Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Phylogenetic network models as graphical models

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Stacked reticulations are statistically invisible in displayed tree network models.

desk verdict Solid theory paper that recasts displayed-tree network models as DAG submodels and proves real nonidentifiability results; one repairable gap in Proposition 4.5 should be fixed before publication. read the letter →

arxiv 2507.23056 v2 pith:B7RAVLTJ submitted 2025-07-30 q-bio.PE math.COmath.STstat.TH

classification q-bio.PEmath.COmath.STstat.TH MSC 92D1562R01
keywords displayedtreemodelphylogeneticnetworksgraphicalmodelslocalmodificationssplittabilitystackedreticulationsflatteningranksequivariant
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 shows that the displayed tree model of evolution on a phylogenetic network is a special case of a directed graphical model, and it uses that connection to prove when two different networks produce the same probability distributions on observed leaves. The central result is that an edge from a reticulation vertex with no other outgoing edges can be contracted without changing the model's distribution family, provided the substitution matrices are multiplicatively closed, closed under convex combinations, and splittable. Since stacked reticulations are exactly such edges, networks that differ only by stacked reticulations cannot be distinguished from sequence data under these models. The same machinery shows that 2-blobs are invisible and yields rank bounds on flattenings of the probability tensor that generalize classic results for phylogenetic trees. A sympathetic reader cares because knowing which network features are identifiable is a precondition for inferring reticulate evolution from data.

What carries the argument

The central object is a local modification of a DAG: a triple (A, B, C) of vertex sets in which every edge into B comes from A or B, every edge out of B goes to B or C, and every edge into C comes from A, B, or C. Theorem 4.4 says that two DAGs that are local modifications with the same conditional family p_{C|A} give the same joint family once B is hidden. Splittability, defined in Definition 3.6, is the closure property on transition matrices that lets the proof reverse an edge contraction: for any finite list of matrices M_i in the model, there must exist a single matrix N in the model such that every M_i $N^{{-1}}$ is also in the model. Splittability holds for general equivariant and open equivariant phylogenetic models, which is why the contraction results apply uniformly across model types rather than only to group-based models or the general Markov model.

What would settle it

For a concrete four-state model satisfying the three conditions, such as the general Markov model with strictly positive transition matrices, compute the dimension of the displayed-tree distribution family for the left and middle networks in Figure 1.1; the paper predicts the two dimensions are exactly equal, so a pair with different dimensions under such a model would contradict Theorem 5.1.

Watch

Extended reading notes

Core claim

The displayed tree phylogenetic network model sits as a natural submodel of the graphical model associated to a directed acyclic graph, and this representation carries the paper's main results. Theorem 5.1 establishes that if the transition model is multiplicatively closed, closed under convex combinations, and splittable, then contracting an edge b -> c where b has outdegree 1 does not change the family of leaf distributions when b is hidden. Because a stacked reticulation is precisely such an edge from one reticulation to another, networks that differ only by stacked reticulations are distributionally equivalent. The same argument, applied repeatedly, implies that a 2-blob can be replaced by a single edge without changing the model family. The paper also derives linear relations among conditional distributions at a reticulation node that cause a dimension loss of (m-1)k parameters, and it uses d-separation to bound flattening ranks of the leaf probability tensor by k^(#E) for any edge cutset E.

Load-bearing premise

The set of allowed substitution matrices must be splittable: for any finite list of matrices M_i in the model, some single matrix N in the model must exist such that every M_i $N^{{-1}}$ is also in the model, and this is what lets the proof undo a contracted edge.

Editorial extensions

If this is right

  • Two binary phylogenetic networks that differ only by a stacked reticulation yield exactly the same family of leaf distributions under multiplicatively closed, convex, splittable models, so those networks cannot be distinguished by displayed-tree data.
  • Any 2-blob in a network can be replaced by a single edge without changing the displayed-tree distribution family, under the same conditions, meaning blob structure is not identifiable from the displayed-tree model alone.
  • For the general Markov model on k states with a reticulation node having m parents, the space of conditional distributions has dimension (k-1)(m(k-1)+1), losing (m-1)k dimensions relative to the number of parameters used to describe it.
  • If removing a set of edges E separates the leaves into groups A and B, then every distribution in the model satisfies rank Flat(A,B)(P) <= k^(#E).
  • If one displayed tree T has parsimony score l_T(A|B), then a generic distribution in the model has rank Flat(A,B)(P) at least min(k^(#A), k^(#B), k^(l_T(A|B))).

