acceptodds
Under review as a conference paper at ICLR 2027

Decision-Sensitive Compression with Learned Actuation

Abstract

A compressed predictor can support accurate decisions without predicting every output accurately. This is harder to certify when the action response is also learned: a full-matrix error bound includes irrelevant output noise. We develop a finite-sample alternative for quadratic decisions. A Gaussian expansion at the fixed true optimal controller, together with calibrated low-dimensional moments, bounds the first-order decision error without a leading factor proportional to the square root of the output dimension; noise products and the plug-in Gram bias remain explicit. The resulting observable policy ball certifies arbitrary fitted linear controllers, and its optimal rank-constrained controller is ordinary plug-in compression. A local-rotation lower bound isolates the gap-dependent difficulty of selecting a representation, while a task-sampling bound separates physical calibration from generalization to fresh objectives. Experiments test irrelevant outputs, observable gap resolution, sampled tasks, and neural nuisance–capacity tradeoffs with learned actuation. Calibration improves markedly in the many-output regime, but remains conservative and is not uniformly tighter at small sample sizes. The analysis does not certify nonlinear pilots or closed-loop latent world models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
decision-sensitive compression, representation learning, learned actuation, finite-sample guarantees, quadratic control

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_sM6GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how to certify the linear-quadratic decision error induced by a low-rank representation when both the prediction matrix and the actuation matrix are estimated from data. Its central contribution is an expansion of the first-order Gaussian error around the fixed true optimal controller, together with a policy ball constructed from observable low-dimensional moments that is simultaneously valid for any fitted linear controller. It also provides a gap lower bound for spectral direction selection and an analysis of sampling error for new tasks. The novelty should be attributed primarily to finite-sample certification, rather than to weighted PCA or low-rank compression itself.

Critical strengths
  1. The central certification result has clear technical substance. Theorem 2 on p. 3 and Sections A.5–A.6 on pp. 16–17 separate the first-order Gaussian term, products of noise terms, and Gram deviations, avoiding the direct use of full-matrix error from irrelevant outputs as a measure of decision difficulty. Expanding around the fixed true controller also avoids treating data-dependent coefficients as fixed quantities in the Gaussian analysis.

  2. The claims align with the experimental scope, and limitations are adequately disclosed. Table 1 on p. 5 compares certificates for the same fitted models: as p increases from 4 to 1024, the stationarity ratio rises from 73.9 to 1229.0, whereas the policy-ball ratio rises from 13.0 to 21.7. The paper explicitly retains the pb² bias term and does not claim complete dimension independence. It also reports on p. 7 that the new certificate performs worse at small sample sizes and that compression quality deteriorates when too few task samples are available.

Critical weaknesses
  1. The scope of applicability and algorithmic novelty are limited. Equation (6) on p. 3 and the limitations paragraph on p. 9 require a correctly specified linear model, known feature second moments, independent isotropic Gaussian OLS errors, and a known noise scale. Correlations in practical trajectories, unknown noise, and closed-loop stability are not certified; the neural-network experiment also does not satisfy the assumptions of the main theorem. The optimum in Corollary 3 remains ordinary Plug-in, and the task representation on p. 4 is equivalent to weighted policy-bank PCA. The authors explicitly acknowledge these facts. The contribution should therefore be assessed as improved statistical certification, rather than credited with broad improvements to general representation-learning algorithms.

  2. The practical utility of the certificates is not yet sufficiently established. The certificates on p. 5 remain 13–21.7 times the true cost. Under one condition on p. 6, the excess bound is approximately 0.02, while the actual excess is approximately 10⁻⁵. The task diagnostic bound on p. 7 is approximately 1.23, compared with a measured excess of 3.20×10⁻⁴. These results clearly establish improvements over the previous bounds, but do not sufficiently demonstrate which ranks the certificates can certify at practical tolerances. All experiments are synthetic, and a direct assessment relating certification success rates, tolerances, and data requirements is missing.

Questions for authors
  1. For representative systems and a tolerance τ, what are the smallest rank accepted by the certificate, the failure rate, and the actual error? Do these results yield useful choices of sample size and compression budget?
  2. Do the benefits of the policy ball persist relative to split-and-certify or tighter decision-error certification methods under the same data budget?
3 comments
AuthorsOpenReview ↗

Response to Reviewer sM6G (1/3): scope, novelty and conservatism

Thank you for a careful and accurate review. Parts 2 and 3 answer Q1 and Q2, and the end of Part 2 explains the 0.02 and 1.23 figures.

