acceptodds
Under review as a conference paper at ICLR 2027

The System Is Out of Control, but Causality Isn't: Leveraging Operational Shifts for Bivariate Causal Discovery

Abstract

Bivariate causal discovery often relies on subtle asymmetries in a single observational distribution, yet these signals become unreliable when causal effects are weak or masked by noise. Multiple environments can offer stronger evidence through distributional changes and invariances, but many applications provide only two regimes: a reference condition and one alternative. We study causal direction identification in this practically important but theoretically underexplored setting. We introduce Kernel Residual Displacement (KRD), which exploits a simple directional asymmetry: under the correct direction, a localized change between the two regimes leaves one residual coordinate invariant, whereas under the reverse direction the change propagates to both residual coordinates. KRD quantifies this asymmetry through kernel distances between residual distributions, requiring neither multiple auxiliary environments nor conditional-independence testing. We establish population identifiability and finite-sample recovery guarantees that explicitly connect successful identification to perturbation magnitude, regression error, sample size, and directional separation. For linear Gaussian systems where causal direction is unidentifiable from a single observational distribution, our analysis further reveals a non-monotone signal-to-noise geometry: identification is strongest when causal signal and downstream noise are balanced and deteriorates toward both near-deterministic and noise-dominated extremes. Experiments across simulations, nonlinear and non-Gaussian mechanisms, and real-world datasets show that even one additional regime can yield substantial causal information beyond conventional single-context discovery.

Primary area:
causal reasoning
Keywords:
Bivariate Causal Discovery, Out of Control, Residual Patterns

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_kepWOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of bivariate causal discovery, that is, given two correlated variables, identifying cause and effect. For this, the authors assume an additive-noise model but allow a distribution shift between environments. This is an additive mean shift to one noise variable, while preserving the structural mechanism. The statistical asymmetries obtained from this shift can provide additional evidence for differentiating cause from effect. The method comes with identifiability guarantees and has been extensively evaluated on synthetic and real-world datasets.

Critical strengths

The paper is a great extension of the theory for cause-effect inference from bivariate data under an additive-noise assumption to scenarios with distribution shifts in noise variables. This can provide additional statistical evidence for inferring the causal direction. Beyond the theoretical analysis, the empirical results look promising.

Critical weaknesses

The identifiability result appears to require stronger assumptions to exclude shifts that preserve the effect distribution. For example, with a symmetric U-shaped relationship, moving a symmetric bell-shaped cause distribution could leave the effect distribution unchanged. This can happen through a mean shift alone, without changing the cause's variance and, thus, both causal directions would then receive the same zero score (not identifiable).

Questions for authors

Overall a great extension of the ANM-based approach! Just a few questions:

  • I might be overlooking something, but, as mentioned above, there seem to be conditions where the proposed score cannot identify the causal direction, for example, a symmetric U-shaped relationship where shifting the cause leaves the effect distribution unchanged. An additional assumption may be needed here. What could that look like and how would the identification result need to be revised (if at all)?
  • What regression models are used for computing KRD's residuals and for the baseline methods, I am assuming you are using causal discovery library where these are implemented, if yes, which one? Appendix B.1 describes the baseline implementations, but I could not find the corresponding details for KRD itself. It would be insightful to see the implementation and configuration used for the experiments.

Otherwise, if the first question can be addressed, I don't have any further concerns as the theoretical guarantees are then clearly given.

2 comments
AuthorsOpenReview ↗

Response to Reviewer kepW

We thank the reviewer for recognizing our extension of additive-noise causal discovery to two environments using a simple residual displacement score. We also appreciate the time and thoughtful feedback devoted to our work. Please see our uploaded revision with changes highlighted in blue.

Response to weaknesses:

We thank the reviewer for this example. The stronger condition excludes the degenerate scenarios in Remark 1, which we may not have stated clearly enough. We have revised it (lines 200–240) to give precise distributional definitions and an exhaustive classification of the invariance cases. In our revised classification, the U-shaped example falls under (a) cause-shift masking, where the cause shifts but the effect distribution remains unchanged.

Response to questions:

  1. As noted above, Remark 1 (revised lines 200–240) specifies the degenerate scenarios excluded by the identification result, including this U-shaped example.

  2. Both KRD variants use our implementation of ridge regression with 64 random Fourier features, an intercept, and a linear term, f^(x)=β0+β1z(x)+∑j=164wjϕj(x)\widehat f(x)=\beta_0+\beta_1z(x)+\sum_{j=1}^{64}w_j\phi_j(x), where xx is the regression input, z(x)z(x) is its IC-standardized value, and ϕj(x)\phi_j(x) are random Fourier features. We have added the implementation and configuration details to Appendix B.1, “Regression implementation in KRD” (lines 1163–1191).

    Our baseline implementations use established Python libraries, including scikit-learn for regression, ITE in Python for entropy estimation, CDT for IGCI, sparse_shift for MSS, causaldag for UT-IGSP, and causal-learn for CD-NOD. Moreover, we have validated our ANM and IGCI implementations against the benchmark of Mooij et al. (2016).

We thank the reviewer again for recognizing the value of our approach and for the suggestions that helped us make our analysis more rigorous.

Reference:

Joris M. Mooij, Jonas Peters, Dominik Janzing, Jakob Zscheischler, and Bernhard Schölkopf. Distinguishing cause from effect using observational data: methods and benchmarks. Journal of Machine Learning Research, 17(32):1–102, 2016.

Thank you for the clarification, in particular of my first question (which was my main concern). I keep my recommendation for acceptance!

Reviewer_e5MtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies bivariate causal discovery from a reference environment and an alternative environment affected by a localized shift. It proposes Kernel Residual Displacement (KRD), which compares changes in residual distributions and favors the direction retaining an invariant residual coordinate. The paper develops population identification and finite-sample guarantees, examines how signal and noise affect identification, and evaluates the method on synthetic and real-world benchmarks.

