--- title: Non-Asymptotic Spectral Bands url: https://www.emergentmind.com/topics/non-asymptotic-spectral-bands type: topic --- # Non-Asymptotic Spectral Bands Searching arXiv for the cited works to ground the article in current arXiv metadata. Non-asymptotic spectral bands are finite-scale, quantitatively controlled descriptions of spectral support, spectral density, or spectral estimation error that hold without passing to a limit such as dimension \(n\to\infty\), bandwidth \(W\to\infty\), coupling \(\lambda\to\infty\), or sample size \(N\to\infty\). Across random operators, periodic graph Laplacians, non-normal Toeplitz perturbations, time-series spectrum estimators, integral operators, and inverse problems, the common objective is to replace asymptotic support statements by explicit inequalities for spectral intervals, spectral measures, eigenvalue locations, or confidence envelopes at fixed problem size. In this sense, “spectral bands” refers variously to actual bands of an operator spectrum, deterministic images such as \(f(S^1)\), high-probability intervals for eigenvalues, and finite-sample uncertainty bands for estimated spectra [1101.4413], [2106.04785], [1312.6510], [1301.2382], [2303.11908]. ## 1. Concept and scope In the non-asymptotic setting for random matrices, one is interested in high-probability “bands” \([\lambda_1(A),\lambda_n(A)]\) within which the spectrum of \(A\) lies, typically of the form \([\mu_1-\Delta,\mu_2+\Delta]\), where \(\mu_1,\mu_2\) are deterministic centering constants and \(\Delta\) is a fluctuation bound of order \(O(f(n))\) [1301.2382]. This formulation treats a spectral band as a finite-sample enclosure of the whole spectrum. A different but related meaning appears for periodic and almost-periodic operators. For Laplacians on periodic equilateral metric graphs, the spectrum of the Laplacian consists of an absolutely continuous part, which is a union of an infinite number of non-degenerated spectral bands, plus an infinite number of flat bands [1312.6510]. For banded Toeplitz matrices, the limiting empirical spectral measure is supported on the plane curve \(f(S^1)\), identified as “the bands” [2106.04785]. For the Kohmoto model, periodic approximants \(H_{p_k/q_k,\lambda}\) have \(q_k\) distinct bands, and all spectral bands admit a hierarchical structure for all non-vanishing coupling constants [2607.06361]. A further meaning arises in statistical spectral estimation. There, a non-asymptotic spectral band is a finite-sample confidence envelope for an estimated spectral density or covariance spectrum. For classical quadratic estimators, one obtains uniform bands of the form \[ \forall\,\omega,\quad \Pr\Bigl(\,S(\omega)\in[\, \hat S(\omega)-\Delta,\;\hat S(\omega)+\Delta\,]\Bigr)\ge 1-\delta \] under explicit variance and bias conditions [2303.11908]. For \(L\)-mixing processes with unknown means, analogous bands are built from explicit \(L_{2q}\)-norm bounds and moment-to-probability conversion [2504.00217]. These usages differ in object and technique, but all are finite-scale and quantitative. *This suggests* that “non-asymptotic spectral bands” is best understood as a unifying methodological category rather than a single formal notion. ## 2. Random operators and finite-scale spectral regularity For random band matrices, Sodin studies a symmetric