acceptodds
Under review as a conference paper at ICLR 2027

Operational Memory: Predictive Spectra and Realization

Abstract

A state representation of a dynamical system is typically judged by how accurately it predicts future observations. In systems with memory, however, two histories with the same observed present can respond differently to the same intervention. A model may therefore predict trajectories accurately while discarding historical information that remains predictive of intervention responses. We call this information Operational Memory. We formalize this distinction and show that accurate trajectory prediction need not imply accurate prediction of response differences under intervention. To characterize Operational Memory, we introduce the Operational Predictive Memory Spectrum (OPMS), which decomposes the dependence of future intervention responses on history into predictive directions; a context-resolved form prevents cancellation when history–response relationships vary across matched contexts. We further develop Operation-Complete Predictive Realization (OCPR), which constructs a finite predictive state whose time and intervention updates can be identified and reused recursively. In two spin–boson environments, updates identified using responses with at most two interventions compose to predict responses after a third intervention, with Causal Skill 0.966/0.971. On superconducting qubit data, context resolution reveals dependence missed by the pooled spectrum; an independent quantum jump dataset supports the same principle.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
non-Markovian dynamics, predictive state representations, system identification, interventional prediction, quantum dynamics

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_F8DjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper introduces several concepts related to predicting the future of a system under interventions. Operational Memory is taken to be the historical information of a system that is predictive of its response to interventions. A CCA decomposition of history features and future response features gives singular values termed the Operational Predictive Memory Spectrum. A context-resolved version, which conditions on context and prevents cancellation between different contexts, is also presented. Lastly, a concrete realization of a finite predictive state variable under linearity assumptions is termed the Operation-Complete Predictive Realization. Evaluation is presented on physical systems, where the primary evaluation metric is an R2R^2 metric of predicted response differences under different histories.

Critical strengths

Generating representations that are predictive of future trajectories under different interventions is an important research direction. The proposed CCA-based approach is sound and reasonable under the stated linearity assumptions.

Critical weaknesses

Major clarity issues: The methods and experiments sections are very difficult to follow. Many concepts are simply presented mathematically and under-explained in the text. Additional terms like context CtC_t, context strata β\beta, and intervention depth are undefined. In the experiments section, there seems to be insufficient detail to interpret all the presented results, at least for non-experts in this precise area. For example, TLS is defined as two-level-system in the caption of Figure 2, but "fitted TLS fraction" is not explained anywhere in the main text or appendix, from what I can see. Table 1 contains raw numbers without context (e.g. 0.006256 > 0.000468). Both sections could benefit from explaining less using more space.

Low significance and novelty: The core methods are CCA (Eq. 3) and CCA with conditional covariances (Eq. 4). I do not believe these methods and associated results are substantially novel and significant to an ICLR audience. Citations are provided for modern sequence-to-sequence SSMs such as S4D and S5, but no meaningful comparison is given, either conceptually or experimentally.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer F8Dj (1/2)

We thank the reviewer for carefully reading our paper and for raising important questions about our method. We have carefully considered the concerns regarding presentation, significance, novelty, and the relationship to modern state-space models, and revised the manuscript accordingly.

W1. What is novel beyond established CCA and Hankel realization tools?

For the unclear presentation in the original manuscript, we have substantially revised the Methods and Experiments. Matched context, context strata, and intervention depth are now defined before use, and experimental quantities such as the fitted TLS fraction and matched-null thresholds are explicitly explained.

Regarding novelty, we agree that CCA, conditional covariance, and classical Hankel realization are established tools, and we do not claim these mathematical tools themselves as new. We also apologize that the original presentation did not sufficiently distinguish these established mathematical ingredients from our actual contributions, which may have made our work more difficult to understand.

Our contribution has three levels.

First, we formalize Operational Memory: historical information that remains predictive of responses to the same future intervention after the observable present is matched. Proposition 1 proves that ordinary prediction does not certify this information: trajectory R2R^2 can approach one while intervention-response recovery becomes arbitrarily poor.

Second, OPMS characterizes this operational dependence. CCA provides its canonical geometry, while the context-resolved construction places residual responses from different context strata into separate blocks,

QCR=eβ⊗Q∘,Q_{\mathrm{CR}}=e_\beta\otimes Q^\circ,

before spectral analysis. This prevents cross-context cancellation. Proposition 3 shows that the context-resolved statistic is zero only when every retained within-context cross-covariance is zero, whereas the pooled statistic may vanish despite predictive dependence in every context.

Third, OCPR is separate from CCA. It constructs a finite predictive state with operation-specific updates identified from limited intervention depth. OCPR combines exact intervention coordinates, center–gap realization, and depth-restricted update identification, together with restricted-rank and finite-horizon perturbation guarantees.

