acceptodds
Under review as a conference paper at ICLR 2027

Causal Echoes: Task Precision Through Physical Cancellation

Abstract

Control actions can improve the measurements used to learn a system. For unknown linear systems with signal-dependent sensing, paired inputs nearly cancel in the plant and leave a small residual with small absolute readout noise. Repeating this physical cancellation sharpens the information needed for a future control task. Under radial relative noise and zero additive floors, actionable packets yield minimax orders for the learning-terminal state and for a future command using a final upload; full physical calibration remains at order in dimension at least two. Matching lower bounds allow adaptive packet timing and silence. A calibrated-drift gain family gives a finite separation from every noninteractive acquisition, including randomized designs, under nonzero readout and deployment floors. Held-out matrix results are consistent with the predicted orders and show how conditioning and noise floors limit the useful refinement depth.

Primary area:
learning theory
Keywords:
adaptive control, system identification, linear dynamical systems, signal-dependent sensing, task-oriented exploration, minimax theory, active experiment design, physical cancellation

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_pwB7OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies task-oriented learning and control of unknown linear dynamical systems with signal-dependent observation noise and limited communication. The proposed physical cancellation approach uses control inputs to produce small residual states with reduced absolute measurement noise, enabling iterative refinement of task-relevant estimates. Theoretical results cover controllable and underactuated systems, adaptive communication, and noise floors. Validation includes simulations on held-out linear systems supporting the predicted error rates, an analytical example demonstrating an advantage over nonadaptive acquisition, and numerical studies examining dimensional scaling, conditioning, and sensitivity to measurement and process noise.

Critical strengths

The idea of using control actions to reduce subsequent measurement noise through physical cancellation is potentially interesting.

Critical weaknesses
  1. Restrictive assumptions and questionable practical significance: The main theoretical improvements rely heavily on a specialized observation model in which measurement noise decreases as the system state approaches zero. The strongest gains require vanishing noise floors and are established for relatively narrow control objectives. The paper provides insufficient justification for the practical relevance of these assumptions. Moreover, the numerical results show diminishing benefits with increasing dimensionality and, in some cases, substantially worse performance with additional refinement rounds when noise floors are present. This raises concerns about whether the theoretical improvements translate into meaningful advantages in realistic settings.

  2. Severely inadequate problem formulation, exposition, and literature positioning: The paper is exceptionally poorly motivated and difficult to approach for readers outside a narrow control-theory audience. Neither the abstract nor the introduction clearly establishes the underlying control and system-identification problem, explains the practical setting, or provides sufficient intuition for why the proposed problem is important. Instead, the authors immediately introduce specialized terminology and concepts while assuming substantial prior familiarity with the field. The introduction contains only a single citation and makes virtually no attempt to situate the work within the extensive literature on adaptive control, active sensing, dual control, and experimental design. Although related work is briefly discussed near the end, this does not compensate for the lack of context and differentiation when introducing the contribution. As a result, it is unnecessarily difficult to understand the motivation, identify what is genuinely novel, or assess the broader significance of the results. These are fundamental shortcomings in the paper's scientific presentation and positioning, not merely stylistic concerns, particularly for a broad machine learning venue such as ICLR.

Questions for authors
  1. Novelty and related work: How does the proposed physical cancellation mechanism fundamentally differ from existing approaches in adaptive experimental design, active sensing, and dual control? Can the authors identify the precise conceptual and technical distinctions from these established research directions and provide a more comprehensive discussion of the relevant literature?

  2. Motivation and sensing assumptions: What concrete applications motivate the particular signal-dependent observation model considered in the paper? Can the authors provide realistic examples of sensing systems in which measurement noise decreases sufficiently near zero for the proposed refinement mechanism to provide meaningful benefits?

  3. Practical relevance and robustness: Given that the numerical results show diminishing improvements with dimensionality and potentially severe degradation under nonzero noise floors, under what realistic operating conditions would the proposed method be preferable to standard estimation and control approaches? Can the authors characterize these conditions more explicitly?

  4. Comparison with existing approaches: Could the authors provide more direct empirical comparisons against established adaptive system-identification and control methods under matched measurement, communication, and energy budgets? The current numerical evidence makes it difficult to assess the practical value of the proposed method relative to existing techniques.

  5. Generality of the task formulation: To what extent can the precision improvements extend beyond minimum-energy zeroing from a fixed initial state? Given the limitations already identified for different initial preparations and other control objectives, what broader classes of practically relevant tasks could benefit from this approach?

