acceptodds
Under review as a conference paper at ICLR 2027

Identifiable Hierarchical Factor Disentanglement in Dual Non-IID Time Series

Abstract

Many nonstationary time series contain two coupled latent scales: a persistent sequence-level environment and recurrent mechanism states within each sequence. Flat switching representations conflate these sources of variation, obscuring the hierarchy, state cardinalities, and cross-environment state correspondence. We formalize this setting as dual non-IID time series and study when the persistent–recurrent hierarchy is identifiable from unlabeled sequences. Under identifiable within-sequence switching and translated invariant transition signatures, we recover the environment partition, both unknown cardinalities, and a shared indexing of recurrent states across environments. With sufficient transition variation, dynamic coordinates are identifiable up to a component permutation and componentwise transformations within contexts. Cross-context anchoring yields a global chart, and unique minimal target sufficiency further identifies the target-relevant subset. We also characterize representative recovery objects and ambiguities at the corresponding identification boundaries. These recovery stages motivate Hierarchical Dual Non-IID Component Estimation (HiDICE), a hierarchical variational state-space model with persistent–recurrent inference and additive signature sharing. Controlled experiments reproduce the predicted hierarchy recovery and stagewise behavior at the identification boundaries. Among evaluated label-free methods, HiDICE performs best under shortcut reversal on HPR-MNIST. It also achieves the lowest mean leave-one-subject-out heart rate error on PPG-DaLiA.

Primary area:
learning on time series and dynamical systems
Keywords:
nonstationary time series, identifiability, hierarchical representation learning, disentangled representation learning, switching dynamical systems

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_xp7AOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the identifiability probelm of the persistent-recurrent hierarchy in dual non-iid time series and proposes a new model for latent factor recovery.

Critical strengths
  1. The paper investigates an important identifiability problem by considering persistent environments and recurrent mechanisms.
Critical weaknesses
  1. The theoretical results rely on numerous conditions and assumptions that are much difficult to understand and lack sufficient intuitions and motivations. Their relationships to existing literature are also unclear, making it difficult to assess whether the proposed method effectively addresses the identifiability problem.
Questions for authors
  1. How are Conditions A-C and Assumptions 1-2 related to the existing literature? Please carefully clarify these conditions to help readers better understand the positioning of this work.
  2. The authors should formulate the theoretical results mathematically rather than relying on verbal descriptions (e.g., Theorems 1 and 2). The current presentation makes these results difficult to understand.
4 comments
AuthorsOpenReview ↗

Response to Reviewer xp7A (Part 1/2)

Response to Reviewer xp7A

We sincerely thank the reviewer for recognizing the importance of the identifiability problem and for highlighting the need for clearer motivations, mathematical statements, and connections to prior work. We have revised the presentation around the distinct recovery stages and made their mathematical relationships more explicit. Manuscript references in our responses use the revised version; reviewer quotations retain their original numbering.

1. Roles of Conditions A–C and Assumptions 1–2

Reviewer question.

How are Conditions A-C and Assumptions 1-2 related to the existing literature? Please carefully clarify these conditions to help readers better understand the positioning of this work.

Response. The conditions address different ambiguities. We now separate the discrete hierarchy result from coordinate recovery and target-dependent refinement, rather than presenting their assumptions as a single undifferentiated requirement.

Condition or assumptionAmbiguity it resolvesRole and connection to prior work
Condition AWithin-sequence switching-state ambiguityA unique minimal path mixture, coverage, and distinct kernels identify local state quotients up to permutation. This uses the switching/finite-mixture identification route associated with Balsells-Rodas et al. [1]. It supplies a local starting point, not cross-environment correspondence.
Assumption 1Persistent environment partition and correspondence among local state quotientsA fixed background-invariant observable functional gives translated rows Qn=bEn+W\mathcal Q_n=b_{E_n}+\mathcal W, with injective environment and recurrent contributions. Translation rigidity and centering then identify the hierarchy across sequences. This is the distinctive additional step in Theorem 1.
Condition BMixing of continuous coordinates and reassignment between dynamic and static factorsFactorized innovations and sufficiently rich contrast derivatives separate coordinates; regularity, connected support, and dynamic indecomposability delimit the admissible alternatives. The derivative-rank route is related to nonlinear ICA and temporal identification [2,3], here applied to the aligned recurrent-state contrasts.
Condition CDifferent coordinate charts across static contextsUnique relative alignment on overlapping contexts and connectivity yield one global chart. A symmetry of the shared transition family can obstruct this alignment, as illustrated by Counterexample 6. This condition is additional to within-context identification.
Assumption 2Ambiguity among target-sufficient coordinate subsetsWith the population joint law of (X,Y)(X,Y) observed, unique inclusion-minimal target sufficiency selects a subset in the global chart. This is a separate target-dependent refinement, not a premise of hierarchy identification.

