Implicit Differentiation for Measurement-Efficient Bilevel Quantum-Classical Optimization

Claudia Linnhoff-Popien; Federico Harjes Ruiloba; Jonas Stein; Markus Baumann; Maximilian Zorn; Tobias Rohe

REVIEW 2 major objections 5 minor 35 references

Implicit Differentiation for Measurement-Efficient Bilevel Quantum-Classical Optimization

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Outer gradients for parametric quantum optimization can be read off measurements already taken during the inner solve, at essentially zero extra circuit cost.

desk verdict Clean theory, but the matched-budget experiments undercount SPSA's true cost, so the headline efficiency gains aren't yet established. read the letter →

arxiv 2608.07717 v1 pith:KO6DA55Z submitted 2026-08-07 quant-ph

Tobias Rohe , Markus Baumann , Federico Harjes Ruiloba , Maximilian Zorn , Jonas Stein , Claudia Linnhoff-Popien This is my paper · ORCID

classification quant-ph

keywords bileveloptimizationvariationalquantumalgorithmsimplicitdifferentiationenvelopetheoremcorrelatorreuseMax-CutQUBOmeasurementefficiency

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

This paper proposes that when a variational quantum algorithm solves a family of parametric QUBO problems—such as weighted Max-Cut with edge weights depending on an outer control $\lambda$—the outer gradient of the optimal value function $F(\lambda)$ can be estimated without any new quantum circuits. The mechanism combines the envelope theorem with the fact that all needed edge-cut probabilities $p_e$ are already measured during the inner energy evaluation. For VQE, where the trial state does not depend on $\lambda$, this reuse yields an unbiased gradient; for QAOA, the state inherits $\lambda$-dependence, introducing a bias that must be traded against measurement cost. Under a matched evaluation budget, the paper reports about 4% better budget-normalized efficiency in one-dimensional outer control and over 14% in multi-dimensional edge-wise control, relative to central finite differences.

What carries the argument

The load-bearing object is the envelope identity (Eq. 27), $dF(\lambda)/d\lambda = \partial J(\phi^*(\lambda),\lambda)/\partial\lambda$, which removes the need to differentiate the inner optimizer map. In this setting it is instantiated as correlator-reuse: the estimator $\hat{g}_{CR}(\lambda)=\sum_e w'_e(\lambda)\hat{p}_e$ (Eq. 30) reweights edge-cut probabilities $\hat{p}_e$ that are computed by classical post-processing from the same bitstrings as the energy estimate. This works because Max-Cut's commuting projectors mean one $Z$-basis measurement batch yields all correlators simultaneously. The remaining machinery is the distinction between VQE, where $\rho$ is $\lambda$-independent, and QAOA, where the cost unitary $e^{-i\gamma_\ell H_C(\lambda)}$ makes the state $\lambda$-dependent, so the full derivative includes $\sum_e w_e(\lambda)\partial_\lambda p_e$ (Eq. 33).

What would settle it

Take one parametric instance, solve the inner problem classically to obtain the true value function $F(\lambda)$, then run CR-ID and central finite differences with the inner solver deliberately stopped after 1, 5, 20, and 100 SPSA iterations; if the CR-ID gradient is not a descent direction or its advantage disappears at low iteration counts, the zero-cost claim fails in that non-stationary regime.

Watch

Extended reading notes

Core claim

The central claim is that for diagonal cost Hamiltonians of the form $H_C(\lambda)=\sum_e w_e(\lambda)\Pi_e$, the derivative of the value function $F(\lambda)=\max_\phi J(\phi,\lambda)$ at an inner optimum reduces, by the envelope theorem, to the partial derivative $\partial_\lambda J = \sum_e w'_e(\lambda) p_e(\phi^*)$. Since every $p_e$ is an edge-cut probability already estimated from the same $Z$-basis shots used to estimate the inner energy, the outer signal is available at essentially zero additional measurement cost. The paper then states the architecture-dependence result: for VQE, $\partial_\lambda \rho(\theta)=0$, so the reuse estimator is exactly unbiased; for QAOA, the cost unitary injects $\lambda$ into the state preparation, producing a state-dependence term that is zero-cost to ignore but biased, or extra-cost to estimate.

Load-bearing premise

The load-bearing premise, acknowledged in the paper's limitations section, is that the inner variational solver has reached a stationary point so the envelope identity applies; the experiments allow only 20 single-shot SPSA iterations, so this is an approximation whose error is not characterized.

