Quantum Annealing Hyperparameter Analysis for Optimal Sensor Placement in Production Environments

Alexander Kuzmany; Daniel Porawski; Jonas Stein; Marvin Erdmann; Nico Kraus

REVIEW 3 major objections 5 minor 24 references

Quantum Annealing Hyperparameter Analysis for Optimal Sensor Placement in Production Environments

T0 review · 3 major / 5 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 Quantum annealing with a binary QUBO encoding, a low penalty weight, reduced chain strength, and spectral-clustering decomposition produces better sensor placements than default settings and scales to a 4,109-qubit instance.

desk verdict An honest, well-run QA study for a real set-cover problem, but the headline tuning gain is in-sample and needs validation. read the letter →

arxiv 2507.16584 v1 pith:TMGRSNTA submitted 2025-07-22 cs.ET quant-ph

Nico Kraus , Marvin Erdmann , Alexander Kuzmany , Daniel Porawski , Jonas Stein This is my paper

classification cs.ETquant-ph

keywords quantumannealingQUBOsetcoverproblemsensorplacementhyperparameteroptimizationspectralclusteringslackvariableencodingD-Wave

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

The paper asks whether a production-style sensor-placement problem can be made tractable on today's quantum annealers by careful parameter selection rather than by waiting for better hardware. It encodes the laser-sensor placement task, a set-cover problem, as a QUBO and runs it on a D-Wave annealer across eleven toy and twelve real-world-inspired instances. Its central claim is that a binary slack encoding combined with a low penalty weight ($\alpha = 0.2$), a reduced chain strength factor ($\beta = 0.5$), and an annealing time of $500\,\mu$s consistently beats the device's default hyperparameters, and that spectral-clustering decomposition into two or four subproblems extends the approach to a 4,109-qubit instance. The paper itself concedes that classical solvers still outperform quantum annealing; the contribution is a parameter recipe and a decomposition strategy that bring quantum annealing closer to industrial instances.

What carries the argument

The central object is the QUBO Hamiltonian $H = \sum_{l \in V_l} x_l + \alpha \sum_{s \in V_s} \bigl(\sum_{l \in N_s} x_l - y_s - 1\bigr)^2$, where $x_l$ marks a LiDAR position and the integer slack $y_s$ counts redundant coverage of street point $s$. The paper's key move is to encode $y_s$ in binary, using $\lceil \log_2 |N_s|\rceil$ qubits, rather than one-hot, and then to tune the penalty weight $\alpha$, the chain strength factor $\beta$ multiplying D-Wave's default chain strength, and the annealing time. A penalty weight below 1 is counterintuitive because it lets infeasible solutions compete, but the experiments show it outperforms stronger penalties, apparently because large coupler weights degrade annealer performance. Spectral clustering is the decomposition mechanism: it partitions the bipartite sensor-street graph into two or four clusters, each embedded and annealed separately, and the union of feasible subproblem solutions is always feasible for the original problem, though not necessarily optimal.

What would settle it

Re-run the decomposed Real 7 through Real 12 instances with a small sweep around the chosen parameters, for example $\alpha \in \{0.1, 0.2, 0.5, 1.0\}$ and $\beta \in \{0.2, 0.5, 1.0\}$, keeping the same embeddings and 1000 samples. If any alternative configuration reaches the relaxed optimum in fewer samples, or if $\alpha = 1.0$ recovers a feasible optimum where $\alpha = 0.2$ does not, then the reported quality for those large instances is an artifact of parameter transfer rather than the annealer's capability.

Watch

Extended reading notes

Core claim

The paper's discovery is that configuration, not just formulation, decides whether a quantum annealer is useful for this problem. On instances from toy layouts up to a real-world-inspired factory setting, the optimized configuration (binary encoding, $\alpha = 0.2$, $\beta = 0.5$, annealing time $500\,\mu$s) finds better feasible solutions than D-Wave's defaults, including optimal solutions for Toy 7 and Real 1 that the defaults miss. The mechanism is that low penalty weights ease the coupling burden on the annealer, and reduced chain strength prevents high-weight interference. The paper further shows that spectral clustering into two subproblems improves solution quality in nearly all scenarios, while four-way decomposition lets the largest instance, Real 12 with 4,109 qubits, be embedded and solved, though not optimally. The claim is not that quantum annealing wins outright; it is that tuned parameters and decomposition make current hardware capable of feasible, and for small instances optimal, sensor placements.

