acceptodds
Withdrawn from ICLR 2027

Fixed-Protocol Amortized MPS Tomography with Conformalized Predictive Uncertainty

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum state tomography is sample-starved, and the states one prepares live on a narrow, learnable manifold. A prior-only control shows that on concentrated families a prior estimate is already near-optimal, so “high fidelity at few measurements” can be family memorization rather than tomography; genuine measurement-efficiency needs a model that conditions on the measurements and demonstrably uses them. Our method is a fixed-protocol amortized matrix-product-state (MPS) estimator: one fixed informative protocol mapped directly to MPS cores, trained once with a gauge-invariant fidelity loss; we deliberately do not rest it on a permutation-invariant set encoder (a plain MLP matches it). The learned generative prior over the same cores, inverted by measurement-guided posterior sampling, is what makes the control bite in the first place, and we retain it only as a complementary route (Approach A, deferred to the appendix). The decisive lever is the measurement design: motivated by the fact that local reduced density matrices determine a -MPS, conditioning on an informative local Pauli set rather than random strings turns a modest, memorization-prone estimator into a high-fidelity one (, up to over prior-only, decisively passing a shuffled-measurement control). A dropout ensemble, conformally recalibrated, gives -coverage intervals—including for observables never measured, where a shot-based interval does not exist. Quality holds as the system grows (fidelity at , gain growing in ; at bond dimension ), the parameterization is polynomial (native contraction to qubits), and we close the loop on IBM hardware ( states at from hardware-measured Paulis). Our prior-only and shuffled controls are a necessary integrity check we argue learned-QST work should adopt.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum state tomography, matrix product states, amortized inference, measurement design, conformal prediction

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_ehQAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a learned MPS-based method for quantum state tomography. The model is trained for a fixed state family and measurement protocol, and then reconstructs new states from their measurement results. The aim is to reduce the optimization cost for each new state. Dropout and conformal calibration provide uncertainty estimates for observables. The evaluation includes simulated state families, comparisons with MPS-MLE, reconstruction experiments up to 10 qubits, feasibility tests at 16 and 20 qubits, and a small IBM hardware experiment.

Critical strengths

The reported reconstruction results are reasonably good. Under the default local measurement design, the model achieves fidelity of about 0.95–0.97 at six qubits and 0.90 at ten qubits in the tested settings.

Critical weaknesses
  1. The comparison with traditional methods is not extensive enough. The main comparison in Table 2 is at six qubits. The eight- and ten-qubit reconstruction experiments do not include a corresponding comparison with MPS-MLE under the main local measurement setting. This makes it difficult to judge how the proposed method compares with traditional methods as the number of qubits increases.
  2. The practical range of the amortization benefit needs stronger validation. The estimated break-even point of approximately 3755 states is based on one reported timing configuration, and the authors explicitly limit it to repeated reconstruction within the same protocol and state family. Although similar average fidelities are reported, a runtime comparison at a common reconstruction-quality target would make the evidence stronger. It remains unclear how the break-even point changes with the fidelity target, number of qubits, or measurement design.
Questions for authors
  1. Clarification: How was comparable reconstruction quality established for the timing comparison? Could the authors explain how the estimated break-even point changes with the fidelity target, number of qubits, or measurement design?
  2. Additional validation: Can the authors provide comparisons with MPS-MLE at eight and ten qubits under the main local measurement setting, reporting both reconstruction quality and runtime? This would help assess the computational advantage as the problem size increases.
Reviewer_Cv1uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies quantum state tomography and aims to bypass the worst-case exponential sample complexity bound by leveraging structural properties of lab-prepared states. Many prior works have studied specific structural assumptions for the unknown states, such as matrix product state, low rank state, and states prepared by shallow circuits. This paper proposes a method to learn the structural priors based on prior lab data.

Critical strengths

Learning properties of well-structured quantum states is an important problem in quantum computing for benchmarking quantum algorithms and devices.