W1: scope and novelty. We agree. The revised abstract and contributions present the work as finite-sample certification of decision error rather than a new compressor. Our certificate-minimizing rank-rr controller is plug-in compression (Corollary 3), and the sampled-task encoder is principal component analysis of a weighted policy bank (Sec. 4.2). Sec. 1 now credits Wagenmaker et al. (2021), whose task-Hessian-weighted risk is the antecedent of our first-order term. Their bounds hold after a burn-in and depend on the unknown system, whereas ours is computed from the data at finite nn.

The linear model with a known feature second moment Exx⊤=I\mathbb E xx^\top=I gives the exact cost formula, Eq. (3). The Gaussian error model, Eq. (6), makes the first-order term exactly Gaussian and gives its second moment (8), its covariance bound (19) and the remainder bound (21), while the known noise scales set every radius. Dependent trajectories, unknown noise scales and feedback control are outside the guarantees (Sec. 9). In new stress tests (Appendix E.9), heavy-tailed noise and a random design caused no violations, but ignored process noise did. The neural experiment (Sec. 8) carries no certificate, and in one cell with few action probes and small nuisance amplitude the weighting does not help (Table 4, now in Appendix C).

W2: practical utility. We agree that the certificate remains conservative, and the revision makes it tighter, explains the remaining gap and gives a floor for any certificate. All experiments remain synthetic because checking a certificate requires the true cost. The new multi-step benchmark is simulated from a physical model (Sec. 5).

  • In Table 1 the rank-2 Plug-in controller equals the uncompressed fitted controller, so the certificate equals our bound on that controller's excess cost. At excitation c=0.3c=0.3 (the ratio of action-probe to input-probe energy), its median is 5.5×10−45.5\times10^{-4} at p=4p=4 and 1.45×10−31.45\times10^{-3} at p=1024p=1024, against median true costs of 4.2×10−54.2\times10^{-5} and 6.7×10−56.7\times10^{-5} (Sec. 5). None of the 160 seeds had a violation.
  • Most of the certificate-to-cost ratio comes from the tail factor on the Gaussian term (geometric mean 6.7–9.9, Appendix E.5). The data-computable caps on unknown traces and norms add a factor of 1.1–1.6, the second-order remainder 1.0–1.9, and using the fitted Hessian at most 1.05.
  • A sharper certificate, PB+ (Prop. 7, Appendix A.15), bounds the Gaussian term with a Laurent–Massart χ2\chi^2 tail and allots that term half of δ\delta in advance. On the same fits it lowers the ratios from 13.0–21.7 to 8.8–16.5 at c=0.3c=0.3 (new column of Table 1) and from 8.1–13.0 to 5.5–9.5 at c=1c=1 (Appendix E.6, Table 14), with no violation.
  • Prop. 8 (Appendix A.16) shows that, when the action map is known, no valid certificate can keep its median below the (1−2δ)(1-2\delta)-quantile of the true cost for every forecast map DD. In the Table 1 geometry the best possible worst-case median lies between 11.77s211.77s^2 and 13.15s213.15s^2, with s2s^2 the per-entry error variance of the fitted forecast map, while the median true cost is 7.67s27.67s^2. Our known-actuation certificate (Eq. 27) gives 20.79s220.79s^2, 1.77 times the lower end of that range and 1.58 times the upper end. With a learned action map we have no such floor.
AuthorsOpenReview ↗

Response to Reviewer sM6G (2/3): certified rank, failure rate and sample size (Q1)

Q1. Conservatism means a higher rank or abstention (certifying no rank). An unsafe rank occurs with probability at most δ\delta.

Because the certificate holds for every linear controller on one event of probability at least 1−δ1-\delta, the rank can be chosen from the same data. On that event the certified rank is never below the oracle rank (the smallest rank that meets τ\tau with the true model), and the deployed controller's true cost is at most τ\tau (Sec. 3, Appendix A.14).

