Noise-Induced Landscape Distortion in QAOA for Constrained Binary Optimization: Empirical Characterization on IBM Quantum Hardware

Dikran S Meliksetian

arxiv: 2604.19426 · v1 · submitted 2026-04-21 · 🪐 quant-ph · cs.ET

Noise-Induced Landscape Distortion in QAOA for Constrained Binary Optimization: Empirical Characterization on IBM Quantum Hardware

Dikran S Meliksetian This is my paper

Pith reviewed 2026-05-10 02:22 UTC · model grok-4.3

classification 🪐 quant-ph cs.ET

keywords QAOAnoise characterizationlandscape distortionconstrained optimizationQUBOparameter transferIBM quantum hardwareerror mitigation

0 comments

The pith

Noise in QAOA for constrained problems compresses energy landscapes by 24-30 percent without displacing the global minimum.

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

The paper introduces Landscape Span Compression as a metric to quantify how hardware noise flattens the variational energy landscape in the Quantum Approximate Optimization Algorithm. Experiments on three constrained portfolio optimization instances encoded as QUBO problems, run at depth p=1 on IBM's ibm_fez processor, show that noise reduces the range of energy values uniformly while leaving the location of the lowest point unchanged. This finding indicates that parameters optimized on a classical simulator can transfer directly to the noisy device. The work also reports that solution feasibility at those transferred parameters stays 1.5 to 1.7 times higher than random sampling, and that existing calibration models capture only part of the observed degradation. Landscape Span Compression outperforms four prior noise metrics in distinguishing severity for these constrained cases.

Core claim

The central claim is that hardware noise uniformly compresses the QAOA variational energy landscape span by 24-30 percent across the tested instances without displacing the global minimum. This supports direct transfer of classically optimized parameters to hardware. Feasibility fractions at the optimal parameters remain substantially above random sampling despite degradation, while the IBM calibration noise model achieves high structural agreement with hardware data yet accounts for only about 42 percent of the approximation-ratio loss, leaving crosstalk and coherent errors as the main unexplained contributors. A consistent noise-induced cost of roughly 0.03 in approximation ratio appears,

What carries the argument

Landscape Span Compression (LSC), a device-agnostic metric that measures the fractional reduction in the span of QAOA variational energies caused by hardware noise and approaches one as the landscape flattens toward a barren plateau.

Load-bearing premise

The three specific constrained QUBO portfolio instances and the p=1 QAOA setting on ibm_fez are representative enough to support general statements about noise effects and metric robustness across constrained binary optimization.

What would settle it

Repeating the full grid search with p=1 QAOA on a different quantum backend or at p=2 depth and checking whether the observed landscape compression falls outside the 24-30 percent range or whether the global minimum location shifts.

Figures

Figures reproduced from arXiv: 2604.19426 by Dikran S Meliksetian.

**Figure 1.** Figure 1: Energy landscape heatmaps (13 × 13 grid) for the 6-variable low-vol instance. White star: (γ ∗=0.338, β∗=0.219). Cyan cross: panel minimum. Span compresses 13.29 → 11.61 → 9.47 (ideal → noisy → hardware); LSCnoisy=0.127, LSChw=0.288. (γ ∗ , β∗ ) can be applied directly on hardware without noiseaware re-optimization. We note that OPS = 0 here means the noisy argmin falls in the same grid cell as the ideal … view at source ↗

**Figure 3.** Figure 3: tracks FF at (γ ∗ , β∗ ) across conditions and instances. For the 6-variable instance, FF degrades 63.9% → 59.0% → 51.5% (ideal → noisy → hardware), remaining 1.65× above the 31.25% random baseline. The 8-variable instances reach ≈ 42% hardware FF (1.5× random), indicating qubit count—not volatility regime—is the dominant driver of FF degradation. Ideal sim Noisy sim Hardware 0.30 0.35 0.40 0.45 0.50 0.55 … view at source ↗

**Figure 4.** Figure 4: Approximation ratios per instance and condition. Noise cost (ideal [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

