On the Identifiability of Poisson Branching Structural Causal Model Under Latent Confounding

Summary

This paper extends the Poisson Branching SCM (PB-SCM) framework to admit unobserved common causes, calling the result LC-PB-SCM. The conceptual move is to translate branching-structure identifiability under latent confounding — a problem where classical linear-non-Gaussian and rank-based tools do not apply — into an algebraic question on the probability generating function (PGF) of count data. The authors show that every monomial in the log-PGF is in one-to-one correspondence with a Pruned Trie obtained by pruning a rooted prefix tree over directed paths of the causal graph. Because the coefficients of these Trie-generated monomials are products of edge-thinning probabilities that survive marginalisation of exogenous latent roots, the monomial set becomes a structural fingerprint one can read off empirical partial derivatives of $\log \hat G$. Local-PGF restrictions collapse the analysis to three observed variables at a time, yielding orientation rules (Theorems 3.7–3.12) that the authors argue are complete under monomial equivalence for the 29 canonical 3-node classes. A three-phase algorithm (skeleton, confounded-edge, triple orientation) is derived at cubic complexity and evaluated on synthetic and two real count-data settings. The contribution is genuinely novel: it is the first identifiability treatment of PB-SCMs with latent confounders, and the Trie/monomial correspondence is an elegant, non-obvious bridge between combinatorial structure and PGF algebra. Executing the released code confirms this theoretical core reproduces cleanly — the symbolic log-PGF, the empirical estimators, the discriminating-monomial argument, and the Case-2 structure recovery all match the paper. The main conceptual limitation is that the entire result is scoped to exogenous, mutually-independent latent roots and to local 3-variable substructures; latent mediators, hierarchies, and any identification needing $>3$-variable PGF terms are deferred. Several proof steps (Types (d1)/(d3), completeness Types 16/17/18) do not close as written, and — a new finding from execution — the headline synthetic F1 numbers did not reproduce and the comparative "beats all baselines" claim is untestable from the artifact.

Strengths

• Conceptual contribution. The Trie / log-PGF-monomial correspondence (Theorem 3.5) is a clean and, to my knowledge, new bridge from branching-process structure to a symbolic-algebraic fingerprint that survives marginalisation of exogenous latent roots. The "each Pruned Trie is one term in the log-PGF" intuition is genuinely useful, and executing the shipped symbolic code reproduces the running-example log-PGF byte-for-byte.
• Identifiability of confounded direct edges. The result that a bow (a direct edge coexisting with a latent common cause of its endpoints) becomes identifiable via the $s_i s_j^2$ signature is non-trivial: this configuration is generically non-identifiable in linear-Gaussian and rank-based settings, and the paper turns it into a testable third-derivative condition. The coefficient estimator for this signature converged to its true value in replication.
• Complete case analysis for local 3-node structures. Enumerating the 29 monomial equivalence classes and pairing each to an orientation rule is exactly the granularity a reader wants, and it makes Phase 3 a mechanical lookup once the empirical monomial set is estimated.
• Practical polynomial-time algorithm. The $O(d^3)$ complexity was confirmed by code inspection of the shipped loop structure (pairwise, edge-wise, and triple enumeration), making the method plausibly usable at moderate $d$.
• Reproducible empirical-PGF backbone. The empirical PGF and its third-derivative estimate concentrate with the reported decreasing-variance trend; running the shipped estimators reproduced Tables 6 and 7 in trend and magnitude, giving a reassuring numerical baseline for the Phase 1–3 tests.

