Introduction
A point cloud is often analysed through a weighted graph on its points, in which the weight of the edge between two points is given by a kernel, a function of their distance and a bandwidth ε \varepsilon ε . The unnormalized graph Laplacian of this graph is the diagonal matrix of degrees minus the matrix of weights, where the degree of a point is the sum of the weights of its edges (see, e.g., [28 ] ). The first (smallest) eigenvalues of this matrix, or of a normalized version of it, and their eigenvectors are used to cluster the points [28 ] , to describe the geometry of the point cloud with diffusion maps [11 ] , and to embed the points in a low-dimensional space with Laplacian eigenmaps [2 ] .
When the points are i.i.d. samples from a density on a manifold, the unnormalized graph Laplacian approximates a weighted Laplace–Beltrami operator [7 , 16 ] , a continuum operator determined by the manifold and the density (see, e.g., [17 ] ). This approximation has been studied through pointwise limits [3 , 20 , 21 , 26 ] , spectral convergence on domains in R d \mathbb{R}^d R d [19 ] , and quantitative spectral estimates for discretizations of a manifold and samples from it [4 , 16 ] . Calder and García Trillos [7 ] prove that, with high probability, the first eigenvalues of the graph Laplacian approximate those of the weighted Laplace–Beltrami operator at rate O ( ε ) O(\varepsilon) O ( ε ) if ε \varepsilon ε tends to zero slowly enough as the number of points grows. They also prove, under the same conditions, that the corresponding unit eigenvectors approximate eigenfunctions of this operator at the same rate, even for a multiple eigenvalue. For a simple eigenvalue and slightly different normalizations, García Trillos, Li and Venkatraman [17 ] improve both rates to O ( ε 2 log ( 1 / ε ) ) O(\varepsilon^2\log(1/\varepsilon)) O ( ε 2 log ( 1/ ε )) at the same bandwidth.
In practice the points are observed with noise, which changes the distances between them and may move them off the manifold. We ask how much the first eigenvalues of the graph Laplacian and their eigenvectors can change when each point is moved by at most δ \delta δ . We consider the unnormalized graph Laplacians of the points and of their noisy observations, built with the same kernel and the same bandwidth, and call them the clean and noisy graph Laplacians. For arbitrary noise bounded by δ \delta δ , we prove that, for each fixed l l l , if δ / ε \delta/\varepsilon δ / ε is small enough, then with high probability, the l l l -th eigenvalue of the noisy graph Laplacian differs by O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) from that of the clean one. If moreover the l l l -th and ( l + 1 ) (l+1) ( l + 1 ) -st eigenvalues of the weighted Laplace–Beltrami operator are distinct, then the orthogonal projection onto the subspace spanned by the first l l l eigenvectors of the noisy graph Laplacian differs by O ( δ / ε ) O(\sqrt{\delta}/\varepsilon) O ( δ / ε ) from the corresponding projection for the clean one. Combined with the rate O ( ε ) O(\varepsilon) O ( ε ) of Calder and García Trillos [7 ] for clean points, our bounds show that the first eigenvalues of the noisy graph Laplacian approximate those of the weighted Laplace–Beltrami operator at rate O ( ε + δ / ε ) O(\varepsilon+\delta/\varepsilon) O ( ε + δ / ε ) , and that the first eigenvectors of the noisy graph Laplacian approximate the eigenfunctions of this operator at rate O ( ε + δ / ε ) O(\varepsilon+\sqrt{\delta}/\varepsilon) O ( ε + δ / ε ) .
Noise that is not small compared with the bandwidth can sometimes be handled through its structure. Several results concern random noise in high ambient dimension. El Karoui [14 ] shows that, if the noise is spherical, the kernel matrix of the noisy points is, to first order, the kernel matrix of the clean points with a different kernel. Ding and Wu [12 ] study how close the eigenvalues of the kernel matrix of the noisy points are to those of the kernel matrix of the clean points, depending on the bandwidth and on the variance of the clean points relative to that of the noise. Landa, Coifman and Kluger [24 ] show that a Gaussian kernel matrix with the doubly stochastic normalization changes little under noise whose covariance may differ from point to point. Cheng and Landa [10 ] prove pointwise consistency under outlier noise for a graph Laplacian with an approximate version of this normalization.
Noise that is not small compared with the bandwidth can also be handled by denoising before building the graph [1 , 18 , 22 , 25 , 27 ] . Each denoised point may depend on all the observations, so the errors that remain after denoising need not be independent. Our results allow such dependence, since they hold for arbitrary noise bounded by a deterministic δ \delta δ . Section 1.4 explains how our results apply after denoising.
Section 1.3 compares our results with those of García Trillos, Sanz-Alonso and Yang [18 ] , of Burago, Ivanov and Kurylev [5 ] , and of Divol [13 ] .
Setting and main results
We use the setting of Section 2. The clean points x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n are i.i.d. samples from a density ρ \rho ρ on an m m m -dimensional manifold M ⊂ R d \mathcal{M} \subset\mathbb{R}^d M ⊂ R d . For a non-increasing Lipschitz kernel profile η \eta η supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] , the scaled unnormalized graph Laplacian L L L acts on u ∈ R n u \in\mathbb{R}^n u ∈ R n by
( L u ) i = 1 n ε m + 2 ∑ j = 1 n ω i j ( u i − u j ) , ω i j = η ( ∥ x i − x j ∥ ε ) . (Lu)_i = \frac{1}{n\varepsilon^{m+2}}\sum_{j=1}^{n}\omega_{ij}(u_i-u_j), \qquad\omega_{ij}=\eta\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}\right). ( Lu ) i = n ε m + 2 1 j = 1 ∑ n ω ij ( u i − u j ) , ω ij = η ( ε ∥ x i − x j ∥ ) .
The noisy graph Laplacian L ~ \widetilde{L} L is built in the same way from y i = x i + δ i y_i=x_i+\delta_i y i = x i + δ i , where ∥ δ i ∥ ≤ δ \lVert\delta_i\rVert\leq\delta ∥ δ i ∥ ≤ δ for a deterministic bound δ \delta δ , possibly depending on n n n . The displacements are otherwise arbitrary and may depend on the clean points.
Write λ l \lambda_l λ l , λ ~ l \widetilde{\lambda}_l λ l and λ l M \lambda_l^{\mathcal{M}} λ l M for the l l l -th smallest eigenvalues of L L L , of L ~ \widetilde{L} L and of the weighted Laplace–Beltrami operator Δ ρ \Delta_\rho Δ ρ defined in Section 2. Fix l l l and p ≥ 1 p\geq1 p ≥ 1 , and choose ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon=C_\varepsilon(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) with C ε C_\varepsilon C ε large enough, as in [7 ] . For δ / ε \delta/\varepsilon δ / ε small enough and n n n large enough, the following bounds hold with probability at least 1 − O ( n − p ) 1 - O(n^{-p}) 1 − O ( n − p ) . Theorem 1 shows that λ ~ l \widetilde{\lambda}_l λ l is close to λ l \lambda_l λ l :
∣ λ l − λ ~ l ∣ = O ( δ / ε ) . |\lambda_l - \widetilde{\lambda}_l| = O(\delta/\varepsilon). ∣ λ l − λ l ∣ = O ( δ / ε ) .
Let P l P_l P l and P ~ l \widetilde{P}_l P l be the orthogonal projections onto the subspaces spanned by the first l l l eigenvectors of L L L and of L ~ \widetilde{L} L . If λ l M < λ l + 1 M \lambda_l^M < \lambda_{l+1}^M λ l M < λ l + 1 M , whether or not λ l M \lambda_l^M λ l M is simple, Theorem 3 shows that P ~ l \widetilde{P}_l P l is close to P l P_l P l :
∥ P l − P ~ l ∥ H S , n = O ( δ / ε ) . \|P_l - \widetilde{P}_l\|_{\mathrm{HS},n} = O(\sqrt{\delta/\varepsilon}). ∥ P l − P l ∥ HS , n = O ( δ / ε ) .
where ∥ ⋅ ∥ H S , n \|\cdot\|_{\mathrm{HS},n} ∥ ⋅ ∥ HS , n is the Hilbert–Schmidt norm, equal to the Frobenius norm (§ 3 ).
For each bound, the event depends only on the clean points, so the bound holds for all displacements with ∥ δ i ∥ ≤ δ \|\delta_i\| \le\delta ∥ δ i ∥ ≤ δ at once (Lemma 42 ).
Corollary 2 and Theorem 4 combine these bounds with the clean-sample results of [7 ] to give the eigenvalue and eigenvector convergence rates stated above.
Idea of the proofs
We bracket both graph Laplacians between two graph Laplacians on the clean points. García Trillos, Sanz-Alonso and Yang [18 Lemma 18] use this idea for the indicator kernel. Moving each point by at most δ \delta δ changes the distance between two points by at most 2 δ 2\delta 2 δ . Since η \eta η is non-increasing, shifting its argument ∥ x i − x j ∥ / ε \|x_i-x_j\|/\varepsilon ∥ x i − x j ∥/ ε up or down by 2 δ / ε 2\delta/\varepsilon 2 δ / ε , with negative arguments replaced by zero, gives lower and upper bounds for both the clean and the noisy weights. The graph Laplacians on the clean points with these weights, scaled as L L L , are the comparison operators A A A and B B B , and
A ⪯ L ⪯ B , A ⪯ L ~ ⪯ B , A \preceq L \preceq B,\qquad A \preceq\widetilde{L} \preceq B, A ⪯ L ⪯ B , A ⪯ L ⪯ B ,
for every point cloud and all displacements with ∥ δ i ∥ ≤ δ \|\delta_i\| \le\delta ∥ δ i ∥ ≤ δ , where S ⪯ T S \preceq T S ⪯ T means ⟨ S u , u ⟩ n ≤ ⟨ T u , u ⟩ n \langle Su,u\rangle_n \le\langle Tu,u\rangle_n ⟨ S u , u ⟩ n ≤ ⟨ T u , u ⟩ n for every u ∈ R n u \in\mathbb{R}^n u ∈ R n , and ⟨ u , v ⟩ n = n − 1 ∑ i u i v i \langle u,v\rangle_n = n^{-1}\sum_i u_i v_i ⟨ u , v ⟩ n = n − 1 ∑ i u i v i is the empirical inner product, with norm ∥ u ∥ n \|u\|_n ∥ u ∥ n . By the Courant–Fischer principle [23 Theorem 4.2.6, p. 236] , the l l l -th eigenvalues of L L L and L ~ \widetilde{L} L therefore both lie between those of A A A and B B B .
Write ω i j − ≤ ω i j + \omega_{ij}^{-} \le\omega_{ij}^{+} ω ij − ≤ ω ij + for the weights of A A A and B B B . These weights differ by O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) , and only for pairs of clean points at distance at most ε + 2 δ \varepsilon+ 2\delta ε + 2 δ ; with high probability each point has O ( n ε m ) O(n\varepsilon^m) O ( n ε m ) such neighbours. For every vector u u u ,
⟨ ( B − A ) u , u ⟩ n = 1 2 n 2 ε m + 2 ∑ i , j ( ω i j − − ω i j + ) ( u i − u j ) 2 . \langle(B-A)u,u\rangle_n = \frac{1}{2n^2\varepsilon^{m+2}}\sum_{i,j}(\omega_{ij}^{-}-\omega_{ij}^{+})(u_i-u_j)^2. ⟨( B − A ) u , u ⟩ n = 2 n 2 ε m + 2 1 i , j ∑ ( ω ij − − ω ij + ) ( u i − u j ) 2 .
so ⟨ ( B − A ) u , u ⟩ n = O ( δ / ε ) \langle(B-A)u,u\rangle_n = O(\delta/\varepsilon) ⟨( B − A ) u , u ⟩ n = O ( δ / ε ) if ∥ u ∥ n = 1 \|u\|_n = 1 ∥ u ∥ n = 1 and the entries u i u_i u i and u j u_j u j differ by O ( ε ) O(\varepsilon) O ( ε ) whenever ∥ x i − x j ∥ ≤ ε + 2 δ \|x_i-x_j\| \le\varepsilon+ 2\delta ∥ x i − x j ∥ ≤ ε + 2 δ . For a unit eigenvector u u u of a graph Laplacian on i.i.d. samples, with eigenvalue below a fixed cutoff, the Lipschitz estimate of Calder, García Trillos and Lewicka [8 Theorem 2.3] bounds ∣ u i − u j ∣ |u_i-u_j| ∣ u i − u j ∣ by a constant multiple of the geodesic distance between x i x_i x i and x j x_j x j plus the bandwidth. The operator A A A is built on the clean points, its shifted kernels satisfy the hypotheses of this estimate, and the constants of the estimate can be chosen uniformly in the shift (§ 4.3 and Appendix § 8 ). With high probability, therefore, the entries of each of the first l l l unit eigenvectors of A A A differ by O ( ε ) O(\varepsilon) O ( ε ) whenever ∥ x i − x j ∥ ≤ ε + 2 δ \|x_i-x_j\| \le\varepsilon+ 2\delta ∥ x i − x j ∥ ≤ ε + 2 δ . Since l l l is fixed, the same holds for every unit vector u u u in their span, so ⟨ ( B − A ) u , u ⟩ n = O ( δ / ε ) \langle(B-A)u,u\rangle_n = O(\delta/\varepsilon) ⟨( B − A ) u , u ⟩ n = O ( δ / ε ) for every such u u u . The Courant–Fischer principle [23 Theorem 4.2.6, p. 236] then bounds the difference between the l l l -th eigenvalues of A A A and B B B , and hence ∣ λ l − λ ~ l ∣ |\lambda_l-\widetilde{\lambda}_l| ∣ λ l − λ l ∣ , by O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) (§ 5 ).
For projectors, let P l + P_l^{+} P l + be the orthogonal projection onto the subspace spanned by the first l l l eigenvectors of A A A . The squared distances ∥ P l − P l + ∥ H S , n 2 \|P_l-P_l^{+}\|_{\mathrm{HS},n}^2 ∥ P l − P l + ∥ HS , n 2 and ∥ P ~ l − P l + ∥ H S , n 2 \|\widetilde{P}_l-P_l^{+}\|_{\mathrm{HS},n}^2 ∥ P l − P l + ∥ HS , n 2 are bounded by a constant multiple of the maximum of ⟨ ( B − A ) u , u ⟩ n \langle(B-A)u,u\rangle_n ⟨( B − A ) u , u ⟩ n over the unit vectors u u u in this subspace, divided by the spectral gap between the l l l -th and ( l + 1 ) (l+1) ( l + 1 ) -st eigenvalues of A A A . The maximum is O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) by the preceding paragraph. If λ l M < λ l + 1 M \lambda_l^M < \lambda_{l+1}^M λ l M < λ l + 1 M , this spectral gap is bounded away from zero with high probability. Taking square roots and applying the triangle inequality give Theorem 3 (§ 6 ); we do not claim that the square root is sharp. For l ≥ 2 l \ge2 l ≥ 2 , the proof of Theorem 4 applies the projector bound at the spectral gaps immediately below and above λ l M \lambda_l^M λ l M and combines the results with the proof of [7 Theorem 2.7(2)] .
Comparison with previous work
Coifman and Lafon [11 ] observe that, for a smooth kernel, a diffusion map remains a valid approximation as long as the bandwidth is larger than the size of the noise, and refer to spectral perturbation theory for the eigenvalues and eigenfunctions. We compare our results with later quantitative bounds, first for the eigenvalues and then for the eigenvectors, where our main contribution lies.
For the eigenvalues, Burago, Ivanov and Kurylev [5 Lemma 5.6] treat two metrics on the same set with the same measure, the situation of clean and noisy points. For the Laplacian with the indicator kernel, normalized by the measure of the ball around each point, they show that, under conditions on the measure, every eigenvalue changes by a relative amount O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) if the distances change by at most 2 δ 2\delta 2 δ . Divol [13 ] builds on their notion of closeness of measures, which holds when mass is moved over distances at most δ \delta δ . He replaces their indicator kernel by continuous kernels and their conditions on the measure by a coarse Poincaré inequality, which he shows empirical measures inherit with high probability. Under his geometric and smallness hypotheses, he bounds the relative change of the first eigenvalues between two close measures [13 Proposition 3.3] .
García Trillos, Sanz-Alonso and Yang [18 ] show that, for i.i.d. centred noise normal to the manifold and bounded by δ \delta δ , averaging nearby observations can reduce the error in the distances between nearby points from 2 δ 2\delta 2 δ to order δ 3 / 2 \delta^{3/2} δ 3/2 , which allows a smaller bandwidth for a given accuracy. For the indicator kernel, they bound each eigenvalue of the graph on the averaged points between multiples of those of two graphs on the clean points, at a smaller and a larger bandwidth [18 Lemma 18] . Combining this bracket with a rate for clean points from [16 ] , they bound the eigenvalues of the graph on the averaged points against those of a weighted Laplace–Beltrami operator [18 Theorem 3] . Corollary 19 carries this bracket over from the indicator kernel to non-increasing Lipschitz kernels, and together with the estimate of ⟨ ( B − A ) u , u ⟩ n \langle(B-A)u,u\rangle_n ⟨( B − A ) u , u ⟩ n above gives Theorem 1 .
For the eigenvectors, Burago, Ivanov and Kurylev [5 Section 6] observe that a small perturbation generally splits a multiple eigenvalue. For eigenvalues much smaller than ε − 2 \varepsilon^{-2} ε − 2 , they describe each eigenvector, after transport to the perturbed space, as close to a linear combination of eigenvectors with nearby eigenvalues. Divol [13 Proposition 6.8] gives an eigenfunction bound under his geometric and smallness hypotheses, for an eigenvalue separated from all the others by a relative gap, and hence simple. The bound controls the distance from an eigenfunction of one measure to the corresponding eigenspace of a close measure, and involves the difference of the two Laplacians applied to the eigenfunction. García Trillos, Sanz-Alonso and Yang remark that statements about the convergence of eigenvectors, “or better yet, spectral projections”, are possible [18 Remark 19] .
Our main contribution is a bound on the change of the subspace spanned by the first l l l eigenvectors under arbitrary bounded noise, with multiple eigenvalues allowed. If δ / ε \delta/\varepsilon δ / ε is small enough and λ l M < λ l + 1 M \lambda_l^M < \lambda_{l+1}^M λ l M < λ l + 1 M , whether or not λ l M \lambda_l^M λ l M is simple, Theorem 3 shows that, with high probability, P ~ l \widetilde{P}_l P l is within O ( δ / ε ) O(\sqrt{\delta/\varepsilon}) O ( δ / ε ) of P l P_l P l . This bound depends on the noise and the bandwidth only through their ratio δ / ε \delta/\varepsilon δ / ε , as in the observation of Coifman and Lafon [11 ] . For clean points against the continuum, eigenvector bounds that allow a multiple eigenvalue are known for discretizations of a manifold [4, Theorem 4] and for samples [7 ] . Combined with the results of [7 ] , our bound gives a rate for the eigenvectors of the noisy graph Laplacian against the eigenfunctions of Δ ρ \Delta_\rho Δ ρ (Theorem 4 ).
Application to denoising
Our results need only one property of the noise: every noisy point lies within a deterministic distance δ \delta δ of its clean point, with δ / ε \delta/\varepsilon δ / ε small enough. They therefore apply to the output of any denoising method that guarantees this with high probability, even when the noise in the observations is larger than the bandwidth. Our bounds are O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) for the eigenvalues and O ( δ / ε ) O(\sqrt{\delta/\varepsilon}) O ( δ / ε ) for the projectors. So if δ = O ( ε 2 ) \delta= O(\varepsilon^2) δ = O ( ε 2 ) , the eigenvalues of the graph Laplacian on the denoised points converge at the rate O ( ε ) O(\varepsilon) O ( ε ) that Calder and García Trillos [7 ] prove for the clean graph, and if δ = O ( ε 3 ) \delta= O(\varepsilon^3) δ = O ( ε 3 ) , so do the eigenvectors.
Appendix B works this out for the projection algorithm of Aizenbud and Sober [1 ] , for observations uniform in a tube of fixed radius around M \mathcal{M} M (Theorem 43 ). Puchkin and Spokoiny [25 ] bound, with high probability, the largest distance from a denoised point to its clean point, for bounded noise whose size tends to zero. A related approach is the local averaging of García Trillos, Sanz-Alonso and Yang [18 ] , which reduces the error in the distances between nearby points (§ 1.3 ).
Organization of the paper
Section 2 fixes the setting and Section 3 states the results. Section 4 constructs the bracket, Section 5 proves the eigenvalue bounds and Section 6 the projector and eigenvector bounds. Section 7 reports numerical experiments on the circle S 1 S^1 S 1 and the unit sphere S 2 S^2 S 2 . Appendix A verifies that the constants of the Lipschitz estimate do not depend on the shift, Appendix B contains the denoising application, and Appendix C states and verifies the hypotheses of the results of [7 ] that we use.
Setup
Let M ⊂ R d \mathcal{M} \subset\mathbb{R}^d M ⊂ R d be a compact, connected, orientable, smooth m m m -dimensional manifold without boundary, equipped with the Riemannian structure induced by the ambient space, and define d M d_\mathcal{M} d M as its geodesic distance. Let Vol M \operatorname{Vol}_{\mathcal{M}} Vol M be its volume measure and let μ = ρ d Vol M \mu= \rho\,d\operatorname{Vol}_{\mathcal{M}} μ = ρ d Vol M be a probability measure, where ρ ∈ C 2 , ϑ ( M ) \rho\in C^{2,\vartheta}(\mathcal{M}) ρ ∈ C 2 , ϑ ( M ) for some ϑ ∈ ( 0 , 1 ] \vartheta\in(0,1] ϑ ∈ ( 0 , 1 ] , that is, ρ \rho ρ is twice continuously differentiable and its second derivatives are ϑ \vartheta ϑ -Hölder continuous, and
0 < ρ min ≤ ρ ≤ ρ max < ∞ . 0 < \rho_{\min} \le\rho\le\rho_{\max} < \infty. 0 < ρ m i n ≤ ρ ≤ ρ m a x < ∞.
Let { x i } i = 1 n \{x_i\}_{i=1}^n { x i } i = 1 n be sampled i.i.d. from the probability measure μ \mu μ . Then the vector ( x 1 , … , x n ) (x_1,\ldots,x_n) ( x 1 , … , x n ) has distribution P n : = μ ⊗ n \mathbb{P}_n := \mu^{\otimes n} P n := μ ⊗ n , the product measure on the n n n -fold product M n : = M × ⋯ × M \mathcal{M}^n := \mathcal{M} \times\cdots\times\mathcal{M} M n := M × ⋯ × M . For an event E ⊆ M n \mathcal{E} \subseteq\mathcal{M}^n E ⊆ M n we write E c : = M n ∖ E \mathcal{E}^c := \mathcal{M}^n \setminus\mathcal{E} E c := M n ∖ E for its complement. Let { y i } i = 1 n ⊂ R d \{y_i\}_{i=1}^n \subset\mathbb{R}^d { y i } i = 1 n ⊂ R d be the noisy points, with y i = x i + δ i y_i = x_i + \delta_i y i = x i + δ i , where the displacements δ i ∈ R d \delta_i \in\mathbb{R}^d δ i ∈ R d satisfy ∥ δ i ∥ ≤ δ \lVert\delta_i\rVert\le\delta ∥ δ i ∥ ≤ δ for every i i i , for some deterministic δ ≥ 0 \delta\ge0 δ ≥ 0 . The displacements are otherwise arbitrary; in particular, they may depend on the clean points.
The clean and noisy graph Laplacians defined below, and the auxiliary graph Laplacians of Section 4, are all built from a fixed kernel profile η : [ 0 , ∞ ) → [ 0 , ∞ ) \eta: [0,\infty) \to[0,\infty) η : [ 0 , ∞ ) → [ 0 , ∞ ) that satisfies
(K1) supp ( η ) ⊆ [ 0 , 1 ] and η ( 1 ) = 0 ; (K2) η is non-increasing ; (K3) η is C η -Lipschitz on [ 0 , 1 ] ; (K4) ∫ B ( 0 , 1 ) ⊂ R m η ( ∣ w ∣ ) d w = 1 ; (K5) η ( 1 / 2 ) > 0. (1) \begin{aligned}
\text{(K1)}\quad&\operatorname{supp}(\eta) \subseteq[0,1]\ \text{and}\ \eta(1)=0;\\
\text{(K2)}\quad&\eta\ \text{is non-increasing};\\
\text{(K3)}\quad&\eta\ \text{is}\ C_{\eta}\text{-Lipschitz on }[0,1];\\
\text{(K4)}\quad&\int_{B(0,1)\subset\mathbb{R}^{m}} \eta(\lvert w\rvert)\,\mathrm{d}w=1;\\
\text{(K5)}\quad&\eta(1/2)>0. \tag*{(1)}
\end{aligned} (K1) (K2) (K3) (K4) (K5) supp ( η ) ⊆ [ 0 , 1 ] and η ( 1 ) = 0 ; η is non-increasing ; η is C η -Lipschitz on [ 0 , 1 ] ; ∫ B ( 0 , 1 ) ⊂ R m η (∣ w ∣) d w = 1 ; η ( 1/2 ) > 0. (1)
In (K3), C η C_{\eta} C η is a positive constant, and η \eta η is C η C_{\eta} C η -Lipschitz on [ 0 , 1 ] [0,1] [ 0 , 1 ] if ∣ η ( s ) − η ( t ) ∣ ≤ C η ∣ s − t ∣ \lvert\eta(s)-\eta(t)\rvert\leq C_{\eta}\lvert s-t\rvert ∣ η ( s ) − η ( t )∣ ≤ C η ∣ s − t ∣ for all s , t ∈ [ 0 , 1 ] s,t\in[0,1] s , t ∈ [ 0 , 1 ] . The kernel is the radial function w ↦ η ( ∣ w ∣ ) w\mapsto\eta(\lvert w\rvert) w ↦ η (∣ w ∣) on R m \mathbb{R}^{m} R m , which appears in (K4). With the hypotheses on M \mathcal{M} M and ρ \rho ρ above, (K1)–(K3) and (K5) imply the hypotheses that [7 Section 2] places on the manifold, the density and the kernel, so its results apply to the graph Laplacian L L L defined below. The Lipschitz estimate of [8 ] is applied in Section 4 to kernel profiles built from η \eta η , and its hypotheses are stated and checked there. We define the second moment of the kernel in the direction of e 1 e_1 e 1 , the first standard basis vector of R m \mathbb{R}^{m} R m , by
σ η : = ∫ B ( 0 , 1 ) ⟨ w , e 1 ⟩ 2 η ( ∣ w ∣ ) d w , \sigma_{\eta}:=\int_{B(0,1)}\langle w,e_1\rangle^2\eta(\lvert w\rvert)\,\mathrm{d}w, σ η := ∫ B ( 0 , 1 ) ⟨ w , e 1 ⟩ 2 η (∣ w ∣) d w ,
and σ κ \sigma_{\kappa} σ κ in the same way for any other kernel profile κ \kappa κ . By (K2) and (K5), η ≥ η ( 1 / 2 ) > 0 \eta\geq\eta(1/2)>0 η ≥ η ( 1/2 ) > 0 on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] . Hence η ( 0 ) > 0 \eta(0)>0 η ( 0 ) > 0 , and σ η > 0 \sigma_{\eta}>0 σ η > 0 because the integrand of σ η \sigma_{\eta} σ η is positive on B ( 0 , 1 / 2 ) B(0,1/2) B ( 0 , 1/2 ) off the hyperplane { ⟨ w , e 1 ⟩ = 0 } \{\langle w,e_1\rangle=0\} {⟨ w , e 1 ⟩ = 0 } .
Define the weighted Laplace–Beltrami operator on a smooth function f f f by
Δ ρ f : = − 1 2 ρ div ( ρ 2 ∇ f ) , \Delta_{\rho}f:=-\frac{1}{2\rho}\operatorname{div}(\rho^2\nabla f), Δ ρ f := − 2 ρ 1 div ( ρ 2 ∇ f ) ,
where the divergence and gradient are intrinsic to M \mathcal{M} M (see, e.g., [7 equation (2.12)] ). The operator Δ ρ \Delta_{\rho} Δ ρ is positive semidefinite on L 2 ( M , μ ) L^2(\mathcal{M},\mu) L 2 ( M , μ ) , and its spectrum consists of eigenvalues of finite multiplicity that tend to infinity [7 Section 2.2.1] . We denote by λ l M \lambda_l^{\mathcal{M}} λ l M the l l l -th eigenvalue of Δ ρ \Delta_{\rho} Δ ρ . Eigenvalues are indexed here and below by l = 1 , 2 , … l=1,2,\ldots l = 1 , 2 , … in non-decreasing order and repeated according to multiplicity, so that λ 1 M ≤ λ 2 M ≤ ⋯ \lambda_1^{\mathcal{M}}\leq\lambda_2^{\mathcal{M}}\leq\cdots λ 1 M ≤ λ 2 M ≤ ⋯ . We write γ l M : = λ l + 1 M − λ l M \gamma_l^{\mathcal{M}}:=\lambda_{l+1}^{\mathcal{M}}-\lambda_l^{\mathcal{M}} γ l M := λ l + 1 M − λ l M for the l l l -th spectral gap of Δ ρ \Delta_{\rho} Δ ρ .
For a bandwidth ε > 0 \varepsilon>0 ε > 0 and every pair 1 ≤ i , j ≤ n 1\leq i,j\leq n 1 ≤ i , j ≤ n , define the clean weights ω i j \omega_{ij} ω ij and the noisy weights ω ~ i j \widetilde{\omega}_{ij} ω ij by
ω i j = η ( ∥ x i − x j ∥ ε ) , ω ~ i j = η ( ∥ y i − y j ∥ ε ) . \omega_{ij}=\eta\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}\right),\qquad
\widetilde{\omega}_{ij}=\eta\left(\frac{\lVert y_i-y_j\rVert}{\varepsilon}\right). ω ij = η ( ε ∥ x i − x j ∥ ) , ω ij = η ( ε ∥ y i − y j ∥ ) .
The corresponding (unnormalized) graph Laplacians are defined by
L = 1 n ε m + 2 ( D − W ) , L ~ = 1 n ε m + 2 ( D ~ − W ~ ) , L=\frac{1}{n\varepsilon^{m+2}}(D-W),\qquad
\widetilde{L}=\frac{1}{n\varepsilon^{m+2}}(\widetilde{D}-\widetilde{W}), L = n ε m + 2 1 ( D − W ) , L = n ε m + 2 1 ( D − W ) ,
where the weight matrix W = ( ( ω i j ) i , j = 1 n ) W=((\omega_{ij})_{i,j=1}^{n}) W = (( ω ij ) i , j = 1 n ) and the diagonal degree matrix D = diag ( ( ∑ j = 1 n ω i j ) i = 1 n ) D=\operatorname{diag}((\sum_{j=1}^{n}\omega_{ij})_{i=1}^{n}) D = diag (( ∑ j = 1 n ω ij ) i = 1 n ) are n × n n\times n n × n matrices, and W ~ , D ~ \widetilde{W},\widetilde{D} W , D are defined in the same way from the ω ~ i j \widetilde{\omega}_{ij} ω ij . Define X n : = { x 1 , … , x n } X_n:=\{x_1,\ldots,x_n\} X n := { x 1 , … , x n } and identify each u ∈ R n u\in\mathbb{R}^{n} u ∈ R n with the function x i ↦ u i x_i\mapsto u_i x i ↦ u i on X n X_n X n . For f : M → R f:\mathcal{M}\to\mathbb{R} f : M → R , define f ∣ X n f|_{X_n} f ∣ X n as the vector with entries f ( x i ) f(x_i) f ( x i ) . We denote the empirical L 2 L^2 L 2 -norm, that is, the norm of L 2 ( X n ) L^2(X_n) L 2 ( X n ) , by ∥ u ∥ n 2 : = 1 n ∑ i = 1 n u i 2 \lVert u\rVert_n^2:=\frac{1}{n}\sum_{i=1}^{n}u_i^2 ∥ u ∥ n 2 := n 1 ∑ i = 1 n u i 2 and the associated inner product by ⟨ u , v ⟩ n : = 1 n ∑ i = 1 n u i v i \langle u,v\rangle_n:=\frac{1}{n}\sum_{i=1}^{n}u_iv_i ⟨ u , v ⟩ n := n 1 ∑ i = 1 n u i v i . The Dirichlet energies associated with L L L and L ~ \widetilde{L} L are defined by
b ε ( u ) = 1 n 2 ε m + 2 ∑ i , j ω i j ( u i − u j ) 2 , b ~ ε ( u ) = 1 n 2 ε m + 2 ∑ i , j ω ~ i j ( u i − u j ) 2 . b_{\varepsilon}(u)=\frac{1}{n^2\varepsilon^{m+2}}\sum_{i,j}\omega_{ij}(u_i-u_j)^2,\qquad
\widetilde{b}_{\varepsilon}(u)=\frac{1}{n^2\varepsilon^{m+2}}\sum_{i,j}\widetilde{\omega}_{ij}(u_i-u_j)^2. b ε ( u ) = n 2 ε m + 2 1 i , j ∑ ω ij ( u i − u j ) 2 , b ε ( u ) = n 2 ε m + 2 1 i , j ∑ ω ij ( u i − u j ) 2 .
We denote by λ l \lambda_l λ l and λ ~ l \widetilde{\lambda}_l λ l the l l l -th eigenvalue of L L L and of L ~ \widetilde{L} L , indexed by the convention above. Since L L L and L ~ \widetilde{L} L are symmetric n × n n \times n n × n matrices, l l l ranges over 1 , … , n 1,\ldots,n 1 , … , n .
Calder and García Trillos [7 Theorem 2.4(2)] show that, for fixed l l l and p ≥ 1 p \ge1 p ≥ 1 , if ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon= C_{\varepsilon}(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) with C ε C_{\varepsilon} C ε large enough, then for n n n large enough, with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) ,
∣ λ l − σ η λ l M ∣ ≤ C l ′ ε , \left|\lambda_l-\sigma_{\eta}\lambda_l^{\mathcal{M}}\right| \le C_l'\varepsilon, λ l − σ η λ l M ≤ C l ′ ε ,
where C l ′ C_l' C l ′ depends only on l l l , M \mathcal{M} M , ρ \rho ρ and η \eta η . The factor σ η \sigma_{\eta} σ η appears because L L L acts on smooth functions approximately as σ η Δ ρ \sigma_{\eta}\Delta_{\rho} σ η Δ ρ [7 Theorem 3.3] . Our results use the same bandwidth. Theorem 1 bounds the difference between the noisy and the clean eigenvalues, and Corollary 2 combines Theorem 1 with this estimate.
We denote the relative noise scale by
a : = 2 δ ε . (2) a := \frac{2\delta}{\varepsilon}. \tag*{(2)} a := ε 2 δ . (2)
The bound δ \delta δ may depend on n n n . In the results of § 3 , ε → 0 \varepsilon\to0 ε → 0 as n → ∞ n \to\infty n → ∞ , and δ / ε \delta/\varepsilon δ / ε is assumed to be sufficiently small.
Main Results
Eigenvalue stability
Theorem 1. Assume the setup of § 2 and fix l , p ≥ 1 l,p \ge1 l , p ≥ 1 . Let ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon= C_{\varepsilon}(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) , where C ε C_{\varepsilon} C ε is large enough depending on p , l , M , ρ p,l,\mathcal{M},\rho p , l , M , ρ and η \eta η , and assume that δ / ε \delta/\varepsilon δ / ε is sufficiently small, depending only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η . Then, for n n n large enough, with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) ,
∣ λ l − λ ~ l ∣ ≤ K l δ ε = O ( δ ε ) . |\lambda_l-\widetilde{\lambda}_l| \le K_l\frac{\delta}{\varepsilon}=O\left(\frac{\delta}{\varepsilon}\right). ∣ λ l − λ l ∣ ≤ K l ε δ = O ( ε δ ) .
Here K l K_l K l is the constant (12) of Lemma 24 . It is deterministic and depends only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η .
Corollary 2. Assume the setup of § 2 and fix l , p ≥ 1 l,p \ge1 l , p ≥ 1 . Let ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon= C_{\varepsilon}(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) , where C ε C_{\varepsilon} C ε is large enough depending on p , l , M , ρ p,l,\mathcal{M},\rho p , l , M , ρ and η \eta η , and assume that δ / ε \delta/\varepsilon δ / ε is sufficiently small, depending only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η . Then, for n n n large enough, with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) ,
∣ λ ~ l − σ η λ l M ∣ ≤ K l δ ε + C l ′ ε = O ( δ ε ) + O ( ε ) . |\widetilde{\lambda}_l-\sigma_{\eta}\lambda_l^{\mathcal{M}}| \le K_l\frac{\delta}{\varepsilon}+C_l'\varepsilon=O\left(\frac{\delta}{\varepsilon}\right)+O(\varepsilon). ∣ λ l − σ η λ l M ∣ ≤ K l ε δ + C l ′ ε = O ( ε δ ) + O ( ε ) .
Here K l K_l K l is the constant of Theorem 1, and C l ′ C_l' C l ′ is the constant from [7 Theorem 2.4(2)] quoted in § 2 , depending only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η . In particular the right-hand side is O ( ε ) O(\varepsilon) O ( ε ) when δ = O ( ε 2 ) \delta=O(\varepsilon^2) δ = O ( ε 2 ) .
Eigenvector stability
Define P l P_l P l as the orthogonal projection of R n \mathbb{R}^n R n onto the subspace spanned by l l l orthonormal eigenvectors of L L L with eigenvalues λ 1 , … , λ l \lambda_1,\ldots,\lambda_l λ 1 , … , λ l , which we call the first l l l eigenvectors of L L L , and P ~ l \widetilde{P}_l P l in the same way for L ~ \widetilde{L} L . We call P l P_l P l and P ~ l \widetilde{P}_l P l spectral projectors. When λ l = λ l + 1 \lambda_l=\lambda_{l+1} λ l = λ l + 1 , P l P_l P l depends on the choice of eigenvectors, and likewise P ~ l \widetilde{P}_l P l when λ ~ l = λ ~ l + 1 \widetilde{\lambda}_l=\widetilde{\lambda}_{l+1} λ l = λ l + 1 ; the results below hold for every choice. We measure the distance between projectors in the Hilbert–Schmidt norm of operators T T T on ( R n , ⟨ ⋅ , ⋅ ⟩ n ) (\mathbb{R}^n,\langle\cdot,\cdot\rangle_n) ( R n , ⟨ ⋅ , ⋅ ⟩ n ) ,
∥ T ∥ H S , n 2 : = ∑ i = 1 n ∥ T e i ∥ n 2 , \|T\|_{\mathrm{HS},n}^{2}:=\sum_{i=1}^{n}\|Te_i\|_{n}^{2}, ∥ T ∥ HS , n 2 := i = 1 ∑ n ∥ T e i ∥ n 2 ,
where e 1 , … , e n e_1,\ldots,e_n e 1 , … , e n is any ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n -orthonormal basis of R n \mathbb{R}^n R n ; the sum does not depend on the basis. The vectors n \sqrt{n} n times the standard basis vectors form such a basis, so ∥ T ∥ H S , n \lVert T\rVert_{\mathrm{HS},n} ∥ T ∥ HS , n equals the Frobenius norm of the matrix of T T T .
Theorem 3. Assume the setup of § 2 and fix l , p ≥ 1 l,p\geq1 l , p ≥ 1 such that the spectral gap γ l M \gamma_l^{\mathcal{M}} γ l M is strictly positive. Let ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon=C_\varepsilon(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) , where C ε C_\varepsilon C ε is large enough depending on p , l , M , ρ p,l,\mathcal{M},\rho p , l , M , ρ and η \eta η , and assume that δ / ε \delta/\varepsilon δ / ε is sufficiently small, depending only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η . Then, for n n n large enough, with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) ,
∥ P l − P ~ l ∥ H S , n ≤ D l δ ε = O ( δ ε ) , D l : = 4 K l σ η γ l M . \lVert P_l-\widetilde{P}_l\rVert_{\mathrm{HS},n}\leq D_l\sqrt{\frac{\delta}{\varepsilon}}=O\left(\sqrt{\frac{\delta}{\varepsilon}}\right),\qquad D_l:=4\sqrt{\frac{K_l}{\sigma_\eta\gamma_l^{\mathcal{M}}}}. ∥ P l − P l ∥ HS , n ≤ D l ε δ = O ( ε δ ) , D l := 4 σ η γ l M K l . where K l K_l K l is the constant of Theorem 1 .
