pith. sign in

arxiv: 2606.11646 · v1 · pith:K5V3PEUYnew · submitted 2026-06-10 · 💻 cs.LG · q-bio.QM· stat.ML

Tree-Structured Orthonormal Decomposition of the Aitchison Simplex

Daisuke Yamada , Qijun Zhang , Travis Pence , Barbara B. Bendlin , Federico Rey , Vikas Singh This is my paper

Pith reviewed 2026-06-27 11:05 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QMstat.ML
keywords compositional dataAitchison geometrytree topologyorthonormal decompositionmicrobiome analysissingle-cell datasoftmax connection
0
0 comments X

The pith

PolyILR produces a canonical orthonormal decomposition of the Aitchison tangent space aligned with any tree topology.

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

The paper presents PolyILR as a method that decomposes compositional data vectors, which represent relative proportions, into a set of coordinates that follow a given tree hierarchy such as taxonomies or phylogenies. It builds this by first defining a weighted local geometry at each internal node to capture the full branching, then combining those into one global orthonormal basis under the Aitchison inner product. Every resulting coordinate maps directly to a specific location in the tree. A reader would care because this handles both the relative nature of the data and the hierarchical structure at once, enabling stable features and multiscale analysis on data like microbiomes or single-cell counts.

Core claim

PolyILR yields a canonical orthonormal decomposition of the Aitchison tangent space aligned with any tree topology, producing stable interpretable features that enable inference at multiscale tree resolution. The construction defines a weighted local geometry at each internal node capturing full branching structure, then lifts these to a global orthonormal basis where every coordinate corresponds to a specific tree location. On microbiome and single-cell benchmarks, PolyILR yields stable, interpretable features and enables inference at multiscale tree resolution. It also establishes a novel theoretical connection to softmax classifiers.

What carries the argument

The PolyILR construction, which defines a weighted local geometry at each internal node to capture full branching structure and lifts it to a global orthonormal basis preserving the Aitchison inner product.

If this is right

  • Every coordinate in the resulting basis corresponds to a specific location on the input tree.
  • The features remain stable across different tree resolutions for microbiome and single-cell data.
  • Inference becomes possible at multiple scales of the same tree without discarding geometry or structure.
  • A direct theoretical link appears between the decomposition and softmax classifiers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same lifting procedure might extend to other hierarchical structures such as ontologies or phylogenies outside the tested domains.
  • The connection to softmax suggests the coordinates could serve as a drop-in replacement for standard inputs in probabilistic models that already use trees.
  • Because the basis is canonical for any tree, it could support direct comparison of models trained on different but related hierarchies.

Load-bearing premise

A weighted local geometry defined at each internal node can be lifted to a single global orthonormal basis while preserving the Aitchison inner product and capturing the full branching structure for arbitrary trees.

What would settle it

An explicit counterexample tree (binary or non-binary) for which the lifted coordinates fail to remain orthonormal under the Aitchison inner product or fail to span the full tangent space.

Figures

Figures reproduced from arXiv: 2606.11646 by Barbara B. Bendlin, Daisuke Yamada, Federico Rey, Qijun Zhang, Travis Pence, Vikas Singh.

Figure 1
Figure 1. Figure 1: Polytomous vs. Binarized Tree. (Left) T with a poly￾tomous root (k = 3). PolyILR assigns k − 1 = 2 basis vectors (column of V ) to the root (red). (Right) Binarized Tb required by PhILR (one coordinate per node) introduces an artificial node (yel￾low) encoding an arbitrary grouping of leaves 1 and 2—a choice not justified by original T . The absence of such a coordinate system extends beyond traditional Co… view at source ↗
Figure 2
Figure 2. Figure 2: PolyILR from T on d = 6 leaves to V ∈ R d×(d−1) . Colors indicate generating nodes; white entries are zeros. wise log-ratio testing (Mandal et al., 2015), or phylogeny￾guided ILR factors via edge selection (Washburne et al., 2017). But these methods yield isolated contrasts rather than a full, node-grouped orthonormal coordinate system canonically tied to a given multifurcating tree. No existing method pro… view at source ↗
Figure 3
Figure 3. Figure 3: Naive vs. weighted spreading (Example 4.1). Uniform spreading (Left) vs. Weighted spreading (Right) of children 1, . . . , m. For example, with ku = 3 children: H(u) =   √ 1 2 √ 1 6 − √ 1 2 √ 1 6 0 − √ 2 6   . The first column contrasts child 2 versus child 1, the second contrasts child 3 versus the average of children 1 and 2. Helmert contrasts provide a canonical choice: given an ordering of childr… view at source ↗
Figure 4
Figure 4. Figure 4: Projection onto top-2 PolyILR features for HMP body site classification (left) and DISCO leukemia classification (right). Axes are the two most predictive contrasts ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Orthogonal subspace partitions. The coordinates of V can be grouped into orthogonal subspaces indexed by tree structure: by individual node, by depth level, or by subtree (one example). Each grouping enables inference at a different resolution. metadata were extracted from TreeSummarizedExperiment objects and merged using mergeData. Only samples present in both the abundance matrix and metadata were retain… view at source ↗
Figure 6
Figure 6. Figure 6: Learning dynamics on CIFAR-100. Left: Gradient norm distribution by tree depth across training. Depth 0 (superclass contrasts) decreases while depth 1 (within-superclass) increases, indicating coarse-to-fine learning. Right: Gradient entropy and test accuracy over training epochs. B.7. Full Experimental Results This appendix provides complete experimental results; biological interpretation is in Section 6.… view at source ↗
read the original abstract

