Pith. sign in

REVIEW 3 major objections 3 minor 24 references

Geometry of the space of phylogenetic trees with non-identical leaves

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper introduces Towering space, a complete metric space that extends BHV tree space to phylogenetic trees with different leaf sets.

desk verdict The Towering space construction is a real new idea with mostly credible proofs, but the paper's headline claim—completeness of the metric—is never proved, and the algorithm's guarantee leans on that missing proof; conditionally worth serious refereeing. read the letter →

arxiv 2508.06747 v1 pith:ZIK2WIIN submitted 2025-08-08 q-bio.PE math.MGq-bio.QM

Maria Alejandra Valdez Cabrera , Amy D Willis This is my paper
classification q-bio.PEmath.MGq-bio.QM MSC 05C0554E3592D15
keywords phylogenetictreesBHVspacemetricleafpruningregraftinggenegainandlossquotientgeodesics
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

Phylogenetic tree comparisons are usually forced to discard any gene tree whose leaf set differs from the rest. This paper introduces Towering space, a metric space that extends the Billera–Holmes–Vogtmann (BHV) tree space—so far defined only for trees with identical leaf sets—to trees with different leaf sets. The construction collects all BHV spaces on subsets of a fixed leaf set, declares pruning and regrafting of leaves zero-cost moves, and takes the Towering distance to be the shortest total BHV length of a path under those moves. The paper's central claim is that this quotient distance is a true metric: distinct trees with strictly positive external edges are separated, and trees at distance zero coincide after finitely many leaf operations. It also characterizes geodesics and gives a first, admittedly inefficient, algorithm for computing the distance, with the stated aim of enabling BHV-style statistical analyses on gene trees with gain and loss.

What carries the argument

The load-bearing mechanism is the leaf-pruning map $\psi(T,M)$: a set $M$ of leaves can be pruned from $T$ only when its external edges have length zero and every internal edge maps to an internal edge of the smaller tree; merged edges receive the $\ell^2$-norm of the merged lengths. Regrafting produces sprouting spaces $\Lambda_L(T)$, the trees in a higher BHV space that prune down to $T$. The equivalence classes formed by these operations, combined with the BHV path-space support-pair distance, produce the quotient metric that defines Towering space.

What would settle it

Run the Section 4.4 iterative algorithm on a 5-leaf tree and a 4-leaf tree while exhaustively enumerating all partitions of the shared leaf set; if any enumeration yields a candidate path shorter than the algorithm's converged value, the claimed guarantee of finding a geodesic fails. A more direct check: attempt to build a Cauchy sequence of trees in Towering space whose equivalence classes converge to an object that is not a tree with $|L|\ge 3$ leaves; any such sequence would falsify completeness.

Watch

Extended reading notes

Core claim

Towering space is the union $\bigcup_{L\subseteq N, |L|\ge 3} \mathcal T_L$ of BHV spaces over all leaf subsets, with an equivalence relation $\simeq$ generated by leaf prunings and regraftings. The Towering distance is the quotient pseudometric of the BHV distance, i.e. the infimum of sums $\sum_i d_{\mathrm{BHV}}(t_i,t'_i)$ over finite sequences with $T_1\simeq t_1$, $t'_i\simeq t_{i+1}$, $t'_k\simeq T_2$. Theorem 5 states that if $T_1$ has all external edges of strictly positive length and $d(T_1,T_2)=0$, then $T_1\simeq T_2$; this is what upgrades the pseudometric to a metric on equivalence classes. Geodesics are characterized as constant-rate morphs contained in the uppermost BHV level:

Load-bearing premise

The whole construction treats pruning and regrafting as zero-cost moves and then minimizes over finite sequences, while completeness of the resulting Towering space—needed to guarantee that the algorithm's search over partitions finds a geodesic—is asserted but never proved.

Editorial extensions

If this is right

  • Comparisons of gene trees with different sampled gene sets become possible with a true metric, so methods that rely on the triangle inequality (Fr\'echet means, principal paths, clustering, confidence sets) can in principle be built on Towering space.
  • A tree with all external edges positive is at distance zero only from trees in its own pruning/regrafting equivalence class, so the metric does not collapse distinct trees.
  • Geodesics can be lifted to the highest shared BHV space, meaning the extra leaves that appear in one tree but not the other do not artificially shorten distances: adding arbitrary leaves never improves the BHV distance between two trees (Theorem 1).
  • Each geodesic has a biological reading as a sequence of gene-loss (pruning) and gene-gain (regrafting) events combined with continuous branch-length change, and the $\ell^2$ merging rule preserves each tree's distance from the origin in its own BHV space.
  • The paper's algorithm, while not efficient enough for large datasets, is the first explicit scheme to compute Towering distances and is argued to be faster than gradient-based alternatives because each update is closed-form.