Theorem 4. Assume the setup of § 2 and fix l ≥ 2 l\geq2 l ≥ 2 , p ≥ 1 p\geq1 p ≥ 1 . Write
λ l − M = ⋯ = λ l M = ⋯ = λ l + M , l − ≤ l ≤ l + , \lambda_{l_-}^{\mathcal{M}}=\cdots=\lambda_l^{\mathcal{M}}=\cdots=\lambda_{l_+}^{\mathcal{M}},\qquad l_-\leq l\leq l_+, λ l − M = ⋯ = λ l M = ⋯ = λ l + M , l − ≤ l ≤ l + , for the multiplicity block of λ l M \lambda_l^{\mathcal{M}} λ l M , so that l − l_- l − and l + l_+ l + are the smallest and the largest index k k k with λ k M = λ l M \lambda_k^{\mathcal{M}}=\lambda_l^{\mathcal{M}} λ k M = λ l M . Hence the two spectral gaps γ l − − 1 M \gamma_{l_--1}^{\mathcal{M}} γ l − − 1 M and γ l + M \gamma_{l_+}^{\mathcal{M}} γ l + M bounding the block are strictly positive, and l − ≥ 2 l_-\geq2 l − ≥ 2 since λ l M \lambda_l^{\mathcal{M}} λ l M is simple. Let ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon=C_\varepsilon(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) , where C ε C_\varepsilon C ε is large enough depending on p , l , M , ρ p,l,\mathcal{M},\rho p , l , M , ρ and η \eta η , and assume that δ / ε \delta/\varepsilon δ / ε is sufficiently small, depending only on l , M , ρ l,\mathcal{M},\rho l , M , ρ and η \eta η . Then, for n n n large enough, with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) , for every eigenvector u ~ l \widetilde{u}_l u l of L ~ \widetilde{L} L with eigenvalue λ ~ l \widetilde{\lambda}_l λ l and ∥ u ~ l ∥ n = 1 \lVert\widetilde{u}_l\rVert_n=1 ∥ u l ∥ n = 1 there is an eigenfunction f l ∈ L 2 ( M , μ ) f_l\in L^2(\mathcal{M},\mu) f l ∈ L 2 ( M , μ ) with Δ ρ f l = λ l M f l \Delta_\rho f_l=\lambda_l^{\mathcal{M}}f_l Δ ρ f l = λ l M f l and ∥ f l ∥ L 2 ( μ ) = 1 \lVert f_l\rVert_{L^2(\mu)}=1 ∥ f l ∥ L 2 ( μ ) = 1 such that
∥ u ~ l − f l ∣ X n ∥ n = O ( δ ε + ε ) . \lVert\widetilde{u}_l-f_l|_{X_n}\rVert_n=O\left(\sqrt{\frac{\delta}{\varepsilon}}+\varepsilon\right). ∥ u l − f l ∣ X n ∥ n = O ( ε δ + ε ) . In particular the right-hand side is O ( ε ) O(\varepsilon) O ( ε ) when δ = O ( ε 3 ) \delta=O(\varepsilon^3) δ = O ( ε 3 ) .
Remark 5 (The index l = 1 l=1 l = 1 ). Theorem 4 is stated for l ≥ 2 l\geq2 l ≥ 2 because its proof uses the spectral gap γ l − − 1 M \gamma_{l_--1}^{\mathcal{M}} γ l − − 1 M , which is not defined when l − = 1 l_-=1 l − = 1 . The case l = 1 l=1 l = 1 is trivial. Since M \mathcal{M} M is connected, λ 1 M = 0 \lambda_1^{\mathcal{M}}=0 λ 1 M = 0 is simple and its unit eigenfunctions are the constants ± 1 \pm1 ± 1 . Under the hypotheses of Corollary 2 with l = 2 l=2 l = 2 , for n n n large and with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) , λ ~ 1 \widetilde{\lambda}_1 λ 1 is simple as well, so every unit noisy eigenvector u ~ 1 \widetilde{u}_1 u 1 at index 1 is constant and ∥ u ~ 1 − f 1 ∣ X n ∥ n = 0 \lVert\widetilde{u}_1-f_1|_{X_n}\rVert_n=0 ∥ u 1 − f 1 ∣ X n ∥ n = 0 for the matching constant eigenfunction f 1 f_1 f 1 . The proof is at the end of § 6.4 .
The Bracketing Construction
In this section we introduce the bracketing construction, which reduces the comparison of the clean and noisy graph Laplacians to a comparison between two auxiliary graph Laplacians on the clean samples. Since ∥ δ i ∥ ≤ δ \lVert\delta_i\rVert\leq\delta ∥ δ i ∥ ≤ δ for every i i i , noise changes each pairwise distance by at most 2 δ = a ε 2\delta=a\varepsilon 2 δ = a ε , and hence the argument of η \eta η in each weight by at most a a a . Because η \eta η is non-increasing, increasing or decreasing that argument by a a a gives lower and upper bounds for both the clean and the noisy weights. We use these bounds to define auxiliary weights satisfying
ω i j + ≤ ω i j ≤ ω ‾ i j , ω i j + ≤ ω ~ i j ≤ ω i j − (3) \omega_{ij}^{+}\leq\omega_{ij}\leq\overline{\omega}_{ij},\qquad\omega_{ij}^{+}\leq\widetilde{\omega}_{ij}\leq\omega_{ij}^{-} \tag*{(3)} ω ij + ≤ ω ij ≤ ω ij , ω ij + ≤ ω ij ≤ ω ij − (3)
for 1 ≤ i , j ≤ n 1\leq i,j\leq n 1 ≤ i , j ≤ n . Section 5 transfers this common bracket to Dirichlet energies and eigenvalues, with the appropriate bandwidths.
Keeping the auxiliary graphs on { x i } ⊂ M \{x_i\} \subset\mathcal{M} { x i } ⊂ M allows us to use the estimates of [8 ] , whose sampling hypotheses need not hold for the noisy cloud. We express the shifted weights through kernel profiles η a ± \eta_a^{\pm} η a ± supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] and bandwidths ε ± = ( 1 ∓ a ) ε \varepsilon^{\pm}=(1\mp a)\varepsilon ε ± = ( 1 ∓ a ) ε . The rest of this section verifies the weight ordering and shows that, after normalization, these kernel profiles satisfy the hypotheses of the Lipschitz estimate of [8 ] , stated as (CL1) and (CL2) in § 4.3 , with bounds that do not depend on a a a when a a a is small enough.
Definition 6. For a parameter a ∈ [ 0 , 1 ) a\in[0,1) a ∈ [ 0 , 1 ) , define the shift maps
ϕ + ( t ) : = a + ( 1 − a ) t , ϕ − ( t ) : = ( ( 1 + a ) t − a ) + . \phi^{+}(t) := a+(1-a)t,\qquad\phi^{-}(t):=((1+a)t-a)_{+}. ϕ + ( t ) := a + ( 1 − a ) t , ϕ − ( t ) := (( 1 + a ) t − a ) + . where ( ⋅ ) + : = max ( ⋅ , 0 ) (\cdot)_{+}:=\max(\cdot,0) ( ⋅ ) + := max ( ⋅ , 0 ) , and the shifted kernel profiles η a ± : = η ∘ ϕ ± \eta_a^{\pm}:=\eta\circ\phi^{\pm} η a ± := η ∘ ϕ ± , that is,
η a + ( t ) = η ( a + ( 1 − a ) t ) , η a − ( t ) = η ( ( ( 1 + a ) t − a ) + ) . \eta_a^{+}(t)=\eta\bigl(a+(1-a)t\bigr),\qquad\eta_a^{-}(t)=\eta\bigl(((1+a)t-a)_{+}\bigr). η a + ( t ) = η ( a + ( 1 − a ) t ) , η a − ( t ) = η ( (( 1 + a ) t − a ) + ) . The auxiliary Laplacians use the value a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε of (2) , which we assume is less than 1. For this value, set
ε + : = ( 1 − a ) ε , ε − : = ( 1 + a ) ε . \varepsilon^{+}:=(1-a)\varepsilon,\qquad\varepsilon^{-}:=(1+a)\varepsilon. ε + := ( 1 − a ) ε , ε − := ( 1 + a ) ε . and define the auxiliary weights
ω i j + = η a + ( ∥ x i − x j ∥ ε + ) = η ( a + ∥ x i − x j ∥ ε ) , \omega_{ij}^{+}=\eta_a^{+}\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon^{+}}\right)=\eta\left(a+\frac{\lVert x_i-x_j\rVert}{\varepsilon}\right), ω ij + = η a + ( ε + ∥ x i − x j ∥ ) = η ( a + ε ∥ x i − x j ∥ ) , ω i j − = η a − ( ∥ x i − x j ∥ ε − ) = η ( ( ∥ x i − x j ∥ ε − a ) + ) . \omega_{ij}^{-}=\eta_a^{-}\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon^{-}}\right)=\eta\left(\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}-a\right)_{+}\right). ω ij − = η a − ( ε − ∥ x i − x j ∥ ) = η ( ( ε ∥ x i − x j ∥ − a ) + ) . Let W ± = ( ω i j ± ) i , j = 1 n W^{\pm}=(\omega_{ij}^{\pm})_{i,j=1}^{n} W ± = ( ω ij ± ) i , j = 1 n and D ± = diag ( ∑ j = 1 n ω i j ± ) i = 1 n D^{\pm}=\operatorname{diag}\left(\sum_{j=1}^{n}\omega_{ij}^{\pm}\right)_{i=1}^{n} D ± = diag ( ∑ j = 1 n ω ij ± ) i = 1 n be the n × n n\times n n × n weight and diagonal degree matrices of these weights, and let L ± = 1 n ( ε ± ) m + 2 ( D ± − W ± ) L^{\pm}=\frac{1}{n(\varepsilon^{\pm})^{m+2}}(D^{\pm}-W^{\pm}) L ± = n ( ε ± ) m + 2 1 ( D ± − W ± ) be the corresponding graph Laplacians on { x i } \{x_i\} { x i } , with Dirichlet energies
b ± ( u ) = 1 n 2 ( ε ± ) m + 2 ∑ i , j ω i j ± ( u i − u j ) 2 . b^{\pm}(u)=\frac{1}{n^{2}(\varepsilon^{\pm})^{m+2}}\sum_{i,j}\omega_{ij}^{\pm}(u_i-u_j)^{2}. b ± ( u ) = n 2 ( ε ± ) m + 2 1 i , j ∑ ω ij ± ( u i − u j ) 2 . Let λ l ± \lambda_l^{\pm} λ l ± denote the l l l -th eigenvalue of L ± L^{\pm} L ± .
The weight bracket
In Lemmas 7 and 8 we verify the weight ordering (3) .
Lemma 7. For all i , j i,j i , j ,
∥ x i − x j ∥ − ∥ y i − y j ∥ ≤ ∥ δ i − δ j ∥ ≤ 2 δ = a ε . \lVert x_i-x_j\rVert-\lVert y_i-y_j\rVert\leq\lVert\delta_i-\delta_j\rVert\leq2\delta=a\varepsilon. ∥ x i − x j ∥ − ∥ y i − y j ∥ ≤ ∥ δ i − δ j ∥ ≤ 2 δ = a ε . Proof. Since y i − y j = ( x i − x j ) + ( δ i − δ j ) y_i-y_j=(x_i-x_j)+(\delta_i-\delta_j) y i − y j = ( x i − x j ) + ( δ i − δ j ) , the reverse triangle inequality gives the first bound, and ∥ δ i − δ j ∥ ≤ ∥ δ i ∥ + ∥ δ j ∥ ≤ 2 δ \lVert\delta_i-\delta_j\rVert\leq\lVert\delta_i\rVert+\lVert\delta_j\rVert\leq2\delta ∥ δ i − δ j ∥ ≤ ∥ δ i ∥ + ∥ δ j ∥ ≤ 2 δ the second; 2 δ = a ε 2\delta=a\varepsilon 2 δ = a ε by (2) .
Lemma 8. For all i , j i,j i , j , ω i j + ≤ ω i j ≤ ω i j − \omega_{ij}^{+}\leq\omega_{ij}\leq\omega_{ij}^{-} ω ij + ≤ ω ij ≤ ω ij − and ω i j + ≤ ω ~ i j ≤ ω i j − \omega_{ij}^{+}\leq\widetilde{\omega}_{ij}\leq\omega_{ij}^{-} ω ij + ≤ ω ij ≤ ω ij − .
Proof. Since η \eta η is non-increasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and a ≥ 0 a\geq0 a ≥ 0 ,
η ( a + ∥ x i − x j ∥ ε ) ≤ η ( ∥ x i − x j ∥ ε ) ≤ η ( ( ∥ x i − x j ∥ ε − a ) + ) . \eta\left(a+\frac{\lVert x_i-x_j\rVert}{\varepsilon}\right)\leq\eta\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}\right)\leq\eta\left(\left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}-a\right)_{+}\right). η ( a + ε ∥ x i − x j ∥ ) ≤ η ( ε ∥ x i − x j ∥ ) ≤ η ( ( ε ∥ x i − x j ∥ − a ) + ) . which is exactly ω i j + ≤ ω i j ≤ ω i j − \omega_{ij}^{+}\leq\omega_{ij}\leq\omega_{ij}^{-} ω ij + ≤ ω ij ≤ ω ij − . For the noisy weights, Lemma 7 gives
∥ x i − x j ∥ ε − a ≤ ∥ y i − y j ∥ ε ≤ ∥ x i − x j ∥ ε + a . \frac{\lVert x_i-x_j\rVert}{\varepsilon}-a\leq\frac{\lVert y_i-y_j\rVert}{\varepsilon}\leq\frac{\lVert x_i-x_j\rVert}{\varepsilon}+a. ε ∥ x i − x j ∥ − a ≤ ε ∥ y i − y j ∥ ≤ ε ∥ x i − x j ∥ + a .
The positive-part map t ↦ ( t ) + = max ( t , 0 ) t \mapsto(t)_{+} = \max(t,0) t ↦ ( t ) + = max ( t , 0 ) is non-decreasing, so applying it to the left inequality and using ( ∥ y i − y j ∥ ε ) + = ∥ y i − y j ∥ ε \left(\frac{\lVert y_i-y_j\rVert}{\varepsilon}\right)_{+}=\frac{\lVert y_i-y_j\rVert}{\varepsilon} ( ε ∥ y i − y j ∥ ) + = ε ∥ y i − y j ∥ gives
( ∥ x i − x j ∥ ε − a ) + ≤ ∥ y i − y j ∥ ε ≤ ∥ x i − x j ∥ ε + a . \left(\frac{\lVert x_i-x_j\rVert}{\varepsilon}-a\right)_{+}\leq\frac{\lVert y_i-y_j\rVert}{\varepsilon}\leq\frac{\lVert x_i-x_j\rVert}{\varepsilon}+a. ( ε ∥ x i − x j ∥ − a ) + ≤ ε ∥ y i − y j ∥ ≤ ε ∥ x i − x j ∥ + a .
Since η \eta η is non-increasing, applying it reverses the chain and yields ω i j + ≤ ω ~ i j ≤ ω i j − \omega_{ij}^{+}\leq\widetilde{\omega}_{ij}\leq\omega_{ij}^{-} ω ij + ≤ ω ij ≤ ω ij − .
Properties of the shifted kernel profiles
Lemmas 9–13 collect the properties of the shifted kernel profiles that are used below. They hold for every value of the parameter a a a in a stated range, not only for a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε . We need this because a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε changes with n n n and δ \delta δ , and Lemmas 14 and 17 need bounds that are uniform in a a a .
Lemma 9. For every a ∈ [ 0 , 1 ) a\in[0,1) a ∈ [ 0 , 1 ) , the shifted kernel profiles η a + \eta_a^{+} η a + and η a − \eta_a^{-} η a − satisfy:
η a ± : [ 0 , ∞ ) → [ 0 , ∞ ) \eta_a^{\pm}:[0,\infty)\to[0,\infty) η a ± : [ 0 , ∞ ) → [ 0 , ∞ ) ;
η a ± \eta_a^{\pm} η a ± is non-increasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) ;
supp ( η a ± ) ⊂ [ 0 , 1 ] \operatorname{supp}(\eta_a^{\pm})\subset[0,1] supp ( η a ± ) ⊂ [ 0 , 1 ] and η a ± ( 1 ) = 0 \eta_a^{\pm}(1)=0 η a ± ( 1 ) = 0 ;
the restriction of η a ± \eta_a^{\pm} η a ± to [ 0 , 1 ] [0,1] [ 0 , 1 ] is Lipschitz, with constant at most ( 1 − a ) C η (1-a)C_{\eta} ( 1 − a ) C η for η a + \eta_a^{+} η a + and ( 1 + a ) C η (1+a)C_{\eta} ( 1 + a ) C η for η a − \eta_a^{-} η a − , where C η C_{\eta} C η is the Lipschitz constant of η \eta η in (K3).
Proof. The map ϕ + \phi^{+} ϕ + of Definition 6 is affine with slope 1 − a ≥ 0 1-a\geq0 1 − a ≥ 0 , and ϕ − \phi^{-} ϕ − is zero on [ 0 , a / ( 1 + a ) ] [0,a/(1+a)] [ 0 , a / ( 1 + a )] and affine with slope 1 + a 1+a 1 + a on [ a / ( 1 + a ) , ∞ ) [a/(1+a),\infty) [ a / ( 1 + a ) , ∞ ) . Thus both maps are non-decreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , with ϕ ± ( 0 ) ∈ [ 0 , a ] \phi^{\pm}(0)\in[0,a] ϕ ± ( 0 ) ∈ [ 0 , a ] and ϕ ± ( 1 ) = 1 \phi^{\pm}(1)=1 ϕ ± ( 1 ) = 1 . Hence ϕ ± ( [ 0 , 1 ] ) ⊆ [ 0 , 1 ] \phi^{\pm}([0,1])\subseteq[0,1] ϕ ± ([ 0 , 1 ]) ⊆ [ 0 , 1 ] and ϕ ± ( t ) ≥ 1 \phi^{\pm}(t)\geq1 ϕ ± ( t ) ≥ 1 for t ≥ 1 t\geq1 t ≥ 1 .
(1) Inherited from η ≥ 0 \eta\geq0 η ≥ 0 .
(2) A non-decreasing map composed with a non-increasing map is non-increasing.
(3) For t ≥ 1 t\geq1 t ≥ 1 , ϕ ± ( t ) ≥ 1 \phi^{\pm}(t)\geq1 ϕ ± ( t ) ≥ 1 , so η a ± ( t ) = η ( ϕ ± ( t ) ) = 0 \eta_a^{\pm}(t)=\eta(\phi^{\pm}(t))=0 η a ± ( t ) = η ( ϕ ± ( t )) = 0 by (K1).
(4) On [ 0 , 1 ] [0,1] [ 0 , 1 ] , ϕ + \phi^{+} ϕ + has Lipschitz constant 1 − a 1-a 1 − a , and ϕ − \phi^{-} ϕ − has Lipschitz constant 1 + a 1+a 1 + a (the positive-part map is 1-Lipschitz, applied to an affine map of slope 1 + a 1+a 1 + a ). Since ϕ ± ( [ 0 , 1 ] ) ⊆ [ 0 , 1 ] \phi^{\pm}([0,1])\subseteq[0,1] ϕ ± ([ 0 , 1 ]) ⊆ [ 0 , 1 ] and η \eta η is C η C_{\eta} C η -Lipschitz on [ 0 , 1 ] [0,1] [ 0 , 1 ] by (K3),
∣ η a ± ( t ) − η a ± ( t ′ ) ∣ ≤ C η ∣ ϕ ± ( t ) − ϕ ± ( t ′ ) ∣ ≤ ( 1 ∓ a ) C η ∣ t − t ′ ∣ . \left|\eta_a^{\pm}(t)-\eta_a^{\pm}(t')\right|\leq C_{\eta}\left|\phi^{\pm}(t)-\phi^{\pm}(t')\right|\leq(1\mp a)C_{\eta}|t-t'|. η a ± ( t ) − η a ± ( t ′ ) ≤ C η ϕ ± ( t ) − ϕ ± ( t ′ ) ≤ ( 1 ∓ a ) C η ∣ t − t ′ ∣.
Lemma 10. For every a ∈ [ 0 , 1 ) a\in[0,1) a ∈ [ 0 , 1 ) and every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] ,
∣ ϕ ± ( t ) − t ∣ ≤ a . |\phi^{\pm}(t)-t|\leq a. ∣ ϕ ± ( t ) − t ∣ ≤ a . Proof. ϕ + ( t ) − t = a ( 1 − t ) ∈ [ 0 , a ] \phi^{+}(t)-t=a(1-t)\in[0,a] ϕ + ( t ) − t = a ( 1 − t ) ∈ [ 0 , a ] . For ϕ − \phi^{-} ϕ − : when t ∈ [ a / ( 1 + a ) , 1 ] t\in[a/(1+a),1] t ∈ [ a / ( 1 + a ) , 1 ] , ϕ − ( t ) − t = a ( t − 1 ) ∈ [ − a / ( 1 + a ) , 0 ] \phi^{-}(t)-t=a(t-1)\in[-a/(1+a),0] ϕ − ( t ) − t = a ( t − 1 ) ∈ [ − a / ( 1 + a ) , 0 ] ; when t ∈ [ 0 , a / ( 1 + a ) ) t\in[0,a/(1+a)) t ∈ [ 0 , a / ( 1 + a )) , ϕ − ( t ) − t = − t ∈ ( − a / ( 1 + a ) , 0 ] \phi^{-}(t)-t=-t\in(-a/(1+a),0] ϕ − ( t ) − t = − t ∈ ( − a / ( 1 + a ) , 0 ] .
Lemma 11. The kernel profile η \eta η is C η C_{\eta} C η -Lipschitz on all of [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , and for every a ∈ [ 0 , 1 ) a\in[0,1) a ∈ [ 0 , 1 ) and every t ∈ [ 0 , ∞ ) t\in[0,\infty) t ∈ [ 0 , ∞ ) ,
∣ η a ± ( t ) − η ( t ) ∣ ≤ C η a . |\eta_a^{\pm}(t)-\eta(t)|\leq C_{\eta}a. ∣ η a ± ( t ) − η ( t ) ∣ ≤ C η a . Proof. For s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] and t ≥ 1 t\geq1 t ≥ 1 , η ( t ) = η ( 1 ) = 0 \eta(t)=\eta(1)=0 η ( t ) = η ( 1 ) = 0 , hence
∣ η ( s ) − η ( t ) ∣ = ∣ η ( s ) − η ( 1 ) ∣ ≤ C η ∣ s − 1 ∣ ≤ C η ∣ s − t ∣ , |\eta(s)-\eta(t)|=|\eta(s)-\eta(1)|\leq C_{\eta}|s-1|\leq C_{\eta}|s-t|, ∣ η ( s ) − η ( t ) ∣ = ∣ η ( s ) − η ( 1 ) ∣ ≤ C η ∣ s − 1∣ ≤ C η ∣ s − t ∣ , where the last step uses ∣ s − t ∣ = t − s ≥ 1 − s |s-t|=t-s\geq1-s ∣ s − t ∣ = t − s ≥ 1 − s . Combined with the (K3) Lipschitz bound on [ 0 , 1 ] [0,1] [ 0 , 1 ] and the trivial bound on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , η \eta η is C η C_{\eta} C η -Lipschitz on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
For t ≥ 1 t\geq1 t ≥ 1 , η a ± ( t ) = η ( t ) = 0 \eta_a^{\pm}(t)=\eta(t)=0 η a ± ( t ) = η ( t ) = 0 (Lemma 9 (3)). For t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , by Lemma 10 and the global Lipschitz bound for η \eta η ,
∣ η a ± ( t ) − η ( t ) ∣ = ∣ η ( ϕ ± ( t ) ) − η ( t ) ∣ ≤ C η ∣ ϕ ± ( t ) − t ∣ ≤ C η a . |\eta_a^{\pm}(t)-\eta(t)|=|\eta(\phi^{\pm}(t))-\eta(t)|\leq C_{\eta}|\phi^{\pm}(t)-t|\leq C_{\eta}a. ∣ η a ± ( t ) − η ( t ) ∣ = ∣ η ( ϕ ± ( t )) − η ( t ) ∣ ≤ C η ∣ ϕ ± ( t ) − t ∣ ≤ C η a .
Lemma 12. For every a ∈ [ 0 , 1 ) a \in[0,1) a ∈ [ 0 , 1 ) ,
∣ ∫ R m η a ± ( ∣ w ∣ ) d w − 1 ∣ ≤ C η ∣ B 1 m ∣ a , ∣ σ η a ± − σ η ∣ ≤ C η ∣ B 1 m ∣ a , \left|\int_{\mathbb{R}^{m}}\eta_{a}^{\pm}(|w|)\,dw-1\right|\le C_{\eta}|B_{1}^{m}|a,\qquad|\sigma_{\eta_{a}^{\pm}}-\sigma_{\eta}|\le C_{\eta}|B_{1}^{m}|a, ∫ R m η a ± ( ∣ w ∣ ) d w − 1 ≤ C η ∣ B 1 m ∣ a , ∣ σ η a ± − σ η ∣ ≤ C η ∣ B 1 m ∣ a , where C η C_{\eta} C η is the Lipschitz constant of η \eta η in (K3) and ∣ B 1 m ∣ |B_{1}^{m}| ∣ B 1 m ∣ is the volume of the unit ball in R m \mathbb{R}^{m} R m . In particular, for a ≤ σ η / ( 2 C η ∣ B 1 m ∣ ) a\le\sigma_{\eta}/(2C_{\eta}|B_{1}^{m}|) a ≤ σ η / ( 2 C η ∣ B 1 m ∣ ) ,
σ η a ± ≥ 1 2 σ η > 0. \sigma_{\eta_{a}^{\pm}}\ge\tfrac{1}{2}\sigma_{\eta}>0. σ η a ± ≥ 2 1 σ η > 0. Proof. Both η \eta η and η a ± \eta_{a}^{\pm} η a ± are supported on { ∣ w ∣ ≤ 1 } \{|w|\le1\} { ∣ w ∣ ≤ 1 } , so the integrand ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ |\eta_{a}^{\pm}(|w|)-\eta(|w|)| ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ vanishes for ∣ w ∣ > 1 |w|>1 ∣ w ∣ > 1 . By Lemma 11 and the unit normalization (K4),
∣ ∫ η a ± ( ∣ w ∣ ) d w − 1 ∣ ≤ ∫ ∣ w ∣ ≤ 1 ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ d w ≤ C η ∣ B 1 m ∣ a . \left|\int\eta_{a}^{\pm}(|w|)\,dw-1\right|\le\int_{|w|\le1}|\eta_{a}^{\pm}(|w|)-\eta(|w|)|\,dw\le C_{\eta}|B_{1}^{m}|a. ∫ η a ± ( ∣ w ∣ ) d w − 1 ≤ ∫ ∣ w ∣ ≤ 1 ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ d w ≤ C η ∣ B 1 m ∣ a . For the second bound, using ⟨ w , e 1 ⟩ 2 ≤ 1 \langle w,e_{1}\rangle^{2}\le1 ⟨ w , e 1 ⟩ 2 ≤ 1 on ∣ w ∣ ≤ 1 |w|\le1 ∣ w ∣ ≤ 1 ,
∣ σ η a ± − σ η ∣ ≤ ∫ ∣ w ∣ ≤ 1 ⟨ w , e 1 ⟩ 2 ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ d w ≤ ∫ ∣ w ∣ ≤ 1 C η a d w = C η ∣ B 1 m ∣ a . |\sigma_{\eta_{a}^{\pm}}-\sigma_{\eta}|\le\int_{|w|\le1}\langle w,e_{1}\rangle^{2}|\eta_{a}^{\pm}(|w|)-\eta(|w|)|\,dw\le\int_{|w|\le1}C_{\eta}a\,dw=C_{\eta}|B_{1}^{m}|a. ∣ σ η a ± − σ η ∣ ≤ ∫ ∣ w ∣ ≤ 1 ⟨ w , e 1 ⟩ 2 ∣ η a ± ( ∣ w ∣ ) − η ( ∣ w ∣ ) ∣ d w ≤ ∫ ∣ w ∣ ≤ 1 C η a d w = C η ∣ B 1 m ∣ a .
Lemma 13. Let c 0 : = η ( 0 ) c_{0}:=\eta(0) c 0 := η ( 0 ) , t 0 : = min ( c 0 / ( 2 C η ) , 1 / 2 ) t_{0}:=\min(c_{0}/(2C_{\eta}),1/2) t 0 := min ( c 0 / ( 2 C η ) , 1/2 ) , and a 0 : = t 0 / 2 a_{0}:=t_{0}/2 a 0 := t 0 /2 . For every a ∈ [ 0 , a 0 ) a\in[0,a_{0}) a ∈ [ 0 , a 0 ) ,
η a ± ( t ) ≥ c 0 2 for every t ∈ [ 0 , t 0 / 2 ] . \eta_{a}^{\pm}(t)\ge\frac{c_{0}}{2}\qquad\text{for every }t\in[0,t_{0}/2]. η a ± ( t ) ≥ 2 c 0 for every t ∈ [ 0 , t 0 /2 ] . In particular, η a ± ( ∣ w ∣ ) ≥ c 0 / 2 \eta_{a}^{\pm}(|w|)\ge c_{0}/2 η a ± ( ∣ w ∣ ) ≥ c 0 /2 for every w w w in the open Euclidean ball S 0 : = B ( 0 , t 0 / 2 ) ⊂ R m S_{0}:=B(0,t_{0}/2)\subset\mathbb{R}^{m} S 0 := B ( 0 , t 0 /2 ) ⊂ R m , and neither the ball nor the bound depends on a a a .
Proof. By (K3) and the definition of t 0 t_{0} t 0 , η ( t ) ≥ η ( 0 ) − C η t = c 0 − C η t ≥ c 0 / 2 \eta(t)\ge\eta(0)-C_{\eta}t=c_{0}-C_{\eta}t\ge c_{0}/2 η ( t ) ≥ η ( 0 ) − C η t = c 0 − C η t ≥ c 0 /2 for every t ∈ [ 0 , t 0 ] t\in[0,t_{0}] t ∈ [ 0 , t 0 ] . For t ∈ [ 0 , t 0 / 2 ] t\in[0,t_{0}/2] t ∈ [ 0 , t 0 /2 ] and a ∈ [ 0 , a 0 ) a\in[0,a_{0}) a ∈ [ 0 , a 0 ) ,
ϕ + ( t ) = a + ( 1 − a ) t ≤ a + t ≤ t 0 2 + t 0 2 = t 0 , \phi^{+}(t)=a+(1-a)t\le a+t\le\frac{t_{0}}{2}+\frac{t_{0}}{2}=t_{0}, ϕ + ( t ) = a + ( 1 − a ) t ≤ a + t ≤ 2 t 0 + 2 t 0 = t 0 , so η a + ( t ) = η ( ϕ + ( t ) ) ≥ c 0 / 2 \eta_{a}^{+}(t)=\eta(\phi^{+}(t))\ge c_{0}/2 η a + ( t ) = η ( ϕ + ( t )) ≥ c 0 /2 .
For η a − \eta_{a}^{-} η a − , let t ∈ [ 0 , t 0 / 2 ] t\in[0,t_{0}/2] t ∈ [ 0 , t 0 /2 ] . Since ( 1 + a ) t − a ≤ t (1+a)t-a\le t ( 1 + a ) t − a ≤ t for t ≤ 1 t\le1 t ≤ 1 , we have 0 ≤ ϕ − ( t ) ≤ t ≤ t 0 0\le\phi^{-}(t)\le t\le t_{0} 0 ≤ ϕ − ( t ) ≤ t ≤ t 0 , and again η a − ( t ) = η ( ϕ − ( t ) ) ≥ c 0 / 2 \eta_{a}^{-}(t)=\eta(\phi^{-}(t))\ge c_{0}/2 η a − ( t ) = η ( ϕ − ( t )) ≥ c 0 /2 .
Lipschitz estimates for the auxiliary Laplacians
Section 5 needs a Lipschitz estimate for the eigenvectors of L + L^{+} L + . Calder, García Trillos and Lewicka [8 Theorem 2.3] prove such an estimate for the graph Laplacian on the clean points built, as L L L is in § 2 , from a single kernel profile κ : [ 0 , ∞ ) → [ 0 , ∞ ) \kappa:[0,\infty)\to[0,\infty) κ : [ 0 , ∞ ) → [ 0 , ∞ ) in place of η \eta η . Besides hypotheses on M \mathcal{M} M and ρ \rho ρ that those of § 2 imply, the estimate assumes the following.
(CL1) κ \kappa κ is non-increasing, supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] and Lipschitz on [ 0 , 1 ] [0,1] [ 0 , 1 ] , and ∫ R m κ ( ∣ w ∣ ) d w = 1 \int_{\mathbb{R}^{m}}\kappa(|w|)\,dw=1 ∫ R m κ ( ∣ w ∣ ) d w = 1 .
(CL2) There are an open set S ⊂ B ( 0 , 1 ) ⊂ R m S\subset B(0,1)\subset\mathbb{R}^{m} S ⊂ B ( 0 , 1 ) ⊂ R m and a constant c > 0 c>0 c > 0 such that κ ( ∣ w ∣ ) ≥ c \kappa(|w|)\ge c κ ( ∣ w ∣ ) ≥ c for every w ∈ S w\in S w ∈ S .
These are [8 (H1)(iii)] , with the support condition that [8 Section 2] imposes throughout, and [8 (9.1)] . The lower bound (CL2) is used only in [8 Section 9] , for pairs of points at geodesic distance at most 3 ε 3\varepsilon 3 ε . There it gives [8 Theorem 9.1] , which bounds ∣ f ( x ) − f ( y ) ∣ |f(x)-f(y)| ∣ f ( x ) − f ( y ) ∣ for such a pair by a factor θ < 1 \theta<1 θ < 1 times the supremum of ∣ f ( x ′ ) − f ( y ′ ) ∣ |f(x')-f(y')| ∣ f ( x ′ ) − f ( y ′ ) ∣ over nearby pairs in a larger ball, up to error terms.
The proof of the global estimate [8 Theorem 10.1] needs θ < 1 \theta< 1 θ < 1 to absorb that supremum. The constants of [8 Theorem 2.3] , including its smallness threshold for the bandwidth, depend on κ \kappa κ . Our kernel profiles depend on a a a , and a = 2 δ / ε a = 2\delta/\varepsilon a = 2 δ / ε changes with n n n and δ \delta δ , so we need these constants to be the same for all of them. For a family of profiles this requires, among other things, that θ \theta θ stay below a fixed number smaller than 1 1 1 , which is why Lemma 13 bounds the shifted profiles from below on one ball, by one constant. Lemma 14 normalizes the shifted kernel profiles and shows that the normalized profiles satisfy (CL1), and (CL2) with S = S 0 S = S_0 S = S 0 , with bounds that do not depend on a a a . Lemma 17 is the estimate of [8 Theorem 2.3] for the normalized auxiliary Laplacians L ‾ ± \overline{L}^{\pm} L ± of Definition 16. What it adds to that theorem is that the constants depend only on M \mathcal{M} M , ρ \rho ρ and η \eta η , which Appendix § 8 verifies.
Lemma 14. Let
a ⋆ : = min ( t 0 2 , 1 2 C η ∣ B 1 m ∣ , σ η 2 C η ∣ B 1 m ∣ , 1 2 ) . a_{\star} := \min\left(\frac{t_0}{2},\frac{1}{2C_{\eta}|B_1^m|},\frac{\sigma_{\eta}}{2C_{\eta}|B_1^m|},\frac{1}{2}\right). a ⋆ := min ( 2 t 0 , 2 C η ∣ B 1 m ∣ 1 , 2 C η ∣ B 1 m ∣ σ η , 2 1 ) . with t 0 t_0 t 0 as in Lemma 13 , C η C_{\eta} C η the Lipschitz constant of (K3) and ∣ B 1 m ∣ |B_1^m| ∣ B 1 m ∣ the volume of the unit ball in R m \mathbb{R}^m R m . It depends only on m m m and η \eta η . For every a ∈ [ 0 , a ⋆ ) a \in[0,a_{\star}) a ∈ [ 0 , a ⋆ ) and for both signs, the quantities
I a ± : = ∫ R m η a ± ( ∣ w ∣ ) d w , η ‾ a ± : = 1 I a ± η a ± I_a^{\pm} := \int_{\mathbb{R}^m}\eta_a^{\pm}(|w|)\,dw,\qquad\overline{\eta}_a^{\pm} := \frac{1}{I_a^{\pm}}\eta_a^{\pm} I a ± := ∫ R m η a ± ( ∣ w ∣ ) d w , η a ± := I a ± 1 η a ± are well-defined, with 1 2 ≤ I a ± ≤ 3 2 \frac{1}{2} \le I_a^{\pm} \le\frac{3}{2} 2 1 ≤ I a ± ≤ 2 3 , and the normalized kernel profiles η ‾ a ± \overline{\eta}_a^{\pm} η a ± satisfy the following bounds, none of which depends on a a a :
supp ( η ‾ a ± ) ⊆ [ 0 , 1 ] , ∫ R m η ‾ a ± ( ∣ w ∣ ) d w = 1 , \operatorname{supp}(\overline{\eta}_a^{\pm}) \subseteq[0,1],\qquad\int_{\mathbb{R}^m}\overline{\eta}_a^{\pm}(|w|)\,dw = 1, supp ( η a ± ) ⊆ [ 0 , 1 ] , ∫ R m η a ± ( ∣ w ∣ ) d w = 1 , Lip [ 0 , 1 ] ( η ‾ a ± ) ≤ 3 C η , ∥ η ‾ a ± ∥ ∞ ≤ 2 η ( 0 ) , \operatorname{Lip}_{[0,1]}(\overline{\eta}_a^{\pm}) \le3C_{\eta},\qquad\|\overline{\eta}_a^{\pm}\|_{\infty} \le2\eta(0), Lip [ 0 , 1 ] ( η a ± ) ≤ 3 C η , ∥ η a ± ∥ ∞ ≤ 2 η ( 0 ) , σ η 3 ≤ σ η ‾ a ± ≤ 3 σ η , η ‾ a ± ( ∣ w ∣ ) ≥ c 0 3 on S 0 . \frac{\sigma_{\eta}}{3} \le\sigma_{\overline{\eta}_a^{\pm}} \le3\sigma_{\eta},\qquad\overline{\eta}_a^{\pm}(|w|) \ge\frac{c_0}{3}\quad\text{on }S_0. 3 σ η ≤ σ η a ± ≤ 3 σ η , η a ± ( ∣ w ∣ ) ≥ 3 c 0 on S 0 . Here Lip [ 0 , 1 ] \operatorname{Lip}_{[0,1]} Lip [ 0 , 1 ] is the smallest Lipschitz constant of the restriction to [ 0 , 1 ] [0,1] [ 0 , 1 ] , ∥ ⋅ ∥ ∞ \|\cdot\|_{\infty} ∥ ⋅ ∥ ∞ is the supremum, and c 0 c_0 c 0 and the ball S 0 = B ( 0 , t 0 / 2 ) S_0 = B(0,t_0/2) S 0 = B ( 0 , t 0 /2 ) are as in Lemma 13 . In particular, η ‾ a ± \overline{\eta}_a^{\pm} η a ± satisfies (CL1), and (CL2) with S = S 0 S = S_0 S = S 0 and c = c 0 / 3 c = c_0/3 c = c 0 /3 .