Compositional data -- vectors encoding relative proportions -- arise across scientific domains, including ecology, geochemistry, and genomics. The features in these data often come with known hierarchical structure (e.g., taxonomies, phylogenies, ontologies), yet existing methods either ignore this structure, discard the intrinsic Aitchison geometry, are designed for binary trees, or yield incomplete coordinate systems. We describe PolyILR, a canonical orthonormal decomposition of the Aitchison tangent space aligned with any tree topology. Our construction defines a weighted local geometry at each internal node capturing full branching structure, then lifts these to a global orthonormal basis where every coordinate corresponds to a specific tree location. On microbiome and single-cell benchmarks, PolyILR yields stable, interpretable features and enables inference at multiscale tree resolution. We also establish a novel theoretical connection to softmax classifiers, suggesting possible applications to probabilistic modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces PolyILR as a canonical orthonormal decomposition of the Aitchison tangent space aligned with arbitrary tree topologies for compositional data. It defines a weighted local geometry at each internal node that captures the full branching structure and lifts these local bases to a single global orthonormal basis in which each coordinate corresponds to a specific tree location. The work reports stable interpretable features on microbiome and single-cell benchmarks, multiscale inference capability, and a theoretical link to softmax classifiers.

Significance. If the lifting step is shown to preserve the Aitchison inner product and produce a complete basis for any tree (including nodes with branching factor >2), the result would supply a principled, tree-aligned coordinate system for compositional data that respects the intrinsic geometry while enabling hierarchical analysis; this would be a useful contribution to compositional data methods in ecology, genomics, and related fields.

major comments (1)
  1. [Main construction (lifting step)] The central construction asserts that weighted local geometries defined at internal nodes can be lifted to a global orthonormal basis of the (D-1)-dimensional Aitchison tangent space while preserving the inner product for arbitrary (including non-binary) trees. The manuscript must supply an explicit verification or inductive argument showing that the lifted vectors remain orthogonal and span the full space when an internal node has three or more children; without this, the claim that the decomposition is canonical and complete for general tree topologies is not secured.
minor comments (2)
  1. [Abstract] The abstract states existence of the construction and benchmark outcomes but supplies no derivation outline, error analysis, or data-exclusion criteria; the full manuscript should include a concise proof sketch or algorithmic pseudocode for the lifting procedure.
  2. [Methods] Notation for the weighted local inner product at each node and the precise definition of the lifting map should be introduced with explicit equations rather than descriptive text only.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying a point where the theoretical justification can be strengthened. We agree that an explicit inductive argument for the lifting step on non-binary trees will improve the manuscript and will add it in revision.

read point-by-point responses

  1. Referee: [Main construction (lifting step)] The central construction asserts that weighted local geometries defined at internal nodes can be lifted to a global orthonormal basis of the (D-1)-dimensional Aitchison tangent space while preserving the inner product for arbitrary (including non-binary) trees. The manuscript must supply an explicit verification or inductive argument showing that the lifted vectors remain orthogonal and span the full space when an internal node has three or more children; without this, the claim that the decomposition is canonical and complete for general tree topologies is not secured.

