{"id":"3d5cbce9-4c89-44dd-8bda-0354c1fd0393","arxiv_id":"2411.18436","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Lanczos coefficient ensembles over random initial operators are shown to follow Wishart and rescaled chi-square statistics, with patterns that distinguish chaotic from integrable billiards.","lead":"By averaging over many random initial operators, the authors show that the Lanczos coefficients controlling Krylov complexity have universal fluctuations: their correlations follow a Wishart distribution and their variance follows a rescaled chi-square distribution. These statistics offer a new probe of quantum chaos, but the result rests on an assumed normality of the coefficients and is verified only through qualitative histogram fits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wishart/chi-square claims rest on an undemonstrated multivariate-normal ansatz; the presented numerics test only diagonal marginals, not the joint Wishart/off-diagonal structure.","rationale":"The paper aims to establish universal statistical signatures of chaos in Krylov-space Lanczos coefficients. The only route from the modeling to the Wishart and rescaled chi-square statements is the multivariate normality assumption for X, asserted in Section 3.2 on central-limit intuition and previous numerics. Eq. (3.13) for the Wishart distribution and Eq. (3.18) for the chi-square distribution are standard consequences of that assumption, so the assumption is genuinely load-bearing. The numerical evidence presented does not close this gap: it checks univariate histograms of diagonal entries and of σ², but not the full matrix distribution, and the claim that the initial-operator ensembles are 'univariate, whose Σ is identity' is not a derivation of the covariance structure of the resulting X vector. The concern is not that the paper disagrees with consensus; it is that the central claim is under-supported by the evidence shown. I therefore keep the reader's conditional verdict: the paper should be accepted only if the multivariate normal ansatz is independently verified or derived, and if the full Wishart and chi-square structure, including the neglected cross terms, is quantitatively tested. The figure-reference swaps in Section 4.2 are secondary and should also be corrected, but they are not the basis for the conditional verdict.","tokens_in":20103,"tokens_out":4553,"duration_ms":41967,"concrete_test":"Using the same Sinai-billiard setup (a = 1, GUE, Nmax = 50, N = 5000), collect the 5000 vectors X = (x_1, ..., x_500) from the Lanczos pipeline and run a formal multivariate normality test (e.g., Mardia skewness/kurtosis or Henze-Zirkler) on X; then compare the empirical density of all off-diagonal entries of the normalized Gram matrix M/N = ⟨x_i x_j⟩ to the Wishart density W_K(Σ, N) with Σ the empirical covariance, and compute explicitly the cross term p^{-2} Σ_{i≠j} x_i x_j in Eq. (3.17). If multivariate normality is rejected or the off-diagonal entries deviate from the Wishart density, the universal distributional claims collapse.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central distributional claims—⟨x_i x_j⟩ ∼ W_K(Σ,N) and σ² ∼ rescaled χ²—follow only from the ansatz in Section 3.2 that the K-vector X = (x_1, ..., x_K) of log-ratio Lanczos coefficients is zero-mean multivariate normal, with Eq. (3.13) turning the sample covariance into a Wishart matrix. This ansatz is never derived; Lemma 1 only shows that the coefficients are functions of the initial matrix and energy levels, not that those functions are Gaussian or jointly Gaussian. The numerical check in Section 4.1 is weaker than claimed: Figure 2(a,c) histograms only the diagonal entries ⟨x_i²⟩ and reduces the Wishart check to a univariate chi-square, while no off-diagonal entry of ⟨x_i x_j⟩ is compared to the Wishart density. The argument that 'the distributions of initial operators are univariate, whose Σ is the identity' conflates the elementwise distribution of the initial random matrix with the covariance of the derived, highly nonlinear K-dimensional vector X. Likewise, Section 3.3 discards the cross terms in Eq. (3.17) on the assertion that they are smaller by at least an order of magnitude, but no supporting calculation is shown; if the x_i are correlated, σ² is not a rescaled chi-square. The Nmax = 5 fits in the appendix test only the marginal histogram and cannot distinguish the full multivariate-normal hypothesis from other distributions with similar marginals.