Optimized Circuit Cutting for QAOA Sampling Tasks

Christian Ufrecht; Daniel D. Scherer; Friedrich Wagner; Martin Braun

REVIEW 4 major objections 6 minor 20 references

Optimized Circuit Cutting for QAOA Sampling Tasks

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper reports that, on the largest MaxCut instance tested, cutting the QAOA circuit into pieces shifted the hardware-sampled solutions closer to the optimum than running the whole circuit.

desk verdict Honest proof-of-principle that circuit cutting can mitigate noise for QAOA sampling, but the central claim rests on a single un-replicated 25-node hardware run. read the letter →

arxiv 2507.06649 v1 pith:PLT6LC4Q submitted 2025-07-09 quant-ph

Friedrich Wagner , Christian Ufrecht , Martin Braun , Daniel D. Scherer This is my paper · ORCID

classification quant-ph

keywords circuitcuttingwireQAOAMaxCutnoisemitigationsamplingvertexseparatorquantumapproximateoptimizationalgorithm

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Circuit cutting splits a quantum circuit into smaller fragments and was originally built for estimating expectation values. This report argues that it can also be used for sampling tasks such as reading out solutions from a trained QAOA circuit, and that doing so can reduce noise rather than only make circuits fit on small hardware. On a 25-node MaxCut instance, the distribution sampled from the cut circuit on a real device was shifted closer to the optimal solution than the distribution from the uncut circuit, even though cutting alone broadens and degrades the distribution in noiseless simulation. The authors conclude that for large enough circuits the noise reduction from smaller width and depth can over-compensate the cutting-induced distribution shift, making optimized circuit cutting a candidate noise-mitigation tool for near-term quantum optimization.

What carries the argument

The load-bearing mechanism is a wire-cutting decomposition applied to each QAOA layer at a single separating vertex, combined with classical post-processing that reconstructs bitstrings of the original circuit from samples of the two fragments. The number of cuts is minimized beforehand by solving an integer program for a minimum balanced vertex separator and then shrinking that separator to one vertex, so each layer needs two cuts total. A central quantitative object is the normalized objective $r = (c - c_0)/(c^* - c_0)$, which rescales every sampled cut value so that $r=0$ is random guessing and $r=1$ is the optimum; the 95th percentile of this quantity is the comparison metric across instances. The paper also uses a theoretical sampling-overhead bound of $\kappa = 12$: reconstructed bitstrings are suppressed by at most $1/12$ relative to the uncut circuit, so a constant factor more samples recovers any bitstring with high probability.

What would settle it

Repeat the 25-node cut-versus-uncut experiment several times on the same device and on other devices, interleaving the two circuits and recording device calibration; the central claim fails if the cut distribution's 95th percentile is not consistently above the uncut distribution, or if the advantage disappears when device noise is low.

Watch

Extended reading notes

Core claim

The discovery is that circuit cutting, originally a technique for exactly reconstructing expectation values, can be repurposed for a pure sampling task and, on sufficiently large or noisy circuits, act as net noise mitigation. The paper samples from two-layer QAOA circuits for MaxCut instances on 10 to 25 nodes, using a minimum balanced vertex separator plus graph shrinking to reduce each circuit to two fragments connected by a single vertex, and cuts that vertex once per layer using two complementary wire-cutting schemes. In noiseless simulation the reconstructed bitstring distribution is broadened and shifted toward suboptimal values relative to the uncut circuit. On the real device, for the 25-node instance the cut distribution is instead shifted closer to the optimum than the uncut distribution, and its 95th percentile of the normalized objective is higher. The paper attributes the reversal to reduced width and depth of the cut fragments, and notes that the same benefit did not appear for the smaller 10-, 15- and 20-node instances.

Load-bearing premise

The load-bearing premise is that the better hardware result for the 25-node cut circuit comes from the reduced noise of the smaller circuit, and not from chance, calibration drift, or some other difference between the two runs, since the report presents a single run of 100,000 samples per method without repeated runs or confidence intervals.