    Authors: We acknowledge that while the manuscript states the construction holds for arbitrary trees and provides the local weighted geometry and lifting procedure, an explicit inductive verification of orthogonality and completeness for nodes with branching factor greater than two is not supplied in detail. In the revised manuscript we will insert a new subsection (or appendix) containing a short inductive argument: the base case for binary nodes follows directly from the local orthonormal construction; the inductive step for a node with k>2 children shows that the weighted local basis vectors remain mutually orthogonal under the Aitchison inner product, that their lifts are orthogonal to all previously lifted vectors from other subtrees, and that the resulting set spans the full (D-1)-dimensional tangent space. This addition will make the canonicity claim fully rigorous without altering any other results or claims. revision: yes

Circularity Check

0 steps flagged

No circularity: construction presented as novel without reduction to inputs or self-citations

full rationale

The provided abstract and description outline a construction that defines weighted local geometry at internal nodes and lifts it to a global orthonormal basis of the Aitchison tangent space. No equations, fitted parameters, or self-citations are shown that would make any claimed prediction or uniqueness equivalent to the inputs by definition. The central claim of a canonical decomposition aligned with arbitrary tree topologies is presented as a new method without load-bearing reliance on prior author work or renaming of known results. This is the typical case of a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed from abstract alone; no free parameters, axioms, or invented entities are described in the provided text.

pith-pipeline@v0.9.1-grok · 5698 in / 970 out tokens · 25429 ms · 2026-06-27T11:05:02.168747+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 2 canonical work pages

  1. [1]

    Mathematical Geology , volume=

    Isometric logratio transformations for compositional data analysis , author=. Mathematical Geology , volume=. 2003 , publisher=

    2003

  2. [2]

    Journal of Scientific Computing , volume=

    Haar-Like Wavelets on Hierarchical Trees , author=. Journal of Scientific Computing , volume=. 2024 , publisher=

    2024

  3. [3]

    Austrian Journal of Statistics , volume=

    Changing the reference measure in the simplex and its weighting effects , author=. Austrian Journal of Statistics , volume=

  4. [4]

    elife , volume=

    A phylogenetic transform enhances analysis of compositional microbiota data , author=. elife , volume=. 2017 , publisher=

    2017

  5. [5]

    PeerJ , volume=

    Phylogenetic factorization of compositional data yields lineage-level associations in microbiome datasets , author=. PeerJ , volume=. 2017 , publisher=

    2017

  6. [6]

    Mathematical Geology , volume=

    Groups of parts and their balances in compositional data analysis , author=. Mathematical Geology , volume=. 2005 , publisher=

    2005

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Fisher flow matching for generative modeling over discrete data , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    arXiv preprint arXiv:2510.27480 , year=

    Simplex-to-Euclidean Bijections for Categorical Flow Matching , author=. arXiv preprint arXiv:2510.27480 , year=

    work page arXiv

  9. [9]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    The statistical analysis of compositional data , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1982 , publisher=

    1982

  10. [10]

    Stochastic Environmental Research and Risk Assessment , volume=

    Geometric approach to statistical analysis on the simplex , author=. Stochastic Environmental Research and Risk Assessment , volume=. 2001 , publisher=

    2001

  11. [11]

    Journal of the American Statistical Association , volume=

    Statistical interpretation of species composition , author=. Journal of the American Statistical Association , volume=. 2001 , publisher=

    2001

  12. [12]

    2018 , publisher=

    Compositional data analysis in practice , author=. 2018 , publisher=

    2018

  13. [13]

    Frontiers in Microbiology , volume=

    Microbiome datasets are compositional: and this is not optional , author=. Frontiers in Microbiology , volume=. 2017 , publisher=

    2017

  14. [14]

    Ecology , volume=

    Compositional data in community ecology: the paradigm or peril of proportions? , author=. Ecology , volume=. 1997 , publisher=

    1997

  15. [15]

    Earth and Planetary Science Letters , volume=

    Calibration of XRF core scanners for quantitative geochemical logging of sediment cores: Theory and application , author=. Earth and Planetary Science Letters , volume=. 2008 , publisher=

    2008

  16. [16]

    2019 , publisher=

    Phylogenetic comparative methods: learning from trees , author=. 2019 , publisher=

    2019

  17. [17]

    The ISME Journal , volume=

    Fast UniFrac: facilitating high-throughput phylogenetic analyses of microbial communities including analysis of pyrosequencing and PhyloChip data , author=. The ISME Journal , volume=. 2010 , publisher=

    2010

  18. [18]

