acceptodds
Under review as a conference paper at ICLR 2027

Causal Partial Separability: Learning Nonlinear DAGs from Multivariate Functional Data

Abstract

In functional data analysis, learning directed acyclic graphs (DAGs) from multivariate functional data is crucial for identifying root causes. While recent causal discovery methods for multivariate functional data primarily rely on linear models, there is a lack of nonlinear acyclic models tailored for such data. The fundamental differences between scalars and functions prevent existing scalar nonlinear causal discovery methods from being directly applicable to functional data. To address these limitations, we propose Causal Partial Separability (CPS), a framework where functional variables admit a shared spectral basis and a common causal ordering, thereby mapping the problem to scalar DAG learning across the shared basis. We analyze the equivalence between functional DAGs and the union of blockwise DAGs, deriving the nonlinear identifiability of functional DAGs under the assumption of nonlinear Additive Noise Models (ANMs). Building on CPS, we introduce cps-fDAG, a functional DAG estimation method that combines shared spectral basis learning, searching for a common causal ordering, and hierarchical edge pruning. We establish the theoretical consistency for both the common ordering and edge pruning and provide a sufficient sample size condition. Experiments on synthetic datasets and biophysical fMRI simulation benchmarks demonstrate that cps-fDAG effectively learns causal DAGs for multivariate functional data.

Primary area:
causal reasoning
Keywords:
Causal discovery, Functional data analysis, Directed acyclic graph, Nonlinear regression, Separability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_p3JuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed a nonlinear functional DAG model for multivariate functional data. They use basis expansion and impose a nonlinear DAG model on each basis coefficient. The identifiability is largely based on existing nonlinear DAG result, which seems almost immediate to me. Then they propose an estimation method with consistency results.

Critical strengths

The consistency results appear rigorous.

Critical weaknesses

My main concern is the lack of methodological novelty. While the nonlinear functional DAG is new (if I have not missed any existing relevant work), its formulation is quite straightforward and expected. It seems to be really just applying ANM (in fact, it should be CAM because ANM allows non-additive causes, but this paper doesn't) multiple times, once to each basis coefficient, and taking the union of each individual resulting DAG. Although they do provide an estimation procedure and its associated consistency result, I don't think this clears the bar of this conference.

Questions for authors

In addition to concern about novelty, a few other questions I had while reading the paper are below.

  1. The paper assumes independence basis coefficients. Is that something implied from orthonormal basis or it is an additional assumption? If it is the latter, how to justify such assumption.
  2. The authors mentioned in Remark 1 that "We assume that the model is not directly formulated as an arbitrary ANM in the time domain as Equation 1, but rather under a partial separability basis as Equations 4 and 5". What's the connection between these two formulations?
  3. The union rule in (7) and (8) seems prone to false discoveries.
  4. Is it reasonable to assume X and E live in the same space given the nonlinear mapping?
Reviewer_ZV17OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose the methodology cps-FDAG for learning DAGs from functional data. It is based on a causal separability assumption which imposes zero covariance between the functional coefficients and noise components, assumes an additive non-linear model and that the block-wise graphs share the same topological ordering. Based on this assumption they build their methodology, consisting of an empirical estimation of the causal ordering and then a hierarchical edge pruning. The authors support it with theoretical analysis and experiments. They include both synthetic data and a simulated fMRI benchmark. An ablation study and robustness analysis are included as well.

Critical strengths

The area of focus of this contribution is interesting and significant. The authors present a methodology that bridges functional causal discovery and nonlinear DAG learning.

The paper structure is solid. The authors have structured and organized the related work, provided sufficient theoretical analysis and experimental evidence. The theoretical analysis, even though not 100% understood by the reviewer, looks significant and the experiments are extensive.

Critical weaknesses

First, the theory is hard to read and understand for this reviewer. In particular, for the proofs of theoretical results assumptions 1,2,3 are used, which are very technical and are not motivated by intuition in the text. It is not clear whether they are realistic in practice, or just facilitate the proofs of the theorems. Moreover, the computational complexity of finding the topological ordering is expensive as stated O(p!). Two strategies (lines 373-377) are followed to overcome this, but only little details are given for them.

The second weakness is in the experimental performance. Even though the experiments are quite extensive in the number of scenarios considered, the performance is not significant.

In particular, one notices in Table 1 an achieved SHD=10 on a graph with 20 nodes and undeclared number of edges. Above in the text it mentions average degrees 1,2,4. If the reported results correspond to average degree 1, then SHD =10 corresponds to half of the edges gotten wrong, which is in essence failure.

The method is only executed in very small graphs and there is no discussion about scalability.

Questions for authors

Some parts of the text seem too technical for this reviewer. For instance, some notions are not introduced or explained, which the reader can be unfamiliar with (e.g. RKHS). If you give me some more intuition I would have been more confident giving you a better rating.

Can you please give some motivation/intuition about the stated assumptions 1,2,3?

Also, Lemma 1 doesn’t make sense to this reviewer. What is given and what is being claimed here. Is the explanation a proof or not?

Can you please explain us whether we understood something wrong on the performance of your method, and the number of edges in the graph? Should it be considered as a failure in the case we describe? (see weakness above).

How are the hyperparameters of the baselines chosen in the experiments? The supplementary material in section B gives the choices, but doesn’t explain how they were chosen.

Reviewer_Sx7sOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper provides a new condition that they show allows identifying a DAG from functional data in the ANM setting without assuming linearity. It performs well compared to alternatives on simulated data, including 'biologically plausible' data and ablations.

Critical strengths
  1. Causal discovery from functional data is an important open problem, and this paper addresses it without assuming linearity like existing approaches.
Critical weaknesses
  1. As an expert in causal discovery and not functional data analysis, I found this paper quite dense and hard to follow. A gentler introduction (even via references or a separate section in the appendix) to the model class would help a lot.
  2. The experimental results don't establish the usefulness of the method. It's not clear how the scale of graphs in the experiments compares to the scale of graphs in relevant application domains, and I don't see any mention of running time.
Questions for authors
  1. Can you provide a clear, concise algorithm description? I would've expected to see this as an algorithm environment in the main text.
  2. What role does assumed sparsity play in the algorithm? The simulations only evaluate on quite sparse DAGs.
  3. How computationally expensive is the method? Can it be run on real fMRI data?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.