The resulting logical order is

\text{A + Assumption 1 + coverage} &\Longrightarrow \text{persistent–recurrent hierarchy},\\\\ \text{the hierarchy result + B} &\Longrightarrow \text{context-wise dynamic coordinates},\\\\ \text{context-wise recovery + C} &\Longrightarrow \text{a global coordinate chart},\\\\ \text{global recovery + the observed }(X,Y)\text{ law + Assumption 2} &\Longrightarrow \text{the target-relevant subset}. \end{aligned}$$ These statements use the population, regularity, and coverage conventions of Section 2.1 and Appendix C.1. The common tools from [1–3] are acknowledged explicitly in the revised Appendix B (p. 16). Our contribution is the additional identification of the unlabeled hierarchy and shared state correspondence, followed by a clear separation of the requirements for context-wise, global, and target-dependent recovery. To provide intuition for Assumption 1, Section 2.3 and the new Appendix C.5 (p. 22) give an observable channel with conditional mean $m_0(h)+u_e+v_r$, independent of the static background. Averaging over common histories produces a translated signature. Nonlinear history dependence and nonadditive conditional covariance remain possible. This supplies a nontrivial sufficient family while keeping the condition's scope explicit.
AuthorsOpenReview ↗

Response to Reviewer xp7A (Part 2/2)

2. Mathematical formulation of Theorems 1 and 2

Reviewer question.

The authors should formulate the theoretical results mathematically rather than relying on verbal descriptions (e.g., Theorems 1 and 2). The current presentation makes these results difficult to understand.

Response. We agree that the identification targets and allowed transformations should be directly visible. Below we state the mathematical content and indicate the corresponding presentation changes.

Theorem 1 concerns one global hierarchy correspondence. Let ϑ∼Xθ\vartheta\sim_X\theta denote admissible parameterizations with the same observed sequence laws under the conventions in Appendix C.1. Under Condition A, Assumption 1, and coverage, the two parameterizations have the same KE,KRK_E,K_R and environment partition up to naming. There exist separate global permutations πE\pi_E and πR\pi_R such that the identified state-indexed objects obey

A~πR(r),πR(r′)(πE(e))=Ar,r′(e),μ~πE(e),πR(r),s~=μe,r,s.\widetilde A^{(\pi_E(e))}_{\pi_R(r),\pi_R(r')}=A^{(e)}_{r,r'},\qquad \widetilde\mu_{\pi_E(e),\pi_R(r),\widetilde s}=\mu_{e,r,s}.

on the corresponding covered contexts and common histories. These relations appear in Eq. (45), Appendix D.2 (p. 25). The same πR\pi_R applies across covered environments, whereas sequence-local recovery permits different permutations. Lemma D.4 identifies KRK_R locally; the translated rows identify the environment partition and KEK_E, then align the recurrent classes. We have expanded the Section 2.3 proof sketch using

