{"id":"34bcfad3-b3fb-4102-bd23-5f6be7c5d193","arxiv_id":"1908.06721","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"First general algorithms compute spectral measures, point/continuous/singular decompositions, functional calculus, and Radon-Nikodym derivatives for self-adjoint or unitary operators with known column decay, with Solvability Complexity Index classifications.","lead":"This mathematics paper supplies the first general algorithms that compute spectral measures and their point, continuous, and singular parts for infinite-dimensional operators whose matrices have columns of known decay, plus a hierarchy classifying how hard each task is. A nonspecialist might read it because spectral measures encode long-time quantum dynamics, wave transport, and operator stability, all previously lacking a general computational method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2's RAGE step asserts a uniform-in-s approximation of e^{-iTs}χ_U(T)x that is only 'easily adapted'; for unbounded U the total-variation justification is not applicable, leaving the main decomposition algorithm's convergence unproven.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption listing three structural premises. I agree that all three are real, but the most load-bearing is the uniform-in-s approximation inside the proof of Theorem 3.2. The first premise (knowing f, α, β and the decay rates) is part of the problem definition and does not undermine the theorem as stated; the second (the PDE inner-product lemma imported from [37]) is an external dependency that affects Theorem 1.1 but not the core l2(N) results; the third is the separation condition in Theorem 4.2, which is explicitly stated as a hypothesis and therefore not a hidden gap. The uniform-in-s bound, by contrast, is an asserted step in the proof of a main classification theorem, and the accompanying explanation — that exp(-iλs)χ_U(λ) has 'known total variation' — is not applicable to unbounded U. Since the RAGE identity is the only route in the paper to the continuous part of the measure, a failure of this step would directly invalidate the inclusion in Theorem 3.2 and propagate to Theorem 5.1. The concern does not force a rejection: the theorem may be true, and the step may be repairable using a careful L^2(μ) approximation with spectral-tail control. But the paper as written does not supply that argument, which is exactly why the conditional verdict is appropriate. The proposed concrete test (multiplication operator with unbounded U) isolates the issue in a minimal setting where the resolvent and the evolution can be computed exactly, so it would settle whether the asserted uniform bound is derivable from the advertised premises or requires additional assumptions.","tokens_in":57713,"tokens_out":8460,"duration_ms":96959,"concrete_test":"Attempt an independent proof of the 'easily adapted' assertion in §3.2.1, Step 1, for an unbounded operator and unbounded U. Take T to be the multiplication operator by λ on L^2(1,∞), U=(1,∞), and x with absolutely continuous spectral measure, e.g. dμ_{x,x}(λ)=e^{-λ}dλ. Using only the resolvent algorithms of Corollary 2.2, derive the minimal n needed to guarantee ||e^{-iTs}χ_U(T)x - Γ_{n,m}|| ≤ C/m uniformly for s∈[0,m]. If the required n must grow faster than any fixed polynomial in m, or if C must depend on spectral-tail information not computable from Λ1 evaluations, then the uniform bound used in Theorem 3.2 fails as written. Alternatively, implement the proposed Γ_{n,m} for this example and check numerically whether lim_{n→∞}lim_{m→∞}Γ_{n,m} equals ||P_c E_T(U)x||^2 = ||x||^2; a divergence would refute the inclusion, while convergence would indicate the gap is repairable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The inclusion proof of Theorem 3.2 (§3.2.1, Step 1) reduces the computation of the continuous-part spectral measure to the RAGE identity (3.3). To turn this into an arithmetic tower, the proof needs a sequence of algorithms ~Γ_{n,m} with ||Q_n e^{-iTs}χ_U(T)x - ~Γ_{n,m}(T,x,U,s)|| ≤ C(T,x,U)/m uniformly for s∈[0,m]. The only justification given is: 'The proof of Theorem 4.1 is easily adapted ... Note that this bound can be made independent of s ... by sufficiently approximating the function λ→ exp(-iλs)χ_U(λ) (it has known total variation for a given s and uniform bound).' This justification is not valid in the stated generality. If U is unbounded, for example U=(a,∞), the function f_s(λ)=e^{-iλs}χ_U(λ) has infinite total variation on U for s>0, so it cannot be uniformly approximated on U by compactly supported piecewise-constant functions in sup norm. The natural replacement is an L^2(μ_{x,x}) approximation, but the error then depends on the spectral tail of x, which is not available to the Λ1-based algorithm beyond the abstract decay β in (1.19). The claimed uniform-in-s bound is load-bearing: it is the mechanism that converts the RAGE limit into a computable three-limit tower. Without it, the Riemann-sum error in Γ_{n,m} may fail to vanish as m→∞, and the inclusion