The new experiment (Sec. 5, Appendix E.2) pads the systems of Tables 2 and 3 with pure-noise outputs up to p=1024p=1024. It uses nn from 256 to 65536 and tolerances τ\tau of 10−110^{-1}, 10−210^{-2} and 10−310^{-3} times trL\mathrm{tr}L, the excess cost of not acting, for 10,240 fits in all. None of the three certificates (ours, the stationarity bound and the simpler moment bound of Eq. (18)) was violated, and PB+ was violated once, consistent with δ\delta (Appendix E.6). The table gives medians for our certificate at p=1024p=1024 (Table 8), where a dash means that no fit certified any rank. "Diagonal" is the diagonal system of Table 2, and "orientation" means the 16 randomly oriented systems of Table 3.

Systemnnτ/trL\tau/\mathrm{tr}LOracle rankCertified rankFits certifying nothingTrue cost over τ\tauCertificate over true cost
diagonal10240.1240%0.0518.9
diagonal10240.01, 0.0014, 5–100%––
diagonal655360.1220%0.6141.13
diagonal655360.01440%0.141.95
diagonal655360.001550%0.134.83
orientation1024all2–3–100%––
orientation655360.1220%0.211.41
orientation655360.01330%0.002522.6
orientation655360.001330%0.02522.6
  • At n=65536n=65536 the certified rank equals the oracle rank in 6,095 of 6,144 pairs of fit and tolerance and is one higher in the rest. The deployed controller's median true cost is at most 0.615τ0.615\tau.
  • Abstention is the main weakness. Our certificate certifies nothing in 70 of the 180 combinations of system family, pp, nn and τ\tau, all with n≤1024n\le1024 or τ≤0.01trL\tau\le0.01\mathrm{tr}L. The stationarity bound abstains in 114 of them, Eq. (18) in 93 and PB+ in 65.
  • With many outputs and few samples, simpler bounds can be far lower. For the orientation systems at p=1024p=1024, n=256n=256 and rank 2, ours is 284, the stationarity bound 31 and Eq. (18) 5.2, against a true cost of 0.31 (Table 10). Even without padding, Table 3 gives ours 2.80 against 1.97 for the stationarity bound at n=256n=256.

Sample size (Appendix E.3). In the dimension study of Table 1, our certificate certified τ\tau on 90% of fits once nn was 4–16 times the sample size at which the median true cost first reaches τ\tau. In five settings the median true cost was already at most τ\tau at the smallest nn tested, so there the factor is only a lower bound. The stationarity bound needed 4–8 times more data than ours at p=4p=4, and at p=1024p=1024 it needed 64–128 times more or did not certify within 2242^{24} samples. PB+ needs half as many samples in 5 of the 18 settings and never more. Recomputing the radius from a preliminary fit at n=256n=256 predicted the required nn on every fit of the dimension study on the factor-two grid, as a planning heuristic rather than a guarantee.

To your last question, the results give usable choices at large nn but often none with n≤1024n\le1024 or tight tolerances.

The two other numbers you quote are diagnostics of other bounds. The excess bound of about 0.02 in Sec. 6 (mean 0.025, against an actual excess of about 10−510^{-5}) is a moment-based eigengap diagnostic. Substituting the true gap lowers it only to 0.018, so most of the looseness comes from the calibration terms rather than the unknown gap. Certifying a rank at a tolerance needs no eigengap. The 1.23 in Sec. 7 is a diagnostic in which task sampling makes up 99.7% of the radius term. The measured excess it bounds is 3.22×10−43.22\times10^{-4}, not 3.20×10−43.20\times10^{-4} as submitted.

AuthorsOpenReview ↗

Response to Reviewer sM6G (3/3): split-and-certify and other methods (Q2)

Q2. Certifying on all the data beat splitting in every setting we ran, although not always with our policy ball.

Theorem 2 holds uniformly over linear controllers, so it needs no held-out data. Split-and-certify (SC) instead fits on half the data and certifies the frozen controller on the other half (Appendix A.7). The table gives median certificates at rank 2 with equal halves, with the median ratio to the true cost in parentheses (Appendix E.4, Table 12). "Dimension" is the Table 1 system at n=4096n=4096, and "orientation" refers to the unpadded systems of Table 3. These are fresh fits, so the values differ slightly from Table 1 and from the means in Table 3.

