Pith. sign in

REVIEW 4 major objections 4 minor 45 references

Bridging Chaos Game Representations and $k$-mer Frequencies of DNA Sequences

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proves that the chaos game representation of a DNA sequence, discretized at resolution $2^k \times 2^k$, is exactly the table of $k$-mer frequencies, and uses that equivalence to reconstruct synthetic sequences with any desired…

desk verdict The formal equivalence between CGR cell counts and k-mer frequencies is cleanly proved and worth having; the reconstruction theorem's proof has gaps that need fixing before its error bound can be trusted. read the letter →

arxiv 2506.22172 v2 pith:U7PPHML7 submitted 2025-06-27 cs.FL

Haoze He , Lila Kari , Pablo Millan Arias This is my paper · ORCID
classification cs.FL MSC 05C4568R1592D20
keywords chaosgamerepresentationfrequencyk-mervectorDeBruijnmultigraphEulerianpathsyntheticDNAsequencegenomicsignaturemarginalconstraints
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to prove that two widely used ways of summarizing a DNA sequence — the chaos game representation (CGR), a scatter of points in a square, and the table of $k$-mer counts — are the same mathematical object viewed differently. The central theorem states that for any sequence and any $k$-mer, the number of times the $k$-mer occurs equals the number of CGR points that fall in the cell associated with that $k$-mer; hence the frequency CGR at resolution $2^k \times 2^k$ is exactly the $k$-mer frequency vector. The paper also shows that the eight symmetries of the square correspond exactly to eight nucleotide permutations, and builds an algorithm that turns any prescribed $k$-mer distribution satisfying the natural overlap constraints into a synthetic DNA sequence whose empirical distribution is within any desired tolerance. If correct, every CGR-based classification, distance, or visualization is mathematically interchangeable with $k$-mer statistics, and synthetic sequence generation from a target profile becomes a controlled process.

What carries the argument

The load-bearing object is the CGR cell $c(w)$: for a $k$-mer $w$, the open square of side $2^{1-k}$ centred at the last point of the CGR of $w$. These cells tile the square in bijection with grid cells of order $k$, and because no CGR point lies on a cell boundary, counting points per cell is unambiguous. The companion mechanism is the De Bruijn multigraph whose vertices are $(k-1)$-mers and whose edge multiplicities are the target $k$-mer counts; an Eulerian cycle exists exactly when the graph is balanced and strongly connected, which is what the marginal constraints and the artificial-edge balancing step are designed to ensure.

What would settle it

Build a target distribution that satisfies the marginal balance constraints but whose $k$-mer overlap graph splits into two balanced components, count the minimum artificial edges needed to connect them, and reconstruct with the sequence length at the theorem's threshold; if the needed edges exceed $4^{k-1}$ or the empirical $\ell^1$ error exceeds $\epsilon$, the theorem's bound fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the grid-cell discretization of a CGR is a lossless encoding of $k$-mer statistics: Theorem 3 states $\operatorname{occ}(s,w) = \operatorname{card}(\operatorname{CGR}(s) \cap c(w))$ for every DNA sequence $s$ and $k$-mer $w$, where $c(w)$ is the cell whose centre is the last CGR point of the word $w$. Corollary 2 lifts this to the whole matrix: the frequency CGR of order $k$, defined by counting points in the $2^k \times 2^k$ grid, equals the table of all $k$-mer counts. The paper further proves that a symmetry of the CGR square is equivalent to applying one of eight nucleotide permutations to the sequence, and that any target distribution satisfying the marginal balance condition can be realized by an Eulerian path on a De Bruijn multigraph with an explicit error bound that tends to zero as the sequence length grows.

Load-bearing premise

The reconstruction guarantee rests on an unproved bound: that a balanced network built from overlapping $k$-mers can always be made fully connected by adding at most $4^{k-1}$ extra links, and on the target distribution satisfying the overlap balance condition at every node.

