pith. sign in

arxiv: 2606.23262 · v1 · pith:IU7JNE5Wnew · submitted 2026-06-22 · 🪐 quant-ph

Tuning Quantum MPS

Anna Leonteva , Maxime Outteryck , Guido Masella This is my paper

Pith reviewed 2026-06-26 07:58 UTC · model grok-4.3

classification 🪐 quant-ph
keywords matrix product statesquantum circuit simulationhyperparameter optimizationCMA-ESmachine learning rankingcircuit features
0
0 comments X

The pith

A circuit-aware hybrid ranking model recovers a meaningful fraction of the gains from offline CMA-ES optimization of MPS simulator hyperparameters.

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

The paper sets out a two-stage process to select better hyperparameters for Matrix Product State simulations of quantum circuits instead of relying on defaults. Stage one runs CMA-ES optimization offline under a fidelity limit to create a database of circuit-configuration pairs and their performance. Stage two extracts static features from circuits that reflect MPS structural properties and trains a hybrid ranking model to suggest suitable configurations for new circuits. Evaluation across circuit families shows the model captures part of the optimization benefit, though results are stronger when circuits are validated by size than when entire families are held out. This matters for practical simulation because manual or default hyperparameter choices often leave performance on the table.

Core claim

Offline single-objective CMA-ES optimization under a fidelity constraint produces configurations that improve over defaults, with the size of the improvement depending on backend, circuit family, and scale. A hybrid ranking model trained on static circuit features recovers a meaningful fraction of those gains, performing better under size-based validation than under leave-one-family-out transfer, yet still falls short of the offline optimum.

What carries the argument

The circuit-aware hybrid ranking model, trained on a set of static circuit features chosen to capture MPS-relevant structural properties, which ranks and recommends hyperparameter configurations for new circuits.

If this is right

  • Offline CMA-ES optimization improves MPS performance over defaults, though the scale of improvement varies with backend, family, and circuit size.
  • The learned predictor captures part of the offline gain when applied to new circuits.
  • Size-based validation produces stronger predictor results than family-based transfer.
  • The predictor still trails the offline optimum, leaving a performance gap.

Where Pith is reading between the lines

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

  • The same feature-plus-ranking approach could be tested on other tensor-network simulators to see whether the static features transfer beyond MPS.
  • Embedding the trained ranker directly inside a simulator package would let users obtain improved settings without running separate optimization.
  • Closing the remaining gap to the offline optimum would require either richer circuit features or a shift from ranking to direct regression on expected performance.

Load-bearing premise

That the chosen static circuit features capture the MPS-relevant structural properties well enough for the hybrid ranking model to generalize across circuits.

What would settle it

If the ranking model recommends configurations whose simulated fidelities match or fall below those of default settings on a fresh set of circuits outside the training distribution, the claim that it recovers a meaningful fraction of the optimization gain would be refuted.

Figures

Figures reproduced from arXiv: 2606.23262 by Anna Leonteva, Guido Masella, Maxime Outteryck.

Figure 1
Figure 1. Figure 1: Speedup (ratio of default params. run time / optimized param. run time) by circuit family for Qiskit (blue) and MIMIQ (orange). Each box plot represents the distribution of speedup values across circuit instances with varying qubit counts. The median is shown by the central line, and the dashed horizontal line at 1 indicates no improvement. The vertical axis is shown on a logarithmic scale [PITH_FULL_IMAG… view at source ↗
Figure 2
Figure 2. Figure 2: Prediction speedup (ratio of default to predicted run time) by circuit family for Qiskit (blue) and MIMIQ (orange) under one-family-out validation. The black line shows the median, boxes the interquartile range, and whiskers the variability across instances. The dashed line at 1 indicates no improvement. The vertical axis is logarithmic [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prediction speedup (ratio of default to predicted run time) by circuit family for Qiskit (blue) and MIMIQ (orange) under size-based validation. The black line shows the median, boxes the interquartile range, and whiskers the variability across instances. The dashed line at 1 indicates no improvement. The vertical axis is shown on a logarithmic scale [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
read the original abstract