The central experiment identifies all updates using responses with at most two interventions, excludes third-intervention responses from identification, and then composes the same updates to predict the unseen third intervention. This achieves Causal Skill 0.993/0.998 with response-derived states and 0.966/0.971 with history-inferred states.

The downstream decision experiment further shows that these distinctions matter: among 12 cases where matched histories require different optimal interventions, history-inferred OCPR recovers both actions in 11 cases and response-derived OCPR in all 12 (App. H.1, Table 7).

We have revised the manuscript to make the separation between established mathematical tools and these problem, construction, realization, and theoretical contributions much clearer.

AuthorsOpenReview ↗

Response to Reviewer F8Dj (2/2)

W2. A meaningful comparison with S4D, S5, and recursive-state models was missing.

The original manuscript cited S4D and S5 without providing a sufficiently meaningful comparison. We have now added a controlled comparison of our implementations of TCN, S4D-style, and S5-style history encoders using identical 64-step inputs, response targets, five random seeds, and a shared 16-dimensional response representation:

EncoderIntermediateStructured
TCN0.910 ± 0.0220.867 ± 0.028
S4D-style0.889 ± 0.0260.849 ± 0.027
S5-style0.915 ± 0.0260.909 ± 0.024

Values are five-seed mean ± sample standard deviation of Causal Skill; higher is better.

The result shows that modern SSM encoders can indeed retain substantial intervention-relevant historical information, with S5-style performing best in this controlled encoder comparison. We therefore do not use this experiment to claim that OCPR simply outperforms S4D/S5.

The comparison addresses a different question. These encoders directly map observed history to a fixed future-response representation. OCPR additionally identifies operation-specific predictive-state updates from limited intervention depth and recursively composes those updates at unseen intervention depth. The S4D/S5 comparison does not use this same recursive-identification protocol, so the scores should not be interpreted as a direct method-level ranking against OCPR (App. G.2, Table 5).

Beyond this encoder comparison, the revision also adds recursive-state baselines that more directly probe OCPR’s central claim. Two learned 16-dimensional recurrent controls retain high trajectory R2R^2 but lose matched-history response-difference geometry at held-out horizons (App. G.1, Fig. 13).

We also added an external OQE/system–memory baseline. At the held-out depth-three endpoint, OQE retains ordinary-response R2=0.915/0.875R^2=0.915/0.875 but obtains Causal Skill −10.15/−1.06-10.15/-1.06, compared with 0.994/0.9970.994/0.997 for center–gap OCPR on the same 30-pair response-derived comparison (App. G.3, Table 6). Because OQE uses a different response representation and information interface, we report it as an external comparison rather than a strictly matched ablation.

This experiment has been added to the supplementary material, and the conceptual distinction is now stated explicitly in the revised Related Work.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer again for the thoughtful comments and suggestions. We hope these revisions address the reviewer’s concerns and that the reviewer will reconsider their assessment in light of the revised manuscript.

Reviewer_3nNmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on the following issue: for the same state representation, different histories may lead to different responses under interventions. Previous research generally uses the accuracy of predicting the future of dynamic systems as the criterion for state representations, but for such compressed state representations of past histories, can they still make accurate predictions when interventions are introduced? This paper proposes the Operational Predictive Memory Spectrum (OPMS) to characterize history, and uses Operation-Complete Predictive Realization (OCPR) to improve prediction results. The authors conduct small-scale experiments on both classical and quantum systems to demonstrate the advantages of their approach. However, the paper is very difficult to read. Many symbols appear suddenly without any explanation, and too much background knowledge is assumed but not introduced, significantly impacting readability.

Critical strengths
  1. The paper points out an important question: whether compressed representations of historical state, under interventions, can still accurately predict the future.
Critical weaknesses
  1. The paper is too heavy on physics background, making it very difficult for a pure machine learning audience to understand. Given this, the authors should have paid more attention to exposition. There should not be abrupt mathematical symbols as in line 50, such as, “In particular, trajectory R2R^{2} can approach,” with no explanation as to what these symbols mean. Also, new terms such as “Causal Skill” in lines 84-85 are introduced without definition, and values like “0.966/0.971” later on are not explained clearly.

  2. Similarly, in Figure 1, used as a main example, uu is not defined; “free evolution” is left unexplained; and the measurement operators in the middle are also not described.

  3. Too many unexplained new terms are introduced, such as “Causal skill.” According to the definition, it seems to just reflect the difference between the model’s predictions and the true differences, normalized by the true differences, but it is not clear what it has to do with “skill.”

