{"id":"53fe3b9a-c852-49cf-84ea-d1519b7d903e","arxiv_id":"2411.14057","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For DAGs with a unique LCA for every leaf subset of allowed size, the clusters that survive the simplification operator are exactly those that are inclusion-minimal for some allowed subset, and the deleted vertex set is unique.","lead":"This mathematics paper shows when a directed acyclic graph used to model evolutionary history can be simplified without losing clusters of species, and proves that the simplifying vertex set is unique. It pins down exactly which clusters survive under the I-LCA framework, generalizing earlier binary and k-ary results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.9's characterization and uniqueness proof depends on imported (S4) from [19]; a failure of (S4) would invalidate both C_{G⊖W}=C_G(I1) and the uniqueness of W.","rationale":"The reader identified exactly the same weakest assumption: Theorem 2.12(S4) is imported from [19] and is not reproven. My own reading of the internal proof of Theorem 3.9 found no additional gap: the inclusions C_G(I1)⊆C_{G⊖W} and C_{G⊖W}⊆C_G(I1) follow from (S0), Lemma 2.3, and the definition of I1-lca vertices; the I1-ary conclusion follows from Lemma 3.6; and the uniqueness argument is valid once (S4) supplies lca_{G⊖W}(A)=lca_G(A). The abstract overclaim about tree/galled-tree preservation is a real presentation issue, but it concerns Theorem 4.5 and does not weaken Theorem 3.9 itself. Because the central theorem inherits its main mechanism from (S4), a targeted verification of (S4) is the check that would settle whether the concern lands. Since the reader already conditioned on this same dependency, the verdict should remain CONDITIONAL.","tokens_in":15630,"tokens_out":15554,"duration_ms":144454,"concrete_test":"Write a small exhaustive checker over all labeled DAGs on up to 4 leaves (or, if that is infeasible, a randomized sampler over DAGs with up to 6 leaves). For every such G with the I1-lca-property, for every I1⊆{1,...,|L(G)|} with 1∈I1, and for every subset W of non-I1-lca vertices, verify that lca_{G⊖W}(A)=lca_G(A) for all A∈X(I1). If every instance passes, the imported (S4) is corroborated and Theorem 3.9 stands; if a counterexample appears, the characterization and uniqueness claims in Theorem 3.9 fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem is proven only relative to Theorem 2.12 of the prior paper [19]. The key dependency is (S4): for any W of non-I1-lca vertices, lca_{G⊖W}(A)=lca_G(A) for every A∈X(I1) for which lca_G(A) is well-defined. In a DAG with the I1-lca-property this applies to every A∈X(I1). The current paper does not reprove or independently verify (S4). This matters twice. First, Lemma 3.3 uses (S4) to conclude that G⊖W retains the I1-lca-property, which is then needed for Lemma 3.6 to certify that C_G(I1) is I1-ary. Second, the uniqueness argument in Theorem 3.9 uses (S4) to obtain lca_G(A)=lca_H(A)=lca_H*(A) for all A∈X(I1); this equality is what forces V(H)=V(H*) and hence W=W*. If (S4) failed for even one DAG with the I1-lca-property, the identity C_{G⊖W}=C_G(I1) could fail and another subset W* could plausibly yield an I1-lca-REL DAG satisfying (S0)-(S3) but not the same vertex set. The internal reasoning from (S4) onward appears sound, but the load-bearing assumption itself is imported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies DAGs within the I1-lca framework, where I1 is a set of leaf-subset sizes containing 1. It distinguishes between I1-lca-relevant DAGs and DAGs with the I1-lca-property, and relates these to pre-I1-ary and I1-ary set systems. The main results are: a characterization of the I1-lca-property for DAGs satisfying path-cluster-comparability (Theorem 3.5); characterizations of grounded pre-I1-ary and I1-ary set systems as cluster systems of such DAGs (Theorem 3.7); a characterization, for DAGs with the I1-lca-property, of the cluster system of the transformed DAG G ⊖ W as the set C_G(I1) of I1-minimal clusters, together with uniqueness of the removed vertex set W (Theorem 3.9); and a consequence for DAGs whose cluster system is tree-like or galled-tree-like, stating that (G ⊖ W)^− is a tree or galled-tree with the same cluster system (Theorem 4.5). The proofs are detailed and the paper generalizes earlier results for binary and k-ary set systems.","tokens_in":15869,"tokens_out":17625,"duration_ms":153064,"significance":"If the results hold, the paper provides a clean structural description of which clusters survive the ⊖-simplification for DAGs with the I1-lca-property, and it establishes uniqueness of the simplification. This is a useful contribution to the theory of LCA-based DAG simplification and generalizes prior work on binary and k-ary clustering systems. The paper is careful with definitions, provides helpful examples (Figures 