Efficient Causal Discovery for Autoregressive Time Series

Achintya Gopal; Mohammad Fesanghary

REVIEW 4 major objections 7 minor 10 references

Efficient Causal Discovery for Autoregressive Time Series

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2507.07898 v1 pith:LAUGMXNR submitted 2025-07-10 cs.LG stat.AP

Mohammad Fesanghary , Achintya Gopal This is my paper

classification cs.LGstat.AP

keywords causaldiscoverytimeseriesconstraint-basedlearningconditionalindependencetestsnonlinearautoregressivedistancecorrelationminimumlagselectiongraphpruning

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 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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

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

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

What'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.

The 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.

To 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.

Referee Report

4 major / 7 minor

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.

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 (4)

[Section 2.2 / Theorem 2.1] 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.
[Algorithm 1 / Section 2.1] 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 2.3] 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 3.3] 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.

minor comments (7)

[References] 'B¨ uhlman' in the reference for Kalisch and Buehlmann should be 'Bühlmann'.
[Section 1] The sentence 'The former attempt find relationships' is missing the infinitive marker 'to' and should read 'The former attempt to find relationships'.
[Algorithm 1] 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 2.2] 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 2.2 / Figure 2] 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 3] 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 4] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SyPI+ builds on an external theorem (Mastakouri et al. 2021); the unproven lag-selection step is a correctness gap, not a definitional circle.

full rationale

The paper's derivation chain is not circular. Theorem 2.1 is imported from Mastakouri et al. (2021), an external source, and is stated as a theorem with proof in that prior work; the present paper does not define the graph-theoretic minimum lag in terms of its own outputs. Section 2.2 replaces that theoretical quantity with a distance-correlation heuristic, but this is an unsupported identification step—a correctness/robustness risk—not an equation that makes the prediction equal to the input by construction. Algorithm 1's two CI tests are the same tests the theorem requires; no fitted parameter is later relabeled as a prediction. The pruning step is standard post-hoc model selection. The only same-author citation is CD-NOTS (Sadeghi et al., 2024), used as a baseline for comparison; it does not carry the proof of SyPI+ and is not load-bearing. The claimed quadratic test count is structural in the algorithm's loop. Therefore no step reduces the central claim to its inputs; the main caveat—whether distance-correlation-selected lags equal Theorem 2.1's minimum lag—is an empirical approximation, not a definitional equivalence, and belongs under correctness risk rather than circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method has no new physical entities; its free parameters are the test thresholds and the lag-selection threshold, none of which are given principled values. The theoretical load rests on the SyPI theorem from prior work plus two unproven adaptations: distance-correlation lag identification and the pruning fix for assumption A6.

free parameters (4)

threshold1
Significance threshold for the first CI test in Algorithm 1; not specified in the paper, yet it controls which edges are added before pruning.
threshold2
Significance threshold for the second CI test in Algorithm 1; not specified.
distance correlation threshold = 0.05
Used in Section 2.2 to filter weak lag dependencies; arbitrary and not validated.
max_lag T = 5 (in experiments)
Maximum lag considered; a user-specified hyperparameter, not a fitted constant but needed by the method.

assumptions (6)

domain assumption Causal Markov condition and faithfulness (A1, A2 from Table 1)
Standard assumptions for constraint-based causal discovery; stated in Table 1 and inherited from SyPI.
domain assumption No backward arrows in time, stationary graph, acyclicity (A3-A5)
Temporal ordering and graph stationarity/acyclicity assumptions stated in Table 1.
domain assumption Every variable has one-lag self-dependency and no multi-lag links (A7, A8)
Graph-connectivity restrictions inherited from SyPI; they bound the structure the method can recover.
domain assumption Theory of SyPI (Theorem 1 of Mastakouri et al. 2021)
The two CI tests in Algorithm 1 are justified solely by this cited theorem; the paper does not re-prove it.
ad hoc to paper Distance-correlation-selected lags equal the graph-theoretic minimum lag w of Theorem 2.1
Section 2.2 implicitly assumes that the lags with the highest distance correlation (above 0.05) give the minimum lag w. This is not proved and is load-bearing for conditioning-set construction.
ad hoc to paper Pruning step restores soundness when A6 is violated
The paper relaxes the sink-node assumption (A6) and claims pruning fixes the resulting spurious edges, but provides no proof that the two-test rule plus pruning is sound.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Causal Discovery for Autoregressive Time Series." pith.science (2026). https://pith.science/paper/LAUGMXNR

@misc{pith2026250707898,
  author       = {Pith},
  title        = {Pith review of: Efficient Causal Discovery for Autoregressive Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAUGMXNR}},
  note         = {Machine review of arXiv:2507.07898}
}

read the original abstract

In this study, we present a novel constraint-based algorithm for causal structure learning specifically designed for nonlinear autoregressive time series. Our algorithm significantly reduces computational complexity compared to existing methods, making it more efficient and scalable to larger problems. We rigorously evaluate its performance on synthetic datasets, demonstrating that our algorithm not only outperforms current techniques, but also excels in scenarios with limited data availability. These results highlight its potential for practical applications in fields requiring efficient and accurate causal inference from nonlinear time series data.

Figures

Figures reproduced from arXiv: 2507.07898 by the authors.