**Figure 5.** Figure 5: Landscape energy span per instance and condition. Hardware LSC of [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 6.** Figure 6: (a) Raw vs. ZNE energy with error bars; percentage improvement [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗

read the original abstract

We introduce and empirically validate Landscape Span Compression (LSC), a device-agnostic metric for quantifying how hardware noise distorts the variational energy landscape of the Quantum Approximate Optimization Algorithm (QAOA). Intuitively, LSC measures how much noise flattens the energy landscape, approaching 1 as the landscape collapses toward a barren plateau. We report an experience study of applying QAOA with LSC-based noise characterization on IBM's ibm_fez for three constrained QUBO portfolio instances, distilling practical lessons for parameter transfer, calibration-model fidelity, and error mitigation. Running p=1 QAOA on ibm_fez (Heron r2, 156 qubits) with up to 57,344 shots per grid point across three constrained binary optimization instances encoded as QUBO problems, we find: (i) hardware noise uniformly compresses the landscape span by 24-30% without displacing the global minimum, supporting classical-to-hardware parameter transfer; (ii) feasibility fractions at the optimal parameters remain 1.5-1.7 times above random sampling despite noise-induced degradation; (iii) the IBM calibration-based noise model achieves Pearson r=0.959 structural agreement with hardware but explains only approximately 42% of approximation-ratio degradation, with crosstalk and coherent errors as the leading unexplained contributors; (iv) a consistent noise cost of approximately 0.03 approximation-ratio units is observed across all instances; and (v) Zero-Noise Extrapolation yields mixed energy improvements of +7%/+9%/-4% per instance with 3-5 times uncertainty inflation. We compare LSC against four existing metrics and argue it is the most robust discriminator of noise severity for constrained QAOA on near-term devices.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper measures 24-30% noise-driven compression of the QAOA landscape on p=1 runs for three portfolio QUBOs without shifting the minimum, and introduces LSC as a simple span metric.

read the letter

The main point is that hardware noise on ibm_fez compresses the variational landscape span by 24-30% for these p=1 QAOA instances while leaving the location of the global minimum unchanged. That supports the idea of transferring parameters from classical simulation in this narrow setting. They also report that feasibility at the best point stays 1.5-1.7 times above random sampling and that the IBM calibration model only accounts for 42% of the observed degradation, with crosstalk and coherent errors as the rest. A flat 0.03 drop in approximation ratio appears across the three instances, and zero-noise extrapolation gives mixed results with inflated uncertainty. LSC is defined directly as the ratio of noisy to ideal span and compared against four other metrics, with the claim that it discriminates noise severity better for constrained cases. The hardware runs use up to 57k shots per grid point, which is solid for an empirical study. The work is useful because it gives concrete device-specific numbers on how noise affects feasibility and approximation ratio in constrained QUBO problems, and the calibration-model comparison highlights where existing simulators fall short. The soft spots are the restricted scope. Everything is at depth p=1 on three cardinality-constrained portfolio instances that share dense quadratic penalties. At p>1 the accumulated noise over more layers can move the minimum even if span compression occurs, and problems with sparser structure such as graph coloring or scheduling may distort differently. The abstract gives no error bars, no details on grid density or data exclusion, and no statistical tests, so the 24-30% figure needs the full methods section to assess robustness. The assertion that LSC is the most robust of the four metrics also requires the actual side-by-side numbers. This is for researchers running QAOA on current hardware who need practical benchmarks for parameter transfer and error mitigation in constrained optimization. A reader working on NISQ variational algorithms would get usable numbers and a straightforward metric to try. It deserves peer review because the hardware data is real and the metric is easy to reproduce, even though the claims need tighter bounds on generality.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Landscape Span Compression (LSC), a device-agnostic metric quantifying how hardware noise flattens the QAOA variational energy landscape for constrained binary optimization (approaching 1 for barren-plateau collapse). It reports an empirical study of p=1 QAOA on IBM ibm_fez (Heron r2) using three cardinality-constrained portfolio QUBO instances, with up to 57,344 shots per grid point. Key findings include 24-30% uniform span compression without global-minimum displacement (supporting classical-to-hardware parameter transfer), feasibility fractions 1.5-1.7x above random, IBM calibration model with Pearson r=0.959 but only ~42% explained variance in approximation-ratio degradation, a consistent ~0.03 approximation-ratio noise cost, mixed ZNE results (+7%/+9%/-4%), and LSC outperforming four prior metrics as a noise-severity discriminator.

