acceptodds
Under review as a conference paper at ICLR 2027

Causal Discovery from Quantile Separation

Abstract

Causal discovery exploits asymmetries in uncertainty measures such as variance. However, many uncertainty-based methods rely on additive-noise assumptions. We extend this principle to structural causal models beyond additive noise using the Pairwise Average Separation Score (PASS), which measures uncertainty through quantile separation. Under a quantile-separation condition, our algorithm uses the same nodewise PASS score for both topological ordering and parent recovery. Under stated conditions, we establish graph-recovery consistency for a growing tensor-spline estimator. The criterion accommodates heavy-tailed distributions, including those with infinite variance, under suitable fractional-moment conditions. Experiments demonstrate robust graph recovery across additive-noise, heteroskedastic-noise, and heavy-tailed settings, with evaluations on three real-data benchmarks.

Primary area:
causal reasoning
Keywords:
causal discvoery, quantile regression, heavy tailed distribution

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GLHTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces PASS, a quantile-based conditional uncertainty score for causal discovery beyond additive-noise models. Under a quantile-separation condition, it establishes identification of causal order and parents, and proves consistency for a tensor-spline estimator under regularity and fractional-moment assumptions. Experiments evaluate a practical additive-spline implementation on additive, heteroskedastic, and heavy-tailed simulations and three real-data benchmarks.

Critical strengths
  1. PASS uses the same conditional uncertainty score for causal ordering and parent recovery, with identification results and consistency guarantees for a tensor-spline estimator under stated assumptions. Its applicability beyond additive noise and finite variance provides a meaningful extension of uncertainty-based causal discovery.

  2. The empirical evaluation covers diverse synthetic settings and real-world benchmarks, complemented by sensitivity analyses and ablations. This breadth provides substantial empirical support for the proposed method.

Critical weaknesses
  1. Restrictive, scale-dependent identification. Recovery relies on an ordering of parent-conditioned uncertainties whose applicability to real systems is insufficiently justified. Variable-specific rescaling substantially degrades ordering accuracy (Table 2), while the main HNM generator favors this condition. This leaves the practical scope of the proposed extension unclear.

  2. Theoretical guarantees do not cover the evaluated algorithm. Theorem 6.1 analyzes growing tensor splines with moment clipping and a fixed threshold; experiments use fixed additive splines, different clipping, and a data-dependent threshold. Consequently, the consistency result does not establish reliability of the implementation responsible for the reported empirical gains.

Questions for authors

What practically relevant class of causal systems is enabled by the proposed extension beyond additive-noise and finite-variance settings? This would help assess the practical significance of the broader theoretical model class.

4 comments

We thank the reviewer for the thorough review. The questions helped us check whether the main simulation favors PASS, bring the implementation closer to the theory, and explain where the extension is useful. Revised passages of the manuscript are marked in blue.

W1-1. Scale-dependent identifiability

The condition of PASS is weaker than the existing equal-variance conditions, and it is natural for variables measured on a common scale, such as examination scores (Park, 2020) or river discharge. Like other dispersion-based methods, PASS is not invariant to arbitrary coordinatewise rescaling, so it is intended for variables on a common scale.

(i) Scale dependence. Methods that identify a linear Gaussian SEM beyond its Markov equivalence class compare error variances across variables, so none of them is invariant to arbitrary coordinatewise rescaling (Peters and Bühlmann, 2014; Ghoshal and Honorio, 2018; Park, 2020; Xu et al., 2024). A common change of unit and variable-specific shifts leave the population PASS order unchanged (Proposition B.1). We stated the scope of PASS explicitly in Section 5.1 and the conclusion of the revised manuscript.

(ii) Restrictiveness. The condition of PASS includes the equal-variance condition of Peters and Bühlmann (2014) and the Gaussian variance-ordering condition of CaPS as special cases. It also covers general SCMs and orderings driven by scale, tail shape, or both. Its ordering step needs only the recursive sink separation condition (5), which is weaker than the edgewise condition of Corollary 5.1. PASS remains the most accurate method when the depth-based scale levels of the Table 1 generator are removed (see our answer to W1-2).

