acceptodds
Under review as a conference paper at ICLR 2027

State-Consuming Experimental Design: Greedy Failure and Amortized Planning

Abstract

An experimental design action normally reveals information and leaves the system available for the next action. A quantum measurement does not: it consumes the state that later measurements needed. We show this is not a quantitative nuisance but a change of problem class. Under back-action the design utility stops being a set function and becomes a function of the measurement sequence, and we prove that greedy selection admits no constant-factor shot-cost guarantee over the stated family with no duration or resource restriction: on an explicit two-parameter family, for every fixed regularizer and every finite some member forces shot overhead above , with the overhead tending to as the frequency scale . The argument loses its premise—the utility is not a set function—rather than a degraded ratio; unbounded cost alone would not falsify the bound, and we locate the regularizer-coupled regime where the normalized ratio itself drops below . Numerically the failure is severe and specifically attributable: greedy pays a shot penalty on an instance whose fresh-copy control shows no gap at all ( on a second, where fresh copies reverse the ranking), and search closes the gap only at -leaf exhaustive cost. The answer the structure dictates is amortization. A policy trained once per instance distribution, spending zero utility evaluations at deployment, beats greedy on all four conditions of the tested random-Hermitian family (– of – paired instances, Holm ), and adding a budget-parity repair sweep at greedy's own evaluation count reaches the exact optimum in – of instances against greedy's –; the policy trains without the exhaustive-search teacher ("teacher-free" below) at lower cost. We state the boundary as plainly: on device-physics instances the improvement requires the simulator-based repair sweep and is therefore not yet hardware-deployable, and "teacher-free" means free of the exhaustive-search teacher, not free of the simulator utility oracle during repair. We preregister a deferred-measurement compilation of the experiment for an ion-trap processor; the experimental artifact package is not included with the current submission.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
experimental design, quantum measurement, measurement back-action, submodularity, amortized policy learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_83N4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission appears to be AI slop

Critical strengths

This submission appears to be AI slop

Critical weaknesses

This submission appears to be AI slop

Questions for authors

This submission appears to be AI slop

Reviewer_ChU3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper appears to be about the impacts of quantum measurement on experimental design, and that the nature of quantum measurements changes the class of design problem. In particular, later outcome distributions depend on earlier measurements, so design utility is a function of an ordered sequence and not a set function. The authors present both theoretical claims and numerical experiments.

Critical strengths

None.

Critical weaknesses

W1. This paper is completely inaccessible as written, and makes it impossible to verify the claims in the paper. Stylistic choices lead me to suspect that it was predominantly (if not entirely) written by AI. Throughout the abstract and introduction, notation (e.g., \rho_D, \epsilon, d and |\Acal|) that is never defined is used to quantify arguments. At many points in the text, there are blocks of text headlines with seemingly no connection to the material beforehand. The problem that the authors want to solve is never formalized. There are 30 sections in the appendix. Many are listed as "Exploratory" and leads me to further doubt the scientific merits of the work.

W2. The results in the body have major issues. A rigorous proof is not provided for Proposition 1. The provided proof makes assertions about the relevant Fisher matrices and determinants, but never derives them; instead it seems some numerical comparison was made with finite differences. A proof might exist, but it is not given.

W3. I do not understand how the title of the Proposition 2 is related to the (one sentence) proof that is given. The instances described in the proof are not made explicit. Putting these issues aside, I am not even sure the computation that the authors do substantiates their claim of "not a worse action multiset".

W4. Curiously, no AI usage statement was included by the authors.

Questions for authors

Can you provide fully rigorous and explicit proofs of your claims?

How are you generating your random Hermitian family?

Why is Prop. 2 titled "The deficit is back-action, not a worse action multiset"?

Please include the random-Hermitian distribution and the exact description of the instances for Prop 2 in the main body.

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

This paper shows that quantum measurement back-action transforms experimental design utility from a set function into a sequence function, causing greedy selection to lose its constant-factor (1−1/e)(1-1/e) regularized D-efficiency guarantee. It proposes an amortized policy requiring zero deployment Fisher-utility evaluations.

Critical strengths

The paper presents no clear strengths.

Critical weaknesses

Most of the writing is unreadable for humans (at least for me). Moreover, despite clear indicators of AI-generated text (Figure1/detailed experimental reports in the main text), there is no statement of AI usage.

Questions for authors

What is the main claim of this work? While the theoretical proof of greedy failure is noted, the amortized planning method only succeeds on synthetic random-Hermitian instances. Instead of defensively writing the negative results on device physics family, could the authors better justify the value of studying this problem on the random-Hermitian family?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.