part of Theorem 3.2 — and hence the SCI=3 classification for decompositions — is not established. This is a gap inside the main proof, not a dependency on an external result or a caveat about the problem definition. It may be repairable through a separate spectral-tail argument, but as written the central claim is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops algorithms for computing spectral measures, their Lebesgue decompositions, functional calculi, Radon–Nikodym derivatives, and the pure point/absolutely continuous/singular continuous spectra of self-adjoint (and unitary) operators on ℓ²(N) whose matrix columns decay at a known asymptotic rate. The main constructive tool is an arithmetic algorithm for the resolvent with asymptotic error control (Theorem 2.1), combined with Stone's formula. The paper formulates these tasks in the Solvability Complexity Index (SCI) hierarchy, proving classifications such as: the full spectral measure on open sets is in Δ₂^A; measure decompositions are in Δ₃^A but not Δ₂^G; and singular continuous spectra require three limits under a bandwidth growth condition. Numerical experiments cover Jacobi, Laguerre, CMV/Geronimus/Rogers–Szegő measures and fractional diffusion on a Penrose-tile graph, and Appendix B extends the results to a class of PDEs through Hermite-function bases.","tokens_in":57914,"tokens_out":7562,"duration_ms":81368,"significance":"If the main proofs are completed, this is a substantial contribution: it provides the first general algorithms for computing spectral measures and spectral decompositions for a broad class of infinite-dimensional operators, with explicit SCI classifications and with detailed numerical demonstrations. The resolvent error bound (2.1) is explicit and checkable, the algorithms are arithmetic (hence implementable with rigorous interval arithmetic), and the numerical section goes well beyond toy examples. The paper also gives credit-worthy honest discussion of where error control is impossible (Theorem 5.2). The main caveat is a proof gap in the inclusion part of Theorem 3.2 that is load-bearing for the SCI classification of measure decompositions; until that gap is repaired, the full strength of the decomposition results is not established.","major_comments":[{"comment":"The proof requires an arithmetic algorithm ~Γ_{n,m} satisfying ||Q_n e^{-iTs} χ_U(T)x − ~Γ_{n,m}(T,x,U,s)|| ≤ C(T,x,U)/m uniformly for s∈[0,m]. The text justifies this by saying that 'the proof of Theorem 4.1 is easily adapted' because the function λ↦e^{-iλs}χ_U(λ) 'has known total variation for a given s and uniform bound'. This justification is not valid in the stated generality. For an unbounded open set U=(a,∞) and s>0, the function has infinite total variation on U, so it cannot be uniformly approximated on U by compactly supported piecewise-constant functions in sup norm; a natural replacement using L²(μ_{x,x}) approximation would require control of the spectral tail of x that is not provided by the hypotheses. Even for bounded U, the total variation grows with |s|, so obtaining a uniform-in-s error O(1/m) for s∈[0,m] needs an argument that the cited 'easy adaptation' does not supply. This uniform bound is load-bearing: it is the mechanism that converts the RAGE limit in (3.3) into the computable double limit Γ_{n,m}, and hence it underpins the inclusion part of Theorem 3.2 and the claimed SCI=3 classification for decompositions. The manuscript should either supply a complete proof of the uniform-in-s approximation (for example, via Stone's formula on bounded subintervals combined with a tail estimate derived from the stated decay assumptions) or restrict Theorem 3.2 to bounded U and make the corresponding adjustment to the classification statement.","section":"§3.2.1, Step 1 (proof of inclusion in Theorem 3.2)"},{"comment":"The reduction of the PDE problem to ℓ²(N) depends on the assertion that the Hermite-basis inner products (B.3)–(B.5) can be computed from point samples with asymptotic error control, a result imported from the companion paper [37]. This lemma is not stated or proved in the present manuscript. Since Theorem 1.1 and the PDE claims in the abstract rest on this step, the paper should either include a proof or a precise statement with a clear pointer, so that a reader can verify that the imported result has the required uniformity over the class Ω_PDE. As written, the PDE extension is conditional on an unstated external result.","section":"Appendix B (proof of Theorem B.1)"}],"minor_comments":[{"comment":"In the displayed definition of Γ_{n,m}, the inner algorithm is written as ~Γ_{m,n}(T,x,U,j/m), although the preceding estimate concerns ~Γ_{n,m}. With the written indices, taking the first limit m→∞ would involve Q_m→0 strongly and would produce 0 rather than the RAGE average; the indices should presumably be ~Γ_{n,m} throughout.","section":"§3.2.1"},{"comment":"The parenthetical claim that the approximating function 'has known total variation for a given s and uniform bound' is at best misleading: the total variation