Critical strengths
  1. The paper addresses the useful setting in which only one additional environment is available. The residual-based score provides a simple way to exploit localized changes, and the linear Gaussian analysis offers an interpretable account of why identification becomes difficult at both signal-to-noise extremes.
  2. The experiments cover varied mechanisms and distinguish shifts affecting the cause from those affecting the effect. The appendix also examines different sample allocations and reports settings in which MSS performs better, providing a balanced picture of the empirical tradeoffs.
Critical weaknesses

-- The general population identification claim is not adequately established. The proof argues that perturbing the reverse prediction changes the reverse residual distribution unless certain degeneracies occur. However, changing a random quantity does not by itself establish that its marginal distribution changes. The discussion of exact cancellation and weak sensitivity does not rule out all distribution-preserving transformations. Consequently, the positive separation required by the identification theorem is asserted without sufficient justification for general nonlinear mechanisms. Explicit conditions ensuring this separation are needed to determine which models the theorem actually covers.

-- The explicit finite-sample recovery threshold relies on an invalid error bound. The derivation bounds the sum of average absolute cause values across the two environments by the shift magnitude plus a sampling term that vanishes as the sample grows. This omits the nonzero contribution of ordinary Gaussian variation, so the claimed inequality does not hold even with arbitrarily large samples. Moreover, the calculation treats the forward regression with known normalization, without establishing the simultaneous bound needed for the estimated residuals in both directions. These gaps undermine the stated recovery threshold, although they do not invalidate the population signal-to-noise calculation.

-- The experiments with inferred environment labels do not establish that the method exploits genuine environmental changes. The partition is constructed from deviations around a curve fitted using the first supplied variable, together with restrictions on its observed range. This selection can itself create residual differences, and exchanging the variables can produce a different partition. Fixing the partition parameters without causal labels avoids direct label tuning, but does not resolve these concerns. The resulting accuracy may partly reflect the partition procedure rather than the proposed causal asymmetry. Evidence that the conclusions remain stable under variable exchange and an independently justified partition would make these results more convincing.

-- The efficiency comparison does not isolate the contribution of the proposed score. The reported comparison uses Gaussian process regression for the additive-noise baselines but a fixed-size regression specification for KRD. Since regression dominates the cost of those baselines, part of the measured speed advantage follows from this implementation difference. The manuscript also leaves KRD's regression specification insufficiently described, making its approximation quality difficult to assess. A comparison using the same regression method would clarify how much speed is gained from residual displacement and whether that gain is retained at comparable prediction quality.

Questions for authors

Please see comments for authors.

2 comments
AuthorsOpenReview ↗

Response to Reviewer e5Mt (1/2)

We appreciate the reviewer’s recognition of the practical value of using only one additional environment. We are grateful for the time and care devoted to this review. The thorough, constructive feedback helped us rethink our assumptions and arguments, strengthening the manuscript’s rigor and completeness. To address the concerns, we made the degenerate cases precise, exhaustive, and intuitive (lines 200–240, Section 4.1), corrected the finite-sample threshold and proof (Appendix A.4), randomized partition orientation in unlabeled real-data experiments (lines 453–517, Section 5.2), and compared decision-stage runtimes excluding regression fitting. These revisions preserve the paper’s main qualitative conclusions and contributions. The changes are highlighted in blue in the uploaded revised manuscript. Below, we address the points raised in the weaknesses.

1. The general population identification claim is not adequately established

Revised Remark 1 (lines 200–240) precisely defines the excluded distribution-preserving transformations through equalities of marginal distributions, with Table 1 (lines 208–215) exhaustively classifying them into cause-shift masking, cause-shift compensation, and effect-shift absorption. These degeneracies capture symmetries and cancellations that hide environmental shifts, intuitively illustrated in Figure 2 (lines 228–240). Excluding these invariances ensures the positive separation required by the population identification theorem for a characteristic kernel.

2. The explicit finite-sample recovery threshold relies on an invalid error bound

We agree that the original bound omitted ordinary Gaussian variation and did not control all residual estimation errors. We have corrected both issues and re-derived the recovery threshold. Specifically,

  • Correcting the invalid absolute-value bound. As the reviewer notes, average absolute Gaussian observations do not vanish with sample size. Our original bound omitted this contribution. The revised proof retains 2/π\sqrt{2/\pi} and establishes$ \frac{1}{n_e}\sum_i |z_{j,i}^{(e)}| \le \sqrt{2/\pi}+|\delta|+2\varepsilon_N,$ for each of the four standardized residual coordinates jj in both IC and OOC environments ee.

  • Extending the forward-only calculation to all estimated residuals. The original proof controlled only the forward OLS coefficient error with known normalization. We now jointly bound the forward and reverse OLS coefficient errors, together with the errors in all four estimated centers and normalization scales.

Combining these bounds gives the full residual estimation bound used to re-derive Corollary 1 (lines 351–363), with the proof updated in Appendix A.4.

3. The experiments with inferred environment labels do not establish that the method exploits genuine environmental changes

We thank the reviewer for pointing out the sensitivity to column order and helping us clarify how accurately KRD can recover causal directions without observed OOC labels. Before addressing the IC/OOC partitioning concern, we want to clarify why we chose DREAM and Tübingen. Our experiments test KRD under progressively weaker assumptions and with less information about OOC. The extended simulations in Section 5.2 include causal mechanisms and noise distributions that depart from our assumptions. TEP provides OOC labels, but its faults may involve changes beyond mean shifts. DREAM and Tübingen go one step further: we must infer the OOC labels from the data. Our goal is to assess KRD’s limits, not to show that it always outperforms the baselines beyond its assumptions.

