Pith. sign in

REVIEW 2 major objections 4 minor 42 references

Two scalable scores tell which fixed quantum reservoirs will actually learn: one measures how Haar-like their outputs are, the other how many usable feature directions reach the classical readout.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 02:56 UTC pith:6FNPZI56

load-bearing objection Solid, usable two-axis diagnostic for quantum reservoirs: multi-basis ORS for expressivity plus R_eff for coverage, with hardware-compatible noise correction that actually works on IBM data. the 2 major comments →

arxiv 2607.09445 v1 pith:6FNPZI56 submitted 2026-07-10 quant-ph

Diagnosing quantum reservoirs at scale based on expressivity and coverage

Laia Domingo , Oriol Ball\'o-Gimbernat , Fernando Vilari\~no This is my paper
classification quant-ph
keywords quantum reservoir computingquantum extreme learning machinesexpressivityorder statisticseffective rankHaar measuredepolarizing noisequantum machine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Quantum reservoirs process data with fixed random quantum dynamics and a cheap classical linear readout, so performance hinges on which random family is chosen. Existing quality checks either explode with system size or only work for special models. This paper gives a two-axis diagnosis that stays practical as the number of qubits grows and works on real hardware. The first axis is an order-statistics (ORS) expressivity score: keep only the few largest output probabilities of each reservoir instance and compare them to the known analytic distribution for Haar-random states; a closed-form correction removes the trivial effect of depolarizing noise. The second axis is the effective rank of the feature matrix actually seen by the readout, which counts how many independent, input-dependent directions are available for learning. On both synthetic and real extreme-learning and reservoir-computing tasks, multi-basis ORS ranks reservoir families by intrinsic expressivity while the effective rank shows when that expressivity becomes usable predictive power. The same noise-corrected ORS gap continues to separate expressive from restricted families under simulated noise and on IBM hardware.

Core claim

A reservoir family is useful when two complementary diagnostics align: its multi-basis order-statistics gap to Haar is near zero (intrinsic expressivity) and the effective rank of the measured feature matrix is large (task-dependent coverage). ORS never needs the full output distribution, is independent of Hilbert-space dimension for fixed top-K ranks, and admits an exact depolarizing correction that remains informative on real devices.

What carries the argument

The order-statistics (ORS) expressivity gap: the harmonic-weighted log-likelihood of the K largest output probabilities of a reservoir ensemble, measured against the closed-form Haar order-statistics density (with optional multi-basis average and depolarizing correction). It is paired with the participation-ratio effective rank of the column-centred feature matrix seen by the linear readout.

Load-bearing premise

The hardware correction treats device noise as a single global depolarizing channel whose fidelity can be estimated from gate and readout calibration data; if real noise is strongly coherent, correlated or non-Markovian, the corrected gap can mis-rank families.

What would settle it

Run the same G3 versus G1 or commuting versus non-commuting IQP families on a device whose noise is known to be highly structured (or under a simulated non-depolarizing channel) and check whether the noise-corrected multi-basis ORS gap still correctly ranks the families while the effective-rank / test-error relationship collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a two-axis diagnostic for quantum reservoirs (QRC and QELM). The first axis is a task-independent order-statistics (ORS) expressivity score that compares only the top-K output probabilities of a reservoir ensemble to an analytical Haar order-statistics baseline (Eqs. 3–5), with a multi-basis extension (Eq. 15) and a closed-form global-depolarizing correction (Eqs. 6–11). The second axis is the task-dependent effective rank R_eff of the feature matrix (Eq. 16). Small-system validation against KL fidelity divergence, level-spacing ratio, and Krylov complexity (Fig. 1) shows a consistent transition to Haar-like behavior. Noise-corrected ORS remains discriminative under simulated depolarizing noise (Table I) and on IBM Aachen hardware (Table II). Across synthetic Fourier/NARMA and real LiH/EMSIG benchmarks (Figs. 2–5), multi-basis ORS ranks reservoir families by intrinsic expressivity while R_eff indicates when that expressivity becomes usable for the linear readout.

