Spectral Projection Estimates Restricted to Uniformly Embedded Submanifolds Zhexing Zhang Abstract . Let M M be a manifold with nonpositive sectional curvature and bounded geometry, and let Σ \Sigma be a uniformly embedded submanifold of M . M. We estimate the L 2 ​ ( M ) → L q ​ ( Σ ) L^{2}(M)\to L^{q}(\Sigma) norm of a log \log -scale spectral projection operator. It is a generalization of result of Chen [ 7 ] to noncompact cases. We also prove sharp spectral projection estimates of spectral windows of any small size restricted to nontrapped geodesics on even asymptotically hyperbolic surfaces with bounded geometry and curvature pinched below 0. 1. Introduction Let (M,g) be a smooth n -dimensional boundaryless complete Riemannian manifold with nonpositive curvature and bounded geometry, and \Sigma be a k -dimensional smooth uniformly embedded submanifold on M . Denote \Delta_{g} the Laplace operator associated with the metric g , and denote P=\sqrt{\Delta_{g}} . Let \mathbf{1}_{I}(P) be the spectral projection operator on the spectral window I\subset\mathbb{R} . Let R_{\Sigma} be the restriction operator from M to \Sigma. Define (1) \displaystyle\mu(q)=\begin{cases}\frac{n-1}{2}-\frac{k}{q},&\text{ if }k\leq n-2,q\geq 2\text{ or }k=n-1,q\geq\frac{2n}{n-1},\\ \frac{n-1}{4}-\frac{k-1}{q},&\text{ if }k=n-1,q<\frac{2n}{n-1}.\end{cases} The first main result of this paper is Theorem 1 . Given any f\in L^{2}(M) , when k=n-1 and q>\frac{2n}{n-1}, or when k\leq n-2 and q>2 , (2) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+\log(\lambda)^{-1}]}(P)f||_{L^{q}(\Sigma)}\lesssim\frac{\lambda^{\mu(q)}}{(\log\lambda)^{1/2}}||f||_{L^{2}(M)}. Reznikov [ 14 ] investigated the spectral projection estimates restricted to curves on compact hyperbolic surfaces. Then, Burq, Gérard and Tzvetkov [ 5 ] proved the following spectral projection estimate. If M is an n -dimensional compact manifold, and \Sigma is a k -dimensional submanifold of M , then (3) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+1]}(P)f||_{L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}(\log\lambda)^{1/2}||f||_{L^{2}(M)} if k=n-2 and q=2 , or k=n-1 and q=\frac{2n}{n-1} . Meanwhile, we have (4) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+1]}(P)f||_{L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}||f||_{L^{2}(M)}, if q\geq 2 otherwise. Thereafter, Hu [ 12 ] proved that we may remove the (\log\lambda)^{1/2} in ( 3 ) when q=\frac{2n}{n-1} and k=n-1 . Chen [ 7 ] refined the unit band estimate in [ 5 ] to a \log -scale estimate on compact manifolds for q>2 when k\leq n-2 and q>\frac{2n}{n-1} when k=n-1 , i.e. (5) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+\log\lambda^{-1}]}(P)||_{L^{2}(M)\to L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}(\log\lambda)^{-1/2}. Our work generalizes Chen’s result to manifolds with bounded geometry and nonpositive sectional curvature. We are also interested in curves in Riemannian surfaces with nonpositive curvature. In [ 16 ] and [ 15 ] , Xi and Zhang proved the \log -scale spectral projection restricted to compact geodesics of compact hyperbolic surfaces. When M is a Riemannian surface with nonpositive curvature and \gamma is a compact geodesic, Chen and