In summary, we revise the IC/OOC partition by randomizing the predictor for each pair to make the evaluation fairer, add two off-the-shelf anomaly detection methods to check whether our conclusions hold under other partitions, and generate synthetic OOC samples to compare KRD’s performance with correctly labeled mean shifts. The details are given below.

  • Column-order sensitivity. Randomizing the predictor for each pair changes the results, especially on Tübingen. We update Table 2 (lines 463–481) with the results for both datasets. The single-context baselines outperform the cross-context methods, including KRD, which may reflect the limits of inferring OOC from the data. Even so, KRD still leads the cross-context methods on DREAM and Tübingen.
  • Alternative OOC inference. We add experiments with two off-the-shelf anomaly detectors, GMM and LOF. Both lead to the same conclusion (Appendix B.4, Table 11).
  • Known mean-shift OOC labels. We generate OOC samples by applying a mean shift of magnitude δ=2\delta=2 to the cause or effect. KRD ranks among the top two methods on both datasets and reaches nearly 100% pair accuracy on Tübingen (Table 3, lines 498–516). This shows that KRD achieves high accuracy if OOC samples from a mean shift are correctly identified.
AuthorsOpenReview ↗

Response to Reviewer e5Mt (2/2)

4. The efficiency comparison does not isolate the contribution of the proposed score

The table below shows that KRD is faster than ANM-HSIC under matched regression. Its identification accuracy is also higher under cause shifts (100% versus 90%) and equal under effect shifts (100% for both). Thus, the speed advantage does not come at the cost of lower identification accuracy.

ShiftMethodAccuracyPrep (ms)Score (ms)Total (ms)OOC RMSE
CauseKRD-Linear1.00015.2090.02415.2320.1801
CauseKRD-RBF1.00015.2097.10822.3170.1801
CauseANM-HSIC0.90015.20949.70264.9110.1801
CauseMSS1.000——1585.440—
EffectKRD-Linear1.00015.0130.02415.0370.1136
EffectKRD-RBF1.00015.0137.00522.0190.1136
EffectANM-HSIC1.00015.01343.07958.0920.1136
EffectMSS1.000——1492.572—

Specifically, we use 500 observations per context, δ=1\delta=1, and 10 trials per shift type. Both KRD variants and ANM-HSIC share our IC-fitted ridge regression with 64 random Fourier features, now described in the newly added “Regression implementation in KRD” subsection of Appendix B.1 (lines 1163–1191). Regression fitting and residual computation are performed and timed once, then reused by all three scoring methods. Each method’s total time includes this preparation time (Prep (ms)) plus its own scoring time (Score (ms)). MSS has no separate regression step to replace, so we report its complete runtime as a reference.

We thank the reviewer again for helping us make the manuscript more rigorous and the scope of our experiments clearer.

Reviewer_gWWxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Kernel Residual Displacement (KRD) for bivariate causal discovery leveraging a reference environment and a second environment featuring localized additive mean shifts. The method compares marginal residual distributions across distinct environments and determines the causal direction based on the minimum displacement value. The key contributions of this work include population identification arguments, finite-sample error bounds, signal-to-noise ratio analysis under the linear-Gaussian setting, and empirical validations on synthetic datasets, standard cause–effect datasets, gene network datasets, and industrial process datasets.

Critical strengths
  • The score is simple and interpretable. Exploiting an unchanged residual coordinate provides a computationally attractive implementation of the sparse-shift principle.

  • The linear-Gaussian population calculation is informative, explaining why separation of KRD deteriorates at both signal-to-noise extremes. Experiments varying shift location, magnitude, and sample allocation help characterize this behavior.

Critical weaknesses
  • The main theoretical guarantees are not established. Under the stated assumptions, a nonconstant periodic reverse regression can preserve the reverse residual under an effect shift, giving both directions zero KRD scores and invalidating the general identification claim in Lemma 2/Theorem 1. Corollary 1 also uses a false bound: under a standard Gaussian model with an effect shift of δ=1\delta=1, the left-hand side of lines 865–866 converges to 1.5961.596, while the right-hand side converges to 11.

  • The Tübingen and DREAM environments are constructed from regression residuals and the range of the first supplied variable (Appendix B.4). This asymmetric selection does not establish the assumed single-component shift and may introduce selection effects. Without column-swap and partition-sensitivity checks, the reported gains cannot be attributed confidently to the proposed causal shift mechanism.

Questions for authors
  • What precise conditions establish Lemma 2 beyond assuming positive reverse-direction displacement? Can you repair Corollary 1, including reverse-regression and scale-estimation errors?

  • Do the Tübingen/DREAM results survive swapping the supplied columns before constructing environments? How sensitive are they to the partitioning procedure?

  • Please specify the KRD regression model, kernel-bandwidth selection, and TEP evaluation protocol. Does the runtime advantage persist when regression models are matched across methods?

2 comments
AuthorsOpenReview ↗

Response to Reviewer gWWx (Weaknesses)

We sincerely thank the reviewer for their time, dedication, and valuable comments. We have carefully addressed the concerns and strengthened our theoretical analysis. The changes are highlighted in blue in the uploaded revised manuscript.

Response to weaknesses

  • Weakness 1

    • (a) Degenerate cases. Our original Remark 1 may not have stated this exclusion clearly enough. As stated in Lemma 2 and Theorem 1, identification is conditional on excluding the degenerate scenarios described in Remark 1. The reviewer’s periodic reverse-regression example falls under (c), effect-shift absorption, in our revised classification (lines 216–227). To make this condition precise and intuitive, we have exhaustively classified the possible distributional invariances into three scenarios: cause-shift masking, cause-shift compensation, and effect-shift absorption. These degeneracies reflect functional or distributional symmetries and cancellations that hide environmental shifts, as illustrated in Figure 2 (lines 228–240).