Editorial extensions

If this is right

  • Every FCGR-based classifier, distance measure, or image feature is mathematically interchangeable with the corresponding $k$-mer frequency analysis; results proved for one transfer to the other.
  • Symmetry transformations of a CGR image can be implemented by applying one of eight nucleotide permutations to the sequence, and an avoided $k$-mer transforms to the permuted $k$-mer.
  • Given any target $k$-mer distribution satisfying the marginal constraints, one can construct a synthetic sequence whose empirical distribution is within $\epsilon$ in $\ell^1$ distance by choosing the sequence length large enough.
  • The reconstruction algorithm enables generating synthetic CGR images from a user-specified $k$-mer profile, supporting data augmentation for machine-learning-based taxonomic classification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: since real linear chromosomes violate the marginal balance at their two boundary $(k-1)$-mers, the theorem's guarantee as stated applies to circularized or constrained targets; empirical reconstructions from linear fragments likely carry a small boundary error the paper does not quantify.
  • Editorial extension: the exact cell-count identity offers a cheap consistency check for any CGR implementation — count points per cell and compare with $k$-mer counts; a mismatch indicates a numerical or coding error rather than a biological signal.
  • Editorial extension: the symmetry theorem implies that augmenting a training set by reflecting or rotating CGR images is equivalent to permuting nucleotide labels in the sequence, which makes the augmentation's biological meaning explicit and could guide label-preserving augmentation design.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper formalizes three main connections: (i) symmetries of the CGR square correspond to the eight letter permutations in the set S (Theorems 1 and 2); (ii) the FCGR of order k, viewed as a 2^k × 2^k discretization of the CGR point set, is entrywise equal to the table of k-mer counts (Theorem 3 and Corollary 2); and (iii) a target k-mer distribution satisfying the marginal constraints can be approximately realized by a synthetic DNA sequence via Eulerian traversal of a De Bruijn multigraph (Theorem 5). The paper also presents numerical experiments on real genomic fragments and on distributions sampled from the constrained simplex, plus an interactive web tool implementing the reconstruction.

Significance. If the central equivalence of Theorem 3 and Corollary 2 holds, the paper supplies the first rigorous formal justification for an identification that is widely assumed in the FCGR literature, which is a genuinely useful contribution. The symmetry characterization in Theorem 1 is also clean and well proved. The reconstruction algorithm of Theorem 5 is a valuable algorithmic idea: a valid proof with a correct L1 guarantee would provide a principled way to generate synthetic DNA and synthetic CGR images for data-augmentation applications. The detailed proofs, the explicit software tool, and the reproducible experiment pipeline are strengths. However, the proof of Theorem 5 as written contains load-bearing gaps, and the experimental validation uses parameters outside the theorem's stated regime, so the reconstruction guarantee is not currently established as stated.

