{"id":"9bf7a285-9581-497b-88e2-e5730f1ec633","arxiv_id":"1908.05097","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"In heavy-tailed linear causal models, the causal tail coefficient is asymmetric along the true causal direction, and the EASE algorithm uses this asymmetry to recover the causal order consistently.","lead":"This paper introduces a causal tail coefficient that reads the direction of cause and effect from extreme events in heavy-tailed data. It also provides a fast, provably consistent algorithm for recovering the causal order, even when unobserved common causes are present.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The comparable-tail assumption in Eq. (4) is load-bearing: if noise slowly varying functions differ, Γ can equal 1 in both directions and Theorem 1's classification fails.","rationale":"The reader's weakest_assumption correctly identifies the comparable-tail condition as the point where the main argument is least secure. The paper proves Theorem 1 and Proposition 1 only under Eq. (4), where all noise variables share a common tail index and slowly varying function. The cancellation of ℓ(x) in Lemma 1 is exactly what makes the closed-form expression independent of the slowly varying components; without it, the ratio in Eq. (14) can be driven to 1 by a heavier-tailed ancestor, breaking the dichotomy in Table 1. The paper's own Section 4.3 demonstrates this for differing tail indices, and the same mechanism applies to differing slowly varying functions. This is a genuine limitation, but it is explicitly assumed and honestly discussed, not a hidden inconsistency. The central claim is correct within its stated model: the proofs of Lemma 1, Theorem 1, and Proposition 1 are detailed and mathematically sound, and the inconsistency the stress-test found in Proposition 2's proof is a minor gap (one should choose v as the argmax in the inequality), not a fatal flaw. The finite-sample choice of k_n is a practical concern but does not affect the asymptotic consistency claim. Since the reader already conditioned the verdict on this assumption and the minor proof gap, my read does not change the verdict; it remains CONDITIONAL.","tokens_in":40047,"tokens_out":24195,"duration_ms":208793,"concrete_test":"Analytically compute Γ12 and Γ21 for the two-variable SCM X1 = ε1, X2 = β ε1 + ε2, where ε1 and ε2 are independent with common tail index α but slowly varying functions ℓ1(x) = log(x) and ℓ2(x) = 1. Using Eq. (14), verify that Γ12 = 1 and Γ21 = 1 for any β > 0 and α > 0. If both limits equal 1, the classification in Theorem 1 is violated for a directed causal pair, confirming that the comparable-tail assumption is load-bearing. As a complementary simulation, generate n = 10^5 observations from this SCM with α = 1.5 and β = 1, estimate Γ with k = n^0.4, and run EASE; the estimated order should be essentially random between the two variables, showing the practical failure when the assumption is relaxed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central identifiability result (Theorem 1) and the correctness of EASE (Proposition 1) both rest on Lemma 1, whose closed form requires cancellation of the common slowly varying function ℓ(x) in Eq. (4). If the noise variables have the same tail index α but non-comparable slowly varying functions, the ratio in Eq. (14) need not converge to the expression in Lemma 1, and Γ can be 1 in both directions for a directed relationship. Concretely, take X1 = ε1 and X2 = β ε1 + ε2 with P(ε1 > x) ∼ log(x) x^{-α} and P(ε2 > x) ∼ x^{-α}. For Γ12, the numerator and denominator are identical, so Γ12 = 1. For Γ21, the denominator includes the term from ε2, but the numerator includes only the term from ε1 as a common ancestor; because log(x) x^{-α} dominates x^{-α}, the ratio tends to 1, so Γ21 = 1. Thus the pair (Γ12, Γ21) = (1, 1) occurs, which Table 1 declares impossible under the assumptions. This is not merely a technical boundary: the paper's own financial application (Section 5.2) reports estimated tail indices that differ (0.31 vs. 0.25 vs. 0.16), and its Section 4.3 concedes that a heavier-tailed cause or confounder masks the causal signal. Therefore, the equal-tail/comparable-slowly-varying assumption is the least secure condition on which the main claim depends, and it is not satisfied