{"id":"1378f1c1-7632-4994-8c7e-05dc777a0c7f","arxiv_id":"2507.07898","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SyPI+ extends SyPI to nonlinear time series via distance-correlation lag selection and a pruning step, reporting better F-score and SHD than PCMCI and CD-NOTS on synthetic data.","lead":"This paper presents SyPI+, an algorithm for finding cause-and-effect relationships in nonlinear time series data that needs far fewer statistical tests than existing methods. It matters because faster, scalable causal discovery could help finance, climate, and healthcare applications with limited data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing gap is that §2.2's distance-correlation lag is not shown to equal the graph-theoretic minimum lag w required by Theorem 2.1, so the two CI tests may be run at wrong lags.","rationale":"SyPI+'s claimed advantage is reducing the number of CI tests to two per candidate, but those two tests are only valid if the time alignment in Theorem 2.1 holds. The theorem from Mastakouri et al. is the only soundness result in the paper, and it is conditional on w_i being the graph-theoretic minimum lag. The paper explicitly says this identification is 'very important for ensuring correct causal connections', yet no formal or empirical bridge is supplied. The pruning step (§2.3) is also under-specified and unproven after relaxing A6, so it is a second risk; however, the minimum-lag step is more clearly load-bearing because it corrupts the conditioning set before pruning. My concern matches the reader's weakest assumption, and the requested ablation (oracle w vs. selected w) would settle it. Other issues (no code, no error bars, no SyPI baseline) affect reproducibility but not the logical core as directly. I therefore keep the reader's CONDITIONAL verdict; the concern is not new, but it is real and should be a stated acceptance condition.","tokens_in":5864,"tokens_out":8275,"duration_ms":89662,"concrete_test":"On the §3 synthetic graphs, compute each ordered pair's true minimum lag w_i* from the ground-truth graph using the collider-free-path definition and compare it with the lag selected by the §2.2 distance-correlation procedure; report agreement rates per graph size and sample size. Then re-run Algorithm 1 with w_i* substituted for the selected lags and compare F-score/SHD to the published pipeline. If the two variants differ on any graph, the distance-correlation lag is not the w required by Theorem 2.1, and the correctness/efficiency claims are not established by the current experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central correctness claim relies on Theorem 2.1, whose conditioning sets are defined in terms of graph-theoretic minimum lags: for each pair (X^i, Y), w_i is the smallest lag over collider-free paths from X^i to Y. Section 2.2 replaces that object by the lag(s) with maximum marginal distance correlation above an arbitrary 0.05 threshold, and Algorithm 1 feeds this estimate into w = min lags(R, Y) and into S_i = {X^j_{t+w_i-w_j-1}}. Marginal distance correlation is not an estimator of the collider-free-path minimum lag. A short indirect or confounded path can make a non-causal lag look most dependent, and then the time indices in equations (2.1)-(2.2) are wrong; the premises no longer imply X^i_t -> Y_{t+w_i}. The paper gives neither a proof nor an identifiability condition connecting the selected lag to w_i, nor an ablation showing the choice does not matter. Since every downstream edge claim passes through these two tests, this is the most load-bearing unverified step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SyPI+, a constraint-based causal discovery algorithm for nonlinear autoregressive time series. The method extends the SyPI algorithm of Mastakouri et al. (2021) by replacing Lasso-based minimum-lag selection with distance-correlation lag plots and by adding a pruning step intended to relax the sink-node assumption A6. The central claims are that SyPI+ requires only a quadratic number of conditional independence tests, matches or outperforms PCMCI and CD-NOTS in accuracy on synthetic data, and works particularly well in low-data regimes. The paper also presents a case study on weekly CDS data of major banks before the 2008 financial crisis.","tokens_in":6172,"tokens_out":4466,"duration_ms":49659,"significance":"If the method's central correctness claims held, SyPI+ would be a practically valuable contribution: it targets a genuine scalability bottleneck in