    • (b) Finite-sample recovery. For Corollary 1, we thank the reviewer for identifying this error, which arose from omitting the nonzero Gaussian absolute moment in our original bound. The revised proof explicitly accounts for the previously omitted Gaussian absolute moment through the term 2/π\sqrt{2/\pi}, alongside separate bounds for the mean shift and finite-sample fluctuation (Appendix A.4, lines 1026–1031). Beyond this correction, we jointly bound errors from both fitted regressions, estimated centers, and normalization scales across all four residual coordinates and both contexts (Appendix A.4, lines 1047–1050). This more complete error analysis establishes a rigorous finite-sample recovery threshold in the revised Corollary 1.

  • Weakness 2

    • (a) Scope of the experiments. We test KRD under progressively weaker assumptions. Our controlled simulations in Section 5.1 test KRD under the conditions of our theory. KRD remains competitive across 18 additional simulation settings with different causal mechanisms and noise distributions, and on TEP, where OOC labels are available but faults may involve changes beyond mean shifts. DREAM and Tübingen go one step further because we must infer the OOC labels. Our goal is to assess KRD’s limits, not to show that it always outperforms the baselines beyond our assumptions.

    • (b) Column order and partition sensitivity. We thank the reviewer for pointing out the sensitivity to column order. We update the DREAM and Tübingen results by randomly choosing the predictor for each pair. Tübingen now shows the same pattern as DREAM. The best single-context methods score higher, likely because the inferred environments provide limited information for cross-context methods. Even so, KRD leads the cross-context methods on Tübingen and DREAM (Table 2). We test two off-the-shelf anomaly detectors, GMM and LOF, and find the same pattern (Appendix B.4, Table 11). We also generate synthetic OOC samples to test KRD with correctly labeled mean shifts. KRD achieves high accuracy on both DREAM and Tübingen (Table 3). The accuracy gap shows how much better KRD can perform with correctly labeled mean shifts than with inferred OOC labels.

AuthorsOpenReview ↗

Response to Reviewer gWWx (Questions)

Response to questions

1. What precise conditions establish Lemma 2 beyond assuming positive reverse-direction displacement? Can you repair Corollary 1, including reverse-regression and scale-estimation errors?

The precise condition excludes the following three distributional invariances.

  • (a) Cause-shift masking. P(zv(c)≤ζ)=P(zv(0)≤ζ)P(z_v^{(\mathrm{c})}\le\zeta)=P(z_v^{(0)}\le\zeta). The cause shifts, but the effect distribution remains unchanged.
  • (b) Cause-shift compensation. P(zu∣v(c)≤ζ)=P(zu∣v(0)≤ζ)P(z_{u\mid v}^{(\mathrm{c})}\le\zeta)=P(z_{u\mid v}^{(0)}\le\zeta). The cause and its reverse prediction change, but their difference retains the same distribution.
  • (c) Effect-shift absorption. P(zu∣v(e)≤ζ)=P(zu∣v(0)≤ζ)P(z_{u\mid v}^{(\mathrm{e})}\le\zeta)=P(z_{u\mid v}^{(0)}\le\zeta). The effect shifts, but the reverse residual distribution remains unchanged.

Here, zvz_v is the standardized effect and zu∣vz_{u\mid v} is the standardized reverse-regression residual. Superscripts 0,c,e0,\mathrm{c},\mathrm{e} denote IC, cause shift, and effect shift. Each equality holds for every threshold ζ∈R\zeta\in\mathbb{R}. Table 1 (lines 208–215) exhaustively classifies these degenerate scenarios.

Yes, we have also repaired Corollary 1. Appendix A.4 now jointly controls forward- and reverse-regression, centering, and scale-estimation errors across all four residual coordinates and both contexts. We corrected the omitted Gaussian absolute moment and re-derived the recovery threshold.

2. Do the Tübingen/DREAM results survive swapping the supplied columns before constructing environments? How sensitive are they to the partitioning procedure?

We now randomize the predictor for each pair. On Tübingen, KRD’s group accuracy drops from above 70% to about 60% (updated Table 2, lines 470–475). As on DREAM, the best single-context methods score higher. This is expected when the inferred split offers little of the information that cross-context methods rely on. However, KRD still leads the cross-context methods on Tübingen and DREAM. Moreover, two off-the-shelf anomaly detectors, GMM and LOF, lead to the same conclusion (Appendix B.4, Table 11). When OOC samples from a controlled mean shift are correctly labeled, KRD reaches nearly 100% pair accuracy on Tübingen (Table 3, lines 498–516).

3. Please specify the KRD regression model, kernel-bandwidth selection, and TEP evaluation protocol. Does the runtime advantage persist when regression models are matched across methods?

  • KRD regression model. Both KRD variants use ridge regression with 64 random Fourier features, an intercept, and a linear term,$ \widehat f(x)=\beta_0+\beta_1z(x)+\sum_{j=1}^{64}w_j\phi_j(x), $ where xx is the regression input, z(x)z(x) its IC-standardized value, and ϕj(x)\phi_j(x) its random Fourier features. Full details are provided in Appendix B.1, “Regression implementation in KRD” (lines 1163–1192).

  • Kernel-bandwidth selection. For KRD-RBF, the IC-only bandwidth rule is λ=median⁡i<j,  zi(0)≠zj(0)∣zi(0)−zj(0)∣,\lambda=\operatorname{median}_{i<j,\;z_i^{(0)}\ne z_j^{(0)}}\left|z_i^{(0)}-z_j^{(0)}\right|, where zi(0)z_i^{(0)} are IC residual values. The bandwidth is selected separately for each residual coordinate. KRD-Linear requires no bandwidth.

  • TEP evaluation protocol. For each TEP fault, we use normal-operation samples as IC and samples from that fault as OOC. We evaluate variable pairs whose causal directions are established by the physical process structure. Cross-context methods receive IC and OOC separately, and single-context baselines use the pooled samples.

  • Runtime. Both KRD variants remain faster than ANM-HSIC with the same regression model.