Proof. Let a < a ⋆ a < a_{\star} a < a ⋆ . Since a < t 0 / 2 = a 0 a < t_0/2 = a_0 a < t 0 /2 = a 0 , a ≤ 1 / ( 2 C η ∣ B 1 m ∣ ) a \le1/(2C_{\eta}|B_1^m|) a ≤ 1/ ( 2 C η ∣ B 1 m ∣ ) and a ≤ σ η / ( 2 C η ∣ B 1 m ∣ ) a \le\sigma_{\eta}/(2C_{\eta}|B_1^m|) a ≤ σ η / ( 2 C η ∣ B 1 m ∣ ) , Lemmas 12 and 13 give
1 2 ≤ I a ± ≤ 3 2 , 1 2 σ η ≤ σ η a ± ≤ 3 2 σ η , η a ± ( ∣ w ∣ ) ≥ c 0 2 on S 0 . \frac{1}{2} \le I_a^{\pm} \le\frac{3}{2},\qquad\frac{1}{2}\sigma_{\eta} \le\sigma_{\eta_a^{\pm}} \le\frac{3}{2}\sigma_{\eta},\qquad\eta_a^{\pm}(|w|) \ge\frac{c_0}{2}\quad\text{on }S_0. 2 1 ≤ I a ± ≤ 2 3 , 2 1 σ η ≤ σ η a ± ≤ 2 3 σ η , η a ± ( ∣ w ∣ ) ≥ 2 c 0 on S 0 . In particular I a ± > 0 I_a^{\pm} > 0 I a ± > 0 . The normalized profile η ‾ a ± \overline{\eta}_a^{\pm} η a ± is the shifted profile η a ± \eta_a^{\pm} η a ± divided by the constant I a ± I_a^{\pm} I a ± . So it has unit mass, and it keeps the support, non-negativity and monotonicity of η a ± \eta_a^{\pm} η a ± from Lemma 9 . Its Lipschitz constant on [ 0 , 1 ] [0,1] [ 0 , 1 ] , its supremum and its lower bound on S 0 S_0 S 0 are those of η a ± \eta_a^{\pm} η a ± multiplied by ( I a ± ) − 1 ∈ [ 2 / 3 , 2 ] (I_a^{\pm})^{-1} \in[2/3,2] ( I a ± ) − 1 ∈ [ 2/3 , 2 ] . By Lemma 9 (4) and a < 1 / 2 a < 1/2 a < 1/2 , Lip [ 0 , 1 ] ( η a ± ) ≤ ( 1 + a ) C η ≤ 3 C η / 2 \operatorname{Lip}_{[0,1]}(\eta_a^{\pm}) \le(1+a)C_{\eta} \le3C_{\eta}/2 Lip [ 0 , 1 ] ( η a ± ) ≤ ( 1 + a ) C η ≤ 3 C η /2 , and ∥ η a ± ∥ ∞ = η ( ϕ ± ( 0 ) ) ≤ η ( 0 ) \|\eta_a^{\pm}\|_{\infty} = \eta(\phi^{\pm}(0)) \le\eta(0) ∥ η a ± ∥ ∞ = η ( ϕ ± ( 0 )) ≤ η ( 0 ) because η \eta η is non-increasing. Hence Lip [ 0 , 1 ] ( η ‾ a ± ) ≤ 3 C η \operatorname{Lip}_{[0,1]}(\overline{\eta}_a^{\pm}) \le3C_{\eta} Lip [ 0 , 1 ] ( η a ± ) ≤ 3 C η , ∥ η ‾ a ± ∥ ∞ ≤ 2 η ( 0 ) \|\overline{\eta}_a^{\pm}\|_{\infty} \le2\eta(0) ∥ η a ± ∥ ∞ ≤ 2 η ( 0 ) , and η ‾ a ± ( ∣ w ∣ ) ≥ 2 3 ⋅ c 0 2 = c 0 3 \overline{\eta}_a^{\pm}(|w|) \ge\frac{2}{3}\cdot\frac{c_0}{2} = \frac{c_0}{3} η a ± ( ∣ w ∣ ) ≥ 3 2 ⋅ 2 c 0 = 3 c 0 on S 0 S_0 S 0 . The second moment is divided by I a ± I_a^{\pm} I a ± as well:
σ η ‾ a ± = σ η a ± I a ± ∈ [ σ η / 2 3 / 2 , 3 σ η / 2 1 / 2 ] = [ σ η / 3 , 3 σ η ] . \sigma_{\overline{\eta}_a^{\pm}} = \frac{\sigma_{\eta_a^{\pm}}}{I_a^{\pm}} \in\left[\frac{\sigma_{\eta}/2}{3/2},\frac{3\sigma_{\eta}/2}{1/2}\right] = [\sigma_{\eta}/3,3\sigma_{\eta}]. σ η a ± = I a ± σ η a ± ∈ [ 3/2 σ η /2 , 1/2 3 σ η /2 ] = [ σ η /3 , 3 σ η ] . Finally, these properties give (CL1), and the lower bound on the open ball S 0 ⊂ B ( 0 , 1 ) S_0 \subset B(0,1) S 0 ⊂ B ( 0 , 1 ) gives (CL2).
Remark 15 (Standing assumption). From now on, a a a is again the relative noise scale 2 δ / ε 2\delta/\varepsilon 2 δ / ε of (2), and we assume a < a ⋆ a < a_{\star} a < a ⋆ . In the results of Section 3, this is part of the assumption that δ / ε \delta/\varepsilon δ / ε is sufficiently small.
Definition 16. Denote by L ˉ ± : = ( I a ± ) − 1 L ± \bar{L}^{\pm} := (I_a^{\pm})^{-1}L^{\pm} L ˉ ± := ( I a ± ) − 1 L ± the normalized auxiliary Laplacians . They are the graph Laplacians on { x i } \{x_i\} { x i } built from the normalized profiles η ˉ a ± \bar{\eta}_a^{\pm} η ˉ a ± at bandwidths ε ± \varepsilon^{\pm} ε ± , since η ˉ a ± = ( I a ± ) − 1 η a ± \bar{\eta}_a^{\pm} = (I_a^{\pm})^{-1}\eta_a^{\pm} η ˉ a ± = ( I a ± ) − 1 η a ± rescales every auxiliary weight by the same factor. The normalization is that of the mass of the kernel profile, as in (K4). Like L L L , the operators L ˉ ± \bar{L}^{\pm} L ˉ ± are unnormalized graph Laplacians in the sense of § 2 .
Lemma 17. There exist constants C C L u n i f , c ⋆ u n i f , C P , c P > 0 C_{\mathrm{CL}}^{\mathrm{unif}},c_{\star}^{\mathrm{unif}},C_{\mathrm{P}},c_{\mathrm{P}}>0 C CL unif , c ⋆ unif , C P , c P > 0 , depending only on M \mathcal{M} M , ρ \rho ρ and the kernel profile η \eta η , and in particular not on a a a , with the following property. Let s ∈ { + , − } s\in\{+,-\} s ∈ { + , − } be one of the two signs and Λ > 0 \Lambda>0 Λ > 0 a cutoff, and assume ε s ≤ c ⋆ u n i f / ( Λ + 1 ) \varepsilon^s\le c_{\star}^{\mathrm{unif}}/(\Lambda+1) ε s ≤ c ⋆ unif / ( Λ + 1 ) . There is an event E n , 1 s ( Λ ) E_{n,1}^{s}(\Lambda) E n , 1 s ( Λ ) , depending only on x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n , of probability at least
1 − C P ( ε s ) − 6 m exp ( − c P n ( ε s ) m + 4 ) − 2 n exp ( − c P n ( Λ + 1 ) − m ) 1-C_{\mathrm{P}}(\varepsilon^s)^{-6m}\exp(-c_{\mathrm{P}}n(\varepsilon^s)^{m+4})-2n\exp(-c_{\mathrm{P}}n(\Lambda+1)^{-m}) 1 − C P ( ε s ) − 6 m exp ( − c P n ( ε s ) m + 4 ) − 2 n exp ( − c P n ( Λ + 1 ) − m )
such that, given E n , 1 s ( Λ ) E_{n,1}^{s}(\Lambda) E n , 1 s ( Λ ) , every nonzero f f f on X n X_n X n with ∥ L ˉ s f ∥ ∞ < Λ ∥ f ∥ ∞ \|\bar{L}^s f\|_{\infty}<\Lambda\|f\|_{\infty} ∥ L ˉ s f ∥ ∞ < Λ∥ f ∥ ∞ satisfies
∣ f ( x i ) − f ( x j ) ∣ ≤ C C L u n i f ( Λ + 1 ) m + 1 ∥ f ∥ 1 , n ( d M ( x i , x j ) + ε s ) for all i , j , |f(x_i)-f(x_j)|\le C_{\mathrm{CL}}^{\mathrm{unif}}(\Lambda+1)^{m+1}\|f\|_{1,n}\left(d_{\mathcal{M}}(x_i,x_j)+\varepsilon^s\right)\qquad\text{for all }i,j, ∣ f ( x i ) − f ( x j ) ∣ ≤ C CL unif ( Λ + 1 ) m + 1 ∥ f ∥ 1 , n ( d M ( x i , x j ) + ε s ) for all i , j ,
where ∥ f ∥ ∞ : = max i ∣ f ( x i ) ∣ \|f\|_{\infty}:=\max_i|f(x_i)| ∥ f ∥ ∞ := max i ∣ f ( x i ) ∣ and ∥ f ∥ 1 , n : = 1 n ∑ i ∣ f ( x i ) ∣ \|f\|_{1,n}:=\frac{1}{n}\sum_i|f(x_i)| ∥ f ∥ 1 , n := n 1 ∑ i ∣ f ( x i ) ∣ .
Proof. Fix the sign s s s . By Definition 16, L ˉ s \bar{L}^s L ˉ s is the graph Laplacian on the clean points built from the profile η ˉ a s \bar{\eta}_a^s η ˉ a s at the bandwidth ε s \varepsilon^s ε s , scaled as L L L is in § 2 . The clean points are i.i.d. samples from μ \mu μ , and by Lemma 14 the profile η ˉ a s \bar{\eta}_a^s η ˉ a s satisfies (CL1), and (CL2) with S = S 0 S=S_0 S = S 0 and c = c 0 / 3 c=c_0/3 c = c 0 /3 . So [8 Theorem 2.3] , with η ˉ a s \bar{\eta}_a^s η ˉ a s and ε s \varepsilon^s ε s in place of η \eta η and ε \varepsilon ε , gives an event and an estimate of the asserted form. Its hypothesis on f f f is λ f < Λ \lambda_f<\Lambda λ f < Λ , where λ f : = ∥ L ˉ s f ∥ ∞ / ∥ f ∥ ∞ \lambda_f:=\|\bar{L}^s f\|_{\infty}/\|f\|_{\infty} λ f := ∥ L ˉ s f ∥ ∞ /∥ f ∥ ∞ , which is the condition in the statement. The theorem has two smallness conditions on the bandwidth: ε s ≤ c / ( Λ + 1 ) \varepsilon^s\le c/(\Lambda+1) ε s ≤ c / ( Λ + 1 ) for a constant c c c , and ε s \varepsilon^s ε s below a threshold.
It remains to choose the constants independently of a a a . The constants of [8 Theorem 2.3] , including c c c and the threshold, depend on M \mathcal{M} M and ρ \rho ρ , and on the profile only through four quantities: an upper bound on its supremum, an upper bound on its Lipschitz constant on [ 0 , 1 ] [0,1] [ 0 , 1 ] , a lower bound on its second moment, and a positive lower bound on a fixed ball. Proposition 36 in Appendix § 8 verifies this. By Lemma 14 ,
∥ η ˉ a s ∥ ∞ ≤ 2 η ( 0 ) , Lip [ 0 , 1 ] ( η ˉ a s ) ≤ 3 C η , σ η ˉ a s ≥ 1 3 σ η , η ˉ a s ( ∣ w ∣ ) ≥ c 0 3 for w ∈ S 0 , \|\bar{\eta}_a^s\|_{\infty}\le2\eta(0),\qquad\operatorname{Lip}_{[0,1]}(\bar{\eta}_a^s)\le3C_{\eta},\qquad\sigma_{\bar{\eta}_a^s}\ge\tfrac{1}{3}\sigma_{\eta},\qquad\bar{\eta}_a^s(|w|)\ge\frac{c_0}{3}\text{ for }w\in S_0, ∥ η ˉ a s ∥ ∞ ≤ 2 η ( 0 ) , Lip [ 0 , 1 ] ( η ˉ a s ) ≤ 3 C η , σ η ˉ a s ≥ 3 1 σ η , η ˉ a s ( ∣ w ∣ ) ≥ 3 c 0 for w ∈ S 0 ,
and these bounds are the same for every a < a ⋆ a<a_{\star} a < a ⋆ and both signs. The source uses the lower bound through [8 Proposition 9.2] , which is stated for an annulus, so Appendix § 8 applies it on the annulus B ( 0 , t 0 / 2 ) \ B ˉ ( 0 , t 0 / 4 ) B(0,t_0/2)\backslash\bar{B}(0,t_0/4) B ( 0 , t 0 /2 ) \ B ˉ ( 0 , t 0 /4 ) , which lies in S 0 S_0 S 0 and inherits the bound. Proposition 36 , applied to the family of Lemma 35 , therefore gives a single prefactor C C L u n i f C_{\mathrm{CL}}^{\mathrm{unif}} C CL unif , a single pair C P , c P C_{\mathrm{P}},c_{\mathrm{P}} C P , c P of probability constants, and a single constant c ⋆ u n i f c_{\star}^{\mathrm{unif}} c ⋆ unif below both c c c and the threshold, valid for every a < a ⋆ a<a_{\star} a < a ⋆ and both signs. Since Λ > 0 \Lambda>0 Λ > 0 , the assumption ε s ≤ c ⋆ u n i f / ( Λ + 1 ) \varepsilon^s\le c_{\star}^{\mathrm{unif}}/(\Lambda+1) ε s ≤ c ⋆ unif / ( Λ + 1 ) implies both smallness conditions. □ \square □
Eigenvalue Stability
In this section we prove Theorem 1 and Corollary 2. By Lemma 18 and Corollary 19 , for every point cloud both λ l \lambda_l λ l and λ ~ l \widetilde{\lambda}_l λ l lie between ( 1 − a ) m + 2 λ l + (1-a)^{m+2}\lambda_l^+ ( 1 − a ) m + 2 λ l + and ( 1 + a ) m + 2 λ l + (1+a)^{m+2}\lambda_l^+ ( 1 + a ) m + 2 λ l + . Proposition 25 shows that, given an event E n E_n E n of high probability, the difference between the upper and the lower bound, and hence ∣ λ l − λ ~ l ∣ |\lambda_l-\widetilde{\lambda}_l| ∣ λ l − λ l ∣ , is O ( δ / ε ) O(\delta/\varepsilon) O ( δ / ε ) . The proof of Proposition 25 applies the Courant–Fischer principle [23 Theorem 4.2.6, p. 236] for λ l − \lambda_l^- λ l − to the span of the first l l l eigenvectors of L + L^+ L + . For u u u in this span, Lemma 24 bounds ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) (1+a)^{m+2}b^-(u)-(1-a)^{m+2}b^+(u) ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) by a constant multiple of a ∥ u ∥ 2 2 a\|u\|_2^2 a ∥ u ∥ 2 2 , using the Lipschitz estimate of [8 ] for these eigenvectors (Proposition 22 ). This estimate applies because L + L^+ L + is built on the clean points (§ 4 ). § 5.2 defines E n E_n E n , bounds its probability and proves the estimate given E n E_n E n , and § 5.3 completes the proofs.
Bracketing of Dirichlet energies and eigenvalues
Lemma 18 transfers the weight bracket of Lemma 8 to the Dirichlet energies, and Corollary 19 transfers it to the eigenvalues by the Courant–Fischer principle [23 Theorem 4.2.6, p. 236] .
Lemma 18. For every u ∈ R n u \in\mathbb{R}^{n} u ∈ R n ,
( 1 − a ) m + 2 b + ( u ) ≤ b ε ( u ) ≤ ( 1 + a ) m + 2 b − ( u ) , (1-a)^{m+2}b^{+}(u) \le b_{\varepsilon}(u) \le(1+a)^{m+2}b^{-}(u), ( 1 − a ) m + 2 b + ( u ) ≤ b ε ( u ) ≤ ( 1 + a ) m + 2 b − ( u ) , and the same inequalities hold with b ε b_{\varepsilon} b ε replaced by b ~ ε \widetilde{b}_{\varepsilon} b ε .
Proof. Using ( ε ± ) m + 2 = ( 1 ∓ a ) m + 2 ε m + 2 (\varepsilon^{\pm})^{m+2}=(1\mp a)^{m+2}\varepsilon^{m+2} ( ε ± ) m + 2 = ( 1 ∓ a ) m + 2 ε m + 2 and Lemma 8,
( 1 − a ) m + 2 b + ( u ) = 1 n 2 ε m + 2 ∑ i , j ω i j + ( u i − u j ) 2 ≤ 1 n 2 ε m + 2 ∑ i , j ω i j ( u i − u j ) 2 = b ε ( u ) , (1-a)^{m+2}b^{+}(u)=\frac{1}{n^{2}\varepsilon^{m+2}}\sum_{i,j}\omega_{ij}^{+}(u_i-u_j)^2\le\frac{1}{n^{2}\varepsilon^{m+2}}\sum_{i,j}\omega_{ij}(u_i-u_j)^2=b_{\varepsilon}(u), ( 1 − a ) m + 2 b + ( u ) = n 2 ε m + 2 1 i , j ∑ ω ij + ( u i − u j ) 2 ≤ n 2 ε m + 2 1 i , j ∑ ω ij ( u i − u j ) 2 = b ε ( u ) , and the upper bound is analogous. The same argument with ω ~ i j \widetilde{\omega}_{ij} ω ij in place of ω i j \omega_{ij} ω ij proves the inequalities for b ~ ε \widetilde{b}_{\varepsilon} b ε .
Corollary 19. For each l l l ,
( 1 − a ) m + 2 λ l + ≤ λ l ≤ ( 1 + a ) m + 2 λ l − , (1-a)^{m+2}\lambda_l^{+}\le\lambda_l\le(1+a)^{m+2}\lambda_l^{-}, ( 1 − a ) m + 2 λ l + ≤ λ l ≤ ( 1 + a ) m + 2 λ l − , and the same inequalities hold for λ ~ l \widetilde{\lambda}_l λ l . Consequently
∣ λ l − λ ~ l ∣ ≤ ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + . (4) |\lambda_l-\widetilde{\lambda}_l|\le(1+a)^{m+2}\lambda_l^{-}-(1-a)^{m+2}\lambda_l^{+}. \tag*{(4)} ∣ λ l − λ l ∣ ≤ ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + . (4) Proof. By the Courant–Fischer principle [23 Theorem 4.2.6, p. 236] and the identity b ε ( u ) = 2 ⟨ L u , u ⟩ n b_{\varepsilon}(u)=2\langle Lu,u\rangle_n b ε ( u ) = 2 ⟨ Lu , u ⟩ n ,
λ l = 1 2 min V ∈ G l max u ∈ V ∖ { 0 } b ε ( u ) ∥ u ∥ 2 , \lambda_l=\frac{1}{2}\min_{V\in\mathcal{G}_l}\max_{u\in V\setminus\{0\}}\frac{b_{\varepsilon}(u)}{\lVert u\rVert^2}, λ l = 2 1 V ∈ G l min u ∈ V ∖ { 0 } max ∥ u ∥ 2 b ε ( u ) , where G l \mathcal{G}_l G l is the family of l l l -dimensional subspaces. Apply Lemma 18 inside the min-max. The same argument applies to λ ~ l \widetilde{\lambda}_l λ l . Both λ l \lambda_l λ l and λ ~ l \widetilde{\lambda}_l λ l lie inside the interval [ ( 1 − a ) m + 2 λ l + , ( 1 + a ) m + 2 λ l − ] [(1-a)^{m+2}\lambda_l^{+},(1+a)^{m+2}\lambda_l^{-}] [( 1 − a ) m + 2 λ l + , ( 1 + a ) m + 2 λ l − ] , yielding (4).
The high-probability event
Corollary 19 holds for every point cloud. To bound the right-hand side of (4) , Section 5.3 needs two properties of the clean points that hold only with high probability: a bound on the number of sample points in each ball B ( x i , ε − ) B(x_i,\varepsilon^{-}) B ( x i , ε − ) , and hence on the number of pairs of sample points at distance at most ε − \varepsilon^{-} ε − , and a Lipschitz estimate for the first l l l orthonormal eigenvectors u u u of L + L^{+} L + , which bounds ∣ u ( x i ) − u ( x j ) ∣ |u(x_i)-u(x_j)| ∣ u ( x i ) − u ( x j ) ∣ by a constant multiple of d M ( x i , x j ) + ε + d_M(x_i,x_j)+\varepsilon^{+} d M ( x i , x j ) + ε + . Proposition 20 below gives the Lipschitz estimate only for eigenvectors whose eigenvalues lie below a cutoff fixed before the clean points are drawn. A third property, the estimate of [7 Theorem 2.4(2)] for the l l l -th eigenvalue of the clean Laplacian L L L , keeps λ l + \lambda_l^{+} λ l + below a deterministic constant, and the cutoff Λ l \Lambda_l Λ l , defined after Proposition 20, is chosen above twice this constant. Definition 21 introduces the corresponding event E n \mathcal{E}_n E n of the clean points, and Proposition 22 bounds its probability and proves the Lipschitz estimate given E n \mathcal{E}_n E n .
Proposition 20. Fix a cutoff Λ > 0 \Lambda> 0 Λ > 0 and assume ε + ≤ c ⋆ u n i f / ( Λ + 1 ) \varepsilon^{+} \le c_{\star}^{\mathrm{unif}}/(\Lambda+ 1) ε + ≤ c ⋆ unif / ( Λ + 1 ) , where c ⋆ u n i f c_{\star}^{\mathrm{unif}} c ⋆ unif and the constants C C L u n i f C_{\mathrm{CL}}^{\mathrm{unif}} C CL unif , C P C_{\mathrm{P}} C P , c P c_{\mathrm{P}} c P below are those of Lemma 17. Let E n , 1 + ( Λ ) ⊆ M n E_{n,1}^{+}(\Lambda) \subseteq\mathcal{M}^{n} E n , 1 + ( Λ ) ⊆ M n be the event of Lemma 17 for the sign s = + s=+ s = + . The event E n , 1 + ( Λ ) E_{n,1}^{+}(\Lambda) E n , 1 + ( Λ ) depends only on the clean points x 1 , … , x n x_{1},\ldots,x_{n} x 1 , … , x n , and
P n ( E n , 1 + ( Λ ) c ) ≤ C P ( ε + ) − 6 m exp ( − c P n ( ε + ) m + 4 ) + 2 n exp ( − c P n ( Λ + 1 ) − m ) . (5) \mathbb{P}_{n}\left(E_{n,1}^{+}(\Lambda)^{c}\right) \le C_{\mathrm{P}}(\varepsilon^{+})^{-6m}\exp\left(-c_{\mathrm{P}}n(\varepsilon^{+})^{m+4}\right)+2n\exp\left(-c_{\mathrm{P}}n(\Lambda+1)^{-m}\right). \tag*{(5)} P n ( E n , 1 + ( Λ ) c ) ≤ C P ( ε + ) − 6 m exp ( − c P n ( ε + ) m + 4 ) + 2 n exp ( − c P n ( Λ + 1 ) − m ) . (5) Given E n , 1 + ( Λ ) E_{n,1}^{+}(\Lambda) E n , 1 + ( Λ ) , the following holds for every eigenvector u u u of L + L^{+} L + with ∥ u ∥ n = 1 \lVert u\rVert_{n}=1 ∥ u ∥ n = 1 whose eigenvalue λ \lambda λ satisfies ( I a + ) − 1 λ < Λ (I_{a}^{+})^{-1}\lambda< \Lambda ( I a + ) − 1 λ < Λ :
∣ u ( x i ) − u ( x j ) ∣ ≤ C C L u n i f ( Λ + 1 ) m + 1 ( d M ( x i , x j ) + ε + ) for all i , j . \lvert u(x_{i})-u(x_{j})\rvert\le C_{\mathrm{CL}}^{\mathrm{unif}}(\Lambda+1)^{m+1}\left(d_{\mathcal{M}}(x_{i},x_{j})+\varepsilon^{+}\right)\qquad\text{for all }i,j. ∣ u ( x i ) − u ( x j )∣ ≤ C CL unif ( Λ + 1 ) m + 1 ( d M ( x i , x j ) + ε + ) for all i , j .
Proof. The probability bound is that of Lemma 17. Let u u u be as in the statement. Since L ˉ + = ( I a + ) − 1 L + \bar{L}^{+}=(I_{a}^{+})^{-1}L^{+} L ˉ + = ( I a + ) − 1 L + , u u u is an eigenvector of L ˉ + \bar{L}^{+} L ˉ + with eigenvalue λ ˉ = ( I a + ) − 1 λ \bar{\lambda}=(I_{a}^{+})^{-1}\lambda λ ˉ = ( I a + ) − 1 λ , so ∥ L ˉ + u ∥ ∞ = λ ˉ ∥ u ∥ ∞ < Λ ∥ u ∥ ∞ \lVert\bar{L}^{+}u\rVert_{\infty}=\bar{\lambda}\lVert u\rVert_{\infty}<\Lambda\lVert u\rVert_{\infty} ∥ L ˉ + u ∥ ∞ = λ ˉ ∥ u ∥ ∞ < Λ ∥ u ∥ ∞ . Lemma 17 therefore applies to u u u , and the bound follows from ∥ u ∥ 1 , n ≤ ∥ u ∥ n = 1 \lVert u\rVert_{1,n}\le\lVert u\rVert_{n}=1 ∥ u ∥ 1 , n ≤ ∥ u ∥ n = 1 .
From now on we fix l ≥ 1 l\ge1 l ≥ 1 and take the bandwidth
ε = C ε ( log n / n ) 1 / ( m + 4 ) , (6) \varepsilon=C_{\varepsilon}(\log n/n)^{1/(m+4)}, \tag*{(6)} ε = C ε ( log n / n ) 1/ ( m + 4 ) , (6)
where C ε > 0 C_{\varepsilon}>0 C ε > 0 is a constant, fixed independently of n n n . Set
B l : = 2 m + 2 ( σ η λ l M + C l ′ ) , Λ l : = 2 B l + 1 , B_{l}:=2^{m+2}(\sigma_{\eta}\lambda_{l}^{\mathcal{M}}+C_{l}^{\prime}),\qquad\Lambda_{l}:=2B_{l}+1, B l := 2 m + 2 ( σ η λ l M + C l ′ ) , Λ l := 2 B l + 1 ,
where λ l M \lambda_{l}^{\mathcal{M}} λ l M is the l l l -th eigenvalue of Δ ρ \Delta_{\rho} Δ ρ and C l ′ C_{l}^{\prime} C l ′ is the constant of [7 Theorem 2.4(2)] . Both are deterministic, depending only on l l l , M \mathcal{M} M , ρ \rho ρ and the kernel profile η \eta η . We assume that ε + ≤ c ⋆ u n i f / ( Λ l + 1 ) \varepsilon^{+}\le c_{\star}^{\mathrm{unif}}/(\Lambda_{l}+1) ε + ≤ c ⋆ unif / ( Λ l + 1 ) , the bandwidth condition of Proposition 20 at Λ = Λ l \Lambda=\Lambda_{l} Λ = Λ l , and that ε ≤ 1 \varepsilon\le1 ε ≤ 1 . Since ε + ≤ ε \varepsilon^{+}\le\varepsilon ε + ≤ ε , the first holds whenever ε ≤ c ⋆ u n i f / ( 2 B l + 2 ) \varepsilon\le c_{\star}^{\mathrm{unif}}/(2B_{l}+2) ε ≤ c ⋆ unif / ( 2 B l + 2 ) , so under (6) both hold for all n n n large enough. Finally, C 1 C_{1} C 1 denotes the constant, depending only on M \mathcal{M} M and ρ \rho ρ , that the proof of Proposition 22 obtains from [8 Corollary 3.3] .
Definition 21 (High-probability event). Let E n ⊆ M n E_{n}\subseteq\mathcal{M}^{n} E n ⊆ M n be the intersection E n , 1 ∩ E n , 2 ∩ E n , 3 E_{n,1}\cap E_{n,2}\cap E_{n,3} E n , 1 ∩ E n , 2 ∩ E n , 3 of the following three events.
(Lipschitz bound for the eigenvectors of L + L^{+} L + .) E n , 1 : = E n , 1 + ( Λ l ) E_{n,1}:=E_{n,1}^{+}(\Lambda_{l}) E n , 1 := E n , 1 + ( Λ l ) , the event of Proposition 20 at Λ = Λ l \Lambda=\Lambda_{l} Λ = Λ l .
(Degree upper bound at radius ε − \varepsilon^{-} ε − .) E n , 2 E_{n,2} E n , 2 is the event that, for every i ∈ { 1 , … , n } i\in\{1,\ldots,n\} i ∈ { 1 , … , n } ,
# { j : ∥ x i − x j ∥ ≤ ε − } ≤ C 1 n ( ε − ) m . \#\{j:\lVert x_{i}-x_{j}\rVert\le\varepsilon^{-}\}\le C_{1}n(\varepsilon^{-})^{m}. # { j : ∥ x i − x j ∥ ≤ ε − } ≤ C 1 n ( ε − ) m . (Eigenvalue convergence for the clean Laplacian L L L .) E n , 3 E_{n,3} E n , 3 is the event that ∣ λ l − σ η λ l M ∣ ≤ C l ′ ε \lvert\lambda_{l}-\sigma_{\eta}\lambda_{l}^{\mathcal{M}}\rvert\le C_{l}^{\prime}\varepsilon ∣ λ l − σ η λ l M ∣ ≤ C l ′ ε .
The constant C l ′ C_{l}^{\prime} C l ′ and the event E n , 3 E_{n,3} E n , 3 come from [7 Theorem 2.4(2)] . Appendix C states the hypotheses of that theorem in our notation, chooses the parameters θ \theta θ and δ e \delta_{e} δ e of [7 ] so that they hold for n n n large enough (Lemma 44), and shows that, for each p ≥ 1 p\ge1 p ≥ 1 , once C ε C_{\varepsilon} C ε is large enough,
P n ( E n , 3 ) ≥ 1 − O ( n − p ) , ∣ λ l − σ η λ l M ∣ ≤ C l ′ ε given E n , 3 . (7) \mathbb{P}_{n}(E_{n,3})\ge1-O(n^{-p}),\qquad\lvert\lambda_{l}-\sigma_{\eta}\lambda_{l}^{\mathcal{M}}\rvert\le C_{l}^{\prime}\varepsilon\ \text{ given }E_{n,3}. \tag*{(7)} P n ( E n , 3 ) ≥ 1 − O ( n − p ) , ∣ λ l − σ η λ l M ∣ ≤ C l ′ ε given E n , 3 . (7)
Proposition 22. The event E n E_{n} E n of Definition 21 has the following properties.
Let p ≥ 1 p\ge1 p ≥ 1 . If C ε C_{\varepsilon} C ε is large enough, depending on p p p , l l l , M \mathcal{M} M , ρ \rho ρ and η \eta η , then P n ( E n ) ≥ 1 − O ( n − p ) \mathbb{P}_{n}(E_{n})\ge1-O(n^{-p}) P n ( E n ) ≥ 1 − O ( n − p ) for n n n large enough.
Given E n \mathcal{E}_n E n , λ l + ≤ B l \lambda_l^+ \le B_l λ l + ≤ B l .
Let u 1 , … , u l u^1,\ldots,u^l u 1 , … , u l be L 2 ( X n ) L^2(X_n) L 2 ( X n ) -orthonormal eigenvectors of L + L^+ L + corresponding to its first l l l eigenvalues. Given E n \mathcal{E}_n E n , for every k ≤ l k\le l k ≤ l and all i , j i,j i , j ,
∣ u k ( x i ) − u k ( x j ) ∣ ≤ C a u x ( B l + 1 ) m + 1 ( d M ( x i , x j ) + ε + ) , |u^k(x_i)-u^k(x_j)|\le C_{\mathrm{aux}}(B_l+1)^{m+1}(d_{\mathcal{M}}(x_i,x_j)+\varepsilon^+), ∣ u k ( x i ) − u k ( x j ) ∣ ≤ C aux ( B l + 1 ) m + 1 ( d M ( x i , x j ) + ε + ) ,
where
C a u x : = 2 m + 1 C C L u n i f (8) C_{\mathrm{aux}}:=2^{m+1}C_{\mathrm{CL}}^{\mathrm{unif}} \tag*{(8)} C aux := 2 m + 1 C CL unif (8)
depends only on M \mathcal{M} M , ρ \rho ρ and η \eta η , through the constant of Lemma 17.
Proof. In this proof the constants C , c > 0 C,c>0 C , c > 0 may differ from display to display, but each depends only on M , ρ \mathcal{M},\rho M , ρ and the kernel profile η \eta η .
The probability bound. For E n , 1 \mathcal{E}_{n,1} E n , 1 , by (5) with Λ = Λ l \Lambda=\Lambda_l Λ = Λ l , the complement has probability at most
P 1 : = C ( ε + ) − 6 m exp ( − c n ( ε + ) m + 4 ) + 2 n exp ( − c n ( Λ l + 1 ) − m ) , (9) P_1:=C(\varepsilon^+)^{-6m}\exp(-cn(\varepsilon^+)^{m+4})+2n\exp(-cn(\Lambda_l+1)^{-m}), \tag*{(9)} P 1 := C ( ε + ) − 6 m exp ( − c n ( ε + ) m + 4 ) + 2 n exp ( − c n ( Λ l + 1 ) − m ) , (9)
and these constants do not depend on a a a .
For E n , 2 \mathcal{E}_{n,2} E n , 2 , by [8 Corollary 3.3] , the complement has probability at most
P 2 : = 2 n exp ( − c n ( ε − ) m ) . (10) P_2:=2n\exp(-cn(\varepsilon^-)^m). \tag*{(10)} P 2 := 2 n exp ( − c n ( ε − ) m ) . (10)
[8 Corollary 3.3] bounds the number of sample points in a ball B ( x , ε − ) B(x,\varepsilon^-) B ( x , ε − ) whose centre x x x does not depend on the sample points. Here the centre is itself a sample point, so for each i i i we condition on x i = x x_i=x x i = x . The other n − 1 n-1 n − 1 points are still independent samples from μ \mu μ , and the corollary applies to the number of them in B ( x , ε − ) B(x,\varepsilon^-) B ( x , ε − ) . For n ( ε − ) m n(\varepsilon^-)^m n ( ε − ) m large enough, the term j = i j=i j = i and the change from n − 1 n-1 n − 1 to n n n points are absorbed into C 1 C_1 C 1 and c c c . Integrating over x x x and taking a union bound over i i i gives (10).
For E n , 3 \mathcal{E}_{n,3} E n , 3 , set
P 3 : = P n ( E n , 3 c ) . P_3:=\mathbb{P}_n(\mathcal{E}_{n,3}^{c}). P 3 := P n ( E n , 3 c ) .
By a union bound over the three events,
P n ( E n ) ≥ 1 − P 1 − P 2 − P 3 = 1 − C ( ε + ) − 6 m exp ( − c n ( ε + ) m + 4 ) − 2 n exp ( − c n ( Λ l + 1 ) − m ) \mathbb{P}_n(\mathcal{E}_n)\ge1-P_1-P_2-P_3
=1-C(\varepsilon^+)^{-6m}\exp(-cn(\varepsilon^+)^{m+4})-2n\exp(-cn(\Lambda_l+1)^{-m}) P n ( E n ) ≥ 1 − P 1 − P 2 − P 3 = 1 − C ( ε + ) − 6 m exp ( − c n ( ε + ) m + 4 ) − 2 n exp ( − c n ( Λ l + 1 ) − m )
− 2 n exp ( − c n ( ε − ) m ) − P 3 . (11) \qquad{}-2n\exp(-cn(\varepsilon^-)^m)-P_3. \tag*{(11)} − 2 n exp ( − c n ( ε − ) m ) − P 3 . (11)
By (7) (Appendix C), P 3 = O ( n − p ) P_3=O(n^{-p}) P 3 = O ( n − p ) . Since a < a ⋆ ≤ 1 / 2 a<a_\star\le1/2 a < a ⋆ ≤ 1/2 (Remark 15), we have 1 2 ε ≤ ε + ≤ 3 2 ε \frac{1}{2}\varepsilon\le\varepsilon^+\le\frac{3}{2}\varepsilon 2 1 ε ≤ ε + ≤ 2 3 ε . After increasing C ε C_\varepsilon C ε if necessary, the remaining terms in (11) are likewise O ( n − p ) O(n^{-p}) O ( n − p ) . Hence P n ( E n ) ≥ 1 − O ( n − p ) \mathbb{P}_n(\mathcal{E}_n)\ge1-O(n^{-p}) P n ( E n ) ≥ 1 − O ( n − p ) .
The bound on λ l + \lambda_l^+ λ l + . By Corollary 19 applied to the clean Laplacian, ( 1 − a ) m + 2 λ l + ≤ λ l (1-a)^{m+2}\lambda_l^+\le\lambda_l ( 1 − a ) m + 2 λ l + ≤ λ l , so λ l + ≤ ( 1 − a ) − ( m + 2 ) λ l \lambda_l^+\le(1-a)^{-(m+2)}\lambda_l λ l + ≤ ( 1 − a ) − ( m + 2 ) λ l . Given E n \mathcal{E}_n E n , item 3 of Definition 21 gives λ l ≤ σ η λ l M + C l ′ \lambda_l\le\sigma_\eta\lambda_l^{\mathcal{M}}+C'_l λ l ≤ σ η λ l M + C l ′ , since ε ≤ 1 \varepsilon\le1 ε ≤ 1 . Since a < a ⋆ ≤ 1 / 2 a<a_\star\le1/2 a < a ⋆ ≤ 1/2 , we have ( 1 − a ) − ( m + 2 ) ≤ 2 m + 2 (1-a)^{-(m+2)}\le2^{m+2} ( 1 − a ) − ( m + 2 ) ≤ 2 m + 2 , whence λ l + ≤ 2 m + 2 ( σ η λ l M + C l ′ ) = B l \lambda_l^+\le2^{m+2}(\sigma_\eta\lambda_l^{\mathcal{M}}+C'_l)=B_l λ l + ≤ 2 m + 2 ( σ η λ l M + C l ′ ) = B l .
The Lipschitz estimate. Item 1 of Definition 21 is the event E n , 1 = E n , 1 + ( Λ l ) \mathcal{E}_{n,1}=\mathcal{E}_{n,1}^+(\Lambda_l) E n , 1 = E n , 1 + ( Λ l ) of Proposition 20 , with Λ l = 2 B l + 1 \Lambda_l=2B_l+1 Λ l = 2 B l + 1 . For k ≤ l k\le l k ≤ l , the bound λ l + ≤ B l \lambda_l^+\le B_l λ l + ≤ B l proved above and I a + ≥ 1 2 I_a^+\ge\frac{1}{2} I a + ≥ 2 1 (Lemma 14 ) give
( I a + ) − 1 λ k + ≤ 2 λ k + ≤ 2 λ l + ≤ 2 B l < 2 B l + 1 = Λ l , (I_a^+)^{-1}\lambda_k^+\le2\lambda_k^+\le2\lambda_l^+\le2B_l<2B_l+1=\Lambda_l, ( I a + ) − 1 λ k + ≤ 2 λ k + ≤ 2 λ l + ≤ 2 B l < 2 B l + 1 = Λ l ,
so the cutoff condition of Proposition 20 is met. Its bandwidth condition ε + ≤ c ⋆ u n i f / ( Λ l + 1 ) \varepsilon^+\le c_\star^{\mathrm{unif}}/(\Lambda_l+1) ε + ≤ c ⋆ unif / ( Λ l + 1 ) is assumed before Definition 21 . Applying the proposition at Λ = Λ l \Lambda=\Lambda_l Λ = Λ l and using Λ l + 1 = 2 ( B l + 1 ) \Lambda_l+1=2(B_l+1) Λ l + 1 = 2 ( B l + 1 ) with C a u x = 2 m + 1 C C L u n i f C_{\mathrm{aux}}=2^{m+1}C_{\mathrm{CL}}^{\mathrm{unif}} C aux = 2 m + 1 C CL unif gives the stated bound.
