pith. sign in

arxiv: 2502.02625 · v2 · submitted 2025-02-04 · 💻 cs.LG · quant-ph

Bayesian Parameter Shift Rule in Variational Quantum Eigensolvers

Samuele Pedrielli , Christopher J. Anders , Lena Funcke , Karl Jansen , Kim A. Nicoli , Shinichi Nakajima This is my paper

Pith reviewed 2026-05-23 03:34 UTC · model grok-4.3

classification 💻 cs.LG quant-ph
keywords Bayesian parameter shift rulevariational quantum eigensolverGaussian processgradient estimationstochastic gradient descentquantum optimizationuncertainty quantification
0
0 comments X

The pith

Bayesian parameter shift rule estimates VQE gradients from arbitrary past observations with uncertainty.

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

The paper introduces a Bayesian variant of the parameter shift rule for gradient estimation in variational quantum eigensolvers. Gaussian processes model the objective function to produce gradient estimates and uncertainty from measurements at any locations rather than only at fixed shift points. This flexibility supports reusing data collected in earlier stochastic gradient descent steps to accelerate convergence. The posterior uncertainty is combined with a gradient confident region concept to decide when fewer new observations suffice. Numerical results show the combined approach speeds up optimization and beats standard methods including sequential minimal optimization.

Core claim

The Bayesian PSR estimates the gradient of the VQE objective using Gaussian process regression, offering estimates from observations at arbitrary locations along with uncertainty quantification. It reduces to the generalized PSR in special cases. In SGD, this flexibility permits reusing observations from prior steps to accelerate optimization, and the posterior uncertainty combined with the gradient confident region concept allows minimizing observation costs per step.

What carries the argument

Bayesian parameter shift rule implemented via Gaussian process regression with appropriate kernels on the VQE objective function

If this is right

  • Past observations can be reused across SGD steps without requiring new circuit evaluations at every iteration.
  • Posterior uncertainty combined with GradCoRe reduces the number of new observations needed while preserving gradient reliability.
  • The method reduces exactly to the generalized parameter shift rule when observations coincide with the required shift locations.
  • Numerical experiments show faster convergence than sequential minimal optimization on the tested VQE instances.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reuse mechanism could be combined with adaptive step-size rules that scale with reported uncertainty.
  • Kernel choice may need adjustment when the objective exhibits sharp features or noise not captured by the current stationarity assumption.
  • The same Gaussian-process construction might apply to gradient estimation in other variational quantum algorithms that rely on parameter shifts.

Load-bearing premise

The VQE objective function is sufficiently smooth and stationary that Gaussian processes with chosen kernels can produce reliable gradient estimates and uncertainty quantification from arbitrary past observation points.

What would settle it

If VQE runs that reuse past observations under Bayesian PSR require the same or greater total evaluations to reach target energies compared with standard PSR, or if gradient estimates from distant points show large systematic deviation from shift-rule values on the same circuit.

Figures

Figures reproduced from arXiv: 2502.02625 by Christopher J. Anders, Karl Jansen, Kim A. Nicoli, Lena Funcke, Samuele Pedrielli, Shinichi Nakajima.