Significance. If the results hold, the work supplies a practical, scalable alternative to full-distribution or full-unitary diagnostics that become unusable as Hilbert-space dimension grows. The analytical Haar baseline, cost independence of D, multi-basis extension that exposes commuting IQP structure, and closed-form depolarizing correction are concrete technical contributions that make the diagnostic usable on near-term hardware. Explicit cross-checks against KL, spectral chaos, and Krylov complexity, plus both synthetic and real QELM/QRC tasks, strengthen the claim that the two axes jointly explain performance. The framework is architecture-agnostic and therefore useful for comparing gate-based and analog reservoirs.

major comments (2)
  1. Sec. IIC and Table II: the hardware demonstration relies on an effective fidelity f_Bq estimated from calibration data under a global-depolarizing model (Eqs. 12–13). The paper correctly notes that real device noise is not exactly global depolarizing. Because Table II is presented as evidence that ORS remains informative on hardware, a short quantitative check of residual sensitivity (e.g., comparison of corrected vs uncorrected gaps, or a simple coherent-error simulation) would make the hardware claim more robust; without it the hardware result remains supportive but secondary.
  2. Sec. III C and Figs. 2–3: for the non-commuting D2 extensions the multi-basis ORS approaches the Haar reference while R_eff and MSE remain suboptimal. The text attributes this to residual correlations not captured by top-K ranks and to memory effects in QRC. A brief ablation (larger K, more bases, or a simple memory-capacity diagnostic) would clarify whether the observed decoupling is fundamental or an artifact of the chosen K and B; the central claim that the two axes are complementary is otherwise well supported.
minor comments (4)
  1. Appendix A: the large-D asymptotic form of Pk(x) is used throughout; a short statement of the n range where the approximation remains accurate (or a finite-D correction) would help readers applying ORS at small n.
  2. Fig. 1: the multiple right-hand axes for different K make visual comparison slightly crowded; a single normalized gap or an inset could improve readability.
  3. Sec. IIF: the mean-degree parameter d of the Erdős–Rényi interaction graph is introduced without an explicit formula for the expected number of ZZ terms; a one-line clarification would aid reproducibility.
  4. Notation: GORS, G(f)_ORS and Gmb(B) are used interchangeably in places; a consistent symbol table or early definition would reduce minor ambiguity.

Circularity Check

0 steps flagged

No significant circularity: ORS and R_eff are independently defined diagnostics whose correlation with held-out performance is empirical, not definitional.

full rationale

The paper defines the ORS expressivity score solely from the top-K ordered output probabilities of a reservoir ensemble versus an analytical Haar order-statistics baseline (Eqs. 3–5, Appendix A); this construction never references any learning task, feature matrix, or performance metric. The multi-basis extension (Eq. 15) and the closed-form depolarizing correction (Eqs. 6–11) likewise depend only on measurement statistics and an effective fidelity parameter. Coverage is quantified separately by the ordinary participation-ratio effective rank R_eff of the column-centred feature matrix (Eq. 16). Predictive performance is then measured by held-out MSE or R^{2} on synthetic and real QELM/QRC benchmarks (Figs. 2–5). The claimed joint predictive power is therefore an empirical correlation, not a tautology; the paper itself reports partial decoupling cases (e.g., D2,XZ / D2,XZY near Haar in multi-basis ORS yet sub-maximal R_eff and MSE). Small-system validation against KL divergence, level-spacing ratios and Krylov complexity (Fig. 1) supplies independent external checks. Self-citations (e.g., to the authors’ prior LiH dataset construction or to Micklitz’s ORS paper) are used only for data or for the original ORS idea; they do not force the present hierarchy or performance claims. No fitted parameter is renamed a prediction, no uniqueness theorem is imported from the authors, and no definitional loop exists between the diagnostics and the reported results.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard quantum information (Haar measure, depolarizing channel, Pauli observables) plus a small set of modeling choices (large-D order-statistics asymptotics, global-noise approximation, fixed K and B) and the adaptation of an external ORS construction. No new physical entities are postulated; free parameters are diagnostic hyperparameters, not fitted performance constants.