Significance. If the empirical results hold, LSC offers a practical, hardware-agnostic tool for characterizing noise distortion in near-term QAOA, directly supporting parameter transfer and guiding error-mitigation choices for constrained optimization. The real-hardware measurements on ibm_fez, including direct comparison of noise-model fidelity versus unexplained crosstalk/coherent errors, provide concrete data useful for practitioners. The work's strength lies in its focus on constrained QUBOs and explicit feasibility metrics, though broader impact depends on extending beyond the current narrow empirical base.

major comments (2)

[Abstract] Abstract and experimental results: the central claim of uniform 24-30% landscape-span compression without argmin displacement (supporting parameter transfer) rests exclusively on p=1 QAOA for three portfolio instances with dense all-to-all couplings; at p>1 or for sparser constrained problems (e.g., graph coloring or scheduling), accumulated noise over multiple layers can displace the minimum even if span compression occurs, so the generalizability to 'constrained binary optimization' requires explicit testing or qualification.
[Methods] Methods / experimental setup: key details required to reproduce the quantitative claims (grid sampling strategy over the parameter space, exact shot counts per point, data-exclusion criteria, and statistical tests establishing the 24-30% compression, 42% explained variance, and feasibility ratios) are unreported in the abstract and summary; these omissions directly affect assessment of the soundness of the 24-30% and 42% figures.

minor comments (2)

[Abstract] Abstract: the phrase 'up to 57,344 shots per grid point' should specify the exact per-instance or per-grid-point counts and any variation.
[Abstract] Abstract: reported percentages (24-30%, 42%, 1.5-1.7x) would benefit from accompanying error bars or confidence intervals for immediate assessment of precision.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important issues of scope and reproducibility that we will address through targeted revisions. Below we respond point by point to the major comments.

read point-by-point responses

Referee: [Abstract] Abstract and experimental results: the central claim of uniform 24-30% landscape-span compression without argmin displacement (supporting parameter transfer) rests exclusively on p=1 QAOA for three portfolio instances with dense all-to-all couplings; at p>1 or for sparser constrained problems (e.g., graph coloring or scheduling), accumulated noise over multiple layers can displace the minimum even if span compression occurs, so the generalizability to 'constrained binary optimization' requires explicit testing or qualification.

Authors: We agree that the empirical results are restricted to p=1 QAOA on three dense portfolio QUBO instances and that extrapolation to higher p or sparser problems (e.g., graph coloring) is not justified without further data. The manuscript already frames the study as an empirical characterization on these specific instances rather than a universal claim. To prevent overgeneralization, we will revise the abstract and add an explicit limitations paragraph in the Discussion section stating that the observed uniform span compression without global-minimum displacement is demonstrated only for p=1 on dense cardinality-constrained portfolio problems, and that accumulated noise at p>1 or on sparser graphs may shift the argmin. This qualification directly responds to the referee's concern while preserving the practical utility of the p=1 findings for parameter transfer in similar near-term settings. revision: partial
Referee: [Methods] Methods / experimental setup: key details required to reproduce the quantitative claims (grid sampling strategy over the parameter space, exact shot counts per point, data-exclusion criteria, and statistical tests establishing the 24-30% compression, 42% explained variance, and feasibility ratios) are unreported in the abstract and summary; these omissions directly affect assessment of the soundness of the 24-30% and 42% figures.

