Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Chaos switches on an information surplus in Wigner-negativity curves

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 · deepseek-v4-flash

2026-08-01 18:14 UTC pith:54HD4HV6

load-bearing objection An impressively careful ML study of Krylov observables with a plausible headline result, but the analytical bound in Sec. 6.2 misuses the R² inequality and the SFF-lossiness claim is limited by model capacity. the 3 major comments →

arxiv 2607.17346 v1 pith:54HD4HV6 submitted 2026-07-19 hep-th cond-mat.dis-nncond-mat.stat-mechmath-phmath.MPquant-ph

The Information Content of Krylov Observables: A Machine Learning Approach

Ritam Basu This is my paper
classification hep-th cond-mat.dis-nncond-mat.stat-mechmath-phmath.MPquant-ph MSC 81Q5081T4068T0515B52 PACS 05.45.Mt03.65.-w02.70.-c
keywords Krylov complexityspread complexityWigner negativitynormalized negativityspectral form factorquantum chaosrandom matrix theorymachine 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.

This paper asks how much information the standard Krylov-space curves—spread complexity C(t), Wigner negativity N(t), and normalized negativity χ(t)=N(t)/|S(t)|—carry about the quantum system that generated them, and about each other. Using small neural networks trained on roughly 57,000 simulated evolutions across GUE, GOE, Poisson, integrable CFT, and an integrable-to-chaos interpolation, it finds that temperature and universality class are recoverable from a single curve, but the fine spectral form factor is not recoverable from any moment observable. The central discovery is that the normalized negativity χ gains an informational surplus over spread complexity exactly as chaos sets in: the asymmetry gap grows from +0.33 to +0.77 as level statistics cross to the GUE value, while the raw-negativity gap decays to zero. The paper derives an analytical mechanism and a quantitative bound for this surplus, identifying the division by the survival amplitude as what preserves the non-self-averaging spectral fine structure.

Core claim

The paper's headline claim is operational: along the interpolation H(ε)=H_SL(2,R)+ε R0 W_GUE, the asymmetry gap R^2(χ→C) − R^2(C→χ) rises monotonically from +0.33 at ε=0.005 to +0.77 at ε≈0.12, precisely where the mean level-spacing ratio ⟨r⟩ first reaches the GUE plateau, then saturates near 0.6–0.7. The raw negativity gap, by contrast, decays to zero. The paper argues this shows the second-moment informational surplus is a signature of chaos carried specifically by the normalized negativity. The mechanism is that χ(t)=N(t)/|S(t)|: division by the survival amplitude injects the erratic, rigidity-sensitive return amplitude into an otherwise smooth observable, so χ retains exactly the spectra

What carries the argument

The central objects are the Krylov chain wavefunction ψ_n(t) and three observables built from it: spread complexity C(t) (first moment of the occupation), discrete Wigner negativity N(t) (integrated negative volume of the Wigner function on the chain's discrete phase space), and normalized negativity χ(t)=N(t)/|S(t)| where S(t) is the survival amplitude. The argument's load-bearing identity is the decomposition Y=log χ = A(t) + B(t) + ξ_D, where A is a smooth envelope slaved to the moments and B is 'speckle'—half the log of the filtered spectral form factor, whose variance is predicted to be π²/6 in the chaotic phase (Lemma 3). Theorem 1 converts this decomposition into a lower bound on the

Load-bearing premise

The claim that the spectral form factor strictly refines both moments rests on treating a tiny residual network as an information oracle: the same network scores R²≈0 when asked to decode the fine SFF from the lossless chain data {a_n,b_n}, so the 'lossiness' could in principle be a decoding limitation of the model class rather than information absent from the moments.

What would settle it

Train a substantially more expressive decoder (e.g., a large transformer or an analytically constructed estimator) on the lossless chain data {a_n,b_n}→log SFF; if it achieves high reconstruction R² on held-out samples, the paper's claim that the SFF is 'information-theoretically' absent from the moments would need revision, and the χ-surplus interpretation would be weakened. Alternatively, measure the χ-over-C gap along an experimental or numerical chaotic transition: it must rise monotonically with the level-spacing ratio ⟨r⟩ and persist with increasing dimension.