Weaknesses

• Type (d1) proof is internally inconsistent with its conclusion. The proof explicitly admits an alternative where the third $s_3$ comes from an additional latent edge $L \to X_3$ with $X_1,X_3$ non-adjacent — a case with no $X_1$–$X_3$ edge — yet concludes $X_1 \circ\!\to X_3$, which presupposes an adjacent arrowhead. Since Theorem 3.13 (completeness) rests on each case closing, this must be repaired or the conclusion softened.
• Type (d3) does not exclude a direct latent edge to $X_2$. The argument that "$X_2$ appearing once" forces $L \to X_1 \to X_2$ (hence $X_1 \to X_2$) does not rule out a direct $L \to X_2$ with no $X_1 \to X_2$ edge; both produce a single $s_2$ copy.
• Completeness entries for Types 16/17/18 look inconsistent. Type 16 derives $X_1 \leftrightarrow X_2$ from the presence of $s_1^2 s_2^2 s_3$ while Type 17 derives the same from its absence; Type 18 claims a directed $X_3 \to X_2$ from just the adjacency monomial $s_2 s_3$, whereas everywhere else a confounded orientation requires the higher-order $s_2^2 s_3$ (Theorem 3.8). These read as citation errors or missing intermediate monomials.
• Theorem 3.9 is missing its monomial condition. The body announces "under the following monomial condition:" and gives none; the conditions live only in Figure 3 subcaptions, inconsistent with Theorems 3.10–3.12.
• Types (e1)/(e2) collapse to the same condition. Both invoke $s_1 s_2^2 s_3^2 \in M(G)$ but conclude opposite-strength targets; the distinguishing structural context exists only as figure drawings. Types (e3)–(e5) inherit similar under-specification.
• Two proof cruxes are appeals to intuition. Theorem 3.8's necessity ("$X_1$ would essentially be disconnected") and Type (c1)'s "$R=X_2$" sub-case ("must pass through $X_1$ since $X_1,X_2$ are adjacent") do not formally exclude latent-only alternative routes; and Type (c1) cites Theorem 3.7 to orient an edge, which that skeleton theorem cannot do.
• Restriction to exogenous latent roots is a substantive scope limit. Latent mediators, latent-to-latent chains, and hierarchies are excluded by Definition 2.1 — precisely the settings much compared work targets. The framing ("bridge this gap") suggests broader coverage than the assumptions deliver.
• Unspecified test protocol and no misspecification analysis. The bootstrap/threshold significance level and any multiple-comparison control across the $O(d^3)$ tests are never stated; absence-based rules (e3, e5, f1–f4) lean silently on faithfulness at exactly the finite-sample regime where they are most fragile. Appendix E studies only convergence under a correctly specified model, with no stress test under over-dispersion, zero-inflation, or additive contamination.
• Evaluation protocol changes between tables. The random-graph experiments switch to a laxer F1 (circle marks compatible) scored against the PADMG rather than the PPADMG, and to a much sparser generative regime ($\mu_i \in [0.001,0.005]$), so F1 values are not comparable across tables and the change is not motivated.
• Real-world experiments are minimal. The shopping-mall setting reduces to a single two-variable orientation once Impressions is declared latent (none of the local 3-node rules run), and the football setting is handed the true skeleton, so Phase 1 is never exercised on real data.
• Global-reconstruction claim overstated. The paper concedes global identifiability needs extra information, then claims local enumeration can "reconstruct the overall causal graph"; overlapping-triple conflicts and $>3$-variable edges are unaddressed, and Phase 2's asymmetric third derivative is applied as if symmetric.
• Minor consistency issues. Latent-mediator equivalence in the main text omits the mediator's Poisson-noise absorption that Appendix F correctly performs; LC-PBSCM (abstract) vs LC-PB-SCM; Des(i) written as membership in $\mathbf{E}$; bold $\mathbf{L}$ reused for both latent confounders and the local-PGF target set; Definition 2.5's domain excludes the latent indices the proofs use.

Reproducibility & code

• Theoretical core reproduces exactly. Running the shipped PGF.py, util.py, and pgf_confounder_partial.py unmodified reproduced: the running-example log-PGF to the last digit ($0.0048 s_1 s_2^2 + 0.034 s_1 s_2 + 0.0532 s_1 + 0.094 s_2 - 0.186$); the empirical PGF converging to the true $0.8315$ with the reported decreasing variance (Table 6); the $s_1 s_2^2$ coefficient converging to $0.0048$ (Table 7); the motivating-example monomial-set difference being exactly $\{s_1 s_2^2 s_3\}$; and the Case-2 recovery $X \circ\!-\!\circ Y$, $X \circ\!\to Z \leftarrow\!\circ Y$. The inference engine is present, runnable, and correct.
• Headline Table 1 numbers did not reproduce. Under the paper's stated protocol the reproduced "Ours" F1 was $0.86/0.61/0.72/0.60$ versus the reported $1.00/0.72/0.84/0.87$ — systematically $\sim$0.11–0.27 lower, with only Case 2 within one reported $\sigma$. Qualitatively the correct PPADMG is recovered on representative repeats, so the gap most plausibly reflects the from-scratch, strict endpoint-mark F1 the reproducer had to write (the paper's own metric, per Appendix C.1, is laxer) and figure-transcribed case topologies — but the exact numbers stand unverified.
• The comparative claim is untestable from the artifact. None of the four baselines (FCI, PB-SCM, GES, OCD), the F1 routine, or the ground-truth-PPADMG builder ship, so the "outperforms all baselines" half of the headline result and all of Tables 2–4 (additional cases and 10-/20-variable sweeps) could not be attempted. Reproducing any comparative cell requires re-implementing four external baselines and the full evaluation harness.
• Real-world results could not be attempted. Both datasets require authenticated download and no loader/preprocessing script maps raw rows to the count matrices used (Clicks/Conversions binning; per-match Y1/Y2/R/S aggregation; latent handling of Impressions/Foul). An independent reproduction would start from different counts.
• The 29-class enumeration and Figure 5 are not scriptable. graph2pgf/pgf2terms compute $M(G)$ per DAG (and matched representative classes exactly), but no enumeration/isomorphism-canonicalisation script emits the 29 classes, and no figure-generation code exists; the per-type monomial-set displays are also missing from the manuscript, so Theorem 3.13's case-by-case content cannot be checked as printed.
• Robustness tables have primitives but no driver. The utilities behind Tables 6–7 are shipped and reproduced, but no driver fixes seeds or runs the sweeps, and Table 7's evaluation point $\mathbf{z}$ is not stated in the code path.