free parameters (5)
  • K (number of retained top ranks) = 4
    Fixed to K=4 after a small-system sweep over {1,4,8}; enters every ORS evaluation and the Haar baseline sum.
  • B (number of random local Pauli bases) = 50
    Set to 50 for multi-basis ORS on diagonal families; controls the strictness of the basis-robustness test.
  • n_obs (size of random Pauli observable pool) = 1000
    Fixed at 1000 for synthetic tasks to saturate R_eff; chosen by hand rather than derived.
  • M (ensemble size) = 100
    Number of independent reservoir instances averaged for each diagnostic point; set to 100.
  • effective hardware fidelity f_Bq = backend-dependent (0.86–0.94 in Table II)
    Estimated from backend 1q/2q/readout errors via exp(-E); used to shift the Haar baseline on IBM runs and is therefore a data-dependent correction parameter.
axioms (4)
  • domain assumption Large-D asymptotic form of Haar order-statistics densities P_k(x) (Eq. 3 / App. A) is accurate enough for the system sizes considered.
    Used to obtain the closed-form Haar baseline without sampling Haar states; validity improves with n but is assumed already at n=6.
  • domain assumption Dominant noise can be approximated by a global depolarizing channel with a single fidelity parameter f.
    Underpins the invariance claim (Eq. 11) and the hardware-aware correction; real devices deviate from this model.
  • domain assumption Haar-random measurement statistics constitute the appropriate maximal-expressivity reference for reservoir ensembles.
    Standard in quantum expressivity literature; adopted without re-derivation.
  • standard math Participation ratio of singular values of the centered feature matrix is a faithful measure of usable coverage for linear readout.
    Ordinary effective-rank definition; no new derivation required.
invented entities (2)
  • Multi-basis order-statistics (ORS) gap G_mb(B) independent evidence
    purpose: Average ORS deviation over random local Pauli bases to expose basis-dependent structure (e.g., commuting IQP families).
    Defined in Sec. IID; independent evidence is the empirical separation of D2 vs D2,XZ/XZY that single-basis ORS would miss.
  • Noise-corrected ORS gap for reservoir ensembles independent evidence
    purpose: Make the expressivity score comparable across different effective fidelities and usable on hardware.
    Adaptation of Micklitz’s single-circuit construction to ensembles plus the hardware fidelity estimator; falsifiable by checking invariance under controlled depolarizing noise (verified in simulation).

pith-pipeline@v1.1.0-grok45 · 19508 in / 3235 out tokens · 38881 ms · 2026-07-13T02:56:47.512469+00:00 · methodology

0 comments
read the original abstract

Quantum reservoirs offer a hardware-friendly route to quantum machine learning, replacing trainable circuits with fixed random dynamics and a classical readout. Because the reservoir is not optimized, performance depends entirely on the choice of reservoir family, yet existing diagnostics demand resources that grow exponentially with system size. We introduce a scalable, hardware-agnostic framework built on two complementary quantities. The first is a task-independent order-statistics (ORS) expressivity score, which compares only the largest output probabilities of a reservoir ensemble against an analytical Haar baseline. It never reconstructs the full output distribution, is cost-independent of Hilbert-space dimension, and admits a closed-form depolarizing noise correction, making it directly usable on hardware. The second is the task-dependent effective rank $R_{\mathrm{eff}}$ of the feature matrix, which measures how much input-dependent information reaches the readout. We validate the ORS score against established complexity diagnostics and confirm it remains informative under simulated noise and on IBM quantum hardware. Across synthetic and real quantum extreme learning machine and quantum reservoir computing benchmarks, ORS captures the intrinsic expressivity hierarchy of reservoir families while $R_{\mathrm{eff}}$ determines when that expressivity becomes usable predictive information.

Figures

Figures reproduced from arXiv: 2607.09445 by Fernando Vilari\~no, Laia Domingo, Oriol Ball\'o-Gimbernat.

Figure 1
Figure 1. Figure 1: Comparison of expressivity and dynamical-complexity diagnostics for the [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Quantum extreme learning machine results on the synthetic Fourier regression task for [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Same as Fig. 2 for the quantum reservoir computing task on the synthetic NARMA dataset for [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: QELM results on the LiH excited-energy prediction task. The expressivity–coverage alignment observed in the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: QRC results on the EMSIG residential energy forecasting task. The test metric is [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

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

42 extracted references · 4 linked inside Pith

  1. [1]

    For fixed reservoir instances and exact probabilities, Eq

    A key consequence of this correction is that, under an exact global depolarizing model, the corrected ORS gap is invariant under the fidelity parameter. For fixed reservoir instances and exact probabilities, Eq. (8) shifts both the reservoir score and the Haar reference by the same amount, so that Λ (f) −Λ (f) Haar = Λ−Λ Haar.(11) This invariance is impor...