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

If this is right

  • A single normalized-negativity curve can serve as a practical chaos diagnostic: the χ-over-C gap tracks the level-statistics crossover point, so it can detect the chaotic transition without spectral unfolding.
  • Fine spectral rigidity (the oscillatory part of the SFF) is not reconstructible from any moment observable under any admissible smoothing; spectral rigidity must be measured directly, not inferred from smooth curves.
  • The coarse e^S plateau—the long-time spectral-form-factor plateau—is largely latent in the first 20% of the complexity curve, giving an 'information budget in time' for the black-hole interior volume scale.
  • In integrable sectors (SL(2,R) primary modules) the observables are informationally equivalent, so any apparent C↔N asymmetry there is a protocol artifact; the paper's dual-metric and logarithmic-channel audits eliminate such artifacts.
  • The χ-surplus sharpens with Hilbert-space dimension and saturates, which the paper interprets as a thermodynamic-limit effect—a testable signature that the mechanism is not a finite-size artefact.

Where Pith is reading between the lines

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

  • The same normalization trick—dividing a smooth observable by a microscopic amplitude—could be applied to other detectors (e.g., out-of-time-order correlators or Loschmidt echoes) to amplify fine spectral information; the paper's speckle decomposition gives a template for predicting when such a normalized observable will beat its raw counterpart.
  • The ceiling experiment suggests a cautious reading of all ML-for-physics lossiness claims: a score of R²≈0 from a tiny network is not proof that information is absent, only that it is not decodable at that model depth. A more expressive decoder might extract the SFF from {a_n,b_n}, which would leave the χ-surplus intact but weaken the 'SFF strictly refines moments' phrasing.
  • If the mechanism is correct, the χ-surplus should port to time-reversal-invariant (GOE-like) chaotic systems, including potentially experimental many-body quantum simulators where single-observable time traces are all one has.
  • The paper's information-budget analysis suggests an experimental protocol: measure the early-time complexity curve of a quench and use it to predict the late-time spectral plateau; deviations from the predicted plateau would signal physics beyond the single-observable description.

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

3 major / 5 minor

Summary. The paper uses small residual networks and boosted trees on ~57,000 labeled Krylov evolutions to measure how much information the spread complexity C(t), Wigner negativity N(t), and normalized negativity χ(t)=N(t)/|S(t)| carry about temperatures, spectral form factors, symmetry classes, and integrability-to-chaos crossover. The results include near-perfect temperature recovery (R²≈0.999), an apparent inability of either moment to reconstruct fine SFF structure (R²≈0.18), recovery of the coarse e^S plateau (R²≈0.861), ensemble classification up to 98%, an integrable-sector slaving where no asymmetry survives, and the headline that the asymmetry gap of χ over C grows from +0.33 to +0.77 as level statistics cross to GUE. The paper also proposes an analytical mechanism and a quantitative lower bound for this χ-surplus (Theorem 1, Eq. 6.13).

Significance. If the empirical results hold, this is a careful and useful operational map of Krylov observables, with a notably rigorous ML protocol: dual metrics, seed averaging, held-out parameter splits, and documented audits of metric and selection artifacts. The exact FFT form of the discrete Wigner negativity and the explicit D-scaling prediction are also strengths. However, the central quantitative bound rests on a misused inequality (trained R² as an upper bound when it is a lower bound), and the Lanczos-ceiling interpretation is internally contradicted by the paper's own Remark 4. These issues are load-bearing for the 'analytical mechanism and quantitative bound' claim, but they are local enough that a major revision—restating the bound as a heuristic or supplying a genuine upper bound, and weakening the information-theoretic language—could make the empirical findings stand.