Reading between the lines

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

  • The text announces completeness of Towering space but gives no proof; the algorithm's guarantee in Section 4.4 ('By the completeness of Towering space...') therefore rests on an unproved premise, and the exhaustive search over partitions does not by itself rule out convergence to a local minimum.
  • A natural next step, which the paper identifies but does not take, is a fast classification of edges into 'prune-preventing' and 'common' groups in the spirit of the Owen–Provan support-pair algorithm; that would turn the preliminary algorithm into a practical one.
  • If the paper's conjecture that geodesics are unique for almost all tree pairs holds, then Fr\'echet means and principal-component-style analyses could be developed on Towering space without the complications of non-unique geodesics.
  • The $\ell^2$ merging rule gives a testable conservation statement: pruning a set of leaves does not change the distance from the tree to the origin, so a gene-loss event alone contributes zero change in total branch-length divergence—an interpretable and checkable property for empirical gene-tree datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper introduces Towering space, a metric space on the union of BHV tree spaces for all leaf subsets L⊂N with |L|≥3. Trees with different leaf sets are connected by zero-cost pruning/regrafting operations; the distance is defined as the quotient pseudometric (Eq. 2) of the BHV distance, with an equivalence relation generated by these operations. The authors prove several structural inequalities: Theorem 1 (monotonicity under pruning), Theorem 2 (optimal single-prune path formula), Theorem 3 (shortest path through a common lower level), Lemma 5 (lifting geodesic segments), and Theorem 5 (separation of trees with positive external edge lengths). They also propose an iterative algorithm for computing the distance in Section 4.4. The abstract and discussion claim that Towering space is a complete metric space extending BHV space and that the algorithm provably finds a geodesic; however, completeness and global convergence of the algorithm are asserted but never proved.

Significance. If the completeness and geodesic-characterization claims are substantiated, the paper fills an important gap: a true metric on trees with non-identical leaf sets that accounts for both topology and branch lengths. The construction is parameter-free, and the lower-bound arguments via the Pythagorean inequality are coherent. The potential biological interpretation (pruning = gene loss, regrafting = HGT) is appealing. However, the manuscript's central advertised properties—completeness and algorithmic guarantee—currently lack proof; the paper is a promising framework rather than a finished complete-metric-space construction.

major comments (3)
  1. [Abstract, Definition 10, Section 4.4, Section 5] Completeness of (T P(N), d) is asserted but never proved. Theorem 5 only shows that for T1 with strictly positive external edges, d(T1,T2)=0 implies T1≃T2; it does not show every Cauchy sequence converges or that the infimum in Eq. (2) is attained. This is load-bearing: Section 4.4 states 'By the completeness of Towering space, Theorem 3 and Theorem 4, this procedure is guaranteed to find at least one geodesic.' Without completeness, there is no guarantee of geodesic existence or global convergence. A proof of completeness of the quotient pseudometric must be supplied, or the claims in the abstract and discussion must be weakened.
  2. [Theorem 4 (Section 4.3)] The geodesic characterization in Theorem 4 is the basis for the algorithm, but its proof in the supplement is a single sentence: 'Any path constructed using Theorem 3 and Lemmas 3 and 4 consists of segments fully contained in some BHV level... By Lemma 5, each of these segments can be lifted to an equal length segment in TN.' This does not prove that an arbitrary geodesic must have the stated alternating form with all positive-length segments in TN, nor that the independent lifts of consecutive segments can be chosen so that they meet at the required transition trees. The compatibility of the lifts is exactly what must be established. Please provide a complete proof or explicitly state Theorem 4 as a partial characterization.
  3. [Section 4.4 (Algorithm)] The iterative procedure (forward/backward passes, repeat until no improvement) is claimed to be 'guaranteed to find at least one geodesic' by completeness, Theorem 3, and Theorem 4. This is unjustified. Alternating minimization over a finite set of partitions and initial topologies may converge to a stationary point that is only locally optimal; no monotone-convergence or global-optimality proof is given. The enumeration over all partitions and all topologies in ΛN(T1), ΛN(T2) does not by itself rule out local minima. The algorithm section should be reframed as a heuristic, or a formal convergence proof supplied.
minor comments (3)
  1. [Section 5] Typographical errors: 'gain of of gene' and 'trees trees' appear in the biological interpretation paragraph. Please proofread.
  2. [Figure 11 caption] The caption says 'passing through (0,x)', but the minimum occurs at (x,0) with x = (ρ1 z2 + ρ2 z1)/(ρ1+ρ2). The notation should be corrected.
  3. [Section 4.1] The text introduces an equivalence relation in Definition 9 and then says 'We can further refine d into a proper metric function on T P(N)/≃ by defining T1≃T2 whenever d(T1,T2)=0.' If this creates a new equivalence relation, subsequent uses of ≃ (including Theorem 5) need to specify which relation is meant. This is not fatal, but the notation should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Towering distance is constructed from BHV distance by an explicit quotient, and the separation theorem is proved rather than assumed.