constraint-based time-series causal discovery, and the synthetic evaluation covers multiple graph sizes, sample sizes, and CI tests. The paper builds on a known theorem rather than proving a new identifiability result, and it does not ship code, data, or machine-checked proofs, so the empirical contribution is not directly reproducible. The main issue is that two load-bearing components of the algorithm—the distance-correlation lag-selection step and the pruning step—are not specified precisely enough to support the method's correctness claims, and no proof or strong ablation connects them to the theorem the paper relies on.","major_comments":[{"comment":"Equations (2.1) and (2.2) are valid only when w_i is the graph-theoretic minimum lag defined in Section 2.1: the smallest lag over collider-free paths from X^i to Y. Section 2.2 replaces this object with the lags that maximize marginal distance correlation above a 0.05 threshold, but the paper provides no proof or identifiability condition that these quantities coincide. A short indirect or confounded path can make a non-causal lag look most dependent in marginal distance correlation, and then the time indices in S_i and in the two CI tests are wrong, so the implication 'X^i_t -> Y_{t+w_i}' does not follow from Theorem 2.1. The manuscript needs either a theorem linking the estimated lag to the graph-theoretic minimum lag under stated assumptions, or an ablation demonstrating that the method is insensitive to this choice.","section":"Section 2.2 / Theorem 2.1"},{"comment":"Algorithm 1 is internally inconsistent as written. It computes a single scalar w = min lags(R, Y), but then defines S_i using per-pair quantities w_i and w_j that are never defined in the pseudocode. Moreover, Section 2.2 says that when multiple lags have similarly high dependency values, 'we retain all of them', yet Theorem 2.1 and the conditioning-set construction require one minimum lag per pair. The paper must specify exactly what min lags returns, how multiple retained lags are combined (e.g., multiple tests, union of conditioning sets, or an additional selection rule), and how that interacts with equations (2.1) and (2.2). Without this, the algorithm is not reproducible.","section":"Algorithm 1 / Section 2.1"},{"comment":"The pruning step is described only in prose using 'simple cycles' and 'common ancestors', with no precise algorithm, no explicit condition set, no pseudocode, and no proof that it restores correctness when Assumption A6 is violated. Since Theorem 2.1 explicitly assumes A6 and the paper's stated contribution is to relax that assumption via pruning, the pruning step is load-bearing for the method's soundness. As written, it is not a well-defined procedure and cannot be evaluated, replicated, or proven to control false positives in a principled way.","section":"Section 2.3"},{"comment":"The claims of superiority over PCMCI and CD-NOTS, especially in low-data regimes, are based on average F-scores and SHD values but the paper reports no standard deviations, confidence intervals, or per-graph variability. With 50 random graphs per node count, some measure of dispersion is necessary to assess whether the reported differences are meaningful. The manuscript should include error bars or per-condition tables, and ideally make the synthetic data generation and evaluation code available.","section":"Section 3.3"}],"minor_comments":[{"comment":"'B¨ uhlman' in the reference for Kalisch and Buehlmann should be 'Bühlmann'.","section":"References"},{"comment":"The sentence 'The former attempt find relationships' is missing the infinitive marker 'to' and should read 'The former attempt to find relationships'.","section":"Section 1"},{"comment":"The notation 'S_nvars j=1,j≠i' is ambiguous; please use standard set-builder notation with explicit indices, and define w_i for all variables appearing in the conditioning set.","section":"Algorithm 1"},{"comment":"The 0.05 threshold for distance correlation is presented without justification or sensitivity analysis; the paper would benefit from reporting how results change with this threshold.","section":"Section 2.2"},{"comment":"The lag-plot description says the second column 'reveals' that the only potential cause is X^0_{t-2}, but the decision rule for 'significant dependency' from a lag plot is not defined; the figure alone is not a reproducible criterion.","section":"Section 2.2 / Figure 2"},{"comment":"The paper states the maximum lag allowed in experiments was five, but it is not clear whether this is the user-specified T in Section 2.2; please clarify the relationship between the experiment's maximum lag and the algorithm's parameter.","section":"Section 3"},{"comment":"The case study is explicitly framed as an illustrative application and the paper disclaims accuracy of the discovered links; this is appropriate, but the framing 'Should Lehman Brothers have been saved?' promises more than the analysis can deliver and should be softened.