pith. sign in

arxiv: 2605.29071 · v1 · pith:VO3542A3new · submitted 2026-05-27 · 🪐 quant-ph

A hidden bottleneck in classical and quantum linear reservoir computing

Johannes Nokkala , Federico Centrone , Francesco Arzani This is my paper

Pith reviewed 2026-06-29 11:23 UTC · model grok-4.3

classification 🪐 quant-ph
keywords reservoir computinglinear dynamicsquantum reservoir computinginformation capacitynon-Gaussian processingcontinuous-variable systemscomputational bottleneck
0
0 comments X

The pith

Linear reservoir dynamics redistribute features but cannot create new fixed-delay expressive power.

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

The paper shows that when features evolve linearly inside a reservoir and the output is produced by linear readout with bias, the capacity available at any fixed delay cannot exceed what the preprocessed input already supplies. Linear dynamics inside the reservoir are therefore limited to redistributing existing features rather than generating additional expressive power at any chosen delay. This per-delay limit remains invisible to global capacity scores because contributions from separate delays can still add together. The same bound holds for covariance-based continuous-variable quantum reservoirs in the Gaussian regime, while single-photon operations are shown to exceed it.

Core claim

When the measured features evolve linearly in the reservoir and the output is formed by linear readout with bias, the capacity available at any fixed delay is limited by what is already present in the preprocessed input. Linear reservoir dynamics can therefore redistribute features, but cannot create new fixed-delay expressive power on their own. This limitation is hidden by global capacity measures, since contributions from different delays can accumulate even when each individual delay is strongly constrained.

What carries the argument

The fixed-delay capacity bound that arises under linear feature evolution and linear readout with bias, which caps expressive power at the level already present in the preprocessed input.

Load-bearing premise

The measured features evolve linearly in the reservoir and the output is formed by linear readout with bias.

What would settle it

An experiment or calculation in which a linear reservoir with linear readout achieves strictly higher capacity at some fixed delay than exists in the preprocessed input would falsify the bound.

Figures

Figures reproduced from arXiv: 2605.29071 by Federico Centrone, Francesco Arzani, Johannes Nokkala.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Witnessing the non-Gaussianity of the reservoir with [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

We identify a hidden bottleneck in the information processing capacity of linear reservoir computers. When the measured features evolve linearly in the reservoir and the output is formed by linear readout with bias, we show that the capacity available at any fixed delay is limited by what is already present in the preprocessed input. Linear reservoir dynamics can therefore redistribute features, but cannot create new fixed-delay expressive power on their own. This limitation is hidden by global capacity measures, since contributions from different delays can accumulate even when each individual delay is strongly constrained. As an experimentally important realization of this general result, we derive the corresponding Gaussian limit for covariance-based continuous-variable quantum reservoirs. Numerical experiments show that experimentally accessible single-photon operations surpass this limit, establishing them as a genuine resource for quantum reservoir computing. The resulting excess capacity also provides an operational witness of non-Gaussian processing in black-box continuous-variable systems under minimal assumptions.

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

0 major / 2 minor

Summary. The manuscript identifies a hidden bottleneck in linear reservoir computing: when measured features evolve linearly inside the reservoir and the output is formed by a linear readout with bias, the capacity available at any fixed delay is bounded by the information already present in the preprocessed input. Linear dynamics can only redistribute features and cannot generate new fixed-delay expressive power. The authors derive the corresponding Gaussian limit for covariance-based continuous-variable quantum reservoirs and present numerical experiments indicating that single-photon (non-Gaussian) operations exceed this bound, thereby providing an operational witness of non-Gaussian processing under minimal assumptions.

Significance. If the central conditional derivation holds, the result supplies a precise, assumption-explicit limitation that explains why aggregate capacity measures can mask per-delay constraints. The specialization to the Gaussian CV case and the numerical demonstration that experimentally accessible non-Gaussian operations surpass the bound constitute a concrete resource characterization for quantum reservoir computing. The explicit linearity assumptions and the falsifiable numerical claim are strengths that make the contribution technically useful.

minor comments (2)
  1. The definition of 'capacity' (e.g., whether it is normalized mutual information, mean-squared error, or another functional) should be stated explicitly in the main text near the first use of the term, rather than only in supplementary material.
  2. Figure captions for the numerical results should include the precise values of reservoir size, delay range, and number of input samples used, to facilitate direct comparison with the derived Gaussian bound.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their careful reading and positive recommendation to accept the manuscript. The report correctly summarizes the central result on the hidden per-delay bottleneck and its implications for both classical linear reservoirs and covariance-based continuous-variable quantum reservoirs.