major comments (4)
  1. [Appendix A.3, proof of Theorem 5] The error calculation omits the rounding error. The proof defines c_i = round((n-k+1)θ_i) but then writes the componentwise error as |c'_i/(|E|+nart) - c_i/|E||, implicitly identifying θ_i with c_i/|E|. The missing term sum_i |θ_i - c_i/(n-k+1)| is at least potentially as large as 4^k/(2(n-k+1)), and if |E| is taken to be sum_i c_i rather than n-k+1, there is an additional normalization mismatch. The displayed threshold n > 2(k-1)4^k/ε + k-1 therefore does not follow from the derived bound.
  2. [Appendix A.3, proof of Theorem 5] The assertion that if G' is not strongly connected, at most 4^{k-1} additional artificial edges are needed to connect the components is stated without proof, and these connectivity edges are not included in the bound nart = sum_{v in V+} (k-1)δ(v) ≤ (k-1)4^k. A target distribution supported on two disjoint De Bruijn cycles is a concrete case in which connectivity repair is required and is not accounted for in the budget, so the inequality ||θhat - θ||_1 ≤ 2 nart/(|E|+nart) is not justified as written.
  3. [Section 4, Figure 3b and surrounding text] The experiments with hit-and-run sampling fix the sequence length at nmin = 2·4^k/ε + k-1, which is a factor k-1 smaller than the theorem's threshold n > 2(k-1)4^k/ε + k-1. For k > 2 the reported empirical regime is outside the theorem's guarantee, so Figure 3b does not validate the stated bound. This discrepancy should be corrected either by using the theorem's longer sequences or by proving a sharper bound that matches the experimental setting.
  4. [Section 4, Theorem 5 and Figure 3a] Theorem 5 assumes equation 7 holds at every (k-1)-mer, but the real-data experiments in Figure 3a use linear genomic fragments whose first and last (k-1)-mers violate equation 7. No degradation bound or modified statement for linear sequences with boundary violations is provided, so the strong agreement seen in Figure 3a is not explained by the theorem as stated. The authors should either circularize the sequences in the real-data experiments, or extend the theorem with an explicit boundary-error term.
minor comments (4)
  1. [Appendix A.2, proof of Theorem 3] The proof does not explicitly treat the initial point p0. Since p0 lies on cell boundaries for every k and the cells in Definitions 6 and 9 are open, p0 is never counted, which is consistent with the statement, but this should be stated explicitly to make the counting argument fully complete.
  2. [Appendix B.7, Proposition 5] The claim that the rows r_v have mutually disjoint supports is false: for example, when k=3, the rows for (k-1)-mers AC and CA both include the k-mers ACA and CAC. The full-rank statement may still be true via a signed-incidence argument, but the proof as written is incorrect.
  3. [Theorem 5, statement] The parameter ε ranges over [0,1], but the length condition n > 2(k-1)4^k/ε + k-1 becomes infinite at ε=0; the statement should restrict to ε>0 or handle the exact-balance case separately.
  4. [Definition 6] The notation is confusing because the cell is written as cell_k(i,j) but the center is given by (x_j,y_i), with indices swapped to match matrix convention. A short remark in the definition is helpful, but a consistent subscript notation would reduce the risk of error for readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CGR/k-mer equivalence is derived from the CGR recursion and cell geometry, and Theorem 5 is an independent constructive guarantee.

full rationale

The paper's central equivalence (Theorem 3 and Corollary 2) is not circular. The proof starts from Definition 9, where the cell c(w) is the open square centered at plast(w) with half-side 2^{-k}, and uses Lemma 1 to show that every prefix of s ending in w contributes a CGR point inside c(w), while the non-overlap of distinct c(w) cells forces every point inside c(w) to come from a prefix ending in w. This is a genuine derivation from the CGR recursion rather than an assumption of the conclusion. Proposition 1 then supplies the bijection between c(w) and the order-k grid cells, so the FCGR-to-k-mer-count statement follows without any fitted parameters or hidden inputs. No load-bearing claim is justified by the authors' prior work: the two self-citations ([1], [22]) appear only as contextual examples of alignment-free methods and are not used to exclude alternatives or to define the CGR-to-k-mer bridge. Theorem 5's constructive guarantee is also not circular: the algorithm builds a De Bruijn multigraph from a target distribution, repairs imbalances, and the displayed bound is derived from the artificial-edge count; the numerical experiments sample independent targets (including hit-and-run over the constraint polytope) and measure the actual L1 reconstruction error, so they would detect a wrongly derived bound. A reviewer concern about the unproved connectivity repair and the omitted rounding term in the proof of Theorem 5 is a correctness gap, not a circularity: the error bound is not obtained by assuming the desired error. The only minor definitional observation is that Definition 7 defines FCGR as the CGR discretization, so the abstract's phrase 'equivalent to a discretization' is true by definition; the substantive equivalence to k-mer frequencies is still proved, and this does not affect the derivation chain.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The theorem statements are proved from the definitions rather than fitted, so there are no data-fitted free parameters. The boundary between theorem and experiment is worth watching: the experiments use real linear sequences whose empirical distributions violate the circular balance hypothesis of Theorem 5 at two boundary (k-1)-mers, and the experimental n_min is smaller than the theorem's requirement, which is why the error bound should not be read as tight.

