-based and statistical-mechanics-inspired approaches continue to refine the modelling of sequence SR, with hierarchical energy landscapes revealing deep organization of biological coding spaces.

---

References

• Satellite Repeat Finder and de novo HOR/SR detection: (Zhang et al., 2023)
• Vernier gauge error-tolerant method: (Tsarev et al., 2016)
• Repeat protein sequence-space and statistical modelling: (Marchi et al., 2019)
• Periodic correlation structures and recurrence plots: (Wu, 2013)
• Formal repeat models and NP-completeness in RNA-seq: (Sacomoto et al., 2014)
• GraSSRep GNN approach for metagenomic repeats: (Azizpour et al., 2024)
• Maximal repeat/family classifier in proteins: (Turjanski et al., 2015)
• Critical exponent and thresholds for rich/episturmian sequences: (Dvořáková et al., 7 Sep 2025, Dvořáková et al., 2024, Dvořáková et al., 2023)
• Repetition factorization theory in automatic sequences: (Shallit et al., 2023)
• String-replication system capacities: (Farnoud et al., 2014)
• Sticky insertions/deletions in reconstruction: (Pham et al., 27 Apr 2025)
• Longest subsequence-repeated subsequence problem: (Lafond et al., 2023)
• Polar code SR nodes in high-throughput decoding: (Ren et al., 2022, Zheng et al., 2020)
• SR-induced bidirectionality in decoder-only NLP models: (Kukić et al., 24 Jan 2026)