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies statistical properties of Lanczos coefficients in Krylov operator complexity, over ensembles of random initial operators. It proposes two quantities: the average correlation matrix ⟨x_i x_j⟩ and the distribution f_{σ²} of the variance σ² of the log-ratio Lanczos coefficients x_i. The central claim, stated in the abstract and developed in Section 3, is that ⟨x_i x_j⟩ follows a Wishart distribution and σ² follows a rescaled chi-square distribution, independent of the distribution of initial operators, and that both become normal for large matrix size. The numerical examples are Sinai and Stadium billiards with GOE, GUE, and uniform-distribution initial operators. The paper reports agreement with these distributional predictions and uses the quantities to distinguish chaotic from nonchaotic behavior, connecting RMT, Anderson localization, and Krylov complexity.","tokens_in":20448,"tokens_out":4192,"duration_ms":40908,"significance":"If the distributional claims were established, they would provide a genuinely useful, potentially universal statistical characterization of Lanczos coefficients, connecting operator growth in Krylov space to random matrix theory and Anderson localization. The paper has notable strengths: Lemma 1 explicitly identifies the dependence of Lanczos coefficients on initial operators and energy levels; the assumption behind the Wishart/chi-square claims is stated transparently; and the numerical appendix is extensive, including fitted distributions, skewness/kurtosis values, and robustness checks with different Lanczos windows. The main weakness is that the load-bearing premise, the multivariate normality of X = (x_1, ..., x_K), is assumed rather than derived, and the numerical verification presented tests only univariate marginals, not the joint Wishart structure. The paper is therefore best read as a phenomenological proposal whose central claim is plausible but not yet conclusively supported.","major_comments":[{"comment":"The Wishart result M = X^T X ~ W_K(Σ, N) is a direct mathematical consequence of assuming that the vectors X_i are independent draws from N_K(0, Σ). It is not derived from Lemma 1, which only shows that each Lanczos coefficient is a deterministic function of the initial operator and energy levels. Lemma 1 does not imply that these nonlinear functions are jointly Gaussian. Since all subsequent distributional claims rest on this ansatz, the paper should either provide a derivation or a quantitative justification of the Gaussian approximation, or explicitly and consistently frame the contribution as a phenomenological model conditional on the normality ansatz.","section":"Section 3.2, Eq. (3.13)"},{"comment":"The numerical verification is weaker than claimed. Figure 2(a,c) and Figure 3(a,c) test only the diagonal marginals ⟨x_i²⟩ and the scalar variance σ²; no off-diagonal entry ⟨x_i x_j⟩ with i ≠ j is compared to the Wishart density, and no test of the joint distribution of M is reported. Consequently, the numerics cannot distinguish the multivariate normal hypothesis from other distributions sharing the same marginals. A direct test of the Wishart structure, for example by comparing the empirical distribution of off-diagonal entries of M to the predicted Wishart density or by using a Wishart likelihood/QQ diagnostic, is needed to support the central claim.","section":"Section 4.1, Figures 2 and 3"},{"comment":"The neglect of the cross term ∑_{i≠j} x_i x_j is justified only by the assertion that it is 'smaller than the first term by at least one order of magnitude', with no supporting calculation. If the x_i are correlated, σ²_O is not a rescaled chi-square variable but a generalized chi-square with parameters depending on the full covariance Σ. Moreover, the simplified chi-square result assumes independence of the x_i, and this independence is not tested in the paper. The authors should quantify the cross-term contribution and provide evidence for independence or fit the full covariance structure.","section":"Section 3.3, Eqs. (3.17)-(3.18)"},{"comment":"The statement that 'the distributions of initial operators are univariate, whose Σ is the identity matrix' conflates the elementwise distribution of the initial random matrix with the covariance structure of the derived vector X. In the model N_K(0, Σ), Σ is the covariance among the x_i components, not the covariance among entries of O(0). This conflation is load-bearing because it is used to reduce the Wishart claim to a univariate chi-square statement. The authors should clarify what exactly is assumed about Σ and provide numerical evidence that Σ is diagonal, or at least that the off-diagonal structure does not affect the conclusions.","section":"Section 3.2, paragraph after Eq. (3.16)"},{"comment":"The claimed independence of the statistics from the initial operator distribution is overstated. Table 1 and Figure 5a show that the fitted mean of σ² depends strongly on the initial operator ensemble (GOE vs. GUE vs. UCP) in the chaotic Sinai billiard. What appears to be universal is only the functional form of the distribution, not its parameters. The paper should state this distinction clearly, since the phrase 'independent of the distributions of initial operators' invites the stronger reading that the entire distribution is ensemble-independent.","section":"Abstract and