Reading between the lines

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

  • A practical consequence the paper does not spell out is that any inference pipeline treating stacked reticulations as distinct evolutionary hypotheses is fitting statistically equivalent models, so likelihood-based support for one network over the other cannot come from the displayed-tree model alone.
  • The flattening rank bounds suggest an implementable model check: compute the flattening of the observed site-pattern tensor for a candidate network and compare its rank to k^(#E); a violation would indicate the network or the substitution model is wrong.
  • The local-modification technique may transfer to other hidden-variable graphical models beyond phylogenetics, wherever a conditional distribution factors through an intermediate variable that can be contracted under a suitable closure condition.
  • The dimension loss at reticulation nodes suggests that even when a reticulation's position in the graph is identifiable, the mixing proportions and branch lengths entering that node may not be separately estimable from displayed-tree data.
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

2 major / 5 minor

Summary. The paper develops a DAG-graphical-model perspective on the displayed tree phylogenetic network model. It introduces local modifications of DAGs and a splittability condition on sets of transition matrices, then proves that, under multiplicative closure, convex closure, and splittability, contracting certain hidden edges—including stacked reticulation edges—does not change the family of distributions. This leads to nonidentifiability results for stacked reticulations and 2-blobs. The paper also derives dimension formulas for reticulation conditional distributions under the general Markov model and proves rank bounds on flattenings that generalize classical tree-model results.

Significance. If the results are fully established, the paper offers a valuable unifying framework: displayed tree models are constrained DAG graphical models, and local modification arguments yield model-uniform statements instead of case-by-case algebraic computations. The splittability property is cleanly defined and proven for general equivariant and open equivariant models, and the central stacked-reticulation contraction (Theorem 5.1) is proved in detail. The rank conditions in Section 7 are a genuine generalization of tree flattening results. The main weakness is an incomplete proof in Proposition 4.5, which is local and repairable but affects later results that rely on the equality of the two distribution families.

major comments (2)
  1. [Section 4, Proposition 4.5] The proof of Proposition 4.5 only establishes one containment: it shows that p_{G',c|A}(x_c|x_A) from the subdivided-edge graph is a special case of p_{G,c|A}(x_c|x_A) by taking M_ac = M_ab M_bc. It never proves the reverse containment, although the proposition states that the two DAGs produce the same family of distributions. This is not merely a presentational shortcut: Proposition 5.4 and Theorem 7.8 both rely on the equality of the two families. The missing direction is straightforward using splittability: for any M_ac in the model, choose N in the transition-matrix set with M_ac N^{-1} in the model, set M_ab = M_ac N^{-1} and M_bc = N, and leave the remaining parent terms and reticulation weights unchanged. Please add this argument, or alternatively restate the proposition as a one-sided containment and adjust the later results accordingly.
  2. [Section 5, Proposition 5.4 and Section 7, Theorem 7.8] Both Proposition 5.4 and Theorem 7.8 depend on the equality stated in Proposition 4.5, not merely on the containment proven there. In Proposition 5.4, the step 'We can assume that all degree 2 vertices within the 2-blob have been contracted by using Proposition 4.5' inherits the gap described above. In Theorem 7.8, if only the containment from the subdivided graph G' to the original graph G were available, then a distribution arising from G need not be representable in G', and the d-separation/rank argument would not apply to all distributions in the model. Once the reverse containment of Proposition 4.5 is supplied, both arguments are sound, but as written they are not fully supported.
minor comments (5)
  1. [Section 6, proof of Proposition 6.1] In the first display of the proof, the second sum is missing the factor pi^c_a: the term should be pi^c_a M_ac(x_c|y_a), not M_ac(x_c|y_a).
  2. [Section 7, proof of Theorem 7.11] The proof states that matrix rank is upper-semicontinuous and uses this to conclude that generic parameter values have rank at least the rank at a particular point. The correct fact is that rank is lower-semicontinuous; the wording should be corrected.
  3. [Section 5, Theorem 5.1] The theorem assumes only that b has outdegree 1, but the reverse construction divides by delta^c_b, which can be zero if A1 union A3 is empty. Since the intended application is to stacked reticulations, where b has indegree greater than one and hence has parents, the theorem statement should either include that assumption explicitly or discuss this edge case.
  4. [Example 4.7] The dimension computation for the Jukes-Cantor model is asserted without details. Please include the calculation or a reference, since the contrast between dimensions 8 and 9 is used to argue that two networks do not yield the same family.
  5. [References [6,7]] The text cites [6] as the source of generalizations of Theorem 6.2 and for level-1 identifiability results, but [6] is listed as 'In preparation'. If these citations are not load-bearing, consider marking them as pointers; if the results are needed, state the relevant statements in the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the displayed-tree/DAG equivalence is a reparametrization, Theorem 5.1 follows from stated splittability assumptions, and self-citations are non-load-bearing; Proposition 4.5 has a repairable proof gap, not a circular step.