Recommended Changes

Essential

• Fix the Type (d1) proof. Either rule out the "additional $L \to X_3$ with $X_1,X_3$ non-adjacent" case by a monomial argument on the other terms it would introduce, or weaken the concluded $X_1 \circ\!\to X_3$ orientation.
• Fix Type (d3) step 1. Add a monomial argument that positively excludes a direct latent edge $L \to X_2$ under the (d3) context, so "$X_2$ appearing once" genuinely implies $X_1 \to X_2$.
• Reconcile completeness Types 16/17/18. Exhibit the differing surrounding monomials that let presence (Type 16) and absence (Type 17) of $s_1^2 s_2^2 s_3$ both yield $X_1 \leftrightarrow X_2$, and correct the Type 18 citation (likely Theorem 3.8 via $s_2^2 s_3$) or supply the missing higher-order monomial.
• Restore the missing content. State Theorem 3.9's monomial conditions inline; restore the per-type monomial-set displays in the Type 1–29 completeness proof; and state the structural context of Types (e1)–(e5) inline so each rule is self-contained.
• Release the evaluation harness and disclose what is missing. Ship the four baselines (or wrappers plus commit hashes), the exact F1 metric, the ground-truth-PPADMG builder, the Case 1–9 graph specifications, and the random-graph drivers for the 10-/20-variable sweeps, so the headline Table 1 numbers and the comparative Tables 2–4 can be checked. Add an artifact note stating the baseline/metric/data layers are not otherwise included.
• Reconcile the reproduced Table 1 shortfall. Report results under the paper's own (documented) metric and confirm the $1.00/0.72/0.84/0.87$ cells regenerate; if the strict-vs-lax metric explains the gap, define the metric precisely and report both.

Suggested

• Add a misspecification stress test. Rerun the synthetic tables with negative-binomial noise or additive contamination to show the presence/absence monomial tests degrade gracefully — addressing the missing-robustness weakness.
• Specify the test protocol. State the bootstrap/threshold significance level, the evaluation point $\mathbf{z}$, and how error is controlled across the $O(d^3)$ simultaneous tests; note faithfulness reliance where absence-based rules (e3, e5, f1–f4) are stated.
• Provide real-world preprocessing scripts and strengthen the real experiments. Map the raw Kaggle/football files to the count matrices used, and add a real dataset with more observed variables so Phase 1 and the local 3-node rules are actually exercised (currently the shopping-mall case is a two-variable orientation and the football case is given the skeleton).
• Hold the evaluation metric and generative regime fixed, or label them. Justify the $\mu_i \in [0.001,0.005]$ / PADMG / compatible-circle changes in the random-graph experiments so cross-table comparison is valid.
• Document Phase 2's pair ordering. State whether both orderings of the asymmetric third derivative are tested and how conflicts are broken, and mirror this in the code; rename the phase to reflect that it detects confounded edges (bows), not all latent confounding.
• Soften the global-reconstruction claim to "build up a local estimate", and describe how overlapping-triple conflicts are resolved.
• Complete Theorem 3.8's sufficiency with the coefficient-non-vanishing/faithfulness invocation used in Theorem 3.7, and make the two intuition-based cruxes (Theorem 3.8 necessity; Type (c1) $R=X_2$) into formal arguments; fix the Type (c1) cross-reference (Theorem 3.7 $\to$ 3.8).
• Fix minor consistency issues. Harmonise LC-PBSCM/PBSCM; rewrite Des(i) as a directed-path condition; rename the local-PGF target set (e.g., $A$) and extend Definition 2.5's domain to $[d+r]$; state the main-text latent-mediator equivalence "after absorbing $L$'s Poisson noise".