major comments (3)
  1. [§6.2, (D2), Eq. (6.13)] Theorem 1's lower bound uses ρ² ≤ R²⋆(C→log SFF) ≲ 0.19, but the only measured quantity is the trained score R²_vw = 0.185 (Sec. 3.2.1). Remark 1 (Sec. 6.1) explicitly states that trained scores are certified lower bounds on R²⋆. A lower bound cannot be used as an upper bound. If a more expressive decoder lifts R²⋆(C→log SFF) above 0.19, then (1−ρ²)F may fall below the observed gap or even become negative, and Eq. (6.13) is not established. Please replace (D2) with a genuine upper bound on ρ² or reframe Theorem 1 as a heuristic decomposition rather than a quantitative bound.
  2. [§3.2.4, Fig. 7, Remark 4 (§6.3)] The ceiling task {a_n,b_n}→log SFF returns R²_vw ≈ −0.025, and Sec. 3.2.4 concludes that the SFF lossiness is 'information-theoretic in character.' But Remark 4 concedes that a fundamentally more expressive decoder could extract more fine SFF from {a_n,b_n}, making the ceiling a statement at fixed model depth. This internal contradiction undermines the stronger reading that the moments 'strictly refine' the SFF and that χ's surplus carries spectral fine structure that is provably inaccessible to C(t). The conclusion should be weakened to a model-class-limited statement, or backed by explicit model-capacity scaling demonstrating the ceiling is stable.
  3. [§5, Fig. 22, Definition 2 (§6.1)] The χ-surplus is substantially by construction: χ(t) = N(t)/|S(t)| includes the sample-specific speckle of |S(t)|, which C(t) cannot predict, so a positive gap in the chaotic regime follows directly from Lemmas 2 and 3. The paper's own mechanism makes this explicit. While the ε=0 slaving control and the D-persistence check are valuable, the stronger claim that the surplus is 'a signature of chaos' rather than a definitional property of the normalization needs a control in which |S| is replaced by an independent or scrambled speckle process of the same statistics. Without such a control, the interpretation is overreach.
minor comments (5)
  1. [§2.1, Eq. (2.3)] The notation `ifftℓ ρ` is undefined; please specify the discrete-Fourier convention used and the meaning of the subscript ℓ.
  2. [§2.3.3] R²_vw is used throughout but never defined with a formula. Please define the variance-weighted coefficient of determination explicitly.
  3. [§5.1, Table 4 and Fig. 22] The ε=0 anchor is mentioned in the text and table but not plotted in Fig. 22; adding the integrable-control point to the figure would make the crossover visually complete.
  4. [§6.2, proof of Theorem 1] The reverse inequality relies on a low-pass filter P_w with t_c ≪ w ≪ t_env, but the scales t_c and t_env are not quantitatively estimated in the text. Please state where these scales are measured or cite a figure.
  5. [Abstract and §7] The abstract's phrase 'The second-moment informational surplus is therefore a signature of chaos' is stronger than the caveats in §6.3 and §7 ('What it is not'). Please align the abstract with the stated scope.

Circularity Check

2 steps flagged

The χ-surplus is substantially built into the definition χ=N/|S|, and Theorem 1's quantitative bound is closed by using the trained C→SFF score as an upper bound despite Remark 1 certifying such scores as lower bounds.

specific steps
  1. self definitional [Sec. 2.1 (definition of χ, Eq. (1.2)) and Secs. 5.2, 6.2 (Theorem 1)]
    "We note for later use that for the TFD at inverse temperature β, |S(t)|^2 coincides with the (filtered) per-sample SFF; this makes some χ-involving comparisons trivially confounded... Mechanically this is natural: |S(t)| carries the erratic, rigidity-sensitive return amplitude, so χ retains fine-grained data that the smooth moments discard."

    χ is defined as N/|S|, and |S|^2 is the per-sample SFF. Since Sec. 3.2.1 already establishes that C cannot reconstruct the fine SFF (R2_vw=0.185), the asymmetry R2(χ→C)−R2(C→χ) is guaranteed by construction once S contains SFF speckle; Theorem 1 only formalizes this decomposition (gap ≥ (1−ρ^2)F). The existence of a surplus is therefore not an independent empirical discovery but a restatement of the definition of χ plus the measured SFF lossiness. The genuinely empirical part is the ε-dependence of F, not the surplus itself.

  2. fitted input called prediction [Sec. 6.2, hypothesis (D2) and Eq. (6.13)]
    "(D2) ρ^2 ≤ R^2_⋆(C→log SFF)≲0.19: B is (one half of ) the log filtered SFF, so its C-predictable share is bounded by the measured SFF score of Sec. 3.2.1."

