acceptodds
Under review as a conference paper at ICLR 2027

Quantile Geometry: Finite-Rank Drift Ratios for Causal Discovery

Abstract

Unknown nonlinear transformations of an effect can obscure the structure used to distinguish cause from effect. We study the drift ratio, which measures how the effect’s conditional quantiles move as the cause changes, relative to their local spacing. This quantity is unchanged by smooth increasing transformations of the effect. We introduce finite-rank drift-ratio (FRDR) models, which assume that the drift-ratio surface has finite rank, meaning that its variation is captured by finitely many shared functions of the quantile level. The framework includes additive noise models, post-nonlinear models, location-scale noise models, and post-nonlinear heteroscedastic noise models. Under suitable conditions, we prove generic identifiability of the causal direction in FRDR models: the reverse drift ratio has rank greater than any fixed finite bound, apart from exceptional distributions of the cause. We derive a quantile geometry (QG) score, QG-FRDR, that applies across these model classes, along with four specialized scores that exploit their additional structure. These scores require neither noise reconstruction nor estimation of the unknown transformation. Experiments demonstrate fast computation for QG-FRDR and strong directional accuracy for its weak formulation on structural-model simulations. The class-specific QG-ANM, QG-LSNM, and QG-PNL scores achieve high accuracy within their target classes.

Primary area:
causal reasoning
Keywords:
causal discovery, post-nonlinear models, monotone transformations, generic identifiability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vbQTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks which of two variables is the cause and which is the effect when the effect may have been transformed in an unknown nonlinear way. It studies how the effect's percentiles shift as the possible cause changes, relative to the spacing between those percentiles. This quantity remains the same after a smooth, increasing transformation of the effect. The authors prove that a simple pattern in this quantity can often identify the causal direction, then turn the idea into several practical scores.

Critical strengths
  1. The transformation-resistant quantity is an interesting way to connect several familiar causal models in one framework. The paper also gives a formal argument for why the causal direction can often be identified under its assumptions.

  2. The authors report both successes and failures. On simulated data, their more robust general score reaches 93.2% accuracy, close to a strong comparator's 95.0% (Table 3/ Panel A), while their simpler score runs much faster.

Critical weaknesses
  1. The two general scores perform poorly on the broader benchmark collection. In Table 4, they reach 50.8% and 54.8% average accuracy across 19 cause-effect benchmark groups, the lowest two results among seven methods. The best comparator reaches 69.1%. The faster score is close to chance, while the more accurate score takes about 2.8 seconds per pair. The strong simulation results therefore do not yet show a practical advantage across varied data. The authors should explain which kinds of data cause these failures and how a user could recognize them in advance.

  2. The theorem does not say how much evidence a finite dataset needs to distinguish the two directions. It says the wrong direction generally has a different pattern, but that difference can be very small. In some simulations, the simpler score chooses the correct direction only 0-4% of the time despite the theoretical distinction. The paper links this to small differences and estimation errors. An analysis of the size of the difference, even in a restricted case, would better connect the theorem to the method used in practice.

Questions for authors
  1. Which benchmark groups account for the low Table 4 averages, and can a user tell when either general score is likely to work?
  2. In the settings where the simpler score gets only 0-4% correct, how small is the difference between directions? Would changing the assumed number of shared patterns improve the result?
2 comments

Thank you for the constructive review. References follow the revised manuscript.

W1 and Q1. Benchmark performance and recognizing failures

All four class-specific scores are now evaluated without class selection. QG-LSNM reaches 69.6% at 0.014 s per pair, comparable to the strongest baselines (68.3–69.1%, 0.047–2.6 s). For LSNMs, its generic identifiability follows from Theorem 1 through Proposition 2.

Table 11 identifies where the FRDR scores fail. Equal-group averages are:

Benchmark groupsQG-FRDRQG-FRDR-WQG-LSNMQPE-kLRQS fix-gCVEL
Four SIM groups, Cha, Multi, Net, Tübingen67.8%68.7%78.2%75.5%75.1%63.8%
Per, Sig, Vex15.7%34.0%55.0%76.3%65.3%86.7%
Four QPE-generated groups48.0%48.0%75.3%57.5%74.0%87.8%
Four DREAM4 groups45.9%49.3%57.8%62.7%53.3%44.2%
All 19 groups50.8%54.8%69.6%69.1%68.7%68.3%

The FRDR deficits are concentrated in seven groups simulated from other model families and in the in-silico DREAM4 networks.