(iii) Applicability to real systems. By Corollary 5.1, the recursive sink separation condition (5) holds whenever the dispersion of each variable given its parents does not decrease along the edges, which is natural for variables from a similar domain (Peters and Bühlmann, 2014; Park, 2020). River discharge is such a case, since all stations measure the same physical quantity. On the Danube discharge network, PASS has the highest F1 among the methods in Table 4 (see our answer to Q1).

W1-2. The main HNM generator

PASS remains the most accurate method in every noise setting when the depth-based scale levels are removed. We removed them from the Table 1 generator and kept everything else.

MethodGaussian AtopA_{\mathrm{top}}Gaussian F1Gaussian SHDPareto(2) AtopA_{\mathrm{top}}Pareto(2) F1Pareto(2) SHDCauchy AtopA_{\mathrm{top}}Cauchy F1Cauchy SHD
PASS0.9190.70229.80.9520.76622.90.9030.76023.1
FICO0.8510.151478.20.9400.237251.00.8810.38488.7
TOPIC0.5820.39875.10.3340.173105.50.3830.073105.6
CAM0.5490.43769.30.3860.252111.70.3030.098127.1
SCORE0.3800.144227.30.6480.101604.20.5710.068684.1
NoGAM0.3730.137231.80.6900.107600.70.5810.072674.5
DirectLiNGAM0.5820.40986.90.7990.51473.80.7810.51749.2
HOST0.7280.362135.80.2080.109168.90.3750.208135.3
SkewScore0.4460.30498.40.5770.319119.10.5410.186137.5

This agrees with Theorem 5.1. The ordering step of PASS needs only the recursive sink separation condition (5), which is weaker than the edgewise condition of Corollary 5.1 that the depth-based scale levels favor. We added the results of all methods without the depth-based scale levels to Appendix G.2.7 of the revised manuscript.

W2. Gap between the theory and the implemented algorithm

The implementation now uses the moment clipping of Theorem 6.1, additive splines match the accuracy of growing tensor splines where both can be fit, and the standard-error threshold selects non-parent candidates at a rate at or below α\alpha for α≥0.01\alpha\ge0.01 and 0.0012 at the default α=0.001\alpha=0.001.

(i) Additive splines. The growing tensor splines of Theorem 6.1 cannot be fit at p=50p=50, since their basis has ⌈n1/5⌉∣A∣\lceil n^{1/5}\rceil^{|A|} functions for a conditioning set AA, at least nn once ∣A∣≥5|A|\ge5. We compared the two on graphs with p=5p=5, where both can be fit (six settings of Table 1, ten seeds). They agree within about 0.02 in AtopA_{\mathrm{top}} and 0.01 in F1 at every nn.

nnAdditive splines AtopA_{\mathrm{top}}Additive splines F1Growing tensor splines AtopA_{\mathrm{top}}Growing tensor splines F1
10000.9650.8940.9820.885
20000.9750.9200.9840.908
40000.9720.9190.9880.909

(ii) Clipping. We reran every table and figure with the moment clipping, and the estimated edge set changed in 66 of the 19,927 runs. The tables of the original submission change by at most 0.2 in SHD and 0.001 in F1.

(iii) Standard-error threshold. The fixed threshold of Theorem 6.1 depends on the unknown smallest population score of a true parent (Theorem 5.2), so the implementation compares each score with z1−αz_{1-\alpha} times its estimated standard error.

Selection rate of non-parent candidates. With the true order supplied (six settings of Table 1), non-parent candidates have zero population score (Theorem 5.2), and the table reports the fraction of them that the threshold selects.

False selection rate at α\alpha0.10.050.010.001
True order0.0420.0220.00580.0012

Sensitivity to α\alpha. The F1 changes by 0.007 between α=0.01\alpha=0.01 and the default α=0.001\alpha=0.001 (estimated order, six settings of Table 1), so the results do not rely on a tuned α\alpha.

α\alphaPrecisionRecallF1
0.10.4590.7990.579
0.050.5920.7710.666
0.010.7840.7230.750
0.0010.8700.6730.757

We added these results to Appendices G.3.4 and G.3.7 of the revised manuscript.