    Remark 1 (Sec. 6.1) states that trained held-out scores are certified lower bounds: R2_test ≤ R2⋆. The value 0.185 quoted from Sec. 3.2.1 is such a trained score, so it cannot be used as an upper bound on ρ^2. The quantitative bound in Eq. (6.13) is thereby closed with a fitted input in the wrong inequality direction; a stronger decoder could lift R2⋆(C→log SFF) above 0.19 and the derived bound would fail. The paper's own Remark 4 concedes that a more expressive model class might extract more SFF from the lossless chain data, undermining the use of this particular fit as a ceiling.

full rationale

The paper contains a large amount of independent numerical content: temperature regression, C↔N mutual reconstruction, ensemble classification, SFF lossiness, plateau budget, and the integrable-sector slaving checks are all self-contained ML experiments with documented protocol. These do not reduce to their inputs. However, the headline claim—that the normalized-negativity surplus over complexity is an analytical mechanism and quantitative bound—is substantially definitional. Because χ=N/|S| and |S|^2 is the per-sample SFF, χ inherits the SFF speckle by construction; Theorem 1's lower bound is a decomposition identity (gap ≥ (1−ρ^2)F) rather than a first-principles derivation. The quantitative closing of the bound uses the trained C→SFF score (0.185) as an upper bound on ρ^2, directly contradicting the paper's own Remark 1 that trained scores are lower bounds on R2⋆. This is a fitted input doing load-bearing work in the wrong direction. Additionally, Sec. 3.2.4's ceiling task ({a_n,b_n}→SFF R2≈0) is interpreted in the main text as 'information-theoretic in character,' but Remark 4 later concedes that a stronger decoder might extract more SFF from the lossless chain data—so the lossiness conclusion that underpins the χ-interpretation is model-class dependent. These issues make the central analytical claim partially circular/unsupported, though the empirical ε-scan and the integrable-control values remain real measurements. Score 6 reflects partial circularity: the surplus's existence is by construction, and the quantitative bound is closed by a fitted score.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

The central claim's quantitative support rests on two measured quantities (F, ρ²) evaluated from the same data used to test the claim, plus heuristic concentration lemmas with uncontrolled constants. No new physical entities are introduced; the 'speckle' is a statistical fluctuation, not an entity.

free parameters (4)
  • ε (chaos interpolation strength)
    Control parameter in H(ε)=H_SL2+ε R0 W_GUE; swept over [0,0.3], not fitted.
  • ML hyperparameters (width 16/32, dropout 0.2–0.3, learning rate 1e-3, batch 32–64, patience 30–40)
    Chosen by hand; robustness to width and data fraction is argued, not a fit to the target.
  • Speckle share F and C-predictable share ρ² = F≈0.93–1.0 (deep chaos), ρ²≤0.19
    Measured from the same interpolation data and used in Theorem 1's bound; the bound is therefore a consistency loop, not a prediction.
  • Smoothing window Δt = up to 0.26
    Admissible window limited by dip-ramp physics; used to show the raw-N gap is noise-borne.
axioms (7)
  • domain assumption Krylov observables depend only on the spectrum; TFD evolution on random-matrix ensembles encodes temperature in curve shapes
    Sec. 2.1; used to define datasets and validate ensemble generation.
  • ad hoc to paper Lemma 2: moments concentrate as O(D^{-1/2}) with weakly correlated residues
    Sec. 6.1 proof sketch with uncontrolled constants; load-bearing for the self-averaging dichotomy.
  • domain assumption Lemma 3: late-time survival amplitude is complex Gaussian (Lindeberg CLT), giving Var(ln|S|²)→π²/6
    Standard RMT phasor limit; numerically checked; basis for speckle universality.
  • ad hoc to paper (D1) Cov(A,B)≈0 envelope/speckle decoupling
    Empirically controlled but fails at the crossover (F overshoots >1); needed for Theorem 1.
  • ad hoc to paper (D2) ρ² ≤ R²(C→SFF)≲0.19
    Uses the measured SFF score to bound speckle predictability.
  • standard math Negative-binomial occupation (4.3) for the SL(2,R)/CFT sector
    Perelomov coherent states; used for the integrable control and exact moment slaving.
  • domain assumption Discrete Wigner construction requires odd prime dimension D
    Wootters phase space; restricts datasets to odd primes and excludes GSE.

pith-pipeline@v1.3.0-alltime-deepseek · 24560 in / 12948 out tokens · 121348 ms · 2026-08-01T18:14:36.399674+00:00 · methodology