assumptions (4)
  • standard math Euler's criterion for directed Eulerian cycles (Theorem 4): a strongly connected digraph with equal in/out degree at every vertex has an Eulerian cycle.
    Invoked in the proof of Theorem 5 as the existence guarantee for the traversal that defines the reconstructed sequence.
  • domain assumption The marginal consistency constraints (equation 7) characterize feasible k-mer distributions of sequences.
    Necessity is argued from the linear overlap structure of DNA; sufficiency is claimed in Section 4 but only established approximately in Theorem 5; the paper cites [18, 24] for the polytope characterization.
  • ad hoc to paper The corner labelling convention (Definition 1) assigns A, C, G, T to specific square corners.
    The cell-to-k-mer correspondence (Propositions 1, 3, 4) and the symmetry-to-permutation map (Theorem 1) are stated relative to this specific labelling; a different labelling would permute which k-mer is associated with which cell.
  • ad hoc to paper The strong-connectivity repair in Theorem 5 adds at most 4^{k-1} artificial edges.
    Asserted without derivation in the proof of Theorem 5; it is needed for the Eulerian cycle guarantee and is not included in the n_art budget calculation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Chaos Game Representations and $k$-mer Frequencies of DNA Sequences." pith.science (2026). https://pith.science/paper/U7PPHML7

@misc{pith2026250622172,
  author       = {Pith},
  title        = {Pith review of: Bridging Chaos Game Representations and $k$-mer Frequencies of DNA Sequences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7PPHML7}},
  note         = {Machine review of arXiv:2506.22172}
}
abstract

This paper establishes formal mathematical foundations linking Chaos Game Representations (CGR) of DNA sequences to their underlying $k$-mer frequencies. We prove that the Frequency CGR (FCGR) of order $k$ is mathematically equivalent to a discretization of CGR at resolution $2^k \times 2^k$, and its vectorization corresponds to the $k$-mer frequencies of the sequence. Additionally, we characterize how symmetry transformations of CGR images correspond to specific nucleotide permutations in the originating sequences. Leveraging these insights, we introduce an algorithm that generates synthetic DNA sequences from prescribed $k$-mer distributions by constructing Eulerian paths on De Bruijn multigraphs. This enables reconstruction of sequences matching target $k$-mer profiles with arbitrarily high precision, facilitating the creation of synthetic CGR images for applications such as data augmentation for machine learning-based taxonomic classification of DNA sequences. Numerical experiments validate the effectiveness of our method across both real genomic data and artificially sampled distributions. To our knowledge, this is the first comprehensive framework that unifies CGR geometry, $k$-mer statistics, and sequence reconstruction, offering new tools for genomic analysis and visualization.

Figures

Figures reproduced from arXiv: 2506.22172 by the authors.

Figure 1
Figure 1. (a) CGR of a 100,000 bp DNA sequence randomly extracted from the complete genome of Pseudomonas aeruginosa strain PAO1 (RefSeq NC_002516.2); (b) CGR of a 100,000 bp DNA sequence randomly selected from human chromosome 4 (GRCh38.p14 primary assembly, RefSeq NC_000004.12); (c) CGR of a computer-generated DNA sequence that is random in all other aspects except that the dinucleotide GC is absent. Independent of CGR rese… view at source ↗
Figure 2
Figure 2. a illustrates c(ACG), the CGR cell associated with the 3-mer ACG. It is easy to see that, for a k-mer w, the last point plast(w) of CGR(w) is the center of the sub-square c(w). Note also that for two different k-mers w1 and w2, the cells c(w1) and c(w2) do not intersect. −1 −0.5 0.5 1 −1 −0.5 0.5 1 A C G T P PA PAC PACG (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. (a) Reconstruction of sequences from k-mer distributions of real DNA sequences. The reconstruction is computed across different values of k, and sequences spanning species from each of the six kingdoms of life. For each kingdom, 100 genomes were selected, and a single 100,000-long DNA fragment was randomly selected from each genome. The mean error was computed for each pair (k, kingdom). (b) Reconstruction of sequen… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Snapshot of the functionality of the computational tool: (a) Input: Dinucleotide (k = 2) distribution selected by a user via interactively adjusting 16 different sliders, each corresponding to one dinucleotide. (b) Output: CGR of a reconstructed DNA sequence whose k-me…
Figure 5
Figure 5. Figure 5: Visual correspondence between the eight elements of the permutation group S4 acting on {A, C, G, T} and the dihedral group D8. Panels (a) to (h) show Chaos Game Representations of a 100 kb randomly selected fragment from human chromosome 4 (GRCh38.p14 primary assembly,…
Figure 6
Figure 6. Figure 6: A visual comparison of FCGRs of order k = 7 according to different definitions. The intensity of the pixel represents the frequencies of 7-mers. Darker pixels mean larger frequencies. (a) CGR of a 100,000 bp DNA sequence randomly selected from human chromosome 4 (GRCh3…
Figure 7
Figure 7. Figure 7: Comparison of De Bruijn graphs and multigraph of a sample string s = ATCGTATCCA, for k = 3. Each vertex represents a unique 2-mer from the string, and each edge represents a unique 3-mer. (a) De Bruijn graph: A walk in the graph corresponds to a substring of s with eac…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