Circularity Check

0 steps flagged

No significant circularity; derivation follows directly from stated linearity assumptions

full rationale

The central claim is explicitly conditional on the assumptions of linear feature evolution inside the reservoir and linear readout with bias. Under those conditions the bottleneck follows as a direct algebraic consequence of the reservoir state being an affine function of input history, so that any linear functional cannot introduce new fixed-delay information. No load-bearing self-citation, fitted parameter renamed as prediction, or self-definitional reduction is present; the result is self-contained against the explicit premises and does not reduce to prior author work or data fitting.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Central claim rests on the domain assumption of linear feature evolution and linear readout; no free parameters, new entities, or additional axioms are indicated in the abstract.

axioms (1)
  • domain assumption Measured features evolve linearly in the reservoir and output uses linear readout with bias
    Explicitly stated in the abstract as the setting in which the bottleneck applies.

pith-pipeline@v0.9.1-grok · 5676 in / 1277 out tokens · 31537 ms · 2026-06-29T11:23:43.967808+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

41 extracted references · 2 canonical work pages

  1. [1]

    Pascanu, T

    R. Pascanu, T. Mikolov, and Y. Bengio, On the difficulty of training recurrent neural networks, inProceedings of the 30th International Conference on Machine Learn- ing, Proceedings of Machine Learning Research, Vol. 28, edited by S. Dasgupta and D. McAllester (PMLR, At- lanta, Georgia, USA, 2013) pp. 1310–1318

    2013

  2. [2]

    Lukoˇ seviˇ cius and H

    M. Lukoˇ seviˇ cius and H. Jaeger, Survey: Reservoir com- puting approaches to recurrent neural network training, Comput. Sci. Rev.3, 127–149 (2009)

    2009

  3. [3]

    Tanaka, T

    G. Tanaka, T. Yamane, J. B. H´ eroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Recent advances in physical reservoir comput- ing: A review, Neural Networks115, 100 (2019)

    2019

  4. [4]

    L. C. Govia, G. J. Ribeill, G. E. Rowlands, and T. A. Ohki, Nonlinear input transformations are ubiquitous in quantum reservoir computing, Neuromorphic Computing and Engineering2, 014008 (2022)

    2022

  5. [5]

    Nokkala, R

    J. Nokkala, R. Mart´ ınez-Pe˜ na, G. L. Giorgi, V. Pa- rigi, M. C. Soriano, and R. Zambrini, Gaussian states of continuous-variable quantum systems provide univer- sal and versatile reservoir computing, Communications Physics4, 53 (2021)

    2021

  6. [6]

    Nokkala, R

    J. Nokkala, R. Mart´ ınez-Pe˜ na, R. Zambrini, and M. C. Soriano, High-performance reservoir computing with fluctuations in linear networks, IEEE Transactions on Neural Networks and Learning Systems33, 2664 (2022)

    2022

  7. [7]

    Garc´ ıa-Beni, G

    J. Garc´ ıa-Beni, G. L. Giorgi, M. C. Soriano, and R. Zam- brini, Scalable photonic platform for real-time quantum reservoir computing, Physical Review Applied20, 014051 (2023)

    2023

  8. [8]

    Garc´ ıa-Beni, G

    J. Garc´ ıa-Beni, G. L. Giorgi, M. C. Soriano, and R. Zam- brini, Squeezing as a resource for time series processing in quantum reservoir computing, Opt. Express32, 6733 (2024)

    2024

  9. [9]

    Nokkala, Online quantum time series processing with random oscillator networks, Scientific Reports13, 7694 (2023)

    J. Nokkala, Online quantum time series processing with random oscillator networks, Scientific Reports13, 7694 (2023)

    2023

  10. [10]

    Nokkala, G

    J. Nokkala, G. L. Giorgi, and R. Zambrini, Retriev- ing past quantum features with deep hybrid classical- quantum reservoir computing, Machine Learning: Sci- ence and Technology5, 035022 (2024)

    2024

  11. [11]

    Paparelle, J

    I. Paparelle, J. Henaff, J. Garcia-Beni, E. Gillet, D. Mon- tesinos, G. L. Giorgi, M. C. Soriano, R. Zambrini, and V. Parigi, Experimental memory control in continuous- variable optical quantum reservoir computing, Nature Photonics20, 413 (2026)

    2026

  12. [12]

    S. D. Bartlett, B. C. Sanders, S. L. Braunstein, and K. Nemoto, Efficient classical simulation of continuous variable quantum information processes, Physical Review Letters88, 097904 (2002)

    2002

  13. [13]