Q1. Systems enabled by the extension

The extension enables dispersion-based causal ordering for heavy-tailed data and for non-additive noise. Heavy-tailed or infinite-variance data are common in finance, insurance, and hydrology [1, 2], and non-additive noise includes heteroskedastic and post-nonlinear mechanisms.

(i) Heavy-tailed data. The PASS score needs only a finite β\beta-moment for some 0<β≤10<\beta\le1, not a finite variance. River discharge networks are one heavy-tailed application, studied in causal discovery on the Danube data we use (Gnecco et al., 2021; Tran et al. [3]) and on CausalRivers (Stein et al., 2025). PASS has the highest F1 on the Danube data and ranks fourth by F1-max on the CausalRivers benchmark, above the time-series methods DynoTears and PCMCI, without using the temporal order (Table 4). PASS also has the lowest SHD under Cauchy noise, which has no finite mean (Table 1).

(ii) Non-additive mechanisms. PASS also applies when the noise does not enter the mechanism additively, since it compares the full conditional distribution rather than a residual variance. Heteroskedastic noise arises when the variability of an effect grows with its causes, as is common for river discharge. For such noise, the condition of PASS compares the expected β\betath power of the noise scale across nodes, which extends the equal-variance condition. PASS has the lowest SHD and the highest F1 in the three heteroskedastic settings of Table 1. Post-nonlinear distortion arises when a sensor or a physical process distorts the mechanism (Zhang and Hyvärinen, 2009). Under such distortion, PASS has the highest AtopA_{\mathrm{top}} for g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5} and the second highest for g(x)=x3g(x)=x^3.

References

[1] Embrechts, P., Klüppelberg, C., and Mikosch, T. (1997). Modelling Extremal Events for Insurance and Finance. Springer.

[2] Katz, R. W., Parlange, M. B., and Naveau, P. (2002). Statistics of extremes in hydrology. Advances in Water Resources, 25(8–12):1287–1304.

[3] Tran, N. M., Buck, J., and Klüppelberg, C. (2024). Estimating a directed tree for extremes. Journal of the Royal Statistical Society Series B, 86(3):771–792.

Thank you for the responses and the revised manuscript. Your response to my review substantially alleviates my main concerns through the HNM results without depth-based scale calibration, the additive-versus-tensor spline comparison, the updated moment clipping, and the threshold diagnostics.

I also appreciate the additional post-nonlinear experiments (reply to Reviewer aKqM) and the explicit discussion of cases where PASS trails the strongest baselines in F1 (revised manuscript).

The practical justification for the separation condition and the theoretical coverage of the implemented algorithm remain limitations. On balance, considering the additional evidence, the paper’s contributions, and my assessment of the other papers I reviewed, I am raising my rating to weak acceptance.

Thank you for reading our responses and the revised manuscript carefully.

In particular, we are glad that the HNM results without the depth-based scaling, the spline comparison, and the threshold diagnostics addressed your main concerns, and that you found the post-nonlinear experiments and the discussion of the F1 results useful.

We also appreciate that, despite the remaining points on the practical justification of the separation condition and the theoretical coverage of the implemented algorithm, you raised your rating, considering the additional evidence and your assessment of the other papers. We will clarify these points in the final version, and we thank you again for your constructive feedback.

Reviewer_bYxZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose PASS, a quantile-based conditional uncertainty score for causal discovery. Under a quantile-separation assumption, they use the same score to remove sinks recursively and then test candidate parents within the recovered order. The paper proves population identifiability and consistency for a growing tensor-spline quantile estimator, including certain infinite-variance distributions. An implemented version using additive spline regressions is tested on additive-noise and heteroskedastic simulations, scale perturbations, and three real-data benchmarks.

Critical strengths
  1. The connection between quantile separation and conditional distribution changes is useful. Proposition 4.1 gives the score a clear interpretation, and using it for both sink ordering and parent selection makes the method conceptually simple.