Matrix Product State (MPS) methods are among the most effective approaches for the classical simulation of quantum circuits, but their practical performance depends strongly on simulator hyperparameters, and default settings are often suboptimal. In this work, we propose a two-stage framework for automatic hyperparameter selection for quantum MPS simulation. In the first stage, we perform offline single-objective CMA-ES optimization under a fidelity constraint and construct a database of circuit--configuration--performance evaluations. In the second stage, we define a set of static circuit features designed to capture MPS-relevant structural properties and train a circuit-aware hybrid ranking model to recommend configurations for different quantum circuits. We evaluate the approach on multiple scalable circuit families using leave-one-family-out and size-based validation. The results show that offline optimization often improves over default settings, although the magnitude of the gain depends strongly on the backend, circuit family, and circuit scale. The learned predictor recovers a meaningful fraction of this gain, with better performance under size-based validation than under family-based transfer, but generally remains below the offline optimum.

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

2 major / 2 minor

Summary. The paper proposes a two-stage framework for automatic hyperparameter tuning in Matrix Product State (MPS) simulation of quantum circuits. Stage one performs offline single-objective CMA-ES optimization under a fidelity constraint to build a database of circuit-configuration-performance triples. Stage two extracts static circuit features intended to capture MPS-relevant structure and trains a circuit-aware hybrid ranking model to recommend configurations. Evaluation on multiple scalable circuit families uses leave-one-family-out and size-based validation splits; the abstract states that offline optimization improves over defaults (with gains varying by backend, family, and scale) and that the learned predictor recovers a meaningful fraction of this gain, performing better under size-based validation than family-based transfer but remaining below the offline optimum.

Significance. If the central claim holds, the work offers a practical, data-driven method to reduce manual hyperparameter tuning in a widely used classical simulation technique for quantum circuits. The two-stage separation of expensive offline optimization from fast online prediction is a reasonable engineering contribution, and the use of both family-based and size-based validation provides some evidence on generalization. Reproducible code or an open database would strengthen the result, but none is mentioned in the provided material.

major comments (2)
  1. [Abstract] Abstract: the claim that 'the learned predictor recovers a meaningful fraction of this gain' is stated without any numerical values, confidence intervals, or description of how the fraction is computed; this absence makes it impossible to judge whether the recovery is practically useful or statistically reliable.
  2. [Evaluation] Evaluation (implied by validation description): the stronger performance under size-based validation versus leave-one-family-out is consistent with the risk that static circuit features primarily encode coarse size or family correlations rather than the dynamic entanglement or contraction properties that actually determine optimal MPS hyperparameters; without an ablation that isolates feature predictive power or reports per-circuit error analysis, the central generalization claim rests on an untested assumption.
minor comments (2)
  1. [Abstract] The abstract would be clearer if it listed the specific circuit families, the range of circuit sizes, and the backends used.
  2. Notation for the hybrid ranking model and the precise definition of the static circuit features should be introduced earlier and with explicit formulas.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive suggestions. We address the major comments below, indicating planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'the learned predictor recovers a meaningful fraction of this gain' is stated without any numerical values, confidence intervals, or description of how the fraction is computed; this absence makes it impossible to judge whether the recovery is practically useful or statistically reliable.

    Authors: We agree with this observation. The abstract will be revised to include quantitative results: specifically, the average fraction of the offline gain recovered by the predictor (computed as (predictor gain / offline gain)), with values and ranges across families and scales, plus reference to standard deviations reported in the results section. This will allow readers to assess practical utility directly. revision: yes

  2. Referee: [Evaluation] Evaluation (implied by validation description): the stronger performance under size-based validation versus leave-one-family-out is consistent with the risk that static circuit features primarily encode coarse size or family correlations rather than the dynamic entanglement or contraction properties that actually determine optimal MPS hyperparameters; without an ablation that isolates feature predictive power or reports per-circuit error analysis, the central generalization claim rests on an untested assumption.