Xi et al. (2025) [1] designed Per/Sig/Vex so that existing model classes are misspecified, and the QPE-generated groups use prescribed-basis QPE models [2].

We have not determined whether these groups have finite-rank drift ratios; estimation error can also cause failures within the model class, as discussed below.

Recognizing failures. Appendix F.6.3 tests abstention using the normalized score gap cc. For QG-LSNM, benchmark accuracy rises from 69.6% to 71.2% at the primary threshold c≥0.05c\geq0.05 (90.6% of pairs decided) and to 76.5% at c≥0.20c\geq0.20 (67.2% decided).

For QG-FRDR-W, c≥0.05c\geq0.05 improves accuracy on the five linear-location LSNMs from 79.8% to 85.4%, retaining 82.6% of pairs.

The gap does not detect the FRDR reversals on Per/Sig/Vex: their more confident decisions are less accurate.

Section 8.3 therefore recommends QG-LSNM for an unknown class; reliable failure detection for the FRDR scores remains open.

W2 and Q2. Population gaps and rank

Gap size. Appendix F.7 (original Table 14, now Table 19) computes population scores numerically on a 50×5050\times50 grid.

For the five failing mechanisms with Gaussian cause and noise, the causal rank-two score is zero and the reverse scores are 0.0007–0.0088: less than 1% of squared norm lies outside the reverse rank-two approximation.

Separately, the noise-balanced experiments at n=1000n=1000 give mean estimated scores of 0.026–0.033 causally and 0.014–0.018 in reverse. This wrong ordering is consistent with direction-dependent estimation error and the observed 0–4% accuracy.

Changing rank. On the transformed versions of these mechanisms, K=2,3,4K=2,3,4 gives QG-FRDR accuracies of 1.6%, 1.4%, and 1.0%, and QG-FRDR-W accuracies of 79.8%, 70.4%, and 66.0% (Table 13). Increasing KK therefore does not resolve these failures.

Once KK covers the causal rank, increasing it can only reduce the reverse population rank score; K=1K=1 instead favors the reverse direction for all five mechanisms (Table 19).

Analysis of gap size. We do not have a theoretical analysis on the size of the directional gap. To our knowledge, the closest methods (CVEL, QPE, LRQS) also share this limitation, which is an important direction for future work.

[1] Xi et al., Distinguishing Cause from Effect with Causal Velocity Models, ICML 2025.

[2] Chen et al., Causal discovery via quantile partial effect, ICLR 2026.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer vbQT,

Thank you again for your review. Could you let us know whether our revision and rebuttal have addressed your concerns and whether you would consider updating your assessment? We are happy to clarify any remaining concerns.

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

This paper tackles the problem of distinguishing between cause and effect in bivariate distributions. The central assumption is a low-rank structure in the drift ratio, which is defined to be the ratio of the partial derivatives of the conditional quantile function with respect to the cause xx and quantile uu respectively. The authors prove that such an assumption leads to a rank bound in the reverse direction for only a finite union of finite-parameter families (i.e. the cause-effect direction is identifiable generically, as the authors state it). The authors derive a quantile geometry score (QG) with the finite-rank drift-ratio (FRDR) assumption for different model classes and demonstrate the efficacy of their methods on both synthetic and established benchmark datasets.

Critical strengths
  1. The paper recognizes and analyzes a very interesting direction for identifying cause-effect structure in bivariate distributions. By making the drift ratio the primary object of study, the authors are able to derive an easier-to-compute estimator for analyzing the causal direction.
  2. Theorem 1 provides a concrete identifiability result for the causal direction when the low-rank structure holds.
  3. Some variants of the proposed method are computationally very efficient.
Critical weaknesses
  1. The presentation of the paper makes it quite challenging to follow along. For example, in the introduction, the authors talk about drift ratio in terms of Q(x,u)Q(x, u) without ever clearly defining what QQ is. The reader has to infer from context (after much effort) that QQ is the conditional quantile. 