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a preliminary working draft: Algorithm 1 has an undefined variable, the pruning section is entirely prose, several figures are referenced but not included in the text I received, and there is no code or data release. The central theoretical gap—the unproven equivalence between distance-correlation-selected lags and the graph-theoretic minimum lag of Theorem 2.1—is serious. I would recommend a major revision that provides a precise algorithm specification, a proof or strong empirical validation of the lag-selection equivalence, a precise pruning procedure with justification, and reproducible experimental details. I would not recommend rejection on novelty grounds, as the core idea of extending SyPI to nonlinear settings with reduced test counts is reasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things upfront. First, this is a genuinely useful incremental paper: it takes SyPI's two-test-per-cause trick and tries to make it work for nonlinear time series, with a wide synthetic evaluation across sample sizes, CI tests, and graph sizes. Second, the two places where it departs from SyPI — lag selection via distance correlation and the pruning step — are both underspecified, and the first one is load-bearing.\n\nWhat's actually new: distance-correlation-based minimum-lag detection for nonlinear dependencies, a pruning heuristic meant to relax SyPI's sink-node assumption, and an extensive empirical comparison against PCMCI and CD-NOTS. The quadratic-vs-exponential test count is inherited from SyPI, so the central complexity claim is not new. The experiments are better than average for this area, and the paper is clearly written. They also include an honest caveat in the case study, acknowledging unmeasured confounders.\n\nThe soft spots, in order of severity. The main theoretical gap: Section 2.2 selects lags by marginal distance correlation with an arbitrary 0.05 threshold, and Algorithm 1 feeds those into the conditioning sets of Theorem 2.1. But that theorem's w_i is a graph-theoretic object — the minimum lag over collider-free paths. No proof, or even a stated identifiability condition, connects the two. If distance correlation picks a different lag, the time indices in the CI tests are wrong and the causal conclusion does not follow. This is not a minor detail; every edge claim passes through it. Second, the pruning in Section 2.3 is prose only. There is no precise algorithm, no condition set, and no proof that pruning restores soundness when A6 is violated. Third, the empirical section compares against two baselines but omits the original SyPI, which is the natural reference for this exact extension. There is also no code or data release, and the F-score/SHD figures have no error bars.\n\nTo be fair, the paper is not sloppy or dishonest. It is a reasonable engineering contribution that leans on a solid theorem. But the load-bearing pieces need work. I would send it to peer review, not desk-reject it, because the topic is legitimate and the empirical scope is substantial. I would ask for a major revision: identify the distance-correlation lag as an estimator of w_i under explicit assumptions, or run a careful ablation; formalize the pruning algorithm; add the SyPI baseline; release code and data; and add error bars. If they can close the identifiability gap, this becomes a paper I'd want in the literature. As is, I'd treat its causal claims cautiously.","headline":"Plausible incremental extension of SyPI to nonlinear series, but the lag-selection step is unproven and the pruning is underspecified; worth a serious look if those gaps can be closed.","tokens_in":6620,"tokens_out":2313,"would_cite":false,"duration_ms":28772,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SyPI+ extends SyPI to nonlinear time series with only quadratic growth in tests, matching or beating PCMCI and CD-NOTS in accuracy, especially with little data.","keywords":["causal discovery","time series","constraint-based learning","conditional independence