    ShiftMethodAccuracyPrep (ms)Score (ms)Total (ms)OOC RMSE
    CauseKRD-Linear1.00015.2090.02415.2320.1801
    CauseKRD-RBF1.00015.2097.10822.3170.1801
    CauseANM-HSIC0.90015.20949.70264.9110.1801
    CauseMSS1.000——1585.440—
    EffectKRD-Linear1.00015.0130.02415.0370.1136
    EffectKRD-RBF1.00015.0137.00522.0190.1136
    EffectANM-HSIC1.00015.01343.07958.0920.1136
    EffectMSS1.000——1492.572—

    We use 500 observations per context, δ=1\delta=1, and 10 trials per shift type. Both KRD variants and ANM-HSIC share IC-fitted ridge regression with 64 random Fourier features (Appendix B.1, lines 1163–1191). Prep is the shared time for regression fitting and residual computation. Score is each method’s scoring time. Total = Prep + Score. For MSS, Total is its full runtime.

We thank the reviewer again for helping us make the manuscript more rigorous and the scope of our experiments clearer. In light of the new results and clarifications, we kindly ask the reviewer to consider increasing the rating. Thank you for your time and consideration.

Reviewer_vqD3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  1. Summary

The paper addresses bivariate causal direction identification when observations are available from only two environments: an in-control (IC) reference regime and an out-of-control (OOC) regime, a setting the authors motivate by statistical process control. In the OOC regime, a mean shift perturbs a single exogenous noise term, either the cause’s or the effect’s, while the causal mechanism remains unchanged. The authors propose Kernel Residual Displacement (KRD), which compares how two standardized residual coordinates change across environments under each candidate direction, i.e., the centered putative cause, and the residual of regressing the putative effect on it. Under the correct direction, one of these coordinates is invariant across regimes. Under the incorrect direction, both generally change. The method measures each coordinate’s change by the Maximum Mean Discrepancy (MMD) between its IC and OOC distributions, using either a linear or an RBF kernel. Each direction is scored by the smaller of its two displacements, and the direction with the lower score is selected. The theory assumes an additive noise model in which the shift affects either the cause or the effect, but not both.

  1. Main contributions

a) New causal discovery method. KRD identifies causal direction from two environments by comparing marginal distributions of standardized residuals, avoiding conditional independence testing. The authors frame it as combining residual invariance with the sparse mechanism shift principle. They introduce a linear-kernel variant, which compares residual means in O(N) time, and an RBF-kernel variant, which uses the full empirical MMD in O(N²) time. b) Theoretical guarantees. The paper establishes population identifiability, excluding a set of degenerate cases, and gives sufficient conditions for finite-sample recovery. These conditions relate the reverse-direction separation to regression error, sample size, and perturbation magnitude. c) Signal-to-noise analysis for linear Gaussian models. The authors derive an explicit separation function ϕ(t) of the signal-to-noise ratio, together with corresponding thresholds on shift magnitude and IC sample size. The separation is largest when causal signal and downstream noise are balanced (t = 1) and vanishes at both extremes.

  1. Validation

The authors evaluate KRD through extensive synthetic experiments and three real-world benchmark families. a) Synthetic experiments. Controlled linear Gaussian and nonlinear non-Gaussian models vary the signal-to-noise ratio, shift magnitude, and IC/OOC sample sizes. Eighteen further settings from five benchmark families include mechanisms outside the additive noise model. b) Real-world benchmarks. The experiments use the Tübingen cause–effect pairs, the DREAM gene-network benchmarks, and Tennessee Eastman Process (TEP) fault data. Only TEP has naturally labeled IC and OOC regimes (normal versus faulty operation). For Tübingen and DREAM, the authors construct the two environments by anomaly detection around a fitted reference curve. KRD achieves about 69.5% overall accuracy on Tübingen, the best among the methods compared. On TEP, it reaches 65% on Fault 1, where the other methods are near chance, and is competitive on Faults 2 and 6. On DREAM, it is near chance, as are the other cross-environment methods. c) Computational efficiency. The authors report about 8 ms per pair for KRD-Linear and 20 ms for KRD-RBF, versus about 50 ms for MSS. The evaluated Gaussian-process and KCI-based baselines scale cubically with sample size.

  1. Overall interpretation

The paper’s main contribution is a computationally efficient method for extracting causal direction information from a single distributional contrast. It is supported by theoretical guarantees and relatively extensive empirical comparisons. The results are most compelling for linear Gaussian systems, where single-environment methods cannot identify causal direction. However, the guarantees rest on a narrow perturbation model and an incompletely characterized exclusion of degenerate cases. In addition, the real-world evidence depends largely on environments constructed from the data rather than observed.