Section 5"}],"minor_comments":[{"comment":"The notation is inconsistent: M = X^T X is the scatter matrix, while the average correlation matrix is ⟨x_i x_j⟩ = M_{ij}/N. The distribution stated, W_K(Σ, N), applies to M, not to M/N, so the rescaled distribution of ⟨x_i x_j⟩ should be stated explicitly.","section":"Section 3.2, Eq. (3.13)"},{"comment":"There is a typo: 'nubmers' should be 'numbers'.","section":"Section 2, Lanczos algorithm"},{"comment":"The y-axis labels in Figures 2(b) and 2(d) read '⟨x_i²⟩' even though the text describes the histogram of ⟨x_i x_j⟩; the labels should be corrected.","section":"Section 4.1, Figure 2"},{"comment":"The phrase 'if K = Σ = 1' is awkward and potentially confusing, since K is the dimension of the Wishart distribution and Σ is its scale matrix; a clearer statement would be 'in the one-dimensional case with unit variance'.","section":"Section 4.1, first paragraph"},{"comment":"The appendix states that for the Stadium billiard there is no overlapping behavior in the nonchaotic case, but Figure 6a indeed shows separated groups; this is consistent, but the wording could be made clearer to avoid apparent contradiction with the Sinai discussion.","section":"Appendix, Section [II]"}],"recommendation":"major_revision","confidential_remarks":"The paper's central distributional claims are conditional on an unproven multivariate-normality ansatz, and the presented numerics verify only marginals, not the joint Wishart structure. I believe this is fixable within the manuscript's scope by adding joint-distribution tests, quantifying the neglected cross terms, and softening the universality claim. The topic is of interest to the quantum-chaos and Krylov-complexity community, and the extensive numerical appendix is a strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the genuinely new thing here is a set of ensemble-level observables for Lanczos coefficients—the mean correlation matrix ⟨xi xj⟩ and the distribution of the variance σ²—plus numerical phenomenology that is striking: for the Sinai billiard these statistics change visibly as the integrability-breaking parameter is tuned, and in the chaotic regime the variance distributions split into two groups depending on whether the initial-operator ensemble is real or complex. That is worth a close look.\n\nWhat the paper does well: it is honest about its method. Section 3.2 is explicitly a qualitative/phenomenological analysis, and the appendix is careful about checking that the normal limit is reached for large Nmax. Lemma 1 is a clean statement that the Lanczos coefficients depend only on the energy levels and the initial operator. The reported patterns—cross-like structures in the nonchaotic Sinai correlation matrix and the two-group separation—are plausible and interesting, and the paper connects them to Anderson localization without overreaching.\n\nThe soft spot is the load-bearing one. The Wishart claim and the chi-square claim both follow from assuming the K-vector X of log-ratio Lanczos coefficients is multivariate normal, plus, for σ², dropping the cross terms in Eq. (3.17). That ansatz is not derived, and the argument that the initial-operator distributions are \"univariate with Σ = I\" conflates the elementwise distribution of the initial random matrix with the covariance of the derived nonlinear vector X. The numerics in Figs. 2 and 3 test only diagonal marginals (histograms of ⟨xi²⟩ and of σ²); no off-diagonal element of the sample covariance is compared to the Wishart density. So the abstract's \"their statistics are the Wishart distribution... independent of the distributions of initial operators\" is a stronger statement than the paper establishes. The cross-term discard is asserted to be good to at least an order of magnitude, but no supporting calculation is shown. Minor issues: Section 4.2 has duplicated \"[I]\" headings, and Table 1's UCP chaotic fit values (2.11016, 0.163631) look like a units or transcription error relative to all other entries. No code or data are released, so the fitted curves are fits, not predictions.\n\nNone of this kills the paper. The observations stand on their own as a candidate statistical probe of chaos. But a serious referee should send it back for: (1) a derivation or a much more qualified statement of the normality ansatz, (2) a test of the full joint distribution including off-diagonal structure, (3) a quantitative justification for neglecting the cross terms in the variance, and (4) cleanup of the small presentation problems. With those, it would be a useful contribution. I would read the revised version.","headline":"Worth engaging for its new ensemble observables and the two-group splitting, but the Wishart/chi-square claims rest on an unproven normality ansatz and the numerics only test marginals.","tokens_in":20912,"tokens_out":3191,"would_cite":true,"duration_ms":29258,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that the statistical