full rationale

The paper's central construction is not circular. Definition 10 defines Towering distance as a quotient pseudometric built from BHV distance and the equivalence relation generated by pruning/regrafting operations (Definition 9). The key separation result, Theorem 5, is proved in the supplement via Lemmas 6–8, not assumed: it shows that d(T1,T2)=0 implies T1 and T2 are equivalent when T1 has strictly positive external edges. The arguments use the BHV metric and explicit projection/pruning inequalities rather than presupposing the Towering distance. There are no fitted constants, no parameter estimated from a subset and then 'predicted', and no load-bearing self-citation: the authors' own prior work (Valdez Cabrera & Willis 2025) is cited only as background on Extension-space dissimilarities, and the proofs do not reduce to it. The main weakness flagged in the text is that completeness of Towering space is asserted in the abstract and Section 5 but never proved, and Section 4.4 relies on it ('By the completeness of Towering space, Theorem 3 and Theorem 4, this procedure is guaranteed to find at least one geodesic'). This is a genuine correctness/rigor gap, but it is not circularity: an unproved completeness claim does not make the distance definition or the separation theorem equivalent to its own inputs. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central construction is a quotient of BHV spaces under zero-cost leaf operations. No numeric parameters are fitted; the main modeling choices are the L2 merging rule and the zero-cost equivalence. Completeness and algorithm convergence are asserted but not proved, so they are treated as gaps rather than axioms.

assumptions (5)
  • standard math BHV tree space is a Hadamard (CAT(0)) complete geodesic space, and the Owen-Provan formula (Eq. 1) gives geodesic lengths.
    Invoked throughout Sections 2.2-2.3 and used for projections, convexity, and geodesic distance formulas.
  • domain assumption Leaf pruning and regrafting are treated as zero-cost operations in the equivalence relation (Definition 9).
    Central modeling choice that makes distances finite between trees with different leaf sets and defines the quotient in Eq. (2). Biological interpretation is gene gain and loss, but no empirical constraint fixes zero cost.
  • ad hoc to paper Edge lengths of merged edges after pruning combine by the L2 norm (Definition 7).
    Chosen to preserve distance to the origin and to make the Pythagorean inequality work. Not forced by BHV geometry or biology; L1 or other norms would give a different metric.
  • domain assumption Only trees with leaf set size at least 3 are included.
    The construction defines P(N) as subsets L of N with |L| at least 3, so small trees and the BHV edge cases for few taxa are excluded.
  • ad hoc to paper Pruning is restricted so that every internal edge maps to an internal edge under the TDR map (Definition 5).
    This restriction prevents zero-distance leaf swaps (Figure 4) and is a design choice rather than a consequence of the data or of BHV space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometry of the space of phylogenetic trees with non-identical leaves." pith.science (2026). https://pith.science/paper/ZIK2WIIN

@misc{pith2026250806747,
  author       = {Pith},
  title        = {Pith review of: Geometry of the space of phylogenetic trees with non-identical leaves},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZIK2WIIN}},
  note         = {Machine review of arXiv:2508.06747}
}
read the original abstract

Phylogenetic trees summarize evolutionary relationships. The Billera-Holmes-Vogtmann (BHV) space for comparing phylogenetic trees has many elegant mathematical properties, but it does not encompass trees with differing leaf sets. To overcome this, we introduce Towering space: a complete metric space that extends BHV space to trees with non-identical leaf sets. Towering space is a structured collection of BHV spaces connected via pruning and regrafting operations. We study the geometry of paths in Towering space and present an algorithm for computing metric distances. By addressing a major limitation of BHV space, Towering space facilitates the analysis of modern phylogenetic datasets such as multi-domain gene trees.