tests","nonlinear autoregressive series","distance correlation","minimum lag selection","graph pruning"],"falsifier":"Run SyPI+ on synthetic autoregressive graphs with known ground-truth minimum lags, arranging a slow common cause so that a spurious early lag has high distance correlation; then compare the selected $w$ to the truth and check whether the two-test rule still produces the correct edges. If choosing the wrong lag changes the output graph materially, the distance-correlation selection step is the breaking point.","tokens_in":5695,"feed_emoji":"📈","tokens_out":12989,"duration_ms":119914,"temperature":0.7,"pith_summary":"This paper tries to establish that causal structure learning for nonlinear autoregressive time series can be made much cheaper without sacrificing accuracy. The proposed algorithm, SyPI+, builds on the SyPI condition-set construction so that each candidate cause is checked with only two conditional-independence tests (statistical checks of dependence after conditioning on other variables); across a network this makes the total number of tests grow quadratically with the number of variables rather than exponentially. To handle nonlinearity, the paper replaces Lasso-based lag selection with distance-correlation lag plots, and it adds a pruning step to remove spurious edges, which also lets the method relax the original sink-node assumption. In synthetic experiments on graphs with 3 to 15 nodes and 50 to 1,000 observations, the paper reports that SyPI+ matches or beats PCMCI and CD-NOTS in F-score and structural Hamming distance, and that its advantage is largest when data are scarce.","feed_headline":"Causal discovery: tests now grow quadratically, not exponentially","feed_subtitle":"A new method matches PCMCI and CD-NOTS accuracy with far fewer tests, and wins in low-data settings.","key_machinery":"The load-bearing mechanism is SyPI's constructed conditioning set plus its two-test criterion (Theorem 2.1 of the original SyPI method). For each ordered pair of time series $(X^i, Y)$, with minimum lag $w_i$, the conditioning set is $S_i = \\{X^j_{t+w_i-w_j-1}: j \\neq i\\}$, and the algorithm performs two conditional-independence tests: a dependence test of $X^i_t$ on $Y_{t+w_i}$ given $S_i$ and $Y_{t+w_i-1}$, and an independence test of $X^i_{t-1}$ on $Y_{t+w_i}$ given $S_i$, $X^i_t$, and $Y_{t+w_i-1}$. SyPI+ supplies the $w_i$ values with distance-correlation lag plots instead of Lasso, which is what adapts the scheme to nonlinear dependencies, and it closes the pipeline with pruning of simple cycles and common-ancestor-induced edges to compensate for relaxing the sink-node assumption.","core_discovery":"The central claim is that the two-test-per-candidate scheme of SyPI, originally designed for linear time series with a sink target, can be extended to nonlinear autoregressive series. The extension replaces Lasso-based minimum-lag estimation with thresholded distance-correlation lag plots and appends a pruning step for simple cycles and common ancestors, which in turn lets the method relax the assumption that the target time series has no children. Under the stated assumptions, the paper argues, the two conditional-independence tests in Theorem 2.1 are still sufficient: if $X^i_t$ is dependent on $Y_{t+w_i}$ given the constructed set $S_i$ and $Y_{t+w_i-1}$, and $X^i_{t-1}$ is independent of $Y_{t+w_i}$ given those plus $X^i_t$, then $X^i_t$ causes $Y_{t+w_i}$. The result is an algorithm whose test count grows quadratically rather than exponentially, with synthetic F-scores and structural Hamming distances that are competitive with or better than PCMCI and CD-NOTS, especially in low-data settings.","pith_inferences":["The paper does not bound the number of lags that survive the distance-correlation threshold; if dependency values are flat across lags, the retained set may grow and the quadratic test-count guarantee would degrade, so the scaling claim is best read as conditional on well-separated lag dependencies.","The pruning step is a heuristic that removes edges whose removal improves fit to conditional independencies for cycles and common ancestors, and the paper does not prove that pruning preserves the soundness of Theorem 2.1, so users should expect some recall loss in dense graphs.","Because SyPI+ runs only two tests per candidate pair, its low-data advantage is likely a statistical-power effect: fewer tests mean less multiple-testing penalty. If so, pairing SyPI+ with a stronger nonlinear CI test should carry the advantage