in some of the paper's own examples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a causal tail coefficient Γ_{jk} that measures extremal dependence between two variables, and shows that for linear structural causal models with positive coefficients and comparable regularly varying noise (same tail index and common slowly varying function), the pair (Γ_{12}, Γ_{21}) identifies whether X1 causes X2, X2 causes X1, there is a common cause, or no causal link. Based on this, the authors propose the EASE algorithm, a greedy procedure that recovers a causal order of the underlying DAG from the matrix of pairwise coefficients. They prove consistency of a nonparametric estimator of Γ, establish the correctness of EASE in the population case and in finite samples, extend the results to real-valued coefficients and to hidden confounders, and discuss the consequences of relaxing the equal-tail assumption. The theoretical results are supported by simulations and two real-data applications, and an R package is provided.","tokens_in":40404,"tokens_out":15438,"duration_ms":134054,"significance":"If the results hold, the paper makes a valuable connection between extreme value theory and causal discovery, offering a computationally efficient method that can exploit tail information and that is robust to hidden confounders under the stated model. The closed-form expression for Γ in Lemma 1, the classification in Theorem 1, and the consistency result in Theorem 2 are elegant and constitute a substantial contribution. The paper is also careful to state the assumptions and to discuss failure modes when tails are not comparable, which increases its credibility. The main value lies in the population-level identifiability result and the accompanying algorithm, which may open new directions for causal analysis of extreme events.","major_comments":[{"comment":"The proof of Lemma 1 defines c = β_{h→j}/β_{h→k} and then uses the event {β_{h→k}ε_h > cx, X_k > cx} as a lower bound for {β_{h→j}ε_h > x}. With this choice of c, the inclusion {β_{h→k}ε_h > cx, X_k > cx} ⊆ {β_{h→j}ε_h > x} holds only if β_{h→j} ≥ β_{h→k}; in general it is false, so the inequality E[F_k(X_k) 1_{β_{h→j}ε_h > x}] ≥ F_k(cx) P(β_{h→k}ε_h > cx, X_k > cx) is not justified. The argument can be repaired by taking c = β_{h→k}/β_{h→j}, which makes the event inclusion exact. Since Lemma 1 underpins Theorem 1, Theorem 3, and Proposition 1, this correction is load-bearing and should be made.","section":"Appendix B.1 (proof of Lemma 1)"},{"comment":"In the proof of Proposition 2, the sentence \"Let v ∈ V~\\{j}, and note that v /∈ An(j,G). Thus, by (35), it follows that Γ̂_{vj} − Γ̂_{ui} ≥ 0\" is unjustified for an arbitrary v. One must take v ∈ argmax_{u'∈V~\\{j}} Γ̂_{u'j} so that Γ̂_{vj} equals the max-score for node j, which is at least the max-score for node i by the defining property (35) of i; the inequality then follows. Please add this specification, as the current text leaves the proof incomplete.","section":"Appendix B.5 (proof of Proposition 2)"},{"comment":"The identifiability result in Theorem 1 requires not merely a common tail index α but comparable slowly varying functions ℓ in condition (4). The abstract's statement that the result \"holds even in the presence of latent common causes that have the same tail index as the observed variables\" is weaker than the actual assumption and is therefore potentially misleading. For example, with P(ε1 > x) ∼ log(x)x^{−α} and P(ε2 > x) ∼ x^{−α}, the pair (Γ12, Γ21) = (1,1) can occur when X1 causes X2, as the paper's own Section 4.3 predicts. Please state the full comparability condition in the abstract and explicitly mark Theorem 1 as valid under Eq. (4).","section":"Section 2.1, Eq. (4); abstract"}],"minor_comments":[{"comment":"The statement that the estimated shape parameters \"seem reasonable\" because their confidence intervals overlap is informal; the later acknowledgment that NESN and NOVN have slightly lighter tails is more accurate. Consider presenting a more explicit check of the comparability assumption (4) for this dataset.","section":"Section 5.2"},{"comment":"The sentence \"However, this does not affect the correctness of the EASE algorithm, on the population level\" is immediately qualified by Example 2, where EASE can make mistakes when Γjk = Γkj = 1 for an ancestor-descendant pair. Rephrase to reflect the conditional nature, for example by saying \"apart from the case in Example 2\".","section":"Section 4.3"},{"comment":"The EASE algorithm description does not specify tie-breaking when multiple nodes attain the minimal score M_i^{(s)}. Adding a sentence that ties are broken arbitrarily would make the algorithm fully deterministic.