Editorial extensions

If this is right

For any diagonal Hamiltonian with $\lambda$-dependent coefficients, an outer-loop gradient update costs only the inner solve at the current $\lambda$, removing the factor-of-$M$ probe overhead (e.g., $M=3$ for central finite differences) per outer step.
Under matched evaluation budgets, the method reports consistent area-under-curve gains, about 4% in one-dimensional outer control and over 14% when the control is edge-wise and multi-dimensional, with trajectory curves reaching the plateau earlier.
The reuse gradient is exact for VQE if the inner solver has reached stationarity; for QAOA, using only the explicit term is biased, and estimating the missing state-dependence term reintroduces measurement overhead—a cost–bias trade-off practitioners must choose between.
Operationally, when solutions are read out by best-of-32 sampling, QAOA's tail quality narrows the expectation-level gap, but VQE retains a higher per-shot probability of drawing a solution within 10% of the optimum, which matters under limited readout budget.

Reading between the lines

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

Editorial extension: for non-diagonal Hamiltonians whose Pauli terms fall into a few commuting groups, the same reuse argument should apply per group, so the zero-additional-cost property should survive with overhead equal to the number of measurement groups rather than the number of Hamiltonian terms.
Editorial extension: a practical estimator for the QAOA state-dependence term could be built from parameter-shift-style shifts in $\lambda$ at the final variational angles, costing only a small constant number of additional inner-solve-equivalent circuits; the paper does not test such an estimator.
Editorial extension: because the reuse signal only needs weight derivatives, a black-box user who can query $w_e(\lambda)$ classically can obtain a fully classical gradient of the weights and keep the quantum cost at zero, even when the response functions are not analytic.
Editorial extension: the envelope identity suggests a quantitative diagnostic for when reuse-only gradients can be trusted: measure the residual inner-gradient norm $\|\nabla_\phi J(\phi,\lambda)\|$; if it is below a threshold, the bias from non-stationarity is expected to be small. The paper does not derive that threshold.

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's central theoretical observation is correct and clean—for diagonal cost Hamiltonians, the outer gradient's partial term can be estimated by reweighting correlators already measured in the inner VQE solve, and this reuse is exact for VQE but biased for QAOA because the control enters the state. The architecture-dependence analysis is the genuinely new part. The empirical claims, though, are not yet credible because the evaluation budget is counted in a way that doesn't match the actual measurement cost.

The envelope derivation (Eqs. 26–32) is sound under stationarity. The specialization to Max-Cut is straightforward but useful. The paper is honest about its limitations, including the fact that the envelope identity only holds at inner stationarity.

The soft spots. First, budget accounting. The stress-test note claims the formula B=T(3I+1) undercounts FD by a factor of three. That specific arithmetic is wrong: 3I+1 bills for three inner solves (each of I evaluations) plus one extra evaluation. What is wrong is the cost model inside each inner solve: SPSA requires two energy evaluations per iteration (one at each perturbed parameter), while the paper counts one. So the true number of shots for an inner solve is about 2I, not I. For FD, with three inner solves per outer step, the true cost is roughly 6I per outer step, not 3I. The reported budget B=1830 is therefore roughly half the true shot count for CR-ID and about a third of the true count for FD. That means the matched-budget comparison doesn't reflect the real measurement burden; the reported 4% and 14% gains could easily shift once the accounting is corrected. This is the main reason I don't trust the empirical conclusions.

Second, stationarity. After 20 single-shot SPSA iterations, the inner solution is nowhere near a stationary point of J. The envelope theorem's condition is not met, so the CR-ID signal is a biased estimate of the outer gradient. The authors acknowledge this but don't characterize the bias. It's possible CR-ID works precisely because the biased signal is a useful heuristic; that would still be worth reporting, but the paper currently frames it as an unbiased gradient method. That overreaches.

Third, the experimental comparison is thin: one baseline (central FD), no code or data, and the multi-dimensional results show a suspiciously low FD tail hit-rate. These need to be checked before the efficiency claims can be taken at face value.

Who should read this: anyone working on parametric VQA optimization or measurement-efficient outer loops. The theoretical core is a legitimate, citable result. But the paper needs major experimental revisions: correct the inner-solve cost model, run FD within a truly matched budget, and quantify the stationarity bias. I would send it to peer review—the theory deserves referee time—but I'd expect the quantitative claims to change substantially.