Figure 1
Figure 1. Figure 1: Illustration of our gradient confident region (GradCoRe) approach. Our goal is to minimize the true energy f ∗ (x) over the set of parameters x ∈ [0, 2π) D, where we use a GP surrogate f(x) for approximating f ∗ (x). Observing f ∗ at points x− and x+ (green circles) along the d-th direction (solid horizontal line) decreases the uncertainty (dashed curves) not only for predicting f(x±), but also for predict… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the behavior of the Bayesian PSR when Vd = 1 (left) and when Vd = 2 (middle). Bayesian PSR prediction (red) coincides with general PSR (green cross) for the designed equidistant observations (magenta crosses). The right plot visualizes the variance (20) of the derivative GP prediction at x ′ , as a function of the shift α of observations when Vd = 1. Although the optimum is at α = π 2 , the… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between SGD with PSR (dashed curves) and SGD with Bayesian PSR (solid curves), as well as GradCoRe (red solid curve), on the Ising Hamiltonian with an (L = 3)-layered (Q = 5)-qubits quantum circuit. The energy (left) and fidelity (right) are plotted as functions of the cumulative Nshots, i.e., the total number of measurement shots. Except GradCoRe equipped with the adaptive shots strategy, the n… view at source ↗
Figure 4
Figure 4. Figure 4: Energy (left) and fidelity (right) achieved within the cumulative number of measurement shots for the Ising Hamiltonian with an (L = 3)-layered (Q = 5)-qubits quantum circuit. The curves correspond to SGLBO (blue), Bayes-NFT (green), EMICoRe (orange), SubsCoRe (purple), and our proposed GradCoRe (red). ditions, H = − P i∈{X,Y,Z} hPQ−1 j=1 (Jiσ i jσ i j+1) + PQ j=1 hiσ i j i , (25) where {σ i j }i∈{X,Y,Z} a… view at source ↗
Figure 5
Figure 5. Figure 5: Gradient estimation error by PSR (dashed curve) and Bayesian PSR (solid curve) for Nshots = 1024, evaluated by the L2-distance between the estimated gradient µe(xb) and the true gradient g ∗ (xb) (computed by the PSR with simulated noiseless measurements). (see [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between NFT (Nakanishi et al., 2020) and Bayes-NFT for the Ising Hamiltonian with an (L = 3)-layered (Q = 5)- qubits quantum circuit. The energy (left) and fidelity (right), in the forms of Eqs.(26) and (27), respectively, are plotted as functions of the cumulative Nshots, i.e., the total number of measurement shots. The number of shots per observation is set to Nshots = 128 (blue), 256 (green),… view at source ↗
Figure 7
Figure 7. Figure 7: Numerical validation of Theorem C.1 under two parameter settings (see above each panel). Given the 2Vd equidistant observations (magenta crosses), the derivative GP prediction (blue curve) with uncertainty (blue shades) is compared to their analytic forms (31) and (32), i.e., the mean function (red curve) and the variance function (red shades), respectively. We observe that our theory perfectly matches the… view at source ↗
Figure 8
Figure 8. Figure 8: shows the behavior of the GradCoRe threshold κ(t) (left), and the number ν(t) of measurement shots (left) that GradCoRe used in each SGD iteration. 0 200 400 600 800 1000 1200 #SGD Step 0.1 0.2 0.3 (t) 0 250 500 750 1000 #SGD Step 0 20000 40000 60000 80000 100000 # of measurement shots [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
read the original abstract

Parameter shift rules (PSRs) are key techniques for efficient gradient estimation in variational quantum eigensolvers (VQEs). In this paper, we propose its Bayesian variant, where Gaussian processes with appropriate kernels are used to estimate the gradient of the VQE objective. Our Bayesian PSR offers flexible gradient estimation from observations at arbitrary locations with uncertainty information and reduces to the generalized PSR in special cases. In stochastic gradient descent (SGD), the flexibility of Bayesian PSR allows the reuse of observations in previous steps, which accelerates the optimization process. Furthermore, the accessibility to the posterior uncertainty, along with our proposed notion of gradient confident region (GradCoRe), enables us to minimize the observation costs in each SGD step. Our numerical experiments show that the VQE optimization with Bayesian PSR and GradCoRe significantly accelerates SGD and outperforms the state-of-the-art methods, including sequential minimal optimization.

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.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a Bayesian variant of the parameter shift rule (PSR) for gradient estimation in variational quantum eigensolvers (VQEs). Gaussian processes with chosen kernels are used to estimate gradients from observations at arbitrary locations, providing uncertainty information. The method is claimed to reduce to the generalized PSR in special cases, enable reuse of prior SGD observations to accelerate optimization, and support a gradient confident region (GradCoRe) concept to reduce per-step observation costs. Numerical experiments indicate that the approach accelerates SGD and outperforms state-of-the-art methods including sequential minimal optimization.

Significance. If the GP-based gradient estimates and uncertainty quantification prove reliable for VQE objectives, the approach could meaningfully lower the number of circuit evaluations required during variational optimization by reusing historical data and adaptively controlling new measurements via GradCoRe. The explicit reduction to existing PSR methods and the uncertainty-aware stopping criterion represent potentially useful extensions of classical shift-rule techniques into a Bayesian setting.