Load-bearing premise

The load-bearing assumption is that hyperparameters picked on small and medium instances remain the right ones for the larger, decomposed subproblems, since the subproblems are comparable in size to the smaller instances.

Editorial extensions

If this is right

If the tuned recipe holds, D-Wave users should treat default hyperparameters as a weak baseline, not a sensible operating point, for QUBO set-cover problems.
Spectral clustering into two subproblems is a viable route to better feasible solutions on real-world-inspired instances, and four-way splitting helps when the composed problem cannot be embedded.
Weakening the penalty below the objective-cost threshold can improve annealer output, so feasibility and solution quality must be tracked separately in QUBO benchmarking.
Current quantum annealing can produce feasible, and on small instances optimal, sensor placements, but classical solvers still dominate; the practical payoff depends on hardware maturity and on avoiding the slack-variable overhead.
The same hyperparameter methodology can be transferred to other constrained QUBO problems whose constraint penalties and chain strengths are otherwise left at defaults.

Reading between the lines

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

Inference beyond the paper: since $\alpha = 0.2$ can make the smallest instance infeasible, the one-size-fits-all penalty likely leaves performance on small problems on the table; a size-aware penalty schedule would be a direct next experiment.
Inference beyond the paper: the four-subproblem decomposition outperforms the composed formulation mainly on the largest instances, which suggests that a hybrid scheme, decomposing only when embedding fails, may dominate any fixed decomposition count.
Inference beyond the paper: the relaxed-optimum metric, which counts samples to a solution within 10% of the optimum, is a practical way to benchmark quantum optimization on industrial instances where true optima are unknown; it could be paired with confidence intervals on the gap rather than only on sample counts.
Inference beyond the paper: because a sensor from one subproblem could cover street points in a neighboring area, the decomposition bound is loose; a post-processing step that checks and repairs boundary coverage could push the composed solution closer to the global optimum.

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

Colleague,

Read the Kraus et al. paper on quantum annealing for sensor placement. The headline: it's a solid, honest empirical study, but the central result—that tuned hyperparameters beat D-Wave defaults—is measured on the same instances used to pick those parameters. That doesn't kill the paper, but it means the improvement numbers are selection artifacts until shown otherwise.

What's genuinely useful: they formulate a set-cover-derived QUBO for LiDAR placement, compare binary vs one-hot slack encodings, and report a clean sweep over penalty weight, chain strength, and annealing time. The metrics are thought through (s_p and a relaxed version), and they give bootstrap confidence intervals. They also test spectral clustering decomposition, pushing to a 4109-qubit instance. Credit where due: the paper never oversells. It states plainly that Gurobi outperforms QA and that industrial instances are an order of magnitude larger. That honesty is real.

The soft spots are centered on validation. The hyperparameters were selected by iterative one-at-a-time sweeps averaged over the Toy and Real categories, then evaluated on exactly those same instances in Section IV-C. Section III-B warns about overfitting in choosing global parameters, but no held-out instances or cross-validation follow. So Fig. 6's 'significant improvement' is in-sample. The transfer of those parameters to the decomposed subproblems is also just assumed, not tested. There's also no code or instance data released, which makes independent checking harder. And the classical baseline is Gurobi on the small instances; we never see how the heuristics actually used in production would do, though that's a minor gap given the paper's scope.

That said, the core message—that parametrization matters and decomposition helps scale—is likely true in general, and the paper is transparent about its limitations. It's a good case study for people working with D-Wave on realistic optimization problems.

My recommendation: send it to peer review. The experiments are reproducible in principle and the claims are appropriately hedged, but I'd ask the authors to add an out-of-sample validation (e.g., tune on half the instances, test on the other half) or to explicitly frame the results as indicative rather than measured gains. Also request data release.

Referee Report

3 major / 5 minor

Summary. The paper formulates the LiDAR sensor placement problem in automotive production as a set-cover QUBO and studies it on D-Wave quantum annealing hardware. It compares one-hot and binary slack-variable encodings, tunes the penalty weight, chain-strength factor, and annealing time by an iterative one-parameter-at-a-time search, and then compares the tuned configuration against D-Wave defaults on toy and real-world-inspired instances up to 879 qubits. For larger instances, spectral clustering decomposes the problem into two or four subproblems, allowing instances with up to 4109 qubits to be embedded. The main claims are that the tuned configuration (binary encoding, α=0.2, β=0.5, 500 µs) improves solution quality over defaults and that decomposition into two subproblems improves results in most cases, while classical solvers remain superior.