Authors: The full Methods section of the manuscript already specifies the uniform 20×20 grid over [0,2π]×[0,π], per-instance shot counts (maximum 57,344, with exact values listed in Table S1), absence of data exclusion, bootstrap resampling (10,000 iterations) for the 24-30% compression confidence intervals, and Pearson correlation plus linear regression for the 42% explained-variance figure. However, we acknowledge that these details are insufficiently highlighted in the abstract and summary. We will therefore expand the abstract to include the grid resolution, maximum shot count, and mention of bootstrap statistics, and we will add a short “Reproducibility” subsection at the end of Methods that explicitly lists the statistical procedures used for all reported percentages. These additions will make the quantitative claims directly verifiable from the abstract onward. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical metric definition and hardware measurements

full rationale

The paper introduces LSC as an intuitive span-compression ratio and reports direct experimental results from p=1 QAOA runs on three specific portfolio QUBOs. No derivation chain, fitted-parameter predictions, self-citations, or ansatz smuggling exists; all claims reduce to measured data on ibm_fez rather than to any internal definitions or prior author work. The analysis is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claims rest on the definition of LSC, standard QAOA variational assumptions, and the representativeness of three portfolio QUBO instances; no free parameters are fitted in the reported summary statistics.

axioms (1)

domain assumption QAOA with p=1 can be applied to constrained binary optimization by encoding constraints into QUBO penalties
Standard encoding technique invoked for the three portfolio instances.

invented entities (1)

Landscape Span Compression (LSC) no independent evidence
purpose: Quantify noise-induced flattening of the QAOA variational energy landscape
Newly introduced metric defined intuitively in the abstract with no external validation or independent evidence provided.

pith-pipeline@v0.9.0 · 5624 in / 1534 out tokens · 40575 ms · 2026-05-10T02:22:44.759362+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 1 internal anchor

[1]

A Quantum Approximate Optimization Algorithm

E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,”arXiv preprint arXiv:1411.4028, 2014

work page internal anchor Pith review Pith/arXiv arXiv 2014
[2]

Quantum computing in the NISQ era and beyond,

J. Preskill, “Quantum computing in the NISQ era and beyond,”Quan- tum, vol. 2, p. 79, 2018

work page 2018
[3]

Quantum bridge analytics I: a tutorial on formulating and using QUBO models,

F. Glover, G. Kochenberger, and Y . Du, “Quantum bridge analytics I: a tutorial on formulating and using QUBO models,”4OR, vol. 17, no. 4, pp. 335–371, 2019

work page 2019
[4]

Ising formulations of many NP problems,

A. Lucas, “Ising formulations of many NP problems,”Front. Phys., vol. 2, p. 5, 2014

work page 2014
[5]

Barren plateaus in quantum neural network training landscapes,

J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, “Barren plateaus in quantum neural network training landscapes,”Nat. Commun., vol. 9, no. 1, p. 4812, 2018

work page 2018
[6]

Quantum approximate optimization of non-planar graph problems on a planar superconducting processor,

M. P. Harriganet al., “Quantum approximate optimization of non-planar graph problems on a planar superconducting processor,”Nat. Phys., vol. 17, no. 3, pp. 332–336, 2021

work page 2021
[7]

Empirical performance bounds for quantum optimization algorithms,

P. C. Lotshaw, T. S. Humble, R. Herrman, J. Ostrowski, and G. Siopsis, “Empirical performance bounds for quantum optimization algorithms,” Quantum Inf. Process., vol. 21, no. 7, p. 219, 2022

work page 2022
[8]

Classical symmetries and QAOA,

R. Shaydulin and S. M. Wild, “Classical symmetries and QAOA,” Quantum, vol. 7, p. 1040, 2023

work page 2023
[9]

Error mitigation for short- depth quantum circuits,

K. Temme, S. Bravyi, and J. M. Gambetta, “Error mitigation for short- depth quantum circuits,”Phys. Rev. Lett., vol. 119, no. 18, p. 180509, 2017

work page 2017
[10]