depends on s and is infinite for unbounded U. The proof should state explicitly how the uniformity in s is obtained, rather than appealing to total variation alone.","section":"§3.2.1"},{"comment":"There is a typo in the first sentence: 'This results of this paper' should read 'These results of this paper'.","section":"§1.3"},{"comment":"The notation Ω_{f,α,β} already encodes pairs (T,x), and Theorem 3.2 then writes the domain as Ω_{f,α,β}×V_β×U with variables (T,x,y,U). This is understandable but slightly confusing; a sentence clarifying that the first factor carries the pair (T,x) and the V_β factor carries y would help the reader.","section":"§1.6 and §3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious and, if the proof gap in Theorem 3.2 is repaired, likely to be an important contribution to computational spectral theory. The numerical experiments are extensive and impressive, but they do not substitute for the missing uniform-in-s argument in Step 1 of the inclusion proof. The editor may also wish to ensure that the companion papers [37] and [39] are available and that the imported lemma on Hermite-basis inner products is verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First things first: this is a substantial paper. It gives the first algorithms I know of that compute projection-valued spectral measures, functional calculi, and Radon–Nikodym derivatives for self-adjoint and unitary operators with known column decay. The resolvent-with-error-control theorem (2.1) is clean and I checked the main bound; Stone's formula reduction is standard and correctly handled. The SCI classifications, including the sharpness results for σpp and σsc, are genuinely new and interesting. The numerical section is honest and useful, though no code or data is shipped.\n\nThe soft spot is in the proof of Theorem 3.2, Step 1. The proof needs a sequence of approximations ~Γ_{n,m} with error C(T,x,U)/m uniform in s∈[0,m] for the time-evolved projected vector e^{-iTs}χ_U(T)x. The justification is a single sentence saying the proof of Theorem 4.1 is 'easily adapted' because the function λ→e^{-iλs}χ_U(λ) 'has known total variation for a given s and uniform bound.' For unbounded U, that function has infinite total variation on R, so the stated reason is wrong. The natural fallback is an L^2(μ_{x,x}) approximation, but the tail error depends on the spectral tail of x, which is not available to the algorithm from the assumed column decay. As written, the uniform-in-s bound is not established, and it is load-bearing: it converts the RAGE limit into the three-limit tower. This leaves the inclusion half of Theorem 3.2—and the SCI=3 classification for decompositions—unproven.\n\nEverything else I looked at holds up. The PDE part is honest about importing a lemma from the companion paper [37]; that's a real dependency but not a hidden one. The numerical convergence-rate claims in Section 6 are empirical, with a promise of proof rather than the proof itself.\n\nWho is this for? Anyone working on computational spectral theory, infinite-dimensional linear algebra, or rigorous numerics. It deserves a serious referee—conditionally, the editor should send it out, but the referee should be asked to nail down Step 1 or provide a replacement argument. My own verdict is conditional: I expect the gap is repairable, but the current text does not close it.","headline":"The paper delivers the first general algorithms for spectral measures on a broad operator class, but the proof of Theorem 3.2 leans on an unjustified uniform-in-s approximation and is incomplete as written.","tokens_in":58670,"tokens_out":6173,"would_cite":true,"duration_ms":66050,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["47A10","47B15","47A60","47B39","65J99"],"pacs":[],"model":"deepseek-v4-flash","headline":"For self-adjoint and unitary operators whose matrix columns decay at a known rate, the paper establishes that spectral measures, spectral types, functional calculus, and absolutely-continuous densities are computable by explicit…","keywords":["spectral measures","spectral decompositions","Solvability Complexity Index","self-adjoint operators","unitary operators","resolvent with error control","functional calculus","Radon-Nikodym derivative"],"falsifier":"Run the one-limit algorithm on a Jacobi matrix whose spectral measure is known to be the pure-point measure with atoms at the nonnegative integers and weights $\\exp(-\\alpha)\\alpha^m/m!