MethodDimension, p=4p=4Dimension, p=1024p=1024Orientation, n=256n=256Orientation, n=1024n=1024
Ours, all data5.5e-4 (13.4)1.45e-3 (20.5)2.72 (30.6)0.507 (6.5)
Stationarity, all data3.1e-3 (76)0.082 (1163)1.90 (21.4)0.668 (8.6)
Eq. (18), all data3.2e-3 (79)4.1e-3 (58)2.82 (31.6)0.905 (11.8)
SC with ours2.2e-3 (29)6.35e-3 (32)9.64 (89)1.11 (13.8)
SC with Eq. (18)8.9e-3 (112)0.0105 (53)6.24 (56)1.68 (20.9)

Across Appendix E.4, SC with our certificate gave values 1.1–5.5 times larger than ours on all the data. In the dimension study, the controller fitted on half the data also had 1.9 times (p=4p=4) and 2.8 times (p=1024p=1024) the true cost of the full-data controller (Table 12).

We also ran ten padded small-sample settings at ranks 1–6 (Appendix E.7, Table 16), including 8 of the 11 settings of Appendix E.2 where a simpler bound is below ours.

  • In 4 of the 10 settings at rank 2, SC with Eq. (18) was below our policy ball. The worst case is the orientation systems at p=1024p=1024 and n=256n=256, where ours is 284 and SC with Eq. (18) is 10.9.
  • Even there, Eq. (18) on all the data gives 5.2. At every rank in all ten settings at least one full-data certificate was lower than every split-and-certify variant. The split controller also had a higher true cost, 0.635 against 0.311.
  • Combining bounds helps more than splitting the data. Computed at the failure probability that PB+ assigns to each tail event, Eq. (18) uses ten events that PB+ already contains. The minimum of the two, PB+M, therefore holds with probability at least 1−δ1-\delta at no extra cost (Appendix A.15). In the worst cell it gives 5.6, slightly above the 5.2 that Eq. (18) gives with its own split of δ\delta. By construction it never exceeds PB+. In the ten settings of Appendix E.7 it was never more than 10% above Eq. (18), so it is the bound we would use in practice.
  • A bound specialized to one frozen controller could be tighter, but we do not derive one.

Other methods. Learn-then-Test and conformal risk control work with observed per-sample losses, and conformal risk control also needs them bounded and monotone in a tuning parameter. Our excess over an unknown optimum has no observable per-sample loss, is unbounded, and need not be monotone in the rank (Eq. 4).

Reviewer_hRgiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies linear control and asks what predictions are necessary for good decisions. In general, one does not need to identify the entire system in order to control it effectively (i.e., achieve low cost). However, the paper is completely incomprehensible, even to an expert. There is very little motivation, poor connection to prior work, way too much technical jargon, and essentially no intuition regarding technical definitions (Basically all the hallmarks of modern AI generated scientific writing). I do not think the paper can be accepted in its current form.

Critical strengths

From what I understand, the problem seems interesting, but I know that prior work studies similar things (e.g., this paper, https://arxiv.org/abs/2001.09576, gets optimal regret bounds for the LQR, which does not require "true" system identification. This paper, https://arxiv.org/pdf/2110.06150, makes a similar point, in structured LQR systems, we do not need to estimate all directions in order to control the system).

Critical weaknesses

The clarity and presentation is quite poor, to the point that an expert cannot really appreciate the paper.

Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Response to Reviewer hRgi: presentation, the setting in control terms, and the cited work

Thank you for the review and the two pointers. We agree that the submitted text was too compressed, defined terms too late, gave too little motivation and left related work to the end.

  • Sec. 1 now opens with the motivation and an example in which added outputs carry only noise, as happens when many sensors observe a low-dimensional system. A "Setting at a glance" paragraph states what is known, learned, compressed and certified (pp. 1–2).
  • A related-work paragraph in Sec. 1 discusses both papers you cite (p. 2).
  • Terms are defined at or near their first use, and in-house terms are replaced by plain ones. For example, "observable" is now "data-computable", "irrelevant outputs" are now "pure-noise outputs", and "orthoprojector" is now "orthogonal projector".
  • Each main result is followed by a paragraph that explains it in words, each section opens by saying what it shows, and Appendix D is a notation table.