Significance. If the central claims hold, the paper would be a useful empirical contribution: it applies quantum annealing to a concrete industrial set-cover variant, uses explicit performance metrics, reports bootstrap confidence intervals, averages over multiple embeddings, and honestly acknowledges that classical solvers outperform the quantum approach. The decomposition experiments also provide a practical scaling strategy for current hardware. The main weakness is that the headline quantitative comparison is in-sample: the hyperparameters are selected on the same instances used for the final default-versus-optimized comparison, so the reported improvement is not a validated out-of-sample estimate. The paper is nevertheless clearly written and reproducible in spirit, with the optimization metrics and experimental protocol described in enough detail to be audited.

major comments (3)

[§III-B and §IV-C] The hyperparameter selection and the final default-versus-optimized comparison use exactly the same problem instances. Section III-B describes an iterative one-parameter-at-a-time search over roughly ten values per parameter, averaging over the Toy and Real categories, and Section IV-C then reports improvements on Toy 1–11 and Real 1–6 on those identical instances. With noisy annealer samples and ten embeddings per instance, selecting the best of many parameter values on the evaluation set systematically inflates the apparent gain. The paper acknowledges the overfitting risk in qualitative terms but provides no held-out instances, no cross-validation, and no selection-aware estimate. This is the load-bearing evidence for the conclusion that tuned parameters improve solution quality, so the authors should either provide an out-of-sample evaluation or explicitly reframe the claim as an in-sample demonstration of tunability rather than a validated improvement.
[§II-B and §IV-B] For the real-world-inspired instances, the unrelaxed success metric is capped: Table I reports 'Samples Real 5000+' for both encodings, and several panels in Figs. 3–5 show flat plateaus at the 5000 cap for the Real category. This means that for many configurations popt is exactly zero over 1000 samples and the reported sp is an artificial upper bound. Choosing β=0.5 and an annealing time of 500 µs based on such capped values may amount to selecting on noise. The authors should report raw success counts or uncapped popt values, and should quantify whether the differences among β and annealing-time settings are statistically significant, for example by confidence intervals on popt rather than on the capped sp.
[§IV-D] The decomposition experiments reuse the hyperparameters tuned on Toy 1–11 and Real 1–6 on the assumption that the decomposed subproblems are 'comparable in size' to the smaller instances. This transfer assumption is not validated: subproblems produced by spectral clustering have different boundary structure, different sensor-to-street-point ratios, and different average degrees than the original Toy and Real instances, so the optimal penalty weight and chain strength may shift. Additionally, the comparison of decomposed solutions against the global optimum found by Gurobi conflates two effects: the quality of the QA solutions within each subproblem and the suboptimality introduced by the decomposition boundary (a sensor in one subproblem can cover street points in a neighboring subproblem). The authors should report per-subproblem optimality gaps, or at least validate the parameter-transfer assumption on held-out instances, before claiming that decomposition improves solution quality.

minor comments (5)

[§II-A, Eq. (7)] The indexing in Eq. (7) uses 'N_{v_i}' while the rest of the section uses N_s; this should be made consistent to avoid confusion about which street point the slack variables belong to.
[§II-B] The quantity sP(X) in Eq. (10) is used before it is defined in the surrounding text. Please define it explicitly or align the notation with reference [11].
[§IV-B, Table I] The entries '5000+' in Table I are not numerical expected sample counts but a capped value. Please state the cap explicitly in the table caption or in the text, and give the raw number of successes whenever the cap is reached.
[§IV-C, Fig. 6] The y-axis label 'Gap to best objective (0 is optimal)' is not fully specified. Please state whether the gap is an absolute number of sensors or a relative gap, and explain how it is computed for cases where no feasible solution is found.
[§II-B] The sentence defining the relaxed optimum says 'within 10% of the optimal value' but does not specify whether this is absolute or relative; for consistency with the figures, please clarify the definition.