2 comments

Your review separates two problems, and we agree with both. The second went beyond style: the submission used its problem before stating it and placed it in the literature only at the end. The revision addresses both. For practical significance we added three kinds of evidence:

  • a matched-budget comparison with eight baselines (§5.1);
  • studies of floors, actuation error and dimension (§5.2);
  • a pH-neutralization rig where relative precision is a property of the instrument (§5.3). There one commissioning batch yields a recipe 75×75\times more precise than an optimized nonadaptive design.

§6 (p. 9) states when the method is the right tool and when it is not.

W2/Q1. The abstract and introduction now set up the problem first, and a scalar example shows the mechanism (p. 2). The introduction now cites 21 works instead of one. The contrast with adaptive control, active sensing and dual control is in §4.3, §7 and App. O (p. 41). The main text is full, so a short version will move into the introduction for the final version.

Fisher-information designs favor large excitation, which does not help when the error grows with the signal. A correctly specified heteroscedastic design (NASD) instead rediscovers cancellation in 91% of its rounds. Active perception moves toward precise sensing but assumes known dynamics. In dual control's trade-off, each probe drives the state far from zero, but the next input brings it back, so one policy attains the optimal orders for both regulating and learning (Theorem 1), at a constant cumulative cost. What is new is the learning-theoretic account. Lower bounds match the exponents for every KK-round policy, including data-dependent round times. For n≥2n\ge2 and norm-proportional errors, no policy identifies the physical matrices beyond order σ\sigma, and Proposition 2 says which tasks inherit σK+1\sigma^{K+1}. The algorithm needs no noise model. For ICLR, this is active learning under an adaptivity budget, as in batched bandits, with minimax lower bounds.

Q2. Table 25 (p. 41) places the sensor classes of §1 on our operating map. Predicted gains are large for bfloat16 and binary16 telemetry, logarithmic quantizers and gravimetric pH dosing, borderline for FP8, and under 3×3\times where actuation is not much more precise than sensing (piston-burette dosing 2.3×2.3\times, relative navigation 0.8×0.8\times). These predictions assume no floor. The pH rig, which has a measured floor, gains 75×75\times.

W1/Q3. On the rig our recipe misses neutrality by 0.079 µM, 2.0×2.0\times / 4.1×4.1\times less than NASD with the measured / estimated noise law (Table 3, p. 9). The margin over the best alternative is modest, and others win where they apply. Feedback dosing with the electrode in the batch reaches 0.018 µM, a trim helps if the batch can wait 5 min, and batch-to-batch ILC matches our single batch after 3. In simulation, ILC with K+1K+1 preparations wins 53 of 54 plant classes. The method is preferable when:

  • deployment has no feedback and no time for a trim;
  • actuation and preparation are much more precise than sensing;
  • the floor is small and the dimension low;
  • the plant's free response outlasts the learning run;
  • re-preparing x0x_0 is costly.

Floors and dimension. With a floor the rate saturates at σd+τ\sigma d+\tau, and the gain falls roughly like 1/d1/d. Extra rounds hurt for one reason: when the command is reconstructed through inverse powers of a contracting plant, these amplify the floor-limited error. The floor-aware rule (§4.2) fixes this. On contracting plants at d=10−3d=10^{-3} the miss stays at 8.2×10−48.2\times10^{-4} instead of growing to 3.0×10−13.0\times10^{-1} by K=5K=5 (Figure 3(a), p. 8). The gain shrinks with dimension. More rounds recover much of it on near-orthogonal plants (3.3×103×3.3\times10^{3}\times at n=8n=8, 62×62\times at n=16n=16, with K=7K=7) but not on generic plants beyond n=8n=8, so we recommend the method for low dimension.

Q4. Table 2 and Figure 2 (p. 7) compare eight baselines with the same horizon, energy cap, readings and number KK of adaptive rounds, which replaces the communication budget (App. L, p. 28). Every non-cancelling baseline stays at order σ\sigma, the best missing 1.5×104×1.5\times10^{4}\times more than ours at σ=1.3×10−3\sigma=1.3\times10^{-3}. On a contracting panel ILC with resets is best.