band operator \(H(W)\) on \(\ell^2(\mathbb Z)\) with bandwidth \(W\) and random \(\pm 1\)-entries scaled by \((2W-1)^{-1/2}\), where \(H(u,v)=0\) unless \(0<|u-v|\le W\), and the nonzero entries are independent Rademacher signs [1101.4413]. The average spectral measure at the origin is the probability measure \(\mu_W\) characterized by its Stieltjes transform \[ G_W(z)=\int_{\mathbb R}(E-z)^{-1}\,d\mu_W(E)=\mathbb E[(H-z)^{-1}(0,0)],\qquad \Im z>0. \] The main theorem fixes \(E_0\in(-1,1)\) and proves that there is \(C=C(E_0)<\infty\), uniform on compact subintervals of \((−1,1)\), so that for every \(W\gg 1\) and every \(\epsilon\ge W^{-0.99}\), \[ \Bigl|\,G_W(E_0+i\epsilon)-\int_{-1}^{1}\frac{a_0(E)\,dE}{E-E_0-i\epsilon}\Bigr|\le \frac{C}{W}, \] where \(a_0(E)=\tfrac{2}{\pi}\sqrt{1-E^2}\) is the semicircle density on \(E\in(-1,1)\) [1101.4413]. Equivalently, \[ \Bigl|\,\mu_W([E_0-\epsilon,E_0+\epsilon])-\int_{E_0-\epsilon}^{E_0+\epsilon}a_0(E)\,dE\Bigr| \le O\!\bigl(\tfrac{\epsilon}{W}\bigr). \] The conclusion given in the source is that at any “mesoscopic” scale \(\epsilon\gg W^{-1}\), the averaged spectral measure \(\mu_W\) is as regular as the semicircle law itself, up to an \(O(1/W)\) error, and that no fine-scale band gaps or singularities can persist above the scale \(W^{-1}\) [1101.4413]. This is a prototypical non-asymptotic band statement: it does not identify individual eigenvalue intervals, but it controls local spectral mass down to \(\epsilon=W^{-0.99}\). The same work further gives a formal series in powers of \((2W-1)^{-1}\), \[ G_W(E_0+i\epsilon)\sim \sum_{k=0}^{m-1}\frac{A_k(E_0,i\epsilon)}{(2W-1)^k}+R_m(W;E_0,\epsilon), \] with \[ |R_m(W;E_0,\epsilon)|\le C_m(E_0)(2W-1)^{-m} \] uniformly for \(\epsilon\ge W^{-0.99}\), and in particular one may take \(m\approx W^{0.99}\) so that \(R_m\) is exponentially small in \(W\) [1101.4413]. This yields finite-mesoscopic asymptotics rather than only leading-order regularity. In the broader non-asymptotic random matrix literature, explicit spectral bands arise from concentration inequalities. Matrix Bernstein gives \[ \Pr\!\Bigl(\Bigl\|\sum_{k=1}^m X_k\Bigr\|\ge t\Bigr)\le 2n\cdot \exp\!\Bigl(\frac{-t^2}{2\sigma^2+(2/3)Mt}\Bigr) \] for independent mean-zero Hermitian random matrices with \(\|X_k\|\le M\) almost surely, while Matrix Chernoff bounds control \(\lambda_{\min}\) and \(\lambda_{\max}\) for sums of independent positive semidefinite matrices [1301.2382]. In examples, Wigner matrices satisfy \[ \Pr(\|A\|\ge 2\sigma\sqrt n+t)\le 2\exp(-c t^2/K^4), \] and rectangular subgaussian matrices satisfy \[ \Pr(s_1(A)\ge \sqrt N+\sqrt n+t)\le 2\exp(-c t^2),\qquad \Pr(s_n(A)\le \sqrt N-\sqrt n-t)\le 2\exp(-c t^2) \] [1301.2382]. These are high-probability spectral bands in the sense of deterministic intervals containing the spectrum. A common misconception is that non-asymptotic theory is merely a coarse precursor to asymptotic edge analysis. The survey explicitly notes that sharpening \(O(\sqrt n)\to O(n^{-2/3})\) requires eigenvalue-specific methods such as Tracy–Widom theory, but the finite-sample bands remain robust and widely applicable in statistics, signal processing, and computational geometry [1301.2382]. The distinction is therefore between general explicit control and edge-optimal fluctuation theory, not between “useful” and “crude” results. ## 3. Deterministic band geometry in periodic and Toeplitz settings For periodic equilateral metric graphs \(\Gamma=(V,E)\), Korotyaev and Saburova work with a connected, locally finite, \(\mathbb Z^d\)-periodic graph embedded in \(\mathbb R^d\), with all edges identified with \([0,1]\), and analyze the momentum operator \(\sqrt{\Delta_M}\) subject to Kirchhoff conditions [1312.6510]. Through Floquet–Bloch decomposition, one obtains fiber operators \(\sqrt{\Delta_M}(\theta)\) on \(L^2(\Gamma_f)\), each with spectrum consisting of \(\nu\) continuous bands plus flat bands, and an associated normalized discrete Laplacian \(A\) on \(\ell^2(V)\). Denoting by \(B_n\) the number of bridges incident to a vertex \(u_n\) in the fundamental graph and by \(x_n\) its degree, the geometric parameter \[ B=\sum_{n=1}^{\nu}\frac{B_n}{x_n} \] governs the total band measure [1312.6510]. Theorem 1.1 gives the two-sided comparison \[ |\Sigma_n(A)|\le |\Sigma_n(2)|\le \tfrac{\pi}{2}\,|\Sigma_n(A)|,\qquad n=1,\dots,\nu, \] and the total-measure bounds \[ |\Sigma(A)|\le 2B,\qquad |\Sigma(2)|\le \pi B. \] If there are \(s\) nontrivial gaps \(\gamma_k(2)\) in \([0,\pi]\), then \[ \sum_{k=1}^s|\gamma_k(2)|>\pi(1-B). \] In particular, if \(B<1\), then \(|\Sigma(2)|<\pi\), so there must exist infinitely many spectral gaps of \(\sqrt{\Delta_M}\), and hence infinitely many gaps of \(\Delta_M\) [1312.6510]. The proofs are explicitly non-asymptotic. No Weyl law is invoked; the estimates are exactly finite-dimensional trace bounds on the \(\nu\times \nu\) Floquet matrices \(A(\theta)\) plus elementary trigonometric distortion [1312.6510]. The source also stresses that they apply to every band in the entire absolutely continuous spectrum, not only to high-energy bands. For banded Toeplitz matrices, the object is non-normal rather than self-adjoint. Let \[ A_n=\mathrm{Toeplitz}_n(a_j)_{|j|\le b},\qquad f(e^{i\theta})=\sum_{j=-b}^{b} a_j e^{ij\theta}, \] and perturb \(A_n\) by a small additive matrix \(\epsilon E_n\), where \(E_n\) may be random with iid entries of mean \(0\), variance \(1\), finite moments or even heavy tails, or a non-random low-rank or bounded-entry adversarial matrix [2106.04785]. Under \[ b=o(n/\log n),\qquad \sum_{j\in\mathbb Z}|j\,a_j|<\infty,\qquad \epsilon_n\|E_n\|\to 0, \] the empirical spectral distribution converges to \(\mu\), the law of \(f(U)\) for \(U\) uniform on \(S^1\), and \(\mu\) is supported on the plane curve \(f(S^1)\), described as “the bands” [2106.04785]. The non-asymptotic local law states that for a smooth compactly supported \(\phi\), point \(z_0\in\mathbb C\), and mesoscopic scale \(a\ge 0\), if \[ \phi_{z_0,a}(z)=\phi(n^a(z-z_0)), \] then under the iid-or-unitary perturbation model, for any \(\epsilon>0\) and \(\kappa>0\) there is \(C<\infty\) so that, with probability at least \(1-O(n^{-\kappa})\), \[ \Bigl|\sum_{j=1}^{n}\phi_{z_0,a}(\lambda_j(A_n+\epsilon E_n))-\sum_{j=1}^{n}\phi_{z_0,a}(f(e^{2\pi i j/n}))\Bigr| \le C\bigl(b\log n+n^{a+1}\|E_n\|_2\,\epsilon\bigr) \] [2106.04785]. When \(b=O(1)\) and \(\epsilon=O(n^{-\gamma})\) with \(\gamma>1\), the error is \(O(\log n)\) uniformly for all \(0\le a\le \gamma-1\). The same paper defines classical locations \[ \gamma_j=f(e^{2\pi i j/n}),\qquad j=1,\dots,n, \] and proves rigidity: with high probability, \[ \min_\pi\Bigl(n^{-1}\sum_j |\lambda_j(A_n+\epsilon E_n)-\gamma_{\pi(j)}|^p\Bigr)^{1/p}=O(n^{-\delta}) \] for some \(\delta>0\), and equivalently \[ |\lambda_j(A_n+\epsilon E_n)-\gamma_{\sigma(j)}|=O(n^{-1+\epsilon}) \] with an optimal near \(1/n\) rate up to logs [2106.04785]. These two examples show two distinct deterministic notions of non-asymptotic bands: total-measure and gap estimates for self-adjoint periodic spectra, and finite-\(n\) localization around a deterministic curve for non-normal Toeplitz matrices. ## 4. Hierarchical and operator-theoretic band structures The Kohmoto model provides a non-asymptotic hierarchical description of spectral bands for a family of discrete Schrödinger operators with Sturmian potentials. For frequency \(\alpha\in[0,1]\) and coupling \(\lambda\ne 0\), \[ (H_{\alpha,\lambda}\psi)(n)=\psi(n+1)+\psi(n-1)+\lambda\,\chi_{[1-\alpha,1)}(n\alpha+\phi \bmod 1)\,\psi(n), \] and if \(\alpha_k=p_k/q_k\) is the \(k\)-th convergent of \(\alpha=[0;a_1,a_2,\dots]\), then \(H_{p_k/q_k,\lambda}\) is \(q_k\)-periodic with \(q_k\) distinct bands [2607.06361]. The band structure is encoded by transfer-matrix trace recurrences. For suitable continued-fraction data \(c\), one defines \(M_c(E,\lambda)\) and its trace \(t_c(E,\lambda)\), with a fundamental recursion of trace-map type and a Chebyshev-form rewriting using the dilated Chebyshev-II polynomials \(S_n(x)\) [2607.06361]. For the periodic approximant \(\alpha_k=p_k/q_k\), the \(q_k\) bands are exactly the intervals \(I_{k,\ell}=[E_{k,\ell}^-,E_{k,\ell}^+]\) on which \(|t_k(E,\lambda)|\le 2\), and their edges satisfy \[ t_k(E_{k,\ell}^{\pm},\lambda)=\pm 2. \] The next-generation edges are determined by \(t_{k+1}(E,\lambda)=\pm 2\), equivalently by an exact algebraic relation involving \(S_{a_{k+1}}\) [2607.06361]. The source states that non-asymptotic width inequalities valid for all finite \(\lambda\) follow from Chebyshev bounds, and that repeated use of the Chebyshev-trace estimates yields non-asymptotic upper and lower bounds for the length of the \(\ell\)-th band at level \(k\), \[ C_1(\lambda)\,q_k^{-1}\,\lambda^{-a_{k+1}}\le \Delta_{k,\ell}(\lambda)\le C_2(\lambda)e^{-c q_k}, \] where the constants \(C_1,C_2,c>0\) can be written explicitly in terms of the first \(k\) continued-fraction entries and \(\lambda\) [2607.06361]. The result is explicitly distinguished from large-coupling or small-coupling expansions: every band interval, at every rational approximation level, is controlled by exact algebraic trace-map recurrences and Chebyshev inequalities, with no passage to \(\lambda\to\infty\) or \(\lambda\to 0\). A different operator-theoretic realization of non-asymptotic bands appears for the Sinc-kernel operator \[ Q_cf(x)=\int_{-1}^{1}\frac{\sin(c(x-y))}{\pi(x-y)}\,f(y)\,dy, \] whose discrete spectrum is \[ 1>\lambda_0(c)\ge \lambda_1(c)\ge \cdots \ge 0 \] with simple eigenvalues and prolate spheroidal wave functions [1804.01257]. The spectrum splits into a bulk region \(0\le n\le n_c-\Delta\), plunge region \(|n-n_c|\le \Delta\), and tail region \(n\ge n_c+\Delta\), where \(n_c=\frac{2c}{\pi}\) and \(\Delta\asymp \log c\). The source gives region-wise non-asymptotic bounds, including a plunge-region two-sided estimate with explicit constants \(A_2=5\), \(B_2=0.069\), \(A_3=1\), \(B_3=0.12\), and \(\Delta=\log c+6\), valid for \(c\ge 22\): \[ A_2\exp(-B_2(n-n_c))\le \lambda_n(c)\le A_3\exp(-B_3|n-n_c|). \] It also gives a tail upper bound \[ \lambda_n(c)\le \exp(-0.12(n-c)) \] for every \(c>0\) and every integer \(n\ge \max\{2,\;c+\log n+9\}\) [1804.01257]. These results illustrate a central feature of the topic: non-asymptotic band theory can refer not only to spectral support intervals, but also to finite-index partitions of the eigenvalue sequence into bulk, transition, and tail zones with explicit numerical inequalities. ## 5. Finite-sample confidence bands in spectral estimation For classical spectrum estimation, non-asymptotic bands quantify the deviation of an estimator \(\hat S(\omega)\) from either its mean \(\mathbb E[\hat S(\omega)]\) or the true spectrum \(S(\omega)\). In the framework of quadratic spectral estimators, \[ \hat S(\omega)=\sum_{k=-(n-1)}^{n-1} b[k]\,\hat R[k]\,e^{-ik\omega}, \] the estimators include Blackman–Tukey, Bartlett, and Welch [2303.11908]. Under either a zero-mean Gaussian stationary process with bounded spectral density and absolutely summable autocovariance, or a sub-Gaussian linear process, one can write \[ \hat S(\omega)=y^\top A(\omega) y \] for a suitable Hermitian matrix \(A(\omega)\) and derive pointwise and uniform bounds. For each fixed \(\omega\), the source states that \[ \Pr(|\hat S(\omega)-\mathbb E[\hat S(\omega)]|>\varepsilon)\le \delta \] whenever \[ \frac1{\xi(A)}\ge \alpha(\varepsilon)\,\beta(\delta), \qquad \xi(A)=\max\{\|A\|_2,\|A\|_F^2\}, \] where \(\alpha(\varepsilon)\) and \(\beta(\delta)\) are explicit functions of constants \(c_{\rm mult},c_{\rm exp},c_{\rm sub}\), \(\|S\|_\infty\), \(\varepsilon\), and \(\delta\) [2303.11908]. For uniform-in-\(\omega\) control, if only \(N_0\) lag-blocks appear and \[ \frac1g\ge \alpha(\varepsilon/2)\,[\ln(5N_0^2)+\beta(\delta/2)], \] then \[ \Pr\!\Bigl(\sup_{\omega\in[-\pi,\pi]}|\hat S(\omega)-\mathbb E[\hat S(\omega)]|>\varepsilon\Bigr)\le \delta. \] Bias is controlled through \[ \mathbb E[\hat S(\omega)]-S(\omega)=\sum_{k=-\infty}^\infty (1-b[k])R[k]e^{-ik\omega}, \] and if \(b[k]\in[0,1]\) and the bias-cutoff condition holds, then \[ \sup_\omega |\mathbb E[\hat S(\omega)]-S(\omega)|\le \varepsilon. \] Combining the two yields the finite-sample uniform band \[ \forall\,\omega,\quad \Pr\Bigl(S(\omega)\in[\hat S(\omega)-\Delta,\hat S(\omega)+\Delta]\Bigr)\ge 1-\delta \] for suitable \(\Delta\) [2303.11908]. Specialized rates are given for Bartlett and Welch. For Bartlett with \(n=LM\), one has \(\xi(A)=\|A\|_2=\|A\|_F^2=M/n\), \(N_0=M\), and \(b[k]=1-|k|/M\). The source states that balancing bias \(M^{-1}\) against variance \(\sqrt{M/n}\) gives \(M\approx n^{1/3}\) and overall band-width \(O(n^{-1/3})\) [2303.11908]. For Welch, variance scales like \(\sqrt{\frac{1+2M/K}{S}}\) and bias like \(O(M^{-1})\), again yielding \(O(S^{-1/3})\) under fixed overlap ratio. For \(L\)-mixing data, the same problem is treated beyond Gaussian and linear-process settings. In the zero-mean case, if \((\mathbf y_k)\) is strictly stationary and \(L\)-mixing with finite \(M_q(\mathbf y)\) and \(\Gamma_{d,q}(\mathbf y)\), then for the block-averaged estimator \(\hat\Phi_k(s)\), \[ \|\hat\Phi_k(s)-\bar Z\|_{L_q}\le \frac{b_q}{\sqrt{k}\,\log_2(\log_2 k)} \] for every integer \(k\ge 4\), where \(b_q\) is