    2019

  2. [2]

    Jaeger and H

    H. Jaeger and H. Haas, Science304, 78 (2004)

    2004

  3. [3]

    Maass and H

    W. Maass and H. Markram, Journal of Computer and System Sciences69, 593 (2004)

    2004

  4. [4]

    Tanaka, T

    G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Networks115, 100 (2019)

    2019

  5. [5]

    Nakajima, Japanese Journal of Applied Physics59, 060501 (2020)

    K. Nakajima, Japanese Journal of Applied Physics59, 060501 (2020)

    2020

  6. [6]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications9(2018)

    2018

  7. [7]

    Cerezo, A.Sone, T

    M. Cerezo, A.Sone, T. Volkoff, L.Cincio,andP. J.Coles, Nature Communications12(2021)

    2021

  8. [8]

    Larocca, S

    M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Bia- monte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo, Nature Reviews Physics7, 174 (2025), arXiv:2405.00781 [quant-ph]

    Pith/arXiv arXiv 2025

  9. [9]

    E. R. Anschuetz and B. T. Kiani, Nature Communica- tions13(2022)

    2022

  10. [10]

    Thanasilp, S

    S. Thanasilp, S. Wang, N. A. Nghiem, P. Coles, and M. Cerezo, Quantum Machine Intelligence5(2023)

    2023

  11. [11]

    Fujii and K

    K. Fujii and K. Nakajima, Phys. Rev. Applied8, 024030 (2017)

    2017

  12. [12]

    Mujal, R

    P. Mujal, R. M.-P. na, J. Nokkala, J. García-Beni, G. L. Giorgi, M. C. Soriano, and R. Zambrini, Advanced Quan- tum Technologies4(2021)

    2021

  13. [13]

    Domingo, G

    L. Domingo, G. G. Carlo, and F. Borondo, Scientific Re- ports13, 8790 (2023)

    2023

  14. [14]

    Sannia, F

    A. Sannia, F. Tacchino, I. Tavernelli, G. L. Giorgi, and R. Zambrini, npj Quantum Information10(2024)

    2024

  15. [15]

    J. Chen, H. I. Nurdin, and N. Yamamoto, Phys. Rev. Applied14(2020)

    2020

  16. [16]

    Kobayashi, K

    K. Kobayashi, K. Fujii, and N. Yamamoto, PRX Quan- tum5(2024)

    2024

  17. [17]

    Y. Hou, J. Hua, Z. Wu, W. Xia, Y. Chen, X. Li, Z. Li, X. Peng, and J. Du, Phys. Rev. Lett.136, 120602 (2026)

    2026

  18. [18]

    Xiong, G

    W. Xiong, G. Facelli, M. Sahebi, O. Agnel, T. Chotibut, S. Thanasilp, and Z. Holmes, Quantum Machine Intelli- gence7, 20 (2025), arXiv:2312.15124 [quant-ph]

    Pith/arXiv arXiv 2025

  19. [19]

    Senanian, S

    A. Senanian, S. Prabhu, V. Kremenetski, S. Roy, Y. Cao, J. Kline, T. Onodera, L. G. Wright, X. Wu, V. Fatemi, 11 and P. L. McMahon, Nature Communications15(2024)

    2024

  20. [20]

    Nokkala, R

    J. Nokkala, R. M.-P. na, G. L. Giorgi, V. Parigi, M. C. Soriano, and R. Zambrini, Communications Physics4 (2021)

    2021

  21. [21]

    L. C. G. Govia, G. J. Ribeill, G. E. Rowlands, H. K. Krovi, and T. A. Ohki, Phys. Rev. Research3(2021)

    2021

  22. [22]

