acceptodds
Under review as a conference paper at ICLR 2027

Temporal Aliasing Is Arithmetic: Optimal Source-Time Design for Unseen-Time Control

Abstract

A controlled finite-state continuous-time Markov process may be observed only through endpoint transitions at legal source holding times, while deployment uses an unqueryable duration. Let be the closed additive group generated by the source times. Equality of source skeletons propagates to : if , the generator is identified, while identifies one base skeleton and its compositional times. Under an explicit rate–time condition, every noncompositional target admits irreducible three-state models with identical adaptive source transcripts but opposite target actions. We characterize the resulting Bellman and arithmetic answer quotients. On finite observable quotients with known legal-source laws, a sparse target-excluding Track-and-Stop rule matches the answer-information lower bound, while Diophantine distance to harmful resonances controls finite-sample conditioning. Finally, on a coupled unknown-damping/logarithm-branch family with known parametric form, a profile-design confidence-sequence learner operates on raw categorical endpoints and attains the pointwise answer-information constant even though exact branch information is zero.

Primary area:
reinforcement learning
Keywords:
theoretical reinforcement learning, continuous-time reinforcement learning, temporal-resolution generalization, controlled continuous-time Markov chains, temporal aliasing, policy identification, sample complexity, optimal experimental design, fixed-confidence sequential learning, Markov semigroups, model identifiability, minimax lower bounds

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_CQ8tOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper is mostly theoretical. The main insight lies in how to choose the correct action within a finite-state continuous-time Markov process at an unseen deployment duration when a continuous-time system can only be observed at other, allowed durations.

Critical strengths
  • The idea is somewhat promising.
  • Some insights have potential. The distinction between identifying dynamics and identifying decisions is useful. The constructive impossibility result is good.
Critical weaknesses

After reading the paper twice, I decided I would not go into analyzing it further. The authors did not even spend time to write it so that an average human reviewer can read it and would feel like reading it, so why should I? The authors of this paper either do not understand how to write a paper, or are trying to actively sabotage the conference with slops like the manuscript submitted here. At least try to write something that resembles a literature review. I do not even want to mention that the paper is also full of LLM-generated jargon. As I said, the idea is promising, and I strongly suggest a withdrawal and re-writing with actual human effort.

Questions for authors

I have no questions, as my assessment of this paper would not change.

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

The manuscript studies continuous-time Markov chains, in particular looking at their algebraic structure (what the authors call a "temporal information group" of the CTMC), and designing information-theoretically optimal learners that can operate when optimal actions are concealed; among other algorithms/guarantees, those resembling Garivier-Kaufmann best-arm identification are considered. Despite spending extensive time reading the manuscript, little else is clear about this paper and its contribution due to its AI slop nature.

Critical strengths

Given the manuscript's current state, it is unfair/not possible for me to list up its "critical strengths"; it is barely possible to figure out the setting in which the paper operates, let alone decipher what its claimed contributions should mean for the literature.

Critical weaknesses

In its current form, this paper in my view represents pure AI slop --- which immediately leads to my strong reject recommendation.

--- The quality of writing is abysmal. The title and abstract should give a sense of how the rest of the paper is written. Take the title: "temporal aliasing"? "arithmetic"? "unseen-time control"? --- Who are these terms addressed to, what is the intended audience that is expected to look at this title and know what is meant and be interested in reading further? The rest of the paper just continues throwing unjustified jargon at the reader while barely defining it (e.g."legal", "compositional", "fiberwise", ...)

--- The paper gets straight into action, introducing notation as early as the abstract, and failing to motivate and situate the problem early on. All the reader gets in the main part is a brief 2-paragraph section 9 at the end of the main part, which, while it cites several related papers, fails to explain what critical gaps the literature has that this submission must fill. To defend its contribution, the authors claim: "We do not claim multirate reconstruction or nonunique logarithms as new. Our contribution is their unseen-time control synthesis: adaptive source equivalence, a rate-qualified action converse, Bellman images of the source fiber, and answer-optimal legal design." (By the way --- this sentence, in my view, represents the paper's writing quality fairly well.)

--- For just another example of the quality of writing, consider the empirical evaluation section --- short and completely useless to the reader; it's entitled "Theorem-driven experiments", and instead of introducing the evaluation setup, it proceeds to throw various numbers at the reader ---- for instance stating right away that the experiments use "master seed 20270925".

Questions for authors

At this point, the paper is a clear reject in my opinion, due to its barely readable nature. The authors should probably start by explaining what it is that this paper would contribute to the ICLR community, and should feel free to address any of the above points --- including their honest assessment as to whether they consider the current state of the paper appropriate for submission. I would not be opposed to reading a human-written (!) and well-motivated account explaining (1) what the paper is trying to solve, (2) how these results are situated within the CTMC literature; and (3) sketches of the main claims' proof techniques.

Reviewer_God6OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes new pure exploration algorithm with theory around it in a specific setting.

In the setting, the learner observes a continuous Markov system and the durations that are emitted, but has to choose an action and another duration. I like the setting because it's quite new and practical. On the paper, they characterize when this observation determines the right decision and when you cannot distinguish dynamics, and you would require opposite actions. Finally, I'll create and propose a way to resolve this ambiguity that is relevant to decision-making, and then also provide asymptotically optimal sample complexity under some assumptions of restricting the model.

Critical strengths

I really like the principle that you learn what changes the optimal action, rather than reconstructing the whole generator. I think this is a good insight that the authors leverage.

They have this tree-state construction that turns ambiguity in the semantics into really genuine impositive results for control. The result on condition also explains why identifiability: you can have it at the same time with extremely poor statistical efficiency.

Also, the cleanest numerical result is 16 times more decision-relevant information from the clocks that are answer-oriented than from the branch identification, though it's not a 16x runtime improvement.

Critical weaknesses

The ranking mainly specializes in established track-and-stop ideas - there is definitely room for improvement there, and the end-of-dynamics results covered only one non-parametric family.

What is also important is that the target duration is 0.5, which is exactly twice the legal source duration of 0.25. I would say that the natural missing comparison is therefore estimating the source transition matrix and squaring it (rather than this being what they do) rather than discovering the generator branch.

Finally, the normalized stopping cost. Remind really well about the asymptotics. Practical finite-budget efficiency is not established at all. Maybe the authors could comment on whether this is a lack of theory or the actual algorithm that could be improved.

Questions for authors

If you have the stopping rule fixed, what does the adaptive allocation gain over uniform sampling become, and what will be the best fixed legal clock?

If you have normalized stopping time, separate it into allocation. Can you take the normalized stopping time overhead and separate it into allocation, confidence threshold, and batching cost? Since, once again, the target duration is twice the shortest legal duration, how does this kernel composition that is aware of the structure compare at the match confidence?

Finally, what should be e numerical accuracy must be achieved by profile and desgn optimizer to preserve correctness as asymptotic optimality?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.