  4. The paper presents experiments with superconducting qubits, but it is unclear how the quantum state is actually obtained. For systems with many qubits (e.g., 100 qubits), quantum state tomography is extremely expensive, and this cost does not seem to be considered in the paper.

Questions for authors
  1. Can the authors clarify the practical use cases for the tasks presented in this paper? It seems that the scenarios are cherry-picked just for the sake of telling a story.

  2. Can the authors explain, in lines 300-301, where it says, “To access the same state from observable history, a temporal encoder maps the 64-step history to features, ϕ(h)∈R64\phi(h) \in \mathbb{R}^{64}”? Why does a 64-step history correspond to a 64-dimensional space? And if a high-dimensional history is compressed to a lower dimension, wouldn’t that cause significant information loss, especially for the histories of multi-qubit systems? How is this compression performed?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 3nNm (1/3)

We thank the reviewer for carefully reading our paper and for the detailed comments. We have revised the manuscript to improve its clarity and strengthen the practical motivation.

W1 / Q1. The presentation was difficult to follow, and the practical motivation was unclear.

We have substantially revised the Introduction, Method, and experimental descriptions to improve clarity. Figure 1 now provides a clearer illustration of the matched-history comparison and future interventions. We also define Causal Skill explicitly and explain its interpretation relative to a zero-difference predictor. A score of 1 denotes exact recovery, 0 matches the reference, and negative values indicate larger errors. The term “skill” reflects predictive improvement over this reference. The revised abstract now expresses the scores 0.966/0.971 as a 96.6%/97.1% reduction in squared response-difference error relative to the reference predictor.

Regarding practical use, our target setting is prediction followed by intervention or control in dynamical systems with memory. In a non-Markovian system, two histories can reach essentially the same observed present while retaining different hidden environmental or memory states. The same future action can therefore produce different responses. A representation optimized only for ordinary trajectory prediction may consequently discard information required for selecting the next action.

This issue arises naturally in model-based control, system identification, and experimental decision making. Our experiments cover complementary settings. Latent-memory processes test known memory structure, generalized Langevin dynamics provide a classical memory process, spin–boson systems allow controlled multi-intervention evaluation, and two independent superconducting datasets test the principle using recorded experiments. We have also added concrete application examples to the revised manuscript to improve readability. For example, in soft wearable robots, material hysteresis can make the same actuator command produce different forces depending on previous motion. This is another setting in which the observed present alone may be insufficient for choosing an action.

We further added a downstream intervention-selection experiment. On 40 held-out matched-history pairs, the model chooses among 18 interventions. In 12 matched-history/target cases, the two histories share the matched present but require different oracle-optimal actions. Response-derived OCPR recovers both actions in 12/12 cases, while history-inferred OCPR recovers both in 11/12 (App. H.1, Table 7).

These results illustrate the practical motivation of Operational Memory. Systems that appear equivalent at the present may still require different actions because of their histories.

AuthorsOpenReview ↗

Response to Reviewer 3nNm (2/3)

Q2. What is compressed from the 64-step history, and how much information is lost?

The 64-step history does not theoretically correspond to a 64-dimensional representation. The original wording made two independent choices appear related.

The 64-step quantity is the temporal input window. Separately, the temporal encoder has 64 output channels and therefore produces a 64-dimensional feature vector. An affine map then maps this feature to the d=16d=16 predictive state. The revised manuscript now explicitly distinguishes these quantities and gives the complete encoder and fitting procedure (Sec. 4.3; App. C.4).

The deeper question about information loss is important. In general, compressing a sufficiently rich history into any finite-dimensional state necessarily discards some information. This is a fundamental limitation of finite-dimensional recurrent, state-space, and predictive-state representations.

Our objective is therefore not lossless storage of the entire past. It is to preserve the information required for a specified family of future intervention responses. This is precisely the distinction studied in our paper: ordinary trajectory prediction can remain highly accurate even after intervention-relevant information has been lost.

Empirically, the 16-dimensional states inferred from observable histories achieve Causal Skill 0.966/0.971 on the held-out third intervention, leaving only 3.4%/2.9% of the squared error of the zero-difference reference, and recover 11/12 downstream action reversals. These results show that the chosen state retains most intervention-relevant information in the studied systems. We do not claim that 16 dimensions are universally sufficient; more complex or multi-qubit systems may require larger or structured predictive states.

As an additional check, our new encoder comparison uses a common train-only 16-dimensional response representation. Under this shared representation, TCN, S4D-style, and S5-style encoders all recover substantial intervention-response information, with mean Causal Skill ranging from approximately 0.849 to 0.915 across the two controlled environments. This provides additional evidence that the chosen low-dimensional response representation is not the dominant bottleneck in these experiments, while we do not claim that finite-dimensional compression is lossless in general (App. G.2, Table 5).