major comments (3)
  1. [§3] The central claim that the Bayesian PSR reduces to the generalized PSR in special cases (abstract and §3) depends on the GP posterior mean recovering the exact shift-rule expression. The manuscript must specify the kernel family, noise model, and limiting procedure under which this equivalence holds exactly; without an explicit derivation or theorem showing the posterior mean matches the finite-difference form of the generalized PSR, the reduction remains unverified.
  2. [§4 and §5] VQE objective functions are trigonometric polynomials whose frequencies are determined by the circuit and Hamiltonian. The assumption that standard stationary kernels yield accurate gradients and calibrated posterior variances from arbitrarily spaced (including past SGD) observation points is load-bearing for both the reuse claim and GradCoRe. The paper should include a dedicated analysis or diagnostic (e.g., posterior predictive checks or gradient error vs. distance from current θ) demonstrating that kernel misspecification does not produce biased gradients or miscalibrated uncertainties on representative VQE landscapes.
  3. [§5] Table 2 and Figure 4 report speedups over SMO and standard PSR, but the experiments lack error bars on iteration counts or wall-clock time, details on the number of independent random seeds, and explicit reporting of the total number of circuit evaluations (including those used for GP hyperparameter fitting). Without these, it is impossible to determine whether the reported acceleration is statistically robust or specific to the chosen problem instances.
minor comments (2)
  1. [§3] Notation for the GP posterior mean and variance should be introduced with explicit dependence on the observation set D_t at each SGD step to clarify how reuse is implemented.
  2. [§4] The definition of GradCoRe (threshold on posterior variance) is introduced without a sensitivity analysis; a brief ablation on the threshold value would improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments. We address each major point below and will revise the manuscript to strengthen the presentation and add the requested clarifications and diagnostics.

read point-by-point responses

  1. Referee: [§3] The central claim that the Bayesian PSR reduces to the generalized PSR in special cases (abstract and §3) depends on the GP posterior mean recovering the exact shift-rule expression. The manuscript must specify the kernel family, noise model, and limiting procedure under which this equivalence holds exactly; without an explicit derivation or theorem showing the posterior mean matches the finite-difference form of the generalized PSR, the reduction remains unverified.

    Authors: We agree that an explicit derivation is required. In the revised manuscript we will insert a new theorem in §3 that states the precise conditions: a periodic kernel whose period matches the known frequencies of the VQE objective, homoscedastic Gaussian noise, and the limiting case in which observation locations coincide with the standard shift-rule points while the length-scale parameters are set to the exact frequency values. Under these conditions the GP posterior mean recovers the generalized PSR finite-difference expression exactly; the proof follows by direct substitution of the kernel into the GP mean formula and algebraic simplification to the shift-rule coefficients. revision: yes

  2. Referee: [§4 and §5] VQE objective functions are trigonometric polynomials whose frequencies are determined by the circuit and Hamiltonian. The assumption that standard stationary kernels yield accurate gradients and calibrated posterior variances from arbitrarily spaced (including past SGD) observation points is load-bearing for both the reuse claim and GradCoRe. The paper should include a dedicated analysis or diagnostic (e.g., posterior predictive checks or gradient error vs. distance from current θ) demonstrating that kernel misspecification does not produce biased gradients or miscalibrated uncertainties on representative VQE landscapes.

    Authors: We accept that a dedicated diagnostic is necessary. The revised §4 will contain a new subsection presenting (i) posterior predictive checks on the VQE landscapes used in the experiments and (ii) plots of gradient error versus Euclidean distance from the current θ, using both newly sampled points and historical SGD observations. These diagnostics will quantify any bias or miscalibration introduced by the chosen kernels on representative trigonometric objectives. revision: yes

  3. Referee: [§5] Table 2 and Figure 4 report speedups over SMO and standard PSR, but the experiments lack error bars on iteration counts or wall-clock time, details on the number of independent random seeds, and explicit reporting of the total number of circuit evaluations (including those used for GP hyperparameter fitting). Without these, it is impossible to determine whether the reported acceleration is statistically robust or specific to the chosen problem instances.

    Authors: We agree that the experimental reporting must be strengthened. In the revised version we will augment Table 2 and Figure 4 with error bars (standard deviation over runs), state that all metrics are averaged over 20 independent random seeds, and add a supplementary table that breaks down the total circuit evaluations into those used for GP hyperparameter fitting, initial observations, and the SGD steps themselves. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper introduces Bayesian PSR via Gaussian processes for VQE gradients, with the explicit claim that it reduces to generalized PSR in special cases serving as a consistency property rather than a definitional loop. No equations or steps in the provided text show a prediction reducing to a fitted input by construction, no self-citation chains are load-bearing for the central claims, and the reuse/GradCoRe features follow directly from standard GP posterior mechanics applied to the VQE setting. The smoothness/stationarity assumption is stated openly as a modeling choice, not smuggled in. Numerical experiments are presented as external validation. This matches the default expectation of a non-circular proposal grounded in prior PSR literature without self-referential reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies insufficient detail to enumerate free parameters, axioms, or invented entities; kernel choice and GP suitability are implicit but unspecified.

pith-pipeline@v0.9.0 · 5692 in / 1113 out tokens · 30524 ms · 2026-05-23T03:34:11.829347+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Bias Analysis and Regularization of Sequential Minimal Optimization in Variational Quantum Eigensolvers

    quant-ph 2026-05 unverdicted novelty 7.0

    Bias in SMO-VQE can be estimated without extra measurements; a regularization method that mimics error accumulation while preserving unbiased estimates improves performance across system sizes and Hamiltonians.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

    work page

  2. [2]