Figures

Figures reproduced from arXiv: 2508.06747 by the authors.

Figure 1
Figure 1. (a) A phylogenetic tree with leaf set N = {𝐴, 𝐵, 𝐶, 𝐷, 𝐸} and internal split 𝑠1 = {𝐴, 𝐵} [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Topologies 𝑆1, 𝑆2 and 𝑆3 share topology 𝑆 as a face. The topology 𝑆𝑖 is composed of internal edges 𝑆𝑖 = {𝑠𝑖 , 𝑠4}, while the topology 𝑆 only has 𝑆 = {𝑠4}; thus O (𝑆) is a face of O (𝑆𝑖) for 𝑖 = 1, 2, 3. (b) The topology orthants O (𝑆1), O (𝑆2) and O (𝑆3) are “glued” at O (𝑆). The dimensions arising from external edges are not shown. 𝑑BHV(𝑇1, 𝑇2) = vut∑︁ 𝑘 𝑖=1 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The projection of 𝑇 (left) with leaf set L = {𝐴, 𝐵, 𝐶, 𝐷, 𝐸, 𝐹, 𝐺} onto the M-trimmable space for M = {𝐹, 𝐺} (right). The edges in 𝑃 ↓M (𝑇) preventing M from being prunable from 𝑇 are shown in orange. However, the same set is prunable from 𝑇 ⊥M. The distance between the two trees is √ 3. When {𝐹, 𝐺} is pruned from 𝑇 ⊥M, the edge length of {𝐴, 𝐵}|{𝐶, 𝐷, 𝐸} is the 𝐿 2 norm of the adjacent edges, √ 2 2 + 2 2 [PITH_FU… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: We impose the restriction that leaves can only be pruned from internal, not external, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Shortest paths via a single pruning and regrafting through a lower level. Shortest paths [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Efficient paths from 𝑇1 ∈ T L1 to 𝑇2 ∈ T L2 . (a) The path length from 𝑇1 to 𝑇2 can be reduced by consolidating consecutive prunings into single operations, and likewise for regraftings. Different BHV spaces on this path are connected through one prune-and-regraft oper…
Figure 7
Figure 7. Figure 7: For the above tree, consider the subset of leaves [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Regrafting an independent maximal set neighboring a leaf [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Regrafting independent maximal sets attached to a common edge [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Regrafting independent maximal sets attached to uncommon edges for [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: A two-dimensional representation of the minimizing function [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages

  1. [1]

    Computing medians and means in hadamard spaces.SIAM Journal on Optimization24, 1542–1566

    Bac´ak, M.(2014). Computing medians and means in hadamard spaces.SIAM Journal on Optimization24, 1542–1566

    work page 2014

  2. [2]

    & Owen, M

    Barden, D., Le, H. & Owen, M. (2013). Central limit theorems for Fr´echet means in the space of phylogenetic trees. Electronic Journal of Probability 18, 1–25

    work page 2013

  3. [3]

    Point estimates in phylogenetic reconstructions.Bioinformatics 30, 534–540

    Benner, P., Bac´ak, M.& Bourguignon, P.-Y.(2014). Point estimates in phylogenetic reconstructions.Bioinformatics 30, 534–540

    work page 2014

  4. [4]

    J., Holmes, S

    Billera, L. J., Holmes, S. P. & Vogtmann, K. (2001). Geometry of the space of phylogenetic trees. Advances in Applied Mathematics 27, 733–767

    work page 2001

  5. [5]

    Bridson, M. R. & Haefliger, A. (1999). Metric Spaces of Non-Positive Curvature , vol. 319 of Grundlehren der mathematischen Wissenschaften. Berlin, Heidelberg: Springer Berlin Heidelberg

    work page 1999

  6. [6]

    Brown, D. G. & Owen, M. (2020). Mean and variance of phylogenetic trees. Systematic Biology 69, 139–154

    work page 2020

  7. [7]