full rationale

The paper's central derivation is self-contained and does not reduce to its own inputs. The displayed-tree model is explicitly described as a DAG graphical model with reticulation conditional distributions of the form p_j|pa(j)(x_j|x_pa(j)) = sum_i pi^i_j M_ij(x_j|x_i); Example 3.2 expands this by the distributive law into a sum over displayed trees. This is a reparametrization of the model definition, not a prediction derived from an independent first-principles result, so there is no self-definitional circularity. Theorem 5.1 (stacked reticulation contraction) is derived from explicit, stated hypotheses: multiplicative closure, closure under convex combinations, and splittability. The forward direction constructs parameters in the contracted graph from the uncontracted one; the reverse direction uses splittability to choose M_bc with M_ab = M_ac (M_bc)^{-1}, and convex closure to handle the A3 terms. The conclusion is not presupposed by the hypotheses; it is a genuine mathematical consequence. Equivariant splittability is proven in the paper (Propositions 3.7 and 3.8) from identity containment or positivity, not imported from the authors' prior work. The two in-preparation self-citations [6,7] appear only as pointers to generalizations ("Generalizations of Theorem 6.2 ... appear in [6]") and in the acknowledgments; they are not load-bearing for any theorem. The rank conditions in Section 7 are based on d-separation (Proposition 7.3, cited to Lauritzen) and external tree-flattening results [4,14], not on the authors' own unverified claims. One non-circular proof gap should be noted: Proposition 4.5 (edge subdivision) proves only that the subdivided-edge distribution is a special case of the unsubdivided one, by taking M_ac = M_ab M_bc, and then stops. The reverse containment is omitted; splittability would supply it, so the gap is repairable. This omission affects the written support for Proposition 5.4 and Theorem 7.8, which rely on Proposition 4.5, but it is a missing argument, not a circular reduction. The core nonidentifiability result (Theorem 5.1) contains its own detailed reverse construction and survives independently. Overall, no circularity is present; the score is 0.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The paper is a pure theory contribution. It introduces no fitted numerical constants and no new physical entities; its new notions (local structure, local modification, splittability) are definitions. The principal assumptions are the closure properties of the transition matrix class (multiplicative, convex, splittable) and the standard machinery of DAG graphical models and tree flattening ranks taken from the cited literature.

assumptions (7)
  • standard math DAG graphical models factor as products of conditional distributions (Eq. 1)
    Background from Lauritzen (1996), used throughout as the definition of the DAG model.
  • standard math d-separation characterizes conditional independence in DAG models (Prop 7.3)
    Standard result from graphical models literature, cited to Lauritzen (1996).
  • domain assumption Equivariant Markov models are closed under matrix multiplication and convex combinations (Section 3.1)
    Assumed property of the model class; used to construct new transition matrices in Theorems 4.4, 5.1.
  • domain assumption Transition matrix sets are splittable (Definition 3.6)
    Proven for equivariant and open equivariant models (Props 3.7, 3.8), but it is a restrictive condition that the nonidentifiability theorems depend on.
  • domain assumption Displayed tree model is the mixture over displayed trees with convex combination at reticulation vertices (Section 3)
    This is the definition of the model under study; the paper shows it is a submodel of the DAG model.
  • domain assumption Leaves are the only observed variables and have no hidden descendants (Section 3)
    Standard assumption in phylogenetic network models; used implicitly in all marginalization arguments.
  • standard math Tree flattening rank lower bound (Casanellas-Fernandez-Sanchez, Snyman et al.)
    Used in Theorem 7.11 as a black box from the cited literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phylogenetic network models as graphical models." pith.science (2026). https://pith.science/paper/B7RAVLTJ