spread of an operator in Krylov space is universal: sampling random initial operators makes the log-ratio Lanczos coefficients approximately multivariate normal, the sample correlation matrix…","keywords":["Krylov complexity","Lanczos coefficients","Wishart distribution","chi-square distribution","random matrix theory","Anderson localization","quantum chaos","operator growth"],"falsifier":"Sample initial operators from a distribution with heavy tails (for example, Student's t with a few degrees of freedom) for a fixed chaotic Hamiltonian and test whether the empirical distribution of $\\sigma^2$ is still rescaled chi-square; a detectable departure would falsify the claimed independence from the initial-operator distribution. A sharper check: apply a multivariate normality test to the sampled vectors $\\vec{X}$; strong rejection at large $N$ would invalidate the Wishart derivation at its root.","tokens_in":19865,"feed_emoji":"🎲","tokens_out":8682,"duration_ms":67624,"temperature":0.7,"pith_summary":"This paper tries to establish that the noise in Krylov operator growth is not arbitrary but obeys universal statistical laws. Concretely, for a fixed Hamiltonian, if you draw many initial operators from standard random-matrix ensembles (GOE, GUE, or uniform), the log-ratio Lanczos coefficients $x_i = \\ln|b_{2i-1}/b_{2i}|$ form an approximately normal random vector, the sample correlation matrix $\\langle x_i x_j\\rangle$ follows a Wishart distribution, and the variance $\\sigma^2$ of those coefficients follows a rescaled chi-square distribution. These laws are claimed to be independent of the initial-operator distribution and to become Gaussian as the operator dimension grows. The payoff is a statistical signature of quantum chaos that ties random matrix theory to Anderson localization on the Krylov chain and to Krylov complexity, and that visibly separates integrable from chaotic billiards.","feed_headline":"Lanczos coefficients obey universal Wishart statistics","feed_subtitle":"Sampling random initial operators reveals a statistical fingerprint of chaos tying random matrices to Anderson localization.","key_machinery":"The object that carries the argument is the vector $\\vec{X}=(x_1,\\dots,x_K)$ of log-ratio Lanczos coefficients, $x_i = \\ln|b_{2i-1}/b_{2i}|$, treated as a zero-mean multivariate normal random vector over the ensemble of initial operators. From this normal ansatz the sample correlation matrix $\\langle x_i x_j\\rangle$ becomes a Wishart matrix, and the variance $\\sigma^2$ becomes a rescaled chi-square variable once the off-diagonal terms in $\\sigma^2$ are dropped. A supporting lemma shows by induction that every Krylov-basis matrix element depends only on the energy levels and the initial random matrix, which supplies the central-limit intuition for why the coefficients should look normal; the mapping of the Krylov dynamics to a one-dimensional chain with hopping amplitudes $b_n$ then links these statistics to Anderson localization.","core_discovery":"The central claim is that the ensemble statistics of Lanczos coefficients are universal for chaotic dynamics. Under the ansatz that $\\vec{X} = (x_1,\\dots,x_K)$ with $x_i = \\ln|b_{2i-1}/b_{2i}|$ is a zero-mean multivariate normal random vector across samples of initial operators, the empirical correlation matrix $\\langle x_i x_j\\rangle$ is a sample covariance matrix and therefore has a Wishart distribution $W_K(\\Sigma,N)$, while the variance $\\sigma^2 = \\mathrm{Var}(x_i)$ collapses to a rescaled chi-square variable when off-diagonal correlations are neglected. The paper argues that both distributions do not depend on which random-matrix ensemble the initial operator is drawn from, and that they become normal distributions for large matrix size because many random numbers enter the Krylov construction. In the Sinai billiard, the two quantities change character between integrable ($a=0$) and chaotic ($a=1$) parameter regimes, while the Stadium billiard behaves like the chaotic Sinai case throughout.","pith_inferences":["Inference: the paper demonstrates the Wishart/chi-square laws only in the isotropic case where the underlying covariance $\\Sigma$ is effectively the identity; a natural test is to engineer a Hamiltonian or a sampling scheme with a non-trivial $\\Sigma$ and check whether the full matrix structure of the Wishart density is reproduced.","Inference: if the normality ansatz holds beyond billiards, the distribution of $\\sigma^2$ could serve as a chaos probe in many-body or holographic settings where traditional level statistics are hard to compute, since the variance only requires Lanczos coefficients from a single operator sample.","Inference: the real-versus-complex two-group separation looks like a manifestation of Dyson's threefold way, and one could test it by adding a time-reversal-breaking term to the Sinai billiard and watching the GUE-like group merge or split."],"forward_implications":["The