    Lancaster, HO , journal=. The. 1965 , publisher=

    1965

  19. [19]

    Biometrika , volume=

    Variable selection in regression with compositional covariates , author=. Biometrika , volume=. 2014 , publisher=

    2014

  20. [20]

    Microbial ecology in health and disease , volume=

    Analysis of composition of microbiomes: a novel method for studying microbial composition , author=. Microbial ecology in health and disease , volume=. 2015 , publisher=

    2015

  21. [21]

    MSystems , volume=

    Balances: a new perspective for microbiome analysis , author=. MSystems , volume=. 2018 , publisher=

    2018

  22. [22]

    MSystems , volume=

    Balance trees reveal microbial niche differentiation , author=. MSystems , volume=. 2017 , publisher=

    2017

  23. [23]

    Ecological Monographs , volume=

    Phylofactorization: a graph partitioning algorithm to identify phylogenetic scales of ecological data , author=. Ecological Monographs , volume=. 2019 , publisher=

    2019

  24. [24]

    2021 , publisher=

    Buettner, Maren and Ostner, Johannes and Mueller, Christian L and Theis, Fabian J and Schubert, Benjamin , journal=. 2021 , publisher=

    2021

  25. [25]

    Brigham, E Oran , year=. The

  26. [26]

    2002 , publisher=

    A theory for multiresolution signal decomposition: the wavelet representation , author=. 2002 , publisher=

    2002

  27. [27]

    Nature Methods , volume=

    Accessible, curated metagenomic data through ExperimentHub , author=. Nature Methods , volume=. 2017 , publisher=

    2017

  28. [28]

    Nature , volume=

    Structure, function and diversity of the healthy human microbiome , author=. Nature , volume=. 2012 , publisher=

    2012

  29. [29]

    BMC systems Biology , volume=

    Polytomy identification in microbial phylogenetic reconstruction , author=. BMC systems Biology , volume=. 2011 , publisher=

    2011

  30. [30]

    Journal of Bacteriology , volume=

    The human oral microbiome , author=. Journal of Bacteriology , volume=. 2010 , publisher=

    2010

  31. [31]

    Proceedings of the National Academy of Sciences , volume=

    Impact of diet in shaping gut microbiota revealed by a comparative study in children from Europe and rural Africa , author=. Proceedings of the National Academy of Sciences , volume=. 2010 , publisher=

    2010

  32. [32]

    Nature , volume=

    Human gut microbiome viewed across age and geography , author=. Nature , volume=. 2012 , publisher=

    2012

  33. [33]

    NPJ Biofilms and Microbiomes , volume=

    Integrated metagenomics identifies a crucial role for trimethylamine-producing Lachnoclostridium in promoting atherosclerosis , author=. NPJ Biofilms and Microbiomes , volume=. 2022 , publisher=

    2022

  34. [34]

    Gut , volume=

    A novel faecal Lachnoclostridium marker for the non-invasive diagnosis of colorectal adenoma and cancer , author=. Gut , volume=. 2020 , publisher=

    2020

  35. [35]

    Science , volume=

    Bacterial community variation in human body habitats across space and time , author=. Science , volume=. 2009 , publisher=

    2009

  36. [36]

    core microbiome

    Defining the healthy" core microbiome" of oral microbial communities , author=. BMC microbiology , volume=. 2009 , publisher=

    2009

  37. [37]

    Nature , volume=

    Enterotypes of the human gut microbiome , author=. Nature , volume=. 2011 , publisher=

    2011

  38. [38]

    Forty-second International Conference on Machine Learning , year=

    From Logits to Hierarchies: Hierarchical Clustering made Simple , author=. Forty-second International Conference on Machine Learning , year=

  39. [39]

    Applied and environmental microbiology , volume=

    UniFrac: a new phylogenetic method for comparing microbial communities , author=. Applied and environmental microbiology , volume=. 2005 , publisher=

    2005

  40. [40]

    Mao, Jialiang and Ma, Li , journal=

  41. [41]

    Nature Genetics , volume=

    Gene ontology: tool for the unification of biology , author=. Nature Genetics , volume=. 2000 , publisher=

    2000

  42. [42]

    Nature Machine Intelligence , volume=

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , author=. Nature Machine Intelligence , volume=. 2019 , publisher=

    2019

  43. [43]

    Frontiers in Microbiology , volume=

    Applications of machine learning in human microbiome studies: a review on feature selection, biomarker identification, disease prediction and treatment , author=. Frontiers in Microbiology , volume=. 2021 , publisher=