  2. The experimental evaluation goes beyond Gaussian additive noise. The results on Cauchy and Pareto noise are a convincing reason to consider a fractional-moment score, and the paper reports failures under variable-specific scaling rather than hiding that limitation.

Critical weaknesses
  1. There is a substantial gap between the estimator covered by the consistency proof and the one used in the experiments. Section 6.3 analyzes growing full tensor-spline spaces, while Appendix F describes fixed-knot additive splines, smoothed pinball loss, cross-fitting, and an empirical standard-error threshold. In particular, no argument is given that the practical parent-selection threshold is calibrated after the data-dependent ordering step. The theory therefore does not yet establish consistency of the reported algorithm.

  2. The ordering signal depends on the relative measurement scales. Table 2 shows that even a moderate variable-specific rescaling reduces ordering accuracy, and the main heteroskedastic generator sets scale levels according to graph depth (Appendix G.1). Although both choices are disclosed, it remains difficult to judge how much of the strong HNM performance comes from a favorable separation pattern rather than from the quantile estimator itself.

Questions for authors
  1. Could the authors compare the implemented additive-spline estimator with a growing tensor-spline version on a small graph where the latter is computationally feasible? Even one sample-size study of ordering accuracy and graph recovery would help show whether the theoretical estimator describes the behavior of the practical implementation.

  2. Table 1 uses depth-calibrated HNM scales, while Appendix G.2.7 considers an uncalibrated setting. Could the authors report the uncalibrated HNM results at the main Table 1 setting (p=50, n=1000), using the same baseline methods? I am especially interested in whether the advantage over FICO remains under a less favorable ordering pattern.

  3. The parent rule in Section 6.2 uses z_(1-alpha) times an estimated standard error after first estimating the ordering. How sensitive are edge precision and recall to alpha, and do null parent scores show approximately the expected rejection rate? This would clarify what the threshold is doing in finite samples.

4 comments

We thank the reviewer for the detailed review and the concrete suggestions. The suggestions helped us check how much the results depend on the simulation design and on the estimator, and explain the role of measurement scales. Revised passages of the manuscript are marked in blue.

W1. Gap between the theory and the implemented algorithm

Additive splines and the smoothed pinball loss make the fit feasible at p=50p=50, and additive splines give accuracy similar to tensor splines where both can be fit. The cross-fitted standard-error threshold replaces the unknown constant of the fixed threshold, and with the true order it selects non-parent candidates at a rate at or below α\alpha for α≥0.01\alpha\ge0.01 and 0.0012 at the default α=0.001\alpha=0.001.

(i) Additive splines and smoothed loss. For a conditioning set AA, a tensor basis has ⌈n1/5⌉∣A∣\lceil n^{1/5}\rceil^{|A|} functions, which is at least nn once ∣A∣≥5|A|\ge5. Additive splines need only 5∣A∣5|A| coefficients, so Algorithm 1 remains feasible at p=50p=50. The softplus-smoothed pinball loss is a differentiable approximation of the pinball loss that allows smooth optimization.

(ii) Cross-fitting and the standard-error threshold. Cross-fitting makes each score an average of terms evaluated on held-out folds, so its standard error is computed as for a sample mean. With the estimated order, we also report the selection rate of candidates that are neither ancestors nor descendants of the target as an empirical diagnostic (see our answer to Q3-1 and Q3-2).

W2-1. Relative measurement scales

PASS is not invariant to arbitrary coordinatewise rescaling, as Table 2 shows. This property is shared by every method that identifies a linear Gaussian SEM beyond its Markov equivalence class, since such a method must compare error variances across variables (Peters and Bühlmann, 2014; Ghoshal and Honorio, 2018; Park, 2020; Xu et al., 2024). Since PASS contains the equal-variance condition of Peters and Bühlmann (2014) as a special case, the rescaling in Table 2 would also break that condition. We recommend keeping the variables on their original scales when they are measured on a common scale, such as examination scores (Park, 2020) or river discharge, since the relative scales then carry the identifying information. The revised manuscript states this scope in Section 5.1 and the conclusion.