Editorial extensions

If this is right

For circuits large enough that hardware noise dominates, cutting can improve the sampled solution quality rather than merely reduce circuit size, so the technique doubles as noise mitigation.
The 95th-percentile comparison suggests the crossover point depends on circuit size and device noise; small instances show no benefit because the uncut circuit is already short and the cutting overhead dominates.
Because the reconstruction overhead is bounded by a constant factor ($\kappa = 12$) independent of instance size, users can compensate for the distribution broadening by drawing more samples rather than by changing the algorithm.
The approach preserves one-way classical communication only, since the two wire-cutting schemes are alternated across layers, so it runs on currently available hardware without real-time links.
With current cutting overheads, practical use for QAOA will require refined, low-overhead cutting schemes; simply cutting wires naively is too expensive as the number of layers grows, and plain classical enumeration over the separator has lower overhead.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the noise-mitigation crossover is real, then the optimal place to cut a QAOA circuit is not only where the separator is cheapest but where the two fragments minimize total device error, suggesting an error-aware search over cut locations rather than only separator cardinality.
The single hardware run for the 25-node instance should be tested for reproducibility: repeating the cut-versus-uncut comparison on several devices and at several times of day would separate a genuine noise-reduction effect from calibration drift.
The constant-factor sampling overhead suggests a natural hybrid: use the uncut circuit when the device noise is low, and switch to a cut circuit when calibration data indicates the depth penalty will cost more than the cutting overhead.
One could test the mechanism directly by artificially increasing gate error rates on the uncut circuit in simulation; if noise reduction is the cause, the cut circuit should overtake the uncut circuit exactly when the simulated error model crosses the observed hardware error level.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the paper is a clean, honest proof-of-principle that wire cutting can shift a QAOA sampling distribution back toward the optimum on real hardware. It does not overclaim. The headline result is one 25-node MaxCut instance where the cut histogram beats the uncut one; the three smaller instances show no benefit, and the authors say so plainly. They also admit that classical cutting and even classical solvers beat the whole approach. That level of candor deserves credit.

What is actually new is the hardware demonstration. The cutting-and-shrinking pipeline is from the authors' previous work, so the novelty is incremental, but the experimental observation is new. The paper also does a good job explaining the integer-programming preprocessing and the specific wire-cutting scheme.

The soft spot is exactly where the stress-test lands: the central claim of noise mitigation is supported by a single 25-node run. There are no error bars, no repeated runs, no interleaving of cut and uncut circuits, and no device calibration data. The 100,000 samples per histogram make the sampling error small in a pure statistical sense, but the difference in the 95th percentile could easily be drift in qubit error rates or readout errors over time. The causal attribution to noise reduction is plausible, but it is not established. The authors label this a proof-of-principle, which is the right word, but a referee should push hard for more evidence before this becomes a broader claim.

The theoretical bound kappa = 1/12 and the reconstruction method come from prior work and are not fitted to the hardware result, so circularity is not a concern. The math and the methodology described are sound; the weakness is purely experimental rigor.

Who is this for? People working on circuit cutting, QAOA, or noise mitigation. It is a short technical report, not a full paper, but it makes a useful observation and is honest about its limits. I would send it to peer review with the expectation of heavy revision, mainly to get repeated runs, interleaved schedules, and calibration data. But even in its current form it is worth reading as a candid status report.

Referee Report

4 major / 6 minor

Summary. The paper reports proof-of-principle experiments applying optimized wire cutting to two-layer QAOA circuits for MaxCut instances on 10, 15, 20, and 25 nodes. The authors use an integer-programming-based balanced vertex separator, graph shrinking, and a combination of two wire-cutting methods to minimize the number of cuts. On a noiseless simulator, cutting broadens and shifts the sampled objective-value distribution toward suboptimal values, as expected. On IBM hardware, the 25-node instance yields a cut distribution with a higher 95th percentile than the uncut distribution, while the three smaller instances show no such improvement. The authors interpret this as evidence that, for sufficiently large circuits, the noise reduction from reduced circuit width and depth outweighs the distribution degradation introduced by cutting. The paper explicitly acknowledges that the experiments are proof-of-principle, that all instances are classically solvable in fractions of a second, and that simple 'classical cutting' has lower sampling overhead.

