acceptodds
Withdrawn from ICLR 2027

Measurement as Cold Diffusion:Pricing the Shot Frontier

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pipelines that reconstruct a quantum state's full Born distribution are usually built by importing classical Gaussian diffusion onto the device, which assumes away the cost that dominates them: the distribution must be read out by measurement, and estimating it over outcomes converges only as in the shot budget . We instead take the measurement to be the forward process, making the shot count the diffusion time. We prove that its control parameter is the only one among standard cold-diffusion corruptions that is a genuine resource—costing shots to hold the error fixed, while the Gaussian, blur and mask knobs are free—with a predicted constant confirmed to and the scaling verified over a range in . A learned prior bends that frontier, flattening it and saving up to of inference-time readout only for fixed-complexity data and only once the prior's acquisition cost is amortized; on ground-state Born distributions of an Ising chain the flattening fails. The corruption is also not interchangeable: restorers move between multinomial and Poisson readout at a cost of – but lose – moving from Gaussian or blur onto real readout. Finally, the prior does not appear able to be quantum. We characterize the class a postselecting circuit restorer realizes—the projectivization of a linear map, closed under coherent composition—and find classical controls with fewer parameters beating it at every scale, with logged curves showing it at a floor rather than out of steps. We are explicit that this is a characterization, not an impossibility proof. On ibm_kawasaki the idealized forward model tracks the device only while shots bind, yet a restorer trained solely on it still halves device error at the state-preparation floor.

Primary area:
generative models
Keywords:
quantum machine learning, cold diffusion, finite-shot measurement, Born distribution

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_z61DOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper makes the assertion that current approaches to quantum generative models do not accurately account for the sampling cost required to achieve a fixed error. It propose a cold diffusion based approach to address this issue

Critical strengths

None

Critical weaknesses

This is not an easily readable paper. The fundamental problem and it's importance is not motivated. The main technical contribution is not clearly explained. Large parts of the text read as entirely LLM generated.

The authors say that the current quantum generative modelling frameworks do not account correctly the number of samples required to get to a certain accuracy measured in TV. But this is true for classical generative models as well. In fact, this is a computationally intractable problem, and saying "..it is physically wrong in a way that matters." is a misrepresentation of the literature.

I don't see any particularly interesting quantum ingredients in the paper. The cold diffusion technique is directly applied to the Born distribution of a quantum state. The savings generated by this technique is interesting but not substantially novel to warrant publication in ICLR.

Questions for authors

Please address the weakness section directly.

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

The manuscript is extremely difficult to follow in its current form. In particular, after reading the abstract and introduction, I still cannot clearly identify the central research question, the main methodological contribution, or the key quantitative findings. The abstract introduces a large number of concepts and claims—such as measurement-as-diffusion, cold-diffusion corruptions, learned priors, quantum restorers, and hardware experiments—without establishing a clear logical connection among them. I strongly recommend that the authors substantially rewrite both the abstract and the introduction. The presentation should clearly separate: (i) the problem being addressed, (ii) the proposed method, (iii) the main theoretical result, and (iv) the main numerical or experimental findings. In its current form, the paper reads more like an insufficiently edited draft than a mature manuscript.

Critical strengths

I cannot follow the authors‘ idea and method, and thus cannot find any strength.

Critical weaknesses

No

Questions for authors

No

Reviewer_11AwOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose that the measurement process can be considered the forward diffusion process for reconstructing Born distributions. They demostrate that the number of measurements is the limiting resource, with exponential scaling. A trained prior can reduce the computational cost for data of constant complexity but not for examples such as the ground states of the Ising chain. Furthermore, the study examines when different perturbations can be transferred between each other and evaluates the limitations of quantum circuit restorers. Overall, the contribution more realistically links machine learning to the actual cost of quantum measurements.

Critical strengths

a. The work shifts the focus from artificial noise in the device to the actual cost of data collection i.e. the measurements. b. The work provides an asymptotic analysis of the \Theta(2^n) bound and identifies which distortion parameters constitute actual resources. c. The results can help in designing more efficient pipelines for quantum data. d. The authros include confirmation of the theoretical constant + verification across a wide range of scales and testing on IBM hardware.

Critical weaknesses

a. The savings hold only when the complexity of the data does not increase as it fails for Born distributions of Ising ground states. b. The failure of circuit restorers is characterized as an indication rather than a strict proof of impossibility. c. The idealized forward model approximates IBM hardware mainly at a low number of measurements. Therefore, further validation is required under more realistic noise and scale conditions. d. Limited significance due to limited experimentations

Questions for authors

a. What further experiments can be performed in the future to scale up and clarify the theoretical results of the paper?

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

Related papers

Discussion (0)

Sign in to comment.