    2021

  44. [44]

    2009 , organization=

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , booktitle=. 2009 , organization=

    2009

  45. [45]

    1995 , publisher=

    Miller, George A , journal=. 1995 , publisher=

    1995

  46. [46]

    2009 , publisher=

    Learning multiple layers of features from tiny images , author=. 2009 , publisher=

    2009

  47. [47]

    Frontiers in microbiology , volume=

    Analysis of microbiome data in the presence of excess zeros , author=. Frontiers in microbiology , volume=. 2017 , publisher=

    2017

  48. [48]

    2022 , publisher=

    Hu, Yingtian and Satten, Glen A and Hu, Yi-Juan , journal=. 2022 , publisher=

    2022

  49. [49]

    2022 , publisher=

    Li, Mengwei and Zhang, Xiaomeng and Ang, Kok Siong and Ling, Jingjing and Sethi, Raman and Lee, Nicole Yee Shin and Ginhoux, Florent and Chen, Jinmiao , journal=. 2022 , publisher=

    2022

  50. [50]

    New England Journal of Medicine , volume=

    Acute myeloid leukemia , author=. New England Journal of Medicine , volume=. 1999 , publisher=

    1999

  51. [51]

    Comprehensive Physiology , volume=

    Liver sinusoidal endothelial cells , author=. Comprehensive Physiology , volume=. 2015 , publisher=

    2015

  52. [52]

    2022 , publisher=

    Phipson, Belinda and Sim, Choon Boon and Porrello, Enzo R and Hewitt, Alex W and Powell, Joseph and Oshlack, Alicia , journal=. 2022 , publisher=

    2022

  53. [53]

    The 28th International Conference on Artificial Intelligence and Statistics , year=

    Hypernym Bias: Unraveling Deep Classifier Training Dynamics through the Lens of Class Hierarchy , author=. The 28th International Conference on Artificial Intelligence and Statistics , year=

  54. [54]

    Nickel, Maximillian and Kiela, Douwe , journal=. Poincar

  55. [55]

    Advances in neural information processing systems , volume=

    Hyperbolic graph convolutional neural networks , author=. Advances in neural information processing systems , volume=

  56. [56]

    Advances in Applied Mathematics , volume=

    Geometry of the space of phylogenetic trees , author=. Advances in Applied Mathematics , volume=. 2001 , publisher=

    2001

  57. [57]

    IEEE/ACM Transactions on Computational Biology and Bioinformatics , volume=

    A fast algorithm for computing geodesic distances in tree space , author=. IEEE/ACM Transactions on Computational Biology and Bioinformatics , volume=. 2010 , publisher=

    2010

  58. [58]

    arXiv preprint arXiv:1805.12400 , year=

    Tropical geometry of phylogenetic tree space: a statistical perspective , author=. arXiv preprint arXiv:1805.12400 , year=

    work page arXiv

  59. [59]

    Data mining and knowledge discovery , volume=

    A survey of hierarchical classification across different application domains , author=. Data mining and knowledge discovery , volume=. 2011 , publisher=

    2011

  60. [60]

    International Conference on Machine Learning , pages=

    Representation tradeoffs for hyperbolic embeddings , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  61. [61]

    Signal Transduction and Targeted Therapy , volume=

    A systematic framework for understanding the microbiome in human health and disease: from basic principles to clinical translation , author=. Signal Transduction and Targeted Therapy , volume=. 2024 , publisher=

    2024

  62. [62]

    Cladistics , volume=

    Reconstructing character evolution on polytomous cladograms , author=. Cladistics , volume=. 1989 , publisher=

    1989

  63. [63]

    Diehl, Alexander D and Meehan, Terrence F and Bradford, Yvonne M and Brush, Matthew H and Dahdul, Wasila M and Dougall, David S and He, Yongqun and Osumi-Sutherland, David and Ruttenberg, Alan and Sarntivijai, Sirarat and others , journal=. The. 2016 , publisher=

    2016

  64. [64]

    Nature Communications , volume=

    Large-scale genome-wide analysis links lactic acid bacteria from food with the gut microbiome , author=. Nature Communications , volume=. 2020 , publisher=

    2020

  65. [65]

    Microbiome , volume=

    Proportion-based normalizations outperform compositional data transformations in machine learning applications , author=. Microbiome , volume=. 2024 , publisher=

    2024