Proofs of the eigenvalue results
It remains to bound the right-hand side of (4) , which Proposition 25 does given E n \mathcal{E}_n E n . This section ends with the proofs of Theorem 1 and Corollary 2. We compare Euclidean and geodesic distances by [8 (3.2)] : there exists R > 0 R > 0 R > 0 , depending only on M \mathcal{M} M , such that for all x , y ∈ M x,y \in\mathcal{M} x , y ∈ M with ∥ x − y ∥ ≤ R / 2 \lVert x-y\rVert\le R/2 ∥ x − y ∥ ≤ R /2 ,
∥ x − y ∥ ≤ d M ( x , y ) ≤ ∥ x − y ∥ + 8 R 2 ∥ x − y ∥ 3 . \lVert x-y\rVert\le d_{\mathcal{M}}(x,y) \le\lVert x-y\rVert+ \frac{8}{R^2}\lVert x-y\rVert^3. ∥ x − y ∥ ≤ d M ( x , y ) ≤ ∥ x − y ∥ + R 2 8 ∥ x − y ∥ 3 .
From here on we also assume that ε ≤ R / 3 \varepsilon\le R/3 ε ≤ R /3 ; under the scaling (6) , this holds for large n n n .
Lemma 23. Let a < 1 / 2 a < 1/2 a < 1/2 and ε ≤ R / 3 \varepsilon\le R/3 ε ≤ R /3 . If x , y ∈ M x,y \in\mathcal{M} x , y ∈ M satisfy ∥ x − y ∥ ≤ ε − = ( 1 + a ) ε \lVert x-y\rVert\le\varepsilon^- = (1+a)\varepsilon ∥ x − y ∥ ≤ ε − = ( 1 + a ) ε , then d M ( x , y ) ≤ 3 ε − ≤ 5 ε d_{\mathcal{M}}(x,y) \le3\varepsilon^- \le5\varepsilon d M ( x , y ) ≤ 3 ε − ≤ 5 ε .
Proof. Since a < 1 / 2 a < 1/2 a < 1/2 and ε ≤ R / 3 \varepsilon\le R/3 ε ≤ R /3 , we have ε − < 3 2 ε ≤ R / 2 \varepsilon^- < \frac{3}{2}\varepsilon\le R/2 ε − < 2 3 ε ≤ R /2 , so the hypothesis ∥ x − y ∥ ≤ ε − \lVert x-y\rVert\le\varepsilon^- ∥ x − y ∥ ≤ ε − puts the pair in the range of [8 (3.2)] . Moreover,
8 R 2 ∥ x − y ∥ 3 = 8 R 2 ∥ x − y ∥ 2 ∥ x − y ∥ ≤ 8 R 2 ( R 2 ) 2 ε − = 2 ε − . \frac{8}{R^2}\lVert x-y\rVert^3 = \frac{8}{R^2}\lVert x-y\rVert^2\lVert x-y\rVert\le\frac{8}{R^2}\left(\frac{R}{2}\right)^2\varepsilon^- = 2\varepsilon^-. R 2 8 ∥ x − y ∥ 3 = R 2 8 ∥ x − y ∥ 2 ∥ x − y ∥ ≤ R 2 8 ( 2 R ) 2 ε − = 2 ε − .
Therefore d M ( x , y ) ≤ 3 ε − d_{\mathcal{M}}(x,y) \le3\varepsilon^- d M ( x , y ) ≤ 3 ε − . Since a < 1 / 2 a < 1/2 a < 1/2 , ε − = ( 1 + a ) ε ≤ 3 ε / 2 \varepsilon^- = (1+a)\varepsilon\le3\varepsilon/2 ε − = ( 1 + a ) ε ≤ 3 ε /2 , hence d M ( x , y ) ≤ 9 ε / 2 ≤ 5 ε d_{\mathcal{M}}(x,y) \le9\varepsilon/2 \le5\varepsilon d M ( x , y ) ≤ 9 ε /2 ≤ 5 ε . □
Lemma 24. Let u 1 , … , u l u^1,\ldots,u^l u 1 , … , u l be L 2 ( X n ) L^2(X_n) L 2 ( X n ) -orthonormal eigenvectors of L + L^+ L + corresponding to its first l l l eigenvalues, and set
S + : = Span { u 1 , … , u l } . S^+ := \operatorname{Span}\{u^1,\ldots,u^l\}. S + := Span { u 1 , … , u l } . Then, given E n \mathcal{E}_n E n , for every u ∈ S + u \in S^+ u ∈ S + ,
( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) ≤ K l a ∥ u ∥ n 2 , (1+a)^{m+2}b^-(u) - (1-a)^{m+2}b^+(u) \le K_l a\lVert u\rVert_n^2, ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) ≤ K l a ∥ u ∥ n 2 , where one may take
K l : = 2 ( 3 / 2 ) m C 1 C η l ( 6 C a u x ( B l + 1 ) m + 1 ) 2 (12) K_l := 2(3/2)^m C_1 C_\eta l\left(6C_{\mathrm{aux}}(B_l+1)^{m+1}\right)^2 \tag*{(12)} K l := 2 ( 3/2 ) m C 1 C η l ( 6 C aux ( B l + 1 ) m + 1 ) 2 (12) with C η C_\eta C η the Lipschitz constant of η \eta η in (K3), C a u x C_{\mathrm{aux}} C aux as in (8) , and B l B_l B l and C 1 C_1 C 1 as in Section 5.2.
Proof. Write u = ∑ k = 1 l β k u k u=\sum_{k=1}^l\beta_k u^k u = ∑ k = 1 l β k u k , so that ∥ u ∥ n 2 = ∑ k β k 2 \lVert u\rVert_n^2=\sum_k\beta_k^2 ∥ u ∥ n 2 = ∑ k β k 2 .
By item 3 of Proposition 22 and ε + ≤ ε \varepsilon^+ \le\varepsilon ε + ≤ ε , given E n \mathcal{E}_n E n we have, for every k ≤ l k\le l k ≤ l and all i , j i,j i , j ,
∣ u k ( x i ) − u k ( x j ) ∣ ≤ C a u x ( B l + 1 ) m + 1 ( d M ( x i , x j ) + ε ) . \left|u^k(x_i)-u^k(x_j)\right| \le C_{\mathrm{aux}}(B_l+1)^{m+1}\left(d_{\mathcal{M}}(x_i,x_j)+\varepsilon\right). u k ( x i ) − u k ( x j ) ≤ C aux ( B l + 1 ) m + 1 ( d M ( x i , x j ) + ε ) .
For pairs ( i , j ) (i,j) ( i , j ) with ∥ x i − x j ∥ ≤ ε − \lVert x_i-x_j\rVert\le\varepsilon^- ∥ x i − x j ∥ ≤ ε − , Lemma 23 gives d M ( x i , x j ) ≤ 5 ε d_{\mathcal{M}}(x_i,x_j) \le5\varepsilon d M ( x i , x j ) ≤ 5 ε . Substituting,
∣ u k ( x i ) − u k ( x j ) ∣ ≤ 6 C a u x ( B l + 1 ) m + 1 ε = : M l ε . (13) \left|u^k(x_i)-u^k(x_j)\right| \le6C_{\mathrm{aux}}(B_l+1)^{m+1}\varepsilon=: M_l\varepsilon. \tag*{(13)} u k ( x i ) − u k ( x j ) ≤ 6 C aux ( B l + 1 ) m + 1 ε =: M l ε . (13)
For such a pair, Cauchy–Schwarz and (13) give
∣ u i − u j ∣ ≤ ∑ k = 1 l ∣ β k ∣ ∣ u k ( x i ) − u k ( x j ) ∣ ≤ M l ε ∑ k = 1 l ∣ β k ∣ ≤ M l ε l ∥ u ∥ n , |u_i-u_j| \le\sum_{k=1}^l|\beta_k|\left|u^k(x_i)-u^k(x_j)\right| \le M_l\varepsilon\sum_{k=1}^l|\beta_k| \le M_l\varepsilon\sqrt{l}\lVert u\rVert_n, ∣ u i − u j ∣ ≤ k = 1 ∑ l ∣ β k ∣ u k ( x i ) − u k ( x j ) ≤ M l ε k = 1 ∑ l ∣ β k ∣ ≤ M l ε l ∥ u ∥ n ,
hence
( u i − u j ) 2 ≤ M l 2 ε 2 l ∥ u ∥ n 2 . (14) (u_i-u_j)^2 \le M_l^2\varepsilon^2l\lVert u\rVert_n^2. \tag*{(14)} ( u i − u j ) 2 ≤ M l 2 ε 2 l ∥ u ∥ n 2 . (14)
Next we compare the weights. The function η \eta η is C η C_\eta C η -Lipschitz on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) (Lemma 11), so for every i , j i,j i , j , with r i j : = ∥ x i − x j ∥ / ε r_{ij}:=\lVert x_i-x_j\rVert/\varepsilon r ij := ∥ x i − x j ∥ / ε ,
ω i j − − ω i j + = η ( ( r i j − a ) + ) − η ( a + r i j ) ≤ C η ∣ ( r i j − a ) + − ( a + r i j ) ∣ ≤ 2 C η a . \omega_{ij}^- - \omega_{ij}^+ = \eta((r_{ij}-a)_+) - \eta(a+r_{ij}) \le C_\eta\left|(r_{ij}-a)_+-(a+r_{ij})\right| \le2C_\eta a. ω ij − − ω ij + = η (( r ij − a ) + ) − η ( a + r ij ) ≤ C η ∣ ( r ij − a ) + − ( a + r ij ) ∣ ≤ 2 C η a .
Moreover ω ˉ i j = 0 \bar{\omega}_{ij}=0 ω ˉ ij = 0 whenever ∥ x i − x j ∥ > ε − \lVert x_i-x_j\rVert>\varepsilon^{-} ∥ x i − x j ∥ > ε − .
Finally, the number of pairs ( i , j ) (i,j) ( i , j ) with ∥ x i − x j ∥ ≤ ε − \lVert x_i-x_j\rVert\le\varepsilon^{-} ∥ x i − x j ∥ ≤ ε − is at most n ⋅ C 1 n ( ε − ) m ≤ C 1 ( 1 + a ) m n 2 ε m ≤ ( 3 / 2 ) m C 1 n 2 ε m n\cdot C_1n(\varepsilon^{-})^m\le C_1(1+a)^m n^2\varepsilon^m\le(3/2)^mC_1n^2\varepsilon^m n ⋅ C 1 n ( ε − ) m ≤ C 1 ( 1 + a ) m n 2 ε m ≤ ( 3/2 ) m C 1 n 2 ε m given E n \mathcal{E}_n E n (item 2 of Definition 21). Combining this with the weight bound and (14),
( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) = 1 n 2 ε m + 2 ∑ ∥ x i − x j ∥ ≤ ε − ( ω ˉ i j − ω i j + ) ( u i − u j ) 2 ≤ 1 n 2 ε m + 2 ⋅ ( 3 / 2 ) m C 1 n 2 ε m ⋅ 2 C η a ⋅ M l 2 ε 2 ∥ u ∥ n 2 = 2 ( 3 / 2 ) m C 1 C η M l 2 ⏟ = : K l a ∥ u ∥ n 2 . \begin{aligned}
(1+a)^{m+2b^{-}(u)}-(1-a)^{m+2b^{+}(u)}
=\frac{1}{n^2\varepsilon^{m+2}}\sum_{\lVert x_i-x_j\rVert\le\varepsilon^{-}}(\bar{\omega}_{ij}-\omega_{ij}^{+})(u_i-u_j)^2 \\
&\le\frac{1}{n^2\varepsilon^{m+2}}\cdot(3/2)^mC_1n^2\varepsilon^m\cdot2C\eta a\cdot M_l^2\varepsilon^2\lVert u\rVert_n^2 \\
&=\underbrace{2(3/2)^mC_1C\eta M_l^2}_{=:K_l}a\lVert u\rVert_n^2.
\end{aligned} ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) = n 2 ε m + 2 1 ∥ x i − x j ∥ ≤ ε − ∑ ( ω ˉ ij − ω ij + ) ( u i − u j ) 2 ≤ n 2 ε m + 2 1 ⋅ ( 3/2 ) m C 1 n 2 ε m ⋅ 2 C η a ⋅ M l 2 ε 2 ∥ u ∥ n 2 = =: K l 2 ( 3/2 ) m C 1 C η M l 2 a ∥ u ∥ n 2 .
Proposition 25. Given E n \mathcal{E}_n E n ,
( 1 + a ) m + 2 λ l − ≤ ( 1 − a ) m + 2 λ l + + 1 2 K l a . (1+a)^{m+2}\lambda_l^{-}\le(1-a)^{m+2}\lambda_l^{+}+\frac{1}{2}K_l a. ( 1 + a ) m + 2 λ l − ≤ ( 1 − a ) m + 2 λ l + + 2 1 K l a .
with K l K_l K l as in Lemma 24.
Proof. The subspace S + S^{+} S + of Lemma 24 is l l l -dimensional, so the Courant–Fischer principle, applied as in the proof of Corollary 19, gives
( 1 + a ) m + 2 λ l − = 1 2 min V ∈ G l max u ∈ V ∖ { 0 } ( 1 + a ) m + 2 b − ( u ) ∥ u ∥ n 2 ≤ 1 2 max u ∈ S + ∖ { 0 } ( 1 + a ) m + 2 b − ( u ) ∥ u ∥ n 2 . (1+a)^{m+2}\lambda_l^{-}
=\frac{1}{2}\min_{V\in\mathcal{G}_l}\max_{u\in V\setminus\{0\}}\frac{(1+a)^{m+2}b^{-}(u)}{\lVert u\rVert_n^2}
\le\frac{1}{2}\max_{u\in S^{+}\setminus\{0\}}\frac{(1+a)^{m+2}b^{-}(u)}{\lVert u\rVert_n^2}. ( 1 + a ) m + 2 λ l − = 2 1 V ∈ G l min u ∈ V ∖ { 0 } max ∥ u ∥ n 2 ( 1 + a ) m + 2 b − ( u ) ≤ 2 1 u ∈ S + ∖ { 0 } max ∥ u ∥ n 2 ( 1 + a ) m + 2 b − ( u ) .
where, as there, G l \mathcal{G}_l G l is the family of l l l -dimensional subspaces. By Lemma 24,
( 1 + a ) m + 2 b − ( u ) ∥ u ∥ n 2 ≤ ( 1 − a ) m + 2 b + ( u ) ∥ u ∥ n 2 + K l a . \frac{(1+a)^{m+2}b^{-}(u)}{\lVert u\rVert_n^2}
\le\frac{(1-a)^{m+2}b^{+}(u)}{\lVert u\rVert_n^2}+K_l a. ∥ u ∥ n 2 ( 1 + a ) m + 2 b − ( u ) ≤ ∥ u ∥ n 2 ( 1 − a ) m + 2 b + ( u ) + K l a .
The maximum of the first term over u ∈ S + ∖ { 0 } u\in S^{+}\setminus\{0\} u ∈ S + ∖ { 0 } is exactly 2 ( 1 − a ) m + 2 λ l + 2(1-a)^{m+2}\lambda_l^{+} 2 ( 1 − a ) m + 2 λ l + , since b + ( u ) = 2 ⟨ L + u , u ⟩ n b^{+}(u)=2\langle L^{+}u,u\rangle_n b + ( u ) = 2 ⟨ L + u , u ⟩ n and S + S^{+} S + is spanned by orthonormal eigenvectors of L + L^{+} L + with eigenvalues λ 1 + , … , λ l + \lambda_1^{+},\ldots,\lambda_l^{+} λ 1 + , … , λ l + . Multiplying the displayed inequality by the outside factor 1 / 2 1/2 1/2 gives the claim.
Proof of Theorem 1. By Proposition 22, P n ( E n ) ≥ 1 − O ( n − p ) \mathbb{P}_n(\mathcal{E}_n)\ge1-O(n^{-p}) P n ( E n ) ≥ 1 − O ( n − p ) . Given E n \mathcal{E}_n E n , combine (4) with Proposition 25:
∣ λ l − λ ~ l ∣ ≤ ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + ≤ 1 2 K l a ≤ K l δ ε . \lvert\lambda_l-\widetilde{\lambda}_l\rvert\le(1+a)^{m+2}\lambda_l^{-}-(1-a)^{m+2}\lambda_l^{+}\le\frac{1}{2}K_l a\le\frac{K_l\delta}{\varepsilon}. ∣ λ l − λ l ∣ ≤ ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + ≤ 2 1 K l a ≤ ε K l δ .
Proof of Corollary 2. By Proposition 22, P n ( E n ) ≥ 1 − O ( n − p ) \mathbb{P}_n(\mathcal{E}_n)\ge1-O(n^{-p}) P n ( E n ) ≥ 1 − O ( n − p ) . Given E n \mathcal{E}_n E n , both estimates below hold: the first by the proof of Theorem 1, and the second by (7), since E n \mathcal{E}_n E n is contained in the Calder–García Trillos event E n , 3 \mathcal{E}_{n,3} E n , 3 of item 3 of Definition 21. The triangle inequality gives
∣ λ ~ l − σ η λ l M ∣ ≤ ∣ λ ~ l − λ l ∣ ⏟ ≤ K l δ / ε + ∣ λ l − σ η λ l M ∣ ⏟ ≤ C l ′ ε ≤ K l δ ε + C l ′ ε . \lvert\widetilde{\lambda}_l-\sigma_\eta\lambda_l^{\mathcal{M}}\rvert
\le\underbrace{\lvert\widetilde{\lambda}_l-\lambda_l\rvert}_{\le K_l\delta/\varepsilon}
+\underbrace{\lvert\lambda_l-\sigma_\eta\lambda_l^{\mathcal{M}}\rvert}_{\le C_l'\varepsilon}
\le K_l\frac{\delta}{\varepsilon}+C_l'\varepsilon. ∣ λ l − σ η λ l M ∣ ≤ ≤ K l δ / ε ∣ λ l − λ l ∣ + ≤ C l ′ ε ∣ λ l − σ η λ l M ∣ ≤ K l ε δ + C l ′ ε .
Eigenvector Stability
In this section we prove Theorems 3 and 4, the analogues of Theorem 1 and Corollary 2 for eigenvectors. Theorem 3 bounds the distance between the spectral projectors P l P_l P l and P ~ l \widetilde{P}_l P l of § 3 , rather than between single eigenvectors. When λ l M \lambda_l^{M} λ l M is a multiple eigenvalue, L L L has several eigenvalues close to λ l \lambda_l λ l , and a perturbation larger than the differences between them can mix their eigenvectors. So an eigenvector of L L L for λ l \lambda_l λ l need not be close to any eigenvector of L ~ \widetilde{L} L . The projector P l P_l P l does not have this problem: Theorem 3 needs only γ l M = λ l + 1 M − λ l M > 0 \gamma_l^{M}=\lambda_{l+1}^{M}-\lambda_l^{M}>0 γ l M = λ l + 1 M − λ l M > 0 , and λ l M \lambda_l^{M} λ l M itself may be a multiple eigenvalue.
The proof of Theorem 3 bounds the distances from P l P_l P l and from P ~ l \widetilde{P}_l P l to P l + P_l^{+} P l + , the orthogonal projection onto the subspace spanned by the first l l l eigenvectors of L + L^{+} L + . Lemma 26 restates the energy bracket of Lemma 18 as the inequalities A ⪯ L ⪯ B A\preceq L\preceq B A ⪯ L ⪯ B and A ⪯ L ~ ⪯ B A\preceq\widetilde{L}\preceq B A ⪯ L ⪯ B , where A = ( 1 − a ) m + 2 L + A=(1-a)^{m+2}L^{+} A = ( 1 − a ) m + 2 L + and B = ( 1 + a ) m + 2 L − B=(1+a)^{m+2}L^{-} B = ( 1 + a ) m + 2 L − are the comparison operators and S ⪯ T S\preceq T S ⪯ T means that T − S T-S T − S is positive semidefinite. Lemma 26 also bounds ⟨ ( B − A ) u , u ⟩ n \langle(B-A)u,u\rangle_n ⟨( B − A ) u , u ⟩ n by 1 2 K l a ∥ u ∥ n 2 \frac{1}{2}K_l a\lVert u\rVert_n^2 2 1 K l a ∥ u ∥ n 2 for every u u u in the range of P l + P_l^{+} P l + , given E n \mathcal{E}_n E n , using Lemma 24. Lemma 28 and Proposition 29 show that P l P_l P l and P ~ l \widetilde{P}_l P l are close to P l + P_l^{+} P l + if 1 2 K l a \frac{1}{2}K_l a 2 1 K l a is small compared with the spectral gap Γ l + \Gamma_l^{+} Γ l + between the l l l -th and ( l + 1 ) (l+1) ( l + 1 ) -st eigenvalues of A A A , defined in (17) . Under (25) , Lemma 30 bounds Γ l + \Gamma_l^{+} Γ l + from below by 1 2 σ η γ l M \frac{1}{2}\sigma\eta\gamma_l^{M} 2 1 σ η γ l M , given the events of Definition 21 at the indices l l l and l + 1 l+1 l + 1 , and the triangle inequality then gives Theorem 3.
In § 6.4 we apply the bound of Theorem 3 on both sides of the multiplicity block l − , … , l + l_{-},\ldots,l_{+} l − , … , l + of λ l M \lambda_l^{M} λ l M , at the indices l − − 1 l_{-}-1 l − − 1 and l + l_{+} l + , and combine it with the proof of [7 Theorem 2.7(2)] to prove Theorem 4. The claim of Remark 5 is proved at the end of that section.
Notation
Throughout Section 6, 1 ≤ l < n 1\le l<n 1 ≤ l < n is fixed and R n \mathbb{R}^{n} R n carries the inner product ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n of L 2 ( X n ) L^2(X_n) L 2 ( X n ) . For an operator T T T on R n \mathbb{R}^{n} R n , T ∗ T^{*} T ∗ denotes its adjoint; since ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n is 1 / n 1/n 1/ n times the Euclidean inner product, T ∗ T^{*} T ∗ is the transpose of T T T . We write Tr ( T ) \operatorname{Tr}(T) Tr ( T ) for the trace of T T T , the sum of the diagonal entries of its matrix in any basis. The trace does not depend on the inner product: in a ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n -orthonormal basis e 1 , … , e n e_1,\ldots,e_n e 1 , … , e n , the i i i -th diagonal entry of the matrix of T T T is ⟨ T e i , e i ⟩ n \langle T e_i,e_i\rangle_n ⟨ T e i , e i ⟩ n , so
Tr ( T ) = ∑ i = 1 n ⟨ T e i , e i ⟩ n . (15) \operatorname{Tr}(T)=\sum_{i=1}^{n}\langle T e_i,e_i\rangle_n. \tag*{(15)} Tr ( T ) = i = 1 ∑ n ⟨ T e i , e i ⟩ n . (15)
The Hilbert–Schmidt norm of Section 3 satisfies ∥ T ∥ H S , n 2 = Tr ( T ∗ T ) \lVert T\rVert_{\mathrm{HS},n}^{2}=\operatorname{Tr}(T^{*}T) ∥ T ∥ HS , n 2 = Tr ( T ∗ T ) . We define the operator norm ∥ T ∥ o p , n : = max ∥ u ∥ n = 1 ∥ T u ∥ n \lVert T\rVert_{\mathrm{op},n}:=\max_{\lVert u\rVert_n=1}\lVert T u\rVert_n ∥ T ∥ op , n := max ∥ u ∥ n = 1 ∥ T u ∥ n , which is the spectral norm of the matrix of T T T . The operator norm is at most ∥ T ∥ H S , n \lVert T\rVert_{\mathrm{HS},n} ∥ T ∥ HS , n : a unit vector u u u can be completed to a ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n -orthonormal basis, and then ∥ T u ∥ n 2 \lVert T u\rVert_n^{2} ∥ T u ∥ n 2 is one of the terms in the sum that defines ∥ T ∥ H S , n 2 \lVert T\rVert_{\mathrm{HS},n}^{2} ∥ T ∥ HS , n 2 in Section 3. For a subspace V V V and v ∈ R n v\in\mathbb{R}^{n} v ∈ R n we set dist n ( v , V ) : = min w ∈ V ∥ v − w ∥ n \operatorname{dist}_n(v,V):=\min_{w\in V}\lVert v-w\rVert_n dist n ( v , V ) := min w ∈ V ∥ v − w ∥ n . For symmetric S , T S,T S , T we write S ⪯ T S\preceq T S ⪯ T if ⟨ S u , u ⟩ n ≤ ⟨ T u , u ⟩ n \langle Su,u\rangle_n\le\langle Tu,u\rangle_n ⟨ S u , u ⟩ n ≤ ⟨ T u , u ⟩ n for every u ∈ R n u\in\mathbb{R}^{n} u ∈ R n , that is, if T − S T-S T − S is positive semidefinite.
We express the bracket of Lemma 18 and Corollary 19 in terms of the comparison operators
A : = ( 1 − a ) m + 2 L + , B : = ( 1 + a ) m + 2 L − . (16) A:=(1-a)^{m+2}L^{+},\qquad B:=(1+a)^{m+2}L^{-}. \tag*{(16)} A := ( 1 − a ) m + 2 L + , B := ( 1 + a ) m + 2 L − . (16)
The eigenvalues of A A A are α k = ( 1 − a ) m + 2 λ k + \alpha_k=(1-a)^{m+2}\lambda_k^{+} α k = ( 1 − a ) m + 2 λ k + , and we define the spectral gap of A A A between its l l l -th and $(l+1)-st eigenvalues by
Γ l + : = α l + 1 − α l . (17) \Gamma_l^{+}:=\alpha_{l+1}-\alpha_l. \tag*{(17)} Γ l + := α l + 1 − α l . (17)
Let P l + P_l^{+} P l + be the orthogonal projection onto the subspace spanned by the first l l l eigenvectors of L + L^{+} L + , that is, by l l l orthonormal eigenvectors of L + L^{+} L + with eigenvalues λ 1 + , … , λ l + \lambda_1^{+},\ldots,\lambda_l^{+} λ 1 + , … , λ l + . Since ( 1 − a ) m + 2 > 0 (1-a)^{m+2}>0 ( 1 − a ) m + 2 > 0 , these are also eigenvectors of A A A , with eigenvalues α 1 , … , α l \alpha_1,\ldots,\alpha_l α 1 , … , α l . If Γ l + > 0 \Gamma_l^+>0 Γ l + > 0 , that is, λ l + < λ l + 1 + \lambda_l^+<\lambda_{l+1}^+ λ l + < λ l + 1 + , the range of P l + P_l^+ P l + is the span of all eigenvectors of L + L^+ L + with eigenvalue at most λ l + \lambda_l^+ λ l + : this span contains the range, and it has dimension l l l because only λ 1 + , … , λ l + \lambda_1^+,\ldots,\lambda_l^+ λ 1 + , … , λ l + are at most λ l + \lambda_l^+ λ l + . So P l + P_l^+ P l + does not depend on the choice of eigenvectors. The same argument shows that P l P_l P l does not depend on the choice when λ l < λ l + 1 \lambda_l<\lambda_{l+1} λ l < λ l + 1 , and P ~ l \widetilde P_l P l when λ ~ l < λ ~ l + 1 \widetilde\lambda_l<\widetilde\lambda_{l+1} λ l < λ l + 1 . If Γ l + = 0 \Gamma_l^+=0 Γ l + = 0 , then P l + P_l^+ P l + depends on the choice of eigenvectors, and we fix one.
The size of B − A B-A B − A on the range of P l + P_l^+ P l + is measured by
τ l : = max 0 ≠ u ∈ Range ( P l + ) ⟨ ( B − A ) u , u ⟩ n ∥ u ∥ n 2 . (18) \tau_l:=\max_{0\ne u\in\operatorname{Range}(P_l^+)}\frac{\langle(B-A)u,u\rangle_n}{\lVert u\rVert_n^2}. \tag*{(18)} τ l := 0 = u ∈ Range ( P l + ) max ∥ u ∥ n 2 ⟨( B − A ) u , u ⟩ n . (18)
For a given point cloud, τ l \tau_l τ l can be computed from L + L^+ L + and L − L^- L − , and Section 7 measures it. The bound τ l ≤ 1 2 K l a \tau_l\le\frac{1}{2}K_l a τ l ≤ 2 1 K l a of Lemma 26 below cannot be evaluated, because K l K_l K l contains the prefactor C C L u n i f C_{\mathrm{CL}}^{\mathrm{unif}} C CL unif of Lemma 17, which is not available in effective form.
Comparing spectral projectors
Lemma 26 proves A ⪯ L ⪯ B A\preceq L\preceq B A ⪯ L ⪯ B , A ⪯ L ~ ⪯ B A\preceq\widetilde L\preceq B A ⪯ L ⪯ B and, given E n \mathcal E_n E n , the bound τ l ≤ 1 2 K l a \tau_l\le\frac{1}{2}K_l a τ l ≤ 2 1 K l a . Lemma 28 shows that if A A A has a positive spectral gap between its l l l -th and ( l + 1 ) (l+1) ( l + 1 ) -st eigenvalues and B − A B-A B − A is small on the span of the first l l l eigenvectors of A A A , then, for every symmetric M M M with A ⪯ M ⪯ B A\preceq M\preceq B A ⪯ M ⪯ B , the orthogonal projection onto the subspace spanned by the first l l l eigenvectors of M M M is close to that of A A A in the Hilbert–Schmidt norm. Lemma 28 holds for arbitrary symmetric A ⪯ B A\preceq B A ⪯ B on ( R n , ⟨ ⋅ , ⋅ ⟩ n ) (\mathbb R^n,\langle\cdot,\cdot\rangle_n) ( R n , ⟨ ⋅ , ⋅ ⟩ n ) and every l l l , not only for the comparison operators (16) and the index fixed in Section 6. The proof of Lemma 28 uses Lemma 27, which expresses the Hilbert–Schmidt distance between two orthogonal projections of the same rank as a trace.
Lemma 26. The comparison operators satisfy
A ⪯ L ⪯ B , A ⪯ L ~ ⪯ B . (19) A\preceq L\preceq B,\qquad A\preceq\widetilde L\preceq B. \tag*{(19)} A ⪯ L ⪯ B , A ⪯ L ⪯ B . (19) Moreover, given E n \mathcal E_n E n , τ l ≤ 1 2 K l a \tau_l\le\frac{1}{2}K_l a τ l ≤ 2 1 K l a , with K l K_l K l as in (12).
Proof. Since b ε ( u ) = 2 ⟨ L u , u ⟩ n b_\varepsilon(u)=2\langle Lu,u\rangle_n b ε ( u ) = 2 ⟨ Lu , u ⟩ n , b ~ ε ( u ) = 2 ⟨ L ~ u , u ⟩ n \widetilde b_\varepsilon(u)=2\langle\widetilde Lu,u\rangle_n b ε ( u ) = 2 ⟨ L u , u ⟩ n and b ± ( u ) = 2 ⟨ L ± u , u ⟩ n b^\pm(u)=2\langle L^\pm u,u\rangle_n b ± ( u ) = 2 ⟨ L ± u , u ⟩ n , the definitions (16) give
( 1 − a ) m + 2 b + ( u ) = 2 ⟨ A u , u ⟩ n , ( 1 + a ) m + 2 b − ( u ) = 2 ⟨ B u , u ⟩ n . (1-a)^{m+2}b^+(u)=2\langle Au,u\rangle_n,\qquad(1+a)^{m+2}b^-(u)=2\langle Bu,u\rangle_n. ( 1 − a ) m + 2 b + ( u ) = 2 ⟨ A u , u ⟩ n , ( 1 + a ) m + 2 b − ( u ) = 2 ⟨ B u , u ⟩ n .
Hence Lemma 18 gives (19), and 2 ⟨ ( B − A ) u , u ⟩ n = ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) 2\langle(B-A)u,u\rangle_n=(1+a)^{m+2}b^-(u)-(1-a)^{m+2}b^+(u) 2 ⟨( B − A ) u , u ⟩ n = ( 1 + a ) m + 2 b − ( u ) − ( 1 − a ) m + 2 b + ( u ) . For the bound on τ l \tau_l τ l , Lemma 24 with S + = Range ( P l + ) S^+=\operatorname{Range}(P_l^+) S + = Range ( P l + ) therefore gives, given E n \mathcal E_n E n ,
2 ⟨ ( B − A ) u , u ⟩ n ≤ K l a ∥ u ∥ n 2 for every u ∈ Range ( P l + ) . 2\langle(B-A)u,u\rangle_n\le K_l a\lVert u\rVert_n^2\qquad\text{for every }u\in\operatorname{Range}(P_l^+). 2 ⟨( B − A ) u , u ⟩ n ≤ K l a ∥ u ∥ n 2 for every u ∈ Range ( P l + ) .
Dividing by 2 ∥ u ∥ n 2 2\lVert u\rVert_n^2 2 ∥ u ∥ n 2 for u ≠ 0 u\ne0 u = 0 and taking the maximum over u u u gives τ l ≤ 1 2 K l a \tau_l\le\frac{1}{2}K_l a τ l ≤ 2 1 K l a .
Lemma 27. Let P , Q P,Q P , Q be ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n -orthogonal projections of the same rank l l l . Then
∥ P − Q ∥ H S , n 2 = 2 Tr ( ( I − P ) Q ) = 2 ( l − Tr ( P Q ) ) , (20) \lVert P-Q\rVert_{\mathrm{HS},n}^2=2\operatorname{Tr}((I-P)Q)=2(l-\operatorname{Tr}(PQ)), \tag*{(20)} ∥ P − Q ∥ HS , n 2 = 2 Tr (( I − P ) Q ) = 2 ( l − Tr ( P Q )) , (20) and ∥ P − Q ∥ H S , n 2 ≤ 2 l \lVert P-Q\rVert_{\mathrm{HS},n}^2\le2l ∥ P − Q ∥ HS , n 2 ≤ 2 l .
Proof. Since P , Q P,Q P , Q are self-adjoint idempotents of trace l l l ,
∥ P − Q ∥ H S , n 2 = Tr ( ( P − Q ) 2 ) = Tr ( P ) + Tr ( Q ) − 2 Tr ( P Q ) = 2 ( l − Tr ( P Q ) ) , \lVert P-Q\rVert_{\mathrm{HS},n}^2=\operatorname{Tr}((P-Q)^2)=\operatorname{Tr}(P)+\operatorname{Tr}(Q)-2\operatorname{Tr}(PQ)=2(l-\operatorname{Tr}(PQ)), ∥ P − Q ∥ HS , n 2 = Tr (( P − Q ) 2 ) = Tr ( P ) + Tr ( Q ) − 2 Tr ( P Q ) = 2 ( l − Tr ( P Q )) ,
and l − Tr ( P Q ) = Tr ( ( I − P ) Q ) l-\operatorname{Tr}(PQ)=\operatorname{Tr}((I-P)Q) l − Tr ( P Q ) = Tr (( I − P ) Q ) because Tr ( Q ) = l \operatorname{Tr}(Q)=l Tr ( Q ) = l . By cyclicity of the trace and P 2 = P P^2=P P 2 = P , Q 2 = Q Q^2=Q Q 2 = Q , we have Tr ( P Q ) = Tr ( ( Q P ) ∗ Q P ) = ∥ Q P ∥ H S , n 2 ≥ 0 \operatorname{Tr}(PQ)=\operatorname{Tr}((QP)^*QP)=\lVert QP\rVert_{\mathrm{HS},n}^2\ge0 Tr ( P Q ) = Tr (( QP ) ∗ QP ) = ∥ QP ∥ HS , n 2 ≥ 0 , which gives the bound.
Lemma 28. Let A ⪯ B A \preceq B A ⪯ B be symmetric operators on ( R n , ⟨ ⋅ , ⋅ ⟩ n ) (\mathbb{R}^{n},\langle\cdot,\cdot\rangle_{n}) ( R n , ⟨ ⋅ , ⋅ ⟩ n ) , let α 1 ≤ ⋯ ≤ α n \alpha_{1} \le\cdots\le\alpha_{n} α 1 ≤ ⋯ ≤ α n be the eigenvalues of A A A , and let τ ≥ 0 \tau\ge0 τ ≥ 0 . For any integer l l l with 1 ≤ l < n 1 \le l < n 1 ≤ l < n , let P P P be the orthogonal projection onto the subspace spanned by l l l orthonormal eigenvectors of A A A with eigenvalues α 1 , … , α l \alpha_{1},\ldots,\alpha_{l} α 1 , … , α l , and assume
Γ : = α l + 1 − α l > 0 and ⟨ ( B − A ) u , u ⟩ n ≤ τ ∥ u ∥ n 2 for all u ∈ Range ( P ) . \Gamma:= \alpha_{l+1}-\alpha_{l}>0 \qquad\text{and} \qquad\langle(B-A)u,u\rangle_{n}\leq\tau\lVert u\rVert_{n}^{2} \quad\text{for all }u\in\operatorname{Range}(P). Γ := α l + 1 − α l > 0 and ⟨( B − A ) u , u ⟩ n ≤ τ ∥ u ∥ n 2 for all u ∈ Range ( P ) . Then, for every symmetric M M M with A ⪯ M ⪯ B A\preceq M\preceq B A ⪯ M ⪯ B , with eigenvalues μ 1 ≤ ⋯ ≤ μ n \mu_{1}\leq\cdots\leq\mu_{n} μ 1 ≤ ⋯ ≤ μ n , and every orthogonal projection Q Q Q onto the subspace spanned by l l l orthonormal eigenvectors of M M M with eigenvalues μ 1 , … , μ l \mu_{1},\ldots,\mu_{l} μ 1 , … , μ l ,
∥ Q − P ∥ H S , n 2 ≤ 2 l τ Γ . (21) \lVert Q-P\rVert_{\mathrm{HS},n}^{2}\leq\frac{2l\tau}{\Gamma}. \tag*{(21)} ∥ Q − P ∥ HS , n 2 ≤ Γ 2 l τ . (21)
Proof. Let u 1 , … , u n u^{1},\ldots,u^{n} u 1 , … , u n be an orthonormal eigenbasis of A A A with A u k = α k u k Au^{k}=\alpha_{k}u^{k} A u k = α k u k and Range ( P ) = Span { u 1 , … , u l } \operatorname{Range}(P)=\operatorname{Span}\{u^{1},\ldots,u^{l}\} Range ( P ) = Span { u 1 , … , u l } , put r k : = ⟨ Q u k , u k ⟩ n = ∥ Q u k ∥ n 2 ∈ [ 0 , 1 ] r_{k}:=\langle Qu^{k},u^{k}\rangle_{n}=\lVert Qu^{k}\rVert_{n}^{2}\in[0,1] r k := ⟨ Q u k , u k ⟩ n = ∥ Q u k ∥ n 2 ∈ [ 0 , 1 ] , and set
s : = Tr ( ( I − P ) Q ) = ∑ k > l r k = ∑ k ≤ l ( 1 − r k ) . s:=\operatorname{Tr}((I-P)Q)=\sum_{k>l}r_{k}=\sum_{k\leq l}(1-r_{k}). s := Tr (( I − P ) Q ) = k > l ∑ r k = k ≤ l ∑ ( 1 − r k ) .
The first equality follows by evaluating the trace in the basis { u k } \{u^{k}\} { u k } , and the last because ∑ k r k = Tr ( Q ) = l \sum_{k}r_{k}=\operatorname{Tr}(Q)=l ∑ k r k = Tr ( Q ) = l . We bound Tr ( A Q ) − Tr ( A P ) \operatorname{Tr}(AQ)-\operatorname{Tr}(AP) Tr ( A Q ) − Tr ( A P ) from below by Γ s \Gamma s Γ s and from above by l τ l\tau l τ .
