It is nice to see the theoretical analysis in the paper, however, while I can see that these analysis (proposition 1 to 3) provide some useful interpretation of the proposed quantum modulator. I do have some concern how and whether they represent substantial new theoretical contributions. Let’s take Proposition 2 as an example, it is about Re-uploading expands potentially accessible spectral support, I feel this result has certain connection (at conceptual level) with what found in Schuld et al. (2021). For this work about classical dynamic model, I thin what particularly interesting should be something more clear more clear about what new theoretical understanding is obtained specifically from the interaction between quantum modulation and the recurrent temporal system.
Another question/concern is about the big gap between the quantum models and the classical liquid backbone. While the author did noticed and claimed that this paper is not about a claim of predictive quantum advantage. I still felt this is a big concern, as this differences is too large to ignore what is the causes. I would like the authors to explore/explain why the liquid backbone performs so poorly compared with the GRU.
It is also noticed that: (Table 1) the classical liquid backbone has 3,396 trainable parameters, while the quantum variants have 4,652 parameters. Therefore, is it possible that the improvement of the no-entanglement quantum model over the classical liquid model come (or partially) from the additional nonlinear transformations and trainable parameters? If not, can the authors please show this with further exploration?
The authors concluded that data re-uploading and entanglement contribute different functional structures. However, the experiment results mainly compare the final classification performance of different circuit variants. I do not think these results are sufficient to show that the proposed theoretical mechanisms are actually responsible for the observed performance differences. While the authors did provide some theoretical analysis (as discussed before), can any corresponding experiments be designed and concluded to support this particularly?
Finally, the experiments are conducted on only one dataset, which is originally regularly sampled (as shown in the summary of the paper at beginning) and then modified to create irregular temporal observations. Although this provides a controlled setting, I think it is not sufficient to demonstrate the general usefulness of the proposed approach for irregular time-series learning. Why not use real irregular datasets? Or at least on both modified and real datasets and at least two each category.