Critical strengths
  1. A simple, well-motivated identification principle with an informative linear-Gaussian analysis. KRD turns the sparse-mechanism-shift intuition into a direct and inexpensive test. It checks whether one of two standardized residual coordinates is invariant across regimes, using MMD on marginal residual distributions instead of conditional independence tests. The idea of using cross-environment changes is not new. MSS and invariance-based methods already exploit it. The paper’s more distinctive theoretical contribution is the explicit linear-Gaussian analysis. There, the reverse-direction separation is |δ|ϕ(t) with ϕ(t) = min{1,t}/√(1+t²). This is maximized when causal signal and downstream noise are balanced and vanishes at both extremes. The non-monotone dependence is a useful and somewhat counterintuitive insight into when a single contrast carries directional information, and the controlled simulations bear it out. This strength weighs significantly in my assessment because it gives the paper a clear conceptual takeaway that does not depend on the parts of the theory I question below.
  2. Broad synthetic evaluation and a clear computational advantage. The synthetic experiments are systematic. They vary the signal-to-noise ratio, shift magnitude, shift location, and IC/OOC sample sizes, and they include 18 settings from five benchmark families, many outside the additive noise model. Across these, KRD is consistently among the top two methods and is comparable to MSS on average. It does this at a fraction of the cost: about 8–20 ms per pair, versus about 50 ms for MSS, with linear or quadratic rather than cubic scaling. This matters for my assessment because it shows that KRD’s accuracy holds up beyond the regime its theory covers. It also shows that the method offers a practical advantage over the strongest competitor even where it does not beat it in accuracy.
Critical weaknesses
  1. The identifiability and finite-sample results are not established as stated. The theory covers a deliberately narrow setting: an additive noise model with a mean shift in exactly one exogenous noise term. That scope is acceptable given the SPC motivation, but within it the main results have gaps. (a) Theorem 1 holds only after excluding the degenerate cases of Remark 1, and those cases are not precisely defined. “Weak causal strength” is an approximate condition (“nearly constant,” “little change”), so it does not define an exceptional set at the population level. “Shift equivariance” is stated as a pointwise identity, whereas invariance of z_{u|v} requires only equality in distribution. The listed cases therefore neither precisely define nor demonstrably exhaust the situations in which a reverse residual is invariant. In effect, the exclusion assumes that both reverse residual distributions change, which is what Theorem 1 needs, so the theorem is close to circular. (b) The proof of Corollary 1 (Appendix A.4) contains an error. It bounds (1/n₀)Σ|x_u^(0)| + (1/n₁)Σ|x_u^(1)| by |δ| + ε_N. For standard Gaussian x_u, however, E|x_u| = √(2/π) ≈ 0.8, so this sum is approximately |δ| + 1.6 and does not vanish with n. The stated threshold on |δ| is therefore incorrect as written, although a corrected version with an additional term may preserve the qualitative conclusion. The proof also bounds only the forward OLS coefficient error. Assumption 3’s ε_fit is a maximum over all four residual coordinates, but the errors from the reverse regression and from the estimated centers and scales are not controlled. (c) Assumption 3 requires accurate residual estimation in the OOC environment. Under a cause shift, this means the IC-fitted f must extrapolate up to δσ_u beyond the IC support, with δ up to 4 in the experiments. For nonlinear f, this is a strong assumption rather than routine regression error. It may also explain KRD’s weaker performance under cause shifts. This weakness is consequential because the theory is the paper’s main claimed advantage over existing two-environment methods such as MSS. As written, the general identifiability result is close to tautological, and the explicit linear-Gaussian recovery guarantee needs correction.
  2. The real-world evidence depends on constructed environments that may favor the method. Only TEP has naturally labeled regimes. For Tübingen and DREAM, the IC/OOC split is obtained by fitting a reference curve of the second supplied variable on the first and assigning points with large deviations, or points outside the main input range, to OOC. This raises three concerns. (a) The resulting OOC set is a selection on residual magnitude and input range, not the mean shift of Assumption 2, so the theory does not apply to it. (b) The construction is direction-asymmetric. Selecting points by their vertical deviation from a curve of the second variable on the first resembles an effect-side perturbation consistent with first → second. Unless the column order was randomized relative to the ground truth, or the partition was symmetrized, this could bias the results toward a particular direction. (c) The margins are small and unquantified. On Tübingen, KRD leads ANM-MML by about two percentage points, with no uncertainty estimates. On TEP, several baselines score far below chance (MSS-hard 10–11%, CD-NOD 7.5%). This suggests that abstentions are counted as errors or that these baselines are misconfigured, which complicates the relative comparison. On DREAM, all cross-environment methods, including KRD, are near chance. KRD-Linear scores 46.9% on D4S2A, where single-context ANM-HSIC reaches 64.2%. This weakness is consequential because the real-world benchmarks are the paper’s only evidence of practical relevance beyond simulation. TEP alone, with mixed results across three faults, is a thin basis for that claim.
Questions for authors
  1. Characterization of the degenerate cases (Lemma 2 / Theorem 1). Can the authors give a precise population-level condition under which each reverse residual distribution is invariant? This should be stated as equality in distribution, not as the approximate condition “weak causal strength” or the pointwise “shift equivariance” identity. Can they then show that the cases in Remark 1 are exhaustive? If not, Theorem 1 should be restated as conditional on both reverse residual distributions changing.

  2. Corollary 1. The proof bounds (1/n₀)Σ|x_u^(0)| + (1/n₁)Σ|x_u^(1)| by |δ| + ε_N, but E|x_u| ≈ 0.8 for standard Gaussian x_u. Can the authors provide a corrected threshold? Can they also extend the argument to control the reverse-regression, centering, and scale-estimation errors that Assumption 3’s ε_fit includes?

  3. Extrapolation under cause shifts. How is f fitted in the nonlinear experiments, and how does accuracy under cause shifts vary with δ when the OOC cause values fall outside the IC support? Is Assumption 3 plausible for δ = 2–4 with nonlinear f?

  4. Direction asymmetry in the constructed partitions. Was the column order of the Tübingen and DREAM pairs randomized relative to the ground truth before partitioning? Please report accuracy separately for pairs whose true direction is first → second and second → first. Alternatively, repeat the experiment with the partition built from the reversed regression. This is the most decisive check of whether the real-world results reflect the method rather than the partition procedure.

  5. Reported uncertainty and baseline behavior. Can the authors report confidence intervals or paired tests for the Tübingen and TEP comparisons? Can they also explain the below-chance accuracies of MSS-hard and CD-NOD on TEP? Are abstentions counted as errors?

2 comments
AuthorsOpenReview ↗

Response to Reviewer vqD3 (Weaknesses)

We thank the reviewer for the careful review and detailed feedback, which helps strengthen our work. The revised manuscript highlights the changes in blue.