Circularity Check

1 steps flagged · score 6.0 of 10

The central claim that tuned QA hyperparameters (binary encoding, α=0.2, β=0.5, 500 µs) improve solution quality over D-Wave defaults is an in-sample comparison: the same Toy/Real instances used in Section IV-B to select the parameters are the instances on which Section IV-C reports the improvement, so the gain is a selection-biased in-sample estimate rather than a validated, transferable result.

fitted input called prediction [Section IV-B (Hyperparameter Analysis) → Section IV-C (Comparing Performances); echoed in Section V (Conclusion)]
"For the hyperparameter analysis, we produced 1000 samples per problem instance and parameter configuration, using ten different, precalculated embeddings for the D-Wave hardware, and calculated the average over all problem instances of the respective category (Toy or Real). ..."

The optimized configuration (binary encoding, α=0.2, β=0.5, 500 µs) is selected by an iterative one-parameter-at-a-time sweep in Section IV-B that averages the evaluation metrics (expected samples to optimal / near-optimal) over the same Toy 1–11 and Real 1–6 instances used for the final comparison.

full rationale

The QUBO formulation and metrics in Section II are standard and externally anchored: the encoding follows Lucas [10], and the popt/tts metrics cited to the authors' own [11] are also supported by the external time-to-target reference [12], so the self-citations are not load-bearing; QUARK [22] is only mentioned as future work. The genuine circularity is the hyperparameter claim. The sweeps for encoding, penalty weight α, chain-strength factor β, and annealing time in Section IV-B are evaluated by averaging expected-sample metrics over the Toy and Real categories, and Section IV-C reports the default-versus-optimized improvement on the very same Toy 1–11 and Real 1–6 instances; the Conclusion then asserts that the tuned combination 'demonstrates the potential to improve the solution quality.' This is the fitted-input-called-prediction pattern: the parameters are fit on the evaluation set, so the reported gain is an in-sample estimate, statistically inflated by selection, and cannot support the transferability the conclusion implies. The paper acknowledges the overfitting risk in Section II-C but performs no held-out validation. The decomposition experiment (Section IV-D) reuses the tuned parameters for larger subproblems by arguing the subproblems are 'comparable in size' to the tuning instances; that is an unvalidated transfer assumption rather than a circular reduction, and the results on Real 7–12 (up to 4109 qubits) are the paper's only genuinely out-of-sample content. Overall: partial circularity — the headline tuning-benefit claim reduces to in-sample selection, while the scaling and decomposition findings retain independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on standard penalty-method QUBO encoding, on Gurobi as ground truth, on the sample-count metric with its 5000 cap, on the representativeness of the synthetic instances, and crucially on the unvalidated transfer of hyperparameters from small training instances to larger decomposed problems. No new physical entities are introduced.

free parameters (5)

penalty weight α = 0.2
Chosen from a sweep of values 0.1 to 20 on the Toy and Real instances (Fig. 3); the best performance was at α=0.1 and 0.2, and 0.2 was selected.
chain strength factor β = 0.5
Chosen from a sweep of factors 0.05 to 10 (Fig. 4); β between 0.2 and 0.5 performed best, and 0.5 was selected.
annealing time = 500 µs
Chosen from a sweep of 2 to 1000 µs (Fig. 5); little dependence was observed, and 500 µs was selected.
QUBO encoding = binary encoding
Binary encoding was selected because it outperformed one-hot encoding in sample counts (Table I).
number of decomposition subproblems = 2
Selected based on results shown in Fig. 7; decomposition into 2 subproblems improved results, while 4 only helped the largest instances.

assumptions (5)

standard math The QUBO objective H = Σ x_l + α Σ_s (Σ_{N_s} x_l - y_s - 1)^2 with slack variables correctly encodes the set cover constraints when minimized.
Basis of the penalty method; standard in QUBO literature (Lucas 2014), but the paper shows it can fail for α<1 (Toy 1).
domain assumption Gurobi finds the true global optimum for the considered instances, so popt and gap metrics are reliable.
Used in Section IV-A to compute optimal solutions and gaps; for small NP-hard instances this is reasonable but unverified for the largest instances.
domain assumption The sample count sp derived from popt accurately ranks solver quality, and the sp=5000 cap for zero-found cases is unbiased.
Equations 9-11 in Section II-B; the cap is an ad hoc convention.
domain assumption The synthetic Toy and Real problem instances are representative of real production environments.
Section III-A; the real-world instances are inspired by a BMW production scenario, but the authors note densities and sizes are lower than in practice.
ad hoc to paper Hyperparameters tuned on Toy 1-11 and Real 1-6 transfer to the larger decomposed subproblems of Real 7-12.
Section IV-D states it is 'sufficient to reuse the previously determined parameters' because subproblems are comparable in size; this transfer is assumed, not validated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Quantum Annealing Hyperparameter Analysis for Optimal Sensor Placement in Production Environments." pith.science (2026). https://pith.science/paper/TMGRSNTA