explicit in \(M_{4q}(\mathbf y)\), \(\Gamma_{d,4q}(\mathbf y)\), \(M\), and \(K\) [2410.02951]. Under the growth condition \(b_q\le c q^r\), a high-probability bound follows: \[ \mathbb P\Bigl(\|\hat\Phi_k(s)-\bar Z\|_F> \frac{c\,\log_2(\log_2 k)} {\sqrt{k}\,e^r\,\max\{1,(\ln \tfrac1\delta)^r/r^r\}}\Bigr)\le \delta. \] Bias bounds are then added to form a finite-sample confidence band for \(\|\hat\Phi_k(s)-\Phi(s)\|_2\) [2410.02951]. For unknown means, batch and online estimators are treated for real-valued or vector-valued \(L\)-mixing processes with mean \(\mu\). The batch estimator \(\hat\Phi_{B,k}(s)\) uses a global sample mean \(\hat\mu_k\), while the online estimator \(\hat\Phi_k(s)\) updates \(\hat\mu_i\) recursively with \(\alpha_i=1/(i+1)\) [2504.00217]. The main \(L_{2q}\)-bounds are \[ \|\hat\Phi_{B,k}(s)-\bar\Phi(s)\|_{L_{2q}}\le C_B(q,M,K)\cdot \frac1{\sqrt{k}}, \] and \[ \|\hat\Phi_k(s)-\bar\Phi(s)\|_{L_{2q}}\le C_O(q,M,K)\cdot \Bigl(\frac1{\sqrt{k}}+D(M,q)\frac{M}{k}\Bigr), \] with explicit formulas for \(C_B,C_O,D\) in terms of \(c_{M,2q},c_{\Gamma,2q},c_{2q},M_{4q}(y)\), and \(\Gamma_{d,2q}(y)\) [2504.00217]. The same source states that the obtained error bounds are of \(O(1/\sqrt{k})\), which are tighter than previous results under the zero-mean assumption. A recurring misconception is that non-asymptotic spectral estimation theory only yields pointwise error bars. The literature summarized here includes both pointwise and worst-case-over-frequency bounds, as well as explicit procedures to construct uniform confidence envelopes [2303.11908], [2410.02951], [2504.00217]. ## 6. Methods and proof architectures The methodologies used to obtain non-asymptotic spectral bands vary by problem class, but several recurrent architectures are explicit in the sources. For random band matrices, the proof begins with a Chebyshev/non-backtracking expansion of the resolvent, \[ (H-z)^{-1}(0,0)= -\sum_{n=0}^\infty T_n(H)(0,0)\,\frac{U_{n-1}(z)}{\sqrt{z^2-1}}, \] where \(T_n(H)(0,0)\) is expressed as a sum over non-backtracking paths of length \(n\) [1101.4413]. Divergence is controlled by a smooth cutoff \(y_\epsilon(n)=y(n\epsilon)\), every non-backtracking path is grouped into topological equivalence classes (“diagrams”), and embeddings of diagrams are counted through Fourier variables \(\xi_e\) subject to Kirchhoff-flux constraints \(\sum_{e\ni v}\mathrm{sgn}_e\,\xi_e=0\). Divided-difference and saddle-point estimates are then combined with the spectral gap \(|w(\xi)|<1-c\min\{\xi,1-\xi\}\) of the non-backtracking walk to show that each genus-\(y\) diagram is bounded by \(O(W^{-y}(\log W)^{y+1})\), and summing over \(y\ge 1\) yields the \(1/W\) remainder control [1101.4413]. For perturbed Toeplitz matrices, the key tools are comparison principles and singular-value control. The source lists a non-asymptotic replacement principle, a rank-comparison estimate for converting Toeplitz matrices to circulant ones, a norm-comparison estimate for small-noise perturbations, and least singular-value bounds ensuring the comparison is valid [2106.04785]. This shows that non-asymptotic band localization in non-normal problems often proceeds via logarithmic potentials and singular-value stability rather than direct eigenvalue perturbation. For periodic metric graphs, the proof is based on Cattaneo’s