W2-2 and Q2. Depth-based scale levels and the uncalibrated HNM

Most of the HNM performance remains without the favorable scale pattern, and PASS retains its advantage over FICO. We set the scale level of every node to one in the Table 1 generator and kept everything else, including the baselines. Removing the scale levels lowers the AtopA_{\mathrm{top}} of PASS by 0.02 to 0.03 and its F1 by 0.03 to 0.05 relative to the depth-based scale levels. Under the uncalibrated generator, PASS has the lowest SHD and the highest F1 in every noise setting. The F1 difference from FICO comes mostly from parent recovery, since the AtopA_{\mathrm{top}} of FICO is within 0.03 of that of PASS under Pareto(2) and Cauchy noise, and under Gaussian noise it also comes from ordering, where the AtopA_{\mathrm{top}} of FICO is lower by 0.07.

MethodGaussian AtopA_{\mathrm{top}}Gaussian F1Gaussian SHDPareto(2) AtopA_{\mathrm{top}}Pareto(2) F1Pareto(2) SHDCauchy AtopA_{\mathrm{top}}Cauchy F1Cauchy SHD
PASS0.9190.70229.80.9520.76622.90.9030.76023.1
FICO0.8510.151478.20.9400.237251.00.8810.38488.7
TOPIC0.5820.39875.10.3340.173105.50.3830.073105.6
CAM0.5490.43769.30.3860.252111.70.3030.098127.1
SCORE0.3800.144227.30.6480.101604.20.5710.068684.1
NoGAM0.3730.137231.80.6900.107600.70.5810.072674.5
DirectLiNGAM0.5820.40986.90.7990.51473.80.7810.51749.2
HOST0.7280.362135.80.2080.109168.90.3750.208135.3
SkewScore0.4460.30498.40.5770.319119.10.5410.186137.5

The small effect of the calibration agrees with Theorem 5.1: the ordering step of PASS needs only the recursive sink separation condition (5), which is weaker than the edgewise condition of Corollary 5.1 that the calibration targets. We added this table to Appendix G.2.7 of the revised manuscript.

Q1. Additive and growing tensor splines

The implemented additive splines and growing tensor splines give similar ordering accuracy and graph recovery on graphs with p=5p=5, where growing tensor splines are feasible. The tensor version replaces only the additive spline basis by a tensor-product spline basis, with ⌈n1/5⌉\lceil n^{1/5}\rceil basis functions per predictor. The table averages the six settings of Table 1 (p=5p=5, ten seeds).

nnAdditive splines AtopA_{\mathrm{top}}Additive splines F1Growing tensor splines AtopA_{\mathrm{top}}Growing tensor splines F1
10000.9650.8940.9820.885
20000.9750.9200.9840.908
40000.9720.9190.9880.909

We added this table to Appendix G.3.7 of the revised manuscript.

Q3. Sensitivity to α\alpha and rejection rate of null scores

In our simulations, decreasing α\alpha trades recall for precision, and with the true order, non-parent candidates are selected at a rate at or below α\alpha for α≥0.01\alpha\ge0.01 and 0.0012 at the default α=0.001\alpha=0.001.

(i) Sensitivity to α\alpha. The F1 changes by 0.007 between α=0.01\alpha=0.01 and the default α=0.001\alpha=0.001 (estimated order, six settings of Table 1). Even at α=0.1\alpha=0.1, the F1 of 0.579 exceeds the average F1 of every baseline in Table 1.

α\alphaPrecisionRecallF1
0.10.4590.7990.579
0.050.5920.7710.666
0.010.7840.7230.750
0.0010.8700.6730.757

(ii) Rejection rate of null scores. With the true order supplied, non-parent candidates have zero population score (Theorem 5.2), and the table reports the fraction of them that the threshold selects (same setting as in (i)). With the estimated order, the table also reports the selection rate of candidates that are neither ancestors nor descendants of the target, which is similar. These candidates need not have zero population score, since conditioning on a common child among the predecessors can make a candidate dependent on the target.

False selection rate at α\alpha0.10.050.010.001
Estimated order0.0390.0200.00490.0012
True order0.0420.0220.00580.0012

We added both tables to Appendix G.3.4 of the revised manuscript.