Q5. By Proposition 2 (p. 4), Lipschitz functions of the task coordinates (B−1AB,B−1x0)(B^{-1}AB,B^{-1}x_0) inherit σK+1\sigma^{K+1}. These include zeroing at every horizon, scaled targets c x0c\,x_0, input-energy costs and maneuvers specified through the inputs and x0x_0. The physical matrices and maneuvers from other starts stay at order σ\sigma (App. I, p. 26).

Theorem 1 now states its quantifiers (p. 4), and §6 says what the bounds do not give. We would be grateful if you reconsidered your assessment in light of the new evidence and the scope stated in §6.

Reviewer_pwB7OpenReview ↗

Response to rebuttal

I appreciate the authors' efforts to address my concerns. The revised manuscript provides substantially better motivation, literature positioning, and exposition. I am therefore raising my score from 1 to 2.

However, I find the extent of the revision somewhat difficult to evaluate, as the manuscript has undergone substantial changes beyond clarifying the original submission, including significant new experimental evidence and a reframing of the contribution. The revised paper feels considerably different from the original submission, making it challenging to assess the new claims within the scope of my initial review. I would defer to the AC and reviewers with stronger domain expertise regarding the technical significance of these additions.

Reviewer_9rLQOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies a linear control system with a special measurement structure. Despite my familiarity with MDPs, reinforcement learning, and linear control, I found the paper difficult to understand even after careful reading. The lack of clarity in the exposition prevents me from fully assessing its technical contributions. Substantial rewriting is needed before the manuscript is suitable for publication.

Critical strengths

The paper has a sufficiently long appendix.

Critical weaknesses
  1. The central research problem is not clearly formulated. Although the authors present three objectives in equation (10), the problem setting is not clearly stated (see questions below), making the definition of these objective mathematically vague. Despite the manuscript’s 76-page length, the authors devote insufficient space to explaining the meaning and motivation of its three central objectives.

  2. Theorem 1 appears to bound the three objective values. However, the dependence of these constants on the system class and resource constraints is insufficiently explained, making the bounds’ finite-noise implications difficult to assess.

Questions for authors
  1. Section 3 says that "It (the sensor) may transmit or remain silent based on ..." Do you also design or learn a transmitting policy for the sensor? In this case, the problem is a multi-agent decision-making problem. The authors should clearly state the problem setting. A policy is typically a function that maps available information to actions, please define the policy explicitly.

  2. The meaning of equation (9) is unclear. Can you explain the notations.

  3. For the physical-model error in equation (10), do you estimate the parameter at the end of the horizon? That is, do you estimate θ\theta using the data collected from TT time steps?

  4. The meaning of equation (12) is unclear. What does the notation mean?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 9rLQ

Thank you for pointing to the exact places where the paper lost you. Your diagnosis is right: the submission used its problem before stating it, and all four questions trace back to that. The revision states the problem before the first theorem. The sensor is in Eq. (1) (p. 1), the plant, the policy class and the three objectives are in §2 (pp. 2–3), and the notation is collected in App. A (p. 16). The paper is now 45 pages instead of 76.

In short, some sensors err in proportion to what they measure (relative precision σ\sigma), so small states are read precisely. A learner allowed KK adaptations pushes the state away with a known input, applies the input its model predicts will cancel it, and reads the small residual, which reveals the model error. It can then keep its state at σK\sigma^K and learn a repeated maneuver's command to σK+1\sigma^{K+1}, and lower bounds match these exponents for every KK-round policy. For n≥2n\ge2 and norm-proportional errors, no policy identifies the physical matrices beyond order σ\sigma. Simulations match the exponents (slopes 2.00/3.00/3.99), and on a pH-neutralization rig one learning batch gives a recipe 75×75\times more precise than an optimized nonadaptive design. A short first reading is the scalar example (p. 2), §2, and Theorem 1 with the two paragraphs after it (p. 4).

Q1. There is no second agent. Definition 1 (p. 3) defines the policy by round-start times t1≤⋯≤tKt_1\le\dots\le t_K that are stopping times of the readings, inputs between them that depend on the readings up to the current round start, and a hard energy budget. KK counts plan changes, as in batched bandits and low-switching RL. The old sensor that may "transmit or remain silent" was one way to implement such a policy. App. L (p. 28) shows that a sensor allowed KK messages is exactly a KK-round policy designed jointly with the actuator, so there is no game.