Referee Report

2 major / 5 minor

Summary. The paper proposes a bilevel optimization framework for parametric diagonal cost Hamiltonians, where an outer loop tunes a continuous parameter λ and an inner VQA solve optimizes circuit parameters. The main methodological contribution is correlator-reuse implicit differentiation (CR-ID), which uses the envelope theorem to estimate outer gradients from edge-cut probabilities already measured during inner energy evaluation, avoiding extra circuit executions. The authors derive unbiased reuse gradients for VQE and identify a state-dependence bias term for QAOA. They report experimental comparisons against finite-difference value probing under a claimed matched evaluation budget, showing approximately 4% and over 14% improvements in 1D and multi-dimensional settings.

Significance. The theoretical core of the paper is a clean application of the envelope theorem to parametric VQAs, and the architecture-dependence result (VQE vs. QAOA) is a useful insight for practitioners. The presentation is generally clear, and the empirical study uses paired per-instance comparisons across multiple coefficient families and operational metrics. The method is falsifiable and the derivations in Eqs. (26)-(32) are transparent. However, the empirical claims rest on a budget accounting that appears internally inconsistent.

major comments (2)

[VI-C] The matched evaluation budget in Section VI-C is defined as B=T(3I+1)=30·(3·20+1)=1830, which charges exactly one inner solve per outer step. This contradicts the FD baseline description in the same section, which requires M=3 inner solves per outer step (center and two perturbation evaluations). If an inner solve costs 3I+1=61 energy evaluations, the correct FD cost is T·3·(3I+1)=5490; if an inner solve costs I+1 evaluations, the correct FD cost is T·3·(I+1)=1890. With the reported B=1830, the FD baseline is credited with only about one-third of its true cost, so the AUC_B improvements in Tables II-IV and Figures 2-3 do not establish a matched-budget advantage. This is a load-bearing error because the headline efficiency claim depends directly on this accounting. Please rerun the experiments with a correctly matched budget (e.g., B=5490 for both methods, or with FD allowed only T/3 outer steps) and report the resulting metrics.
[V-C, VIII-A] The envelope identity (Eq. 27) and the unbiasedness of the CR-ID estimator (Eq. 32) require the inner solver to reach a stationary point, ∇_φ J(φ*(λ), λ)=0. With I=20 SPSA iterations and a single shot per energy evaluation, the inner solution is almost certainly far from stationarity, so the CR-ID gradient is biased for VQE as well. The paper acknowledges this in the Limitations but does not characterize the magnitude of the bias or demonstrate robustness with respect to the inner budget. I ask for either a theoretical bound on the bias in terms of the inner optimality gap or a sensitivity study with increased inner iterations (e.g., I=50, 100, 200). Without this, the 'exact reuse gradient' claim in Section V-E is only an asymptotic statement and cannot be assessed against the finite-budget experiments.

minor comments (5)

[VI-C] The parenthetical 'three objective evaluations per SPSA iteration plus one final evaluation per outer step' is confusing; please clarify the cost of an inner solve and reconcile it with Eq. (24).
[V-A] Equation (24) defines C_outer ≈ M·N_inner, but the budget formula in Section VI-C appears to set M=1; please ensure the notation and cost model are consistent throughout.
[Table II] Use consistent subscript notation 'AUC_B' rather than 'AU C_B' across tables and text.
[References] Reference [32] is an abstract in the Bulletin of the American Physical Society; consider citing a full journal or arXiv version if one exists.
[Acknowledgment] The disclosure that generative AI was used to create text, tables, graphs, code, and citations is unusual; please verify that it complies with the journal's policy on AI use and disclosure.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the CR-ID gradient estimator is a direct application of the envelope theorem and reuses measured correlators without fitting or self-referential input.