    Abraham, H. et al. Q iskit: A n open-source framework for quantum computing. Zenodo, 2019. doi:10.5281/zenodo.2562111

    work page doi:10.5281/zenodo.2562111 2019

  3. [3]

    Acharya, D

    Acharya, R., Abanin, D. A., et al. Quantum error correction below the surface code threshold. Nature, 2024. doi:10.1038/s41586-024-08449-y

    work page doi:10.1038/s41586-024-08449-y 2024

  4. [4]

    J., Nicoli, K., Wu, B., Elosegui, N., Pedrielli, S., Funcke, L., Jansen, K., Kuhn, S., and Nakajima, S

    Anders, C. J., Nicoli, K., Wu, B., Elosegui, N., Pedrielli, S., Funcke, L., Jansen, K., Kuhn, S., and Nakajima, S. Adaptive observation cost control for variational quantum eigensolvers. In Proceedings of 41st International Conference on Machine Learning (ICML2024), 2024. doi:10.5555/3692070.3692133

    work page doi:10.5555/3692070.3692133 2024

  5. [5]

    Bluvstein, S

    Bluvstein, D., Evered, S. J., Geim, A. A., Li, S. H., Zhou, H., Manovitz, T., Ebadi, S., Cain, M., Kalinowski, M., Hangleiter, D., et al. Logical quantum processor based on reconfigurable atom arrays. Nature, pp.\ 1--3, 2023. doi:10.1038/s41586-023-06927-3

    work page doi:10.1038/s41586-023-06927-3 2023

  6. [6]

    C., Endo, S., Huggins, W

    Cai, Z., Babbush, R., Benjamin, S. C., Endo, S., Huggins, W. J., Li, Y., McClean, J. R., and O'Brien, T. E. Quantum error mitigation. Rev. Mod. Phys., 95: 0 045005, Dec 2023. doi:10.1103/RevModPhys.95.045005

    work page doi:10.1103/revmodphys.95.045005 2023

  7. [7]

    M., Figgatt, C., Landsman, K

    Debnath, S., Linke, N. M., Figgatt, C., Landsman, K. A., Wright, K., and Monroe, C. Demonstration of a small programmable quantum computer with atomic qubits. Nature, 536 0 (7614): 0 63--66, 2016. doi:10.1038/nature18648

    work page doi:10.1038/nature18648 2016

  8. [8]

    A Tutorial on Bayesian Optimization

    Frazier, P. A tutorial on B ayesian optimization. ArXiv e-prints, 2018. doi:10.48550/arXiv.1807.02811

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1807.02811 2018

  9. [9]

    and Jansen, K

    Iannelli, G. and Jansen, K. Noisy B ayesian optimization for variational quantum eigensolvers. ArXiv e-prints, 2021. doi:10.48550/arXiv.2112.00426

    work page doi:10.48550/arxiv.2112.00426 2021

  10. [10]

    S., Christiansen, O., Yao, Y.-X., and Lanat\`a, N

    Jiang, T., Rogers, J., Frank, M. S., Christiansen, O., Yao, Y.-X., and Lanat\`a, N. Error mitigation in variational quantum eigensolvers using tailored probabilistic machine learning. Phys. Rev. Res., 6: 0 033069, Jul 2024. doi:10.1103/PhysRevResearch.6.033069

    work page doi:10.1103/physrevresearch.6.033069 2024

  11. [11]

    Kielpinski, D., Monroe, C., and Wineland, D. J. Architecture for a large-scale ion-trap quantum computer. Nature, 417 0 (6890): 0 709--711, 2002. doi:10.1038/nature00784

    work page doi:10.1038/nature00784 2002

  12. [12]

    S., Sitdikov, I., Salcedo, C., Seif, A., and Minev, Z

    Liao, H., Wang, D. S., Sitdikov, I., Salcedo, C., Seif, A., and Minev, Z. K. Machine learning for practical quantum error mitigation. Nature Machine Intelligence, 6 0 (12): 0 1478–1486, November 2024. ISSN 2522-5839. doi:10.1038/s42256-024-00927-2. URL http://dx.doi.org/10.1038/s42256-024-00927-2

    work page doi:10.1038/s42256-024-00927-2 2024

  13. [13]

    R., Romero, J., Babbush, R., et al

    McClean, J. R., Romero, J., Babbush, R., et al. The theory of variational hybrid quantum-classical algorithms. New Journal of Physics, 18 0 (2): 0 023023, 2016. doi:10.1088/1367-2630/18/2/023023

    work page doi:10.1088/1367-2630/18/2/023023 2016

  14. [14]