Response to weaknesses

  • Weakness 1(a), degenerate cases. We apologize that the original Remark 1 lacks precise distributional definitions and an exhaustive classification of non-identifiable scenarios. We have revised it to address both issues (lines 200–240). We emphasize that these exclusions capture symmetries and cancellations that hide environmental shifts, rather than arbitrarily selected counterexamples. Their role is analogous to faithfulness in DAG-based causal discovery, which rules out cancellations that conceal causal dependencies. As in faithfulness-based DAG discovery, the exclusion is an explicit premise for identification, so the proof establishes a conditional result rather than reasoning circularly.

  • Weakness 1(b), Corollary 1. We have corrected the invalid absolute-moment bound to retain 2/π\sqrt{2/\pi} and extended the analysis beyond the forward OLS coefficient error to include reverse-regression, centering, and scale-estimation errors. These are controlled simultaneously across all four residual coordinates and both contexts. Corollary 1 has been revised in lines 351–363 of the main text, with its proof updated in Appendix A.4.

  • Weakness 1(c), Assumption 3. Assumption 3 does more than state sufficient conditions for recovery. It supports Theorem 2, which, together with our error analysis in Appendix A.4, accounts for regression, centering, and scale estimation errors, helping practitioners recognize possible causes of identification failure, including the large regression errors caused by extrapolation noted by the reviewer. This goes beyond proposing an identification rule and provides a finite-sample error analysis absent from the MSS work.

    Other theoretical studies also need strong conditions to control estimation errors and provide guarantees, though the restrictions vary by problem. Mooij et al. (2016) require vanishing squared regression errors in both directions, although within the same distribution. Shah and Peters (2020) impose specific convergence-rate conditions on combinations of regression errors, beyond requiring them merely to vanish.

  • Weakness 2, real-data experiments.

    • (a) Scope of the experiments. We clarify that our experiments are organized to test KRD under progressively weaker assumptions and with less information about the OOC context. Specifically, the extended simulations in Section 5.2 depart from our assumptions about the causal mechanism and noise distribution. TEP provides OOC labels, but its faults may involve changes beyond mean shifts. DREAM and Tübingen go a step further because we must infer the OOC labels from the data. The concern in 2(a) is exactly what we aim to examine in both real-data settings.

    • (b) Sensitivity to column order and partitioning. We add three analyses.

      1. We randomize the predictor for each pair. KRD’s group accuracy on Tübingen drops from above 70% to about 60% (Table 2, lines 470–475). As on DREAM, the best single-context methods score higher, which is expected when the inferred split provides little information about changes across environments. However, KRD still leads the cross-context methods on DREAM and Tübingen.
      2. We use two off-the-shelf anomaly detectors, GMM and LOF, to infer OOC. Both give the same qualitative result as the smooth partition (Appendix B.4, Table 11).
      3. We generate OOC samples with a controlled mean shift and provide their labels. KRD reaches nearly 100% pair accuracy on Tübingen, highlighting the importance of correctly identifying OOC samples from a mean shift (Table 3, lines 498–516).
    • (c) Uncertainty and benchmark results.

      1. Uncertainty. We add 95% confidence intervals for DREAM and Tübingen in Appendix B.4.
      2. TEP abstentions. MSS-hard leaves the direction undecided when the two directions tie, and CD-NOD can leave a pair unoriented. On TEP Fault 2, these two methods return a direction for only 14% of pairs. These outcomes follow from the decision rules and are not, by themselves, evidence of misconfiguration. We count abstentions as incorrect and also report MSS-soft, which uses the underlying pp-values to resolve ties.
      3. DREAM. As clarified above, we use DREAM to test the limits of KRD with inferred OOC labels. The best single-context methods perform better, as expected when the inferred split gives cross-context methods little information.
      4. Number of TEP experiments. We select the original three faults because they are the only ones for which the physical process structure clearly establishes the causal directions of the evaluated pairs. We add two more faults to Table 3, even though the process structure alone cannot establish the causal directions for these faults. KRD remains among the top methods.
AuthorsOpenReview ↗

Response to Reviewer vqD3 (Questions)

Response to questions

  1. Can the authors give a precise population-level condition under which each reverse residual distribution is invariant? Can they then show that the cases in Remark 1 are exhaustive?

    Yes. In the revised manuscript, lines 216–227 define precisely when each reverse residual distribution stays unchanged, and Table 1 (lines 208–215) exhaustively classifies the possible invariance scenarios.

  2. Can the authors provide a corrected threshold? Can they also extend the argument to control the reverse-regression, centering, and scale-estimation errors that Assumption 3's εfit\varepsilon_{\mathrm{fit}} includes?

    Yes. The revised proof in Appendix A.4 explicitly includes the previously omitted Gaussian absolute moment 2/π\sqrt{2/\pi} and simultaneously controls both OLS slopes, estimated centers, and normalization scales across all four residual coordinates and both contexts.

  3. How is ff fitted in the nonlinear experiments, and how does accuracy under cause shifts vary with δ\delta when the OOC cause values fall outside the IC support? Is Assumption 3 plausible for δ=2–4\delta=2\text{--}4 with nonlinear ff?

    We fit ff on IC data using ridge regression with a linear term and random Fourier features, tuned by leave-one-out validation. We add these details in Appendix B.1 (lines 1163–1192). Figure 4 shows lower accuracy of KRD at δ=4\delta=4 than at δ=2\delta=2 in the nonlinear cause-shift experiments, consistent with larger shifts making extrapolation harder. Theorem 2 guarantees recovery only when the total estimation error stays within its bound, as explained in the response to Weakness 1(c).

  4. Was the column order of the Tübingen and DREAM pairs randomized relative to the ground truth before partitioning?

    In the revised experiments, we choose the predictor uniformly at random for each pair before partitioning. KRD’s group accuracy on Tübingen drops from above 70% to about 60% (Table 2, lines 470–475). The best single-context methods score higher on both datasets, as expected when the inferred environments provide limited information for cross-context methods. However, KRD still leads the cross-context methods on Tübingen and remains competitive on DREAM. We also test two off-the-shelf anomaly detectors, GMM and LOF, and find the same qualitative pattern (Appendix B.4, Table 11). With correctly labeled OOC samples from a controlled mean shift, KRD reaches nearly 100% pair accuracy on Tübingen (Table 3, lines 498–516).