\qquad \Delta_{n,i}-\bar\Delta_n=w_r-\bar w,$$ which shows why correspondence need not be supplied before alignment. **Theorem 2 concerns coordinate equivalence.** Under Theorem 1, the population regime, and Condition B, every admissible $X$-equivalent parameterization satisfies $$\widetilde S=\kappa(S),\qquad \widetilde Z_{t,j}=a_{S,j}\\!\left(Z_{t,\pi_S(j)}\right),$$ where $\kappa$ is a bimeasurable bijection, $\pi_S$ is a coordinate permutation, and each $a_{S,j}$ is a componentwise $C^2$ diffeomorphism. Under Condition C, a single permutation and a single set of componentwise maps apply globally. We have moved this equivalence, previously stated immediately after the theorem, into Theorem 2 itself as Eq. (12) (p. 5), together with the definitions of the transformations. For completeness, Corollary 1 additionally uses equality of the joint $(X,Y)$ law and Assumption 2. Under the global coordinate permutation, the identified target subset transforms as $\widetilde J_C=\pi^{-1}(J_C)$. This is representational subset identification; a causal interpretation requires the additional structural condition described in Section 3.2. The proofs and representative ambiguities are given in Appendices D–F. These revisions make the assumptions, identifiable objects, and remaining transformations explicit without changing the distinction between population guarantees and finite-sample estimation. We thank the reviewer again and hope the clarified mathematical statements and literature connections support a reassessment of the theoretical contribution. ## References 1. Balsells-Rodas, C., Wang, Y., and Li, Y. [On the Identifiability of Switching Dynamical Systems](https://proceedings.mlr.press/v235/balsells-rodas24a.html). ICML, 2024. 2. Hyvärinen, A., Sasaki, H., and Turner, R. E. [Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning](https://proceedings.mlr.press/v89/hyvarinen19a.html). AISTATS, 2019. 3. Yao, W., Chen, G., and Zhang, K. [Temporally Disentangled Representation Learning](https://proceedings.neurips.cc/paper_files/paper/2022/hash/a938292feb86b94ebe3e6200ff7786ef-Abstract-Conference.html). NeurIPS, 2022.
Reviewer_xp7AOpenReview ↗

Response to authors

I thank the authors for their detailed responses and revisions. The theoretical results are now clearer. However, I still find the overall presentation difficult to follow. The assumptions are too dense, making it difficult to assess the significance of the contribution. In addition, the changes are not highlighted in the revised paper. I will keep my original score.

Thank you for your time and thoughtful feedback. We sincerely appreciate your comments and suggestions.

Reviewer_nhJ1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HiDICE, a hierarchical framework for identifying and disentangling persistent environments and recurrent mechanism states in nonstationary time series. It addresses a fundamental question: Can we mathematically guarantee that a model recovers the underlying latent factors rather than learning an arbitrary representation that predicts well? The authors establish identifiability guarantees under specific assumptions and evaluate the method on synthetic and real-world datasets. Overall, the paper presents a potentially meaningful extension of existing identifiability theory by establishing conditions under which persistent environments and recurrent mechanisms can be separately identified across environments.

Critical strengths
  1. A meaningful identifiability result for persistent–recurrent hierarchical structure. The strongest contribution is the theoretical treatment of jointly identifying persistent environments and recurrent mechanism states when their labels and cardinalities are unknown. The translated-signature formulation provides an interesting way to establish cross-environment state correspondence.

  2. HiDICE incorporates architectural biases motivated by the theoretical analysis, including hierarchical state inference, signature alignment, and factorized dynamics. The controlled experiments evaluate different aspects of latent-factor recovery, while the experiments involving shortcut reversal and generalization to unseen subjects provide additional evidence of predictive robustness under distribution shifts. The ablation studies and experiments examining the boundaries of the theoretical assumptions further strengthen the evaluation.

Critical weaknesses

The central identification result depends on a potentially restrictive additive-signature assumption. Assumption 1 requires an observable functional of the state-conditional transition kernels to admit an additive decomposition into environment-specific and recurrent-state-specific contributions. Although the model allows environment-dependent recurrent transitions and the synthetic generator includes an interaction term, it remains unclear how broadly this signature-level assumption holds when environments and mechanisms interact in more general nonlinear or context-dependent ways.

Questions for authors
  1. While Figure 5 provides evidence of robustness to nuisance perturbations, can you further demonstrate that the inferred persistent and recurrent states correspond to meaningful real-world mechanisms? It would be helpful to visualize the learned state assignments and transition patterns, particularly their consistency across participants or environments (considering real-world data), to provide stronger evidence of hierarchical disentanglement.

  2. Several baseline methods have missing entries for recurrent-state recovery and target leakage in Table 2(a). Could you explain whether these metrics are unavailable because the corresponding methods do not explicitly estimate recurrent states or because they were not evaluated? Where possible, could comparable diagnostic evaluations or adapted baseline results be provided?

  3. (relevant to the above limitation) How restrictive is Assumption 1 in real-world data, especially when environments and mechanisms interact nonlinearly? Could the authors evaluate how recovery changes when the additive-signature assumption is gradually violated?