  1. [1]

    BMC Genomics 25, 1–17 (2024)

    Alipour, F., Hill, K.A., Kari, L.: CGRclust: Chaos game representation for twin contrastive clustering of unlabelled DNA sequences. BMC Genomics 25, 1–17 (2024)

    work page 2024

  2. [2]

    et al.: Accurate and fast clade assignment via deep learning and frequency chaos game representation

    Avila Cartes, J. et al.: Accurate and fast clade assignment via deep learning and frequency chaos game representation. GigaScience12, giac119 (2023)

    work page 2023

  3. [3]

    et al.: Spades: A new genome assembly algorithm and its applications to single-cell sequencing

    Bankevich, A. et al.: Spades: A new genome assembly algorithm and its applications to single-cell sequencing. J. Comput. Biol.19(5), 455–477 (2012)

    work page 2012

  4. [4]

    et al.: Genomic signature: characterization and classification of species assessed by chaos game representation of sequences

    Deschavanne, P.J. et al.: Genomic signature: characterization and classification of species assessed by chaos game representation of sequences. Molecular Biology and Evolution16(10), 1391–1399 (1999)

    work page 1999

  5. [5]

    Archaea2, 159–167 (2008)

    Dexter Dyer, B., Kahnand, M.J., LeBlanc, M.D.: Classification and regression tree (CART) analyses of genomic signatures reveal sets of tetramers that Bridging CGR andk-mer Frequencies of DNA 15 discriminate temperature optima of archaea and bacteria. Archaea2, 159–167 (2008)

    work page 2008

  6. [6]

    John Wiley & Sons, Inc., Hoboken, NJ, third edn

    Dummit, D.S., Foote, R.M.: Abstract algebra. John Wiley & Sons, Inc., Hoboken, NJ, third edn. (2004)

    work page 2004

  7. [7]

    Dutta, C., Das, J.: Mathematical characterization of chaos game representation: New algorithms for nucleotide sequence analysis. J. Molecular Biology 228(3), 715–719 (1992)

    work page 1992

  8. [8]

    et al.: Genomic signature in evolutionary biology: A review

    De la Fuente, R. et al.: Genomic signature in evolutionary biology: A review. Biology 12(2) (2023)

    work page 2023

Show all 45 references
  1. [9]

    Nucleic Acids Research21(10), 2487–2491 (1993)

    Goldman, N.: Nucleotide, dinucleotide and trinucleotide frequencies explain patterns observed in chaos game representations of DNA sequences. Nucleic Acids Research21(10), 2487–2491 (1993)

    1993

  2. [10]

    Physica A: Statistical Mechanics and its Applications282(1), 225–246 (2000)

    Hao, B.L.: Fractals from genomes – exact solutions of a biology-inspired problem. Physica A: Statistical Mechanics and its Applications282(1), 225–246 (2000)

    2000

  3. [11]

    Hill, K.A., Schisler, N.J., Singh, S.M.: Chaos game representation of coding regions of human globin genes and alcohol dehydrogenase genes of phylogenetically divergent species. J. Molecular Evolution35, 261–269 (1992)

    1992

  4. [12]

    Genomics 108(3–4), 134–142 (2016)

    Hoang, T., Yin, C., Yau, S.S.T.: Numerical encoding of DNA sequences by chaos game representation with application in similarity comparison. Genomics 108(3–4), 134–142 (2016)