@misc{pith2026250716584,
  author       = {Pith},
  title        = {Pith review of: Quantum Annealing Hyperparameter Analysis for Optimal Sensor Placement in Production Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TMGRSNTA}},
  note         = {Machine review of arXiv:2507.16584}
}

read the original abstract

To increase efficiency in automotive manufacturing, newly produced vehicles can move autonomously from the production line to the distribution area. This requires an optimal placement of sensors to ensure full coverage while minimizing the number of sensors used. The underlying optimization problem poses a computational challenge due to its large-scale nature. Currently, classical solvers rely on heuristics, often yielding non-optimal solutions for large instances, resulting in suboptimal sensor distributions and increased operational costs. We explore quantum computing methods that may outperform classical heuristics in the future. We implemented quantum annealing with D-Wave, transforming the problem into a quadratic unconstrained binary optimization formulation with one-hot and binary encoding. Hyperparameters like the penalty terms and the annealing time are optimized and the results are compared with default parameter settings. Our results demonstrate that quantum annealing is capable of solving instances derived from real-world scenarios. Through the use of decomposition techniques, we are able to scale the problem size further, bringing it closer to practical, industrial applicability. Through this work, we provide key insights into the importance of quantum annealing parametrization, demonstrating how quantum computing could contribute to cost-efficient, large-scale optimization problems once the hardware matures.

Figures

Figures reproduced from arXiv: 2507.16584 by the authors.

**Figure 2.** Exemplary problem instances. B. Hyperparameter Optimization Numerous approaches for hyperparameter optimization have been documented in the literature [17], [18]. Our experiments with automatic hyperparameter optimization tools revealed suboptimal performance, primarily due to significant noise in the results. In contrast, experiments with manually selected hyperparameters tend to produce more meaningful results, as… view at source ↗

**Figure 5.** Annealing times and expected number of samples for [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Chain strength factors β and expected number of samples for 99% probability to include the optimum (a) and relaxed optimum (b), i.e. within 10% of the optimal. Error bars represent 95 % confidence intervals estimated via bootstrapping. 3) Chain Strength Factor: To find…

**Figure 7.** Figure 7: Comparison of optimized parameter setups without [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 13 canonical work pages

[1]

Industry quantum computing applications,

A. Bayerstadler, G. Becquin, J. Binder, et al., “Industry quantum computing applications,” EPJ Quantum Tech- nology, vol. 8, no. 1, p. 25, Nov. 2021,ISSN : 2196-0763. DOI: 10.1140/epjqt/s40507-021-00114-x

doi:10.1140/epjqt/s40507-021-00114-x 2021
[2]

[Online]

BMW Group, High-tech in production: BMW Group enables automated driving for new vehicles , 2024. [Online]. Available: https : / / www . press . bmwgroup . com/global/article/detail/T0446493EN/high- tech- in- production:- bmw- group- enables- automated- driving- for-new-vehicles?language=en (visited on 03/2025)

work page 2024
[3]

Algorithms for quantum computation: Discrete logarithms and factoring,

P. Shor, “Algorithms for quantum computation: Discrete logarithms and factoring,” in Proceedings 35th Annual Symposium on Foundations of Computer Science, 1994, pp. 124–134. DOI: 10.1109/SFCS.1994.365700

arXiv 1994
[4]

Quantum annealing in the transverse ising model,

T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse ising model,” Physical Review E, vol. 58, no. 5, pp. 5355–5363, Nov. 1998, ISSN : 1095-3787. DOI: 10.1103/physreve.58.5355

doi:10.1103/physreve.58.5355 1998
[5]

H. M. Bauza and D. A. Lidar, Scaling advantage in approximate optimization with quantum annealing ,

work page
[6]