Lower bound. Expanding both traces in the basis { u k } \{u^{k}\} { u k } and using α k ≥ α l + 1 \alpha_{k}\geq\alpha_{l+1} α k ≥ α l + 1 for k > l k>l k > l , α k ≤ α l \alpha_{k}\leq\alpha_{l} α k ≤ α l for k ≤ l k\leq l k ≤ l , and r k , 1 − r k ≥ 0 r_{k},1-r_{k}\geq0 r k , 1 − r k ≥ 0 ,
Tr ( A Q ) − Tr ( A P ) = ∑ k > l α k r k − ∑ k ≤ l α k ( 1 − r k ) ≥ α l + 1 s − α l s = Γ s . \operatorname{Tr}(AQ)-\operatorname{Tr}(AP)=\sum_{k>l}\alpha_{k}r_{k}-\sum_{k\leq l}\alpha_{k}(1-r_{k})\geq\alpha_{l+1}s-\alpha_{l}s=\Gamma s. Tr ( A Q ) − Tr ( A P ) = k > l ∑ α k r k − k ≤ l ∑ α k ( 1 − r k ) ≥ α l + 1 s − α l s = Γ s .
Upper bound. Let S ⪯ T S\preceq T S ⪯ T be symmetric and R R R an orthogonal projection of rank l l l . Evaluating both traces by (15) in an orthonormal basis whose first l l l vectors v 1 , … , v l v_{1},\ldots,v_{l} v 1 , … , v l span Range ( R ) \operatorname{Range}(R) Range ( R ) gives
Tr ( S R ) = ∑ i = 1 l ⟨ S v i , v i ⟩ n ≤ ∑ i = 1 l ⟨ T v i , v i ⟩ n = Tr ( T R ) . \operatorname{Tr}(SR)=\sum_{i=1}^{l}\langle Sv_{i},v_{i}\rangle_{n}\leq\sum_{i=1}^{l}\langle Tv_{i},v_{i}\rangle_{n}=\operatorname{Tr}(TR). Tr ( S R ) = i = 1 ∑ l ⟨ S v i , v i ⟩ n ≤ i = 1 ∑ l ⟨ T v i , v i ⟩ n = Tr ( T R ) .
By the Ky Fan minimum principle [23 Corollary 4.3.39, p. 248] , Tr ( M Q ) \operatorname{Tr}(MQ) Tr ( M Q ) is the minimum of Tr ( M R ) \operatorname{Tr}(MR) Tr ( M R ) over rank-l l l orthogonal projections R R R , so it is at most Tr ( M P ) \operatorname{Tr}(MP) Tr ( M P ) . Combining this with the inequality above for R = Q R=Q R = Q and for R = P R=P R = P gives
Tr ( A Q ) ≤ Tr ( M Q ) ≤ Tr ( M P ) ≤ Tr ( B P ) . \operatorname{Tr}(AQ)\leq\operatorname{Tr}(MQ)\leq\operatorname{Tr}(MP)\leq\operatorname{Tr}(BP). Tr ( A Q ) ≤ Tr ( M Q ) ≤ Tr ( M P ) ≤ Tr ( B P ) .
Subtracting Tr ( A P ) \operatorname{Tr}(AP) Tr ( A P ) and evaluating the resulting trace in the basis { u k } \{u^{k}\} { u k } ,
Tr ( A Q ) − Tr ( A P ) ≤ Tr ( ( B − A ) P ) = ∑ k = 1 l ⟨ ( B − A ) u k , u k ⟩ n ≤ l τ . \operatorname{Tr}(AQ)-\operatorname{Tr}(AP)\leq\operatorname{Tr}((B-A)P)=\sum_{k=1}^{l}\langle(B-A)u^{k},u^{k}\rangle_{n}\leq l\tau. Tr ( A Q ) − Tr ( A P ) ≤ Tr (( B − A ) P ) = k = 1 ∑ l ⟨( B − A ) u k , u k ⟩ n ≤ l τ .
The two bounds give Γ s ≤ l τ \Gamma s\leq l\tau Γ s ≤ l τ , and by (20) , ∥ Q − P ∥ H S , n 2 = 2 s ≤ 2 l τ / Γ \lVert Q-P\rVert_{\mathrm{HS},n}^{2}=2s\leq2l\tau/\Gamma ∥ Q − P ∥ HS , n 2 = 2 s ≤ 2 l τ /Γ . □ \square □
Proof of Theorem 3
Proposition 29 applies Lemma 28 to the comparison operators, with M = L M=L M = L and with M = L ~ M=\widetilde{L} M = L .
Proposition 29. For every point cloud with Γ l + > 0 \Gamma_{l}^{+}>0 Γ l + > 0 and every choice of P l P_{l} P l and P ~ l \widetilde{P}_{l} P l ,
max { ∥ P l − P l + ∥ H S , n , ∥ P ~ l − P l + ∥ H S , n } ≤ 2 l η l Γ l + . (22) \max\left\{\lVert P_{l}-P_{l}^{+}\rVert_{\mathrm{HS},n},\lVert\widetilde{P}_{l}-P_{l}^{+}\rVert_{\mathrm{HS},n}\right\}\leq\sqrt{\frac{2l\eta_{l}}{\Gamma_{l}^{+}}}. \tag*{(22)} max { ∥ P l − P l + ∥ HS , n , ∥ P l − P l + ∥ HS , n } ≤ Γ l + 2 l η l . (22)
Proof. By Lemma 26, A ⪯ L ⪯ B A \preceq L \preceq B A ⪯ L ⪯ B and A ⪯ L ~ ⪯ B A \preceq\widetilde{L} \preceq B A ⪯ L ⪯ B . In particular A ⪯ B A \preceq B A ⪯ B , so τ l ≥ 0 \tau_l \ge0 τ l ≥ 0 , and by (18), ⟨ ( B − A ) u , u ⟩ n ≤ τ l ∥ u ∥ n 2 \langle(B-A)u,u\rangle_n \le\tau_l\|u\|_n^2 ⟨( B − A ) u , u ⟩ n ≤ τ l ∥ u ∥ n 2 for every u ∈ Range ( P l + ) u \in\operatorname{Range}(P_l^+) u ∈ Range ( P l + ) . Since Γ l + > 0 \Gamma_l^+ > 0 Γ l + > 0 , Lemma 28 applies to the comparison operators (16) with P = P l + P=P_l^+ P = P l + , Γ = Γ l + \Gamma=\Gamma_l^+ Γ = Γ l + and τ = τ l \tau=\tau_l τ = τ l . With M = L M=L M = L and Q = P l Q=P_l Q = P l , and with M = L ~ M=\widetilde{L} M = L and Q = P ~ l Q=\widetilde{P}_l Q = P l , Lemma 28 gives, for every choice of P l P_l P l and P ~ l \widetilde{P}_l P l ,
∥ P l − P l + ∥ H S , n 2 ≤ 2 l τ l Γ l + , ∥ P ~ l − P l + ∥ H S , n 2 ≤ 2 l τ l Γ l + . (23) \|P_l-P_l^+\|_{\mathrm{HS},n}^2 \le\frac{2l\tau_l}{\Gamma_l^+}, \qquad\|\widetilde{P}_l-P_l^+\|_{\mathrm{HS},n}^2 \le\frac{2l\tau_l}{\Gamma_l^+}. \tag*{(23)} ∥ P l − P l + ∥ HS , n 2 ≤ Γ l + 2 l τ l , ∥ P l − P l + ∥ HS , n 2 ≤ Γ l + 2 l τ l . (23)
Taking square roots in (23) proves (22).
Next, Lemma 30 gives a lower bound on Γ l + \Gamma_l^+ Γ l + . The proof of Lemma 30 uses the results of Section 5 at the two indices l l l and l + 1 l+1 l + 1 , so it needs the event of Definition 21 for both indices at once. Since the event E n \mathcal{E}_n E n of Definition 21 depends on the index l l l , we write E n ( k ) \mathcal{E}_n^{(k)} E n ( k ) for the event of Definition 21 with l l l replaced by k k k , so that E n = E n ( l ) \mathcal{E}_n=\mathcal{E}_n^{(l)} E n = E n ( l ) . The event E n ( k ) \mathcal{E}_n^{(k)} E n ( k ) is defined once the bandwidth assumptions made before Definition 21 hold with l l l replaced by k k k , which under (6) is the case for n n n large enough. Set
E n p r o j : = E n ( l ) ∩ E n ( l + 1 ) . (24) \mathcal{E}_n^{\mathrm{proj}} := \mathcal{E}_n^{(l)} \cap\mathcal{E}_n^{(l+1)}. \tag*{(24)} E n proj := E n ( l ) ∩ E n ( l + 1 ) . (24)
Given E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj , the results of Section 5 hold at both indices.
Lemma 30. Assume that γ l M > 0 \gamma_l^{\mathcal{M}} > 0 γ l M > 0 and
( C l ′ + C l + 1 ′ ) ε + 1 2 K l + 1 a ≤ 1 2 σ η γ l M . (25) (C'_l+C'_{l+1})\varepsilon+\frac{1}{2}K_{l+1}a \le\frac{1}{2}\sigma_\eta\gamma_l^{\mathcal{M}}. \tag*{(25)} ( C l ′ + C l + 1 ′ ) ε + 2 1 K l + 1 a ≤ 2 1 σ η γ l M . (25) Then, given E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj , the spectral gap Γ l + \Gamma_l^+ Γ l + of A A A satisfies
Γ l + ≥ 1 2 σ η γ l M > 0. \Gamma_l^+ \ge\frac{1}{2}\sigma_\eta\gamma_l^{\mathcal{M}} > 0. Γ l + ≥ 2 1 σ η γ l M > 0.
Proof. The lower bound of Corollary 19 reads α l ≤ λ l \alpha_l \le\lambda_l α l ≤ λ l . The upper bound of Corollary 19 at index l + 1 l+1 l + 1 , followed by Proposition 25 at index l + 1 l+1 l + 1 (which holds given E n ( l + 1 ) \mathcal{E}_n^{(l+1)} E n ( l + 1 ) ), gives
λ l + 1 ≤ ( 1 + a ) m + 2 λ ~ l + 1 ≤ α l + 1 + 1 2 K l + 1 a . \lambda_{l+1} \le(1+a)^{m+2}\widetilde{\lambda}_{l+1} \le\alpha_{l+1}+\frac{1}{2}K_{l+1}a. λ l + 1 ≤ ( 1 + a ) m + 2 λ l + 1 ≤ α l + 1 + 2 1 K l + 1 a .
Moreover, (7) at index l + 1 l+1 l + 1 , given E n ( l + 1 ) \mathcal{E}_n^{(l+1)} E n ( l + 1 ) , and at index l l l , given E n ( l ) \mathcal{E}_n^{(l)} E n ( l ) , gives λ l + 1 − λ l ≥ σ η γ l M − ( C l ′ + C l + 1 ′ ) ε \lambda_{l+1}-\lambda_l \ge\sigma_\eta\gamma_l^{\mathcal{M}}-(C'_l+C'_{l+1})\varepsilon λ l + 1 − λ l ≥ σ η γ l M − ( C l ′ + C l + 1 ′ ) ε . Combining these bounds with (25),
Γ l + = α l + 1 − α l ≥ λ l + 1 − λ l − 1 2 K l + 1 a ≥ σ η γ l M − ( C l ′ + C l + 1 ′ ) ε − 1 2 K l + 1 a ≥ 1 2 σ η γ l M . \Gamma_l^+ = \alpha_{l+1}-\alpha_l \ge\lambda_{l+1}-\lambda_l-\frac{1}{2}K_{l+1}a \ge\sigma_\eta\gamma_l^{\mathcal{M}}-(C'_l+C'_{l+1})\varepsilon-\frac{1}{2}K_{l+1}a \ge\frac{1}{2}\sigma_\eta\gamma_l^{\mathcal{M}}. Γ l + = α l + 1 − α l ≥ λ l + 1 − λ l − 2 1 K l + 1 a ≥ σ η γ l M − ( C l ′ + C l + 1 ′ ) ε − 2 1 K l + 1 a ≥ 2 1 σ η γ l M .
Proof of Theorem 3. Since γ l M > 0 \gamma_l^{\mathcal{M}}>0 γ l M > 0 , condition (25) holds for n n n large enough and δ / ε \delta/\varepsilon δ / ε small enough. Lemma 30 then gives Γ l + ≥ 1 2 σ η γ l M > 0 \Gamma_l^+ \ge\frac{1}{2}\sigma_\eta\gamma_l^{\mathcal{M}}>0 Γ l + ≥ 2 1 σ η γ l M > 0 given E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj . The event E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj is contained in E n \mathcal{E}_n E n , so given E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj , Proposition 29, the triangle inequality, the bound τ l ≤ 1 2 K l a \tau_l \le\frac{1}{2}K_l a τ l ≤ 2 1 K l a of Lemma 26 and a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε yield
∥ P l − P ~ l ∥ H S , n ≤ ∥ P l − P l + ∥ H S , n + ∥ P ~ l − P l + ∥ H S , n \|P_l-\widetilde{P}_l\|_{\mathrm{HS},n} \le\|P_l-P_l^+\|_{\mathrm{HS},n}+\|\widetilde{P}_l-P_l^+\|_{\mathrm{HS},n} ∥ P l − P l ∥ HS , n ≤ ∥ P l − P l + ∥ HS , n + ∥ P l − P l + ∥ HS , n
≤ 2 2 l τ l Γ l + ≤ 2 2 K l l a σ η γ l M = D l δ ε . (26) \le2\sqrt{\frac{2l\tau_l}{\Gamma_l^+}} \le2\sqrt{\frac{2K_l l a}{\sigma_\eta\gamma_l^{\mathcal{M}}}}=D_l\sqrt{\frac{\delta}{\varepsilon}}. \tag*{(26)} ≤ 2 Γ l + 2 l τ l ≤ 2 σ η γ l M 2 K l l a = D l ε δ . (26)
Proposition 22 at the indices l l l and l + 1 l+1 l + 1 and a union bound give P n ( E n p r o j ) ≥ 1 − O ( n − p ) \mathbb{P}_n(\mathcal{E}_n^{\mathrm{proj}})\ge1-O(n^{-p}) P n ( E n proj ) ≥ 1 − O ( n − p ) once C ε C_\varepsilon C ε is large enough. This proves the theorem.
Since ∥ T ∥ o p , n ≤ ∥ T ∥ H S , n \|T\|_{\mathrm{op},n}\le\|T\|_{\mathrm{HS},n} ∥ T ∥ op , n ≤ ∥ T ∥ HS , n , the bound of Theorem 3 also holds in the operator norm.
Proof of Theorem 4
Throughout Section 6.4, l ≥ 2 l \ge2 l ≥ 2 and p ≥ 1 p \ge1 p ≥ 1 are as in Theorem 4, and l − , … , l + l_{-},\ldots,l_{+} l − , … , l + is the multiplicity block of λ : = λ l M \lambda:= \lambda_{l}^{M} λ := λ l M . Since λ 1 M = 0 \lambda_{1}^{M}=0 λ 1 M = 0 is simple and l ≥ 2 l \ge2 l ≥ 2 , λ > 0 \lambda> 0 λ > 0 . For L L L and L ~ \widetilde{L} L , choose orthonormal eigenbases ordered by increasing eigenvalue and let P k P_{k} P k and P ~ k \widetilde{P}_{k} P k be the orthogonal projections onto the spans of their first k k k vectors, so that Range ( P j ) ⊆ Range ( P k ) \operatorname{Range}(P_{j}) \subseteq\operatorname{Range}(P_{k}) Range ( P j ) ⊆ Range ( P k ) and Range ( P ~ j ) ⊆ Range ( P ~ k ) \operatorname{Range}(\widetilde{P}_{j}) \subseteq\operatorname{Range}(\widetilde{P}_{k}) Range ( P j ) ⊆ Range ( P k ) for j ≤ k j \le k j ≤ k . Define
P l − : l + : = P l + − P l − − 1 , P ~ l − : l + : = P ~ l + − P ~ l − − 1 . P_{l_{-}:l_{+}} := P_{l_{+}} - P_{l_{-}-1}, \qquad\widetilde{P}_{l_{-}:l_{+}} := \widetilde{P}_{l_{+}} - \widetilde{P}_{l_{-}-1}. P l − : l + := P l + − P l − − 1 , P l − : l + := P l + − P l − − 1 .
These are the orthogonal projections onto the subspaces spanned by the basis vectors with indices l − , … , l + l_{-},\ldots,l_{+} l − , … , l + . We define S n : = Range ( P l − : l + ) S_{n} := \operatorname{Range}(P_{l_{-}:l_{+}}) S n := Range ( P l − : l + ) , the span of the eigenvectors of L L L with these indices; E λ : = ker ( Δ ρ − λ I ) E_{\lambda} := \ker(\Delta_{\rho} - \lambda I) E λ := ker ( Δ ρ − λ I ) , the eigenspace of Δ ρ \Delta_{\rho} Δ ρ for λ \lambda λ , of dimension r : = l + − l − + 1 r := l_{+} - l_{-} + 1 r := l + − l − + 1 ; and J l : = { l − − 1 , l − , l + , l + + 1 } J_{l} := \{l_{-} - 1, l_{-}, l_{+}, l_{+} + 1\} J l := { l − − 1 , l − , l + , l + + 1 } , the set of indices at which we use the events E n ( j ) \mathcal{E}_{n}^{(j)} E n ( j ) .
Corollary 31 shows that every unit vector in Range ( P ~ l − : l + ) \operatorname{Range}(\widetilde{P}_{l_{-}:l_{+}}) Range ( P l − : l + ) is close to S n S_{n} S n , and Lemma 32 shows that every unit vector in S n S_{n} S n is close to a unit eigenfunction in E λ E_{\lambda} E λ . The statement of [7 Theorem 2.7(2)] gives the second property only for eigenvectors of L L L . The proof of Theorem 4 needs this property for a vector v ∈ S n v \in S_{n} v ∈ S n that is a linear combination of eigenvectors of L L L for λ l − , … , λ l + \lambda_{l_{-}},\ldots,\lambda_{l_{+}} λ l − , … , λ l + . These eigenvalues are close to each other but in general distinct, so v v v need not be an eigenvector of L L L . Lemma 32 therefore uses the proof of [7 Theorem 2.7(2)] instead of its statement, and assumes the hypotheses of that proof, (A1)–(A4) and (S2N _N N ) of Appendix C with N = l + + 1 N = l_{+} + 1 N = l + + 1 .
Corollary 31. Suppose that (25) holds at the indices l − − 1 l_{-} - 1 l − − 1 and l + l_{+} l + . Then, given ⋂ j ∈ J l E n ( j ) \bigcap_{j\in J_{l}}\mathcal{E}_{n}^{(j)} ⋂ j ∈ J l E n ( j ) ,
∥ P l − : l + − P ~ l − : l + ∥ H S , n ≤ C l − , l + δ ε , C l − , l + : = D l + + D l − − 1 . (27) \left\|P_{l_{-}:l_{+}}-\widetilde{P}_{l_{-}:l_{+}}\right\|_{\mathrm{HS},n} \le C_{l_{-},l_{+}}\sqrt{\frac{\delta}{\varepsilon}}, \qquad C_{l_{-},l_{+}} := D_{l_{+}} + D_{l_{-}-1}. \tag*{(27)} P l − : l + − P l − : l + HS , n ≤ C l − , l + ε δ , C l − , l + := D l + + D l − − 1 . (27) Consequently every unit u ~ ∈ Range ( P ~ l − : l + ) \widetilde{u}\in\operatorname{Range}(\widetilde{P}_{l_{-}:l_{+}}) u ∈ Range ( P l − : l + ) satisfies dist n ( u ~ , S n ) ≤ C l − , l + δ / ε \operatorname{dist}_{n}(\widetilde{u},S_{n})\le C_{l_{-},l_{+}}\sqrt{\delta/\varepsilon} dist n ( u , S n ) ≤ C l − , l + δ / ε .
Proof. The spectral gaps γ l − − 1 M \gamma_{l_{-}-1}^{M} γ l − − 1 M and γ l + M \gamma_{l_{+}}^{M} γ l + M are strictly positive, since l − l_{-} l − and l + l_{+} l + are the ends of the multiplicity block, and ⋂ j ∈ J l E n ( j ) \bigcap_{j\in J_{l}}\mathcal{E}_{n}^{(j)} ⋂ j ∈ J l E n ( j ) is contained in E n ( k ) ∩ E n ( k + 1 ) \mathcal{E}_{n}^{(k)}\cap\mathcal{E}_{n}^{(k+1)} E n ( k ) ∩ E n ( k + 1 ) for k = l + k=l_{+} k = l + and for k = l − − 1 k=l_{-}-1 k = l − − 1 . Hence (26) holds with l l l replaced by l + l_{+} l + and by l − − 1 l_{-}-1 l − − 1 , and (27) follows from the triangle inequality
∥ P l − : l + − P ~ l − : l + ∥ H S , n ≤ ∥ P l + − P ~ l + ∥ H S , n + ∥ P l − − 1 − P ~ l − − 1 ∥ H S , n . \left\|P_{l_{-}:l_{+}}-\widetilde{P}_{l_{-}:l_{+}}\right\|_{\mathrm{HS},n}
\le\left\|P_{l_{+}}-\widetilde{P}_{l_{+}}\right\|_{\mathrm{HS},n}
+\left\|P_{l_{-}-1}-\widetilde{P}_{l_{-}-1}\right\|_{\mathrm{HS},n}. P l − : l + − P l − : l + HS , n ≤ P l + − P l + HS , n + P l − − 1 − P l − − 1 HS , n .
For the bound on dist n ( u ~ , S n ) \operatorname{dist}_{n}(\widetilde{u},S_{n}) dist n ( u , S n ) , the identity P ~ l − : l + u ~ = u ~ \widetilde{P}_{l_{-}:l_{+}}\widetilde{u}=\widetilde{u} P l − : l + u = u gives
dist n ( u ~ , S n ) = ∥ ( I − P l − : l + ) u ~ ∥ n = ∥ ( P ~ l − : l + − P l − : l + ) u ~ ∥ n ≤ ∥ P l − : l + − P ~ l − : l + ∥ o p , n ≤ ∥ P l − : l + − P ~ l − : l + ∥ H S , n . \operatorname{dist}_{n}(\widetilde{u},S_{n})
=\left\|(I-P_{l_{-}:l_{+}})\widetilde{u}\right\|_{n}
=\left\|(\widetilde{P}_{l_{-}:l_{+}}-P_{l_{-}:l_{+}})\widetilde{u}\right\|_{n}
\le\left\|P_{l_{-}:l_{+}}-\widetilde{P}_{l_{-}:l_{+}}\right\|_{\mathrm{op},n}
\le\left\|P_{l_{-}:l_{+}}-\widetilde{P}_{l_{-}:l_{+}}\right\|_{\mathrm{HS},n}. dist n ( u , S n ) = ∥ ( I − P l − : l + ) u ∥ n = ( P l − : l + − P l − : l + ) u n ≤ P l − : l + − P l − : l + op , n ≤ P l − : l + − P l − : l + HS , n .
Lemma 32. Let N : = l + + 1 N := l_{+} + 1 N := l + + 1 , and assume that ε \varepsilon ε , θ \theta θ and δ ~ e \widetilde{\delta}_{e} δ e satisfy (A1)–(A4) and (S2N _N N ) of Appendix C. Then there exist an event V n , λ \mathcal{V}_{n,\lambda} V n , λ and constants C λ , c λ > 0 C_{\lambda},c_{\lambda}>0 C λ , c λ > 0 , depending only on λ \lambda λ , M M M , ρ \rho ρ and η \eta η , such that, given V n , λ \mathcal{V}_{n,\lambda} V n , λ , for every v ∈ S n v\in S_{n} v ∈ S n with ∥ v ∥ n = 1 \|v\|_{n}=1 ∥ v ∥ n = 1 there is an f ∈ E λ f\in E_{\lambda} f ∈ E λ with
∥ f ∥ L 2 ( μ ) = 1 , ∥ v − f ∥ X n ≤ r C λ ε . (28) \|f\|_{L^{2}(\mu)}=1,\qquad\|v-f\|_{X_{n}}\le\sqrt{r}\,C_{\lambda}\varepsilon. \tag*{(28)} ∥ f ∥ L 2 ( μ ) = 1 , ∥ v − f ∥ X n ≤ r C λ ε . (28) The failure probability of V n , λ \mathcal{V}_{n,\lambda} V n , λ is at most C λ n exp ( − c λ n θ 2 δ ~ e 2 m ) + C λ n exp ( − c λ n ε m + 4 ) C_{\lambda}n\exp(-c_{\lambda}n\theta^{2}\widetilde{\delta}_{e}^{2m})+C_{\lambda}n\exp(-c_{\lambda}n\varepsilon^{m+4}) C λ n exp ( − c λ n θ 2 δ e 2 m ) + C λ n exp ( − c λ n ε m + 4 ) , as in [7 Theorem 2.7(2)] .
Proof. We use the proof of [7 Theorem 2.7(2)] , which treats an eigenvalue λ > 0 \lambda>0 λ > 0 of Δ ρ \Delta_{\rho} Δ ρ , as here. In its notation the block of λ \lambda λ has the indices i + 1 , … , i + k i+1,\ldots,i+k i + 1 , … , i + k , so that i = l − − 1 i=l_{-}-1 i = l − − 1 and k = r k=r k = r . That proof assumes (S2N) for some N ≥ i + k N \ge i+k N ≥ i + k and uses the resulting bound [7 (4.13)] also at the index i + k + 1 = l + + 1 i+k+1=l_++1 i + k + 1 = l + + 1 , which is why we take N = l + + 1 N=l_++1 N = l + + 1 . Under (A1)–(A4) and (S2N), [7 Theorem 2.4(1)] gives, with probability at least 1 − C n exp ( − c n θ 2 δ ε ) 1-Cn\exp(-cn\theta^2\delta_{\varepsilon}) 1 − C n exp ( − c n θ 2 δ ε ) ,
∣ λ j − σ η λ j M ∣ ≤ 1 2 σ η g j , j = 1 , … , N . |\lambda_j-\sigma_\eta\lambda_j^{\mathcal{M}}| \le\frac{1}{2}\sigma_\eta g_j,\qquad j=1,\ldots,N. ∣ λ j − σ η λ j M ∣ ≤ 2 1 σ η g j , j = 1 , … , N .
This is (4.13) of that proof, with the factor σ η \sigma_\eta σ η discussed in Appendix C. Since g l − − 1 ≤ γ l − − 1 M g_{l_--1}\le\gamma_{l_--1}^{\mathcal{M}} g l − − 1 ≤ γ l − − 1 M and g l + + 1 ≤ γ l + M g_{l_++1}\le\gamma_{l_+}^{\mathcal{M}} g l + + 1 ≤ γ l + M , (4.13) places λ l − − 1 \lambda_{l_--1} λ l − − 1 and λ l + + 1 \lambda_{l_++1} λ l + + 1 , and hence every eigenvalue of L L L whose index lies outside the block, at distance at least 1 2 σ η g l \frac{1}{2}\sigma_\eta g_l 2 1 σ η g l from σ η λ \sigma_\eta\lambda σ η λ . Here g l g_l g l , defined in Appendix C, equals min { γ l − − 1 M , γ l + M } \min\{\gamma_{l_--1}^{\mathcal{M}},\gamma_{l_+}^{\mathcal{M}}\} min { γ l − − 1 M , γ l + M } because l − ≤ l ≤ l + l_-\le l\le l_+ l − ≤ l ≤ l + .
Let f 1 , … , f r f_1,\ldots,f_r f 1 , … , f r be an L 2 ( μ ) L^2(\mu) L 2 ( μ ) -orthonormal basis of E λ E_\lambda E λ . Using this separation, [7 (4.14)] gives, for each j j j ,
dist n ( f j ∣ X n , S n ) ≤ 2 σ η g l ∥ L f j ∣ X n − σ η ( Δ ρ f j ) ∣ X n ∥ n . \operatorname{dist}_n\left(f_j|_{X_n},S_n\right)\le\frac{2}{\sigma_\eta g_l}\left\|Lf_j|_{X_n}-\sigma_\eta(\Delta_\rho f_j)|_{X_n}\right\|_n. dist n ( f j ∣ X n , S n ) ≤ σ η g l 2 ∥ L f j ∣ X n − σ η ( Δ ρ f j ) ∣ X n ∥ n .
Pointwise consistency [7 Theorem 3.3] , the approximation of σ η Δ ρ f \sigma_\eta\Delta_\rho f σ η Δ ρ f by L f ∣ X n Lf|_{X_n} L f ∣ X n described in § 2 , applies to the f j f_j f j , which are C 3 C^3 C 3 [7 Section 2.3] . It bounds the norms on the right-hand side by a constant multiple of ε \varepsilon ε , with probability at least 1 − 2 r n exp ( − c n ε m + 4 ) 1-2rn\exp(-cn\varepsilon^{m+4}) 1 − 2 r n exp ( − c n ε m + 4 ) . The proof of [7 Theorem 2.7(2)] then combines this bound with (4.14) and concludes that, with probability at least 1 − C n exp ( − c n θ 2 δ ε m ) − 2 r n exp ( − c n ε m + 4 ) 1-Cn\exp(-cn\theta^2\delta_{\varepsilon}^{m})-2rn\exp(-cn\varepsilon^{m+4}) 1 − C n exp ( − c n θ 2 δ ε m ) − 2 r n exp ( − c n ε m + 4 ) , there is an ⟨ ⋅ , ⋅ ⟩ n \langle\cdot,\cdot\rangle_n ⟨ ⋅ , ⋅ ⟩ n -orthonormal basis v 1 , … , v r v_1,\ldots,v_r v 1 , … , v r of S n S_n S n with
∥ v j − f j ∣ X n ∥ n ≤ C λ ε , j = 1 , … , r . (29) \left\|v_j-f_j|_{X_n}\right\|_n\le C_\lambda\varepsilon,\qquad j=1,\ldots,r. \tag*{(29)} ∥ v j − f j ∣ X n ∥ n ≤ C λ ε , j = 1 , … , r . (29)
We take V n , λ \mathcal{V}_{n,\lambda} V n , λ to be the event that such a basis exists. This event is defined without θ \theta θ and δ ε \delta_\varepsilon δ ε , and the probability bound just quoted gives the failure probability in the statement. The last sentence of that proof uses this basis to derive the statement of [7 Theorem 2.7(2)] , which concerns eigenvectors of L L L . We use the basis itself, which gives the bound for every unit vector of S n S_n S n .
Now take any unit v ∈ S n v\in S_n v ∈ S n . There are coefficients β 1 , … , β r ∈ R \beta_1,\ldots,\beta_r\in\mathbb{R} β 1 , … , β r ∈ R such that
v = ∑ j = 1 r β j v j , ∑ j = 1 r β j 2 = 1. v=\sum_{j=1}^{r}\beta_jv_j,\qquad\sum_{j=1}^{r}\beta_j^2=1. v = j = 1 ∑ r β j v j , j = 1 ∑ r β j 2 = 1.
Define f : = ∑ j = 1 r β j f j f:=\sum_{j=1}^{r}\beta_jf_j f := ∑ j = 1 r β j f j . Since all f j f_j f j belong to E λ E_\lambda E λ and are orthonormal in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , f ∈ E λ f\in E_\lambda f ∈ E λ and ∥ f ∥ L 2 ( μ ) = 1 \|f\|_{L^2(\mu)}=1 ∥ f ∥ L 2 ( μ ) = 1 . Moreover,
∥ v − f ∣ X n ∥ n ≤ ∑ j = 1 r ∣ β j ∣ ∥ v j − f j ∣ X n ∥ n ≤ C λ ε ∑ j = 1 r ∣ β j ∣ ≤ r C λ ε , \left\|v-f|_{X_n}\right\|_n\le\sum_{j=1}^{r}|\beta_j|\left\|v_j-f_j|_{X_n}\right\|_n\le C_\lambda\varepsilon\sum_{j=1}^{r}|\beta_j|\le\sqrt{r}C_\lambda\varepsilon, ∥ v − f ∣ X n ∥ n ≤ j = 1 ∑ r ∣ β j ∣ ∥ v j − f j ∣ X n ∥ n ≤ C λ ε j = 1 ∑ r ∣ β j ∣ ≤ r C λ ε ,
where the last inequality is Cauchy–Schwarz. This proves (28). □
Proof of Theorem 4. Since the spectral gaps γ l − − 1 M \gamma_{l_--1}^{\mathcal{M}} γ l − − 1 M and γ l + M \gamma_{l_+}^{\mathcal{M}} γ l + M are strictly positive, (25) holds at these two indices for n n n large enough and δ / ε \delta/\varepsilon δ / ε small enough. Let N : = l + + 1 N:=l_++1 N := l + + 1 , and take θ : = θ N \theta:=\theta_N θ := θ N and δ ε : = θ N ε \delta_\varepsilon:=\theta_N\varepsilon δ ε := θ N ε as in Lemma 44 of Appendix C. For n n n large enough, (A1)–(A4) and (S2N) hold, which are the hypotheses of Lemma 32. With V n , λ \mathcal{V}_{n,\lambda} V n , λ the event of Lemma 32, set
H n , l : = V n , λ ∩ ⋂ j ∈ J l E n ( j ) . \mathcal{H}_{n,l}:=\mathcal{V}_{n,\lambda}\cap\bigcap_{j\in J_l}\mathcal{E}_n^{(j)}. H n , l := V n , λ ∩ j ∈ J l ⋂ E n ( j ) .
Given H n , l \mathcal{H}_{n,l} H n , l , let u ~ l \widetilde{u}_l u l be an eigenvector of L ~ \widetilde{L} L with eigenvalue λ ~ l \widetilde{\lambda}_l λ l and ∥ u ~ l ∥ n = 1 \|\widetilde{u}_l\|_n=1 ∥ u l ∥ n = 1 . We show that u ~ l \widetilde{u}_l u l is close to a unit vector v ∈ S n v\in S_n v ∈ S n , and that v v v is close to a unit eigenfunction of Δ ρ \Delta_\rho Δ ρ for λ \lambda λ . Finally we bound P n ( H n , l ) \mathbb{P}_n(\mathcal{H}_{n,l}) P n ( H n , l ) . The eigenbasis of L ~ \widetilde{L} L chosen at the start of § 6.4 was arbitrary. Since u ~ l \widetilde{u}_l u l is a unit eigenvector for the l l l -th eigenvalue λ ~ l \widetilde{\lambda}_l λ l , we may take it as the l l l -th vector of this basis. Then u ~ l \widetilde{u}_l u l lies in Range ( P ~ l − : l + ) \operatorname{Range}(\widetilde{P}_{l_{-}:l_{+}}) Range ( P l − : l + ) , because l − ≤ l ≤ l + l_{-} \le l \le l_{+} l − ≤ l ≤ l + . The hypothesis of Corollary 31 , (25) at the indices l − − 1 l_{-}-1 l − − 1 and l + l_{+} l + , holds by the first paragraph of this proof, and H n , l \mathcal{H}_{n,l} H n , l is contained in the event ⋂ j ∈ J l E n ( j ) \bigcap_{j\in J_l}\mathcal{E}_n^{(j)} ⋂ j ∈ J l E n ( j ) of that corollary. The corollary gives
t : = dist n ( u ~ l , S n ) ≤ C l − , l + δ ε . t := \operatorname{dist}_n(\widetilde{u}_l,S_n) \le C_{l_{-},l_{+}}\sqrt{\frac{\delta}{\varepsilon}}. t := dist n ( u l , S n ) ≤ C l − , l + ε δ .
For δ / ε \delta/\varepsilon δ / ε small enough, t < 1 t<1 t < 1 , so z : = P l − : l + u ~ l z:=P_{l_{-}:l_{+}}\widetilde{u}_l z := P l − : l + u l is nonzero. Set v : = z / ∥ z ∥ n v:=z/\lVert z\rVert_n v := z / ∥ z ∥ n , a unit vector in S n S_n S n . The orthogonal decomposition u ~ l = z + ( I − P l − : l + ) u ~ l \widetilde{u}_l=z+(I-P_{l_{-}:l_{+}})\widetilde{u}_l u l = z + ( I − P l − : l + ) u l gives ∥ z ∥ n 2 = 1 − t 2 \lVert z\rVert_n^2=1-t^2 ∥ z ∥ n 2 = 1 − t 2 and ⟨ u ~ l , v ⟩ n = ∥ z ∥ n \langle\widetilde{u}_l,v\rangle_n=\lVert z\rVert_n ⟨ u l , v ⟩ n = ∥ z ∥ n . Since ∥ z ∥ n ≤ 1 \lVert z\rVert_n\le1 ∥ z ∥ n ≤ 1 , we have ∥ z ∥ n ≥ ∥ z ∥ n 2 \lVert z\rVert_n\ge\lVert z\rVert_n^2 ∥ z ∥ n ≥ ∥ z ∥ n 2 , so
∥ u ~ l − v ∥ n 2 = 2 − 2 ⟨ u ~ l , v ⟩ n = 2 − 2 ∥ z ∥ n ≤ 2 − 2 ∥ z ∥ n 2 = 2 t 2 . \lVert\widetilde{u}_l-v\rVert_n^2=2-2\langle\widetilde{u}_l,v\rangle_n=2-2\lVert z\rVert_n\le2-2\lVert z\rVert_n^2=2t^2. ∥ u l − v ∥ n 2 = 2 − 2 ⟨ u l , v ⟩ n = 2 − 2 ∥ z ∥ n ≤ 2 − 2 ∥ z ∥ n 2 = 2 t 2 .
and therefore
∥ u ~ l − v ∥ n ≤ 2 C l − , l + δ ε . (30) \lVert\widetilde{u}_l-v\rVert_n\le\sqrt{2}C_{l_{-},l_{+}}\sqrt{\frac{\delta}{\varepsilon}}. \tag*{(30)} ∥ u l − v ∥ n ≤ 2 C l − , l + ε δ . (30)
By Lemma 32 there is an f l ∈ E λ f_l\in E_{\lambda} f l ∈ E λ , that is, with Δ ρ f l = λ l M f l \Delta_{\rho}f_l=\lambda_l^{\mathcal{M}}f_l Δ ρ f l = λ l M f l , such that ∥ f l ∥ L 2 ( μ ) = 1 \lVert f_l\rVert_{L^2(\mu)}=1 ∥ f l ∥ L 2 ( μ ) = 1 and
∥ v − f l ∣ X n ∥ n ≤ r C λ ε . \lVert v-f_l|_{X_n}\rVert_n\le\sqrt{r}\,C_{\lambda}\varepsilon. ∥ v − f l ∣ X n ∥ n ≤ r C λ ε .
Combining this bound with (30) by the triangle inequality gives
∥ u ~ l − f l ∣ X n ∥ n ≤ 2 C l − , l + δ ε + r C λ ε . (31) \lVert\widetilde{u}_l-f_l|_{X_n}\rVert_n\le\sqrt{2}C_{l_{-},l_{+}}\sqrt{\frac{\delta}{\varepsilon}}+\sqrt{r}\,C_{\lambda}\varepsilon. \tag*{(31)} ∥ u l − f l ∣ X n ∥ n ≤ 2 C l − , l + ε δ + r C λ ε . (31)
The constants C l − , l + C_{l_{-},l_{+}} C l − , l + , C λ C_{\lambda} C λ and r r r do not depend on n n n or on the displacements, so (31) is the asserted O ( δ / ε + ε ) O(\sqrt{\delta/\varepsilon}+\varepsilon) O ( δ / ε + ε ) estimate. If δ = O ( ε 3 ) \delta=O(\varepsilon^3) δ = O ( ε 3 ) , then δ / ε = O ( ε ) \sqrt{\delta/\varepsilon}=O(\varepsilon) δ / ε = O ( ε ) and the right-hand side is O ( ε ) O(\varepsilon) O ( ε ) .
It remains to bound P n ( H n , l ) P_n(\mathcal{H}_{n,l}) P n ( H n , l ) . By Lemma 32, the failure probability of V n , λ \mathcal{V}_{n,\lambda} V n , λ is at most
C λ n exp ( − c λ n θ ε 2 δ ) + C λ n exp ( − c λ n ε m + 4 ) . C_{\lambda}n\exp(-c_{\lambda}n\theta_{\varepsilon}^2\delta)+C_{\lambda}n\exp(-c_{\lambda}n\varepsilon^{m+4}). C λ n exp ( − c λ n θ ε 2 δ ) + C λ n exp ( − c λ n ε m + 4 ) .