Q2 (old Eq. 9). This is now Eq. (4) (p. 3), the minimum-energy inputs that drive a fresh run from x0x_0 to zero in HH steps,

UH⋆(θ)=−CH(θ)†AHx0,U^\star_H(\theta)=-\mathcal C_H(\theta)^\dagger A^Hx_0,

with CH=[AH−1B,…,B]\mathcal C_H=[A^{H-1}B,\dots,B] the controllability matrix and †\dagger the pseudo-inverse. A learned U^\hat U misses by ∥CH(θ)(U^−UH⋆)∥\|\mathcal C_H(\theta)(\hat U-U^\star_H)\| when CH\mathcal C_H has full row rank, and this miss is what we report.

Q3. Yes. U^\hat U and θ^\hat\theta are computed after the run and may use all readings and inputs (Definition 1 and Table 1, p. 3). For FF the infimum is over policies with feedback after every reading, which only strengthens its lower bound.

Q4 (old Eq. 12).

  • a≍ba\asymp b means cb≤a≤Cbcb\le a\le Cb with constants independent of the noise levels (σ,d,τ,β)(\sigma,d,\tau,\beta). They may depend on Θ\Theta, x0x_0, KK, TT, HH, EE and pp, as §2 and App. A state, and for the algorithm also on its parameters a,r0a,r_0.
  • RKR_K is the rate of Eq. (5) (p. 5), which is σK\sigma^K in the clean case.
  • "Both readers" meant the norm-proportional and the coordinatewise sensing models. The main text uses the first, and App. J (p. 27) the second.

W1. Table 1 (p. 3) puts the objectives side by side. DKD_K is how small the learner leaves its own state, QKQ_K how precisely it learns a command repeated open loop from x0x_0, and FF ordinary calibration. QKQ_K matters when deployment has no feedback, as on the pH rig. The rates differ, σK\sigma^K, σK+1\sigma^{K+1} and σ\sigma, and the command gains a power because the last round's readings still improve the model after the run (p. 4).

W2. Theorem 1 (p. 4) now states explicit conditions: the clean case, σ≤1/10\sigma\le1/10, a horizon T≥L+K(2n+1)+1T\ge L+K(2n+1)+1 and an energy budget E≥E0E\ge E_0 with E0E_0 explicit (App. C, p. 17). Page 5 extends it to floors and process noise, where the submission said only "sufficiently small noise levels". The paragraph "What the theory does not give" (p. 10) explains how the constants scale. The lower ones decay like (c/(nT2Gˉ))K+1(c/(nT^2\bar G))^{K+1}, so they fix exponents only. The explicit one-round bound (Proposition 9, p. 19) applies on our main panel only for σ≤1.0×10−4\sigma\le1.0\times10^{-4}, and as a diagnostic. At our experimental noise levels the per-round factor, about 25σ25\sigma where we measured it, is therefore an empirical finding. The learner can estimate it online to within 14% (Table 10, p. 34).

If any definition in §2 can still be read two ways, please tell us which, and we will fix it. We would also be grateful if you could look at the revision again with the formulation in hand.

Reviewer_9rLQOpenReview ↗

Response to Rebuttal

I thank the authors for the substantial revision, which has improved the paper. However, I still find the presentation confusing, and several aspects of the problem formulation remain unclear to me.

  1. Relationship among the three objectives. Table 1 presents DKD_K, QKQ_K and FF as three objectives for one learning run, but how they relate is not clear. QKQ_K appears to be the main objective throughout the paper. DKD_K seems to serve mainly as a means to achieve QKQ_K (a small state at the anchor is what allows the σK+1\sigma^{K+1} rate), and the paper notes that myopic certainty equivalence already attains DK≍σKD_K \asymp \sigma^K. The role of FF is less clear. FF is the standard system identification objective, which arguably has the most direct practical meaning, yet it appears mainly as a lower bound contrasting with QKQ_K. Moreover, FF is defined over a different policy class (ΠT−1\Pi_{T-1}) than DKD_K and QKQ_K (ΠK\Pi_K). I would ask the authors to clarify whether these are meant as joint objectives or as one target objective plus supporting and contrasting quantities, and to justify each accordingly.

  2. Noise in the system model and the open-loop setting. The plant in Eq. (3) is linear with Gaussian process noise, but the target command in Eq. (4) is defined from the noiseless dynamics, and deployment is "idealized as a noiseless run." Process noise during deployment also does not seem to reappear in Proposition 3, which adds only execution and preparation errors. Since the authors refer to open-loop deployment, could they clarify whether the paper is about learning an open-loop control policy? If so, please discuss how process noise during deployment affects the achievable miss, since an open-loop command cannot compensate for it.