Sogge [ 9 ] proved that (6) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\log\lambda^{-1}]}(P)||_{L^{2}(M)\to L^{4}(\gamma)}=o(\lambda^{1/4}). Thereafter, Blair [ 2 ] showed (7) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\log\lambda^{-1}]}(P)||_{L^{2}(M)\to L^{4}(\gamma)}\lesssim\lambda^{1/4}(\log\lambda)^{-1/4}. We state the following results for history and perspectives. Let M be a compact congruence arithmetic hyperbolic surface, let \gamma be a compact geodesic and let \Psi_{\lambda} be an L^{2}(M) normalized Hecke-Maass form associated to the eigenvalue \lambda. Marshall [ 13 ] proved (8) \displaystyle||R_{\gamma}\Psi_{\lambda}||_{L^{2}(\gamma)}\lesssim\lambda^{3/14+\epsilon}, for any \epsilon>0 . Let M be a 3-dimensional compact congruence arithmetic hyperbolic space and let \Sigma be a totally geodesic surface of M and let \Psi_{\lambda} be an L^{2}(M) normalized Hecke-Maass form associated to the eigenvalue \lambda. Hou [ 11 ] proved that (9) \displaystyle||R_{\Sigma}\Psi_{\lambda}||_{L^{2}(\Sigma)}\lesssim\lambda^{1/4-1/1220+\epsilon}. In addition to manifolds with constant negative curvature, flat manifolds have also been investigated. Let \mathbb{T}^{n} be a flat torus of dimension n, and let \Sigma be a smooth hypersurface of \mathbb{T}^{n} . Let \Psi_{\lambda} be an L^{2}(M) normalized Laplacian eigenfunction associated with the eigenvalue \lambda. When n=2,3, Bourgain and Rudnick [ 4 ] proved (10) \displaystyle||R_{\Sigma}\Psi_{\lambda}||_{L^{2}(\Sigma)}\sim 1. Our work mainly considers nontrapped geodesics of 2-dimensional even asymptotically hyperbolic manifolds. A 2-dimensional manifold, (M,g) , is a even asymptotically hyperbolic manifold, if there exists a compactification \overline{M} , which is a smooth manifold with boundary \partial M , and the metric near the boundary takes the form (11) \displaystyle g=\frac{dx_{1}^{2}+g_{1}(x_{1}^{2})}{x_{1}^{2}}, where x_{1}|_{\partial M}=0 , dx_{1}|_{\partial M}\neq 0 and g_{1}(x^{2}_{1}) is a smooth family of metrics on \partial M . Huang, Sogge, Tao and the author [ 6 ] proved the lossless spectral projection on any even asymptotically hyperbolic surface with curvature pinched below 0 with small spectral windows. Assume (M,g) is an even asymptotically hyperbolic surface with curvature pinched below 0. Following the idea of [ 6 ] , we may construct a simply connected asymptotically hyperbolic background manifold, (\tilde{M},\tilde{g}) , which agrees with (M,g) at infinity. We may use the kernel estimates of the spectral measure obtained by Chen and Hassell [ 8 ] on \tilde{M} to obtain spectral projection estimates with arbitrarily small spectral windows restricted to geodesics of \tilde{M} . We may also obtain a \log -scaled spectral projection estimate on M restricted to its compact geodesic segments derived from [ 2 ] . Meanwhile, we say that a geodesic \gamma in M is nontrapped, if \displaystyle\lim_{t\to\infty}\gamma(t)\to\infty\text{ and }\lim_{t\to\infty}\gamma(-t)\to\infty. We can combine these ingredients to prove the following lossless spectral projection estimate restricted to any nontrapped geodesic in M . Theorem 2 . Let (M,g) be an even asymptotically hyperbolic surface with curvature pinched