Significance. If the single positive hardware observation were robust, the claim that circuit cutting can act as a noise-mitigation tool for QAOA sampling would be a genuinely interesting and counterintuitive result with practical relevance for near-term quantum optimization. The paper is honest in reporting that three smaller instances showed no benefit and in listing the limitations of the approach, including exponential sampling overhead and the availability of lower-overhead classical alternatives. The theoretical bounds from the authors' prior work provide a non-fitted, independent grounding for the sampling budget. However, the central empirical claim rests on a single hardware run without error bars, repeated runs, or calibration data, so the significance is currently preliminary rather than established.

major comments (4)

[Section II d, Figure 2] The central claim that noise reduction outweighs the derogative effects of cutting is supported by exactly one hardware comparison: the 25-node instance in Figure 2. No error bars, no repeated runs, and no device calibration or drift information are reported for either the cut or uncut execution. With a single run per method, the observed advantage could be a statistical fluctuation or a time-dependent drift in qubit or readout errors, especially since the cut and uncut circuits were presumably not interleaved. To make the causal attribution to noise reduction load-bearing, the authors should report repeated runs (ideally interleaved), confidence intervals on the 95th percentiles, and device calibration data, and apply a significance test to the difference.
[Section II d, Figure 3] The paper states that for the 10-, 15-, and 20-node instances 'we did not observe this improvement,' but no statistical quantification is provided for these null results. Without confidence intervals or effect sizes, the absence of an observed improvement cannot be distinguished from a small true effect or from run-to-run noise. Since the abstract's claim is specifically about 'large circuits' and the evidence is a size trend, the authors should quantify the uncertainty in the 95th percentile differences for all four instance sizes.
[Abstract and Section III] The abstract concludes that circuit cutting can 'both scale problem size and mitigate noise for near-term quantum optimization,' but the experiments do not demonstrate scaling in any meaningful sense: the largest instance has 25 nodes and is, as the authors note, solvable to proven optimality by classical solvers in fractions of a second. The demonstrated effect is limited to a single noise-mitigation observation. The conclusion should be reworded to avoid overstating the scaling aspect, or additional evidence of scaling beyond trivially classically solvable sizes should be provided.
[Section II b and II c] The manuscript lacks the device and circuit details needed to support the noise-reduction explanation: no execution date, no information on the number of physical qubits used, no transpiled gate counts or circuit depths for the cut versus uncut circuits, and no two-qubit error rates or readout error rates for ibm_aachen. These data are essential for evaluating whether the observed improvement is plausibly caused by reduced circuit size rather than by unrelated device conditions.

minor comments (6)

[Section II (intro)] There is a typo in 'incured' (should be 'incurred') in the sentence describing the distribution broadening.
[Abstract and Section II d] The phrase 'derogative effects' is unusual; 'deleterious effects' or 'degradation' would be clearer.
[Figure 3 caption] The caption mentions 'dotted' and 'dashed' lines, but the figure appears to use markers (crosses and circles); please clarify the visual encoding and ensure the caption matches the rendered figure.
[Equation (1)] The definition c0 = |E|/2 assumes that a uniformly random cut has expected value |E|/2, which is true for unweighted graphs; this assumption should be stated explicitly.
[References] Reference [16] is a consortium best-practice guide; for the claim that MaxCut models key-customer identification in social networks, a standard peer-reviewed reference would be more appropriate.
[Overall] No data availability statement is included; providing the raw histograms, the transpiled circuit data, and the device calibration data would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is an empirical hardware comparison, with prior-work self-citations used only as background method and not fitted to the measured result.

full rationale

The report is an experimental comparison, not a derivation. The central claim that noise reduction outweighs distribution broadening for large circuits rests on the observed 25-node hardware histograms (Figure 2), which are not constructed from the paper's inputs. The QAOA parameters are precomputed once and used identically for the cut and uncut circuits, and no parameter is fitted to the hardware outcome. The cited prior work [1] supplies the wire-cutting/reconstruction method and the 1/kappa = 1/12 sampling-overhead bound, but that bound is not used to predict the hardware shift; it only motivates the sample budget. These self-citations are background method and are not load-bearing for the empirical comparison. The absence of error bars, repeated runs, and calibration data on the single positive 25-node run is a statistical robustness concern, not evidence of circularity. Accordingly, no step reduces by construction to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a prior theoretical framework (Ref [1]) for circuit-cutting-based sampling, plus a small set of experimental choices (instance selection, sample budget, QAOA parameters). No new entities are invented.

free parameters (2)

QAOA angles (p=2) = not reported
Four variational parameters (gamma1, beta1, gamma2, beta2) optimized via COBYLA on a noiseless simulator; the outcome of the hardware comparison depends on these choices, although they are shared by cut and uncut circuits.
Sample budget = 100,000 per method
Chosen by hand; the cut method would need more samples to compensate for the 1/12 suppression bound, but the budget was kept equal for both methods.

