- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
Finite-interval sampling obscures the distinction between dynamical propagation and synchronous dependencies, making joint structure recovery from irregularly sampled data challenging. This work introduces CAST, a framework that separates these mechanisms.
Critical strengths
- The research direction appears interesting.
- At first glance, the mathematics appears correct; however, I have not verified all the proofs.
Critical weaknesses
- I believe the paper is not reproducible in its current form. It should include a description of the model architecture, as neither the main text nor the appendix makes it clear that a neural network is used. I inspected the accompanying source code and found that the model uses a simple MLP, but this should be explicitly described in the paper.
- The paper presents numerous proofs and a novel methodology. However, even after reading it several times, I remained unsure about the research questions and the motivation behind the proposed approach. I suggest stating the research questions explicitly in the introduction and explaining how they motivate the methodology.
Questions for authors
- The paper contains quite a few theorems and propositions. How do they relate to the experimental evaluation?
- The paper states: “The directed dynamics graph describes how interactions propagate across time, whereas the undirected synchronous dependence graph encodes conditional dependence among innovations at the same instant. We refer to them as the dynamics and synchronous graphs, respectively.” How does this description relate to Figure 1? The figure appears to distinguish the two graphs only by whether their edges are directed or undirected, without clearly illustrating the temporal aspect.
- Line 119: What specific conditions are meant by “standard regularity conditions”? Please state these conditions explicitly or provide an appropriate reference.
- Equation 5: Please consider adjusting the formatting so that the brace labelled “demixing” does not appear between the numerator and the fraction bar.
- Line 319: The evaluation includes several baselines. Please explain why these particular methods were selected.
- In my view, placing the related work section after the experiments is unusual. Presenting it earlier could help readers understand the context and positioning of the proposed method.
4 comments
I want to thank you for the detailed reply. However, please take note that ICLR allows updating the submission with explicitly marked corrections. I remain unconvinced until explicit changes are made in the text.
The revised PDF is uploaded, with changes in blue. Each item of our earlier reply is now in the text.
Section 4.2 now describes the score network and its training. The network is a three-hidden-layer MLP with SiLU activations. Tables 8 and 9 in Appendix E.2 give the training and readout settings for CAST and the matched Gaussian, including the settings that vary by benchmark.
The introduction now starts with the difficulty of separating the two graphs when each observed transition mixes their effects. This motivates the two-block readout. The introduction then explains why we compare the learned score with direct estimates of the transition mean and a Gaussian transition model.
Table 6 in Appendix E explains how the theoretical results relate to the experiments. For example, Proposition 4.1 shows that at a fixed gap, the cross block alone cannot identify both graphs. The readout ablation measures what curvature adds in finite samples. The fraction of pairs with only a synchronous edge that are incorrectly reported as dynamics edges falls from 10% with the raw cross block to 2% with full demixing.
Figure 1 now has a time axis. Dynamics edges run across a short time step, and synchronous edges connect innovations at the same instant. Section 3 states the regularity conditions, and Appendix A proves the two limits in Eq. (2) and checks the conditions for our generator. We moved the demixing brace in Eq. (5) so it covers only the demixing product. Section 6 explains which family each baseline represents, and related work has moved to Section 2.
I want to thank you for addressing my remarks. I need some time to reevaluate the manuscript. However, I am grateful to the authors for the work done.
Response to reviewer Mxmh
Thank you for checking the code. For the main synthetic experiment, the score network is an MLP with three hidden layers, SiLU activations, and width max(192, 16D), where D is the state dimension. It takes the noised increment, the starting state, and two 32-dimensional sinusoidal embeddings for the log gap and noise level. It predicts the injected noise. We use a 100-step linear DDPM schedule with beta ranging from 0.0001 to 0.02. We train with Adam at a learning rate of 0.0003 for 6,000 updates, with batches of 256 and without weight decay or EMA.
Training uses 40,000 transition pairs from 100 segments at eight log-spaced gaps between 0.02 and 0.5. The readout uses 512 held-out states per gap at gaps of 0.076 and 0.2, with eight perturbation draws per state at each of four noise levels (alpha-bar of 0.988, 0.908, 0.769, and 0.610). The Gaussian control uses the same data, optimizer, and training budget, with an MLP of the same depth and width for the mean and a covariance head that depends only on the gap. We will add these settings to the paper.
The introduction will ask whether both graphs can be recovered from one transition model, how curvature separates dynamical sensitivity from innovation precision, and when a learned score improves on direct Gaussian estimation. Table 2 compares the raw cross block, diagonal scaling, and full demixing, with dynamics AUPRC of 0.793, 0.854, and 0.967. The appendix checks the finite-gap expansion on the population readout. The irregular-sampling experiment compares direct transition learning with impute-then-differentiate controls. The strong-noise DblMass experiment uses larger readout gaps, as motivated by Proposition C.7. The main tables report empirical recovery accuracy. Corollary 4.1 gives the conditions for exact recovery. We will add a table showing how the theoretical results relate to these experiments.
We will add a time axis to Figure 1. The dynamics arrows will connect successive times. The synchronous edges will connect innovation components at the same time. The figure will also show the observed transition over a gap, which combines both effects.
For the limits in Eq. (2), one sufficient condition is a continuously differentiable, globally Lipschitz drift and a constant positive-definite innovation covariance. The main synthetic drift includes quadratic terms. Its acyclic ordering and linear self-decay allow moment bounds to be obtained recursively along the ordering. In the simulator, each coordinate is clipped to the interval from -8 to 8 after each integration step, as described in the appendix. We will make the distinction between the continuous-time assumptions and the numerical protocol explicit.
PCMCI+ and VARLiNGAM represent conditional-independence testing and non-Gaussian structural VAR, respectively. DYNOTEARS provides a score-based dynamic Bayesian network comparison. NGM and CaDyT cover continuous-time neural ODE modeling and derivative-based discovery. The matched Gaussian tests whether the learned score improves on direct Gaussian transition estimation. We will explain these choices, move related work before the method, and move the demixing brace outside the fraction in Eq. (5).