full rationale

The derivation is self-contained. Eq. (21) defines the inner objective as a weighted sum of edge-cut probabilities, Eq. (27) is the envelope identity obtained from the chain rule and the stationarity condition, and Eq. (28) is the corresponding partial derivative with respect to the outer parameter. Eq. (30) replaces the exact probabilities by their sample estimates, so Eq. (32) states unbiasedness by linearity of expectation. No parameter is fitted to make the identity hold, and no claimed prediction is constructed from the data it is tested against. The cited foundation [32] is work by Ahmed, Killoran, and Carrasquilla, not by the present authors, and the envelope theorem is attributed to Milgrom and Segal [14]. The only reference to work from the present group is a background example [7] on vehicle routing, which is not load-bearing. The acknowledged stationarity limitation and the possible budget-accounting inconsistency in Section VI-C are genuine correctness or fairness concerns, but they do not make the theoretical derivation circular: the validity of the CR-ID estimator does not depend on how the FD baseline's evaluation budget is computed.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The theoretical derivation introduces no free parameters and no invented entities; it rests on the envelope theorem, the diagonal commuting Hamiltonian structure, and the architectural assumption that VQE states are lambda-independent at fixed parameters. The empirical efficiency numbers are protocol-dependent simulation results, not derived bounds.

free parameters (4)

Number of inner SPSA iterations I = 20 (10 for Experiment 9)
Hand-chosen; the stationarity condition of the envelope theorem is not met, and the bias of CR-ID under finite I is not analyzed.
SPSA perturbation/step-size hyperparameters = not specified
Not reported; needed to reproduce the empirical efficiency gains.
Finite-difference perturbation c_t = not specified
Used in the FD baseline; the reported gain depends on this choice.
Tail threshold epsilon and target probability = epsilon=0.1, p_target=0.99
Used to define the tail hit-rate metric; not fitted to the method.

assumptions (5)

standard math Envelope theorem (Milgrom-Segal) applies to the value function F(lambda)=max_phi J(phi,lambda), requiring differentiability and inner stationarity.
Invoked in Eq. (26)-(27) to drop the dphi*/dlambda term.
domain assumption The parametric Max-Cut Hamiltonian is diagonal in the computational basis, so all edge projectors commute and a single Z-basis measurement estimates all edge correlators p_e simultaneously.
Used throughout Section V-D to justify correlator reuse at zero extra measurement settings.
domain assumption For VQE with fixed circuit parameters, the quantum state does not depend on the outer parameter lambda.
Used in Eq. (28) and Section V-E to claim exact reuse gradients for VQE; does not hold for QAOA.
domain assumption Edge weight functions w_e(lambda) are differentiable and positive on the domain, and their derivatives dw_e/dlambda are available analytically or via classical finite differences.
Needed for Eq. (28)-(30); positivity ensures valid weighted Max-Cut for all lambda.
ad hoc to paper The inner VQA solver reaches (approximately) a stationary point of J(phi,lambda) within the allocated budget.
Required for the envelope identity (Eq. 27) to hold; the experiments use 20 SPSA iterations, so this is an approximation whose error is not quantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Implicit Differentiation for Measurement-Efficient Bilevel Quantum-Classical Optimization." pith.science (2026). https://pith.science/paper/KO6DA55Z

@misc{pith2026260807717,
  author       = {Pith},
  title        = {Pith review of: Implicit Differentiation for Measurement-Efficient Bilevel Quantum-Classical Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KO6DA55Z}},
  note         = {Machine review of arXiv:2608.07717}
}

read the original abstract

Quantum optimization has shown promising results for quadratic unconstrained binary optimization (QUBO) problems. Real-world applications, however, often involve polynomial coefficients that depend on tunable external factors - such as demand forecasts or risk preferences - giving rise to bilevel optimization structures. We show how variational quantum algorithms (VQAs) can efficiently handle such parametric problems, making three contributions. First, we propose a bilevel optimization model for diagonal cost Hamiltonians where coefficients depend on a tunable outer parameter: an outer loop adjusts this parameter - reshaping the cost landscape - while an inner VQA optimizes circuit variables. Second, since derivative-free probing methods incur a multiplicative overhead when each outer evaluation requires a complete inner solve, we develop correlator-reuse implicit differentiation (CR-ID), which obtains outer gradients by reusing quantum measurements already collected during inner energy estimation, requiring essentially no additional circuit executions. Experiments across three coefficient families show that CR-ID consistently improves budget-normalized efficiency by ~4\% in 1D and over 14\% in multi-dimensional settings, showing a significant performance advantage compared to finite-difference methods. Third, we show that this property is architecture-dependent: VQE admits exact reuse gradients, whereas QAOA introduces a state-dependent term that creates a cost-bias trade-off.

Figures

Figures reproduced from arXiv: 2608.07717 by the authors.