By Lemma 44 with c = c λ c=c_{\lambda} c = c λ , the first term is at most C λ n − p C_{\lambda}n^{-p} C λ n − p for n n n large, and the second is at most C λ n − p C_{\lambda}n^{-p} C λ n − p once c λ C m + 4 ε ≥ p + 1 c_{\lambda}C^{m+4}\varepsilon\ge p+1 c λ C m + 4 ε ≥ p + 1 . These two bounds, Proposition 22 at each index j ∈ J l j\in J_l j ∈ J l and a union bound give P n ( H n , l ) ≥ 1 − O ( n − p ) P_n(\mathcal{H}_{n,l})\ge1-O(n^{-p}) P n ( H n , l ) ≥ 1 − O ( n − p ) . □ \square □
Proof of Remark 5. Integration by parts gives
⟨ Δ ρ f , f ⟩ L 2 ( μ ) = 1 2 ∫ M ρ 2 ∣ ∇ f ∣ 2 d Vol M . \langle\Delta_{\rho}f,f\rangle_{L^2(\mu)}=\frac{1}{2}\int_{\mathcal{M}}\rho^2\lvert\nabla f\rvert^2\,d\operatorname{Vol}_{\mathcal{M}}. ⟨ Δ ρ f , f ⟩ L 2 ( μ ) = 2 1 ∫ M ρ 2 ∣ ∇ f ∣ 2 d Vol M .
Since M \mathcal{M} M is connected, the eigenfunctions for λ 1 M = 0 \lambda_1^{\mathcal{M}}=0 λ 1 M = 0 are the constant functions, so λ 1 M \lambda_1^{\mathcal{M}} λ 1 M is simple, and since μ \mu μ is a probability measure, its unit eigenfunctions are the constants ± 1 \pm1 ± 1 . The noisy graph Laplacian has λ ~ 1 = 0 \widetilde{\lambda}_1=0 λ 1 = 0 , with the constant vectors as eigenvectors. Given E n ( 2 ) \mathcal{E}_n^{(2)} E n ( 2 ) , which has probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) by Proposition 22 , the proof of Corollary 2 at index 2 gives ∣ λ ~ 2 − σ n λ 2 M ∣ ≤ K 2 δ / ε + C 2 ′ ε \lvert\widetilde{\lambda}_2-\sigma_n\lambda_2^{\mathcal{M}}\rvert\le K_2\delta/\varepsilon+C'_2\varepsilon ∣ λ 2 − σ n λ 2 M ∣ ≤ K 2 δ / ε + C 2 ′ ε . Since λ 2 M > 0 \lambda_2^{\mathcal{M}}>0 λ 2 M > 0 , this gives λ ~ 2 > 0 \widetilde{\lambda}_2>0 λ 2 > 0 for n n n large enough and δ / ε \delta/\varepsilon δ / ε small enough, so λ ~ 1 \widetilde{\lambda}_1 λ 1 is simple. Every eigenvector u ~ 1 \widetilde{u}_1 u 1 for λ ~ 1 \widetilde{\lambda}_1 λ 1 with ∥ u ~ 1 ∥ n = 1 \lVert\widetilde{u}_1\rVert_n=1 ∥ u 1 ∥ n = 1 is therefore a constant vector ( c , … , c ) (c,\ldots,c) ( c , … , c ) with c 2 = ∥ u ~ 1 ∥ n 2 = 1 c^2=\lVert\widetilde{u}_1\rVert_n^2=1 c 2 = ∥ u 1 ∥ n 2 = 1 . The constant function f 1 ≡ c f_1\equiv c f 1 ≡ c is then a unit eigenfunction of Δ ρ \Delta_{\rho} Δ ρ for λ 1 M \lambda_1^{\mathcal{M}} λ 1 M whose restriction is f 1 ∣ X n = u ~ 1 f_1|_{X_n}=\widetilde{u}_1 f 1 ∣ X n = u 1 , so
∥ u ~ 1 − f 1 ∣ X n ∥ n = 0. \lVert\widetilde{u}_1-f_1|_{X_n}\rVert_n=0. ∥ u 1 − f 1 ∣ X n ∥ n = 0.
Numerical Experiments
This section reports numerical experiments on the bounds from which Theorems 1 and 3 are proved. Let
W l ( a ) : = ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + , W_l(a) := (1+a)^{m+2}\lambda_l^- - (1-a)^{m+2}\lambda_l^+, W l ( a ) := ( 1 + a ) m + 2 λ l − − ( 1 − a ) m + 2 λ l + ,
the difference between the l l l -th eigenvalues of the comparison operators B B B and A A A of (16). Recall from (17) that Γ l + \Gamma_l^+ Γ l + is the difference between the ( l + 1 ) (l+1) ( l + 1 ) -st and the l l l -th eigenvalue of A A A , and from (18) that τ l \tau_l τ l is the largest value of ⟨ ( B − A ) u , u ⟩ n / ∥ u ∥ n 2 \langle(B-A)u,u\rangle_n/\lVert u\rVert_n^2 ⟨( B − A ) u , u ⟩ n / ∥ u ∥ n 2 over nonzero u u u in the range of P l + P_l^+ P l + , the span of the first l l l eigenvectors of L + L^+ L + . The proofs of both theorems pass through the chains
∣ λ l − λ ~ l ∣ ≤ W l ( a ) ⏟ computable ≤ 1 2 K l a , ∥ P l − P ~ l ∥ H S , n ≤ 2 2 l τ l / Γ l + ⏟ computable ≤ D l δ / ε . (32) \lvert\lambda_l-\widetilde{\lambda}_l\rvert\le\underbrace{W_l(a)}_{\text{computable}} \le\frac{1}{2}K_l a,\qquad\lVert P_l-\widetilde{P}_l\rVert_{\mathrm{HS},n} \le2\underbrace{\sqrt{2l\tau_l/\Gamma_l^+}}_{\text{computable}} \le D_l\sqrt{\delta/\varepsilon}. \tag*{(32)} ∣ λ l − λ l ∣ ≤ computable W l ( a ) ≤ 2 1 K l a , ∥ P l − P l ∥ HS , n ≤ 2 computable 2 l τ l / Γ l + ≤ D l δ / ε . (32)
We call ∣ λ l − λ ~ l ∣ \lvert\lambda_l-\widetilde{\lambda}_l\rvert ∣ λ l − λ l ∣ the eigenvalue error , ∥ P l − P ~ l ∥ H S , n \lVert P_l-\widetilde{P}_l\rVert_{\mathrm{HS},n} ∥ P l − P l ∥ HS , n the projector distance , and 2 2 l τ l / Γ l + 2\sqrt{2l\tau_l/\Gamma_l^+} 2 2 l τ l / Γ l + the computable projector bound . For a given point cloud both middle terms can be computed from L + L^+ L + and L − L^- L − , and we call them the computable bounds .
The experiments measure the computable bounds and stop there. The constant K l K_l K l does not depend on ε \varepsilon ε or n n n , but it contains the prefactor C L u n i f C_L^{\mathrm{unif}} C L unif of Lemma 17, which is not available in effective form, so the right-hand sides of (32) cannot be evaluated. The first inequality in each chain holds for every point cloud (for the projectors, every point cloud with Γ l + > 0 \Gamma_l^+>0 Γ l + > 0 ). These first inequalities are Corollary 19, and Proposition 29 with the triangle inequality, so checking them tests the implementation. The second inequalities, Proposition 25 and (26), hold for a < a ⋆ a<a^\star a < a ⋆ (Remark 15), given the event E n E_n E n for the eigenvalues and, under (25), given the event E n p r o j E_n^{\mathrm{proj}} E n proj of (24) for the projectors.
There are three experiments, and the first inequalities of (32) hold up to rounding error for every point cloud in all of them. The first experiment (Figure 1 and Figure 2 ) varies a a a for fixed clean points and bandwidth: W l ( a ) W_l(a) W l ( a ) and τ l \tau_l τ l grow linearly in a a a , and the computable projector bound grows roughly like a \sqrt{a} a . Varying a a a alone cannot show how the computable bounds depend on ε \varepsilon ε , which needs a a a fixed while ε \varepsilon ε decreases. The other two experiments (Figure 3 and Figure 4 ) do this on S 1 S^1 S 1 and S 2 S^2 S 2 . Both computable bounds stay nearly constant while ε \varepsilon ε decreases from 0.240 to 0.024 on S 1 S^1 S 1 and from 0.24 to 0.10 on S 2 S^2 S 2 . Under noise built from the clean points to make the eigenvalue error large, this error is between 14% and 16% of W l ( a ) W_l(a) W l ( a ) on S 1 S^1 S 1 and between 18% and 20% on S 2 S^2 S 2 , so it stays of order δ / ε \delta/\varepsilon δ / ε as ε \varepsilon ε decreases. In these two experiments n n n grows so that every point keeps the same expected number of neighbours. Then ε \varepsilon ε is of order n − 1 / m n^{-1/m} n − 1/ m and decreases much faster than the bandwidth ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon=C_\varepsilon(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) of the theorems, along which a tenfold decrease of ε \varepsilon ε would need n n n to grow more than 10 m + 4 10^{m+4} 1 0 m + 4 -fold. The faster decrease is legitimate because the first inequalities of (32) hold at every bandwidth, so the computable bounds still bound the eigenvalue error and the projector distance.
Figure 1. The bracket on one point cloud (S 1 S^1 S 1 , n = 4000 n=4000 n = 4000 , ε = 0.15 \varepsilon=0.15 ε = 0.15 , a = 0.032 a=0.032 a = 0.032 , global rescaling). Left: at every index, λ l \lambda_l λ l and λ ~ l \tilde{\lambda}_l λ ~ l lie between ( 1 − a ) m + 2 λ l + (1-a)^{m+2}\lambda_l^+ ( 1 − a ) m + 2 λ l + and ( 1 + a ) m + 2 λ l − (1+a)^{m+2}\lambda_l^- ( 1 + a ) m + 2 λ l − . The eigenvalues come in near pairs because every nonzero eigenvalue of the circle has multiplicity two. Right: the computable bound W l ( a ) W_l(a) W l ( a ) and the eigenvalue error it bounds. Index 1 is omitted there, since λ 1 = λ ~ 1 = 0 \lambda_1=\tilde{\lambda}_1=0 λ 1 = λ ~ 1 = 0 up to rounding.
Figure 2. Dependence on a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε (S 1 S^1 S 1 , n = 4000 n=4000 n = 4000 , ε = 0.15 \varepsilon=0.15 ε = 0.15 , l = 3 l=3 l = 3 , 5 seeds). Left: the eigenvalue error under global rescaling and isotropic noise, and the computable bound W l ( a ) W_l(a) W l ( a ) . Right: the projector distance and the computable bound 2 2 l τ l / Γ l + 2\sqrt{2l\tau_l/\Gamma_l^+} 2 2 l τ l / Γ l + , both divided by 2 l \sqrt{2l} 2 l , so the dash-dotted line at one is the trivial bound. The vertical line is at a ⋆ a_\star a ⋆ , and the sloping grey lines have the slopes given in the legends. Shaded bands show the scatter between seeds; where no band is visible it is narrower than the curve.
Figure 3. Fixed a = 0.0208 a = 0.0208 a = 0.0208 and expected degree 40 as ε \varepsilon ε decreases and n n n grows (S 1 S^1 S 1 , l = 3 l = 3 l = 3 , 10 seeds). The tick labels on the horizontal axis give n n n and ε \varepsilon ε . Left: the eigenvalue errors under the three kinds of noise, and the computable bound W l ( a ) W_l(a) W l ( a ) . The dashed grey line has slope − 1 -1 − 1 against n n n , that of a quantity proportional to ε \varepsilon ε . Right: the projector distance under the same noise and the computable bound 2 2 l τ l / Γ l + 2\sqrt{2l\tau_l/\Gamma_l^+} 2 2 l τ l / Γ l + , both divided by 2 l \sqrt{2l} 2 l , so the dash-dotted line at one is the trivial bound.
Figure 4. The experiment of Figure 3 on the unit sphere: fixed a = 0.0125 a = 0.0125 a = 0.0125 and expected degree 20 as ε \varepsilon ε decreases and n n n grows (S 2 S^2 S 2 , l = 4 l = 4 l = 4 , 10 seeds). The panels, kinds of noise and units are those of Figure 3 . Since n ε 2 n\varepsilon^2 n ε 2 is constant along the sweep, a quantity proportional to ε \varepsilon ε has slope − 1 / 2 -1/2 − 1/2 against n n n , the slope of the dashed grey line.
Setup. The experiments of Figure 1 –Figure 3 sample i.i.d. points from the uniform distribution on the circle S 1 ⊂ R 2 S^1\subset\mathbb{R}^2 S 1 ⊂ R 2 , and that of Figure 4 from the uniform distribution on the unit sphere S 2 ⊂ R 3 S^2\subset\mathbb{R}^3 S 2 ⊂ R 3 . Both manifolds satisfy the hypotheses of § 2 . Other manifolds and non-constant densities are not tested. The kernel profile is η ( t ) = ( 1 − t ) + \eta(t)=(1-t)_+ η ( t ) = ( 1 − t ) + , normalized to unit mass on B ( 0 , 1 ) ⊂ R m B(0,1)\subset\mathbb{R}^m B ( 0 , 1 ) ⊂ R m , which satisfies (K1)–(K5). The index l l l is the end of a multiplicity block, so that γ l M > 0 \gamma_l^M>0 γ l M > 0 as Theorem 3 requires. Noise is applied in the ambient space, R 2 \mathbb{R}^2 R 2 or R 3 \mathbb{R}^3 R 3 , and the noisy points are not projected back to the manifold. The expected degree is the expected number of other sample points within distance ε \varepsilon ε of a given sample point, where the distance is that of the ambient space. The theorems hold for any displacements δ 1 , … , δ n \delta_1,\ldots,\delta_n δ 1 , … , δ n with ∥ δ i ∥ ≤ δ \lVert\delta_i\rVert\le\delta ∥ δ i ∥ ≤ δ , including noise that depends on the clean points (Lemma 42 ). We use three kinds of noise, each with max i ∥ δ i ∥ = δ \max_i\lVert\delta_i\rVert=\delta max i ∥ δ i ∥ = δ . The global rescaling y i = x ˉ + ( 1 + δ / R ) ( x i − x ˉ ) y_i=\bar{x}+(1+\delta/R)(x_i-\bar{x}) y i = x ˉ + ( 1 + δ / R ) ( x i − x ˉ ) , with x ˉ \bar{x} x ˉ the centroid of the clean points and R = max i ∥ x i − x ˉ ∥ R=\max_i\lVert x_i-\bar{x}\rVert R = max i ∥ x i − x ˉ ∥ , multiplies every pairwise distance by 1 + δ / R 1+\delta/R 1 + δ / R . Isotropic noise moves each point by δ \delta δ in an independent, uniformly distributed direction. Steepest-descent noise, used only in Figure 3 and Figure 4 and defined before them, is built from the clean points to make the eigenvalue error large.
Each experiment is run on several independent point clouds, one for each seed. Plotted values are means over the seeds, and a shaded band shows the seed-to-seed scatter. Since each plotted quantity is positive and each axis logarithmic, the band runs from μ e − s \mu e^{-s} μ e − s to μ e s \mu e^s μ e s about the mean μ \mu μ , where s s s is the standard deviation of the logarithms of the seed values. The band displays the observed scatter and is not a confidence interval. Every exponent quoted below is a least-squares fit in logarithmic coordinates over the plotted grid, or over its part with a < a ⋆ a<a_\star a < a ⋆ where stated. The sloping grey lines in the figures are reference lines, with the slopes given in the legends.
The bracket. Figure 1 shows Corollary 19 on one point cloud, at the first ten indices. In the right panel, W l ( a ) W_l(a) W l ( a ) and the eigenvalue error both grow with l l l , and the error is nearly the same fraction of W l ( a ) W_l(a) W l ( a ) at every index. The bracket uses only a < 1 a<1 a < 1 and the monotonicity of η \eta η , so it also holds at the values of a a a above a ⋆ a_\star a ⋆ in Figure 2 .
Dependence on a a a . Figure 2 varies a a a at n = 4000 n=4000 n = 4000 , ε = 0.15 \varepsilon=0.15 ε = 0.15 and l = 3 l=3 l = 3 . Each point cloud is the same at every value of a a a , and so are the directions of the isotropic noise, so only δ \delta δ changes. In the left panel, W l ( a ) W_l(a) W l ( a ) is parallel to the reference line of slope one: its fitted exponent in a a a is 1.00. The observed eigenvalue error is at most about 3% of W l ( a ) W_l(a) W l ( a ) under both kinds of noise. The theorems do not predict how that error depends on a a a , and under isotropic noise it varies strongly between seeds.
The right panel shows the projectors. The computable projector bound is built from τ l \tau_l τ l , which is not plotted. In this experiment the fitted exponent of τ l \tau_l τ l in a a a is 1.00. The computable projector bound has exponent 0.56 (0.52 over a < a ⋆ a<a_\star a < a ⋆ ) and stays close to the reference line of slope 1/2, as the square root in (32) suggests. It crosses the trivial bound 2 l \sqrt{2l} 2 l of Lemma 27 between a = 0.064 a=0.064 a = 0.064 and a = 0.128 a=0.128 a = 0.128 , so at this sample size and bandwidth it is informative only for smaller a a a .
The observed projector distances have exponent 1.00 under global rescaling and 1.00 under isotropic noise. They are measured at one bandwidth, whereas Theorem 3 bounds the distance uniformly in n n n and ε \varepsilon ε , so they do not show that its square root can be dropped.
Steepest-descent noise. Let u u u be an eigenvector of L L L with eigenvalue λ l \lambda_l λ l and ∥ u ∥ n = 1 \lVert u\rVert_n=1 ∥ u ∥ n = 1 , and set
δ i = δ v i ∥ v i ∥ , v i = ∑ j ≠ i : ∥ x i − x j ∥ < ε ( u i − u j ) 2 e i j , e i j = x i − x j ∥ x i − x j ∥ . \delta_i=\delta\frac{v_i}{\lVert v_i\rVert},\qquad v_i=\sum_{j\ne i:\,\lVert x_i-x_j\rVert<\varepsilon}(u_i-u_j)^2e_{ij},\qquad e_{ij}=\frac{x_i-x_j}{\lVert x_i-x_j\rVert}. δ i = δ ∥ v i ∥ v i , v i = j = i : ∥ x i − x j ∥ < ε ∑ ( u i − u j ) 2 e ij , e ij = ∥ x i − x j ∥ x i − x j .
with δ i = 0 \delta_i=0 δ i = 0 if v i = 0 v_i=0 v i = 0 . So every nonzero δ i \delta_i δ i has length δ \delta δ and points in the direction of the sum of the vectors e i j e_{ij} e ij , weighted by ( u i − u j ) 2 (u_i-u_j)^2 ( u i − u j ) 2 . Each e i j e_{ij} e ij points away from the neighbour x j x_j x j . Moving x i x_i x i away from x j x_j x j lowers the weight ω i j \omega_{ij} ω ij , and so lowers ⟨ L u , u ⟩ n \langle Lu,u\rangle_n ⟨ Lu , u ⟩ n , in which ω i j \omega_{ij} ω ij appears with the factor ( u i − u j ) 2 (u_i-u_j)^2 ( u i − u j ) 2 .
This noise is meant to make the eigenvalue error ∣ λ l − λ ~ l ∣ \lvert\lambda_l-\widetilde{\lambda}_l\rvert ∣ λ l − λ l ∣ large. Since δ \delta δ is small compared with ε \varepsilon ε (a = 2 δ / ε a=2\delta/\varepsilon a = 2 δ / ε is 0.0208 in Figure 3 and 0.0125 in Figure 4), we look at how ⟨ L u , u ⟩ n \langle Lu,u\rangle_n ⟨ Lu , u ⟩ n , which equals λ l \lambda_l λ l , changes when every clean point starts to move in a chosen direction. Let h 1 , … , h n ∈ R d h_1,\ldots,h_n\in\mathbb{R}^d h 1 , … , h n ∈ R d with ∥ h i ∥ ≤ 1 \lVert h_i\rVert\le1 ∥ h i ∥ ≤ 1 , and let L ( s ) L(s) L ( s ) be the graph Laplacian of the points x i + s h i x_i+sh_i x i + s h i , so that L ( 0 ) = L L(0)=L L ( 0 ) = L . Write η ( t ) = c ( 1 − t ) \eta(t)=c(1-t) η ( t ) = c ( 1 − t ) on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with c > 0 c>0 c > 0 the normalizing constant. If the x i x_i x i are distinct and no ∥ x i − x j ∥ \lVert x_i-x_j\rVert ∥ x i − x j ∥ equals ε \varepsilon ε , which holds almost surely here, the chain rule and e j i = − e i j e_{ji}=-e_{ij} e j i = − e ij give
d d s ⟨ L ( s ) u , u ⟩ n ∣ s = 0 = − c n 2 ε m + 3 ∑ i = 1 n ⟨ h i , v i ⟩ . \left.\frac{d}{ds}\langle L(s)u,u\rangle_n\right|_{s=0}=-\frac{c}{n^2\varepsilon^{m+3}}\sum_{i=1}^{n}\langle h_i,v_i\rangle. d s d ⟨ L ( s ) u , u ⟩ n s = 0 = − n 2 ε m + 3 c i = 1 ∑ n ⟨ h i , v i ⟩ .
By the Cauchy–Schwarz inequality, this derivative is smallest for h i = δ i / δ h_i = \delta_i/\delta h i = δ i / δ . So the noise moves the points in the directions in which ⟨ L ( s ) u , u ⟩ n \langle L(s)u,u\rangle_n ⟨ L ( s ) u , u ⟩ n decreases fastest at s = 0 s=0 s = 0 , which gives the noise its name.
Fixed a a a on S 1 S^1 S 1 . Figure 3 holds a = 0.0208 = 0.5 a ⋆ a = 0.0208 = 0.5\,a_\star a = 0.0208 = 0.5 a ⋆ fixed while ε \varepsilon ε decreases from 0.240 to 0.024, so that δ = a ε / 2 \delta= a\varepsilon/2 δ = a ε /2 decreases from 0.00250 to 0.000250. The sample size grows from 522 to 5236, which keeps the expected degree at 40. Then ε \varepsilon ε is of order n − 1 / m n^{-1/m} n − 1/ m , which does not satisfy the condition ε ≫ ( log n / n ) 1 / m \varepsilon\gg(\log n/n)^{1/m} ε ≫ ( log n / n ) 1/ m under which Calder and García Trillos prove that the eigenvalues of L L L converge to those of σ η Δ ρ \sigma_\eta\Delta_\rho σ η Δ ρ [7 the discussion after Theorem 2.4] . The probability bound of Proposition 22 does not apply either, so only the first inequalities of (32) are assured.
In the left panel, W l ( a ) W_l(a) W l ( a ) stays nearly constant: fitted against n n n , it scales as n − 0.05 n^{-0.05} n − 0.05 . This is consistent with Proposition 25, which bounds W l ( a ) W_l(a) W l ( a ) by 1 2 K l a \frac{1}{2}K_l a 2 1 K l a given E n \mathcal{E}_n E n , a bound that does not depend on ε \varepsilon ε . Under steepest-descent noise the eigenvalue error, averaged over the seeds, is between 14% and 16% of W l ( a ) W_l(a) W l ( a ) at every bandwidth, so for this noise the error stays of order δ / ε \delta/\varepsilon δ / ε as ε \varepsilon ε decreases. Under global rescaling and isotropic noise the error is a decreasing fraction of W l ( a ) W_l(a) W l ( a ) . The global-rescaling error scales as n − 1.03 n^{-1.03} n − 1.03 , parallel to the reference line of slope − 1 -1 − 1 , that is, in proportion to ε \varepsilon ε . So it is of order δ \delta δ rather than δ / ε \delta/\varepsilon δ / ε . This is expected: the rescaling changes every weight by at most C η δ / R C_\eta\delta/R C η δ / R , a factor ε / ( 2 R ) \varepsilon/(2R) ε / ( 2 R ) smaller than the bound C η a C_\eta a C η a that holds for any noise. The isotropic error also decreases, but it varies too much between seeds to fit a power law.
Since every choice of displacements with ∥ δ i ∥ ≤ δ \lVert\delta_i\rVert\leq\delta ∥ δ i ∥ ≤ δ has eigenvalue error at most W l ( a ) W_l(a) W l ( a ) , the largest error over all such choices lies between the error under steepest-descent noise and W l ( a ) W_l(a) W l ( a ) ; these experiments do not locate it. The expected degree is the same at every bandwidth, so the experiment does not show that the error under steepest-descent noise stays of order δ / ε \delta/\varepsilon δ / ε along the bandwidth of the theorems, where the expected degree grows with n n n .
The right panel of Figure 3 shows the projectors on the same point clouds. The computable projector bound also stays nearly constant: it lies between 0.54 and 0.60 times the trivial bound 2 l \sqrt{2l} 2 l at every bandwidth and, fitted against n n n , scales as n − 0.05 n^{-0.05} n − 0.05 . This is consistent with the bound D l δ / ε = D l a / 2 D_l\sqrt{\delta/\varepsilon}=D_l\sqrt{a}/2 D l δ / ε = D l a /2 of (32) , which holds given E n p r o j E_n^{\mathrm{proj}} E n proj and depends on a a a alone. The observed distances decrease overall under all three kinds of noise. The distance under steepest-descent noise is the largest of the three, and the computable bound exceeds it by a factor that grows from 79 to 311. Since steepest-descent noise is not chosen to move the projector, the experiment does not show how large the projector distance can be at fixed a a a .
Fixed a a a on S 2 S^2 S 2 . Figure 4 repeats this experiment on the unit sphere S 2 S^2 S 2 , which, unlike S 1 S^1 S 1 , has dimension m = 2 m = 2 m = 2 and is not intrinsically flat. Here l = 4 l = 4 l = 4 , the end of the first nonzero block, and a = 0.0125 = 0.5 a ⋆ a = 0.0125 = 0.5a_\star a = 0.0125 = 0.5 a ⋆ , with a ⋆ a_\star a ⋆ taken at m = 2 m = 2 m = 2 . To hold the expected degree fixed, n n n must now grow like ε − 2 \varepsilon^{-2} ε − 2 , so a tenfold decrease of ε \varepsilon ε would need a hundredfold increase of n n n , beyond the reach of the dense eigensolvers used here. The range is therefore shorter: ε \varepsilon ε decreases from 0.24 to 0.10 and δ \delta δ from 0.0015 to 0.000625, while n n n grows from 1389 to 8000 at expected degree 20. Over this shorter range, the fitted exponents are less informative than on S 1 S^1 S 1 .
The results agree with those on S 1 S^1 S 1 . Both computable bounds stay nearly constant: fitted against n n n , W l ( a ) W_l(a) W l ( a ) scales as n − 0.03 n^{-0.03} n − 0.03 , and the computable projector bound lies between 0.60 and 0.69 times the trivial bound and scales as n − 0.08 n^{-0.08} n − 0.08 . Under steepest-descent noise the eigenvalue error, averaged over the seeds, is between 18% and 20% of W l ( a ) W_l(a) W l ( a ) at every bandwidth. The global-rescaling error scales as n − 0.52 n^{-0.52} n − 0.52 ; since ε \varepsilon ε is proportional to n − 1 / 2 n^{-1/2} n − 1/2 here, it is again proportional to ε \varepsilon ε . The isotropic error decreases overall, with a large seed-to-seed scatter. The observed projector distances decrease. The distance under steepest-descent noise is again the largest, and the computable bound exceeds it by a factor that grows from 126 to 202.
Remark 33 (Scope of these experiments). The experiments do not test the probability estimates for E n E_n E n and E n p r o j E_n^{\mathrm{proj}} E n proj , or the uniformity of the constants of Lemma 17, which Appendix § 8 verifies, and no experiment shows that a bound is sharp.
Uniformity of the Eigenvector Lipschitz Estimate over a Kernel Profile Family
The constants in the eigenvector Lipschitz estimate of Calder, García Trillos and Lewicka [8 Theorem 2.3] are allowed to depend on the kernel profile. To apply that result to the family { η ˉ a ± : a < a ⋆ } \{\bar{\eta}_a^\pm: a < a_\star\} { η ˉ a ± : a < a ⋆ } , this appendix verifies that the constants, probability constants and smallness thresholds of that theorem can be chosen uniformly over an admissible kernel profile family (Definition 34).
This appendix is not a new proof of that theorem. It follows the argument for [8 Theorem 2.3] in the source’s own order and records, for each step, which quantities of the profile its constants depend on and which smallness thresholds it imposes, including the implicit one behind each “ε ≪ 1 \varepsilon\ll1 ε ≪ 1 ”. The claims below can only be checked against the source, and the appendix is meant to be read with [8 ] open. For m = 1 m = 1 m = 1 two steps of the source need a small change, made where they occur, in Sections A.7 and A.8.
Throughout, the manifold M \mathcal{M} M , its induced Riemannian metric g g g and the density ρ \rho ρ are fixed. We denote by κ \kappa κ a generic member of an admissible family K \mathcal{K} K (Definition 34), and use A ε , κ A_{\varepsilon,\kappa} A ε , κ , A ˉ ε , κ \bar{A}_{\varepsilon,\kappa} A ˉ ε , κ , Δ ε , κ \Delta_{\varepsilon,\kappa} Δ ε , κ , d ε , κ d_{\varepsilon,\kappa} d ε , κ , d ε , κ , X n d_{\varepsilon,\kappa,X_n} d ε , κ , X n , σ κ \sigma_\kappa σ κ for the objects of [8 ] built with κ \kappa κ in place of η \eta η . A result is called uniform when its constants and its implicit smallness thresholds can be chosen once for all κ ∈ K \kappa\in\mathcal{K} κ ∈ K . Constants are named after their place in the source: C 3.6 C_{3.6} C 3.6 is the constant of [8 Lemma 3.6] , c 3.4 ( i i ) c_{3.4(ii)} c 3.4 ( ii ) the exponential constant of [8 Proposition 3.4(ii)] , and C 8.8 C_{8.8} C 8.8 the constant of display (8.8) in [8 ] . An unnamed C C C or c c c is a generic constant of the source, and what it depends on is stated where it appears.
Admissible families
Definition 34 (Admissible kernel profile family). Let 0 < s 1 < s 2 < 1 0 < s_1 < s_2 < 1 0 < s 1 < s 2 < 1 , and let S = B ( 0 , s 2 ) ∖ B ˉ ( 0 , s 1 ) S = B(0,s_2) \setminus\bar{B}(0,s_1) S = B ( 0 , s 2 ) ∖ B ˉ ( 0 , s 1 ) be the corresponding annulus. (The source writes δ 1 , δ 2 \delta_1,\delta_2 δ 1 , δ 2 for s 1 , s 2 s_1,s_2 s 1 , s 2 ; we do not, because δ i \delta_i δ i denotes the displacements.) A family K \mathcal{K} K of profiles κ : [ 0 , ∞ ) → [ 0 , ∞ ) \kappa: [0,\infty) \to[0,\infty) κ : [ 0 , ∞ ) → [ 0 , ∞ ) is admissible with data ( K ∞ , K L , σ ⋆ , S , c S ) (K_\infty, K_L, \sigma_\star, S, c_S) ( K ∞ , K L , σ ⋆ , S , c S ) if every κ ∈ K \kappa\in\mathcal{K} κ ∈ K satisfies:
(U1) (structure ) κ ≥ 0 \kappa\ge0 κ ≥ 0 , supp ( κ ) ⊆ [ 0 , 1 ] \operatorname{supp}(\kappa) \subseteq[0,1] supp ( κ ) ⊆ [ 0 , 1 ] , κ \kappa κ is non-increasing, and ∫ B ( 0 , 1 ) κ ( ∣ w ∣ ) d w = 1 \int_{B(0,1)} \kappa(|w|)\,dw = 1 ∫ B ( 0 , 1 ) κ ( ∣ w ∣ ) d w = 1 ;
(U2) (size ) ∥ κ ∥ ∞ ≤ K ∞ \|\kappa\|_\infty\le K_\infty ∥ κ ∥ ∞ ≤ K ∞ ;
(U3) (regularity ) Lip [ 0 , 1 ] ( κ ) ≤ K L \operatorname{Lip}_{[0,1]}(\kappa) \le K_L Lip [ 0 , 1 ] ( κ ) ≤ K L ;
(U4) (second moment ) σ κ ≥ σ ⋆ > 0 \sigma_\kappa\ge\sigma_\star> 0 σ κ ≥ σ ⋆ > 0 ;
(U5) (positivity ) κ ( ∣ w ∣ ) ≥ c S > 0 \kappa(|w|) \ge c_S > 0 κ ( ∣ w ∣ ) ≥ c S > 0 for every w ∈ B ( 0 , s 2 ) w \in B(0,s_2) w ∈ B ( 0 , s 2 ) .
For every κ ∈ K \kappa\in\mathcal{K} κ ∈ K , condition (U1), together with (U3), gives (CL1) of Section 4.3, and (U5) gives (CL2), which is [8 (9.1)] , with the annulus S S S and c = c S c = c_S c = c S . An annulus is the form of set for which [8 Proposition 9.2] is stated. The lower bound on the rest of the ball B ( 0 , s 2 ) B(0,s_2) B ( 0 , s 2 ) is used in Section A.8. The kernel w ↦ κ ( ∣ w ∣ ) w \mapsto\kappa(|w|) w ↦ κ ( ∣ w ∣ ) is radial for every profile, so radial symmetry is not listed as a condition.
Lemma 35. The family { η ˉ a ± : a ∈ [ 0 , a ⋆ ) } \{\bar{\eta}_a^\pm: a \in[0,a_\star)\} { η ˉ a ± : a ∈ [ 0 , a ⋆ )} is admissible with
K ∞ = 2 η ( 0 ) , K L = 3 C η , σ ⋆ = 1 3 σ η , S = B ( 0 , t 0 / 2 ) ∖ B ˉ ( 0 , t 0 / 4 ) , c S = c 0 3 . K_\infty= 2\eta(0), \qquad K_L = 3C_\eta, \qquad\sigma_\star= \frac{1}{3}\sigma_\eta, \qquad S = B(0,t_0/2) \setminus\bar{B}(0,t_0/4), \qquad c_S = \frac{c_0}{3}. K ∞ = 2 η ( 0 ) , K L = 3 C η , σ ⋆ = 3 1 σ η , S = B ( 0 , t 0 /2 ) ∖ B ˉ ( 0 , t 0 /4 ) , c S = 3 c 0 .
Proof. By Lemma 9, η a ± \eta_a^{\pm} η a ± is non-negative and non-increasing, and so is η ˉ a ± \bar{\eta}_a^{\pm} η ˉ a ± , which is η a ± \eta_a^{\pm} η a ± divided by the positive constant I a ± I_a^{\pm} I a ± . The rest of (U1) and (U2)–(U5) are the bounds of Lemma 14, with B ( 0 , s 2 ) = B ( 0 , t 0 / 2 ) B(0,s_2)=B(0,t_0/2) B ( 0 , s 2 ) = B ( 0 , t 0 /2 ) .
Proposition 36. Let K \mathcal{K} K be admissible with data ( K ∞ , K L , σ ∗ , S , c S ) (K_\infty,K_L,\sigma_*,S,c_S) ( K ∞ , K L , σ ∗ , S , c S ) . Then the constants, the probability constants and the smallness thresholds, stated and implicit, of [8 Theorem 2.3] can be chosen to depend only on M \mathcal{M} M , ρ \rho ρ and the data. In particular they can be chosen the same for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K .
Sections A.2–A.9 prove Proposition 36. Lemma 17 is Proposition 36 applied to the family of Lemma 35.
Concentration and counting
[8 Lemma 3.2] is a Bernstein estimate for ∑ i ψ ( x i ) 1 { ∣ x i − x ∣ ≤ ε } \sum_i\psi(x_i)\mathbf{1}_{\{|x_i-x|\le\varepsilon\}} ∑ i ψ ( x i ) 1 { ∣ x i − x ∣ ≤ ε } , with ψ \psi ψ a bounded Borel function, valid for ε 2 ≤ t ≤ 1 \varepsilon^2\le t\le1 ε 2 ≤ t ≤ 1 . Its constant depends only on M \mathcal{M} M . The density enters through an explicit factor ∥ ρ ∥ C 0 \|\rho\|_{C^0} ∥ ρ ∥ C 0 , and ψ \psi ψ enters only through a factor ∥ ψ ∥ ∞ \|\psi\|_\infty ∥ ψ ∥ ∞ in the size of the deviation. In particular, the exponent of the probability bound does not involve ψ \psi ψ . We use this below, where the lemma is applied to functions ψ \psi ψ built from κ \kappa κ . The requirement ε ≪ 1 \varepsilon\ll1 ε ≪ 1 of Lemma 3.2 is geometric: it comes from the inclusions between Euclidean and geodesic balls and from volume estimates for small balls. [8 Corollary 3.3] applies the lemma with ψ ≡ 1 \psi\equiv1 ψ ≡ 1 at one fixed value of t t t depending only on M \mathcal{M} M and ρ \rho ρ , and [8 Propositions 3.4 and 3.5] involve no kernel. So all four results are uniform.
Degree control
[8 Lemma 3.6] gives
P ( sup x ∈ M ∣ d ε , κ , X n ( x ) − d ε , κ ( x ) ∣ ≤ C 3.6 t ) ≥ 1 − C ( t ε ) − m exp ( − c t 2 n ε m ) . \mathbb{P}\left(\sup_{x\in\mathcal{M}}\left|d_{\varepsilon,\kappa,X_n}(x)-d_{\varepsilon,\kappa}(x)\right|\le C_{3.6}t\right)\ge1-C(t\varepsilon)^{-m}\exp(-ct^2n\varepsilon^m). P ( x ∈ M sup ∣ d ε , κ , X n ( x ) − d ε , κ ( x ) ∣ ≤ C 3.6 t ) ≥ 1 − C ( tε ) − m exp ( − c t 2 n ε m ) .
with constants depending on M \mathcal{M} M , ρ \rho ρ , ∥ κ ∥ ∞ \|\kappa\|_\infty ∥ κ ∥ ∞ and Lip [ 0 , 1 ] ( κ ) \operatorname{Lip}_{[0,1]}(\kappa) Lip [ 0 , 1 ] ( κ ) , which (U2) and (U3) bound. The sup-norm enters when Lemma 3.2 is applied to ψ = κ ( ∣ ⋅ − x ∣ / ε ) \psi=\kappa(|\cdot-x|/\varepsilon) ψ = κ ( ∣ ⋅ − x ∣/ ε ) . The Lipschitz constant enters in [8 (3.11)] , which bounds the difference between κ ( ∣ y − x ∣ / ε ) \kappa(|y-x|/\varepsilon) κ ( ∣ y − x ∣/ ε ) and κ ( d M ( x , y ) / ε ) \kappa(d_{\mathcal{M}}(x,y)/\varepsilon) κ ( d M ( x , y ) / ε ) and also uses monotonicity. Both enter again when the estimate is extended from a net to all x ∈ M x\in\mathcal{M} x ∈ M . The requirement ε ≪ 1 \varepsilon\ll1 ε ≪ 1 of Lemma 3.6 comes from Lemma 3.2, Corollary 3.3, Proposition 3.4 and the bound [8 (3.2)] between Euclidean and geodesic distance, so its threshold depends only on M \mathcal{M} M and ρ \rho ρ .
[8 Corollary 3.7] adds the expansion d ε , κ = ρ + O ( ε 2 ) d_{\varepsilon,\kappa}=\rho+O(\varepsilon^2) d ε , κ = ρ + O ( ε 2 ) of [8 (3.9)] , with a constant that does not depend on κ \kappa κ . Normalization gives the leading term ρ \rho ρ , the first-order term vanishes by radial symmetry, and the second-order term is at most C ε 2 ∫ κ ( ∣ w ∣ ) d w = C ε 2 C\varepsilon^2\int\kappa(|w|)\,dw=C\varepsilon^2 C ε 2 ∫ κ ( ∣ w ∣ ) d w = C ε 2 . Hence, given the event of Lemma 3.6, sup M ∣ d ε , κ , X n − ρ ∣ ≤ C 3.7 t \sup_{\mathcal{M}}|d_{\varepsilon,\kappa,X_n}-\rho|\le C_{3.7}t sup M ∣ d ε , κ , X n − ρ ∣ ≤ C 3.7 t . Since C 3.7 C_{3.7} C 3.7 is uniform, the requirement t ≪ 1 t\ll1 t ≪ 1 of the source has a uniform threshold, below which d ε , κ , X n ∈ [ 1 2 ρ min , 2 ρ max ] d_{\varepsilon,\kappa,X_n}\in[\frac{1}{2}\rho_{\min},2\rho_{\max}] d ε , κ , X n ∈ [ 2 1 ρ m i n , 2 ρ m a x ] . [8 Remark 3.8] , which bounds the number of sample points in a ball and is used in [8 Theorem 4.2] , concerns only the indicator profile and does not involve κ \kappa κ .