  3. Motivation for closed-loop learning with open-loop deployment. The learning policy uses feedback (KK adaptive rounds), while the deployed command runs open loop. The setting is therefore closed-loop learning for open-loop control. This setup needs stronger motivation.

  4. Presentation of Theorem 1. Theorem 1(i) in Section 3 refers to Algorithm 1, but Algorithm 1 is only introduced in Section 4.1, after the theorem. Please introduce the algorithm before the result that analyzes it, or at least describe it briefly before stating the theorem.

Overall, I believe the presentation still needs improvement. In particular, the learning objectives should be more carefully discussed and justified, and the problem setup should be better motivated. I therefore maintain my current rating.

Reviewer_S4QWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Additive noise is what is usually studied in linear control systems, but sometimes things have a multiplicative aspect. This paper studies situations in which there is state-size dependent additive observation noise. (Such models have also been studied in the literature going back decades, but because I've decided to write this review without any AI assistance, I will leave it to the authors to consult the AIs to get the deep dive into the literature. It is certainly cited by some of the papers that you cite.)

The key insight of the paper is that the additive observation noise is smallest around zero and so using pairs of control actions that try to move back towards the zero state has the advantage of getting the system to be in a place where the noise is the smallest.

Although the paper is written with Gaussian observation noise, it can be conceptually easier to think about this as learning from low-precision floating point observations where the exponent term is free but we have only very few mantissa bits. (The paper does not say this but the AC deserves to know what this paper is conceptually about.) The analogy is not exact, but it conveys the core issue. To learn fast in such a setting, we would want to get the most useful information out of the mantissa bits. And that means making them depend on the things that one does not know in a way that lets you solve out for the things that you don't know. As one learns dynamics better, one can think of this as learning the most-significant bits of the dynamics. When those most-significant bits are known with confidence, we would prefer if our observations did not depend solely on them --- because that would give us no information. The trick therefore is to get as much of that dependence to cancel away as possible.

Of course, the devil is in the details and that is what the paper works to do. It claims that one gets pretty much the most one could hope for: we can learn the system matrices to the precision of the mantissa bits. And given more time, we can basically wipe off the mantissa count worth of bits from the state at each step. They also say that there is essentially no conflict between making the state small and learning the dynamics.

Critical strengths

As stated, it is a very niche problem of interest to only a few people (I happen to be among those few). But because of the implicit spiritual connection to finite precision arithmetic, the potential interest is broader.

If the writing were much much much better, then there is an interesting control-theoretic story to tell here.

Moving beyond linear systems, the general setting would have some place in state space where there is less observation noise. The idea would then be how to take advantage of this "I can see clearly here" spot while controlling the unknown system while trying to learn it. The naive approach would be to always try to stay in the "see clearly here" spot. The paper's insight is that no --- it can be useful to leave the clear sight place temporarily and then try to come back into it. This will let you better excite the dynamics to extract information about the dynamics while still landing you often in the "see clearly" zone.

Critical weaknesses

The writing is extremely unclear. And this starts in the title, and by the time one gets to the abstract, one is already confused.

The introduction does a very bad job of setting up the problem and the model does not introduce complexity in layers to build out the insights clearly. There is a substantial amount of extraneous stuff that is not as fundamental as the core issue: like the silences and packets sent.

It feels like the AI aid in writing turned what could have been a short and sweet clear story (possibly with some involved proofs to improve constants) into an unreadable mess.

This is a paper about relative noise. One has a degree of confidence about the direction of vectors and only relative confidence about magnitude. In additive noise, the degree of confidence in direction goes to zero if the vector is small. Not here --- one always has some confidence in the direction.

In this case, I find myself feeling like the poor quality of the writing only enables me to see the direction but not the details. And the authors should iterate on the paper so that it is crystal clear.

Questions for authors

For baselines, what does the most naive approach do in this setting? (i.e. drive the state with pure noise inputs and try to learn) How does it perform? Show the plots.

Next, what if I just myopically tried to control the system towards zero and did nothing else. How does that approach do? Flesh it out completely and show me the plots.

Next, show how your approach does better.

And now, take an iron-man approach of the best approaches in the literature to this problem. How do they do? What are they missing?

I think that the poor quality of the writing makes this paper a lost cause for this review cycle.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.