**Figure 2.** Bilevel optimization diagnostics for the periodic 1D family showing per-instance comparisons between CR-ID and central finite-difference value probing (FD) under a matched evaluation budget B = 1830. Each point corresponds to one problem instance (N = 300), with color indicating the edge density pedge. The black diagonal denotes equal performance. Points above the diagonal indicate instances where ID outperforms FD… view at source ↗

**Figure 1.** Bilevel optimization diagnostics VQE on periodic family. (a) Classical envelope Fˆ∗ (λ) with visited outer iterates overlaid (representative instance). (b) Active-branch structure across λ; switch points mark changes of the optimal bitstring. • Readout best normalized objective at budget (best-of-32 samples); • Tail hit-rate pgood at threshold ε = 0.1. For readout-facing experiments, we additionally report bestof-S… view at source ↗

**Figure 3.** Budget efficiency in the multi-dimensional (λ) setting under a matched evaluation budget. Here, the outer variable is edge-wise, i.e., λ ∈ R |E| assigns one control parameter λe per edge. (a) Best-so-far normalized objective J/J∗ versus cumulative energy evaluations. (b) Budgetmatched best-of-S readout J/J∗ across instances in the multidimensional setting, comparing ID and FD. sizes: CR-ID consistently attains it… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: VQE versus QAOA under a matched objectiveevaluation budget. (a) Representative instance: classical envelope Fb(λ) (black) with the outer iterates {λt} visited by VQE (red) and QAOA (blue); filled markers show expectation estimates and open markers indicate best-of-S…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 22 canonical work pages

[1]

A quantum approximate optimization algorithm,

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

arXiv 2014
[2]

A variational eigenvalue solver on a photonic quantum processor,

A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’brien, “A variational eigenvalue solver on a photonic quantum processor,”Nature communications, vol. 5, no. 1, p. 4213, 2014

2014
[3]

Variational quantum algorithms,

M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,”Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

2021
[4]

Filtering variational quantum algorithms for combinatorial optimization,

D. Amaro, C. Modica, M. Rosenkranz, M. Fiorentini, M. Benedetti, and M. Lubasch, “Filtering variational quantum algorithms for combinatorial optimization,”Quantum Science and Technology, vol. 7, no. 1, p. 015021, 2022

work page 2022
[5]

A review on Quantum Approximate Optimization Algorithm and its variants,

K. Blekos, D. Brand, A. Ceschini, C.-H. Chou, R.-H. Li, K. Pandya, and A. Summer, “A review on Quantum Approximate Optimization Algorithm and its variants,”Physics Reports, vol. 1068, pp. 1–66, 2024

work page 2024
[6]

Best practices for portfolio optimization by quantum computing, ex- perimented on real quantum devices,

G. Buonaiuto, F. Gargiulo, G. De Pietro, M. Esposito, and M. Pota, “Best practices for portfolio optimization by quantum computing, ex- perimented on real quantum devices,”Scientific Reports, vol. 13, no. 1, p. 19434, 2023

work page 2023
[7]

A hybrid solution method for the capacitated vehicle routing problem using a quantum annealer,

S. Feld, C. Roch, T. Gabor, C. Seidel, F. Neukart, I. Galter, W. Mauerer, and C. Linnhoff-Popien, “A hybrid solution method for the capacitated vehicle routing problem using a quantum annealer,”Frontiers in ICT, vol. 6, p. 13, 2019

work page 2019
[8]

The time-dependent traveling salesman problem and its application to the tardiness problem in one-machine scheduling,

J.-C. Picard and M. Queyranne, “The time-dependent traveling salesman problem and its application to the tardiness problem in one-machine scheduling,”Operations research, vol. 26, no. 1, pp. 86–110, 1978

work page 1978

Show all 35 references

[9]

A review on bilevel optimization: From classical to evolutionary approaches and applications,

A. Sinha, P. Malo, and K. Deb, “A review on bilevel optimization: From classical to evolutionary approaches and applications,”IEEE transactions on evolutionary computation, vol. 22, no. 2, pp. 276–295, 2017

2017
[10]

Real-coded quantum optimization-based bi-level dispatching strategy of integrated power and heat systems,

B. Deng, Y . Teng, Q. Hui, T. Zhang, and X. Qian, “Real-coded quantum optimization-based bi-level dispatching strategy of integrated power and heat systems,”IEEE Access, vol. 8, pp. 47 888–47 899, 2020