  5. Can the authors report confidence intervals or paired tests for the Tübingen and TEP comparisons? Can they also explain the below-chance accuracies of MSS-hard and CD-NOD on TEP? Are abstentions counted as errors?

    We report 95% confidence intervals in Appendix B.4, Table 10. As explained in our response to Weakness 2(c)(2), MSS-hard and CD-NOD often leave the causal direction undecided under their decision rules. Thus, we count these abstentions as incorrect when computing overall accuracy. Moreover, we also report MSS-soft, which uses the underlying pp-values to resolve the ties.

References

  1. Joris M. Mooij, Jonas Peters, Dominik Janzing, Jakob Zscheischler, and Bernhard Schölkopf. Distinguishing cause from effect using observational data: Methods and benchmarks. Journal of Machine Learning Research, 17(32):1–102, 2016.
  2. Rajen D. Shah and Jonas Peters. The hardness of conditional independence testing and the generalised covariance measure. 2020.

We thank the reviewer again for the care devoted to our manuscript and for the invaluable suggestions. Your comments give us a lot to think about and help us make the paper clearer and more rigorous. If the new results and clarifications address your concerns, we would be grateful if you would consider raising your rating. Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers

We thank all reviewers for their valuable feedback and for recognizing KRD’s theoretical insight and practical value. Please see our uploaded revision with changes highlighted in blue.


Strengths highlighted by reviewers

1. A simple score using one additional environment. Residual displacement turns a single environmental shift into evidence for causal direction. (e5Mt, vqD3, gWWx, kepW)

2. Linear-Gaussian insight. The two causal directions are easiest to distinguish when the causal signal and downstream noise are balanced, and harder to distinguish at either extreme. (vqD3, gWWx, e5Mt)

3. Competitive accuracy at lower cost. Across 18 simulation settings drawn from the published literature, many beyond our model assumptions, KRD is comparable to MSS on average and has lower reported runtime. (vqD3, e5Mt)


Shared concerns and our responses

The revisions clarify the scope of our claims while preserving the paper’s main conclusions and contributions.

  • Theoretical refinements (2)

    • Unclear characterization of degenerate cases (vqD3, gWWx, e5Mt, kepW). We use equality in distribution to give a precise and exhaustive classification of the three cases. These cases arise when structural cancellations or symmetries mask a shift, not from specially chosen counterexamples.
    • Error in the proof of Corollary 1 (vqD3, gWWx, e5Mt). We restore the missing Gaussian absolute moment and rederive the recovery threshold. The revised proof jointly controls errors from both regressions, centering, and scale estimation across all four residual coordinates and both environments.
  • Clarifications (1)

    • Lower accuracy with inferred OOC labels (vqD3, gWWx, e5Mt). The best single-context methods outperform KRD on DREAM and Tübingen. This is expected when the inferred environments provide little information about what changed.

      We clarify that our experiments are designed to test the limits of KRD by progressively relaxing the assumptions and reducing the information available about OOC. We do not aim to show that KRD always outperforms the baselines beyond our assumptions. Specifically, experiments are organized into the following four levels:

      • Controlled simulations. Our model assumptions hold.
      • Extended simulations. We test other causal mechanisms and noise distributions.
      • Labeled real data. TEP provides OOC labels, but OOC may involve more than mean shifts.
      • Unlabeled real data. For DREAM and Tübingen, we also have to infer OOC labels.
  • Corrections (1)

    • Column order in IC/OOC partitioning (vqD3, gWWx, e5Mt). We rerun DREAM and Tübingen with a randomly chosen predictor for each pair before constructing the environments.
  • Additional analyses (2)

    • Runtime with matched regression (gWWx, e5Mt). Both KRD variants have lower scoring and total runtimes than ANM-HSIC when all three use the same regression model.
    • Alternative IC/OOC partitions and known OOC labels. Two off-the-shelf anomaly detectors show the same pattern. With correctly labeled synthetic mean shifts, KRD achieves the highest accuracy on both DREAM and Tübingen.

Changes in the revised manuscript

  • Remark 1, Table 1, and Figure 2 define, classify, and illustrate the three degenerate cases. (vqD3, gWWx, e5Mt, kepW)
  • Corollary 1 and Appendix A.4 provide the corrected recovery threshold and proof. (vqD3, gWWx, e5Mt)
  • Table 2 reports corrected results with randomized predictors. Appendix B.4 adds GMM and LOF comparisons in Table 11 (vqD3, gWWx, e5Mt) and 95% confidence intervals in Table 10 (vqD3).
  • Table 3 adds synthetic OOC results for DREAM and Tübingen. Appendix B.5 describes how we generate and label those samples.
  • Table 3 also adds two exploratory TEP faults. (vqD3)
  • Appendix B.1 adds a subsection describing KRD’s regression implementation. (vqD3, gWWx, e5Mt, kepW)

Our individual responses address each reviewer’s specific concerns. We appreciate the detailed feedback, which helps us clarify our claims, strengthen the theoretical analysis, and broaden the evaluation.

AuthorsOpenReview ↗

Thank You to the Reviewers and Chairs

We appreciate the time and effort of every reviewer, AC, SAC, and PC member in supporting the review process and our community.

During the discussion, Reviewer kepW maintains a score of 3 with confidence 4 after considering our clarifications on concerns shared by other reviewers. Given the limited discussion time, we hope our responses can still reach the other three reviewers. Many thanks to all of you for your time and consideration.

We understand the challenges of reviewing in the AI era and value this community as a place where insights come together and inspire researchers across domains.

Thank you all again for your dedication!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.