Increasing safety of automated driving by infrastructure-based sensors,

T. D. Borba, O. Vacul ´ın, H. Marzbani, and R. N. Jazar, “Increasing safety of automated driving by infrastructure-based sensors,” IEEE Access , vol. 11, pp. 94 974–94 991, 2023. DOI: 10.1109/ACCESS.2023. 3311136

doi:10.1109/access.2023 2023
[7]

On the optimal placement of cameras for surveillance and the underlying set cover problem,

J. Kritter, M. Br ´evilliers, J. Lepagnot, and L. Idoumghar, “On the optimal placement of cameras for surveillance and the underlying set cover problem,” Applied Soft Computing , vol. 74, pp. 133–153, 2019, ISSN : 1568-4946. DOI: https://doi.org/10.1016/j.asoc. 2018.10.025

doi:10.1016/j.asoc 2019
[8]

Theory versus practice in annealing- based quantum computing,

C. C. McGeoch, “Theory versus practice in annealing- based quantum computing,” Theoretical Computer Sci- ence, vol. 816, pp. 169–183, 2020, ISSN : 0304-3975. DOI: https://doi.org/10.1016/j.tcs.2020.01.024

work page doi:10.1016/j.tcs.2020.01.024 2020

Show all 24 references

[9]

Ushijima-Mwesigwa, C

H. Ushijima-Mwesigwa, C. F. A. Negre, and S. M. Mniszewski, Graph partitioning using quantum anneal- ing on the d-wave system , 2017. arXiv: 1705 . 03082 [quant-ph]

2017
[10]

Ising formulations of many np problems,

A. Lucas, “Ising formulations of many np problems,” Frontiers in Physics , vol. 2, 2014, ISSN : 2296-424X. DOI: 10.3389/fphy.2014.00005

2014
[11]

Bucher, N

D. Bucher, N. Kraus, J. Blenninger, M. Lachner, J. Stein, and C. Linnhoff-Popien, Towards robust bench- marking of quantum optimization algorithms , 2024. arXiv: 2405.07624 [quant-ph]

2024 arXiv
[12]

Lubinski, C

T. Lubinski, C. Coffrin, C. McGeoch, P. Sathe, J. Apanavicius, and D. E. B. Neira, Optimization appli- cations as quantum performance benchmarks , 2024. arXiv: 2302.02278 [quant-ph]

2024 arXiv
[13]

Algorithm engi- neering for a quantum annealing platform,

A. D. King and C. C. McGeoch, “Algorithm engi- neering for a quantum annealing platform,” CoRR, vol. abs/1410.2628, 2014. arXiv: 1410.2628

2014 arXiv
[14]

Power of Pausing: Advancing Understanding of Thermalization in Experimental Quantum Annealers,

J. Marshall, D. Venturelli, I. Hen, and E. G. Rief- fel, “Power of Pausing: Advancing Understanding of Thermalization in Experimental Quantum Annealers,” Physical Review Applied , vol. 11, no. 4, p. 044 083, Apr. 2019, Publisher: American Physical Society. DOI: 10 . 1103 / Phy...

2019
[15]

A tutorial on spectral clustering,

U. von Luxburg, “A tutorial on spectral clustering,” Statistics and Computing , pp. 395–416, 2007. DOI: 10. 1007/s11222-007-9033-z

2007
[16]

An efficient heuristic procedure for partitioning graphs,

B. W. Kernighan and S. Lin, “An efficient heuristic procedure for partitioning graphs,” The Bell System Technical Journal, vol. 49, no. 2, pp. 291–307, 1970. DOI: 10.1002/j.1538-7305.1970.tb01770.x

1970
[17]

Bench- marking advantage and d-wave 2000q quantum anneal- ers with exact cover problems,

D. Willsch, M. Willsch, C. D. G. Calaza, et al., “Bench- marking advantage and d-wave 2000q quantum anneal- ers with exact cover problems,” Quantum Information Processing, vol. 21, no. 4, p. 141, 2022, ISSN : 1573-

2022
[18]

Optimizing embedding-related quantum annealing pa- rameters for reducing hardware bias,

A. Barbosa, E. Pelofske, G. Hahn, and H. N. Djidjev, “Optimizing embedding-related quantum annealing pa- rameters for reducing hardware bias,” in Parallel Ar- chitectures, Algorithms and Programming , L. Ning, V . Chau, and F. Lau, Eds., Singapore: Springer Singapore, 2021, pp...