    Roeland, S

    G. Roeland, S. Kaali, V. R. Rodriguez, N. Treps, and V. Parigi, Mode-selective single-photon addition to a multimode quantum field, New Journal of Physics24, 043031 (2022)

    2022

  14. [14]

    Wenger, R

    J. Wenger, R. Tualle-Brouri, and P. Grangier, Non- Gaussian Statistics from Individual Pulses of Squeezed Light, Physical Review Letters92, 153601 (2004), pub- lisher: American Physical Society

    2004

  15. [15]

    Zavatta, S

    A. Zavatta, S. Viciani, and M. Bellini, Quantum-to- Classical Transition with Single-Photon-Added Coher- ent States of Light, Science306, 660 (2004), publisher: American Association for the Advancement of Science

    2004

  16. [16]

    Ourjoumtsev, R

    A. Ourjoumtsev, R. Tualle-Brouri, J. Laurat, and P. Grangier, Generating Optical Schr¨ odinger Kittens for Quantum Information Processing, Science312, 83 (2006), publisher: American Association for the Ad- vancement of Science

    2006

  17. [17]

    Averchenko, C

    V. Averchenko, C. Jacquard, V. Thiel, C. Fabre, and N. Treps, Multimode theory of single-photon subtrac- tion, New Journal of Physics18, 083042 (2016), pub- lisher: IOP Publishing

    2016

  18. [18]

    Y.-S. Ra, A. Dufour, M. Walschaers, C. Jacquard, T. Michel, C. Fabre, and N. Treps, Non-Gaussian quan- tum states of a multimode light field, Nature Physics16, 144 (2020), publisher: Nature Publishing Group

    2020

  19. [19]

    Stornati, A

    P. Stornati, A. Acin, U. Chabaud, A. Dauphin, V. Parigi, and F. Centrone, Variational quantum simulation using non-gaussian continuous-variable systems, Physical Re- view Research6, 043212 (2024)

    2024

  20. [20]

    Dambre, D

    J. Dambre, D. Verstraeten, B. Schrauwen, and S. Massar, Information processing capacity of dynamical systems, Scientific reports2, 514 (2012)

    2012

  21. [21]

    Wishart, The generalised product moment distribu- tion in samples from a normal multivariate population, 11 Biometrika , 32 (1928)

    J. Wishart, The generalised product moment distribu- tion in samples from a normal multivariate population, 11 Biometrika , 32 (1928)

    1928

  22. [22]

    H¨ ulser, F

    T. H¨ ulser, F. K¨ oster, K. L¨ udge, and L. Jaurigue, Deriving task specific performance from the information process- ing capacity of a reservoir computer, Nanophotonics12, 937 (2023)

    2023

  23. [23]

    Wringe, M

    C. Wringe, M. Trefzer, and S. Stepney, Reservoir com- puting benchmarks: a tutorial review and critique, Inter- national Journal of Parallel, Emergent and Distributed Systems40, 313 (2025)

    2025

  24. [24]

    Borghi, S

    M. Borghi, S. Biasi, and L. Pavesi, Reservoir computing based on a silicon microring and time multiplexing for bi- nary and analog operations, Scientific Reports11, 15642 (2021)

    2021

  25. [25]

    Leonhardt and H

    U. Leonhardt and H. Paul, Measuring the quantum state of light, Progress in Quantum Electronics19, 89–130 (1995)

    1995

  26. [26]

    Centrone, F

    F. Centrone, F. Grosshans, and V. Parigi, Cost and rout- ing of continuous-variable quantum networks, Phys. Rev. A108, 042615 (2023)

    2023

  27. [27]

    Krasimirov-Ivanov, A

    T. Krasimirov-Ivanov, A. Cervera-Lierta, P. Stornati, and F. Centrone, Hardware-inspired continuous vari- ables quantum optical neural networks, arXiv preprint arXiv:2512.05204 (2025)

    work page arXiv 2025

  28. [28]

    Polo and F

    B. Polo and F. Centrone, Non-gaussian enhancement of precision in quantum batteries, arXiv preprint arXiv:2505.24604 (2025)

    work page arXiv 2025

  29. [29]

    Hahto and J

    M. Hahto and J. Nokkala, Smarter usage of measure- ment statistics can greatly improve continuous variable quantum reservoir computing, New Journal of Physics 27, 094510 (2025)

    2025

  30. [30]

    Jaeger,Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the echo state network approach, Vol

    H. Jaeger,Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the echo state network approach, Vol. 5 (GMD-Forschungszentrum Information- stechnik Bonn, 2002)

    2002

  31. [31]

    Montavon, G

    G. Montavon, G. Orr, and K.-R. M¨ uller,Neural networks: tricks of the trade, Vol. 7700 (springer, 2012)

    2012

  32. [32]

    Arzani, N

    F. Arzani, N. Treps, and G. Ferrini, Polynomial approxi- mation of non-gaussian unitaries by counting one photon at a time, Phys. Rev. A95, 052352 (2017). 12 Appendix A: Reservoir computing Reservoir computing is concerned with transformations between two time series. Lets={. . . , s t−2, st−1, st}be the input time series up to some fixed timesteptando={....