average correlation matrix and the variance distribution of Lanczos coefficients become ensemble-level probes of chaos that do not require choosing a particular initial operator.","In the large-matrix limit these distributions are normal, so the mean and variance of $\\sigma^2$ are sufficient statistics for the ensemble, simplifying the numerical characterization of chaos.","The cross-like pattern in $\\langle x_i x_j\\rangle$ for the integrable Sinai billiard, locally resembling two-electron Anderson-localization correlations, ties the statistics directly to localization physics on the Krylov chain.","The split of $f_{\\sigma^2}$ into two well-separated groups (GOE/URE/UIM versus GUE/UCP) in the chaotic Sinai billiard indicates a real-versus-complex distinction in the random initial operators that persists in the statistics.","The proposed duality between random initial operators with fixed Hamiltonian and fixed operators with random-matrix Hamiltonians suggests that the same ensemble statistics control the late-time saturation of Krylov complexity."],"supporting_citations":[{"why":"establishes the Krylov operator space construction and the definition of Krylov complexity that the paper extends to ensemble statistics.","marker":"[1]"},{"why":"introduced the variance $\\sigma^2$ of Lanczos coefficients and its Anderson-localization interpretation on the Krylov chain.","marker":"[5]"},{"why":"provides the billiard-system numerics and the conventions for matrix size and Lanczos set size used in the simulations.","marker":"[13]"},{"why":"defines the Wishart distribution that the paper claims for the average correlation matrix $\\langle x_i x_j\\rangle$.","marker":"[23]"},{"why":"gives prior numerical evidence of the erratic and approximately normal behavior of Lanczos coefficients in disordered Krylov chains.","marker":"[4]"},{"why":"supplies the two-electron Anderson-localization correlation pattern compared with the cross-like structures seen in the Sinai billiard.","marker":"[18]"},{"why":"reports the integrability-to-chaos transition of Krylov complexity saturation and the RMT-based separation that motivates the proposed duality.","marker":"[6]"},{"why":"frames the average correlation matrix as the sample covariance matrix estimator used in random matrix theory.","marker":"[22]"},{"why":"defines the GOE and GUE ensembles from which the initial operators are sampled.","marker":"[24]"}],"fun_headline_variants":["Universal Wishart law for Krylov complexity","Chaos yields universal Lanczos statistics","Random operators reveal universal Lanczos statistics in chaos","Chaos imprints universal Wishart stats on Krylov coefficients","Wishart and chi-square: universal signatures of quantum chaos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if the log-ratio Lanczos coefficients do not behave as a zero-mean multivariate normal random vector across the sampled initial operators, an assumption the paper motivates but does not derive from the dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Universal Wishart law for Krylov complexity","Chaos yields universal Lanczos statistics","Random operators reveal universal Lanczos statistics in chaos","Chaos imprints universal Wishart stats on Krylov coefficients","Wishart and chi-square: universal signatures of quantum chaos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000396,"raw_usage":{"total_tokens":2052,"prompt_tokens":901,"completion_tokens":1151,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":1077}},"tokens_in":517,"tokens_out":1151,"duration_ms":8148,"temperature":1.0,"reasoning_tokens":1077,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:13:14.076316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample initial operators from a distribution with heavy tails (for example, Student's t with a few degrees of freedom) for a fixed chaotic Hamiltonian and test whether the empirical distribution of $\\sigma^2$ is still rescaled chi-square; a detectable departure would falsify the claimed independence from the initial-operator distribution. A sharper check: apply a multivariate normality test to the sampled vectors $\\vec{X}$; strong rejection at large $N$ would invalidate the Wishart derivation at its root.","supporting_citations":[{"cited_title":"Wishart,The generalised product moment distribution in samples from a normal multivariate population, Biometrika (1928) 32","cited_arxiv_id":null,"evidence_quote":"defines the Wishart distribution that the paper claims for the average correlation matrix $\\langle x_i x_j\\rangle$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the two-electron Anderson-localization correlation pattern compared with the cross-like structures seen in the Sinai billiard."},{"cited_title":"Mehta,Random Matrices, Academic Press (1991)","cited_arxiv_id":null,"evidence_quote":"defines the GOE and GUE ensembles from which the initial operators are sampled."}],"review_version":1}