    2016

  5. [13]

    Nucleic Acids Research 18(8), 2163–2170 (1990)

    Jeffrey, H.J.: Chaos game representation of gene structure. Nucleic Acids Research 18(8), 2163–2170 (1990)

    1990

  6. [14]

    Computers & Graphics 16(1), 25–33 (1992)

    Jeffrey, H.: Chaos game visualization of sequences. Computers & Graphics 16(1), 25–33 (1992)

    1992

  7. [15]

    BMC Bioinformatics7, 1–10 (2006)

    Joseph, J., Sasikumar, R.: Chaos game representation for comparison of whole genomes. BMC Bioinformatics7, 1–10 (2006)

    2006

  8. [16]

    et al.: An investigation into inter-and intragenomic variations of graphic genomic signatures

    Karamichalis, R. et al.: An investigation into inter-and intragenomic variations of graphic genomic signatures. BMC Bioinformatics16(1) (2015)

    2015

  9. [17]

    Trends in Genetics11(7), 283 – 290 (1995)

    Karlin, S., Burge, C.: Dinucleotide relative abundance extremes: a genomic signature. Trends in Genetics11(7), 283 – 290 (1995)

    1995

  10. [18]

    et al.: Unsupervised statistical clustering of environmental shotgun sequences

    Kislyuk, A. et al.: Unsupervised statistical clustering of environmental shotgun sequences. BMC Bioinformatics10(1), 316 (2009)

    2009

  11. [19]

    BMC Bioinformatics20, 1–17 (2019)

    Lichtblau, D.: Alignment-free genomic sequence comparison using FCGR and signal processing. BMC Bioinformatics20, 1–17 (2019)

    2019

  12. [20]

    Löchel, H.F., Heider, D.: Chaos game representation and its applications in bioinformatics. Comput. Struct. Biotechnol. J.19, 6263–6271 (2021)

    2021

  13. [21]

    Lovász, L., Vempala, S.: Hit-and-run from a corner. SIAM J. on Computing 35(4), 985–1005 (2006)

    2006

  14. [22]

    et al.: DeLUCS: Deep learning for unsupervised clustering of DNA sequences

    Millán Arias, P.A. et al.: DeLUCS: Deep learning for unsupervised clustering of DNA sequences. PLOS ONE17(1), e0261531 (2022)

    2022

  15. [23]

    Genomics95(6), 315–327 (2010)

    Miller, J.R., Koren, S., Sutton, G.: Assembly algorithms for next-generation sequencing data. Genomics95(6), 315–327 (2010)

    2010

  16. [24]

    et al.: Improved metagenome binning and assembly using deep variational autoencoders

    Nissen, J.N. et al.: Improved metagenome binning and assembly using deep variational autoencoders. Nature Biotechnology39(5), 555–560 (2021) 16 He, Kari, Millan Arias

    2021

  17. [25]

    et al.: Entropic profiles of DNA sequences through chaos-game- derived images

    Oliver, J. et al.: Entropic profiles of DNA sequences through chaos-game- derived images. J. Theoretical Biology160(4), 457–470 (1993)

    1993

  18. [26]

    PNAS98(17), 9748–9753 (2001)

    Pevzner, P.A., Tang, H., Waterman, M.S.: An Eulerian path approach to DNA fragment assembly. PNAS98(17), 9748–9753 (2001)

    2001

  19. [27]

    et al.: Machine learning using intrinsic genomic signatures for rapid classification of novel pathogens: COVID-19 case study

    Randhawa, G.S. et al.: Machine learning using intrinsic genomic signatures for rapid classification of novel pathogens: COVID-19 case study. PLOS ONE 15(4), e0232391 (2020)

    2020

  20. [28]

    et al.: Classification experiments of DNA sequences by using a deep neural network and chaos game representation

    Rizzo, R. et al.: Classification experiments of DNA sequences by using a deep neural network and chaos game representation. In: Proc. of the 17th Intl. Conference on Computer Systems and Technologies. pp. 222–228 (2016)