below 0. Let q>2, \lambda\geq 1 and \eta\in(0,1] . Let \gamma be a nontrapped geodesic in M , then (12) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\eta]}(P)f||_{L^{q}(\gamma)}\lesssim{\lambda^{\mu(q)}}{\eta^{1/2}}||f||_{L^{2}(M)}. Some examples of even asymptotically hyperbolic surfaces with curvature pinched below 0 are convex cocompact hyperbolic surfaces. As stated in [ 3 ] , they are hyperbolic surfaces with finitely many funnels and no cusps. Anker, Germain and Léger [ 1 ] proved the lossless spectral projection with arbitrarily small spectral window on the hyperbolic surfaces satisfying the pressure condition, which are hyperbolic surfaces with limit sets of Hausdorff dimensions less than \frac{1}{2} . This inspires us to use the explicit kernel of spectral measure to prove the sharp spectral projection estimate with arbitrarily small spectral window restricted to nontrapped geodesics of a hyperbolic cylinder, as well as compact curves of hyperbolic surfaces satisfying the pressure condition. Section 2 discusses the uniformly embedded submanifold and proves Theorem 1 . Section 3 proves Theorem 2 . Section 5 gives some examples to illustrate the sharpness of the above two theorems. Notation . For any nonnegative quantity A and B, A\lesssim B and A=O(B) both mean A\leq cB for some constant c>0 only depending on the submanifold and its ambient manifold. We use A\sim B to denote A\lesssim B and B\lesssim A. 2. Theorem 1 2.1. Uniformly embedded submanifold In this subsection, we recall some properties of manifolds with bounded geometry and their uniformly embedded submanifolds. These results can be found in Chapter 2 of [ 10 ] . Definition 2.1 (Manifold with bounded geometry) . A manifold (M,g) is a manifold with bounded geometry, if: 1. The injectivity radius of M is positive. 2. The sectional curvature of M and its derivative of any order are uniformly bounded. There is a \delta(M)>0 so that the coordinate charts given by the exponential map at x in M , \exp^{M}_{x} , are defined on all geodesic balls B_{M}(x,\delta(M)) in M centered at x with radius \delta(M) . In addition, in the resulting normal coordinates, if we let d_{g} denote the Riemannian distance in (M,g) , then we have c|(\exp^{M}_{x})^{-1}(x_{1})-(\exp^{M}_{x})^{-1}(x_{2})|0, we use \Sigma(x,\delta) to denote the image under \iota of the connected component of x contained in B_{M}(\iota(x),\delta)\cap\iota(\Sigma) . We say that \Sigma is a uniformly embedded submanifold, if there exists a \varpi(\Sigma)>0 , such that for all x\in\Sigma , 1. the connected component \Sigma(x,\varpi(\Sigma)) is represented in normal coordinates on B_{M}(\iota(x),\varpi(\Sigma)) by the graph of a function h_{x}:T_{x}\Sigma\to N_{x} and the family of functions h has uniform continuity and boundedness estimates independent of x . 