0 comments
read the original abstract

We employ machine learning to quantify the information carried by three Krylov-space observables: the spread complexity $\mathcal{C}(t)$, the discrete Wigner negativity $N(t)$, and the normalized negativity $\chi(t)=N(t)/|S(t)|$, with $S(t)$ the survival amplitude, recently proposed as a second-moment infall probe (arXiv:2607.04065). Small residual networks (16-32 neurons) and boosted trees are trained on half of $\sim 57{,}000$ labeled evolutions spanning the GUE, GOE and Poisson ensembles, the integrable $SL(2,\mathbb{R})$/CFT sector, and the chaos interpolation $H(\varepsilon)=H_{SL(2,\mathbb{R})}+\varepsilon R_0 W_{GUE}$. Either moment determines the thermofield temperature at $R^2\simeq 0.999$. Neither reconstructs the fine spectral form factor ($R^2\simeq 0.18$ in every ensemble); smoothing the target does not repair this, and windows wide enough to help erase the dip-ramp physics itself: the SFF strictly refines both moments. The coarse $e^S$ plateau is nevertheless recovered at $R^2=0.861$, mostly from the first 20% of $\mathcal{C}(t)$. A single curve identifies the symmetry class at up to 98% accuracy. In the integrable sector the observables are informationally equivalent, as exact negative-binomial slaving demands, while the negativity best resolves the $(h,\alpha)$ degeneracy ($N\to h$: 0.999). Along the interpolation the asymmetry gap of $\chi$ over $\mathcal{C}$ switches on with chaos, growing from +0.33 to +0.77 as the level statistics cross to GUE, while the raw-$N$ gap decays to zero. The second-moment informational surplus is therefore a signature of chaos, carried specifically by the normalized negativity, and we derive an analytical mechanism and a quantitative bound for it.

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

43 extracted references · 33 linked inside Pith

  1. [1]

    Basu,Wigner negativity as a second-moment probe of infall in ads 3,2607.04065

    R. Basu,Wigner negativity as a second-moment probe of infall in ads 3,2607.04065

    Pith/arXiv arXiv

  2. [2]

    Balasubramanian, P

    V. Balasubramanian, P. Caputa, J.M. Magan and Q. Wu,Quantum chaos and the complexity of spread of states,Phys. Rev. D106(2022) 046007 [2202.06957]

    Pith/arXiv arXiv 2022

  3. [3]

    Parker, X

    D.E. Parker, X. Cao, A. Avdoshkin, T. Scaffidi and E. Altman,A universal operator growth hypothesis,Phys. Rev. X9(2019) 041017 [1812.08657]

    Pith/arXiv arXiv 2019

  4. [4]

    Barb´ on, E

    J.L.F. Barb´ on, E. Rabinovici, R. Shir and R. Sinha,On the evolution of operator complexity beyond scrambling,JHEP10(2019) 264 [1907.05393]

    Pith/arXiv arXiv 2019

  5. [5]

    Rabinovici, A

    E. Rabinovici, A. S´ anchez-Garrido, R. Shir and J. Sonner,Operator complexity: a journey to the edge of krylov space,JHEP06(2021) 062 [2009.01862]

    Pith/arXiv arXiv 2021

  6. [6]

    Dymarsky and A

    A. Dymarsky and A. Gorsky,Quantum chaos as delocalization in krylov space,Phys. Rev. B 102(2020) 085137 [1912.12227]

    Pith/arXiv arXiv 2020

  7. [7]

    Dymarsky and M

    A. Dymarsky and M. Smolkin,Krylov complexity in conformal field theory,Phys. Rev. D 104(2021) L081702 [2104.09514]

    Pith/arXiv arXiv 2021

  8. [8]

    Rabinovici, A

    E. Rabinovici, A. S´ anchez-Garrido, R. Shir and J. Sonner,Krylov localization and suppression of complexity,JHEP03(2022) 211 [2112.12128]

    Pith/arXiv arXiv 2022

  9. [9]

    Rabinovici, A

    E. Rabinovici, A. S´ anchez-Garrido, R. Shir and J. Sonner,Krylov complexity from integrability to chaos,JHEP07(2022) 151 [2207.07701]