The setting in control terms. We run nn experiments on a linear system, each from a chosen initial state, fit the HH-step free response DD and the impulse response BB by least squares, and plan open loop by certainty equivalence. The planner sees the initial state only through an rr-dimensional linear code, and we ask how small rr can be. From the data and the known noise level, we bound how much the plan's quadratic cost can exceed the optimum, with probability at least 1−δ1-\delta and for every plan u=−Txu=-Tx linear in the initial state xx. This is a one-shot suboptimality gap rather than online regret. Our bound (Theorem 2) covers measurement noise only, and a new stress test shows that process noise breaks it (Appendix E.9).

Why this is useful. In the new benchmark of Sec. 5, a chain of four masses driven by two forces is observed by 64 sensors per step, and 56 of the 64 sensor directions see only noise. A bound built on the usual operator-norm confidence set for the fitted impulse response, of the kind Dean et al. (2020) use for robust synthesis, is about 150 times the true excess cost there. At n=4096n=4096 ours is 6.0 and 6.9 times the true excess cost for horizons 5 and 10. It also certifies that a 3-dimensional code of the 8-dimensional initial state keeps the excess cost within 10% of that of not acting.

Why pure-noise outputs drop out. Such an operator-norm bound grows like the square root of the number of outputs pp. In Table 1, its median ratio to the true cost grows from 73.9 to 1229.0 as pp goes from 4 to 1024. We instead expand the fitted controller's error at the true optimal controller (Eq. 7). In the first-order term, each output's noise is multiplied by that output's row of BB or of the output map under the optimal plan, so an output with no signal and no actuation contributes nothing. Our ratio grows only from 13.0 to 21.7, which is flatter but still about an order of magnitude above the true cost. With many outputs and few samples, second-order terms dominate and simpler bounds are far tighter (Appendix E.2).

The cited work. Simchowitz & Foster (2020) show that certainty-equivalent control with continual re-estimation is minimax-optimal for the online linear-quadratic regulator, and Efroni et al. (2022) show that in structured partially controllable systems the optimal controller ignores a block of irrelevant states. Both show that control performance depends on only part of the identification problem, and we start from that principle rather than claim it. Our setting is narrower, with one open-loop plan fitted offline and no exploration, regret or feedback. Our guarantee is a bound computed from the data and valid for every linear plan, whereas their results are rates stated in terms of the unknown system. It needs no sparsity assumption, because it is unchanged when the output space is rotated. Outputs that carry cost but cannot be actuated (the output-side analogue of Efroni et al.'s block) still count in our bound, which reports their cost rather than removing it (Sec. 3). Theorems 4 and 5 address two questions these papers do not pose: how hard it is to choose the code when two initial-state directions matter almost equally for the decision, and how well a code serves costs not seen when it was fitted.

Reviewer_hRgiOpenReview ↗

Thanks

I appreciate the update but I think this paper needs a full round of reviews to fully assess the improvements in clarity. I'll keep my score

AuthorsOpenReview ↗

Response to Reviewer hRgi (follow-up)

Thank you for reading our response and for the frank assessment.

We understand that clarity is best judged by a full reading. To make that reading quick, the revision keeps the submission's numbering, and each of the four problems in your summary is addressed at a specific place in the revised PDF.

Motivation. Section 1 now opens with a concrete case of outputs that carry only noise and shows that the usual bound fails there. Its median ratio to the true cost rises from 73.9 to 1229.0 as the number of outputs grows from 4 to 1024, against 13.0 to 21.7 for our certificate (pages 1 and 2).

Prior work. The related-work paragraph on page 2 discusses both papers you cite and states that we do not study regret or exploration. Our contribution is a bound computed from the data and valid for every linear plan.

Jargon. Terms are defined at or near their first use, in-house terms are replaced by standard ones, and Appendix D is a notation table.

Intuition. The main results are followed by explanations in words, for example the paragraph after Theorem 2 on page 4.

Reviewer_JAa2OpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the following problem: estimate the forecast map D and the actuation matrix B from data, rank-r project D, and then compute the certainty-equivalent controller K_C D̂ P from these estimates. Then, they compute an upper bound from data on the excess cost with respect to the optimal controller, as a function of the noise scales, the estimates, and the distance from the uncompressed fitted controller. The bound holds for any linear controller, and its leading term does not grow with the output dimension p. They validate this mechanism in Fig. 1. They also show that the rank-r controller that minimizes this bound is the standard certainty-equivalent controller (Corollary 3). They give a lower bound on the error of choosing the r directions from noisy data (Theorem 4, Fig. 2), and a bound for new tasks (Theorem 5, Fig. 3). They also test implementations with neural networks instead of OLS (Fig. 4).