@misc{pith2026250723056,
  author       = {Pith},
  title        = {Pith review of: Phylogenetic network models as graphical models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B7RAVLTJ}},
  note         = {Machine review of arXiv:2507.23056}
}
read the original abstract

The displayed tree phylogenetic network model is shown to sit as a natural submodel of the graphical model associated to a directed acyclic graph (DAG). This representation allows to derive a number of results about the displayed tree model. In particular, the concept of a local modification to a DAG model is developed and applied to the displayed tree model. As an application, some nonidentifiability issues related to the displayed tree models are highlighted as they relate to reticulation edges and stacked reticulations in the networks. We also derive rank conditions on flattenings of probability tensors for the displayed tree model, generalizing classic results for phylogenetic tree models.

Figures

Figures reproduced from arXiv: 2507.23056 by the authors.

Figure 1.1
Figure 1.1. Since the two binary networks have stacked reticulations, they then give the same distributions on the observed leaves α, β, γ, δ as the (non-binary) network in the middle. One goal for this note is to make the connection between the displayed tree model and the DAG models more widely known. We use this fact to prove some straightforward results about the displayed tree model, and we want to advertise this perspecti… view at source ↗
Figure 2.1
Figure 2.1. A directed four-cycle C4 3. The displayed tree model Phylogenetic network models arise as special cases of the general DAG graphical model by putting multiple types of restrictions on the DAGs that can arise, the particular structure of the conditional distributions that are used, and the fact that many of the variables are unobserved random variables (i.e. hidden random variables or latent random variables). Throug… view at source ↗
Figure 3.1
Figure 3.1. A 6 sunlet network, and its two displayed trees. Example 3.3. For a more phylogenetics relevant example, consider the network in [PITH_FULL_IMAGE:figures/full_fig_p006_3_1.png] view at source ↗
Figures from the paper (4 more)
Figure 4.1
Figure 4.1. Figure 4.1: The diagram give the idea of a local structure. Note that there can be directed edges within each of the groups A, B, and C. so that ∥N −1 ∥ ≤ 1 1 − ∥A∥ ≤ 2. since we assume that ∥A∥ is very small. Then ∥I − N −1 ∥ = ∥N −1 (N − I)∥ ≤ ∥N −1 ∥∥N − I∥ ≤ ϵ µ Then for eac…
Figure 4.2
Figure 4.2. Figure 4.2: Example of an unusual containment between phylogenetic network models Hence, we see that each conditional distribution from G produces a distribution from G′ by taking π c a = π b a for all a, and Mac = MabMbc for all a. This is valid because we assumed that the mode…
Figure 5.1
Figure 5.1. Figure 5.1: A stacked reticulation contraction of the edge 4 → 5. All three networks produce the same probability distributions on α, β, γ Since the substitution model is closed under matrix multiplication and convex combinations, this produces reticulation probabilities and tra…
Figure 5.2
Figure 5.2. Figure 5.2: A network with a 2-blob consisting of vertices {3, 4, 5} Example 5.2. Consider the graphs in [PITH_FULL_IMAGE:figures/full_fig_p015_5_2.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. On Tree-Network Distinguishability and Full Identifiability of Phylogenetic Networks

    q-bio.PE 2026-07 unverdicted novelty 7.0 of 10

    Under JC, K2P, and K3P substitution models, the topology of a level-1 phylogenetic network is fully identifiable from leaf-pattern distributions, and trees can be distinguished from networks unless the network is a tr...