    Domingo, G

    L. Domingo, G. Carlo, and F. Borondo, Phys. Rev. E 106, L043301 (2022)

    2022

  23. [23]

    Kawai and Y

    H. Kawai and Y. Nakagawa, Mach. Learn.: Sci. Technol. 1(2020)

    2020

  24. [24]

    Domingo, M

    L. Domingo, M. Djukic, C. Johnson, and F. Borondo, Scientific Reports13, 17951 (2023)

    2023

  25. [25]

    R. M.-P. na, G. L. Giorgi, J. Nokkala, M. C. Soriano, and R. Zambrini, Phys. Rev. Lett.127, 100502 (2021)

    2021

  26. [26]

    J. I. Latorre and M. A. Martín-Delgado, Phys. Rev. A 66, 022305 (2002)

    2002

  27. [27]

    R. O. Vallejos, F. de Melo, and G. G. Carlo, Phys. Rev. A104, 012602 (2021)

    2021

  28. [28]

    Shaffer, C

    D. Shaffer, C. Chamon, A. Hamma, and E. R. Mucci- olo, Journal of Statistical Mechanics: Theory and Ex- periment2014, P12007 (2014)

    2014

  29. [29]

    Domingo, F

    L. Domingo, F. Borondo, G. Scialchi, A. J. Roncaglia, G. G. Carlo, and D. A. Wisniacki, Phys. Rev. A110, 022446 (2024)

    2024

  30. [30]

    S. Sim, P. D. Johnson, and A. Aspuru-Guzik, Advanced Quantum Technologies2, 1900070 (2019)

    2019

  31. [31]

    Micklitz, Simulation-free fidelity estimation via quan- tum output order statistics (2025), arXiv:2510.13026 [quant-ph]

    T. Micklitz, Simulation-free fidelity estimation via quan- tum output order statistics (2025), arXiv:2510.13026 [quant-ph]

    arXiv 2025

  32. [32]

    Sannia, G

    A. Sannia, G. L. Giorgi, and R. Zambrini, Exponential concentration and symmetries in quantum reservoir com- puting (2025), arXiv:2505.10062 [quant-ph]

    Pith/arXiv arXiv 2025

  33. [33]

    S.Thanasilp, S.Wang, M.Cerezo,andZ.Holmes,Nature Communications15(2024)

    2024

  34. [34]

    Gottesman, inGroup22: Proceedings of the XXII In- ternational Colloquium on Group Theoretical Methods in Physics, edited by S

    D. Gottesman, inGroup22: Proceedings of the XXII In- ternational Colloquium on Group Theoretical Methods in Physics, edited by S. P. Corney, R. Delbourgo, and P. D. Jarvis (International Press, Cambridge, MA, 1999) pp. 32–43, arXiv:quant-ph/9807006 [quant-ph]

    Pith/arXiv arXiv 1999

  35. [35]

    Clark, R

    S. Clark, R. Jozsa, and N. Linden, Quantum Information & Computation8, 106 (2008)

    2008

  36. [36]

    M. V. D. Nest, Quantum Information & Computation 10, 258 (2010)

    2010

  37. [37]

    Jozsa and M

    R. Jozsa and M. V. D. Nest, Quantum Information & Computation14, 633 (2014)

    2014

  38. [38]

    D. E. Koh, Quantum Information & Computation17, 262 (2017)

    2017

  39. [39]

    M. J. Bremner, R. Jozsa, and D. J. Shepherd, Proc. R. Soc. A467, 459 (2011)

    2011

  40. [40]

    Ni and M

    X. Ni and M. V. D. Nest, Quantum Information & Com- putation13, 54 (2013)

    2013

  41. [41]

    Fujii and T

    K. Fujii and T. Morimae, New Journal of Physics19, 033003 (2017)

    2017

  42. [42]

    Brucke, S

    K. Brucke, S. Schmitz, D. Köglmayr, S. Baur, C. Räth, E. Ansari, and P. Klement, Energy and Buildings314, 114236 (2024). Appendix A: Analytical Haar baseline for ORS The ORS diagnostic compares the log-likelihood of the observed top-ranked probabilities with its Haar expecta- tion. This Haar reference can be evaluated analytically in the large-Dapproximat...

    2024