Critical strengths

The paper appears sound and technically accurate. The paper is clear about its limitations. Fig 1 give a clear demonstration of the main claim: As p increases from 4 to 1024, the certificate ratio of the policy ball increases slightly, but the ratio of the full-action stationarity bound increases dramatically.

Critical weaknesses
  • The paper is hard to read, both from language and math equations. For math, the authors consistently rely on the reader to infer mathematical definitions. For example, the authors do not define the objective "R", which is the regret wrt the optimal policy. They also write E and V without definition and never use them again. For language, many terms are used before they are defined or are never defined in the main text: “forecast,” “Plug-in,” “Prediction,” the “full-action stationarity” bound (the main comparator in Fig. 1), “tail units,” “pilot,” “student,” and “excitation.”

  • The setting is narrow, limiting the significance of the actual result. For example, the setting has a one step forecast from a known reset, and has no multistep dynamics or closed loop behavior.

Questions for authors

Please clarify if I misunderstood any weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JAa2: the definitions, and the multi-step and closed-loop setting

Thank you for the careful reading. Both weaknesses are accurate, except that the setting is not limited to one-step forecasts (W2 below). One qualification to your summary is that the bound's second-order terms still grow with pp and dominate with many outputs and few samples (Appendix E.2).

W1: definitions. Each term you list is now defined in the main text at or near its first use; Appendix D is a notation table.

  • The forecast (Sec. 2) is DxDx, the outputs expected for input xx without acting. DD is the forecast map, BB the action map, and D^\hat D and CC their least-squares estimates.
  • R(T)\mathcal R(T) (Sec. 2, Eq. 3) is the expected excess cost of u=−Txu=-Tx over the optimal controller KBDK_BD (your regret). With Exx⊤=I\mathbb E xx^\top=I it equals ∥HB1/2(T−KBD)∥F2\Vert H_B^{1/2}(T-K_BD)\Vert_F^2, where HB=I+B⊤BH_B=I+B^\top B.
  • The revision drops the factorization F=VEF=VE, so a rank-rr forecast is a matrix FF of rank at most rr (Sec. 2), and the sans-serif E\mathsf E of Prop. 1 is a separate error matrix defined there.
  • Plug-in (Sec. 2) keeps the input directions that matter most for the fitted decision, the top-rr eigenvectors of L^=D^⊤WCD^\hat L=\hat D^\top W_C\hat D, where WCW_C (Eq. 2) weights each output direction by how much cost actions can remove under CC. Prediction ignores the decision and keeps the top-rr principal components of the fitted forecast. Both are deployed through KCK_C, the certainty-equivalent gain of the fitted action map.
  • The full-action stationarity bound, the comparator in Fig. 1, is named in Sec. 1 and the caption and defined in Sec. 5. It bounds the gradient of the true cost using a confidence ball for the whole action-map error, whose radius grows like p\sqrt p.
  • A tail unit (Sec. 5) is one Gaussian or χ2\chi^2 tail event. The policy ball splits its failure probability δ\delta equally over twelve of them (Sec. 3).
  • The excitation cc (Sec. 3) is the energy of the action probes relative to that of the input probes, so with nn samples and output-noise standard deviation σ\sigma the action-noise scale is b=σ/ncb=\sigma/\sqrt{nc}.
  • The pilot (Sec. 8) is an uncompressed teacher network trained on 1,024 labelled inputs, and students are bottleneck networks distilled from it with an unweighted squared error (Prediction) or a WCW_C-weighted one (Learned geometry).

Every main-text caption now also says what its figure or table shows.

W2: narrow setting. We partly agree, because multi-step open-loop plans are covered and now tested, but feedback and single-trajectory data are not.