2020
[11]

Bi-level multi-objective optimization scheduling for regional integrated energy systems based on quantum evolutionary algorithm,

W. Fan, Q. Liu, and M. Wang, “Bi-level multi-objective optimization scheduling for regional integrated energy systems based on quantum evolutionary algorithm,”Energies, vol. 14, no. 16, p. 4740, 2021

2021
[12]

A quantum-inspired bilevel optimization algorithm for the first responder network design problem,

A. Karahalios, S. Tayur, A. Tenneti, A. Pashapour, F. S. Salman, and B. Yıldız, “A quantum-inspired bilevel optimization algorithm for the first responder network design problem,”INFORMS Journal on Computing, vol. 37, no. 1, pp. 172–188, 2025

2025
[13]

Eval- uating analytic gradients on quantum hardware,

M. Schuld, V . Bergholm, C. Gogolin, J. Izaac, and N. Killoran, “Eval- uating analytic gradients on quantum hardware,”Physical Review A, vol. 99, no. 3, p. 032331, 2019

2019
[14]

Envelope theorems for arbitrary choice sets,

P. Milgrom and I. Segal, “Envelope theorems for arbitrary choice sets,” Econometrica, vol. 70, no. 2, pp. 583–601, 2002

2002
[15]

Reducibility among combinatorial problems,

R. M. Karp, “Reducibility among combinatorial problems,” inComplex- ity of Computer Computations, ser. The IBM Research Symposia Series, R. E. Miller, J. W. Thatcher, and J. D. Bohlinger, Eds. Boston, MA: Springer US, 1972, pp. 85–103

1972
[16]

Optimization, approximation, and complexity classes,

C. H. Papadimitriou and M. Yannakakis, “Optimization, approximation, and complexity classes,”Journal of Computer and System Sciences, vol. 43, pp. 425–440, 1991

1991
[17]

Some optimal inapproximability results,

J. H ˚astad, “Some optimal inapproximability results,”Journal of the ACM (JACM), vol. 48, no. 4, pp. 798–859, 2001

2001
[18]

Improved approximation algo- rithms for maximum cut and satisfiability problems using semidefinite programming,

M. X. Goemans and D. P. Williamson, “Improved approximation algo- rithms for maximum cut and satisfiability problems using semidefinite programming,”Journal of the ACM (JACM), vol. 42, no. 6, pp. 1115– 1145, 1995

1995
[19]

On the power of unique 2-prover 1-round games,

S. Khot, “On the power of unique 2-prover 1-round games,” inProceed- ings of the thiry-fourth annual ACM symposium on Theory of computing, 2002, pp. 767–775

2002
[20]

Ising formulations of many np problems,

A. Lucas, “Ising formulations of many np problems,”Frontiers in physics, vol. 2, p. 5, 2014

2014
[21]

In finding the correction for the open end of an organ- pipe,

J. Rayleigh, “In finding the correction for the open end of an organ- pipe,”Phil. Trans, vol. 161, no. 77, p. 1870, 1870
[22]

¨Uber eine neue Methode zur L ¨osung gewisser Variationsprob- leme der mathematischen Physik,

W. Ritz, “ ¨Uber eine neue Methode zur L ¨osung gewisser Variationsprob- leme der mathematischen Physik,”Journal f ¨ur die reine und angewandte Mathematik, vol. 135, pp. 1–61, 1908

1908
[23]

Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,

J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,”IEEE transactions on automatic control, vol. 37, no. 3, pp. 332–341, 2002

2002
[24]

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,”Nature communications, vol. 9, no. 1, p. 4812, 2018

2018
[25]

Evolving objective function for improved variational quantum optimization,

I. Kolotouros and P. Wallden, “Evolving objective function for improved variational quantum optimization,”Physical Review Research, vol. 4, no. 2, p. 023225, 2022

2022
[26]

Optimizing millions of hyper- parameters by implicit differentiation,

J. Lorraine, P. Vicol, and D. Duvenaud, “Optimizing millions of hyper- parameters by implicit differentiation,” inInternational conference on artificial intelligence and statistics. PMLR, 2020, pp. 1540–1552

2020
[27]

Meta-learning with implicit gradients,

A. Rajeswaran, C. Finn, S. M. Kakade, and S. Levine, “Meta-learning with implicit gradients,”Advances in neural information processing systems, vol. 32, 2019