  4. Can HiDICE handle mechanisms that occur only once, are missing in some environments, or are active simultaneously? In particular, can it separate multiple independent mechanisms rather than treating their combinations as different states?

4 comments
AuthorsOpenReview ↗

Response to Reviewer nhJ1 (Part 1/2)

Response to Reviewer nhJ1

We sincerely thank the reviewer for recognizing the value of identifying the persistent–recurrent hierarchy when labels and cardinalities are unknown. We have added evidence on nonadditive sensitivity and clarified metric availability and the interpretation of recurrent states. Manuscript references in our responses use the revised version; reviewer quotations retain their original numbering.

1. Scope of Assumption 1 and nonadditive interactions

Reviewer question.

(relevant to the above limitation) How restrictive is Assumption 1 in real-world data, especially when environments and mechanisms interact nonlinearly? Could the authors evaluate how recovery changes when the additive-signature assumption is gradually violated?

Response. Assumption 1 requires additivity of a selected observable transition signature. It permits environment-dependent recurrent-state transition laws and does not require every aspect of the dynamics to be additive. We have made its scope more concrete in Appendix C.5 (p. 22). The sufficient family has a fixed observable channel with conditional mean m0(h)+ue+vrm_0(h)+u_e+v_r, independent of the static background. Averaging over a common history measure yields be+wrb_e+w_r. The history dependence can be nonlinear, and the conditional covariance can contain environment–mechanism interactions. Interactions that destroy the selected signature's additive structure are outside the exact guarantee; the example does not establish that arbitrary real-world data satisfy the assumption.

To examine empirical sensitivity, we have added the perturbation sweep in Table 2 (p. 9) and Appendix H.2.3 (p. 37). It perturbs the generated signature array as

\qquad \lambda\in\\{0,0.25,0.5,1,2\\}.$$ Under the matched protocol, recovery changes as follows, with mean ± s.d. over five seeds. | Interaction strength $\lambda$ | Environment ARI | Global recurrent-state F1 | |---|---:|---:| | 0.00 | 0.886 ± 0.046 | 0.852 ± 0.033 | | 0.25 | 0.821 ± 0.042 | 0.783 ± 0.035 | | 0.50 | 0.842 ± 0.038 | 0.724 ± 0.047 | | 1.00 | 0.789 ± 0.049 | 0.696 ± 0.041 | | 2.00 | 0.733 ± 0.037 | 0.561 ± 0.028 | Recovery remains informative at the smaller tested perturbations, while stronger perturbations impair global state correspondence. Environment recovery also decreases overall. This directly characterizes sensitivity to departures from the shared additive template. The residual in Table 2 measures the departure of the fixed generated signatures from additivity; it is not an estimate of an arbitrary real-data observable functional. We present this sweep as an empirical diagnostic, without extending Theorem 1 to general nonadditive settings. Under its stated conditions, Theorem 1 identifies the persistent–recurrent hierarchy and a shared recurrent-state indexing across environments without supplied environment labels, state labels, or cardinalities. This additional identification result remains the central theoretical contribution. ## 2. Learned states and real-world mechanism semantics **Reviewer question.** > While Figure 5 provides evidence of robustness to nuisance perturbations, can you further demonstrate that the inferred persistent and recurrent states correspond to meaningful real-world mechanisms? It would be helpful to visualize the learned state assignments and transition patterns, particularly their consistency across participants or environments (considering real-world data), to provide stronger evidence of hierarchical disentanglement. **Response.** We agree that the controlled factor-swap results and real-world mechanism semantics require different evidence. The HPR-MNIST swap controls examine factor roles in a constructed setting. The PPG-DaLiA experiments evaluate predictive transfer across participant–activity compositions and held-out participants; they do not establish physiological identities for the inferred states. We have not added the requested learned-state visualizations or an external physiological validation. Accordingly, Appendix H.4 (p. 40) and the Conclusion now explicitly delimit the real-data interpretation. Consistency of learned assignments could be informative about representation structure, but independently validating a physiological mechanism would require evidence beyond state labels or predictive performance alone. Our theoretical claim remains identification of the covered latent hierarchy up to the stated equivalences under the specified conditions. The real-data contribution is the demonstrated predictive transfer, rather than a claim to have discovered physiologically validated mechanisms.
AuthorsOpenReview ↗

Response to Reviewer nhJ1 (Part 2/2)

3. Missing state-recovery and leakage entries

Reviewer question.