    2016

  21. [29]

    et al.: Similarity studies of corona viruses through chaos game representation

    Sengupta, D.C. et al.: Similarity studies of corona viruses through chaos game representation. Computational Molecular Bioscience10(3), 61 (2020)

    2020

  22. [30]

    et al.: An open-source k-mer based machine learning tool for fast and accurate sub-typing of HIV-1 genomes

    Solis-Reyes, S. et al.: An open-source k-mer based machine learning tool for fast and accurate sub-typing of HIV-1 genomes. PLOS ONE13(11), e0206409 (2018)

    2018

  23. [31]

    et al.: The spectrum of genomic signatures: from dinucleotides to chaos game representation

    Wang, Y. et al.: The spectrum of genomic signatures: from dinucleotides to chaos game representation. Gene346, 173–185 (2005)

    2005

  24. [32]

    et al.: Benchmarking of alignment-free sequence comparison methods

    Zielezinski, A. et al.: Benchmarking of alignment-free sequence comparison methods. Genome Biology20(1), 144 (2019)

    2019

  25. [33]

    et al.: Alignment-free sequence comparison: benefits, applications, and tools

    Zielezinski, A. et al.: Alignment-free sequence comparison: benefits, applications, and tools. Genome Biology18(1), 186 (2017) Appendix A: Technical Appendix A.1 Proofs in Section 2 Theorem 1. Let u, w∈ Σn be two DNA sequences of lengthn, and letσ ∈ S be one of the morphisms i...

    2017

  26. [34]

    =⇒" a = σ(b) = σ(G) = G implies label(a) = label(G) = (1, 1) and f −1(σ) =e. Thus, f −1(σ) · label(b) =e · label(G) =label(G) =label(a). “⇐=

    σ = (), the identity permutation onΣ. “=⇒" a = σ(b) = σ(G) = G implies label(a) = label(G) = (1, 1) and f −1(σ) =e. Thus, f −1(σ) · label(b) =e · label(G) =label(G) =label(a). “⇐=" f −1(σ) = 1 0 0 1 , which implies label(a) = f −1(σ) · label(b) = (1, 1), further implying thata...

  27. [35]

    =⇒" a = σ(G) =T implies label(a) =label(T ) = (1, −1) and f −1(σ) =r3. Thus, f −1(σ) · label(b) =r3 · (1, 1) = 0 1 −1 0 · (1, 1) = (1, −1) =label(a). “⇐=

    σ = (A C G T). “=⇒" a = σ(G) =T implies label(a) =label(T ) = (1, −1) and f −1(σ) =r3. Thus, f −1(σ) · label(b) =r3 · (1, 1) = 0 1 −1 0 · (1, 1) = (1, −1) =label(a). “⇐=" f −1(σ) = 0 1 −1 0 , which implieslabel(a) =f −1(σ) · label(b) = (1, −1), which further implies thata = T = σ(b)

  28. [36]

    =⇒" a = σ(G) =A implies label(a) =label(A) = (−1, −1) and f −1(σ) =r2. Thus, f −1(σ) · label(b) =r2 · (1, 1) = −1 0 0 −1 · (1, 1) = (−1, −1) =label(a). “⇐=

    σ = (A G)(C T). “=⇒" a = σ(G) =A implies label(a) =label(A) = (−1, −1) and f −1(σ) =r2. Thus, f −1(σ) · label(b) =r2 · (1, 1) = −1 0 0 −1 · (1, 1) = (−1, −1) =label(a). “⇐=" f −1(σ) = −1 0 0 −1 implies label(a) = f −1(σ) · label(b) = (−1, −1), which further implies thata = A = σ(b)

  29. [37]

    =⇒" a = σ(G) =C implies label(a) =label(C) = (−1, 1) and f −1(σ) =r. Thus, f −1(σ) · label(b) =r · (1, 1) = 0 −1 1 0 · (1, 1) = (−1, 1) =label(a). “⇐=

    σ = (A T G C). “=⇒" a = σ(G) =C implies label(a) =label(C) = (−1, 1) and f −1(σ) =r. Thus, f −1(σ) · label(b) =r · (1, 1) = 0 −1 1 0 · (1, 1) = (−1, 1) =label(a). “⇐=" f −1(σ) = 0 −1 1 0 implies l(a) = f −1(σ) · label(b) = (−1, 1), which further implies thata = C = σ(b)