    Authors: We acknowledge the potential limitation in feature analysis. Our static features include several designed to proxy dynamic properties, such as estimated entanglement entropy and contraction cost metrics. The superior size-based validation performance is expected as it tests within-family scaling. To address the concern, we will include an ablation study removing size-correlated features and report the resulting drop in ranking accuracy. We will also add per-circuit performance breakdowns in the supplementary material to provide finer-grained analysis of generalization. revision: yes

Circularity Check

0 steps flagged

No circularity: two-stage framework uses independent optimization data to train a separate predictor

full rationale

The derivation consists of an offline CMA-ES stage that populates an external database of circuit-configuration-performance triples, followed by training a hybrid ranking model on independently defined static circuit features. Neither stage reduces to the other by construction: the model's recommendations are learned from the database rather than being a renaming or direct reuse of the optimization outputs, and no self-citation, uniqueness theorem, or ansatz is invoked to force the result. The reported recovery of a fraction of the offline gain is therefore an empirical outcome of the learned mapping, not a definitional identity. This is the most common honest finding for a method that separates data generation from model training and evaluates via cross-validation splits.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities. The approach rests on standard CMA-ES and supervised ranking techniques whose details are not provided.

pith-pipeline@v0.9.1-grok · 5701 in / 1086 out tokens · 25279 ms · 2026-06-26T07:58:31.768169+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

29 extracted references · 4 linked inside Pith

  1. [1]

    X., Van Den Berg, E., Rosenblatt, S.,

    Kim, Y., Eddins, A., Anand, S., Wei, K. X., Van Den Berg, E., Rosenblatt, S., ... ,Kandala, A. Evidence for the utility of quantum computing before fault tolerance. Nature, 618(7965), 500-505, (2023)

    2023

  2. [2]

    A., Azhar, M

    Cicero, A., Maleki, M. A., Azhar, M. W., Kockum, A. F., Trancoso, P. Simulation of quantum computers: Review and acceleration opportunities. ACM Transactions on Quantum Computing, 7(1), 1-35, (2025)

    2025

  3. [3]

    Efficient classical simulation of slightly entangled quantum computations

    Vidal, G. Efficient classical simulation of slightly entangled quantum computations. Physical review letters, 91(14), 147902, (2003)

    2003

  4. [4]

    L., Shi, Y

    Markov, I. L., Shi, Y. Simulating quantum computation by contracting tensor net- works. SIAM Journal on Computing, 38(3), 963-981, (2008)

    2008

  5. [5]

    A practical introduction to tensor networks: Matrix product states and projected entangled pair states

    Orús, R. A practical introduction to tensor networks: Matrix product states and projected entangled pair states. Annals of physics, 349, 117-158, (2014)

    2014

  6. [6]

    ACM Transactions on Quantum Computing, 7(2), 1-29, (2026)

    Leonteva,A.,Masella,G.,Outteryck,M.,PeiroOrioli,A.,Whitlock,S.Comparative benchmarking of utility-scale quantum emulators. ACM Transactions on Quantum Computing, 7(2), 1-29, (2026)

    2026

  7. [7]

    S., Hill, C

    Wang, D. S., Hill, C. D., Hollenberg, L. C. Simulations of Shor’s algorithm using matrix product states. Quantum Information Processing, 16(7), 176, (2017)

    2017

  8. [8]

    D., Hollenberg, L

    Dang, A., Hill, C. D., Hollenberg, L. C. Optimising matrix product state simulations of Shor’s algorithm. Quantum, 3, 116, (2019)

    2019

  9. [9]

    Hyper-optimized tensor network contraction

    Gray, J., Kourtis, S. Hyper-optimized tensor network contraction. Quantum, 5, 410, (2021)

    2021

  10. [10]

    Riemannian optimization of isometric tensor networks

    Hauru, M., Van Damme, M., Haegeman, J. Riemannian optimization of isometric tensor networks. Scipost physics, 10(2), 040, (2021)

    2021

  11. [11]

    The CMA evolution strategy: a comparing review

    Hansen, N. The CMA evolution strategy: a comparing review. Towards a new evolutionary computation: Advances in the estimation of distribution algorithms, 75-102, (2006). Tuning Quantum MPS 15

    2006

  12. [12]

    CMA-ES for dis- crete and mixed-variable optimization on sets of points

    Uchida, K., Hamano, R., Nomura, M., Saito, S., Shirakawa, S. CMA-ES for dis- crete and mixed-variable optimization on sets of points. In International Conference on Parallel Problem Solving from Nature (pp. 236-251). Cham: Springer Nature Switzerland, (2024, September)

    2024

  13. [13]