Several baseline methods have missing entries for recurrent-state recovery and target leakage in Table 2(a). Could you explain whether these metrics are unavailable because the corresponding methods do not explicitly estimate recurrent states or because they were not evaluated? Where possible, could comparable diagnostic evaluations or adapted baseline results be provided?

Response. The dashes indicate that the corresponding comparable output is not defined by the evaluated method; they do not indicate failed runs. The original HPR-MNIST subtable is now Table 3. Recurrent-state F1 requires a recurrent-state output, while the reported leakage diagnostic uses the designated residual/static representation (Zresid,S)(Z^{\mathrm{resid}},S). A predictor or representation method without those objects cannot be assigned the same diagnostic by substituting an arbitrary hidden vector or clustering rule.

For baselines that define the relevant objects, the diagnostics are reported. For example, the flat AR-HMM has both globally aligned recurrent-state F1 and target-leakage entries in Table 3. The post-hoc hierarchy mapping used where applicable is identified by the existing superscript convention. Appendix H.1 (p. 35) now explicitly explains the output requirements and missing entries. This revision clarifies the existing comparison; we do not claim new adapted-baseline results. All methods in Table 3 retain their reported shortcut-reversal performance, which provides the common downstream comparison.

4. One-time, missing, and simultaneously active mechanisms

Reviewer question.

Can HiDICE handle mechanisms that occur only once, are missing in some environments, or are active simultaneously? In particular, can it separate multiple independent mechanisms rather than treating their combinations as different states?

Response. These cases have different implications, which we now distinguish in Appendix F.1.

  • A state seen once in a finite sample. It may still have positive population probability and belong to the model. The population identifiability result does not guarantee reliable recovery from one finite-sample occurrence.
  • A state absent from some population environments. The full shared-row result relies on the stated state and environment coverage. Under partial visitation, only supported correspondences can be constrained, and disconnected overlap patterns can retain separate alignment ambiguities. Appendix F.1 characterizes these partial-support limitations.
  • Several independent mechanisms active simultaneously. A categorical recurrent state can represent their joint activation pattern, but our theorem does not separately recover the independent factors within that pattern. Such factorial mechanism identification is a distinct extension beyond the current recurrent-state quotient.

Thus the present guarantee concerns covered persistent environments and recurrent states, with an explicit boundary on independent simultaneous-mechanism decomposition. This distinction preserves the interpretation of the identified hierarchy without overstating what a categorical state model establishes.

We sincerely appreciate the reviewer's constructive suggestions. We hope the new sensitivity experiment and clearer statements about outputs, coverage, and mechanism semantics address the remaining concerns and strengthen confidence in the paper's contributions.

Thank you for providing the sensitivity experiment that tests recovery under non-additive interactions. The authors have clarified the limitations of the work and explained the theoretical boundaries. However, it is still unclear how violations of the additive assumption affect environment partition recovery in real-world data, and whether the sensitivity observed in the synthetic experiments also holds in real-world settings. In Table 4, the improvements on PPG-DaLiA compared to PULSE (2026) and REx (2021) are relatively small, and it is unclear whether these differences are statistically significant. The limited gains on real-world data, together with the lack of direct evidence that the learned state assignments correspond to meaningful physiological mechanisms, are the main reasons I maintain my original score.

We sincerely thank the reviewer for the thoughtful follow-up, careful consideration of our revisions, and constructive feedback throughout the review process.

Our work primarily focuses on establishing the identifiability of the persistent–recurrent hierarchy under explicit conditions. The controlled recovery experiments, ablation studies, and real-world distribution-shift evaluations provide complementary evidence for the proposed framework, with HiDICE achieving the lowest mean MAE across all three PPG-DaLiA evaluation settings.

Regarding physiological interpretation, our theoretical guarantees concern latent structure identifiability rather than the physiological semantics of individual states. Establishing such physiological interpretations and their consistency across participants is not covered by our current theoretical guarantees and represents a valuable direction for future work.

We are grateful for the reviewer's insightful suggestions, which have helped us clarify the scope of our contributions and strengthen the manuscript. Thank you again for your time and consideration.