  30. [38]

    =⇒" a = σ(G) =T implies label(a) =label(T ) = (1, −1) and f −1(σ) =s. Thus, f −1(σ) · label(b) =s · (1, 1) = 1 0 0 −1 · (1, 1) = (1, −1) =label(a). “⇐=

    σ = (A C)(G T). “=⇒" a = σ(G) =T implies label(a) =label(T ) = (1, −1) and f −1(σ) =s. Thus, f −1(σ) · label(b) =s · (1, 1) = 1 0 0 −1 · (1, 1) = (1, −1) =label(a). “⇐=" f −1(σ) = 1 0 0 −1 implies label(a) =f −1(σ) · label(b) = (1, −1), which further implies thata = T = σ(b). ...

  31. [39]

    =⇒" a = σ(G) =C implies label(a) =label(C) = (−1, 1) and f −1(σ) =sr2. Thus, f −1(σ) · label(b) =sr2 · (1, 1) = −1 0 0 1 · (1, 1) = (−1, 1) =label(a). “⇐=

    σ = (A T)(C G). “=⇒" a = σ(G) =C implies label(a) =label(C) = (−1, 1) and f −1(σ) =sr2. Thus, f −1(σ) · label(b) =sr2 · (1, 1) = −1 0 0 1 · (1, 1) = (−1, 1) =label(a). “⇐=" f −1(σ) = −1 0 0 1 implies label(a) =f −1(σ) · label(b) = (−1, 1), which further implies thata = C = σ(b)

  32. [40]

    =⇒" a = σ(G) =G implies label(a) =label(G) = (1, 1) and f −1(σ) =sr3. Thus, f −1(σ) · label(b) =sr3 · (1, 1) = 0 1 1 0 · (1, 1) = (1, 1) =label(a). “⇐=

    σ = (C T). “=⇒" a = σ(G) =G implies label(a) =label(G) = (1, 1) and f −1(σ) =sr3. Thus, f −1(σ) · label(b) =sr3 · (1, 1) = 0 1 1 0 · (1, 1) = (1, 1) =label(a). “⇐=" f −1(σ) = 0 1 1 0 implies label(a) = f −1(σ) · label(b) = (1, 1), which further implies thata = G = σ(b)

  33. [41]

    =⇒" a = σ(G) =A implies label(a) =label(A) = (−1, −1) and f −1(σ) =sr. Thus, f −1(σ) · label(b) =sr · (1, 1) = 0 −1 −1 0 (1, 1) = (−1, −1) =label(a). “⇐=

    σ = (A G). “=⇒" a = σ(G) =A implies label(a) =label(A) = (−1, −1) and f −1(σ) =sr. Thus, f −1(σ) · label(b) =sr · (1, 1) = 0 −1 −1 0 (1, 1) = (−1, −1) =label(a). “⇐=" f −1(σ) = 0 −1 −1 0 implies label(a) = f −1(σ) · label(b) = (−1, −1), which further implies thata = A = σ(b). ...

  34. [42]

    via SVD)

    Compute an orthonormal basisN ∈ R4k×(4k−r) of ker(B) (e.g. via SVD)

  35. [43]

    Initialize θ(0) = (1/4k, . . . ,1/4k)

  36. [44]

    , T: (a) Draw z ∼ N(0, I4k−r) and set d = N z/∥N z∥

    For each iterationt = 1, . . . , T: (a) Draw z ∼ N(0, I4k−r) and set d = N z/∥N z∥. (b) Find tmin ≤ tmax so that θ(t−1) + t d ≥ 0 coordinate-wise. (c) Sample t ∼ Unif[tmin, tmax]. (d) Update θ(t) = θ(t−1) + t d

  37. [45]

    Return θ(T ), which lies in the simplex and satisfiesBθ = b. At each step, the directiond is drawn from a continuous distribution spanning all of ker(B), so any two interior points can be connected by a finite sequence of positive-probability moves. This procedure corresponds ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.