Reference graph

Works this paper leans on

16 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    Allman, Hector Baños, Marina Garrote-Lopez, and John A

    Elizabeth S. Allman, Hector Baños, Marina Garrote-Lopez, and John A. Rhodes,Identifiability of level-1 species networks from gene tree quartets, Bull. Math. Biol.86 (2024), no. 9, Paper No. 110, 48. MR 4778334

  2. [2]

    Allman, Hector Baños, Jonathan D

    Elizabeth S. Allman, Hector Baños, Jonathan D. Mitchell, and John A. Rhodes,The tree of blobs of a species network: identifiability under the coalescent, J. Math. Biol.86 (2023), no. 1, Paper No. 10, 26. MR 4519611

  3. [3]

    Allman and John A

    Elizabeth S. Allman and John A. Rhodes,Phylogenetic ideals and varieties for the general Markov model, Adv. in Appl. Math.40 (2008), no. 2, 127–148. MR 2388607

  4. [4]

    Marta Casanellas and Jesús Fernández-Sánchez,Relevant phylogenetic invariants of evolutionary models, J. Math. Pures Appl. (9)96 (2011), no. 3, 207–229. MR 2831580

  5. [5]

    , Phylogenetic invariants: straightforward from the general markov to equivariant models , arxiv.org/abs/2310.18053, 2023

  6. [6]

    Marta Casanellas, Jesús Fernández-Sánchez, Elizabeth Gross, Benjamin Hollering, and Seth Sullivant,Phy- logenetic networks evolving underg-equivariant models, In preparation, 2025

  7. [7]

    Aviva Englander, Martin Frohn, Elizabeth Gross, Niels Holtgrefe, Mark Jones, Leo van Iersel, and Seth Sullivant, Identifiability of phylogenetic level-2 networks under the jukes-cantor model, In preparation, 2025

  8. [8]

    Joseph Felsenstein,Inferring phylogenies, Sinauer, 2003

Show all 16 references
  1. [9]

    Elizabeth Gross and Colby Long, Distinguishing phylogenetic networks, SIAM J. Appl. Algebra Geom. 2 (2018), no. 1, 72–93. MR 3763914

  2. [10]

    Elizabeth Gross, Leo van Iersel, Remie Janssen, Mark Jones, Colby Long, and Yukihiro Murakami,Distin- guishing level-1 phylogenetic networks on the basis of data generated by Markov processes, J. Math. Biol.83 (2021), no. 3, Paper No. 32, 24. MR 4311049

  3. [11]

    Sungsik Kong, Joan Carles Pons, Laura Kubatko, and Kristina Wicke,Classes of explicit phylogenetic net- works and their biological and mathematical significance, J. Math. Biol.84 (2022), no. 6, Paper No. 47, 44. MR 4417412

  4. [12]

    Lauritzen,Graphical models, Oxford Statistical Science Series, vol

    Steffen L. Lauritzen,Graphical models, Oxford Statistical Science Series, vol. 17, The Clarendon Press, Oxford University Press, New York, 1996, Oxford Science Publications. MR 1419991

  5. [13]

    Rhodes, Hector Baños, Jingcheng Xu, and Cécile Ané,Identifying circular orders for blobs in phylo- genetic networks, Adv

    John A. Rhodes, Hector Baños, Jingcheng Xu, and Cécile Ané,Identifying circular orders for blobs in phylo- genetic networks, Adv. in Appl. Math.163 (2025), Paper No. 102804. MR 4828752

  6. [14]

    Jandre Snyman, Colin Fox, and David Bryant,Parsimony and the rank of a flattening matrix, J. Math. Biol. 86 (2023), no. 3, Paper No. 44, 14. MR 4548024 PHYLOGENETIC NETWORK MODELS AS GRAPHICAL MODELS 21

  7. [15]

    89, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,

    Mike Steel,Phylogeny—discrete and random processes in evolution, CBMS-NSF Regional Conference Series in Applied Mathematics, vol. 89, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,

  8. [16]

    Email address: smsulli2@ncsu.edu Department of Mathematics, North Carolina State University, Raleigh, NC 27695

    Yan Yu, James H Degnan, and Luay Nakhleh,The probability of a gene tree topology within a phylogenetic network with applications to hybridization detection, PLOS Genetics8 (2012), e1002660. Email address: smsulli2@ncsu.edu Department of Mathematics, North Carolina State Univer...

Pith tools

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