$, letting $n$ grow as $\\epsilon\\downarrow 0$: the output must concentrate at the integer points with weights converging to those values, or the central claim fails. For the decomposition classification, take a discrete Schrödinger operator with sparse potential whose known theory predicts purely singular continuous spectrum on $(0,4)$; the two-limit algorithm must return $\\sigma_{sc}\\cap(0,4)=(0,4)$ with $\\sigma_{ac}$ and $\\sigma_{pp}$ empty, otherwise the sharp lower bound is wrong.","tokens_in":57294,"feed_emoji":"📐","tokens_out":15516,"duration_ms":149603,"temperature":0.7,"pith_summary":"The paper claims the first general algorithms for computing spectral measures and spectral types of infinite-dimensional self-adjoint and unitary operators, under one structural assumption: the operator is supplied as a matrix whose columns decay at a known asymptotic rate, together with a matching decay rate for the vector. For this class, the projection-valued spectral measure of any open set is computed by a single convergent limit of arithmetic algorithms, and scalar spectral measures follow by inner products. Functional calculus and, on sets separated from the singular and point supports, the Radon--Nikodym derivative of the absolutely continuous part are also one-limit computations. The paper additionally proves that decomposing measures into pure point, absolutely continuous and singular continuous parts is genuinely harder: two limits are necessary and sufficient, and the singular continuous spectrum needs three limits once the bandwidth grows fast enough. Because the operators may be unbounded, the same machinery transfers to partial differential operators such as linear evolution operators on $L^2(\\mathbb{R}^d)$, where only point samples of the coefficients are needed.","feed_headline":"Infinite matrices yield spectral measures in one limit","feed_subtitle":"With known column decay, resolvent arithmetic computes spectra; separating spectral types provably needs extra limits.","key_machinery":"The load-bearing object is the resolvent engine of Theorem 2.1: rather than taking square truncations of the infinite matrix, it solves a rectangular least-squares system that approximates $R(z,T)x$ and proves a tail bound using the known decay profile and $\\operatorname{dist}(z,\\sigma(T))$. Spectral measures are then accessed through the boundary-limit identity of Proposition 2.3, which expresses the projection-valued measure as a Poisson-kernel convolution of the resolvent; the algorithms integrate this convolution with quadrature while letting the smoothing parameter tend to zero. The Solvability Complexity Index (SCI) hierarchy, which counts the minimum number of successive limits any algorithm must take, is the classification tool that separates the one-limit results from the intrinsically harder decomposition and spectral-set problems.","core_discovery":"The central claim is that for every $T$ in the class $\\Omega_{f,\\alpha,\\beta}$ of self-adjoint or unitary operators with known column decay $\\|(I-P_{f(n)})TP_n\\|=O(\\alpha_n)$, and every vector with decay $\\|P_nx-x\\|=O(\\beta_n)$, the map $(T,x,U)\\mapsto E_T(U)x$ is computable in one limit by arithmetic algorithms; the scalar measures $\\mu^T_{x,y}(U)=\\langle E_T(U)x,y\\rangle$ come from inner products. The engine that makes this possible is a resolvent algorithm with error control, built from rectangular least-squares truncations rather than square truncations. The same engine powers one-limit algorithms for the functional calculus $F(T)x$ and for $L^1$ approximation of the Radon--Nikodym derivative on open sets strictly separated from the singular and point supports. The paper's classifications are sharp: Theorem 3.2 places the measure-decomposition problems in $\\Delta^A_3\\setminus\\Delta^G_2$, and Theorem 5.1 places $\\sigma_{ac}$ and $\\sigma_{pp}$ in $\\Delta^A_3\\setminus\\Delta^G_2$ while $\\sigma_{sc}$ lies in $\\Delta^A_4$ and requires three limits when $f(n)-n\\geq\\sqrt{2n}+1/2$. For partial differential operators whose coefficients are polynomially bounded and of locally bounded variation, the same resolvent engine transfers, giving computable spectral measures, functional calculus, and densities, along with the corresponding decomposition towers, from point-sample data.","pith_inferences":["The same resolvent-plus-Poisson-kernel template should extend to other spectral observables expressible as boundary integrals of the resolvent, such as local densities of states or autocorrelation spectra, whenever the kernel has enough regularity for quadrature convergence.","The lower bounds on the singular continuous spectrum suggest a guiding principle for numerical work on quasiperiodic or random operators: any method that obtains this spectrum in practice must either exploit extra structure or accept an intrinsically slower, multi-limit convergence.","Because the paper leaves general normal operators with interior spectral points open, a natural test is whether its generalized boundary-integral formula can be turned into a one-limit algorithm for operators whose spectrum is a rectifiable curve; the paper does not claim this.","One could benchmark the algorithms against exactly solvable critical models with singular continuous spectra, using the SCI lower bounds to predict where convergence must slow; this is an editorial suggestion, not a paper claim."],"forward_implications":["Any self-adjoint or unitary operator with a known column-decay profile now has its spectral measures, functional calculus, and absolutely-continuous densities computable by a single convergent limit, so spectral computations no