2019
[28]

Bilevel programming for hyperparameter optimization and meta-learning,

L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil, “Bilevel programming for hyperparameter optimization and meta-learning,” in International conference on machine learning. PMLR, 2018, pp. 1568– 1577

2018
[29]

Quantum circuit learning,

K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, “Quantum circuit learning,”Physical Review A, vol. 98, no. 3, p. 032309, 2018

2018
[30]

General parameter- shift rules for quantum gradients,

D. Wierichs, J. Izaac, C. Wang, and C. Y .-Y . Lin, “General parameter- shift rules for quantum gradients,”Quantum, vol. 6, p. 677, 2022

2022
[31]

Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule,

L. Banchi and G. E. Crooks, “Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule,”Quantum, vol. 5, p. 386, 2021

2021
[32]

Implicit differentiation of variational quantum algorithms,

S. Ahmed, N. Killoran, and J. Carrasquilla, “Implicit differentiation of variational quantum algorithms,”Bulletin of the American Physical Society, vol. 68, 2023

2023
[33]

On the evolution of random graphs,

P. Erd ˝os and A. R ´enyi, “On the evolution of random graphs,”Pub- lications of the Mathematical Institute of the Hungarian Academy of Sciences, vol. 5, no. 1, pp. 17–60, 1960

1960
[34]

Quantum approximate optimization algorithm: Performance, mechanism, and im- plementation on near-term devices,

L. Zhou, S.-T. Wang, S. Choi, H. Pichler, and M. D. Lukin, “Quantum approximate optimization algorithm: Performance, mechanism, and im- plementation on near-term devices,”Physical Review X, vol. 10, no. 2, p. 021067, 2020

2020
[35]

Multiobjective optimization and network routing with near-term quantum computers,

S.-H. Chiew, K. Poirier, R. Mishra, U. Bornheimer, E. Munro, S. H. Foon, C. W. Chen, W. S. Lim, and C. W. Nga, “Multiobjective optimization and network routing with near-term quantum computers,” IEEE Transactions on Quantum Engineering, vol. 5, pp. 1–19, 2024

2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[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

arXiv 2014

[2] [2]

A variational eigenvalue solver on a photonic quantum processor,

A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’brien, “A variational eigenvalue solver on a photonic quantum processor,”Nature communications, vol. 5, no. 1, p. 4213, 2014

2014

[3] [3]

Variational quantum algorithms,

M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,”Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

2021

[4] [4]

Filtering variational quantum algorithms for combinatorial optimization,

D. Amaro, C. Modica, M. Rosenkranz, M. Fiorentini, M. Benedetti, and M. Lubasch, “Filtering variational quantum algorithms for combinatorial optimization,”Quantum Science and Technology, vol. 7, no. 1, p. 015021, 2022

work page 2022

[5] [5]

A review on Quantum Approximate Optimization Algorithm and its variants,

K. Blekos, D. Brand, A. Ceschini, C.-H. Chou, R.-H. Li, K. Pandya, and A. Summer, “A review on Quantum Approximate Optimization Algorithm and its variants,”Physics Reports, vol. 1068, pp. 1–66, 2024

work page 2024

[6] [6]

Best practices for portfolio optimization by quantum computing, ex- perimented on real quantum devices,

G. Buonaiuto, F. Gargiulo, G. De Pietro, M. Esposito, and M. Pota, “Best practices for portfolio optimization by quantum computing, ex- perimented on real quantum devices,”Scientific Reports, vol. 13, no. 1, p. 19434, 2023

work page 2023

[7] [7]

A hybrid solution method for the capacitated vehicle routing problem using a quantum annealer,

S. Feld, C. Roch, T. Gabor, C. Seidel, F. Neukart, I. Galter, W. Mauerer, and C. Linnhoff-Popien, “A hybrid solution method for the capacitated vehicle routing problem using a quantum annealer,”Frontiers in ICT, vol. 6, p. 13, 2019

work page 2019

[8] [8]

The time-dependent traveling salesman problem and its application to the tardiness problem in one-machine scheduling,

J.-C. Picard and M. Queyranne, “The time-dependent traveling salesman problem and its application to the tardiness problem in one-machine scheduling,”Operations research, vol. 26, no. 1, pp. 86–110, 1978

work page 1978