into moderate sample sizes.","A natural stress test beyond the paper's experiments is to apply the same distance-correlation lag selection to nonstationary or regime-switching series; the authors list non-stationarity as future work, and we would expect the two-test logic to need an additional lag-selection check when dependencies shift over time."],"forward_implications":["The number of conditional-independence tests scales quadratically with the number of variables rather than exponentially, so the method is practical for constraint-based discovery on networks larger than earlier exhaustive approaches allow.","In synthetic comparisons, SyPI+ achieves higher F-scores than PCMCI and CD-NOTS when samples are scarce (around 50 to 150 time points), suggesting limited observational data can still support causal claims.","For small and mid-sized graphs, SyPI+'s recovered graphs are closer to ground truth (lower SHD) than the baselines, with CD-NOTS slightly better only at the largest tested size.","The method copes with linear and nonlinear functional forms (quadratic, exponential, sine), so it extends the range of data types covered by constraint-based time-series discovery.","The bank CDS case study illustrates a use for policymakers: causal links from Lehman Brothers to Goldman Sachs, Morgan Stanley, and Bank of America appear after removing systematic market factors."],"supporting_citations":[{"why":"Provides the SyPI algorithm and the two-test theorem (Theorem 2.1) that SyPI+ inherits as its identification criterion.","marker":"(Mastakouri et al., 2021)"},{"why":"Defines PCMCI, the principal baseline compared in the synthetic experiments, and the nearest-neighbor conditional mutual information estimator used as CMIknn.","marker":"(Runge, 2018)"},{"why":"Defines CD-NOTS, the second baseline against which F-score and SHD are measured.","marker":"(Sadeghi et al., 2024)"}],"fun_headline_variants":["Causal discovery: tests now grow quadratically, not exponentially","From exponential to quadratic: faster causal discovery","Nonlinear time series causal discovery gets scalable","Quadratic test count enables low-data causal inference","Causal discovery: fewer tests, better low-data performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method must choose the same minimum lag $w$ that the SyPI theorem requires, and it chooses that lag by thresholding distance-correlation values; if the highest-dependence lag is not the true causal minimum lag, the two conditional-independence tests in Algorithm 1 stop identifying causes.","fun_headline_variants_meta":{"raw":{"variants":["Causal discovery: tests now grow quadratically, not exponentially","From exponential to quadratic: faster causal discovery","Nonlinear time series causal discovery gets scalable","Quadratic test count enables low-data causal inference","Causal discovery: fewer tests, better low-data performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1651,"prompt_tokens":839,"completion_tokens":812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":737}},"tokens_in":455,"tokens_out":812,"duration_ms":9024,"temperature":1.0,"reasoning_tokens":737,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:29:36.230128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SyPI+ on synthetic autoregressive graphs with known ground-truth minimum lags, arranging a slow common cause so that a spurious early lag has high distance correlation; then compare the selected $w$ to the truth and check whether the two-test rule still produces the correct edges. If choosing the wrong lag changes the output graph materially, the distance-correlation selection step is the breaking point.","supporting_citations":[{"cited_title":"Necessary and sufficient conditions for causal feature selection in time series with latent common causes","cited_arxiv_id":null,"evidence_quote":"Provides the SyPI algorithm and the two-test theorem (Theorem 2.1) that SyPI+ inherits as its identification criterion."},{"cited_title":"Conditional independence testing based on a nearest-neighbor estimator of con- ditional mutual information","cited_arxiv_id":null,"evidence_quote":"Defines PCMCI, the principal baseline compared in the synthetic experiments, and the nearest-neighbor conditional mutual information estimator used as CMIknn."},{"cited_title":"Causal discovery from nonstationary time series","cited_arxiv_id":null,"evidence_quote":"Defines CD-NOTS, the second baseline against which F-score and SHD are measured."}],"review_version":1}