assumptions (4)

standard math Integer programming can find balanced vertex separators quickly (SCIP)
Section II c states the use of SCIP for the integer program; this is a standard solver and the NP-hardness of separator finding is cited.
domain assumption Graph shrinking preserves the QUBO solution structure (Ref [2])
Section II b invokes the linear programming approach of Ref [2] to shrink the vertex separator to a single vertex, assuming the reconstruction step yields solutions to the original graph.
domain assumption Wire cutting methods of Peng et al. and Harada et al., combined with the reconstruction from Ref [1], yield a valid sampling distribution for QAOA bitstrings
Section II b describes applying the two cutting schemes and reversing the decomposition to retrieve solutions; this relies on the correctness of these methods for sampling, which is the core premise of the paper.
ad hoc to paper The 1/12 suppression bound from Ref [1] justifies the sample budget
Section II c quotes the bound kappa=1/12 from the authors' prior work; it is used to argue the sample budget is sufficient, though the bound itself is not re-derived here.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimized Circuit Cutting for QAOA Sampling Tasks." pith.science (2026). https://pith.science/paper/PLT6LC4Q

@misc{pith2026250706649,
  author       = {Pith},
  title        = {Pith review of: Optimized Circuit Cutting for QAOA Sampling Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLT6LC4Q}},
  note         = {Machine review of arXiv:2507.06649}
}

read the original abstract

Circuit cutting was originally designed to retrieve the expectation value of an observable with respect to a large quantum circuit by executing smaller circuit fragments. In this work, however, we demonstrate the application of circuit cutting to a pure sampling task. In particular, we sample solutions to an optimization problem from a trained QAOA circuit. Here, circuit cutting leads to a broadening and shift of the bitstring distribution towards suboptimal values compared to the uncut case. To reduce this effect, we minimize the number of required cuts via integer programming methods. On the other hand, cutting reduces the circuit size and thus the impact of noise. Our experiments on quantum hardware reveal that, for large circuits, the effect of noise reduction outweighs the derogative effects on the bitstring distribution. The study therefore provides evidence that circuit cutting combined with optimized cutting schemes can both scale problem size and mitigate noise for near-term quantum optimization.

Figures

Figures reproduced from arXiv: 2507.06649 by the authors.

**Figure 2.** Histograms retrieved from cut (//) and uncut ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** 95th percentiles for the distributions retrieved from [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages

[1]

L. S. Herzog, F. Wagner, C. Ufrecht, L. Palackal, A. Plinge, C. Mutschler, and D. D. Scherer, Improving quantum and classical decomposition methods for vehicle routing, arXiv:2404.05551 (2024)

arXiv 2024
[2]

Wagner, J

F. Wagner, J. N¨ ußlein, and F. Liers, Enhancing quantum algo- rithms for quadratic unconstrained binary optimization via integer programming, ACM Transactions on Quantum Computing 6, 1–28 (2025)

work page 2025
[3]

H. F. Hofmann, How to simulate a universal quantum computer using negative probabilities, J. Phys. A: Math. Theor.42, 275304 (2009)

work page 2009
[4]

Mitarai and K

K. Mitarai and K. Fujii, Constructing a virtual two-qubit gate by sampling single-qubit operations, New J. Phys. 23, 023021 (2021)

work page 2021
[5]

Ufrecht, M

C. Ufrecht, M. Periyasamy, S. Rietsch, D. D. Scherer, A. Plinge, and C. Mutschler, Cutting multi-control quantum gates with ZX calculus, Quantum 7, 1147 (2023)

work page 2023
[6]

Piveteau and D

C. Piveteau and D. Sutter, Circuit knitting with classical commu- nication, IEEE Trans. Inf. Theory 70, 2734 (2024)

work page 2024
[7]

Schmitt, C

L. Schmitt, C. Piveteau, and D. Sutter, Cutting circuits with multiple two-qubit unitaries (2024), arXiv:2312.11638

arXiv 2024
[8]

A. W. Harrow and A. Lowe, Optimal quantum circuit cuts with ap- plication to clustered Hamiltonian simulation, arXiv:2403.01018 (2024)

work page Pith review arXiv 2024

Show all 20 references

[9]

T. Peng, A. W. Harrow, M. Ozols, and X. Wu, Simulating large quantum circuits on a small quantum computer, Phys. Rev. Lett. 125, 150504 (2020)