Reviewer_2DLvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies time series with environments that remain fixed within each sequence, switching mechanisms that can recur over time, and static backgrounds specific to each sequence. It provides conditions for identifying these hidden factors, determining the unknown numbers of environments and mechanism states, and aligning the factors across sequences. These results guide the design of HiDICE, a hierarchical variational state-space model that uses additive signature sharing and factorized dynamics, with an option to select factors relevant to the prediction target. Experiments on synthetic data, HPR-MNIST, and PPG-DaLiA show improvements in latent-factor recovery and prediction under distribution shifts.

Critical strengths
  1. The paper distinguishes between recovering switching states up to arbitrary labels and identifying shared mechanisms across environments. It provides a specific sufficient condition based on signatures for matching these states across environments. This distinction is conceptually useful because it clarifies which aspects of the hierarchical structure require additional assumptions.
  2. The experiments evaluate latent-factor recovery on synthetic data, shortcut reversal on HPR-MNIST, and cross-subject heart-rate prediction on PPG-DaLiA. The reported gains across these settings provide solid evidence of the practical value of the overall HiDICE framework and strengthen the paper’s contribution.
Critical weaknesses
  1. The theoretical novelty and significance are not sufficiently established. The within-sequence recovery stage builds on existing identifiability results for switching systems, while the derivative-rank argument for continuous coordinates closely follows work on auxiliary-variable nonlinear ICA. The most distinctive contribution appears to be the alignment of states across environments using an exact, observable additive signature that is invariant to static backgrounds (Assumption 1). However, the paper does not clearly explain when this strong assumption holds in the motivating applications. In addition, the current discussion does not provide a sufficiently precise theorem-level comparison with [1–3], particularly regarding assumptions, identified objects, and proof techniques. The significance of the contribution depends on whether the new assumption holds for a meaningful, well-motivated class of models and whether the resulting decomposition enables useful capabilities.
  2. The experiments do not isolate the practical contribution of signature-based cross-environment alignment. The signatures regularized in Eq. (14) are derived from fitted latent transition distributions, but their relationship to the observable, background-invariant functional required by Assumption 1 remains unclear. Table 8 shows that removing signature decomposition hurts state recovery on synthetic data. However, it does not report how this removal affects prediction on HPR-MNIST or PPG-DaLiA. Thus, the reported gains do not establish whether the proposed alignment contributes beyond the architecture and other training objectives. This gap limits the evidence for the practical significance of the paper’s central theoretical contribution.

References:

[1] Carles Balsells-Rodas, Yixin Wang, and Yingzhen Li. On the identifiability of switching dynamical systems. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp. 2639–2672. PMLR, 2024.

[2] Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. Nonlinear ICA using auxiliary variables and generalized contrastive learning. In Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Learning Research, pp. 859–868. PMLR, 2019.

[3] Weiran Yao, Guangyi Chen, and Kun Zhang. Temporally disentangled representation learning. In Advances in Neural Information Processing Systems, volume 35, pp. 26492–26503, 2022.

Questions for authors
  1. Could you explicitly define the observable functional Γ\Gamma for the controlled SCM in Eqs. (54)–(55), show that it is invariant to SS and admits the required additive decomposition, and explain how it relates to the fitted latent signatures in Eq. (14)?
  2. In the anchoring experiment, what changes in the generative model cause Condition C to fail? Context-specific reparameterizations that preserve the observed distribution do not, by themselves, change identifiability within the same admissible model class. Please clarify which assumptions change in this experiment.
  3. How does removing signature decomposition affect HPR-MNIST reversal accuracy and PPG-DaLiA LOSO error when the rest of the architecture and training objectives are held fixed?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 2DLv (Part 1/3)

Response to Reviewer 2DLv

We sincerely thank the reviewer for recognizing the distinction between local switching recovery and shared mechanisms across environments, and for identifying where the theoretical positioning and experimental attribution needed to be clearer. We have strengthened both aspects. Manuscript references in our responses use the revised version; reviewer quotations retain their original numbering.

1. Theoretical novelty and relation to prior work

Response. Our central theoretical contribution is the joint identification of the persistent–recurrent hierarchy from sequences whose environment labels, recurrent-state labels, and cardinalities are unknown. Under Condition A, Assumption 1, and environment coverage, Theorem 1 identifies the environment partition, both cardinalities, shared cross-environment recurrent-state indexing, transition laws, and state-conditional kernels, up to separate environment and state permutations.