The authors also state:

For Y~=h(Y)\widetilde{Y}=h(Y) with h’>0h’ > 0, both derivatives acquire the factor h’(Q(x,u))h’(Q(x,u)), which cancels from RR.

This is very difficult to understand without context and only makes sense to the readers familiar with the proof of Lemma 1.

In general, I felt like understanding the paper was challenging since it lacked the proper exposition to the concepts used and discussed in the right order. The description of the estimation (Section 7) could also benefit from an algorithm block.

  1. The empirical results are not very convincing on the real-world benchmark datasets, especially compared to the synthetic datasets. This leads me to believe that the low-rank assumption, although interesting from a theoretical perspective, might not be expected to hold realistically. I also find the lack of experimentation with the rank KK quite surprising. The authors use K=2K=2 almost everywhere throughout the paper. It seems to be a very important parameter so an ablation study with higher values of KK would be very illustrative.

  2. I don’t quite understand the point of the different class-specific causal scores. It seems like they are only applicable for certain synthetic family settings, however I don’t see how we would know when the data comes from this family apriori.

Questions for authors
  1. Do the authors have an explanation for why their models perform poorly on the real-world datasets? Any insights on which assumptions of their model are violated would be very useful.
  2. What happens when higher values of rank KK are used?
  3. When would we expect to use the different class-specific causal scores?
  4. How would the method handle cases where XX and YY are independent?
2 comments

Thank you for the constructive review. References follow the revised manuscript.

W1. Presentation

Section 1 now defines Q(x,u)Q(x,u) at first use, displays the invariance derivation, and motivates FRDR through familiar models before defining it. Section 5.2 derives the weak objective step by step. Section 7 contains Algorithm 1 and the new Algorithm 2 for QG-FRDR-W.

W2 and Q1. Benchmark performance and assumptions

All four class-specific scores are now evaluated on every benchmark without class selection. QG-LSNM reaches 69.6% at 0.014 s per pair, comparable to the strongest baselines (68.3–69.1%, 0.047–2.6 s).

Tübingen is the only group made up entirely of real measurements. QG-LSNM reaches 86.9% there (86/99 pairs, unweighted), QG-FRDR 63.6%, and QG-FRDR-W 48.5%.

The FRDR deficits are concentrated in seven groups simulated from other model families (Per/Sig/Vex and the QPE-generated groups) and in the in-silico DREAM4 networks.

On the other eight groups, QG-FRDR and QG-FRDR-W average 67.8% and 68.7%, fourth and third of the seven general-purpose methods (Table 11).

We have not established which groups violate the rank assumption.

Failures also occur within the model class, where Appendix F.7 points to small directional gaps and estimation error.

W2 and Q2. Larger rank

The default K=2K=2 covers all four structural classes. Appendix F.4 compares K=2,3,4K=2,3,4 on identical samples (Tables 12–13):

EvaluationQG-FRDR, K=2/3/4K=2/3/4QG-FRDR-W, K=2/3/4K=2/3/4
Seven selected settings88.3 / 87.3 / 87.4%88.1 / 88.1 / 86.9%
Five linear-location PNL-HNM settings1.6 / 1.4 / 1.0%79.8 / 70.4 / 66.0%

Higher rank gives no overall gain here and increases the weak score's runtime from 2.8 to 21.4 s per pair on the seven settings. The separate K=1K=1 experiment lowers QG-FRDR's overall accuracy from 75.6% to 69.4% (Table 15).

W3 and Q3. Specialized scores

These scores test stronger identities of their respective classes, with generic identification following from Theorem 1 through Proposition 2. Stronger constraints can improve finite-sample separation: on five linear-location LSNM settings, QG-FRDR achieves 0–4%, QG-FRDR-W 69–88%, and QG-LSNM 97–100% (Tables 8–9).

Section 8.3 recommends a class-specific score when its class is justified, QG-LSNM for an unknown class, and QG-FRDR-W when an unknown increasing effect transformation is plausible. Automatic selection remains open.