    & Dessimoz, C

    Gori, K., Suchan, T., Alvarez, N., Goldman, N. & Dessimoz, C. (2016). Clustering genes of common evolutionary history. Molecular Biology and Evolution 33, 1590–1605

    work page 2016

  8. [8]

    Grindstaff, G.& Owen, M. (2019). Representations of partial leaf sets in phylogenetic tree space. SIAM Journal on Applied Algebra and Geometry 3, 691–720. 26 Valdez-Cabrera & Willis

    work page 2019

Show all 24 references
  1. [9]

    & Zhang, K

    Hein, J., Jiang, T., Wang, L. & Zhang, K. (1996). On the complexity of comparing evolutionary trees. Discrete Applied Mathematics 71, 153–169

    1996

  2. [10]

    Jain, R., Rivera, M. C. & Lake, J. A. (1999). Horizontal gene transfer among genomes: The complexity hypothesis. Proceedings of the National Academy of Sciences 96, 3801–3806

    1999

  3. [11]

    V., Makarova, K

    Koonin, E. V., Makarova, K. S. & Aravind, L. (2001). Horizontal Gene Transfer in Prokaryotes: Quantification and Classification. Annual Review of Microbiology 55, 709–742

    2001

  4. [12]

    & Tahiri, N

    Li, W., Koshkarov, A. & Tahiri, N. (2024). Comparison of phylogenetic trees defined on different but mutually overlapping sets of taxa: A review. Ecology and Evolution 14, e70054

    2024

  5. [13]

    & Provan, J

    Miller, E., Owen, M. & Provan, J. S.(2015). Polyhedral computational geometry for averaging metric phylogenetic trees. Advances in Applied Mathematics 68, 51–91

    2015

  6. [14]

    Nye, T. (2011). Principal components analysis in the space of phylogenetic trees. The Annals of Statistics 39, 2716–2739

    2011

  7. [15]

    & Yoshida, R.(2017)

    Nye, T., Tang, X., Weyenberg, G. & Yoshida, R.(2017). Principal component analysis and the locus of the Fr´echet mean in the space of phylogenetic trees. Biometrika 104, 901–922

    2017

  8. [16]

    Computing geodesic distances in tree space.SIAM Journal on Discrete Mathematics25, 1506–1529

    Owen, M.(2011). Computing geodesic distances in tree space.SIAM Journal on Discrete Mathematics25, 1506–1529

    2011

  9. [17]

    S.(2011)

    Owen, M.& Provan, J. S.(2011). A fast algorithm for computing geodesic distances in tree space.IEEE Transactions on Computational Biology and Bioinformatics 8, 2–13

    2011

  10. [18]

    & Sturmfels, B

    Pachter, L. & Sturmfels, B. (2005). Algebraic statistics for computational biology

    2005

  11. [19]

    A., Monical, C., Delcourt, M., Guzman, R

    Ren, Y., Zha, S., Bi, J., Sanchez, J. A., Monical, C., Delcourt, M., Guzman, R. K. & Davidson, R. (2017). A combinatorial method for connecting bhv spaces representing different numbers of taxa. arXiv:1708.02626

    2017 arXiv

  12. [20]

    Sturm, K.-T. (2003). Probability measures on metric spaces of nonpositive curvature. Contemporary Mathematics 338, 357–390. Valdez Cabrera, M. A. & Willis, A. D. (2025). Distances between extension spaces of phylogenetic trees. IEEE Transactions on Computational Biology and Bi...

    2003

  13. [21]

    M., Schardl, C

    Weyenberg, G., Huggins, P. M., Schardl, C. L., Howe, D. K. & Yoshida, R. (2014). Kdetrees: non-parametric estimation of phylogenetic tree distributions. Bioinformatics 30, 2280–2287

    2014

  14. [22]

    Willis, A. (2019). Confidence sets for phylogenetic trees. Journal of the American Statistical Association 114, 235–244

    2019

  15. [23]

    & Bell, R

    Willis, A. & Bell, R. (2018). Uncertainty in phylogenetic tree estimates. Journal of Computational and Graphical Statistics 27, 542–552

    2018

  16. [24]

    , Khiabanian, H

    Zairis, S. , Khiabanian, H. , Blumberg, A. J. & Rabadan, R. (2016). Genomic data analysis in tree spaces. arXiv:1607.07503

    2016 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.