The comparison with prior work is:

Prior workConditions, identified objects, and proof toolsHiDICE’s additional identification
Balsells-Rodas et al. [1]Identifies Markov switching models up to state permutation via finite-mixture arguments, including separated analytic Gaussian transitions; identifies switching dynamical systems up to affine transformations under additional observation-map conditions.These guarantees do not by themselves determine a persistent–recurrent factorization, the persistent environment partition, or shared recurrent-state correspondence across independently identified local models.
Hyvärinen et al. [2]Uses conditional independence given an observed auxiliary variable and sufficiently rich derivative variation to identify nonlinear components up to permutation and componentwise invertible transformations.Does not identify the unknown persistent–recurrent hierarchy. HiDICE identifies state correspondence before constructing transition contrasts for coordinate recovery.
Yao et al. [3]Uses conditional independence given latent history and sufficient derivative variation across histories or contexts to identify temporal components up to permutation and componentwise invertible transformations.Does not establish the joint persistent-environment partition and shared recurrent-state correspondence in Theorem 1.

Our proof explicitly connects local and global identification. Finite-mixture component matching is applied to within-sequence path laws; disintegration and state-kernel separation then recover the local recurrent-state quotient and KRK_R. These steps leave a separate label permutation for each sequence. Assumption 1 supplies the observable structure that resolves this remaining correspondence ambiguity. Each sequence yields an unordered signature row satisfying

\bar\Delta_n=b_{E_n}+\bar w,\qquad \Delta_{n,i}-\bar\Delta_n=w_r-\bar w.$$ Equality of complete rows identifies environments because a finite nonempty set has no nonzero translation symmetry; the number of distinct rows identifies $K_E$. Centering matches the distinct recurrent signatures without requiring their correspondence in advance, yielding one shared state indexing across environments. Equal centroid weights accommodate environment-dependent recurrent-state frequencies. The resulting correspondence aligns both the state-conditional kernels and their associated transition laws. Appendix C.5 makes Assumption 1 concrete through a sufficient family with a fixed observable channel whose conditional mean is $m_0(h)+u_e+v_r$, independent of the static background. Averaging over common histories yields the required translated signature. This family permits nonlinear history dependence and nonadditive environment–mechanism interactions in conditional covariance. The guarantee concerns kernels admitting this observable signature structure; the construction and its relation to fitted signatures are detailed below. The translated-signature argument thus establishes a common discrete hierarchy and supplies the aligned transition contrasts for the subsequent coordinate analysis. Theorem 2 uses these contrasts under Condition B, with Condition C providing the additional requirements for a global chart. Section 2.3 and Appendix B clarify the contribution and comparison; Appendix D gives the hierarchy-identification proof.
AuthorsOpenReview ↗

Response to Reviewer 2DLv (Part 2/3)

2. Observable signatures, the controlled SCM, and fitted features

Reviewer question.

Could you explicitly define the observable functional Γ for the controlled SCM in Eqs. (54)–(55), show that it is invariant to S and admits the required additive decomposition, and explain how it relates to the fitted latent signatures in Eq. (14)?

Response. We agree that the original presentation did not sufficiently distinguish a population observable functional from the fitted features used by the estimator. The revision gives an explicit sufficient observable family and separately identifies the nonlinear controlled SCM as a recovery benchmark. We do not claim that the construction below proves Assumption 1 for the arbitrary nonlinear observation mixing used in that benchmark.

In Appendix C.5, let Dt=ψ(Xt)D_t=\psi(X_t) be a fixed observable channel satisfying