Q4. Independent variables

Under independence, both drift ratios vanish and all population scores tie at zero; Theorem 1's nondegeneracy condition fails. Our procedure orients dependent pairs. A separate independence test can screen out independent pairs first.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer NYep,

Thank you again for your review. Could you let us know whether our revision and rebuttal have addressed your concerns and whether you would consider updating your assessment? We are happy to clarify any remaining concerns.

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

This paper propose finite-rank drift-ratio (FRDR) models for bivariate causal discovery without hidden confounders. The key idea is to use a ratio derived from conditional quantiles that is invariant to smooth increasing transformations of the effect, bringing several noise models into on framework. Under suitable assumptions, the authors establish basis-free generic identifiability and develop causal scores without reconstructing noise or estimating the unknown transformation.

Critical strengths

S1. A unified theoretical perspective. FRDR brings several causal model into one framework and establishes generic identifiability under assumptions without prescribing the reverse basis function.

S2. Efficient scoring. The scores avoid noise reconstruction and output transformation estimation and QG-FRDR is the fastest among the methods.

Critical weaknesses

W1. The model-specific scores are evaluated with oracle knowledge of the model class. QG-ANM, QG-LSNM, QG-PNL, and QG-PNL-HNM are evaluated only on their matched synthetic families, so the reported 99.3–100% accuracies assume the true model class is known in advance. The paper provides no procedure for choosing among the five scores without this knowledge. It also does not evaluate the class-specific scores on mismatched families or on the real benchmarks. As a result, these accuracies are not comparable to the general-purpose methods in Panel A, and their practical value remains unclear.

Questions for authors

Q1. What explains QG-FRDR’s below-chance accuracy on Per, Vex, and Sig? Do normalized score gaps distinguish correct from incorrect predictions well enough to support abstention?

Q2. Holding p(Y∣X)p(Y\mid X) fixed, how do QG-FRDR and QG-FRDR-W perform as pXp_X changes from Gaussian to uniform, skewed, or multimodal distributions?

Q3. When the model family is unknown, is there a data-driven criterion for choosing between general-purpose and class-specific scores? How do the class-specific scores perform on mismatched synthetic families and heterogeneous benchmarks?

2 comments

Thank you for the constructive review. References follow the revised manuscript.

W1 and Q3. Class knowledge and score selection

All four class-specific scores are now evaluated on all 48 synthetic settings and 19 benchmark groups, with fixed settings and no class selection (Tables 3–4):

ScoreSynthetic accuracyBenchmark accuracy
QG-ANM42.5%59.4%
QG-LSNM80.9%69.6%
QG-PNL74.1%53.9%
QG-PNL-HNM73.8%48.7%

QG-LSNM is comparable to the strongest benchmark baselines (68.3–69.1%) at 0.014 s per pair versus 0.047–2.6 s. Mismatch can hurt: QG-ANM falls to 33–41% on three other families, and QG-LSNM reaches 59.7% on PNL.

Section 8.3 recommends QG-LSNM when the class is unknown and QG-FRDR-W when an unknown increasing effect transformation is plausible. The latter reaches 99.4% on PNL and 93.4% on PNL-HNM. Automatic selection remains open (Section 9).

Q1. Systematic reversals and abstention

The reversal extends beyond the rank criterion: on Per/Sig/Vex, QG-PNL and QG-PNL-HNM reach only 3–28% and 4–18%, while QG-LSNM reaches 61%, 50%, and 54% (Table 11).

Several estimated criteria therefore favor the reverse direction. Whether this ordering comes from population structure or estimation error remains unresolved.

Similar failures occur within the model class, where Appendix F.7 points to direction-dependent estimation error.

Appendix F.6.3 tests abstention using

c=∣S(X→Y)−S(Y→X)∣S(X→Y)+S(Y→X).c=\frac{|S(X\to Y)-S(Y\to X)|}{S(X\to Y)+S(Y\to X)}.

For QG-LSNM, the benchmark area under the decision-rate curve is 80.2%, against 69.6% all-pair accuracy.

At the primary threshold c≥0.05c\geq0.05, accuracy is 71.2% with 90.6% of pairs decided; at c≥0.20c\geq0.20, it is 76.5% with 67.2% decided.