longer require a closed-form expression for the measure.","The sharp SCI classifications mean the two-limit and three-limit towers are not an implementation deficiency: no algorithm, in any model of computation, can perform the decompositions with fewer limits.","For linear evolution equations on $L^2(\\mathbb{R}^d)$ in the stated coefficient class, the semigroup action can be computed with guaranteed convergence from point samples of the coefficients, not from analytic spectral data.","Given the recurrence coefficients of orthogonal polynomials, the associated measure can be recovered numerically, giving computational substance to the classical correspondences between such coefficients and measures.","On quasicrystal graph models with growing bandwidth, where powering the matrix is infeasible, the functional-calculus algorithm solves fractional diffusion by contour integrals of the resolvent, with exponential convergence in the holomorphic case."],"supporting_citations":[{"why":"Supplies the SCI-hierarchy framework and the Baire-category impossibility result used in the three-limit lower bound for the singular continuous spectrum.","marker":"[12]"},{"why":"Shows the inner products needed to transfer the Hilbert-space algorithms to partial differential operators can be evaluated with asymptotic error control from point samples.","marker":"[37]"},{"why":"Provides the perturbed Anderson-localization result used to construct potentials for which point-spectrum computations cannot converge in one limit.","marker":"[67]"},{"why":"Gives the sparse-potential dichotomy between purely absolutely continuous and purely singular continuous spectra used in the lower bounds for those spectral sets.","marker":"[92]"},{"why":"Shows finitely supported potentials carry no point spectrum on $(0,4)$, a step in the impossibility proof for computing the point spectrum.","marker":"[114]"},{"why":"Supplies the closest prior algorithm class, tridiagonal compact perturbations of Toeplitz operators, that the new resolvent method generalizes.","marker":"[151]"}],"fun_headline_variants":["First general algorithm for spectral measures","Resolvent arithmetic gives spectral measures","One-limit spectral measures for infinite operators","Spectral measures via rectangular truncations","General computability of spectral decompositions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the algorithm is handed the exact asymptotic decay profile of the matrix columns and vector (the function $f$ and null sequences $\\alpha,\\beta$), so that the resolvent tail bound is certified; if that structural information is missing or inaccurate, the one-limit computability results do not apply.","fun_headline_variants_meta":{"raw":{"variants":["First general algorithm for spectral measures","Resolvent arithmetic gives spectral measures","One-limit spectral measures for infinite operators","Spectral measures via rectangular truncations","General computability of spectral decompositions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000383,"raw_usage":{"total_tokens":2151,"prompt_tokens":1192,"completion_tokens":959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":808,"completion_tokens_details":{"reasoning_tokens":898}},"tokens_in":808,"tokens_out":959,"duration_ms":9607,"temperature":1.0,"reasoning_tokens":898,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:39:46.859933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the one-limit algorithm on a Jacobi matrix whose spectral measure is known to be the pure-point measure with atoms at the nonnegative integers and weights $\\exp(-\\alpha)\\alpha^m/m!$, letting $n$ grow as $\\epsilon\\downarrow 0$: the output must concentrate at the integer points with weights converging to those values, or the central claim fails. For the decomposition classification, take a discrete Schrödinger operator with sparse potential whose known theory predicts purely singular continuous spectrum on $(0,4)$; the two-limit algorithm must return $\\sigma_{sc}\\cap(0,4)=(0,4)$ with $\\sigma_{ac}$ and $\\sigma_{pp}$ empty, otherwise the sharp lower bound is wrong.","supporting_citations":[{"cited_title":"Schr¨ odinger operators with sparse potentials: asymptotics of the Fourier transform of the spectral measure","cited_arxiv_id":null,"evidence_quote":"Gives the sparse-potential dichotomy between purely absolutely continuous and purely singular continuous spectra used in the lower bounds for those spectral sets."},{"cited_title":"The absolutely continuous spectrum of one-dimensional Schr¨ odinger operators with decaying potentials","cited_arxiv_id":null,"evidence_quote":"Shows finitely supported potentials carry no point spectrum on $(0,4)$, a step in the impossibility proof for computing the point spectrum."},{"cited_title":"Spectra of Jacobi operators via connection coefficient matrices","cited_arxiv_id":null,"evidence_quote":"Supplies the closest prior algorithm class, tridiagonal compact perturbations of Toeplitz operators, that the new resolvent method generalizes."}],"review_version":1}