Critical weaknesses
  1. The main weakness is the presentation. I find it very hard to understand the specific problem setup, as well as the algorithm the authors propose. The presentation is not very self-contained, with many undefined expressions and terminology. For example, the authors mention k=0k=0 in the abstract and line 56, but no explanation of its meaning near where this notation is first introduced. The experimental setup and results are also not explained clearly, so it is very difficult to assess the novelty and effectiveness of the proposed method.
  2. The main body of the paper lacks thorough comparison and discussion of related works. The paper does not have a related works section. The experiment results do not show the performance of existing works.
  3. The problem setup assumes observing expectation values of Pauli observables: this corresponds to the two-outcome Pauli observable measurement, and is well known to be suboptimal for learning quantum states. At least the author should consider Pauli basis measurements.
  4. Full state tomography is often a significant overshoot in practice. Except for some very well-structured states (like product states), learning the full state requires exponential time in the worst case even for pure states. Practitioners are often more interested in testing whether the state satisfies desired properties, or learning expectation values of a set of observables (shadow tomography).
Questions for authors

Please provide a clear explanation of the problem setup, as well as a clear description of the algorithm/protocol proposed.

Reviewer_3b2DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a fixed-protocol amortized quantum state tomography method that directly maps noisy local Pauli measurements to matrix product state cores using a gauge-invariant fidelity objective. By prioritizing informative measurements over random configurations, the method achieves 0.95–0.97 reconstruction fidelity on six-qubit states and scales up to twenty qubits via tensor-network contraction. Furthermore, it provides calibrated predictive uncertainty through dropout ensembles and split-conformal calibration, offering an efficient approach for repeatedly measured quantum state families.

Critical strengths
  1. The most compelling contribution is the explicit use of zero-measurement prior-only and shuffled-measurement controls. The authors demonstrate that high reconstruction fidelity on concentrated state families can arise almost entirely from learned priors, without meaningful use of measurement information. On more diverse families, their estimator achieves substantially higher fidelity with correct measurements than with shuffled measurements, providing convincing evidence of genuine measurement conditioning.
  2. The paper demonstrates that local measurement design dramatically improves reconstruction fidelity, while the MPS parameterization supports efficient state representation and contraction. The conformalized uncertainty estimates achieve approximately 90% marginal coverage, including for selected observables not directly supplied to the estimator.
Critical weaknesses
  1. Limited novelty and incomplete theoretical characterization of the measurement design. The primary performance improvement comes from using informative local Pauli measurements, a principle already motivated by established MPS tomography theory. The amortized estimator uses relatively standard neural architectures, and the conformal calibration procedure applies established split-conformal methodology. Furthermore, Proposition 1 does not rigorously establish the sufficiency of the particular local measurement design used in the main experiments. The proposed window-length estimate and tangent-space spanning property require additional assumptions, while the cheaper default protocol remains empirically motivated. Consequently, the contribution is primarily an effective integration and evaluation framework rather than a fundamentally new tomography or uncertainty-estimation principle.
  2. The practical advantages are demonstrated under relatively restrictive conditions. The method assumes a fixed measurement protocol and a sufficiently representative training distribution. While high-fidelity reconstruction is shown up to ten qubits, the twenty-qubit experiment demonstrates only computational feasibility, with low reconstruction fidelity under the reduced measurement budget. Additionally, the amortized approach requires approximately 3,755 reconstructed states to recover its training cost relative to classical MPS-MLE. The five-state IBM experiment is encouraging but does not establish accurate reconstruction of the actual noisy device state. These limitations leave the practical advantage over conventional reconstruction methods less convincing for settings with limited repeated measurements or distribution shift.
Questions for authors
  1. Clarification on Proposition 1: The paper derives a local window-length estimate and assumes the associated measurement map generically spans the MPS tangent space. Can the authors clarify which parts follow rigorously from existing MPS identifiability results and which remain additional assumptions? In particular, what theoretical justification supports the sufficiency of the default two-local measurement design beyond the empirical results?
  2. Clarification on hardware validation: The reported IBM hardware fidelities are measured relative to ideal circuit targets. Can the authors clarify what evidence establishes reconstruction of the actual noisy hardware states rather than recovery of the intended ideal preparations? How should this distinction affect the interpretation of the hardware results?
  3. Clarification on practical efficiency: The amortization analysis gives a break-even point of approximately 3,755 states. Does this comparison include the complete uncertainty-estimation procedure and equivalent uncertainty estimation for the classical baseline? Which practical experimental settings are expected to involve sufficiently many reconstructions from the same state family and protocol?
  4. Clarification on predictive uncertainty: The reported conformal guarantees provide marginal coverage under exchangeability. Can the authors clarify whether the empirical intervals remain informative under the distribution shifts considered in the appendix, particularly when reconstruction fidelity deteriorates? This would help distinguish calibrated in-distribution uncertainty from robustness to unfamiliar quantum states.

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

Related papers

Discussion (0)

Sign in to comment.