\qquad \mathbb E[\varepsilon_t\mid H_t,E=e,R_t=r,S=s]=0.$$ For a fixed probability measure $\nu$ on a common history region where the kernels are identifiable and the integrals are finite, define $$\Gamma(\mu)=\int\\!\left[\int\psi(x)\\,\mu(dx\mid h)\right]\nu(dh).$$ Then $$\Gamma(\mu_{e,r,s}) =\underbrace{\int m_0(h)\nu(dh)+u_e}_{b_e} +\underbrace{v_r}_{w_r}.$$ This functional uses neither latent labels nor static-context values as inputs. It is background invariant, and injective environment and state contributions satisfy the required separation. The history function can be nonlinear, and the conditional covariance can contain nonadditive environment–mechanism interactions. Thus the additive requirement concerns a selected observable summary, while substantial nonlinear structure remains possible outside that summary. The fitted conditional-mean and log-marginal-variance features in Section 4 are model-based surrogates regularized toward the same translated-row geometry. Equality to the population functional is not assumed. Section 4 (pp. 6–7), Appendix G.1, and Appendix H.2.1 now make this distinction explicit. In particular, the controlled SCM provides finite-sample recovery and sensitivity evidence; its latent signature checks do not establish an observable functional after arbitrary nonlinear observation mixing. This separates the mathematical sufficiency claim, its nontrivial example, and the practical inductive bias being evaluated. ## 3. What the cross-context diagnostic establishes **Reviewer question.** > In the anchoring experiment, what changes in the generative model cause Condition C to fail? Context-specific reparameterizations that preserve the observed distribution do not, by themselves, change identifiability within the same admissible model class. Please clarify which assumptions change in this experiment. **Response.** We agree with this distinction and have corrected the interpretation. The numerical experiment does not establish that Condition C alone has been violated. It applies context-specific coordinate permutations and componentwise invertible transformations, with the observation and transition representations composed accordingly, and compares separate within-context matchings with one global matching. Section 5.2 and Figure 3(c) therefore now describe a coordinate-alignment diagnostic, replacing the original “without anchoring” interpretation. Appendix H.2.2 (p. 36) explicitly states its evidential limit. The reported within-context and global MCC values, $0.844\pm0.033$ and $0.593\pm0.071$, diagnose the distinction between local recovery and a shared chart; they are not presented as an isolated intervention on Condition C. The formal nonuniqueness argument is given separately in Counterexample 6 (Appendix F.3, p. 31). A transition family with a sign symmetry admits the identity chart in one static context and a sign-reversed chart in another while preserving the observed law. No single chart covers both contexts. This identifies the symmetry excluded by Condition C, independently of the numerical diagnostic.
AuthorsOpenReview ↗

Response to Reviewer 2DLv (Part 3/3)

4. Signature decomposition beyond the hierarchical architecture

Reviewer question.

How does removing signature decomposition affect HPR-MNIST reversal accuracy and PPG-DaLiA LOSO error when the rest of the architecture and training objectives are held fixed?

Response. We have added the requested downstream comparison. The no-signature-decomposition control retains both discrete branches and removes signature-based alignment. Appendix H.1 specifies the matched architecture and evaluation protocol, including the temporal backbone, decoder family, dataset partitions, model-selection procedure, and five matched seeds.

The results are reported in Table 10, Panel A (Appendix H.5, p. 42), and summarized in Section 5.4. Values are mean ± s.d.; the last column reports the difference of the displayed means.

EndpointFull HiDICEWithout signature decompositionAblated − Full
Controlled SCM environment ARI0.886 ± 0.0460.647 ± 0.091−0.239
Controlled SCM global recurrent-state F10.852 ± 0.0330.702 ± 0.068−0.150
HPR-MNIST reversal accuracy0.788 ± 0.0490.657 ± 0.074−0.131
PPG-DaLiA LOSO MAE, BPM4.513 ± 0.1264.982 ± 0.166+0.469

The same control affects both hierarchy recovery and downstream performance under two different shifts. These results support the practical contribution of signature alignment beyond retaining the hierarchical branches. We interpret this as component-level evidence for the proposed inductive bias, without attributing all performance gains to a population identifiability guarantee.

We thank the reviewer again for the constructive feedback. We hope the clearer identification argument, explicit sufficient family, corrected diagnostic interpretation, and additional downstream ablation address the main concerns and support a reassessment of the contribution.

References

  1. Balsells-Rodas, C., Wang, Y., and Li, Y. On the Identifiability of Switching Dynamical Systems. ICML, 2024.
  2. Hyvärinen, A., Sasaki, H., and Turner, R. E. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. AISTATS, 2019.
  3. Yao, W., Chen, G., and Zhang, K. Temporally Disentangled Representation Learning. NeurIPS, 2022.

Comments on the paper

1 comment

Dear Reviewers, The authors have now submitted their responses to the reviews. There is only a limited amount of time left for you to review their replies and provide any further comments.

Your feedback is crucial to ensure that the review process proceeds smoothly. We would greatly appreciate it if you could share your reactions before the discussion period closes.

Thank you very much for your work. Best regards.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.