Concentration of the double convolution
[8 Lemma 4.1] bounds sup x , y ∣ g n , ε ( x , y ) − g ε ( x , y ) ∣ \sup_{x,y}|g_{n,\varepsilon}(x,y)-g_\varepsilon(x,y)| sup x , y ∣ g n , ε ( x , y ) − g ε ( x , y ) ∣ by C 4.1 t ε − m C_{4.1}t\varepsilon^{-m} C 4.1 t ε − m , where g n , ε g_{n,\varepsilon} g n , ε and g ε g_\varepsilon g ε are the empirical and the continuum double convolution of the kernel, defined in its proof. With κ ε ( s ) : = ε − m κ ( s / ε ) \kappa_\varepsilon(s):=\varepsilon^{-m}\kappa(s/\varepsilon) κ ε ( s ) := ε − m κ ( s / ε ) , applying [8 Lemma 3.2] to ψ = κ ε ( ∣ ⋅ − x ∣ ) κ ε ( ∣ ⋅ − y ∣ ) / ( η ρ ) \psi=\kappa_\varepsilon(|\cdot-x|)\kappa_\varepsilon(|\cdot-y|)/(\eta\rho) ψ = κ ε ( ∣ ⋅ − x ∣ ) κ ε ( ∣ ⋅ − y ∣ ) / ( η ρ ) produces the factor ∥ κ ∥ ∞ 2 \|\kappa\|_\infty^2 ∥ κ ∥ ∞ 2 , and passing from a t ε t\varepsilon tε -net to arbitrary points produces Lip [ 0 , 1 ] ( κ ) ∥ κ ∥ ∞ \operatorname{Lip}_{[0,1]}(\kappa)\|\kappa\|_\infty Lip [ 0 , 1 ] ( κ ) ∥ κ ∥ ∞ and ∥ κ ∥ ∞ 2 \|\kappa\|_\infty^2 ∥ κ ∥ ∞ 2 . These factors appear only in the constant C 4.1 C_{4.1} C 4.1 . The requirement ε ≪ 1 \varepsilon\ll1 ε ≪ 1 of the lemma is that of Lemma 3.2, Corollary 3.3 and Proposition 3.4, so its threshold depends only on M \mathcal{M} M and ρ \rho ρ .
The exponent in Lemma 3.2 does not involve ψ \psi ψ . Hence c 4.1 = min { c 3.2 , c 3.3 , c 3.4 ( i i ) } c_{4.1}=\min\{c_{3.2},c_{3.3},c_{3.4(ii)}\} c 4.1 = min { c 3.2 , c 3.3 , c 3.4 ( ii ) } does not depend on κ \kappa κ , while (U2) and (U3) bound the remaining constants of Lemma 4.1. Lemma 4.1 uses neither the normalization in (U1), nor (U4), nor (U5).
Interpolation and the discrete-to-nonlocal transfer
The estimate of [8 Theorem 4.2] holds given the intersection of the three events of [8 (4.5)] : the event that the bound of Remark 3.8 holds for the number of sample points in each ball of radius 2 ε 2\varepsilon 2 ε , the event of Lemma 3.6, and the event of Lemma 4.1. The profile enters the constant of Theorem 4.2 through C 3.6 C_{3.6} C 3.6 and C 4.1 C_{4.1} C 4.1 and, in the proof itself, through ∥ κ ∥ ∞ \|\kappa\|_\infty ∥ κ ∥ ∞ , Lip [ 0 , 1 ] ( κ ) \operatorname{Lip}_{[0,1]}(\kappa) Lip [ 0 , 1 ] ( κ ) and the normalization. No other quantity of the profile appears. Theorem 4.2 requires ε ≪ 1 \varepsilon\ll1 ε ≪ 1 for Remark 3.8, Lemmas 3.6 and 4.1 and [8 (3.2)] , and t ≪ 1 t \ll1 t ≪ 1 to bound the degree from below as in Section A.3. Both thresholds are uniform.
[8 Theorem 2.9] is then Theorem 4.2 with t = ε 2 t=\varepsilon^2 t = ε 2 , combined with the two algebraic identities [8 (4.10)] . With I ε , X n I_{\varepsilon,X_n} I ε , X n the interpolation operator of the source, these express f − I ε , X n f f-I_{\varepsilon,X_n}f f − I ε , X n f and I ε , X n f − A ε , κ ( I ε , X n f ) I_{\varepsilon,X_n}f-A_{\varepsilon,\kappa}(I_{\varepsilon,X_n}f) I ε , X n f − A ε , κ ( I ε , X n f ) through the graph Laplacian and the nonlocal Laplacian. The event of Theorem 2.9, which we denote by E 2.9 E_{2.9} E 2.9 , and its probability constants are those of Theorem 4.2. The identities add to the constant of Theorem 2.9 only the degree bounds of Section A.3. Substituting t = ε 2 t=\varepsilon^2 t = ε 2 in the probability bound of Theorem 4.2 gives ( t ε ) − 2 m = ε − 6 m (t\varepsilon)^{-2m}=\varepsilon^{-6m} ( tε ) − 2 m = ε − 6 m and t 2 n ε m = n ε m + 4 t^2n\varepsilon^m=n\varepsilon^{m+4} t 2 n ε m = n ε m + 4 , so
P ( E 2.9 ) ≥ 1 − C 2.9 p r o b ε − 6 m exp ( − c 2.9 p r o b n ε m + 4 ) . \mathbb{P}(E_{2.9}) \ge1-C_{2.9}^{\mathrm{prob}}\varepsilon^{-6m}\exp\left(-c_{2.9}^{\mathrm{prob}}n\varepsilon^{m+4}\right). P ( E 2.9 ) ≥ 1 − C 2.9 prob ε − 6 m exp ( − c 2.9 prob n ε m + 4 ) .
Conditions (U1)–(U3) bound the constants of Theorem 2.9. The theorem uses neither (U4) nor (U5).
Geometry and averaging operators
[8 Lemma 5.1] and [8 Proposition 5.2] involve no kernel, so their constants and thresholds are the same for every κ \kappa κ . [8 Lemma 6.3] bounds the difference between A ε , κ A_{\varepsilon,\kappa} A ε , κ and A ‾ ε , κ \overline{A}_{\varepsilon,\kappa} A ε , κ ,
A ε , κ f ( x ) − A ‾ ε , κ f ( x ) = O ( ε 2 ) ∥ f ∥ L ∞ ( B M ( x , ε ) ) , A_{\varepsilon,\kappa}f(x)-\overline{A}_{\varepsilon,\kappa}f(x)=O(\varepsilon^2)\|f\|_{L^\infty(B_{\mathcal{M}}(x,\varepsilon))}, A ε , κ f ( x ) − A ε , κ f ( x ) = O ( ε 2 ) ∥ f ∥ L ∞ ( B M ( x , ε )) ,
with a constant depending only on ( M , g ) (\mathcal{M},g) ( M , g ) and ρ \rho ρ . Its proof uses the normalization, radial symmetry, non-negativity and support of κ \kappa κ , and the second moment σ κ \sigma_\kappa σ κ , which these bound by 1. Consequently
∥ A ‾ ε , κ f − f ∥ ∞ ε 2 ≤ ∥ A ε , κ f − f ∥ ∞ ε 2 + C ∥ f ∥ ∞ , \frac{\|\overline{A}_{\varepsilon,\kappa}f-f\|_\infty}{\varepsilon^2}\leq\frac{\|A_{\varepsilon,\kappa}f-f\|_\infty}{\varepsilon^2}+C\|f\|_\infty, ε 2 ∥ A ε , κ f − f ∥ ∞ ≤ ε 2 ∥ A ε , κ f − f ∥ ∞ + C ∥ f ∥ ∞ ,
with one constant for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K . Moreover, the identity Δ ε , κ f = − d ε , κ ( A ε , κ f − f ) / ε 2 \Delta_{\varepsilon,\kappa}f=-d_{\varepsilon,\kappa}(A_{\varepsilon,\kappa}f-f)/\varepsilon^2 Δ ε , κ f = − d ε , κ ( A ε , κ f − f ) / ε 2 of [8 Remark 6.2(iii)] and the lower bound d ε , κ ≥ 1 2 ρ min d_{\varepsilon,\kappa}\geq\frac{1}{2}\rho_{\min} d ε , κ ≥ 2 1 ρ m i n , which follows from d ε , κ = ρ + O ( ε 2 ) d_{\varepsilon,\kappa}=\rho+O(\varepsilon^2) d ε , κ = ρ + O ( ε 2 ) for ε \varepsilon ε below a uniform threshold, give ∥ A ε , κ f − f ∥ ∞ / ε 2 ≤ 2 ρ min − 1 ∥ Δ ε , κ f ∥ ∞ \|A_{\varepsilon,\kappa}f-f\|_\infty/\varepsilon^2\leq2\rho_{\min}^{-1}\|\Delta_{\varepsilon,\kappa}f\|_\infty ∥ A ε , κ f − f ∥ ∞ / ε 2 ≤ 2 ρ m i n − 1 ∥ Δ ε , κ f ∥ ∞ . The source uses these two steps to deduce [8 Theorem 2.11] from [8 Theorem 10.1] , which is stated with A ε , κ A_{\varepsilon,\kappa} A ε , κ . They add no dependence on the profile.
The coupled random walk
[8 Lemma 7.1] is an identity obtained by a change of variables, which uses only radial symmetry and the normalization. [8 Proposition 7.2] identifies the one-step transition operator of the walk, before the walk is stopped, with A ˉ ε , κ \bar{A}_{\varepsilon,\kappa} A ˉ ε , κ . Since supp ( κ ) ⊆ [ 0 , 1 ] \operatorname{supp}(\kappa) \subseteq[0,1] supp ( κ ) ⊆ [ 0 , 1 ] , every jump has length at most ε \varepsilon ε . The construction also requires ε ∥ ∇ log ρ ∥ ∞ < 1 2 \varepsilon\lVert\nabla\log\rho\rVert_{\infty} < \frac{1}{2} ε ∥ ∇ log ρ ∥ ∞ < 2 1 , which does not involve κ \kappa κ .
For m = 1 m=1 m = 1 the field of isometries Q Q Q of [8 Section 7.1] needs a small change. The source builds it from a rotation Q 0 ∈ S O ( m ) Q_0 \in SO(m) Q 0 ∈ S O ( m ) with Q 0 ( − e 1 ) = e 1 Q_0(-e_1)=e_1 Q 0 ( − e 1 ) = e 1 , and S O ( 1 ) SO(1) S O ( 1 ) contains none. Lemma 7.1 requires only a linear isometry, so we take Q 0 = − 1 Q_0=-1 Q 0 = − 1 . Then Q ( x ) Q(x) Q ( x ) is multiplication by the sign of ⟨ ∇ log ρ ( x ) , e 1 ( x ) ⟩ \langle\nabla\log\rho(x),e_1(x)\rangle ⟨ ∇ log ρ ( x ) , e 1 ( x )⟩ , and by + 1 +1 + 1 where this vanishes, with e 1 ( x ) e_1(x) e 1 ( x ) the parallel unit frame of the source. So ∣ ∇ log ρ ( x ) ∣ Q ( x ) e 1 ( x ) = ∇ log ρ ( x ) \lvert\nabla\log\rho(x)\rvert Q(x)e_1(x)=\nabla\log\rho(x) ∣ ∇ log ρ ( x )∣ Q ( x ) e 1 ( x ) = ∇ log ρ ( x ) , which is [8 (7.1)] , for either choice of e 1 ( x ) e_1(x) e 1 ( x ) , and in particular in the frame of [8 (8.2)] .
Let L ρ L_\rho L ρ be the Lipschitz constant of ∇ log ρ \nabla\log\rho ∇ log ρ , with tangent vectors compared by parallel transport, and let r r r be the radius of [8 Theorem 8.1] . In the case of [8 (8.11)] , where the drifts at the positions of both walks exceed σ κ / 4 \sigma_\kappa/4 σ κ /4 in norm, the source bounds the distance between their unit directions, compared by parallel transport, by 8 L ρ r / σ κ 8L_\rho r/\sigma_\kappa 8 L ρ r / σ κ . This is less than 2 when r ≤ c σ ∗ r \le c\sigma_\ast r ≤ c σ ∗ , a condition imposed below. For m = 1 m=1 m = 1 two unit vectors at one point are equal or opposite, so the two directions are equal. Since Q ( x ) Q(x) Q ( x ) sends t e 1 ( x ) t e_1(x) t e 1 ( x ) to t t t times the unit direction of the drift at x x x , for either choice of e 1 ( x ) e_1(x) e 1 ( x ) , the difference of the two isometries that enters the drift term is zero. The other case uses only that each Q ( x ) Q(x) Q ( x ) is an isometry. Hence [8 (8.12)] holds as in the source, with no new quantity of the profile.
Let ( X j ) j ≥ 0 (X_j)_{j\geq0} ( X j ) j ≥ 0 be the walk started at x 0 x_0 x 0 (the source writes the time index as n n n ), and let τ \tau τ be a stopping time that is at most the exit time of the walk from B M ( x 0 , r ) B_{\mathcal{M}}(x_0,r) B M ( x 0 , r ) . [8 Lemma 7.3] gives
E [ d M ( X τ , x 0 ) 2 ] ≤ 2 m σ κ ε 2 E [ τ ] . \mathbb{E}[d_{\mathcal{M}}(X_\tau,x_0)^2] \leq\frac{2m\sigma_\kappa\varepsilon}{2}\mathbb{E}[\tau]. E [ d M ( X τ , x 0 ) 2 ] ≤ 2 2 m σ κ ε E [ τ ] .
The lemma holds when C 7.3 ( ε + r ) ≤ 1 2 m σ ∗ C_{7.3}(\varepsilon+r)\leq\frac{1}{2}m\sigma_\ast C 7.3 ( ε + r ) ≤ 2 1 m σ ∗ , which is the same condition for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K .
In the statement of [8 Theorem 8.1] the profile appears only in the coefficient 4 r / σ κ 4r/\sigma_\kappa 4 r / σ κ , which is at most 4 r / σ ∗ 4r/\sigma_\ast 4 r / σ ∗ , and in the operator A ˉ ε , κ \bar{A}_{\varepsilon,\kappa} A ˉ ε , κ . Besides that of Lemma 7.3, the hypothesis ε ≪ r ≪ 1 \varepsilon\ll r \ll1 ε ≪ r ≪ 1 of Theorem 8.1 contains three requirements that involve the profile. The lower bound 4 ε 2 σ κ + O ( ε 2 r + ε 3 ) ≥ 3 ε 2 σ κ 4\varepsilon^2\sigma_\kappa+O(\varepsilon^2r+\varepsilon^3)\geq3\varepsilon^2\sigma_\kappa 4 ε 2 σ κ + O ( ε 2 r + ε 3 ) ≥ 3 ε 2 σ κ of [8 (8.8)] holds when C 8.8 ( r + ε ) ≤ σ ∗ C_{8.8}(r+\varepsilon)\leq\sigma_\ast C 8.8 ( r + ε ) ≤ σ ∗ . In the case where both drifts exceed σ κ / 4 \sigma_\kappa/4 σ κ /4 in norm, [8 (8.11)] requires 8 L ρ r / σ κ 8L_\rho r/\sigma_\kappa 8 L ρ r / σ κ to be small, which holds when r ≤ c σ ∗ r\leq c\sigma_\ast r ≤ c σ ∗ . The supermartingale inequality before [8 (8.14)] requires 1 2 σ κ + C ( r + ε ) ≤ σ κ / r \frac{1}{2}\sigma_\kappa+C(r+\varepsilon)\leq\sigma_\kappa/r 2 1 σ κ + C ( r + ε ) ≤ σ κ / r , which holds when r ≤ 1 r\leq1 r ≤ 1 and C ( r + ε ) ≤ 1 2 σ ∗ C(r+\varepsilon)\leq\frac{1}{2}\sigma_\ast C ( r + ε ) ≤ 2 1 σ ∗ . These three requirements and that of Lemma 7.3 hold when r ≤ c σ ∗ r\leq c\sigma_\ast r ≤ c σ ∗ and ε ≤ c ′ r \varepsilon\leq c'r ε ≤ c ′ r , with constants c , c ′ c,c' c , c ′ depending only on M \mathcal{M} M and ρ \rho ρ , as do C 7.3 C_{7.3} C 7.3 , C 8.8 C_{8.8} C 8.8 and C C C . The remaining requirements do not involve the profile. We fix a radius r 0 ≤ c σ ∗ r_0\leq c\sigma_\ast r 0 ≤ c σ ∗ that satisfies them. Theorem 8.1 then holds with r = r 0 r=r_0 r = r 0 for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K , for ε \varepsilon ε below one threshold.
The contraction estimate
Condition (U5) is used only in this section. [8 Proposition 9.2] is stated for an annulus on which the kernel is bounded below by a positive constant. We apply it with the annulus S S S and the constant c S c_S c S of Definition 34. With H N : = { w : ⟨ w , e 1 ⟩ ∈ ( 1 4 N , 1 2 N ) } H_N:=\{w:\langle w,e_1\rangle\in(\frac{1}{4N},\frac{1}{2N})\} H N := { w : ⟨ w , e 1 ⟩ ∈ ( 4 N 1 , 2 N 1 )} and α κ : = ∫ H N κ ( ∣ w ∣ ) d w \alpha_\kappa:=\int_{H_N}\kappa(\lvert w\rvert)\,dw α κ := ∫ H N κ (∣ w ∣) d w , Proposition 9.2 asserts that there is an integer N > 1 N>1 N > 1 such that (i) α κ ∈ ( 0 , 1 ) \alpha_\kappa\in(0,1) α κ ∈ ( 0 , 1 ) and (ii) ∣ S ∩ φ ε , x 0 , y 0 ( S ) ∣ > 1 2 ∣ S ∣ \lvert S\cap\varphi_{\varepsilon,x_0,y_0}(S)\rvert>\frac{1}{2}\lvert S\rvert ∣ S ∩ φ ε , x 0 , y 0 ( S )∣ > 2 1 ∣ S ∣ whenever d M ( x 0 , y 0 ) ≤ ε / N d_{\mathcal{M}}(x_0,y_0)\leq\varepsilon/N d M ( x 0 , y 0 ) ≤ ε / N and ε ≪ 1 \varepsilon\ll1 ε ≪ 1 . Here φ ε , x 0 , y 0 \varphi_{\varepsilon,x_0,y_0} φ ε , x 0 , y 0 is the map of [8 (9.2)] .
Property (ii) and its proof involve only the set S S S , the map φ ε , x 0 , y 0 \varphi_{\varepsilon,x_0,y_0} φ ε , x 0 , y 0 and the expansion [8 (9.3)] , whose constants depend only on ( M , g ) (\mathcal{M},g) ( M , g ) . The kernel does not appear in them. So (ii) holds for every N N N above a threshold that depends only on S S S , and its threshold for ε \varepsilon ε depends only on S S S and ( M , g ) (\mathcal{M},g) ( M , g ) . The same is true of the estimate ∣ S △ φ ε R e f φ ε ( S ) ∣ = O ( ε 2 ) \lvert S\mathbin{\triangle}\varphi_\varepsilon\mathop{\rm Ref}\varphi_\varepsilon(S)\rvert=O(\varepsilon^2) ∣ S △ φ ε Ref φ ε ( S )∣ = O ( ε 2 ) , which the proof of [8 Theorem 9.1] obtains by the same argument (φ ε \varphi_\varepsilon φ ε and R e f \mathop{\rm Ref} Ref are as in the source).
For property (i) the source takes N > 1 / s 2 N>1/s_2 N > 1/ s 2 . Then 1 2 N < s 2 \frac{1}{2N}<s_2 2 N 1 < s 2 , so the open set B ( 0 , s 2 ) ∩ H N B(0,s_2)\cap H_N B ( 0 , s 2 ) ∩ H N contains t e 1 t e_1 t e 1 for every t ∈ ( 1 4 N , 1 2 N ) t\in(\frac{1}{4N},\frac{1}{2N}) t ∈ ( 4 N 1 , 2 N 1 ) . Since κ ( ∣ w ∣ ) ≥ c S \kappa(\lvert w\rvert)\geq c_S κ (∣ w ∣) ≥ c S on B ( 0 , s 2 ) B(0,s_2) B ( 0 , s 2 ) by (U5), for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K
α κ ≥ α ∗ : = c S ∣ B ( 0 , s 2 ) ∩ H N ∣ > 0. (33) \alpha_\kappa\geq\alpha_\ast:= c_S\lvert B(0,s_2)\cap H_N\rvert> 0. \tag*{(33)} α κ ≥ α ∗ := c S ∣ B ( 0 , s 2 ) ∩ H N ∣ > 0. (33)
For m = 1 m = 1 m = 1 the ball is needed here: S = ( − s 2 , − s 1 ) ∪ ( s 1 , s 2 ) S = (-s_{2}, -s_{1}) \cup(s_{1}, s_{2}) S = ( − s 2 , − s 1 ) ∪ ( s 1 , s 2 ) and H N = ( 1 4 N , 1 N ) H_{N} = \left(\frac{1}{4N}, \frac{1}{N}\right) H N = ( 4 N 1 , N 1 ) are disjoint once N ≥ 1 / ( 2 s 1 ) N \ge1/(2s_{1}) N ≥ 1/ ( 2 s 1 ) . We fix N N N , depending only on S S S , such that (i) and (ii) hold.
The proof of [8 Theorem 9.1] uses at most L = 4 ( 3 N − 1 ) L = 4(3N - 1) L = 4 ( 3 N − 1 ) steps, each of which shortens the distance between the two points by at least ε / ( 4 N ) \varepsilon/(4N) ε / ( 4 N ) , and its final estimate holds with N ˉ = 60 N \bar{N} = 60N N ˉ = 60 N . The contraction factor is θ κ = 1 − c S ∣ S ∣ 2 α κ L \theta_{\kappa} = 1 - \frac{c_{S}|S|}{2}\alpha_{\kappa}^{L} θ κ = 1 − 2 c S ∣ S ∣ α κ L , so by (33)
θ κ ≤ θ ∗ : = 1 − c S ∣ S ∣ 2 α ∗ L < 1 , \theta_{\kappa} \le\theta_{*} := 1 - \frac{c_{S}|S|}{2}\alpha_{*}^{L} < 1, θ κ ≤ θ ∗ := 1 − 2 c S ∣ S ∣ α ∗ L < 1 ,
and the theorem holds with the contraction factor θ ∗ \theta_{*} θ ∗ for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K . The proof bounds the coefficients of [8 (9.11)] by 24 N 24N 24 N in its final estimate, so no bound on ( 1 − α κ ) − 1 (1-\alpha_{\kappa})^{-1} ( 1 − α κ ) − 1 enters. The remaining constant C 9.1 C_{9.1} C 9.1 and the thresholds for ε \varepsilon ε come from the geometric expansions, from property (ii), and from requiring the O ( ε 2 ) O(\varepsilon^{2}) O ( ε 2 ) error of each step to be at most ε / ( 4 N ) \varepsilon/(4N) ε / ( 4 N ) . They depend only on ( M , g ) (\mathcal{M},g) ( M , g ) , S S S , c S c_{S} c S and N N N . Theorem 9.1 involves no sample points.
Global regularity and the eigenvector estimate
The proof of [8 Theorem 10.1] combines Theorems 8.1 and 9.1 into [8 (10.1)] , valid for d M ( x 0 , y 0 ) < r d_{\mathcal{M}}(x_{0},y_{0}) < r d M ( x 0 , y 0 ) < r , where r r r is the radius of Theorem 8.1. We take r = r 0 r = r_{0} r = r 0 , the radius of Section A.7. The finitely many smallness conditions of Theorems 8.1 and 9.1, together with the requirement 5 ε < r 0 5\varepsilon< r_{0} 5 ε < r 0 , hold for ε \varepsilon ε below one uniform threshold. The proof takes the supremum of (10.1) over pairs of points at distance at most 5 ε 5\varepsilon 5 ε and absorbs the term with the factor θ κ \theta_{\kappa} θ κ into the left-hand side. This uses 1 / ( 1 − θ κ ) ≤ 1 / ( 1 − θ ∗ ) 1/(1-\theta_{\kappa}) \le1/(1-\theta_{*}) 1/ ( 1 − θ κ ) ≤ 1/ ( 1 − θ ∗ ) , and the coefficient 4 r 0 / σ κ 4r_{0}/\sigma_{\kappa} 4 r 0 / σ κ of Theorem 8.1 is at most 4 r 0 / σ ∗ 4r_{0}/\sigma_{*} 4 r 0 / σ ∗ . The general case follows by compactness of M \mathcal{M} M , as in the source, with constants depending only on M \mathcal{M} M and r 0 r_{0} r 0 . [8 Theorem 2.11] then follows by the two steps of Section A.6. Theorems 10.1 and 2.11 involve no sample points.
The estimate of [8 Theorem 2.1] holds given the intersection of E 2.9 E_{2.9} E 2.9 with the event of Lemma 3.6 at t = ε 2 t = \varepsilon^{2} t = ε 2 . By a union bound and ( t ε ) − m = ε − 3 m ≤ ε − 6 m (t\varepsilon)^{-m} = \varepsilon^{-3m} \le\varepsilon^{-6m} ( tε ) − m = ε − 3 m ≤ ε − 6 m , the probability of this intersection satisfies the bound displayed for E 2.9 E_{2.9} E 2.9 in Section A.5, with uniform constants C 2.1 p r o b C_{2.1}^{\mathrm{prob}} C 2.1 prob and c 2.1 p r o b c_{2.1}^{\mathrm{prob}} c 2.1 prob in place of C 2.9 p r o b C_{2.9}^{\mathrm{prob}} C 2.9 prob and c 2.9 p r o b c_{2.9}^{\mathrm{prob}} c 2.9 prob . The Lipschitz constant C 2.1 C_{2.1} C 2.1 of the theorem is a combination of the constants of Theorems 2.9 and 2.11 and of the lower bound on the degree. Each of these is uniform by the preceding sections.
Let Λ \Lambda Λ be the cutoff of Theorem 2.3. The estimate of [8 Theorem 2.3] holds given the intersection of the event of Theorem 2.1 with the event of [8 (11.2)] that every ball B M ( x i , r ) B_{\mathcal{M}}(x_{i},r) B M ( x i , r ) contains at least C ′ n r m C' nr^{m} C ′ n r m sample points. This r r r is not the radius r 0 r_{0} r 0 . The proof needs r + ε ≤ 1 / ( 2 C 2.1 ( Λ + 1 ) ) r+\varepsilon\le1/(2C_{2.1}(\Lambda+1)) r + ε ≤ 1/ ( 2 C 2.1 ( Λ + 1 )) and takes r r r of order 1 / ( Λ + 1 ) 1/(\Lambda+1) 1/ ( Λ + 1 ) . We take r = c r / ( Λ + 1 ) r = c_{r}/(\Lambda+1) r = c r / ( Λ + 1 ) and require ε ≤ c ε / ( Λ + 1 ) \varepsilon\le c_{\varepsilon}/(\Lambda+1) ε ≤ c ε / ( Λ + 1 ) , with c r c_{r} c r and c ε c_{\varepsilon} c ε chosen so that this inequality holds and r r r is below the threshold of Corollary 3.3. Since C 2.1 C_{2.1} C 2.1 is uniform, so are c r c_{r} c r and c ε c_{\varepsilon} c ε . The source applies Corollary 3.3 at radius r / 2 r/2 r /2 around each sample point, so the event of (11.2) fails with probability at most 2 n exp ( − c n ( Λ + 1 ) − m ) 2n\exp(-cn(\Lambda+1)^{-m}) 2 n exp ( − c n ( Λ + 1 ) − m ) , where c c c depends only on M \mathcal{M} M , ρ \rho ρ and c r c_{r} c r . Through c r c_{r} c r , the constant c c c depends on the data of the family, and it is the same for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K .
Altogether, for every κ ∈ K \kappa\in\mathcal{K} κ ∈ K and every ε ≤ c ε / ( Λ + 1 ) \varepsilon\le c_{\varepsilon}/(\Lambda+1) ε ≤ c ε / ( Λ + 1 ) below the uniform thresholds, the estimate of [8 Theorem 2.3] holds with one prefactor and with probability at least
1 − C 2.1 p r o b ε − 6 m exp ( − c 2.1 p r o b n ε m + 4 ) − 2 n exp ( − c n ( Λ + 1 ) − m ) . 1 - C_{2.1}^{\mathrm{prob}}\varepsilon^{-6m}\exp\left(-c_{2.1}^{\mathrm{prob}}n\varepsilon^{m+4}\right) - 2n\exp\left(-cn(\Lambda+1)^{-m}\right). 1 − C 2.1 prob ε − 6 m exp ( − c 2.1 prob n ε m + 4 ) − 2 n exp ( − c n ( Λ + 1 ) − m ) .
This is the bound of Lemma 17, with C P = C 2.1 p r o b C_{\mathbb{P}} = C_{2.1}^{\mathrm{prob}} C P = C 2.1 prob and c P c_{\mathbb{P}} c P the smaller of c 2.1 p r o b c_{2.1}^{\mathrm{prob}} c 2.1 prob and c c c . Every constant and threshold above has been chosen in terms of M \mathcal{M} M , ρ \rho ρ and the data of K \mathcal{K} K alone. This proves Proposition 36.
Denoising with the Projection Algorithm of Aizenbud and Sober
This appendix applies our results to the model of Aizenbud and Sober [1 ] , in which the observations are uniform in a tube of fixed radius σ \sigma σ around M \mathcal{M} M . We take the nearest points of the observations on M \mathcal{M} M as the clean points. The displacements are then bounded by δ = σ \delta=\sigma δ = σ , which does not decrease with n n n , so δ / ε \delta/\varepsilon δ / ε is not small and our results do not apply to the observations directly. We therefore first denoise the observations with the projection algorithm of [1 ] , which estimates the nearest point on M \mathcal{M} M of a given query point from a dataset drawn from the distribution of the observations, and converges to it as the dataset grows (Theorem 39). We apply it at each observation, always with the same additional dataset, which is independent of the observations. If, for each n n n , the additional dataset is large enough, then with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) the graph on the denoised points has the rates that [7 ] gives for the clean graph, in the number n n n of graph vertices.
We change some of the notation of [1 ] to avoid clashes with ours. Their ambient and intrinsic dimensions D D D and d d d , number n n n of data points, query point r r r , failure probability δ \delta δ , reach τ \tau τ and exponent r 1 r_1 r 1 are here d d d , m m m , N N N , y y y , q q q , τ M \tau_M τ M and r k r_k r k .
The tube model and the clean points
Let M ⊂ R d \mathcal{M}\subset\mathbb{R}^{d} M ⊂ R d satisfy the hypotheses of § 2 : compact, connected, orientable, smooth, of dimension m m m , without boundary. Its reach τ M \tau_M τ M is the largest τ \tau τ such that every point of R d \mathbb{R}^{d} R d at distance less than τ \tau τ from M \mathcal{M} M has a unique nearest point on M \mathcal{M} M [15 Definition 4.1] ; see also [1 Definition 1] . For a compact smooth submanifold it is positive [15 p. 432 and Theorem 4.12] . For 0 < σ < τ M 0<\sigma<\tau_M 0 < σ < τ M let
M σ : = { y ∈ R d : dist ( y , M ) < σ } \mathcal{M}_{\sigma}:=\{y\in\mathbb{R}^{d}:\operatorname{dist}(y,\mathcal{M})<\sigma\} M σ := { y ∈ R d : dist ( y , M ) < σ }
be the open tube of radius σ \sigma σ around M \mathcal{M} M , and let π \pi π denote the nearest-point projection onto M \mathcal{M} M , which is defined on M σ \mathcal{M}_{\sigma} M σ because σ < τ M \sigma<\tau_M σ < τ M . In the model of [1 Section 2.1] , the observations are uniform in M σ \mathcal{M}_{\sigma} M σ , and we take their nearest points on M \mathcal{M} M as the clean points:
y 1 , … , y n sampled i.i.d. from Unif ( M σ ) , x i : = π ( y i ) , δ i : = y i − x i , (34) y_1,\ldots,y_n\text{ sampled i.i.d. from }\operatorname{Unif}(\mathcal{M}_{\sigma}),\qquad x_i:=\pi(y_i),\qquad\delta_i:=y_i-x_i, \tag*{(34)} y 1 , … , y n sampled i.i.d. from Unif ( M σ ) , x i := π ( y i ) , δ i := y i − x i , (34)
where Unif ( M σ ) \operatorname{Unif}(\mathcal{M}_{\sigma}) Unif ( M σ ) is Lebesgue measure on the tube normalized to a probability. Only the y i y_i y i are observed. Throughout this appendix M \mathcal{M} M and σ \sigma σ are fixed, and only the numbers of observations grow.
For x ∈ M x\in\mathcal{M} x ∈ M , let N x M N_x\mathcal{M} N x M be the normal space of M \mathcal{M} M at x x x . For v ∈ N x M v\in N_x\mathcal{M} v ∈ N x M , the shape operator S v ( x ) S_v(x) S v ( x ) is the symmetric linear map of the tangent space T x M T_x\mathcal{M} T x M with ⟨ S v ( x ) u , w ⟩ = ⟨ I I x ( u , w ) , v ⟩ \langle S_v(x)u,w\rangle=\langle\mathrm{II}_x(u,w),v\rangle ⟨ S v ( x ) u , w ⟩ = ⟨ II x ( u , w ) , v ⟩ , where I I \mathrm{II} II is the second fundamental form. It measures how M \mathcal{M} M curves in the direction v v v , and it enters the density of the clean points through the change of variables y = x + v y=x+v y = x + v .
Lemma 37. Let 0 < σ < τ M 0<\sigma<\tau_M 0 < σ < τ M and let y y y have distribution Unif ( M σ ) \operatorname{Unif}(\mathcal{M}_{\sigma}) Unif ( M σ ) . Then x = π ( y ) x=\pi(y) x = π ( y ) has distribution μ σ = ρ σ d Vol M \mu_{\sigma}=\rho_{\sigma}\,\mathrm{d}\operatorname{Vol}_{\mathcal{M}} μ σ = ρ σ d Vol M , with density
ρ σ ( x ) = 1 ∣ M σ ∣ ∫ { v ∈ N x M : ∥ v ∥ < σ } det ( I − S v ( x ) ) d v , (35) \rho_{\sigma}(x)=\frac{1}{|\mathcal{M}_{\sigma}|}\int_{\{v\in N_x\mathcal{M}:\|v\|<\sigma\}}\det(I-S_v(x))\,\mathrm{d}v, \tag*{(35)} ρ σ ( x ) = ∣ M σ ∣ 1 ∫ { v ∈ N x M : ∥ v ∥ < σ } det ( I − S v ( x )) d v , (35) where ∣ M σ ∣ |\mathcal{M}_{\sigma}| ∣ M σ ∣ is the Lebesgue measure of the tube and d v \mathrm{d}v d v is Lebesgue measure on N x M N_x\mathcal{M} N x M . The density ρ σ \rho_{\sigma} ρ σ is smooth and strictly positive on M \mathcal{M} M , hence bounded above and below by positive constants.
Proof. Since σ < τ M \sigma< \tau_{\mathcal{M}} σ < τ M , the map Ψ ( x , v ) = x + v \Psi(x,v)=x+v Ψ ( x , v ) = x + v , defined for x ∈ M x\in\mathcal{M} x ∈ M and v ∈ N x M v\in N_x\mathcal{M} v ∈ N x M with ∥ v ∥ < σ \lVert v\rVert<\sigma ∥ v ∥ < σ , parametrizes M σ \mathcal{M}_\sigma M σ , and π ( Ψ ( x , v ) ) = x \pi(\Psi(x,v))=x π ( Ψ ( x , v )) = x [15 Theorem 4.8(12)] . By the Weingarten formula [9 (3.3)] , the derivative of a normal vector field with value v v v at x x x in a tangent direction u u u has tangential part − S v ( x ) u -S_v(x)u − S v ( x ) u . Hence the Jacobian determinant of Ψ \Psi Ψ with respect to d Vol M ( x ) d v \mathrm{d}\operatorname{Vol}_{\mathcal{M}}(x)\,\mathrm{d}v d Vol M ( x ) d v is det ( I − S v ( x ) ) \det(I-S_v(x)) det ( I − S v ( x )) , as in the first step of the proof of Weyl’s tube formula [6 Section 2] . It follows from [15 Theorem 4.8(7)] that ∥ S v ( x ) ∥ o p ≤ ∥ v ∥ / τ M \lVert S_v(x)\rVert_{\mathrm{op}}\leq\lVert v\rVert/\tau_{\mathcal{M}} ∥ S v ( x ) ∥ op ≤ ∥ v ∥ / τ M , so
det ( I − S v ( x ) ) ≥ ( 1 − σ / τ M ) m > 0. \det(I-S_v(x))\geq(1-\sigma/\tau_{\mathcal{M}})^m>0. det ( I − S v ( x )) ≥ ( 1 − σ / τ M ) m > 0.
The change of variables y = Ψ ( x , v ) y=\Psi(x,v) y = Ψ ( x , v ) gives (35) . In a local orthonormal frame of the normal bundle that depends smoothly on x x x , (35) integrates a function that is smooth in ( x , v ) (x,v) ( x , v ) over a fixed ball. Hence ρ σ \rho_\sigma ρ σ is smooth, and it is positive because the integrand is. Since M \mathcal{M} M is compact, ρ σ \rho_\sigma ρ σ is bounded above and below by positive constants.
Lemma 38. Under (34), the clean points x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n are i.i.d. with distribution μ σ \mu_\sigma μ σ , and ∥ δ i ∥ < σ \lVert\delta_i\rVert<\sigma ∥ δ i ∥ < σ for every i i i . Thus the hypotheses of Section 2 on the manifold, the sampling density and the displacements hold with ρ = ρ σ \rho=\rho_\sigma ρ = ρ σ and δ = σ \delta=\sigma δ = σ .
Proof. Each clean point x i = π ( y i ) x_i=\pi(y_i) x i = π ( y i ) is the image of y i y_i y i alone under the same measurable map π \pi π . Since the observations are i.i.d., so are the clean points, with common distribution μ σ \mu_\sigma μ σ by Lemma 37. That lemma also gives the smoothness of ρ σ \rho_\sigma ρ σ and its bounds. Finally, ∥ δ i ∥ = dist ( y i , M ) < σ \lVert\delta_i\rVert=\operatorname{dist}(y_i,\mathcal{M})<\sigma ∥ δ i ∥ = dist ( y i , M ) < σ because y i ∈ M σ y_i\in\mathcal{M}_\sigma y i ∈ M σ .
The continuum operator is Δ ρ σ \Delta_{\rho_\sigma} Δ ρ σ , defined as in Section 2 with ρ = ρ σ \rho=\rho_\sigma ρ = ρ σ . Although the observations are uniform in the tube, ρ σ \rho_\sigma ρ σ need not be constant when m ≥ 2 m\geq2 m ≥ 2 . For a surface in R 3 \mathbb{R}^3 R 3 , for example, (35) gives ρ σ ( x ) = ( 2 σ + 2 3 σ 3 K ( x ) ) / ∣ M σ ∣ \rho_\sigma(x)=(2\sigma+\frac{2}{3}\sigma^3K(x))/\lvert\mathcal{M}_\sigma\rvert ρ σ ( x ) = ( 2 σ + 3 2 σ 3 K ( x )) / ∣ M σ ∣ , where K K K is the Gauss curvature of M \mathcal{M} M . For m = 1 m=1 m = 1 , ρ σ \rho_\sigma ρ σ is constant: the tangent space is one-dimensional, so det ( I − S v ( x ) ) − 1 \det(I-S_v(x))-1 det ( I − S v ( x )) − 1 is linear in v v v and integrates to zero over the ball.