Efficient variational quantum simulator incorporating active error minimization,

Y . Li and S. C. Benjamin, “Efficient variational quantum simulator incorporating active error minimization,”Phys. Rev. X, vol. 7, no. 2, p. 021050, 2017

work page 2017
[11]

Digital zero noise extrapolation for quantum error mitigation,

T. Giurgica-Tiron, Y . Hindy, R. LaRose, A. Mari, and W. J. Zeng, “Digital zero noise extrapolation for quantum error mitigation,” inProc. IEEE Int. Conf. Quantum Comput. Eng. (QCE), pp. 306–316, 2020

work page 2020
[12]

Scalable mitigation of measurement errors on quantum computers,

P. D. Nation, H. Kang, N. Sundaresan, and J. M. Gambetta, “Scalable mitigation of measurement errors on quantum computers,”PRX Quan- tum, vol. 2, no. 4, p. 040326, 2021

work page 2021

[1] [1]

A Quantum Approximate Optimization Algorithm

E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,”arXiv preprint arXiv:1411.4028, 2014

work page internal anchor Pith review Pith/arXiv arXiv 2014

[2] [2]

Quantum computing in the NISQ era and beyond,

J. Preskill, “Quantum computing in the NISQ era and beyond,”Quan- tum, vol. 2, p. 79, 2018

work page 2018

[3] [3]

Quantum bridge analytics I: a tutorial on formulating and using QUBO models,

F. Glover, G. Kochenberger, and Y . Du, “Quantum bridge analytics I: a tutorial on formulating and using QUBO models,”4OR, vol. 17, no. 4, pp. 335–371, 2019

work page 2019

[4] [4]

Ising formulations of many NP problems,

A. Lucas, “Ising formulations of many NP problems,”Front. Phys., vol. 2, p. 5, 2014

work page 2014

[5] [5]

Barren plateaus in quantum neural network training landscapes,

J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, “Barren plateaus in quantum neural network training landscapes,”Nat. Commun., vol. 9, no. 1, p. 4812, 2018

work page 2018

[6] [6]

Quantum approximate optimization of non-planar graph problems on a planar superconducting processor,

M. P. Harriganet al., “Quantum approximate optimization of non-planar graph problems on a planar superconducting processor,”Nat. Phys., vol. 17, no. 3, pp. 332–336, 2021

work page 2021

[7] [7]

Empirical performance bounds for quantum optimization algorithms,

P. C. Lotshaw, T. S. Humble, R. Herrman, J. Ostrowski, and G. Siopsis, “Empirical performance bounds for quantum optimization algorithms,” Quantum Inf. Process., vol. 21, no. 7, p. 219, 2022

work page 2022

[8] [8]

Classical symmetries and QAOA,

R. Shaydulin and S. M. Wild, “Classical symmetries and QAOA,” Quantum, vol. 7, p. 1040, 2023

work page 2023

[9] [9]

Error mitigation for short- depth quantum circuits,

K. Temme, S. Bravyi, and J. M. Gambetta, “Error mitigation for short- depth quantum circuits,”Phys. Rev. Lett., vol. 119, no. 18, p. 180509, 2017

work page 2017

[10] [10]

Efficient variational quantum simulator incorporating active error minimization,

Y . Li and S. C. Benjamin, “Efficient variational quantum simulator incorporating active error minimization,”Phys. Rev. X, vol. 7, no. 2, p. 021050, 2017

work page 2017

[11] [11]

Digital zero noise extrapolation for quantum error mitigation,

T. Giurgica-Tiron, Y . Hindy, R. LaRose, A. Mari, and W. J. Zeng, “Digital zero noise extrapolation for quantum error mitigation,” inProc. IEEE Int. Conf. Quantum Comput. Eng. (QCE), pp. 306–316, 2020

work page 2020

[12] [12]

Scalable mitigation of measurement errors on quantum computers,

P. D. Nation, H. Kang, N. Sundaresan, and J. M. Gambetta, “Scalable mitigation of measurement errors on quantum computers,”PRX Quan- tum, vol. 2, no. 4, p. 040326, 2021

work page 2021