  • Multi-step plans (Sec. 2). For zt+1=Azt+B0utz_{t+1}=Az_t+B_0u_t and yt=C0zty_t=C_0z_t with an observed z0z_0, the HH stacked outputs equal Oz0+Mu\mathcal Oz_0+\mathcal Mu, where O\mathcal O maps the initial state to the outputs and M\mathcal M is the block-Toeplitz matrix of impulse responses, the standard batch form of finite-horizon linear-quadratic control (Borrelli et al., 2017). With quadratic stage costs and a known Ez0z0⊤\mathbb E z_0z_0^\top, a change of variables gives exactly the one-step form of Eq. (1), and one reset experiment is one HH-step episode. Theorem 2 applies when DD and BB are fitted by unrestricted least squares from an orthogonal design and the cost-weighted outputs have independent Gaussian noise of known variance, for example with unit output weights and only measurement noise. Since the numbers of outputs and actions grow linearly with HH, so can the bound.
  • New test (Sec. 5, Appendix E.8). On a four-mass spring–damper chain with horizons 5 and 10 and up to 640 outputs, no certificate was violated in 2,400 fits. At n=4096n=4096 and full rank, the policy ball is 4.9–6.9 times the true cost, against 39–151 for the stationarity bound, and at a tolerance of 10% of the excess cost of not acting it certifies the smallest sufficient rank, 3, on every fit.
  • Process noise correlates the stacked outputs and is not covered. In a new stress test (Appendix E.9), state noise as large as the measurement noise made the policy ball fail on 77% of fits. The stationarity bound extends to a known noise covariance (Appendix A.13) and stayed valid, but in practice that covariance depends on the unknown system.
  • Feedback. The excess cost of a feedback gain is weighted by the state covariance under that gain and by the unknown Riccati solution (Eq. 14). Neither Eq. (3) nor our ball carries over, and stability would also need certifying (Appendix A.11). Without disturbances an optimal open-loop plan loses nothing to feedback over the horizon, so the limitation matters with disturbances or model error.
  • Reset data make the least-squares errors exactly Gaussian (Eq. 6), and single-trajectory data would need different tools (Appendix A.12).

The setting is stylized, but its exact cost formula (Eq. 3) shows why a decision certificate with a learned action map need not pay for outputs carrying only noise. We do not claim an extension to closed-loop latent world models.

I thank the authors for their detailed response. The revision's presentation is significantly improved, and the multi-step work extends the result so I am increasing my score.

AuthorsOpenReview ↗

Response to Reviewer JAa2 (follow-up)

Thank you for reading the revision and for raising your score.

Every term listed in your review is now defined in the main text at or near its first use, and Appendix D is a notation table. Multi-step open-loop plans for linear systems are a special case of Equation (1), and a new benchmark on a four-mass spring–damper chain with horizons 5 and 10 tests them (Section 5 and Appendix E.8). No certificate was violated on its 2,400 fits, and with 4,096 samples and a tolerance of 10% of the excess cost of not acting, the certified rank equals the oracle rank, 3, on every fit. Feedback and process noise remain outside the guarantee, as Section 9 states.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of the revision

We thank the reviewers for their careful reading. The revised PDF keeps the submission's numbering and revises every section for clarity. New material is mainly in Secs. 1–3, 5 and 9 and in Appendices A.13–A.16, C, D and E. Table 4 and Fig. 4 moved to Appendix C for space.

The paper bounds, from data, the excess cost of every linear controller when both the forecast map (predicting outputs without action) and the action map (the effect of actions on the outputs) are fitted by least squares. Its contribution is this certificate, whose leading term does not grow with outputs that carry only noise. The compression method itself is standard.

  1. Presentation and related work (hRgi, JAa2). The abstract and introduction are rewritten in plain language, with the motivation first and a related-work paragraph, and every term JAa2 listed is now defined at or near its first use.
  2. Scope (sM6G, JAa2). The abstract and Sec. 1 state the assumptions: a linear model, Gaussian least-squares errors with known noise levels, and open-loop plans, which may span many steps but exclude feedback.
  3. New results (sM6G, JAa2). Prop. 7 sharpens the certificate, and Prop. 8 limits how tight any certificate can be when the action map is known. New synthetic experiments (Sec. 5, Appendix E) measure certified ranks and sample sizes, compare with split-and-certify (fitting on half the data and certifying on the other half), add a multi-step benchmark, and stress-test the assumptions. They also show that the certificate is conservative, often certifies nothing at small sample sizes, is far looser than simpler bounds with many outputs and few samples, and fails under process noise, which the theory excludes.
  4. One correction. The measured excess cost on new tasks in Sec. 7 is 3.22×10−43.22\times10^{-4}, not 3.20×10−43.20\times10^{-4} as submitted.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.