    Pith/arXiv arXiv 2022

  10. [10]

    Caputa, J.M

    P. Caputa, J.M. Magan and D. Patramanis,Geometry of krylov complexity,Phys. Rev. Res. 4(2022) 013041 [2109.03824]

    Pith/arXiv arXiv 2022

  11. [11]

    Avdoshkin, A

    A. Avdoshkin, A. Dymarsky and M. Smolkin,Krylov complexity in quantum field theory, and beyond,JHEP06(2024) 066 [2212.14429]

    Pith/arXiv arXiv 2024

  12. [12]

    Erdmenger, S.-K

    J. Erdmenger, S.-K. Jian and Z.-Y. Xian,Universal chaotic dynamics from krylov space, JHEP08(2023) 176 [2303.12151]

    Pith/arXiv arXiv 2023

  13. [13]

    Balasubramanian, J.M

    V. Balasubramanian, J.M. Magan and Q. Wu,Quantum chaos, integrability, and late times in the krylov basis,Phys. Rev. E111(2025) 014218 [2312.03848]

    Pith/arXiv arXiv 2025

  14. [14]

    Nandy, A.S

    P. Nandy, A.S. Matsoukas-Roubeas, P. Mart ´ ınez-Azcona, A. Dymarsky and A. del Campo, Quantum dynamics in krylov space: Methods and applications,Phys. Rept.1125-1128 (2025) 1 [2405.09628]

    Pith/arXiv arXiv 2025

  15. [15]

    Susskind,Computational complexity and black hole horizons,Fortsch

    L. Susskind,Computational complexity and black hole horizons,Fortsch. Phys.64(2016) 24 [1402.5674]

    Pith/arXiv arXiv 2016

  16. [16]

    Stanford and L

    D. Stanford and L. Susskind,Complexity and shock wave geometries,Phys. Rev. D90(2014) 126007 [1406.2678]

    Pith/arXiv arXiv 2014

  17. [17]

    Brown, D.A

    A.R. Brown, D.A. Roberts, L. Susskind, B. Swingle and Y. Zhao,Holographic complexity equals bulk action?,Phys. Rev. Lett.116(2016) 191301 [1509.07876]

    Pith/arXiv arXiv 2016

  18. [18]

    Iliesiu, M

    L.V. Iliesiu, M. Mezei and G. S´ arosi,The volume of the black hole interior at late times, JHEP07(2022) 073 [2107.06286]. – 39 –

    Pith/arXiv arXiv 2022

  19. [19]

    Heller et al.,Krylov spread complexity as a holographic complexity measure,Phys

    M.P. Heller et al.,Krylov spread complexity as a holographic complexity measure,Phys. Rev. Lett.135(2025) 151602

    2025

  20. [20]

    R. Basu, A. Ganguly, S. Nath and O. Parrikar,Complexity growth and the krylov-wigner function,JHEP(2024) [2402.13694]

    arXiv 2024

  21. [21]

    Basu,Wigner negativity, random matrices and gravity,JHEP(2026) [2506.02110]

    R. Basu,Wigner negativity, random matrices and gravity,JHEP(2026) [2506.02110]

    arXiv 2026

  22. [22]

    Kharkov, V.E

    Y.A. Kharkov, V.E. Sotskov, A.A. Karazeev, E.O. Kiktenko and A.K. Fedorov,Revealing quantum chaos with machine learning,Phys. Rev. B101(2020) 064406 [1902.09216]

    Pith/arXiv arXiv 2020

  23. [23]

    Bak, S.-H

    D. Bak, S.-H. Kim, S. Park and J.-P. Song,Machine learns quantum complexity,2501.02005

    arXiv

  24. [24]

    Maldacena,Eternal black holes in anti-de sitter,JHEP04(2003) 021 [hep-th/0106112]

    J.M. Maldacena,Eternal black holes in anti-de sitter,JHEP04(2003) 021 [hep-th/0106112]

    Pith/arXiv arXiv 2003

  25. [25]

    Wootters,A wigner-function formulation of finite-state quantum mechanics,Annals Phys.176(1987) 1

    W.K. Wootters,A wigner-function formulation of finite-state quantum mechanics,Annals Phys.176(1987) 1

    1987

  26. [26]