The projection algorithm and the residual
For k ≥ 2 k\geq2 k ≥ 2 , q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , a query point y ∈ M σ y\in\mathcal{M}_\sigma y ∈ M σ and a dataset R R R of N N N points of M σ \mathcal{M}_\sigma M σ , let A k , q ( y ; R ) A_{k,q}(y;R) A k , q ( y ; R ) denote the output of Algorithms 1 and 2 of [1 ] , in the form analysed in their theorems. It is an estimate of π ( y ) \pi(y) π ( y ) , obtained by fitting polynomials of degree at most k − 1 k-1 k − 1 to the points of R R R near y y y , over an estimate of the tangent space of M \mathcal{M} M at π ( y ) \pi(y) π ( y ) that is refined at each iteration. As in their analysis, the fits use the robust regression and derivative estimates of [1 (12)] , and the number of iterations is that of [1 Lemma 27] , which depends on N N N and q q q . The algorithm also needs m m m , σ \sigma σ and a suitable lower bound on the reach.
We take A k , q ( y ; R ) A_{k,q}(y;R) A k , q ( y ; R ) to be a single point of R d \mathbb{R}^d R d for every y y y and R R R , and we assume that it is a measurable function of the pair ( y , R ) (y,R) ( y , R ) . The proof of Proposition 41 uses this to integrate over a random query.
Theorem 39 ([1 Theorems 1 and 3] ). Assume that M ∈ C k \mathcal{M}\in C^k M ∈ C k is a compact m m m -dimensional submanifold of R d \mathbb{R}^d R d without boundary, that τ M / σ > C τ d log d \tau_{\mathcal{M}}/\sigma>C_\tau\sqrt{d\log d} τ M / σ > C τ d log d for the constant C τ C_\tau C τ of the cited theorem, and that R R R consists of N N N data points sampled i.i.d. from Unif ( M σ ) \operatorname{Unif}(\mathcal{M}_\sigma) Unif ( M σ ) . Let y ∈ M σ y\in\mathcal{M}_\sigma y ∈ M σ . There are a constant C 2 C_2 C 2 and, for each q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , a threshold N 0 ( y , q ) N_0(y,q) N 0 ( y , q ) such that for every N > N 0 ( y , q ) N>N_0(y,q) N > N 0 ( y , q ) , with probability at least 1 − q 1-q 1 − q ,
∥ A k , q ( y ; R ) − π ( y ) ∥ ≤ C 2 ln ( 1 / q ) ( N ( ln ln N ) 2 ) − r k , r k : = k − 1 2 k + m . (36) \left\lVert A_{k,q}(y;R)-\pi(y)\right\rVert\leq C_2\ln(1/q)\left(\frac{N}{(\ln\ln N)^2}\right)^{-r_k},\qquad r_k:=\frac{k-1}{2k+m}. \tag*{(36)} ∥ A k , q ( y ; R ) − π ( y ) ∥ ≤ C 2 ln ( 1/ q ) ( ( ln ln N ) 2 N ) − r k , r k := 2 k + m k − 1 . (36)
Aizenbud and Sober [1 Theorem 1] also bound the distance from A k , q ( y ; R ) \mathcal{A}_{k,q}(y;R) A k , q ( y ; R ) to M \mathcal{M} M , with the larger exponent k / ( 2 k + m ) k/(2k+m) k / ( 2 k + m ) . We use the bound (36) on the distance to π ( y ) \pi(y) π ( y ) , because the results of § 3 need each denoised point to be close to its own clean point. Since M \mathcal{M} M is smooth, it is C k \mathcal{C}^{k} C k for every k k k , so k ≥ 2 k \ge2 k ≥ 2 is a free parameter of the algorithm. Enlarging N 0 ( y , q ) N_{0}(y,q) N 0 ( y , q ) if necessary, we assume that N 0 ( y , q ) ≥ 3 N_{0}(y,q) \ge3 N 0 ( y , q ) ≥ 3 , so that ln ln N > 0 \ln\ln N > 0 ln ln N > 0 in (36) .
In Theorem 39 the query point y y y is fixed, and the probability is over the dataset R R R . To apply the theorem at the observations y i y_i y i , which are random, we run the algorithm on an additional dataset, independent of the observations.
Definition 40 (Denoiser and denoised Laplacian). Under (34), fix k ≥ 2 k \ge2 k ≥ 2 , q ∈ ( 0 , 1 ) q \in(0,1) q ∈ ( 0 , 1 ) and an integer N ≥ 1 N \ge1 N ≥ 1 . Let the additional dataset R N R_N R N consist of N N N data points sampled i.i.d. from Unif ( M σ ) \operatorname{Unif}(\mathcal{M}_{\sigma}) Unif ( M σ ) , independently of y 1 , … , y n y_1,\ldots,y_n y 1 , … , y n , and set
x ^ i : = A k , q ( y i ; R N ) , i = 1 , … , n . (37) \hat{x}_i := \mathcal{A}_{k,q}(y_i;R_N), \qquad i=1,\ldots,n. \tag*{(37)} x ^ i := A k , q ( y i ; R N ) , i = 1 , … , n . (37) All n n n queries use the same additional dataset. The residual
δ ^ : = max 1 ≤ i ≤ n ∥ x ^ i − x i ∥ (38) \hat{\delta} := \max_{1 \le i \le n} \lVert\hat{x}_i-x_i \rVert \tag*{(38)} δ ^ := 1 ≤ i ≤ n max ∥ x ^ i − x i ∥ (38) is the largest distance from a denoised point to its clean point. Let L ^ \hat{L} L ^ be the graph Laplacian of § 2 built on x ^ 1 , … , x ^ n \hat{x}_1,\ldots,\hat{x}_n x ^ 1 , … , x ^ n , let λ ^ 1 ≤ ⋯ ≤ λ ^ n \hat{\lambda}_1 \le\cdots\le\hat{\lambda}_n λ ^ 1 ≤ ⋯ ≤ λ ^ n be its eigenvalues, and define P ^ l \hat{P}_l P ^ l from L ^ \hat{L} L ^ as P l P_l P l is defined from L L L in § 3 .
Proposition 41. Assume (34) with τ M / σ > C τ d log d \tau_{\mathcal{M}}/\sigma> C_{\tau}\sqrt{d\log d} τ M / σ > C τ d log d , and let k ≥ 2 k \ge2 k ≥ 2 and p ≥ 1 p \ge1 p ≥ 1 . For n ≥ 2 n \ge2 n ≥ 2 set q n : = n − ( p + 1 ) q_n := n^{-(p+1)} q n := n − ( p + 1 ) , and run the denoiser of Definition 40 with q = q n q=q_n q = q n . For every n ≥ 2 n \ge2 n ≥ 2 and t > 0 t>0 t > 0 there is an integer N 1 ( n , t ) N_1(n,t) N 1 ( n , t ) such that, for every size N ≥ N 1 ( n , t ) N \ge N_1(n,t) N ≥ N 1 ( n , t ) of the additional dataset,
P ( δ ^ > t ) ≤ 2 n − p , \mathbb{P}(\hat{\delta}>t) \le2n^{-p}, P ( δ ^ > t ) ≤ 2 n − p , where the probability is over the observations y 1 , … , y n y_1,\ldots,y_n y 1 , … , y n and the additional dataset R N R_N R N .
Proof. Fix n ≥ 2 n \ge2 n ≥ 2 and t > 0 t>0 t > 0 , and write ν : = Unif ( M σ ) \nu:= \operatorname{Unif}(\mathcal{M}_{\sigma}) ν := Unif ( M σ ) , so that each y i y_i y i has distribution ν \nu ν and R N R_N R N has distribution ν ⊗ N \nu^{\otimes N} ν ⊗ N , the product measure on M σ N \mathcal{M}_{\sigma}^{N} M σ N . We bound the probability that one distance ∥ x ^ i − x i ∥ \lVert\hat{x}_i-x_i\rVert ∥ x ^ i − x i ∥ exceeds t t t , first with the random query y i y_i y i replaced by a fixed query, using Theorem 39 , and then by integrating that bound against the distribution of y i y_i y i . A union bound over i i i then multiplies the bound by n n n , and n q n = n − p nq_n=n^{-p} n q n = n − p .
For N ≥ 1 N \ge1 N ≥ 1 , y ∈ M σ y \in\mathcal{M}_{\sigma} y ∈ M σ and R ∈ M σ N R \in\mathcal{M}_{\sigma}^{N} R ∈ M σ N , let
g N ( y , R ) : = ∥ A k , q n ( y ; R ) − π ( y ) ∥ g_N(y,R):=\lVert\mathcal{A}_{k,q_n}(y;R)-\pi(y)\rVert g N ( y , R ) := ∥ A k , q n ( y ; R ) − π ( y )∥
be the error of the algorithm at the query y y y with dataset R R R , and let
h N ( y ) : = ν ⊗ N ( { R : g N ( y , R ) > t } ) h_N(y):=\nu^{\otimes N}\left(\left\{R:g_N(y,R)>t\right\}\right) h N ( y ) := ν ⊗ N ( { R : g N ( y , R ) > t } )
be the probability that this error exceeds t t t when the query y y y is fixed. The function g N g_N g N is measurable on M σ × M σ N \mathcal{M}_{\sigma}\times\mathcal{M}_{\sigma}^{N} M σ × M σ N , because A k , q n \mathcal{A}_{k,q_n} A k , q n is jointly measurable and π \pi π is continuous [15 Theorem 4.8(4)] . By Tonelli’s theorem, h N h_N h N is therefore measurable on M σ \mathcal{M}_{\sigma} M σ , and 0 ≤ h N ≤ 1 0 \le h_N \le1 0 ≤ h N ≤ 1 .
First fix a query y ∈ M σ y \in\mathcal{M}_{\sigma} y ∈ M σ . Since ln ( 1 / q n ) = ( p + 1 ) ln n \ln(1/q_n)=(p+1)\ln n ln ( 1/ q n ) = ( p + 1 ) ln n , Theorem 39 with q = q n q=q_n q = q n gives, for every N > N 0 ( y , q n ) N>N_0(y,q_n) N > N 0 ( y , q n ) ,
g N ( y , R ) ≤ C 2 ( p + 1 ) ln n ( N ( ln ln N ) 2 ) − r k g_N(y,R) \le C_2(p+1)\ln n\left(\frac{N}{(\ln\ln N)^2}\right)^{-r_k} g N ( y , R ) ≤ C 2 ( p + 1 ) ln n ( ( ln ln N ) 2 N ) − r k
for all R R R outside a set of ν ⊗ N \nu^{\otimes N} ν ⊗ N -measure at most q n q_n q n . Since r k > 0 r_k > 0 r k > 0 , the right-hand side tends to zero as N → ∞ N \to\infty N → ∞ , so it is at most t t t for all large N N N . Hence h N ( y ) ≤ q n h_N(y) \le q_n h N ( y ) ≤ q n for all sufficiently large N N N .
Next we integrate over the query. By the previous paragraph, ( h N − q n ) + → 0 (h_N-q_n)_+ \to0 ( h N − q n ) + → 0 pointwise on M σ \mathcal{M}_\sigma M σ as N → ∞ N \to\infty N → ∞ . Since 0 ≤ ( h N − q n ) + ≤ 1 0 \le(h_N-q_n)_+ \le1 0 ≤ ( h N − q n ) + ≤ 1 and ν \nu ν is a probability measure, dominated convergence gives ∫ M σ ( h N − q n ) + d ν → 0 \int_{\mathcal{M}_\sigma}(h_N-q_n)_+\,d\nu\to0 ∫ M σ ( h N − q n ) + d ν → 0 . Hence there is an integer N 1 ( n , t ) N_1(n,t) N 1 ( n , t ) such that, for every N ≥ N 1 ( n , t ) N \ge N_1(n,t) N ≥ N 1 ( n , t ) ,
∫ M σ h N d ν ≤ q n + ∫ M σ ( h N − q n ) + d ν ≤ 2 q n . \int_{\mathcal{M}_\sigma} h_N\,d\nu\le q_n+\int_{\mathcal{M}_\sigma}(h_N-q_n)_+\,d\nu\le2q_n. ∫ M σ h N d ν ≤ q n + ∫ M σ ( h N − q n ) + d ν ≤ 2 q n .
Now let N ≥ N 1 ( n , t ) N \ge N_1(n,t) N ≥ N 1 ( n , t ) and fix i i i . By (34) and (37), ∥ x ^ i − x i ∥ = g N ( y i , R N ) \|\hat{x}_i-x_i\|=g_N(y_i,R_N) ∥ x ^ i − x i ∥ = g N ( y i , R N ) . Since R N R_N R N is independent of y i y_i y i , the pair ( y i , R N ) (y_i,R_N) ( y i , R N ) has the product distribution ν ⊗ ν ⊗ N \nu\otimes\nu^{\otimes N} ν ⊗ ν ⊗ N . Tonelli’s theorem, applied to the indicator of { ( y , R ) : g N ( y , R ) > t } \{(y,R):g_N(y,R)>t\} {( y , R ) : g N ( y , R ) > t } , therefore gives
P ( ∥ x ^ i − x i ∥ > t ) = ∫ M σ ∫ M σ N 1 { g N ( y , R ) > t } d ν ⊗ N ( R ) d ν ( y ) = ∫ M σ h N d ν ≤ 2 q n . \begin{aligned}
\mathbb{P}(\|\hat{x}_i-x_i\|>t)=\int_{\mathcal{M}_\sigma}\int_{\mathcal{M}_\sigma^N}\mathbf{1}_{\{g_N(y,R)>t\}}\,d\nu^{\otimes N}(R)\,d\nu(y) \\
&=\int_{\mathcal{M}_\sigma}h_N\,d\nu\le2q_n.
\end{aligned} P ( ∥ x ^ i − x i ∥ > t ) = ∫ M σ ∫ M σ N 1 { g N ( y , R ) > t } d ν ⊗ N ( R ) d ν ( y ) = ∫ M σ h N d ν ≤ 2 q n .
A union bound over i i i gives
P ( δ ^ > t ) ≤ ∑ i = 1 n P ( ∥ x ^ i − x i ∥ > t ) ≤ 2 n q n = 2 n − p . \mathbb{P}(\hat{\delta}>t)\le\sum_{i=1}^{n}\mathbb{P}(\|\hat{x}_i-x_i\|>t)\le2nq_n=2n^{-p}. P ( δ ^ > t ) ≤ i = 1 ∑ n P ( ∥ x ^ i − x i ∥ > t ) ≤ 2 n q n = 2 n − p .
The denoised spectrum
The displacements x ^ i − x i \hat{x}_i-x_i x ^ i − x i of the denoised points are random, and they depend on the observations and on the additional dataset. The results of § 3 still apply to them, because we assume only that the displacements are bounded by a deterministic δ \delta δ ; the displacements themselves may be random. Lemma 42 makes this precise.
Lemma 42. Let δ ≥ 0 \delta\ge0 δ ≥ 0 be deterministic, and assume the hypotheses of one of the results of § 3 . The proof of that result gives its bound given an event that depends only on the clean points x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n , namely E n \mathcal{E}_n E n for Theorem 1 and Corollary 2, E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj for Theorem 3 , H n , l \mathcal{H}_{n,l} H n , l for Theorem 4 and E n ( 2 ) \mathcal{E}_n^{(2)} E n ( 2 ) for Remark 5. Given that event, the bound holds, with the same constants, for the graph Laplacian of every cloud z 1 , … , z n ∈ R d z_1,\ldots,z_n\in\mathbb{R}^d z 1 , … , z n ∈ R d with ∥ z i − x i ∥ ≤ δ \|z_i-x_i\|\le\delta ∥ z i − x i ∥ ≤ δ for all i i i , and for every choice of its spectral projectors and unit eigenvectors.
Proof. In the proofs of the results of § 3 , we use the noisy points only through Lemma 8 , which holds for every cloud as in the statement. Everything else we use, namely L ± L^\pm L ± , the comparison operators A A A , B B B , P l + P_l^+ P l + , τ l \tau_l τ l , Γ l + \Gamma_l^+ Γ l + and the events, is built from x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n , ε \varepsilon ε and δ \delta δ alone, so the events depend only on the clean points. Lemma 28 holds for every symmetric M M M with A ≼ M ≼ B A\preccurlyeq M\preccurlyeq B A ≼ M ≼ B , and Corollary 31 for every choice of eigenbases. So each proof applies without change to the graph Laplacian of z 1 , … , z n z_1,\ldots,z_n z 1 , … , z n . □
It is therefore enough that, with high probability, the residual is at most a deterministic bound. We take this bound to be ε n 3 \varepsilon_n^3 ε n 3 , where ε n \varepsilon_n ε n is the bandwidth, since this is the size of displacement at which Theorem 4 gives the eigenvector rate O ( ε n ) O(\varepsilon_n) O ( ε n ) of [7 Theorem 2.7(2)] . Proposition 41 makes the residual at most ε n 3 \varepsilon_n^3 ε n 3 with probability at least 1 − 2 n − p 1-2n^{-p} 1 − 2 n − p once the additional dataset is large enough.
Theorem 43. Assume the hypotheses of Proposition 41, and run the denoiser of Definition 40 with q = q n q=q_n q = q n , as there. Let η \eta η satisfy the hypotheses of § 2 , and fix l ≥ 1 l\geq1 l ≥ 1 . Use the notation and constants of § 2 and § 3 for the density ρ σ \rho_\sigma ρ σ , and build L ^ \widehat{L} L at ϵ n = C ϵ ( log n / n ) 1 / ( m + 4 ) \epsilon_n=C_\epsilon(\log n/n)^{1/(m+4)} ϵ n = C ϵ ( log n / n ) 1/ ( m + 4 ) , where C ϵ C_\epsilon C ϵ is large enough depending on p , l , M , ρ σ p,l,\mathcal{M},\rho_\sigma p , l , M , ρ σ and η \eta η , as in § 3 . For each n ≥ 2 n\geq2 n ≥ 2 , let the size N ( n ) N(n) N ( n ) of the additional dataset be deterministic, with
N ( n ) ≥ N 1 ( n , ϵ n 3 ) (39) N(n)\geq N_1(n,\epsilon_n^3) \tag*{(39)} N ( n ) ≥ N 1 ( n , ϵ n 3 ) (39) for the size N 1 N_1 N 1 of Proposition 41. Then, for n n n large enough, each of the following holds with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) , over the observations and the additional dataset.
The l l l -th eigenvalues satisfy
∣ λ l − λ ^ l ∣ ≤ K l ϵ n 2 , ∣ λ ^ l − σ η λ l M ∣ ≤ K l ϵ n 2 + C l ′ ϵ n . (40) \lvert\lambda_l-\widehat{\lambda}_l\rvert\leq K_l\epsilon_n^2,\qquad\lvert\widehat{\lambda}_l-\sigma_\eta\lambda_l^{\mathcal{M}}\rvert\leq K_l\epsilon_n^2+C_l'\epsilon_n. \tag*{(40)} ∣ λ l − λ l ∣ ≤ K l ϵ n 2 , ∣ λ l − σ η λ l M ∣ ≤ K l ϵ n 2 + C l ′ ϵ n . (40) If γ l M > 0 \gamma_l^{\mathcal{M}}>0 γ l M > 0 , then for every choice of P l P_l P l and P ^ l \widehat{P}_l P l ,
∥ P l − P ^ l ∥ H S , n ≤ D l ϵ n . (41) \lVert P_l-\widehat{P}_l\rVert_{\mathrm{HS},n}\leq D_l\epsilon_n. \tag*{(41)} ∥ P l − P l ∥ HS , n ≤ D l ϵ n . (41) If l ≥ 2 l\geq2 l ≥ 2 , then every eigenvector u ^ l \widehat{u}_l u l of L ^ \widehat{L} L with eigenvalue λ ^ l \widehat{\lambda}_l λ l and ∥ u ^ l ∥ n = 1 \lVert\widehat{u}_l\rVert_n=1 ∥ u l ∥ n = 1 admits an eigenfunction f l f_l f l of Δ ρ σ \Delta_{\rho_\sigma} Δ ρ σ with eigenvalue λ l M \lambda_l^{\mathcal{M}} λ l M and ∥ f l ∥ L 2 ( μ σ ) = 1 \lVert f_l\rVert_{L^2(\mu_\sigma)}=1 ∥ f l ∥ L 2 ( μ σ ) = 1 such that
∥ u ^ l − f l ∣ X n ∥ n = O ( ϵ n ) . (42) \lVert\widehat{u}_l-f_l|_{X_n}\rVert_n=O(\epsilon_n). \tag*{(42)} ∥ u l − f l ∣ X n ∥ n = O ( ϵ n ) . (42) Here λ l \lambda_l λ l and P l P_l P l belong to the Laplacian L L L built on the clean points x i = π ( y i ) x_i=\pi(y_i) x i = π ( y i ) , and f l ∣ X n f_l|_{X_n} f l ∣ X n is the vector with entries f l ( x i ) f_l(x_i) f l ( x i ) . The clean points are not observed.
Proof. By Lemma 38 , the clean points satisfy the hypotheses of § 2 with ρ = ρ σ \rho=\rho_\sigma ρ = ρ σ . We apply the results of § 3 with the deterministic displacement bound δ n : = ϵ n 3 \delta_n:=\epsilon_n^3 δ n := ϵ n 3 . Since δ n / ϵ n = ϵ n 2 → 0 \delta_n/\epsilon_n=\epsilon_n^2\to0 δ n / ϵ n = ϵ n 2 → 0 , the condition of § 3 that δ / ϵ \delta/\epsilon δ / ϵ be sufficiently small holds at δ = δ n \delta=\delta_n δ = δ n for n n n large enough. For each item, let G n \mathcal{G}_n G n be the event that the proofs of the corresponding results use at δ = δ n \delta=\delta_n δ = δ n , namely E n \mathcal{E}_n E n , E n p r o j \mathcal{E}_n^{\mathrm{proj}} E n proj and H n , l \mathcal{H}_{n,l} H n , l of § 5 and § 6 for items 1, 2 and 3.
First we bound the probabilities. The event G n \mathcal{G}_n G n depends only on the clean points (Lemma 42 ), which are i.i.d. with distribution μ σ \mu_\sigma μ σ (Lemma 38 ). The probability bounds proved in those sections therefore apply and give P ( G n ) ≥ 1 − O ( n − p ) \mathbb{P}(\mathcal{G}_n)\geq1-O(n^{-p}) P ( G n ) ≥ 1 − O ( n − p ) . By (39) , Proposition 41 with t = δ n t=\delta_n t = δ n gives P ( δ ^ > δ n ) ≤ 2 n − p \mathbb{P}(\widehat{\delta}>\delta_n)\leq2n^{-p} P ( δ > δ n ) ≤ 2 n − p . A union bound then gives probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) for G n ∩ { δ ^ ≤ δ n } \mathcal{G}_n\cap\{\widehat{\delta}\leq\delta_n\} G n ∩ { δ ≤ δ n } .
Given { δ ^ ≤ δ n } \{\widehat{\delta}\leq\delta_n\} { δ ≤ δ n } , every denoised point satisfies ∥ x ^ i − x i ∥ ≤ δ n \lVert\widehat{x}_i-x_i\rVert\leq\delta_n ∥ x i − x i ∥ ≤ δ n . By Lemma 42 , the bounds of Theorem 1 , Corollary 2, Theorem 3 and Theorem 4 at δ = δ n \delta=\delta_n δ = δ n therefore hold for L ^ \widehat{L} L given G n ∩ { δ ^ ≤ δ n } \mathcal{G}_n\cap\{\widehat{\delta}\leq\delta_n\} G n ∩ { δ ≤ δ n } . These bounds are K l δ n / ϵ n K_l\delta_n/\epsilon_n K l δ n / ϵ n , K l δ n / ϵ n + C l ′ ϵ n K_l\delta_n/\epsilon_n+C_l'\epsilon_n K l δ n / ϵ n + C l ′ ϵ n , D l δ n / ϵ n D_l\sqrt{\delta_n/\epsilon_n} D l δ n / ϵ n and O ( δ n / ϵ n + ϵ n ) O(\sqrt{\delta_n/\epsilon_n}+\epsilon_n) O ( δ n / ϵ n + ϵ n ) , and substituting δ n / ϵ n = ϵ n 2 \delta_n/\epsilon_n=\epsilon_n^2 δ n / ϵ n = ϵ n 2 gives the asserted bounds. □
Item 3 excludes l = 1 l=1 l = 1 , as Theorem 4 does. For l = 1 l=1 l = 1 , the same argument with the event E n ( 2 ) \mathcal{E}_n^{(2)} E n ( 2 ) used for Remark 5 shows that, for n n n large enough and with probability at least 1 − O ( n − p ) 1-O(n^{-p}) 1 − O ( n − p ) , every eigenvector u ^ 1 \widehat{u}_1 u 1 of L ^ \widehat{L} L with eigenvalue λ ^ 1 \widehat{\lambda}_1 λ 1 and ∥ u ^ 1 ∥ n = 1 \lVert\widehat{u}_1\rVert_n=1 ∥ u 1 ∥ n = 1 is constant, so that ∥ u ^ 1 − f 1 ∣ X n ∥ n = 0 \lVert\widehat{u}_1-f_1|_{X_n}\rVert_n=0 ∥ u 1 − f 1 ∣ X n ∥ n = 0 for the matching constant eigenfunction f 1 f_1 f 1 .
Hypotheses of the Spectral Convergence Results of Calder and García Trillos
The constant C l ′ C'_l C l ′ and the event E n , 3 \mathcal{E}_{n,3} E n , 3 of Section 5.2 come from [7 Theorem 2.4(2)] , and Section 6.4 also uses the proof of [7 Theorem 2.7(2)] . We state the hypotheses of these results in our notation. They are of three kinds.
The manifold, the density and the kernel. [7 Section 2] assumes that M \mathcal{M} M is a compact, connected, orientable, smooth m m m -dimensional submanifold of R d \mathbb{R}^{d} R d , that ρ ∈ C 2 , ϑ ( M ) \rho\in C^{2,\vartheta}(\mathcal{M}) ρ ∈ C 2 , ϑ ( M ) with 0 < ρ min ≤ ρ ≤ ρ max 0 < \rho_{\min} \le\rho\le\rho_{\max} 0 < ρ m i n ≤ ρ ≤ ρ m a x , and that η \eta η is non-increasing, supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] , Lipschitz on [ 0 , 1 ] [0,1] [ 0 , 1 ] and positive at 1 2 \frac{1}{2} 2 1 . Section 2 assumes all of this.
Deterministic restrictions on the parameters. Besides ε \varepsilon ε , the results of [7 ] involve two parameters θ , δ ε > 0 \theta, \delta_{\varepsilon} > 0 θ , δ ε > 0 that do not enter the definition of L L L . The first controls the error in approximating the density, the second the transport distance from that approximation to the empirical measure; the event on which both are controlled is described below. Here and below, C , c > 0 C,c > 0 C , c > 0 denote constants of [7 ] ; they do not depend on n , ε , θ n,\varepsilon,\theta n , ε , θ or δ ε \delta_{\varepsilon} δ ε . The restrictions of [7 Assumption 2.3] are:
(A1) ε \varepsilon ε is small enough, and in particular 2 ε < min { 1 , ι 0 , K − 1 / 2 , R / 2 } 2\varepsilon< \min\{1,\iota_{0},K^{-1/2},R/2\} 2 ε < min { 1 , ι 0 , K − 1/2 , R /2 } , where ι 0 > 0 \iota_{0} > 0 ι 0 > 0 is a lower bound on the injectivity radius of M \mathcal{M} M , K > 0 K > 0 K > 0 an upper bound on the absolute values of its sectional curvatures, and R R R its reach;
(A2) δ ε ≤ ε / 4 \delta_{\varepsilon} \le\varepsilon/4 δ ε ≤ ε /4 ;
(A3) δ ε > n − 1 / m \delta_{\varepsilon} > n^{-1/m} δ ε > n − 1/ m ;
(A4) C ( θ + δ ε ) ≤ ρ min / 2 C(\theta+\delta_{\varepsilon}) \le\rho_{\min}/2 C ( θ + δ ε ) ≤ ρ m i n /2 , where C C C depends only on m m m , ρ \rho ρ and M \mathcal{M} M .
Each result also has a smallness condition at the eigenvalue indices it uses. For an index j j j , let g j : = min { ∣ λ k M − λ j M ∣ : λ k M ≠ λ j M } > 0 g_j := \min\{|\lambda_k^{\mathcal{M}}-\lambda_j^{\mathcal{M}}|:\lambda_k^{\mathcal{M}}\ne\lambda_j^{\mathcal{M}}\}>0 g j := min { ∣ λ k M − λ j M ∣ : λ k M = λ j M } > 0 be the distance from λ j M \lambda_j^{\mathcal{M}} λ j M to the other eigenvalues of Δ ρ \Delta_{\rho} Δ ρ , and let c j > 0 c_j>0 c j > 0 be the smaller of the thresholds that [7 Theorem 2.4(1) and (2)] give at index j j j .
(S1j _j j ) [7 Theorem 2.4(2)] at index j j j requires ( ε λ j M + δ ε / ε + ε + θ ) λ j M ≤ c j (\varepsilon\sqrt{\lambda_j^{\mathcal{M}}}+\delta_{\varepsilon}/\varepsilon+\varepsilon+\theta)\lambda_j^{\mathcal{M}}\le c_j ( ε λ j M + δ ε / ε + ε + θ ) λ j M ≤ c j .
(S2N _N N ) The proof of [7 Theorem 2.7(2)] fixes N ∈ N N\in\mathbb{N} N ∈ N and requires, before its equation (4.13), that for every j ≤ N j\le N j ≤ N the hypothesis ε λ j M + C ( θ + δ ε ) ≤ c j \varepsilon\sqrt{\lambda_j^{\mathcal{M}}}+C(\theta+\delta_{\varepsilon})\le c_j ε λ j M + C ( θ + δ ε ) ≤ c j of [7 Theorem 2.4(1)] holds and that the error bound of that theorem satisfies C ( ε λ j M + δ ε / ε + ε + θ ) λ j M ≤ 1 2 σ η g j C(\varepsilon\sqrt{\lambda_j^{\mathcal{M}}}+\delta_{\varepsilon}/\varepsilon+\varepsilon+\theta)\lambda_j^{\mathcal{M}}\le\frac{1}{2}\sigma_{\eta}g_j C ( ε λ j M + δ ε / ε + ε + θ ) λ j M ≤ 2 1 σ η g j .
The proof in [7 ] writes 1 2 g j \frac{1}{2}g_j 2 1 g j in place of 1 2 σ η g j \frac{1}{2}\sigma_{\eta}g_j 2 1 σ η g j and omits σ η \sigma_{\eta} σ η in the displays after (4.13). Restoring the factor changes only the constant C C C , and the argument needs it, since (4.13) compares λ j \lambda_j λ j with σ η λ j M \sigma_{\eta}\lambda_j^{\mathcal{M}} σ η λ j M .
Probabilistic events. The estimates of [7 ] are proved on the event of [7 Proposition 2.12] : under (A1)–(A4), with probability at least 1 − n exp ( − C n θ 2 δ ε m ) 1-n\exp(-Cn\theta^{2}\delta_{\varepsilon}^{m}) 1 − n exp ( − C n θ 2 δ ε m ) there are a probability density ρ ~ n \widetilde{\rho}_n ρ n on M \mathcal{M} M with ∥ ρ − ρ ~ n ∥ L ∞ ≤ C ( θ + δ ε ) \|\rho-\widetilde{\rho}_n\|_{L^\infty}\le C(\theta+\delta_{\varepsilon}) ∥ ρ − ρ n ∥ L ∞ ≤ C ( θ + δ ε ) and a map T : M → X n T:\mathcal{M}\to X_n T : M → X n that pushes ρ ~ n d Vol M \widetilde{\rho}_n\,d\operatorname{Vol}_{\mathcal{M}} ρ n d Vol M forward to the empirical measure of X n X_n X n , with d M ( x , T ( x ) ) ≤ δ ε d_{\mathcal{M}}(x,T(x))\le\delta_{\varepsilon} d M ( x , T ( x )) ≤ δ ε for every x x x . The event E n , 3 \mathcal{E}_{n,3} E n , 3 , and the event of Lemma 32, are defined without θ \theta θ and δ ε \delta_{\varepsilon} δ ε . These parameters enter only the conditions under which [7 ] bounds the failure probabilities. Under (A1)–(A4) and (S1j _j j ), [7 Theorem 2.4(2)] gives
P n ( E n , 3 c ) ≤ 2 C n exp ( − c n ε m + 4 ) + C n exp ( − c n θ 2 δ ε m ) , \mathbb{P}_n(\mathcal{E}_{n,3}^{c})\le2Cn\exp(-cn\varepsilon^{m+4})+Cn\exp(-cn\theta^{2}\delta_{\varepsilon}^{m}), P n ( E n , 3 c ) ≤ 2 C n exp ( − c n ε m + 4 ) + C n exp ( − c n θ 2 δ ε m ) ,
and Lemma 32 gives a bound of the same form under (A1)–(A4) and (S2N _N N ). Lemma 44 chooses θ \theta θ and δ ε \delta_{\varepsilon} δ ε so that these conditions hold for finitely many indices at once and both terms are O ( n − p ) O(n^{-p}) O ( n − p ) : the term with exp ( − c n θ 2 δ e m ) \exp(-cn\theta^2\delta_e^m) exp ( − c n θ 2 δ e m ) for every C ε > 0 C_\varepsilon> 0 C ε > 0 once n n n is large enough, and the term with exp ( − c n ε m + 4 ) \exp(-cn\varepsilon^{m+4}) exp ( − c n ε m + 4 ) once C ε C_\varepsilon C ε is large enough.
Lemma 44. Fix N ≥ 1 N \ge1 N ≥ 1 and p ≥ 1 p \ge1 p ≥ 1 , and let ε = C ε ( log n / n ) 1 / ( m + 4 ) \varepsilon= C_\varepsilon(\log n/n)^{1/(m+4)} ε = C ε ( log n / n ) 1/ ( m + 4 ) with C ε > 0 C_\varepsilon> 0 C ε > 0 . There is a θ N ∈ ( 0 , 1 4 ] \theta_N \in(0,\frac{1}{4}] θ N ∈ ( 0 , 4 1 ] , depending only on N , M , ρ N,\mathcal{M},\rho N , M , ρ and η \eta η , such that the choice
θ : = θ N , δ e : = θ N ε \theta:= \theta_N,\qquad\delta_e := \theta_N\varepsilon θ := θ N , δ e := θ N ε satisfies ( A 1 ) (A1) ( A 1 ) –( A 4 ) (A4) ( A 4 ) , ( S 1 j ) (S1_j) ( S 1 j ) for every j ≤ N j \le N j ≤ N , and ( S 2 N ) (S2_N) ( S 2 N ) , for all n n n large enough. Moreover, for every constant c > 0 c > 0 c > 0 , n exp ( − c n θ 2 δ e m ) ≤ n − p n\exp(-cn\theta^2\delta_e^m) \le n^{-p} n exp ( − c n θ 2 δ e m ) ≤ n − p for all n n n large enough, and n exp ( − c n ε m + 4 ) ≤ n − p n\exp(-cn\varepsilon^{m+4}) \le n^{-p} n exp ( − c n ε m + 4 ) ≤ n − p for all n n n if c C ε m + 4 ≥ p + 1 cC_\varepsilon^{m+4} \ge p+1 c C ε m + 4 ≥ p + 1 .
Proof. Each inequality in ( A 4 ) (A4) ( A 4 ) , ( S 1 j ) (S1_j) ( S 1 j ) and ( S 2 N ) (S2_N) ( S 2 N ) , for j ≤ N j \le N j ≤ N , bounds a linear combination of ε \varepsilon ε , θ \theta θ , δ e / ε \delta_e/\varepsilon δ e / ε and δ e \delta_e δ e with nonnegative coefficients by a positive number, such as c j c_j c j or 1 2 σ η g j \frac{1}{2}\sigma_{\eta}g_j 2 1 σ η g j . The coefficients and these numbers depend only on j j j , M \mathcal{M} M , ρ \rho ρ and η \eta η , and there are finitely many such inequalities. Choose θ N ∈ ( 0 , 1 4 ] \theta_N \in(0,\frac{1}{4}] θ N ∈ ( 0 , 4 1 ] so small that in each of them, with θ = δ e / ε = θ N \theta= \delta_e/\varepsilon= \theta_N θ = δ e / ε = θ N , the terms in θ \theta θ and δ e / ε \delta_e/\varepsilon δ e / ε add up to at most half of the right-hand side. The remaining terms are multiples of ε \varepsilon ε and δ e = θ N ε \delta_e = \theta_N\varepsilon δ e = θ N ε , which tend to zero as n → ∞ n \to\infty n → ∞ , so each inequality holds for n n n large. So does ( A 1 ) (A1) ( A 1 ) , since ε → 0 \varepsilon\to0 ε → 0 . ( A 2 ) (A2) ( A 2 ) holds since θ N ≤ 1 4 \theta_N \le\frac{1}{4} θ N ≤ 4 1 , and ( A 3 ) (A3) ( A 3 ) holds for n n n large since
n 1 / m δ e = θ N C ε ( log n ) 1 / ( m + 4 ) n 1 / m − 1 / ( m + 4 ) → ∞ . n^{1/m}\delta_e = \theta_N C_\varepsilon(\log n)^{1/(m+4)}n^{1/m-1/(m+4)} \to\infty. n 1/ m δ e = θ N C ε ( log n ) 1/ ( m + 4 ) n 1/ m − 1/ ( m + 4 ) → ∞.
For the first probability bound, n θ 2 δ e m = θ N m + 2 C ε m n 4 / ( m + 4 ) ( log n ) m / ( m + 4 ) n\theta^2\delta_e^m = \theta_N^{m+2}C_\varepsilon^m n^{4/(m+4)}(\log n)^{m/(m+4)} n θ 2 δ e m = θ N m + 2 C ε m n 4/ ( m + 4 ) ( log n ) m / ( m + 4 ) , which is at least c − 1 ( p + 1 ) log n c^{-1}(p+1)\log n c − 1 ( p + 1 ) log n for n n n large, and then n exp ( − c n θ 2 δ e m ) ≤ n ⋅ n − ( p + 1 ) = n − p n\exp(-cn\theta^2\delta_e^m) \le n\cdot n^{-(p+1)} = n^{-p} n exp ( − c n θ 2 δ e m ) ≤ n ⋅ n − ( p + 1 ) = n − p . For the second, n ε m + 4 = C ε m + 4 log n n\varepsilon^{m+4} = C_\varepsilon^{m+4}\log n n ε m + 4 = C ε m + 4 log n , so n exp ( − c n ε m + 4 ) = n 1 − c C ε m + 4 ≤ n − p n\exp(-cn\varepsilon^{m+4}) = n^{1-cC_\varepsilon^{m+4}} \le n^{-p} n exp ( − c n ε m + 4 ) = n 1 − c C ε m + 4 ≤ n − p . □ \square □
By Lemma 44 with N = l N=l N = l , the hypotheses of [7 Theorem 2.4(2)] at index l l l hold for n n n large, and both terms in the bound on P n ( E n , 3 c ) \mathbb{P}_n(\mathcal{E}_{n,3}^c) P n ( E n , 3 c ) are at most 2 C n − p 2Cn^{-p} 2 C n − p for n n n large once c C ε m + 4 ≥ p + 1 cC_\varepsilon^{m+4} \ge p+1 c C ε m + 4 ≥ p + 1 . Hence, for C ε C_\varepsilon C ε large enough, P n ( E n , 3 ) ≥ 1 − O ( n − p ) \mathbb{P}_n(\mathcal{E}_{n,3}) \ge1-O(n^{-p}) P n ( E n , 3 ) ≥ 1 − O ( n − p ) , which is the probability bound in (7). At several indices, each of these events has this probability bound, and Lemma 44, with N N N the largest index, gives one choice of θ \theta θ and δ e \delta_e δ e that serves all of them.