2020
[10]

A. Lowe, M. Medvidovi ´c, A. Hayes, L. J. O’Riordan, T. R. Bromley, J. M. Arrazola, and N. Killoran, Fast quantum circuit 4 cutting with randomized measurements, Quantum 7, 934 (2023)

2023
[11]

Brenner, C

L. Brenner, C. Piveteau, and D. Sutter, Optimal wire cutting with classical communication (2023), arXiv:2302.03366

2023
[12]

Harada, K

H. Harada, K. Wada, and N. Yamamoto, Doubly optimal parallel wire cutting without ancilla qubits, PRX Quantum 5, 040308 (2024)

2024
[13]

T. N. Bui and C. Jones, Finding good approximate vertex and edge partitions is np-hard, Information Processing Letters 42, 153 (1992)

1992
[14]

H. Y. Althoby, M. Didi Biha, and A. Sesbo¨ u´e, Exact and heuristic methods for the vertex separator problem, Computers & Industrial Engineering 139, 106135 (2020)

2020
[15]

Fischer, M

V. Fischer, M. Passek, F. Wagner, J. R. Finˇ zgar, L. Palackal, and C. B. Mendl, Quantum and classical correlations in shrinking algorithms for optimization (2024), arXiv:2404.17242 [quant- ph]

2024 arXiv
[16]

The QuaST consortium, The QuaST Best Practice Guide - A Comprehensive Summary of Research Results (2025)

2025
[17]

S. H. Sack and M. Serbyn, Quantum annealing initialization of the quantum approximate optimization algorithm, Quantum 5, 491 (2021)

2021
[18]

Bolusani, M

S. Bolusani, M. Besan c ¸on, K. Bestuzheva, A. Chmiela, J. Dion ´ısio, T. Donkiewicz, J. van Doornmalen, L. Eifler, M. Ghannam, A. Gleixner, C. Graczyk, K. Halbig, I. Hedtke, A. Hoen, C. Hojny, R. van der Hulst, D. Kamp, T. Koch, K. Kofler, J. Lentz, J. Manns, G. Mexi, E. M¨ u...

2024
[19]

Javadi-Abhari, M

A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lish- man, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. Johnson, and J. M. Gambetta, Quantum computing with qiskit (2024), arXiv:2405.08810 [quant-ph]

2024 arXiv
[20]

M. J. D. Powell, A direct search optimization method that models the objective and constraint functions by linear interpolation, in Advances in Optimization and Numerical Analysis , edited by S. Gomez and J.-P. Hennart (Springer Netherlands, Dordrecht,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

L. S. Herzog, F. Wagner, C. Ufrecht, L. Palackal, A. Plinge, C. Mutschler, and D. D. Scherer, Improving quantum and classical decomposition methods for vehicle routing, arXiv:2404.05551 (2024)

arXiv 2024

[2] [2]

Wagner, J

F. Wagner, J. N¨ ußlein, and F. Liers, Enhancing quantum algo- rithms for quadratic unconstrained binary optimization via integer programming, ACM Transactions on Quantum Computing 6, 1–28 (2025)

work page 2025

[3] [3]

H. F. Hofmann, How to simulate a universal quantum computer using negative probabilities, J. Phys. A: Math. Theor.42, 275304 (2009)

work page 2009

[4] [4]

Mitarai and K

K. Mitarai and K. Fujii, Constructing a virtual two-qubit gate by sampling single-qubit operations, New J. Phys. 23, 023021 (2021)

work page 2021

[5] [5]

Ufrecht, M

C. Ufrecht, M. Periyasamy, S. Rietsch, D. D. Scherer, A. Plinge, and C. Mutschler, Cutting multi-control quantum gates with ZX calculus, Quantum 7, 1147 (2023)

work page 2023

[6] [6]

Piveteau and D

C. Piveteau and D. Sutter, Circuit knitting with classical commu- nication, IEEE Trans. Inf. Theory 70, 2734 (2024)

work page 2024

[7] [7]

Schmitt, C

L. Schmitt, C. Piveteau, and D. Sutter, Cutting circuits with multiple two-qubit unitaries (2024), arXiv:2312.11638

arXiv 2024

[8] [8]

A. W. Harrow and A. Lowe, Optimal quantum circuit cuts with ap- plication to clustered Hamiltonian simulation, arXiv:2403.01018 (2024)

work page Pith review arXiv 2024