Thank you for your rebuttal. Some of my concerns have been addressed. The authors should provide sufficient background to make the paper easier to follow. Definitions of some terms are missing, which makes the paper not easy to follow. I addition,  If the data do not follow the model's setup (HNM), e.g., X_3 = f(X_1, X_2 \epsilon), and f is nonlinear. what will happen to the performance? Could the authors conduction some experiments to illustrate it?

Thank you for this suggestion, and we apologize that several basic terms were not defined in the main text. We added their definitions where they first appear, and we did experiments on data that do not follow an HNM.

(i) Background and definitions. Section 3 now defines ANMs, HNMs, topological orderings, sinks, the local Markov property, d-separation, and faithfulness, and Sections 5 to 7 now explain in words the recursive sink separation condition (5) used for ordering recovery, the sample notation, the cross-fitting folds, and the simulation settings. We tried to cover all the terms that a reader needs, and if you let us know any term that remains unclear, we will improve its explanation further.

(ii) Mechanisms beyond HNM. We used the post-nonlinear model Xj=g(fj(PAj)+εj)X_j=g(f_j(\mathrm{PA}_j)+\varepsilon_j), a standard non-HNM benchmark in which the error enters a nonlinear function together with the causes, with g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5} and with g(x)=x3g(x)=x^3 as in Montagna et al. (2023a), on the Table 1 setting with Gaussian errors and no calibration (ER1, p=50p=50, n=1000n=1000, ten seeds).

ggMetricPASSFICOTOPICCAMSCORENoGAMDirectLiNGAMHOSTSkewScore
sign(x)∣x∣1.5\mathrm{sign}(x)\lvert x\rvert^{1.5}AtopA_{\mathrm{top}} ↑0.9480.9260.6820.7950.4600.4550.3560.8160.494
F1 ↑0.7040.8190.5190.7460.1710.1640.2130.6670.362
SHD ↓25.721.652.928.9228.4238.4125.741.088.2
x3x^3AtopA_{\mathrm{top}} ↑0.7570.7300.5890.6800.2020.2190.2080.8430.250
F1 ↑0.4230.4650.4140.4780.0430.0460.0820.6050.157
SHD ↓41.168.062.860.4395.4390.5133.144.9108.7

PASS has the highest AtopA_{\mathrm{top}} for g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5} and the second highest for g(x)=x3g(x)=x^3, and the lowest SHD for g(x)=x3g(x)=x^3 and the second lowest for g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5}. Its lower F1 comes from missed edges, since it selects few false edges (precision 0.946 and 0.909). These results are in Section 7 and Appendix G.2.8 of the revised manuscript.

Reviewer_aKqMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates a causal discovery method beyond additive noise models and proposes a Pairwise Average Separation Score (PASS). PASS measures conditional uncertainty using the average beta-power pairwise distance between conditional quantiles. Under the condition of edge quantile separation, the method recursively identifies sink nodes to recover topological order and utilizes the reduction of the same score for parent node selection. The paper demonstrates the graph identifiability and consistency of the growing tensor spline estimator under appropriate regularity and moment conditions. Experiments evaluate the proposed method on additive noise, heteroscedastic noise, and heavy-tailed synthetic datasets, as well as on three real-world benchmark datasets.

Critical strengths
  1. The paper extends uncertainty-based causal discovery beyond additive-noise models by introducing a quantile-separation criterion. This addresses a practically relevant limitation of methods based on conditional variance.

  2. The same PASS score is used for ordering and parent recovery. The paper provides theoretical results on graph identifiability and consistency under explicit separation and regularity assumptions.

  3. The method is evaluated on several noise distributions and three real-world benchmarks, with competitive results in heavy-tailed settings.

Critical weaknesses
  1. For full-tensor spline quantile regression with a fixed parent threshold, consistent results have been established, but the actual implementation uses additive splines and data-related thresholds. Could the authors provide further explanation for these differences?