2021
[19]

[Online]

Gurobi Optimization, LLC, Gurobi Optimizer Reference Manual, 2024. [Online]. Available: https://www.gurobi. com

2024
[20]

[On- line]

D-Wave, The AdvantageTM Quantum Computer . [On- line]. Available: https : / / www . dwavequantum . com / solutions-and-products/systems/
[21]

Quantum annealing with inequality constraints: The set cover problem,

H. N. Djidjev, “Quantum annealing with inequality constraints: The set cover problem,”Advanced Quantum Technologies, vol. 6, no. 11, p. 2 300 104, 2023. DOI: https://doi.org/10.1002/qute.202300104

2023 doi
[22]

QUARK: A Framework for Quantum Computing Application Benchmarking,

J. R. Fin ˇzgar, P. Ross, L. H ¨olscher, J. Klepsch, and A. Luckow, “QUARK: A Framework for Quantum Computing Application Benchmarking,” in 2022 IEEE International Conference on Quantum Computing and Engineering (QCE), Sep. 2022, pp. 226–237. DOI: 10. 1109/QCE53715.2022.00042

2022
[1332]

DOI: 10.1007/s11128-022-03476-y

doi
[2024]

arXiv: 2401.07184 [quant-ph]

arXiv

Pith tools

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

[1] [1]

Industry quantum computing applications,

A. Bayerstadler, G. Becquin, J. Binder, et al., “Industry quantum computing applications,” EPJ Quantum Tech- nology, vol. 8, no. 1, p. 25, Nov. 2021,ISSN : 2196-0763. DOI: 10.1140/epjqt/s40507-021-00114-x

doi:10.1140/epjqt/s40507-021-00114-x 2021

[2] [2]

[Online]

BMW Group, High-tech in production: BMW Group enables automated driving for new vehicles , 2024. [Online]. Available: https : / / www . press . bmwgroup . com/global/article/detail/T0446493EN/high- tech- in- production:- bmw- group- enables- automated- driving- for-new-vehicles?language=en (visited on 03/2025)

work page 2024

[3] [3]

Algorithms for quantum computation: Discrete logarithms and factoring,

P. Shor, “Algorithms for quantum computation: Discrete logarithms and factoring,” in Proceedings 35th Annual Symposium on Foundations of Computer Science, 1994, pp. 124–134. DOI: 10.1109/SFCS.1994.365700

arXiv 1994

[4] [4]

Quantum annealing in the transverse ising model,

T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse ising model,” Physical Review E, vol. 58, no. 5, pp. 5355–5363, Nov. 1998, ISSN : 1095-3787. DOI: 10.1103/physreve.58.5355

doi:10.1103/physreve.58.5355 1998

[5] [5]

H. M. Bauza and D. A. Lidar, Scaling advantage in approximate optimization with quantum annealing ,

work page

[6] [6]

Increasing safety of automated driving by infrastructure-based sensors,

T. D. Borba, O. Vacul ´ın, H. Marzbani, and R. N. Jazar, “Increasing safety of automated driving by infrastructure-based sensors,” IEEE Access , vol. 11, pp. 94 974–94 991, 2023. DOI: 10.1109/ACCESS.2023. 3311136

doi:10.1109/access.2023 2023

[7] [7]

On the optimal placement of cameras for surveillance and the underlying set cover problem,

J. Kritter, M. Br ´evilliers, J. Lepagnot, and L. Idoumghar, “On the optimal placement of cameras for surveillance and the underlying set cover problem,” Applied Soft Computing , vol. 74, pp. 133–153, 2019, ISSN : 1568-4946. DOI: https://doi.org/10.1016/j.asoc. 2018.10.025

doi:10.1016/j.asoc 2019

[8] [8]

Theory versus practice in annealing- based quantum computing,

C. C. McGeoch, “Theory versus practice in annealing- based quantum computing,” Theoretical Computer Sci- ence, vol. 816, pp. 169–183, 2020, ISSN : 0304-3975. DOI: https://doi.org/10.1016/j.tcs.2020.01.024

work page doi:10.1016/j.tcs.2020.01.024 2020