correspondence between the momentum spectrum and the discrete spectrum under the map \(z\mapsto -\cos z\), a bridge-trace bound on the Floquet matrices \(A(\theta)\), and a cosine-map distortion estimate [1312.6510]. The essential point is that the band-measure estimates are reduced to computable finite-dimensional quantities attached to the fundamental cell. For classical spectrum estimators, the generic mechanism is to represent the estimator as a quadratic form \(y^\top A(\omega)y\), derive concentration in terms of \(\|A\|_2\) and \(\|A\|_F\), and then control uniformity over frequency through a finite covering of the frequency domain, which produces the factor \(\ln(5N_0^2)\) [2303.11908]. In the \(L\)-mixing works, the central tool is an \(L_{2q}\) norm inequality for weighted sums of zero-mean \(L\)-mixing processes, combined with explicit bookkeeping of overlap, windowing, mean-estimation error, and bias terms [2410.02951], [2504.00217]. For passive imaging through convolutive channels, Lee, Krahmer, and Romberg analyze a cross-correlation matrix \(C_y=Y^*Y\) and its subspace-constrained version \[ \widehat C=\Phi^*(Y^*Y-\sigma_w^2(M-1)L I)\Phi. \] The stability of the estimator is controlled by the spectral gap \[ \Delta=\lambda_2(C)-\lambda_1(C), \] and under the subspace model the ideal gap satisfies \[ \mathrm{Gap}=\lambda_2(C)-\lambda_1(C)\ge \frac{K^2\|x\|_2^2\|u\|_2^2}{2}. \] If \(\|\Delta C\|\le \mathrm{Gap}/5\), then Davis–Kahan gives \[ \sin\Theta(u,\hat u)\le 4\|\Delta C\|/\mathrm{Gap} \] [1708.04343]. The proof decomposes \(\Delta C\) into signal-signal, cross, and noise-noise terms, and bounds each block through concentration of chaos processes. This is a non-asymptotic band argument in the sense of gap-versus-perturbation control. Across these examples, spectral bands are controlled by one of four mechanisms: combinatorial expansions, Floquet decompositions, comparison principles based on singular values or ranks, and concentration/perturbation theory. *A plausible implication is* that the term “band” masks a deeper methodological unity: explicit control of spectral structure through finite-dimensional or finite-scale surrogates. ## 7. Applications, limitations, and interpretive boundaries Non-asymptotic spectral bands are used in several application domains explicitly mentioned in the sources. In time-series analysis, finite-sample spectral envelopes are relevant to economics, astronomy, climatology, speech analysis, seismology, and control systems [2504.00217], [2410.02951]. In passive imaging, subspace-constrained spectral methods improve robustness in multichannel blind deconvolution and are evaluated numerically in underwater-acoustics subspaces [1708.04343]. For periodic metric graphs, the band-measure bounds are described as useful in inverse spectral problems on quantum graphs and in wave-guide design [1312.6510]. For the Sinc kernel operator, non-asymptotic eigenvalue bounds are applied to Remez and Turán–Nazarov inequalities and to the GUE hole probability [1804.01257]. The Sinc-kernel application is especially explicit. If \(T_2(n,E)\) denotes the best constant in the \(L^2\)-Remez inequality, then \[ T_2(n,E)\le 2\sqrt{\lambda_n\!\bigl(\tfrac{n+1}{E}\bigr)}, \] which yields \[ T_2(n,E)\le 2\exp\!\Bigl(-\tfrac12\,n\bigl[\ln((n+1)/E)-1\bigr]\Bigr), \qquad 0