2. \Sigma(x,\varpi(\Sigma)) is the unique component of \Sigma\cap B_{M}(\iota(x),\varpi(\Sigma)) . Remark . If \Sigma satisfies condition 1 in Definition 2.2 , \Sigma is said to be uniformly immersed. We recall some useful facts of uniformly embedded submanifolds of manifolds with bounded geometry from [ 10 ] . Lemma 2.1 (Local equivalence of distance) . Let \Sigma be a uniformly immersed submanifold of the bounded geometry manifold (M,g) . Let d_{\Sigma} denote the distance function of \Sigma with the induced metric from (M,g). Then d_{g} and d_{\Sigma} are locally equivalent. In other words, for all c>1 , there exists a \nu_{c}>0 , such that for all d_{\Sigma}(x_{1},x_{2})<\nu_{c} , we have the local converse d_{\Sigma}(x_{1},x_{2})\leq cd_{g}(x_{1},x_{2}) . Lemma 2.2 (Uniformly locally finite cover of M ) . Let (M,g) be a Riemannian manifold of bounded geometry. Then for \delta(M)>0 small enough and any 0<\delta\leq\delta(M) , M has a countable covering \{B_{M}(x_{m},\delta)\}_{m\geq 1} such that 1. For all m\neq j , d_{g}(x_{m},x_{j})\geq\delta. 2. There exists an explicit global bound K\in\mathbb{N}, such that for each x\in\Sigma , \#\{m:B_{M}(x,\delta(M))\cap B_{M}(x_{m},\delta(M))\neq\emptyset\}\leq K. Lemma 2.3 (Submanifold of bounded geometry) . Let \Sigma be a uniformly embedded submanifold of the bounded geometry manifold (M,g) . Then, \Sigma with the induced metric is a Riemannian manifold with bounded geometry. We use the above facts to specify a covering on \Sigma. We aim to cover \Sigma by a locally finite covering \{A_{j}\} and cover a 1-neighborhood of \Sigma in M by a locally finite covering \{B_{j}\} , such that an 1-neighborhood of A_{j} is a subset B_{j} for each j. This would allow us to localize our problem. Proposition 2.1 . Let \Sigma , M be defined as above. We can fix a small \vartheta(\Sigma)>0 , such that there exists a covering, \{A_{j}\} , of \Sigma , and a covering, \{B_{j}\} , of \{x:d_{g}(x,\Sigma)\leq\vartheta(\Sigma)/4\} in M . In addition, \{B_{j}\} is uniformly locally finite in M and \{x\in\Sigma:d_{g}(x,A_{j})<\vartheta(\Sigma)/4\}\subset B_{j} for every j . Proof. Recall \varpi(\Sigma)>0 in Definition 2.2 and \delta(M)>0 in Lemma 2.2 . Choose \vartheta(\Sigma)=\frac{1}{2}\min\{\delta(M),\varpi(\Sigma)\} . By Lemma 2.2 , we can find a uniformly locally finite cover B_{M}(y_{m},\vartheta(\Sigma)/8) of M , such that B_{M}(y_{m},\vartheta(\Sigma)) is also a uniformly locally finite cover of M . For every m_{j} , such that B_{M}(y_{m_{j}},\vartheta(\Sigma)/8)\cap\Sigma\neq\emptyset , pick an x_{m_{j}}\in B_{M}(y_{m_{j}},\vartheta(\Sigma)/8)\cap\Sigma . By Definition 2.2 , we know that \Sigma(x_{m_{j}},\vartheta(\Sigma)/4) is the unique component of \Sigma in B_{M}(x_{m_{j}},\vartheta(\Sigma)/4) . Therefore, B_{M}(y_{m_{j}},\vartheta(\Sigma)/8)\cap\Sigma\subset B_{M}(x_{m_{j}},\vartheta(\Sigma)/4)\cap\Sigma=\Sigma(x_{m_{j}},\vartheta(\Sigma)/4). Since B_{M}(y_{m},\vartheta(\Sigma)/8) covers M , we know \{\Sigma(x_{m_{j}},\vartheta(\Sigma)/4)\}_{{j}\geq 1} covers \Sigma . We may choose (13) \displaystyle A_{j}:=\Sigma(x_{m_{j}},\vartheta(\Sigma)/2), so that \{A_{j}\} covers \Sigma. Then, we choose (14) \displaystyle B_{j}:=B_{M}(y_{m_{j}},\vartheta(\Sigma)). Notice that for each j, if x\in M and d_{g}(x,A_{j})<\vartheta(\Sigma)/4 , then d_{g}(x,y_{m_{j}})<\vartheta(\Sigma)/2+\vartheta(\Sigma)/4+\vartheta(\Sigma)/8 . So, \{x\in M:d_{g}(x,A_{j})<\vartheta(\Sigma)/4\}\subset B_{j} . Finally, \{B_{j}\}_{j\geq 1} is uniformly locally finite in M , since \{B_{M}(y_{m},\vartheta(\Sigma))\} is a uniformly locally finite cover of M . ∎ By the above proposition, we can choose a smooth partition of unity \{\psi_{j}\}_{j\geq 1} on \cup_{k}B_{j} and \{\phi_{j}\}_{j\geq 1} on \Sigma respectively. We require \sum_{j}\psi_{j} to be uniformly bounded, \psi_{j}\equiv 1 on \{x\in M:d_{g}(x,A_{j})<\vartheta(\Sigma)/8\} , and \text{supp}(\psi_{j})\subset B_{j} . We also require \sum_{j}\phi_{j}=1 and \text{supp}(\phi_{j})\subset A_{j} . Note that \{\psi_{j}\} is subordinate to \{B_{j}\} , and \{\phi_{j}\} is subordinate to \{A_{j}\} . 