1–3), and proves its main theorems in a coherent manner. The main caveat is that the central Theorem 3.9 leans on property (S4) of Theorem 2.12 imported from the authors' companion paper [19]; the paper does not reprove (S4). In addition, the abstract and Section 5 overstate Theorem 4.5 by omitting the hypothesis 1 < k ≤ κ_G. These issues are correctable and do not undermine the body's main derivation, but they do affect how the results are presented.","major_comments":[{"comment":"The abstract and the Summary and Outlook section state that for a DAG G with the I1-lca-property whose cluster system is tree-like or galled-tree-like, the shortcut-free transformed DAG is always a tree or galled-tree and that C_H = C_G. This omits the hypothesis in Theorem 4.5 that I1 contains an integer k with 1 < k ≤ κ_G, where κ_G is the size of the smallest non-singleton cluster. Without this hypothesis the statement is false: if I1 = {1}, then every DAG has the I1-lca-property, and for a non-trivial tree G the set W of all non-I1-lca vertices deletes all internal vertices, so (G ⊖ W)^− is an edgeless graph whose cluster system is not C_G. The abstract and Section 5 should either include the condition 1 < k ≤ κ_G or restrict the claimed consequence to the cases actually covered by Theorem 4.5.","section":"Abstract and Section 5"},{"comment":"The proof of Theorem 3.9 depends essentially on property (S4) of the imported Theorem 2.12 from [19], which asserts that lca_{G⊖W}(A) = lca_G(A) for every A ∈ X(I1) for which lca_G(A) is well-defined. This property is used in Lemma 3.3 to show that G ⊖ W retains the I1-lca-property, and it is used in the uniqueness argument of Theorem 3.9 to conclude that V(H) = V(H*) from the equality of lca values. The paper does not prove (S4) or justify it beyond citing [19]. Since the equality C_{G⊖W} = C_G(I1) and the uniqueness of W both rest on this imported property, the authors should either include a proof of (S4) or state explicitly that the main results are contingent on [19, Thm. 2.12] and give a precise location of its proof.","section":"Theorem 3.9 and Lemma 3.3"}],"minor_comments":[{"comment":"The abstract uses the symbol I, while the paper consistently uses I1 with the convention 1 ∈ I1. Please align the notation in the abstract with the body.","section":"Abstract and Introduction"},{"comment":"The phrase \"cannot not serve as LCAs\" should read \"cannot serve as LCAs\"; the double negative is confusing.","section":"Introduction, paragraph 2"},{"comment":"There is a typo: \"Then. G′ := (V,E \\ {e}) is a DAG\" should have a colon or comma after \"Then\".","section":"Section 2, Lemma 2.2"},{"comment":"The condition \"for some |I1| > 1\" is awkward because I1 is a set; it would be clearer to write \"for some set I1 containing an integer k > 1\".","section":"Section 3, Lemma 3.2"},{"comment":"There is a typo in the statement: \"with with 1 < k ≤ κG\" should be \"with 1 < k ≤ κG\".","section":"Section 4, Theorem 4.5"},{"comment":"The word \"examplify\" should be \"exemplify\".","section":"Figure 2 caption"},{"comment":"The sentence \"G⊖W and (G⊖W)− are I1-lca-REL DAGs on X with the I1-lca-property that satisfies (S0)–(S4)\" has a subject-verb agreement error; it should read \"that satisfy (S0)–(S4)\".","section":"Section 3, Theorem 3.9"}],"recommendation":"major_revision","confidential_remarks":"The body of the paper is mathematically sound in its main lines, and the reliance on Theorem 2.12 of [19] is a normal citation rather than a circularity. However, the abstract and Section 5 overstate Theorem 4.5 by dropping the necessary condition on k, and the dependence of the uniqueness result on imported property (S4) deserves to be made explicit. These are fixable with a revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is solid and genuinely new. For a DAG with the I1-lca-property, Theorem 3.9 identifies exactly which clusters survive the ⊖ simplification: they are C_G(I1), the clusters that are inclusion-minimal for some A of size in I1. Equally new is the proof that the deletion set W is unique under the (S0)-(S4) preservation requirements. The paper also sharpens the connection to pre-I-ary and I-ary set systems, generalizing the binary and k-ary cases from Shanavas et al. The proofs of Theorems 3.5, 3.7, 3.9, and 4.5 are detailed and internally coherent; I did not spot a gap in the reasoning from the stated assumptions.\n\nThe main problem is presentation, not mathematics. The abstract and the Section 5 summary claim that the transformed DAG is always a tree or galled-tree whenever C_G is tree-like or galled-tree-like and G has the I1-lca-property, with no further condition. Theorem 4.5 in the body requires I1 to contain an integer k with 1 < k ≤ κ_G. Without that hypothesis the statement is false, as the degenerate I1 = {1} case shows. This is a real overclaim in the public-facing parts of the paper, even though the formal theorem statement is correct.