AuthorsOpenReview ↗

Response to Reviewer 3nNm (3/3)

Q3. How are the hardware responses obtained, and is full-state tomography used?

We do not perform full-state tomography in either hardware OPMS analysis. In our hardware analysis, we use recorded measurement results from public datasets and construct the response statistics needed in our paper from those results.

For Giarmatzi, the responses are expectation values obtained from recorded Pauli-outcome counts. For Gosling, they are future excitation indicators obtained from recorded quantum-jump traces. These recorded statistics are used for the OPMS analyses performed here (App. D.3).

We agree that unrestricted state tomography becomes prohibitively expensive as the number of qubits grows. The revised manuscript now states this limitation explicitly: local or few-qubit observables can probe the Operational Memory accessible to those measurements, while obtaining a sufficiently informative state for multi-qubit OCPR remains an open measurement problem.

Reviewer_11tnOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies a neglected problem in dynamical representation learning: good trajectory prediction does not imply that the model's state retains history information relevant to intervention responses. The authors define this information as operational memory, propose the OPMS spectral diagnostic based on conditional canonical correlation analysis, and design a context-resolved version to eliminate cross-context signal cancellation. They further propose OCPR: a Hankel-matrix-based predictive state realization that identifies the recurrent update operator using only responses of finite intervention depth, and then composes operators to predict deeper intervention sequences.

Critical strengths
  1. The paper draws an important conceptual distinction between trajectory performance and intervention-response recovery, which is a foundational insight for non-Markovian dynamical representation learning and has received insufficient attention in the existing sequence modeling literature.

  2. The analysis is relatively comprehensive and is supported by corresponding numerical experiments.

Critical weaknesses

The framework relies heavily on an idealized setting with matched contexts: the entire definition of operational memory depends on the case where context histories are matched. It is difficult to construct strictly matched contexts from noisy real-world time-series data.

Questions for authors

Matched contexts are central to this framework. When noisy real-world data cannot provide strictly matched context sample pairs, are there heuristic methods for approximate matching? Can the robustness properties of approximately matched samples be discussed?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 11tn (1/2)

We thank the reviewer for carefully reading our paper and for the constructive comments. We have carefully considered the concerns raised in the review and revised the manuscript accordingly.

W1. Clarity of presentation.

We have substantially revised the Introduction, Method, and experimental descriptions to improve clarity. The revision provides clearer definitions of the main concepts and notation, additional explanations of the key equations, and more direct interpretations of the experimental results.

AuthorsOpenReview ↗

Response to Reviewer 11tn (2/2)

Q1. How can matched contexts be approximated robustly with noisy continuous observations?

We agree that exact matching is difficult with noisy continuous observations. In our formulation, exact matching defines the comparison of interest. The observed present and other relevant conditions are held fixed so that response differences reflect information carried by history. Practical estimation does not require exact duplicates of observed trajectories.

For continuous observations, we now describe matching based on nearest neighbors and calipers, followed by context residualization. If the conditional response is LFL_F-Lipschitz in context and the residual mismatch is bounded by ϵ\epsilon, the resulting response difference has an error bounded by LFϵL_F\epsilon.

We further added a controlled sensitivity experiment on approximate matching, using 20 independent datasets with 2,000 history pairs each.

SettingResult
q=2q=2, residualized, mismatch 0.010.01–0.40σ0.40\sigmaApproximately 100% median spectral strength retention, with 20/20 detections at every tested level
q=0q=0, exact matching, residualized3/20 rejections
q=0q=0, history correlated mismatch 0.40σ0.40\sigma17/20 rejections without residualization, reduced to 3/20 with residualization

These results show that the memory signal remains stable across the tested mismatch range, while residualization substantially reduces spurious dependence. At the largest mismatch, the rejection count falls from 17/20 to 3/20, matching the exact matching reference. With 20 independent datasets, this does not establish exact calibration at the nominal 5% level, but it demonstrates a substantial reduction in spurious detections.

The experiment evaluates sensitivity to residual context mismatch under controlled pairing, providing evidence for the robustness of OPMS under approximate matching. The practical matching procedure, theoretical error bound, and new sensitivity analysis are now included in the revised manuscript (Sec. 5.2; Apps. C.1 and D.2).

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer again for the thoughtful comments and suggestions. We hope these revisions address the reviewer’s concerns and that the reviewer will reconsider their assessment in light of the revised manuscript.

Thanks for the author's response. Since I'm not really familiar with the domain of this work (quantum dynamics), I'd suggest focusing mostly on the other reviewers' points. As my confidence score indicates, I think the other two reviewers' feedback should carry more weight here.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.