  2. The ordering criteria depend on a variable-specific measurement scale, and Table 2 shows a significant degradation after rescaling. Could the authors explore practical methods for selecting scales or evaluating quantile separation conditions?

  3. Besides additive noise, the synthetic experiments primarily consider heteroscedastic location-scale models. Could the authors include experiments on more general non-location-scale causal mechanisms to clarify the applicability of this method?

Questions for authors

See Weakness

2 comments

We thank the reviewer for the careful review and the constructive questions. The questions helped us explain how the implementation relates to the theory, clarify the role of measurement scales, and test PASS beyond location-scale models. Revised passages of the manuscript are marked in blue.

W1. Theory and the implemented estimator

We use additive splines because the computational cost of tensor splines grows exponentially with the size of the conditioning set, and the two give similar accuracy where both can be fit. We use the standard-error threshold to replace the unknown constant of the fixed threshold, and with the true order it selects non-parent candidates at a rate at or below α\alpha for α≥0.01\alpha\ge0.01 and 0.0012 at the default α=0.001\alpha=0.001.

(i) Additive splines. For a conditioning set AA, a tensor basis has ⌈n1/5⌉∣A∣≥n∣A∣/5\lceil n^{1/5}\rceil^{|A|}\ge n^{|A|/5} functions, so it has at least nn coefficients once ∣A∣≥5|A|\ge5, while an additive basis has 5∣A∣5|A|. On graphs with p=5p=5, where both can be fit, the two give similar accuracy (six settings of Table 1, ten seeds).

nnAdditive splines AtopA_{\mathrm{top}}Additive splines F1Growing tensor splines AtopA_{\mathrm{top}}Growing tensor splines F1
10000.9650.8940.9820.885
20000.9750.9200.9840.908
40000.9720.9190.9880.909

(ii) Standard-error threshold. The threshold of Theorem 6.1 lies below the smallest population score of a true parent (Theorem 5.2), a constant that depends on the unknown distribution. The standard-error threshold compares each score with z1−αz_{1-\alpha} times its estimated standard error, so it needs only the level α\alpha and adapts to the scale of each score.

Sensitivity to α\alpha. The results are stable in α\alpha (estimated order, six settings of Table 1). Decreasing α\alpha trades recall for precision, the F1 changes by 0.007 between α=0.01\alpha=0.01 and the default α=0.001\alpha=0.001, and even at α=0.1\alpha=0.1 the F1 of 0.579 exceeds the average F1 of every baseline in Table 1, at most 0.482 (DirectLiNGAM).

α\alphaPrecisionRecallF1
0.10.4590.7990.579
0.050.5920.7710.666
0.010.7840.7230.750
0.0010.8700.6730.757

Rejection rate of null scores. Given the true order, non-parent candidates have zero population score (Theorem 5.2), and the threshold selects them at about the nominal rate (six settings of Table 1, 70,224 candidates).

False selection rate at α\alpha0.10.050.010.001
True order0.0420.0220.00580.0012

We added these results to Appendices G.3.4 and G.3.7 of the revised manuscript.

W2. Measurement scale and quantile separation

Like other dispersion-based methods, PASS is not invariant to arbitrary coordinatewise rescaling. We therefore recommend keeping the variables on their original scales when they share a common scale, such as the same physical quantity. Condition (5) cannot be tested directly, so we judge it with domain knowledge through its edgewise sufficient condition (Corollary 5.1), which is plausible, for example, in river networks.

(i) Scale dependence of this line of methods. Peters and Bühlmann (2014), Ghoshal and Honorio (2018), Park (2020), and CaPS (Xu et al., 2024) identify linear Gaussian SEMs beyond their Markov equivalence class by comparing error variances across variables, so none of them is invariant to arbitrary coordinatewise rescaling. PASS contains the conditions of Peters and Bühlmann (2014) and CaPS as special cases.

(ii) Choice of scale. PASS is most natural when the variables are measured on a common scale, such as examination scores (Park, 2020) or river discharge at the stations of the Danube and CausalRivers networks. A common change of unit and variable-specific shifts leave the population PASS order unchanged (Proposition B.1). We stated this scope explicitly in Section 5.1 and the conclusion of the revised manuscript.