\n\nThe stress-test note raises the dependence on property (S4) from the authors' earlier paper [19]. That is a legitimate dependency: the uniqueness of W and the equality C_{G⊖W} = C_G(I1) both rest on (S4). I do not think this is a defect of the current paper, since building on prior published theorems is standard. But because the load is heavy, a referee should verify that (S4) is proved correctly in [19] or ask the authors to include a short independent proof. The paper would be stronger if it did.\n\nMinor issues: a few typos (e.g., \"with with\" in Theorem 4.5), and the paper does not reprove imported machinery, which is fine but worth flagging.\n\nWho should read this: people working on LCA-based DAG simplification, phylogenetic network encodings, and cluster systems. It is a meaningful advance in that subfield. I would cite it if I worked there. Send it to peer review; the overclaim in the abstract is fixable and the technical contribution deserves referee time.","headline":"The main theorem is real and the proofs look sound, but the abstract and Section 5 overstate Theorem 4.5 by dropping the k > 1 hypothesis.","tokens_in":16457,"tokens_out":1715,"would_cite":true,"duration_ms":18125,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C20","05C75"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper characterizes exactly which clusters survive the I1-LCA simplification of a DAG: those that are inclusion-minimal containers of some allowed leaf set, with a unique choice of removed vertices.","keywords":["DAGs","least common ancestors","clusters","I-ary set systems","phylogenetic networks","galled trees","Hasse diagrams","DAG simplification"],"falsifier":"Build a small DAG with the $\\{1,2\\}$-lca-property, list the vertices that are not the unique LCA of any two leaves, delete all of them with $\\ominus$, then remove shortcut edges. If the resulting cluster system differs from the set of clusters that are inclusion-minimal containers of some two leaves, or if the shortcut-free graph is not isomorphic to the Hasse diagram of that cluster set, Theorem 3.9 is false. A brute-force enumeration of all DAGs on four leaves would settle this completely.","tokens_in":15355,"feed_emoji":"🧬","tokens_out":8581,"duration_ms":67128,"temperature":0.7,"pith_summary":"Directed acyclic graphs used in phylogenetics are often simplified by deleting vertices that are not the least common ancestor of any leaf set of an allowed size. This paper investigates that simplification in the setting where every allowed leaf set already has a unique LCA. It claims that the clusters that survive are exactly the clusters that are inclusion-minimal containers for some allowed leaf set, and that the set of vertices deleted is uniquely determined. It then derives that if the original cluster system is tree-like or galled-tree-like, the simplified graph is always a tree or galled-tree with the same clusters. These results matter because they turn cluster loss under simplification from an artifact of vertex choice into a structural property of the cluster system.","feed_headline":"Which clusters survive a DAG simplification? Exact answer","feed_subtitle":"For DAGs with unique LCAs, surviving clusters are precisely the inclusion-minimal ones, and the deleted vertex set is unique","key_machinery":"The machinery is the pair consisting of the cluster subsystem $\\mathcal{C}_G(I_1)$ and the $\\ominus$-operator. For a set system $\\mathcal{C}$, the subsystem $\\mathcal{C}(I_1)$ collects exactly the clusters that satisfy property (I1-C): each is the unique inclusion-minimal element of $\\mathcal{C}$ containing some subset $A$ of leaves with $|A|\\in I_1$. The $\\ominus$-operator deletes a vertex and joins each parent to each child; earlier work guarantees properties (S0)--(S4), which say the deletion introduces no new clusters, keeps the same leaves, keeps the same vertex set minus the deleted vertices, preserves the ancestor order, and preserves well-defined LCAs of allowed leaf sets. These properties let the proof show that the clusters of the simplified graph are exactly $\\mathcal{C}_G(I_1)$, and that this subsystem is $I_1$-ary.","core_discovery":"The central result is Theorem 3.9. Let $G$ be a DAG with the $I_1$-lca-property on leaf set $X$, and let $W$ be the set of all vertices that are not $I_1$-lca vertices. Then the shortcut-free graph $(G \\ominus W)^-$ is isomorphic to the Hasse diagram $H(\\mathcal{C}_G(I_1))$, where $\\mathcal{C}_G(I_1)$ consists of those clusters in $\\mathcal{C}_G$ that are the unique inclusion-minimal cluster containing some leaf set of size in $I_1$. In particular, $\\mathcal{C}_{G\\ominus W}=\\mathcal{C}_{(G\\ominus W)^-}=\\mathcal{C}_G(I_1)$ is an $I_1$-ary set system, and $W$ is the unique, hence smallest, subset