On Per/Sig/Vex, it rises from 55.0% to 56.3% (83.7% decided) and to 65.2% (54.7% decided) at the same two thresholds (Tables 17–18).

The gap gives only limited gains for the FRDR scores and fails to detect their reversals on Per/Sig/Vex, where more confident decisions are less accurate.

Q2. Changing the cause distribution

Appendix F.5 fixes P(Y∣X)P(Y\mid X) for eight analytic LSNMs with Gaussian noise and varies the cause distribution, each with mean zero and unit variance (Table 14):

Cause distributionQG-FRDRQG-FRDR-W
Gaussian39.9%87.9%
Uniform41.9%89.9%
Skewed Beta(2,5)41.0%84.6%
Bimodal Gaussian mixture49.5%89.3%

Differences across mechanisms dominate those across cause distributions: QG-FRDR's failures on the five linear-location mechanisms persist under all four causes.

Thank you for the thorough response and the additional experiments. I have checked the revised results for each of my points.

W1/Q3. Evaluating all four class-specific scores on every synthetic setting and benchmark group without class selection addresses my concern. The guidance in Section 8.3 (QG-LSNM when the class is unknown, QG-FRDR-W when an effect transformation is plausible) is helpful, and I appreciate the clear acknowledgment that automatic selection remains open. One observation: the strongest benchmark result now comes from QG-LSNM (69.6%) rather than the general FRDR scores (50.8% and 54.8%). It would help to state this relationship explicitly in the revision, so readers understand the respective roles of the general FRDR framework and the class-specific scores in practice.

Q1. Thank you for the candid analysis. The finding that the reversals on Per/Sig/Vex extend beyond the rank criterion, and that confidence-based abstention does not detect them, is useful information for readers. I encourage including this failure mode explicitly in the limitations.

Q2. The experiment varying the cause distribution with P(Y∣X)P(Y\mid X) fixed directly answers my question. The results show that performance is driven more by the mechanism than by the cause distribution.

Overall, the response addresses my questions, and I maintain my positive assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

We thank the reviewers for their constructive comments and for recognizing the unified perspective and the identifiability result. The revision addresses the shared concerns about class knowledge, benchmark performance, and exposition. All references follow the revised manuscript. The main changes in the revised manuscript are highlighted in blue.

  1. Evaluation without class knowledge (Tables 3, 4, and 11). All four class-specific scores are now evaluated on all 48 synthetic settings and 19 benchmark groups, with fixed settings and no class selection. QG-LSNM reaches 80.9% and 69.6%, respectively, at 0.014 s per pair. The strongest benchmark baselines reach 68.3–69.1% at 0.047–2.6 s per pair.

  2. When the scores work (Sections 8.3 and 9). We recommend QG-LSNM for an unknown class and QG-FRDR-W when an unknown increasing effect transformation is plausible: it reaches 99.4% on PNL and 93.4% on PNL-HNM. Table 11 shows that the FRDR failures are concentrated in seven groups simulated from other model families (Per/Sig/Vex and the QPE-generated groups) and in the in-silico DREAM4 networks. Automatic score selection remains open.

  3. Robustness and abstention (Appendices F.4–F.6). Higher ranks K=3,4K=3,4 give no overall gain on the tested subsets. With eight conditional mechanisms fixed, QG-FRDR-W averages 84.6–89.9% across Gaussian, uniform, skewed, and bimodal causes. Confidence-based abstention raises QG-LSNM's benchmark accuracy from 69.6% to 71.2% at the primary confidence threshold (90.6% of pairs decided) and to 76.5% at the strictest one (67.2% decided), but does not detect the FRDR reversals on Per/Sig/Vex.

  4. Exposition (Sections 1, 5.2, 6, and 7). We define conditional quantiles at first use, display the invariance derivation, explain the specialized scores, and derive the weak objective step by step. Both algorithms now appear in Section 7. The abstract and conclusion reflect the expanded evaluation. The proof sketch of Theorem 1 was removed from the main text to make room; the full proof remains in Appendix C.2, and the theoretical results are unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.