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the central idea is promising. The proof of Lemma 1 contains a genuine but easily fixable error in the choice of the constant c, and the proof of Proposition 2 has a missing maximizer specification. Both issues are load-bearing because they support the main identifiability and consistency statements, so I recommend major revision rather than rejection. The comparability assumption should also be stated more carefully in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers what it promises: a new causal tail coefficient, a closed-form expression for it under a heavy-tailed linear SCM, and an identifiability theorem that works even with latent common causes sharing the same tail index. The EASE algorithm is simple, fast, and consistent, and the proofs are mostly careful. The two real-data applications are a bonus, and the R package makes the method easy to try. This is a real contribution to the causal-discovery-from-extremes gap.\n\nThe main soft spot is the comparability assumption in Eq. (4). The closed form in Lemma 1 and the classification in Theorem 1 require every noise variable to have the same slowly varying function. If that fails, the ratios in (14) need not converge to the claimed values, and (1,1) becomes possible for a directed cause, breaking the table. The stress-test example with log(x)x^{-α} versus x^{-α} is correct and not exotic. To the paper's credit, Section 4.3 and the remark after Table 1 openly discuss this limitation, and the financial application even finds tail index estimates that differ across variables. So the authors are not hiding it, but the central claim is more fragile than the abstract suggests. The theory is correct under its stated assumptions, and the paper is clear about those assumptions; still, a reader should not come away thinking the method works whenever the data are heavy-tailed.\n\nA smaller issue: the proof of Proposition 2 does not explicitly choose v as the maximizer of the corresponding score for node j. Without that choice, the inequality Γ̂_vj ≥ Γ̂_ui does not follow. This is a minor fix—just pick v to be an argmax—but as written it is a gap. The choice of k_n via a preliminary simulation is practical and not circular, since the consistency proof allows any intermediate sequence; it is just a slightly ad hoc part of the empirical procedure.\n\nOverall, this is a solid paper with a clear contribution and honest limitations. It deserves a serious referee and, after a small proof repair, publication. I would cite it if working in causal discovery or extreme value methods, and I would send it to peer review without hesitation.","headline":"A genuinely new tail-based causal discovery method with clean population-level theory, but the equal-tail-index assumption is doing more work than the paper lets on.","tokens_in":40866,"tokens_out":3556,"would_cite":true,"duration_ms":36059,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62H22","62D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A tail-asymmetry coefficient reveals which variable causes which in heavy-tailed systems, and a greedy algorithm built on it recovers the causal order consistently.","keywords":["causal discovery","heavy-tailed distributions","regular variation","extreme value theory","structural causal models","causal tail coefficient","EASE algorithm","causal order"],"falsifier":"Simulate a two-variable heavy-tailed linear SCM with $X_1=\\varepsilon_1$ and $X_2=\\varepsilon_1+\\varepsilon_2$, where $\\varepsilon_1$ and $\\varepsilon_2$ are independent Pareto variables with the same tail index, and compute the population limits $\\Gamma_{12}$ and $\\Gamma_{21}$; Theorem 1 predicts $\\Gamma_{12}=1$ and $\\Gamma_{21}\\in(1/2,1)$, and any other pair of values would refute the identification claim in its own setting.","tokens_in":39878,"feed_emoji":"📈","tokens_out":9904,"duration_ms":91969,"temperature":0.7,"pith_summary":"The