of vertices whose deletion makes the graph $I_1$-lca-relevant while preserving the structural properties (S0)--(S4).","pith_inferences":["A computational search over small DAGs could test whether cluster survival for DAGs lacking the $I_1$-lca-property admits a similar description in terms of inclusion-minimal clusters or requires tracking the full LCA structure.","For the orthology-motivated case $I_1=\\{1,2\\}$, the theorem identifies a canonical simplification that preserves exactly the pairwise-LCA information, which may serve as a normal form for orthology-aware network simplification.","The paper leaves open how the $\\ominus$-operator relates to normalization by visible vertices; the uniqueness of $W$ here suggests that for $I_1$-lca-property DAGs the two procedures may agree more often than in general.","A testable algorithmic extension: compute $\\mathcal{C}_G(I_1)$ and compare it with $\\mathcal{C}_G$ to decide in polynomial time whether a given DAG with the $I_1$-lca-property loses any clusters under simplification."],"forward_implications":["For any DAG with the $I_1$-lca-property, the clusters lost under the simplification are exactly $\\mathcal{C}_G \\setminus \\mathcal{C}_G(I_1)$, so the loss is determined by the cluster system alone, not by the particular choice of $W$.","The vertex set $W$ whose deletion produces an $I_1$-lca-relevant DAG while preserving (S0)--(S4) is unique, hence the transformation is canonical for this class.","If $\\mathcal{C}_G$ is tree-like, then $(G \\ominus W)^-$ is a phylogenetic tree with the same cluster system; if $\\mathcal{C}_G$ is galled-tree-like, it is a galled-tree with the same clusters.","The theorem extends earlier binary and $k$-ary set-system characterizations to arbitrary sets $I_1$ of allowed leaf-set sizes.","Since $(G \\ominus W)^-$ is always regular, the simplification ends in a canonical Hasse-diagram object determined by $\\mathcal{C}_G(I_1)$."],"supporting_citations":[{"why":"Supplies the $\\ominus$-operator, Theorem 2.12 with properties (S0)--(S4), and the result that shortcut-free $I_1$-lca-relevant DAGs are regular.","marker":"[19]"},{"why":"Introduced the $k$-lca-DAGs and unique-LCA/cluster results that this paper generalizes to arbitrary sets $I_1$.","marker":"[21]"},{"why":"Established the binary set-system case that the current $I_1$-ary characterization extends.","marker":"[3]"},{"why":"Provides the characterization of tree-like and galled-tree-like clustering systems and the Hasse-diagram framework for phylogenetic networks.","marker":"[13]"},{"why":"Classical reference for tree-like cluster systems and Hasse diagrams of hierarchies.","marker":"[20]"},{"why":"Defines regular DAGs and the Hasse diagram representation used in the main theorem.","marker":"[2]"}],"fun_headline_variants":["Only inclusion-minimal clusters survive DAG transform","Exact survival set: minimal clusters and unique W","DAG simplification preserves only minimal clusters","Unique vertex set deletes non-LCA vertices, minimal clusters remain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that deleting vertices that are not unique LCAs of any allowed-size leaf set never changes the unique LCA of an allowed-size leaf set when that LCA already exists. If that preservation lemma from the earlier framework fails, the characterization of surviving clusters and the uniqueness of the deleted set both collapse.","fun_headline_variants_meta":{"raw":{"variants":["Only inclusion-minimal clusters survive DAG transform","Exact survival set: minimal clusters and unique W","DAG simplification preserves only minimal clusters","Unique vertex set deletes non-LCA vertices, minimal clusters remain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1861,"prompt_tokens":1204,"completion_tokens":657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":820,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":820,"tokens_out":657,"duration_ms":6593,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:35:12.470202+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a small DAG with the $\\{1,2\\}$-lca-property, list the vertices that are not the unique LCA of any two leaves, delete all of them with $\\ominus$, then remove shortcut edges. If the resulting cluster system differs from the set of clusters that are inclusion-minimal containers of some two leaves, or if the shortcut-free graph is not isomorphic to the Hasse diagram of that cluster set, Theorem 3.9 is false. A brute-force enumeration of all DAGs on four leaves would settle this completely.","supporting_citations":[{"cited_title":"Bulletin of Mathematical Biology 87(3):44, DOI 10.1007/ s11538-025-01419-z","cited_arxiv_id":null,"evidence_quote":"Supplies the $\\ominus$-operator, Theorem 2.12 with properties (S0)--(S4), and the result that shortcut-free $I_1$-lca-relevant DAGs are regular."}],"review_version":1}