    Gibbons, M.J

    K.S. Gibbons, M.J. Hoffman and W.K. Wootters,Discrete phase space based on finite fields, Phys. Rev. A70(2004) 062101 [quant-ph/0401155]

    Pith/arXiv arXiv 2004

  27. [27]

    Gross,Hudson ’s theorem for finite-dimensional quantum systems,J

    D. Gross,Hudson ’s theorem for finite-dimensional quantum systems,J. Math. Phys.47 (2006) 122107 [quant-ph/0602001]

    Pith/arXiv arXiv 2006

  28. [28]

    Kenfack and K

    A. Kenfack and K. ˙Zyczkowski,Negativity of the wigner function as an indicator of nonclassicality,J. Opt. B6(2004) 396 [quant-ph/0406015]

    Pith/arXiv arXiv 2004

  29. [29]

    Veitch, C

    V. Veitch, C. Ferrie, D. Gross and J. Emerson,Negative quasi-probability as a resource for quantum computation,New J. Phys.14(2012) 113011 [1201.1256]

    Pith/arXiv arXiv 2012

  30. [30]

    Mari and J

    A. Mari and J. Eisert,Positive wigner functions render classical simulation of quantum computation efficient,Phys. Rev. Lett.109(2012) 230503 [1208.3660]

    Pith/arXiv arXiv 2012

  31. [31]

    Howard, J

    M. Howard, J. Wallman, V. Veitch and J. Emerson,Contextuality supplies the ‘magic’ for quantum computation,Nature510(2014) 351 [1401.4174]

    Pith/arXiv arXiv 2014

  32. [32]

    Basu and S

    R. Basu and S. Paul,On the stabilizer complexity of hawking radiation,2510.18967

    arXiv

  33. [33]

    Oganesyan and D.A

    V. Oganesyan and D.A. Huse,Localization of interacting fermions at high temperature,Phys. Rev. B75(2007) 155111 [cond-mat/0610854]

    Pith/arXiv arXiv 2007

  34. [34]

    Y.Y. Atas, E. Bogomolny, O. Giraud and G. Roux,Distribution of the ratio of consecutive level spacings in random matrix ensembles,Phys. Rev. Lett.110(2013) 084101 [1212.5611]

    Pith/arXiv arXiv 2013

  35. [35]

    K. He, X. Zhang, S. Ren and J. Sun,Deep residual learning for image recognition, inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), p. 770, 2016 [1512.03385]

    Pith/arXiv arXiv 2016

  36. [36]

    Mehta,Random Matrices, Elsevier, 3rd ed

    M.L. Mehta,Random Matrices, Elsevier, 3rd ed. (2004)

    2004

  37. [37]

    T. Guhr, A. M¨ uller-Groeling and H.A. Weidenm¨ uller,Random matrix theories in quantum physics: Common concepts,Phys. Rept.299(1998) 189 [cond-mat/9707301]

    Pith/arXiv arXiv 1998

  38. [38]

    Bohigas, M.-J

    O. Bohigas, M.-J. Giannoni and C. Schmit,Characterization of chaotic quantum spectra and universality of level fluctuation laws,Phys. Rev. Lett.52(1984) 1

    1984

  39. [39]

    Berry and M

    M.V. Berry and M. Tabor,Level clustering in the regular spectrum,Proc. Roy. Soc. Lond. A 356(1977) 375. – 40 –

    1977

  40. [40]

    Cotler, G

    J.S. Cotler, G. Gur-Ari, M. Hanada, J. Polchinski, P. Saad, S.H. Shenker et al.,Black holes and random matrices,JHEP05(2017) 118 [1611.04650]

    Pith/arXiv arXiv 2017

  41. [41]

    Saad, S.H

    P. Saad, S.H. Shenker and D. Stanford,Jt gravity as a matrix integral,1903.11115

    Pith/arXiv arXiv 1903

  42. [42]

    Cover and J.A

    T.M. Cover and J.A. Thomas,Elements of Information Theory, Wiley, 2nd ed. (2006)

    2006

  43. [43]

    D. Guo, S. Shamai and S. Verd´ u,Mutual information and minimum mean-square error in gaussian channels,IEEE Trans. Inf. Theory51(2005) 1261 [cs/0412108]. – 41 –

    Pith/arXiv arXiv 2005