    Quantum amplitude amplification and estimation

    Brassard, G., Hoyer, P., Mosca, M., Tapp, A. Quantum amplitude amplification and estimation. arXiv preprint quant-ph/0005055, (2000)

    Pith/arXiv arXiv 2000

  14. [14]

    A., Chuang, I

    Nielsen, M. A., Chuang, I. L. Quantum computation and quantum information. Cambridge university press, (2010)

    2010

  15. [15]

    Kitaev, A. Y. Quantum measurements and the Abelian stabilizer problem. arXiv preprint quant-ph/9511026, (1995)

    Pith/arXiv arXiv 1995

  16. [16]

    M., Horne, M

    Greenberger, D. M., Horne, M. A., Zeilinger, A. Going beyond Bell’s theorem. In Bell’s theorem, quantum theory and conceptions of the universe (pp. 69-72). Dordrecht: Springer Netherlands, (1989)

    1989

  17. [17]

    Dür, W., Vidal, G., Cirac, J. I. Three qubits can be entangled in two inequivalent ways. Physical Review A, 62(6), 062314, (2000)

    2000

  18. [18]

    H., Zhou, X

    Peruzzo, A., McClean, J., Shadbolt, P., Yung, M. H., Zhou, X. Q., Love, P. J., ... & O’brien, J. L. A variational eigenvalue solver on a photonic quantum processor. Nature communications, 5(1), 4213, (2014)

    2014

  19. [19]

    M., Cleve, R., Deotto, E., Farhi, E., Gutmann, S., Spielman, D

    Childs, A. M., Cleve, R., Deotto, E., Farhi, E., Gutmann, S., Spielman, D. A. Exponential algorithmic speedup by a quantum walk. In Proceedings of the thirty- fifth annual ACM symposium on Theory of computing (pp. 59-68), (2003)

    2003

  20. [20]

    C., Barends, R.,

    Arute, F., Arya, K., Babbush, R., Bacon, D., Bardin, J. C., Barends, R., ... & Mar- tinis, J. M. Quantum supremacy using a programmable superconducting processor. nature, 574(7779), 505-510, (2019)

    2019

  21. [21]

    The quest for a quantum neural network

    Schuld, M., Sinayskiy, I., Petruccione, F. The quest for a quantum neural network. Quantum Information Processing, 13(11), 2567-2586, (2014)

    2014

  22. [22]

    Quantum Emulator Benchmarking Made Easy

    Leonteva, A., Outteryck, M., Masella, G. Quantum Emulator Benchmarking Made Easy. In International Conference on Quantum Engineering Sciences and Technolo- gies for Industry and Services (pp. 280-288). Cham: Springer Nature Switzerland, (2025, December)

    2025

  23. [23]

    MQT Bench: Benchmarking software and design automation tools for quantum computing

    Quetschlich, N., Burgholzer, L., Wille, R. MQT Bench: Benchmarking software and design automation tools for quantum computing. Quantum, 7, 1062, (2023)

    2023

  24. [24]

    J., Lishman, J., Gacon, J., ..., Gambetta, J

    Javadi-Abhari, A., Treinish, M., Krsulich, K., Wood, C. J., Lishman, J., Gacon, J., ..., Gambetta, J. M. Quantum computing with Qiskit. arXiv preprint arXiv:2405.08810, (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Open quantum assembly language

    Cross, Andrew W., et al. "Open quantum assembly language." arXiv preprint arXiv:1707.03429, (2017)

    Pith/arXiv arXiv 2017

  26. [26]

    Measuring the capabilities of quantum computers

    Proctor, T., Rudinger, K., Young, K., Nielsen, E.,Blume-Kohout, R. Measuring the capabilities of quantum computers. Nature Physics, 18(1), 75-79, (2022)

    2022

  27. [27]

    Liu, T. Y. Learning to rank for information retrieval. Foundations and Trends® in Information Retrieval, 3(3), 225-331, (2009)

    2009

  28. [28]

    Random forests

    Breiman, L. Random forests. Machine learning, 45(1), 5-32, (2001)

    2001

  29. [29]

    Extremely randomized trees

    Geurts, P., Ernst, D., Wehenkel, L. Extremely randomized trees. Machine learning, 63(1), 3-42, (2006)

    2006