paper establishes that in heavy-tailed systems the direction of causation can be read from the asymmetry of a simple tail expectation, and that a greedy algorithm based on this signal recovers the causal order of the underlying graph. The central object is the causal tail coefficient $\\Gamma_{jk}$, the limiting average rank of $X_k$ given that $X_j$ exceeds a high quantile. Under a linear structural causal model whose independent noise variables are regularly varying with comparable tails, the paper proves $\\Gamma_{jk}=1$ when $X_j$ causes $X_k$, $\\Gamma_{jk}\\in(1/2,1)$ when the variables share a common cause, and $\\Gamma_{jk}=1/2$ when there is no causal link. A consistent nonparametric estimator and the $O(p^2)$ extremal ancestral search (EASE) algorithm turn this into a practical discovery method that also handles hidden common causes and real-valued coefficients, and the paper demonstrates it on simulated data, financial returns, and river discharge records.","feed_headline":"Extreme tail values reveal the true causal direction","feed_subtitle":"A tail-based coefficient and a fast greedy algorithm recover the causal order of heavy-tailed graphs.","key_machinery":"The load-bearing object is the causal tail coefficient $\\Gamma_{jk}=\\lim_{u\\to 1^{-}}\\mathbb{E}\\,[F_k(X_k)\\mid F_j(X_j)>u]$, together with the greedy algorithm EASE that consumes the matrix of pairwise coefficients. The coefficient works because, under regular variation, an extreme value of $X_j$ is produced almost entirely by one dominant noise term; max-sum equivalence lets the conditional expectation be written as a ratio of sums of $\\alpha$-powered path coefficients, so the asymmetry $\\Gamma_{jk}-\\Gamma_{kj}$ signals which direction the tail effect flows. EASE identifies source nodes one at a time, at each step choosing the node $i$ that minimizes $\\max_{j\\neq i}\\Gamma_{ji}$, and runs in $O(p^2)$.","core_discovery":"On its own terms, the paper's discovery is that extremal dependence in a heavy-tailed linear SCM is asymmetric in a way that exactly mirrors ancestry. For a model $X_j=\\sum_k \\beta_{jk}X_k+\\varepsilon_j$ with positive coefficients and $\\varepsilon_j$ regularly varying with a common index $\\alpha$, the causal tail coefficient has closed form $\\Gamma_{jk}=\\frac12+\\frac12\\,\\frac{\\sum_{h\\in A_{jk}}\\beta_{h\\to j}^{\\alpha}}{\\sum_{h\\in \\mathrm{An}(j,G)}\\beta_{h\\to j}^{\\alpha}}$, where $A_{jk}$ is the set of common ancestors of $X_j$ and $X_k$. Theorem 1 turns this into a decision table: if $X_1$ causes $X_2$ then $\\Gamma_{12}=1$ and $\\Gamma_{21}$ lies strictly between $1/2$ and $1$; if neither causes the other and there is no common cause, both coefficients equal $1/2$; if a hidden node causes both, both lie strictly between $1/2$ and $1$. Proposition 1 shows that the EASE algorithm, which repeatedly removes the node with the smallest maximum incoming coefficient, outputs a causal order of the DAG, and Corollary 1 extends this to consistency from finitely many samples under the von Mises condition.","pith_inferences":["Beyond the paper: the same tail-asymmetry signal can serve as a diagnostic for whether bulk and tail causality diverge; comparing $\\Gamma$ with a non-extremal method's output on the same data would test the extremal mechanism hypothesis directly.","Beyond the paper: plugging an estimated tail index $\\alpha$ into the closed-form ratio gives a stability check for $\\Gamma_{jk}-\\Gamma_{kj}$, and a data-adaptive rule for choosing $k_n$ is a natural next step that the paper leaves open.","Beyond the paper: because $\\Gamma$ is invariant under strictly monotone marginal transformations, EASE should transfer to ordinal or censored heavy-tailed data, a setting the paper does not explore.","Beyond the paper: for time series, applying EASE to vector autoregression residuals would yield a tail-based Granger test, following the paper's own suggestion to combine the two approaches."],"forward_implications":["A causal order among $p$ heavy-tailed variables can be recovered from pairwise tail coefficients alone, without conditioning on other variables or knowing the graph, and the recovery runs in $O(p^2)$ time.","The nonparametric estimator is consistent for any intermediate threshold $k_n\\to\\infty$ with $k_n/n\\to 0$, so the method needs no parametric model of the noise beyond comparable regular variation.","Hidden common causes do not break the result: as long as the confounder's tail is no heavier than the observed variables', EASE returns a correct causal order of the observed subgraph.","With real-valued coefficients, the symmetric two-tail coefficient $\\Psi$ extends the same identification guarantee, so causal discovery applies to signed linear effects.","Empirically, setting $k_n\\approx n^{0.4}$ works across tail indices, and EASE is orders of magnitude faster than the non-extremal baselines it is compared with."],"supporting_citations":[{"why":"Defines linear structural causal models, directed paths, and ancestor terminology on which Theorem 1 and Proposition 1 rest.","marker":"[36]"},{"why":"Supplies regular variation and max-sum equivalence results used to derive the closed-form expression for the causal tail coefficient.","marker":"[11]"},{"why":"Provides the tail empirical process results and von Mises condition used in the consistency proof of the estimated coefficient.","marker":"[10]"},{"why":"Gives the underlying two-sample max-sum equivalence that Lemma 3 of the paper extends to $p$ variables.","marker":"[16]"},{"why":"Defines the LiNGAM causal discovery paradigm and serves as the principal non-extremal baseline the paper builds on and compares against.","marker":"[47]"},{"why":"Supplies the pairwise likelihood-ratio LiNGAM baseline used in the simulations and financial application.","marker":"[24]"}],"fun_headline_variants":["Heavy tails unveil causal direction","Tail coefficients decode causal order","Extremes expose causal links","Causal discovery from extreme events","Unmasking hidden causes via heavy tails"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that all noise variables in the system have tails of essentially the same thickness: the same tail index $\\alpha$ and comparable slowly varying functions, so that no single ancestor or common cause can dominate the tail signal in the wrong direction.","fun_headline_variants_meta":{"raw":{"variants":["Heavy tails unveil causal direction","Tail coefficients decode causal order","Extremes expose causal links","Causal discovery from extreme events","Unmasking hidden causes via heavy tails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2071,"prompt_tokens":982,"completion_tokens":1089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1031}},"tokens_in":598,"tokens_out":1089,"duration_ms":9221,"temperature":1.0,"reasoning_tokens":1031,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:24:43.206836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a two-variable heavy-tailed linear SCM with $X_1=\\varepsilon_1$ and $X_2=\\varepsilon_1+\\varepsilon_2$, where $\\varepsilon_1$ and $\\varepsilon_2$ are independent Pareto variables with the same tail index, and compute the population limits $\\Gamma_{12}$ and $\\Gamma_{21}$; Theorem 1 predicts $\\Gamma_{12}=1$ and $\\Gamma_{21}\\in(1/2,1)$, and any other pair of values would refute the identification claim in its own setting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines linear structural causal models, directed paths, and ancestor terminology on which Theorem 1 and Proposition 1 rest."},{"cited_title":"Embrechts, C","cited_arxiv_id":null,"evidence_quote":"Supplies regular variation and max-sum equivalence results used to derive the closed-form expression for the causal tail coefficient."},{"cited_title":"de Haan and A","cited_arxiv_id":null,"evidence_quote":"Provides the tail empirical process results and von Mises condition used in the consistency proof of the estimated coefficient."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the underlying two-sample max-sum equivalence that Lemma 3 of the paper extends to $p$ variables."},{"cited_title":"Shimizu, P","cited_arxiv_id":null,"evidence_quote":"Defines the LiNGAM causal discovery paradigm and serves as the principal non-extremal baseline the paper builds on and compares against."},{"cited_title":"Pairwise likelihood ratios for estimation of non-gaussian structural equation models","cited_arxiv_id":null,"evidence_quote":"Supplies the pairwise likelihood-ratio LiNGAM baseline used in the simulations and financial application."}],"review_version":1}