    Quantum circuit learning

    Mitarai, K., Negoro, M., Kitagawa, M., et al. Quantum circuit learning. Phys. Rev. A, 98: 0 032309, 2018. doi:10.1103/PhysRevA.98.032309

    work page doi:10.1103/physreva.98.032309 2018

  15. [15]

    M., Fujii, K., and Todo, S

    Nakanishi, K. M., Fujii, K., and Todo, S. Sequential minimal optimization for quantum-classical hybrid algorithms. Phys. Rev. Res., 2: 0 043158, 2020. doi:10.1103/PhysRevResearch.2.043158

    work page doi:10.1103/physrevresearch.2.043158 2020

  16. [16]

    A., Anders, C

    Nicoli, K. A., Anders, C. J., Funcke, L., Hartung, T., Jansen, K., Kuhn, S., M \"u ller, K.-R., Stornati, P., Kessel, P., and Nakajima, S. Physics-informed B ayesian optimization of variational quantum circuits. In Advances in Neural Information Processing Systems (NeurIPS2023), 2023 a

    work page 2023

  17. [17]

    A., Anders, C

    Nicoli, K. A., Anders, C. J., et al. EMICoRe : E xpected maximum improvement over confident regions. https://github.com/emicore/emicore, 2023 b

    work page 2023

  18. [18]

    A., Wagner, L., and Funcke, L

    Nicoli, K. A., Wagner, L., and Funcke, L. Machine-learning-enhanced optimization of noise-resilient variational quantum eigensolvers. ArXiv e-prints, 2025. doi:10.48550/arXiv.2501.17689

    work page doi:10.48550/arxiv.2501.17689 2025

  19. [19]

    A variational eigenvalue solver on a photonic quantum processor

    Peruzzo, A., McClean, J., Shadbolt, P., et al. A variational eigenvalue solver on a photonic quantum processor. Nature Communications, 5 0 (1): 0 4213, 2014. doi:10.1038/ncomms5213

    work page doi:10.1038/ncomms5213 2014

  20. [20]

    Sequential minimal optimization : A fast algorithm for training support vector machines

    Platt, J. Sequential minimal optimization : A fast algorithm for training support vector machines. Microsoft Research Technical Report, 1998

    work page 1998

  21. [21]

    Quantum computing in the NISQ era and beyond

    Preskill, J. Quantum computing in the NISQ era and beyond. Quantum , 2: 0 79, August 2018. doi:10.22331/q-2018-08-06-79

    work page internal anchor Pith review doi:10.22331/q-2018-08-06-79 2018

  22. [22]

    Rasmussen, C. E. and Williams, C. K. I. G aussian Processes for Machine Learning . MIT Press, Cambridge, MA, USA, 2006. doi:10.7551/mitpress/3206.001.0001

    work page doi:10.7551/mitpress/3206.001.0001 2006

  23. [23]

    Quantum error correction: A n introductory guide

    Roffe, J. Quantum error correction: A n introductory guide. Contemporary Physics, 60 0 (3): 0 226--245, 2019. doi:10.1080/00107514.2019.1667078

    work page doi:10.1080/00107514.2019.1667078 2019

  24. [24]

    Principles of Mathematical Analysis

    Rudin, W. Principles of Mathematical Analysis. McGraw-Hill, 1964. doi:10.1017/S0013091500008889

    work page doi:10.1017/s0013091500008889 1964

  25. [25]

    and Yamasaki, H

    Tamiya, S. and Yamasaki, H. Stochastic gradient line B ayesian optimization for efficient noise-robust optimization of parameterized quantum circuits. npj Quantum Information, 8 0 (1): 0 90, 2022. doi:10.1038/s41534-022-00592-6

    work page doi:10.1038/s41534-022-00592-6 2022

  26. [26]

    The variational quantum eigensolver: A review of methods and best practices

    Tilly, J., Chen, H., Cao, S., et al. The variational quantum eigensolver: A review of methods and best practices. Physics Reports, 986: 0 1--128, 2022. doi:https://doi.org/10.1016/j.physrep.2022.08.003

    work page doi:10.1016/j.physrep.2022.08.003 2022

  27. [27]

    Wierichs, D., Izaac, J., Wang, C., and Lin, C. Y.-Y. General parameter-shift rules for quantum gradients. Quantum , 6: 0 677, March 2022. ISSN 2521-327X. doi:10.22331/q-2022-03-30-677

    work page doi:10.22331/q-2022-03-30-677 2022