    2017

  33. [33]

    This causes no essential change because the output already includes a freely tunable bias term

    Affine reservoir updates In the most general case, the linear reservoir transformation may be affine rather than strictly linear, xt =Ax t−1 +Bg t +c,(B1) 13 wherecis a constant vector. This causes no essential change because the output already includes a freely tunable bias term. Indeed, define the augmented state and input vectors ˜xt = xt 1 ,˜g t = gt ...

  34. [34]

    The only requirement is fading memory

    Case A: preprocessing with memory, memoryless reservoir In this caseA=0andg t is allowed to depend on input history{s t−τ }τ≥0, rather than only on the most recent inputs t. The only requirement is fading memory. The information-processing capacity (IPC) is upper bounded by the number of available linearly independent functions of the input, and in the as...

  35. [35]

    Lemma 1.Letτ ′ be fixed and letz t =P d(st−τ ′)withd >0

    Auxiliary lemma on fixed delays We next formalize the fact that functions supported on different delays do not contribute to the reconstruction of a target at a fixed delay. Lemma 1.Letτ ′ be fixed and letz t =P d(st−τ ′)withd >0. LetU τ ′ be the span of the constant function together with all functions of the single variables t−τ ′, and letV τ ′ be the s...

  36. [36]

    The solution to Eq

    Case B: memoryless preprocessing, reservoir with memory In this caseg t =g(s t) depends only on the most recent input, whereas the reservoir recurrence carries memory. The solution to Eq. (6) is xt =A tx0 + t−1X τ=0 Aτ Bg(st−τ),(B11) wherex 0 is an arbitrary initial state. By fading memory, lim t→∞ Atx0 = 0.(B12) Hence in the asymptotic regime the observa...

  37. [37]

    Case A In the memoryless-reservoir case, σout t =Sσ tS⊤.(C1) After vectorization this becomes vec(σout t ) = (S⊗S) vec(σ t),(C2) where⊗is the Kronecker product. Thus Eq. (6) applies with xt = vec(σout t ), g t = vec(σt),A=0,B=S⊗S.(C3)

  38. [38]

    The reservoir modes evolve according to σR t =P Rσout t P⊤ R,(C5) whereP R projects onto the reservoir modes

    Case B In the memoryful-reservoir case, write σout t =S σR t−1 ⊕σ t S⊤,(C4) whereσ R t−1 is the reservoir covariance matrix and⊕denotes the direct sum. The reservoir modes evolve according to σR t =P Rσout t P⊤ R,(C5) whereP R projects onto the reservoir modes. To avoid confusion with the abstract matricesAandB, decompose the symplectic matrix as S= M N P...

  39. [39]

    Non-Gaussian states built from Gaussian ones Let ˆρG be anN-mode Gaussian state and let ˆO= nY r=1 ˆa#r sr ,# r ∈ {·,†},(D3) be a finite ordered product of annihilation and creation operators acting on selected modess r. The corresponding non-Gaussian state is ˆρnG = ˆOˆρG ˆO† Tr h ˆOˆρG ˆO† i = ˆOˆρG ˆO† K ,(D4) with normalization K= Tr h ˆOˆρG ˆO† i = T...

  40. [40]

    Let ˆb1,

    Wick expansion For a zero-mean Gaussian state, expectation values of odd numbers of centered field operators vanish, while even moments factorize according to Wick’s theorem. Let ˆb1, . . . ,ˆbm be operators each equal to some ˆaj or ˆa† j. Then Tr h ˆb1 · · ·ˆb2q+1 ˆρG i = 0,(D8) 17 and Tr h ˆb1 · · ·ˆb2q ˆρG i = X P Y (u,v)∈P Tr h ˆbuˆbv ˆρG i ,(D9) whe...

  41. [41]

    These are reconstructed from Eq

    From ladder-operator moments to the covariance matrix The observables used in the simulations are entries of the covariance matrix of selected quadratures in the non- Gaussian state. These are reconstructed from Eq. (D6) by choosing ˆMto be linear or quadratic in the ladder operators and then converting to quadrature moments. The first moments are ⟨ˆxj⟩nG...