**Figure 1.** In this figure, min lag is 2. Algorithm 1 SYPI+ Pseudocode Require: set of variables X for each Y in X do R = X \ {Y } w = min lags(R, Y ) ▷ See [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Lag Graph: Each row represents a candidate cause variable, and each column corresponds [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Pruning removes spurious links. All direct connections between Mek, Raf, Jnk, and P38 are [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Runtime evaluation for SyPI+ for different CI tests [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Number of CI tests executed by each algorithm across different test types [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: F-score evaluation for different methods [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: SHD for different methods 4 Case Study: Causal Connections Between Major Banks In this case study, we aim to uncover causal relationships between major Wall Street banks using weekly Credit Default Swap (CDS) data from end of January 2004 to September 2008, just before…

**Figure 8.** Figure 8: CDS network for major banks, just before the collapse of Lehman Brothers [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 10 canonical work pages

[1]

An approximation to the distribution of quadratic forms in normal random variables

Buckley, M. J. and G. K. Eagleson (1988). “An approximation to the distribution of quadratic forms in normal random variables”. In: Australian Journal of Statistics 30.1, pp. 150–159

work page 1988
[2]

Optimal structure identification with greedy search

Chickering, D. M. (2002). “Optimal structure identification with greedy search”. In:Journal of machine learning research 3.Nov, pp. 507–554

work page 2002
[3]

Chi squared approximations to the distribution of a sum of independent random variables

Hall, P. (1983). “Chi squared approximations to the distribution of a sum of independent random variables”. In: The Annals of Probability , pp. 1028–1036

work page 1983
[4]

Estimating high-dimensional directed acyclic graphs with the PC-algorithm

Kalisch, M. and P. B¨ uhlman (2007). “Estimating high-dimensional directed acyclic graphs with the PC-algorithm.” In: Journal of Machine Learning Research 8.3

work page 2007
[5]

Necessary and sufficient conditions for causal feature selection in time series with latent common causes

Mastakouri, A. A., B. Sch¨ olkopf, and D. Janzing (2021). “Necessary and sufficient conditions for causal feature selection in time series with latent common causes”. In:International Conference on Machine Learning. PMLR, pp. 7502–7511

work page 2021
[6]

Conditional independence testing based on a nearest-neighbor estimator of con- ditional mutual information

Runge, J. (2018). “Conditional independence testing based on a nearest-neighbor estimator of con- ditional mutual information”. In: International Conference on Artificial Intelligence and Statistics . Pmlr, pp. 938–947

work page 2018
[7]

Causal discovery from nonstationary time series

Sadeghi, A., A. Gopal, and M. Fesanghary (2024). “Causal discovery from nonstationary time series”. In: International Journal of Data Science and Analytics , pp. 1–27

work page 2024
[8]

An approximate distribution of estimates of variance components

Satterthwaite, F. E. (1946). “An approximate distribution of estimates of variance components”. In: Biometrics bulletin 2.6, pp. 110–114

work page 1946

Show all 10 references

[9]

Learning Bayesian networks with the bnlearn R package

Scutari, M. (2010). “Learning Bayesian networks with the bnlearn R package”. In:Journal of statistical software 35, pp. 1–22

2010
[10]

The significance of the difference between two means when the population variances are unequal

Welch, B. L. (1938). “The significance of the difference between two means when the population variances are unequal”. In: Biometrika 29.3/4, pp. 350–362. 10

1938

Pith tools

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

[1] [1]

An approximation to the distribution of quadratic forms in normal random variables

Buckley, M. J. and G. K. Eagleson (1988). “An approximation to the distribution of quadratic forms in normal random variables”. In: Australian Journal of Statistics 30.1, pp. 150–159

work page 1988

[2] [2]

Optimal structure identification with greedy search

Chickering, D. M. (2002). “Optimal structure identification with greedy search”. In:Journal of machine learning research 3.Nov, pp. 507–554

work page 2002

[3] [3]

Chi squared approximations to the distribution of a sum of independent random variables

Hall, P. (1983). “Chi squared approximations to the distribution of a sum of independent random variables”. In: The Annals of Probability , pp. 1028–1036

work page 1983

[4] [4]

Estimating high-dimensional directed acyclic graphs with the PC-algorithm

Kalisch, M. and P. B¨ uhlman (2007). “Estimating high-dimensional directed acyclic graphs with the PC-algorithm.” In: Journal of Machine Learning Research 8.3

work page 2007

[5] [5]

Necessary and sufficient conditions for causal feature selection in time series with latent common causes

Mastakouri, A. A., B. Sch¨ olkopf, and D. Janzing (2021). “Necessary and sufficient conditions for causal feature selection in time series with latent common causes”. In:International Conference on Machine Learning. PMLR, pp. 7502–7511

work page 2021

[6] [6]

Conditional independence testing based on a nearest-neighbor estimator of con- ditional mutual information

Runge, J. (2018). “Conditional independence testing based on a nearest-neighbor estimator of con- ditional mutual information”. In: International Conference on Artificial Intelligence and Statistics . Pmlr, pp. 938–947

work page 2018

[7] [7]

Causal discovery from nonstationary time series

Sadeghi, A., A. Gopal, and M. Fesanghary (2024). “Causal discovery from nonstationary time series”. In: International Journal of Data Science and Analytics , pp. 1–27

work page 2024

[8] [8]

An approximate distribution of estimates of variance components

Satterthwaite, F. E. (1946). “An approximate distribution of estimates of variance components”. In: Biometrics bulletin 2.6, pp. 110–114

work page 1946