(iii) Evaluating the separation condition. Condition (5) asks that, at each removal step, a sink has the largest dispersion given the other remaining variables. It involves the unknown graph, so it cannot be tested directly from data. By Corollary 5.1, it holds whenever the dispersion of each variable given its direct causes does not decrease along the edges. This edgewise condition can be judged with domain knowledge, for example when the variables are measured on a common scale and their error dispersions are similar (Peters and Bühlmann, 2014; Park, 2020).

W3. Mechanisms beyond location-scale

We added two post-nonlinear settings, and PASS has the highest ordering accuracy for g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5} and the second highest for g(x)=x3g(x)=x^3. We generated the post-nonlinear model Xj=g(fj(PAj)+εj)X_j=g(f_j(\mathrm{PA}_j)+\varepsilon_j) of Zhang and Hyvärinen (2009), with g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)|x|^{1.5} and with the function g(x)=x3g(x)=x^3 used by Montagna et al. (2023a, Section 3.1.1). The data use the mechanisms fjf_j of Table 1 and Gaussian noise εj\varepsilon_j (ER1, p=50p=50, n=1000n=1000, ten seeds), without the depth-based scale levels of Table 1, so the generator does not target the separation condition (5).

ggPASSFICOTOPICCAMSCORENoGAMDirectLiNGAMHOSTSkewScore
g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)\vert x\vert ^{1.5}AtopA_{\mathrm{top}}0.9480.9260.6820.7950.4600.4550.3560.8160.494
F10.7040.8190.5190.7460.1710.1640.2130.6670.362
g(x)=x3g(x)=x^3AtopA_{\mathrm{top}}0.7570.7300.5890.6800.2020.2190.2080.8430.250
F10.4230.4650.4140.4780.0430.0460.0820.6050.157

In both settings, PASS selects few false edges (precision 0.946 and 0.909), so its lower F1 comes from missed edges. We added both settings to Appendix G.2.8 of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their constructive comments, which helped us improve the paper. We summarize the main concerns raised in the reviews, where all changes from the submitted version are marked in blue.

  • Theory and the implemented algorithm. The implementation now uses the moment clipping of Theorem 6.1. Additive splines, which we use because the cost of tensor splines grows exponentially with the conditioning set, give accuracy similar to growing tensor splines where both can be fit (p=5p=5). We use a standard-error threshold because the fixed threshold of Theorem 6.1 depends on the unknown smallest population score of a true parent, and this threshold selects non-parent candidates at a rate at or below α\alpha for α≥0.01\alpha\ge0.01, with F1 changing by 0.007 between α=0.01\alpha=0.01 and the default 0.0010.001 (Appendices G.3.4 and G.3.7).

  • Scale dependence. PASS is not invariant to arbitrary coordinatewise rescaling, but this property is shared by every method that identifies a linear Gaussian SEM beyond its Markov equivalence class by comparing error variances. A common change of unit and variable-specific shifts leave the PASS order unchanged (Proposition B.1), and the revision states this scope (Section 5.1 and the conclusion).

  • Depth-uncalibrated HNMs. Without the depth-based scaling, PASS still has the lowest SHD and the highest F1 in every HNM noise setting, with F1 between 0.70 and 0.77 against at most 0.52 for the next best method, and it keeps its advantage over FICO (Table 7A).

  • Mechanisms beyond HNMs. We added experiments on the post-nonlinear model Xj=g(fj(PAj)+εj)X_j=g(f_j(\mathrm{PA}_j)+\varepsilon_j), which is not an HNM, with g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)\lvert x\rvert ^{1.5} and with g(x)=x3g(x)=x^3 as in Montagna et al. (2023a), on the Table 1 setting without calibration. PASS has the highest AtopA_{\mathrm{top}} for g(x)=sign(x)∣x∣1.5g(x)=\mathrm{sign}(x)\lvert x\rvert ^{1.5} and the lowest SHD for g(x)=x3g(x)=x^3, and it ranks second in the other of these two metrics in each setting (Section 7 and Appendix G.2.8).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.