2.2. Unit band Projection Estimate To obtain the \log -scale spectral projection estimates, we need the unit band spectral projection estimates. Proposition 2.2 . Let M be an n -dimensional manifold with bounded geometry and non-positive secitonal curvature, and let \Sigma be a k -dimensional uniformly embedded submanifold of M. Let \mu(q) be defined as in ( 1 ). When q=\frac{2n}{n-1} and k=n-1 , or q=2 and k\leq n-2 , we have (15) \displaystyle||\mathbf{1}_{[\lambda,\lambda+1]}(P)||_{L^{2}(M)\to L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}(\log\lambda)^{1/2}. Meanwhile, if q\geq 2 otherwise, we have (16) \displaystyle||\mathbf{1}_{[\lambda,\lambda+1]}(P)||_{L^{2}(M)\to L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}. Proof. Recall the definition of \nu_{2} from Lemma 2.1 . For some small 0<\epsilon<\frac{1}{8}\min\{{\vartheta(\Sigma)},\nu_{2}\} , which we are going to choose later, there exist \rho\in\mathcal{S}(\mathbb{R}) satisfying (17) \displaystyle\rho(0)=1\text{ and }\hat{\rho}(t)=0\text{ if }t\notin\left[\frac{\epsilon}{2},{\epsilon}\right]. We define the local operator (18) \displaystyle\sigma_{\lambda}=\rho(\lambda-P)+\rho(\lambda+P). We know (19) \displaystyle\sigma_{\lambda}=\pi^{-1}\int^{\epsilon}_{0}\hat{\rho}(t)e^{i\lambda t}\cos(tP)dt. By finite propagation speed, \phi_{j}\sigma_{\lambda} is supported in a neighborhood of size \epsilon of A_{j}. Therefore, we may deduce (20) \displaystyle\text{supp}(\phi_{j}\circ R_{\Sigma}\circ\sigma_{\lambda})\subset\{x\in M:d_{g}(x,A_{j})<\vartheta(\Sigma)/8\}. We can follow the argument of Theorem 3 of [ 5 ] , (i.e. ( 3 ) and ( 4 )), to obtain (21) \displaystyle||\phi_{j}\circ R_{\Sigma}\circ\sigma_{\lambda}||_{L^{2}(B_{j})\to L^{q}(A_{j})}\leq C_{\Sigma}\log(\lambda)^{1/2}\lambda^{\mu(q)}, if k=n-2,n-1 and q=2 , or k=n-1 and q=\frac{2n}{n-1} . Meanwhile, if q\geq 2 otherwise, we could obtain (22) \displaystyle||\phi_{j}\circ R_{\Sigma}\circ\sigma_{\lambda}||_{L^{2}(B_{j})\to L^{q}(A_{j})}\leq C_{\Sigma}\lambda^{\mu(q)}. We sketch the proof here for the completeness. For some \tilde{x}\in A_{j} , let a coordinate chart on M be given by the exponential map \exp^{M}_{\tilde{x}}:\mathbb{R}^{n}\to M . By Theorem 4 of [ 5 ] , there exists \epsilon>0 , such that for any x\in\mathbb{R}^{n} with |x|\leq c\epsilon, (23) \displaystyle\sigma_{\lambda}(f)(x)=\lambda^{\frac{n-1}{2}}\int_{\mathbb{R}^{n}}e^{-i\lambda d_{g}(\exp^{M}_{\tilde{x}}(x),\exp^{M}_{\tilde{x}}(x^{\prime}))}a(x,x^{\prime})f(x^{\prime})dx^{\prime}+R(f)(x). with |\partial^{\alpha}_{x,x^{\prime}}a(x,x^{\prime})|=O(1) for all \alpha . In addition, a(x,x^{\prime}) is supported in \{|x|\leq c_{0}\epsilon\leq|x^{\prime}|\leq c_{1}\epsilon<1\} and does not vanish in d_{g}(\exp^{M}_{\tilde{x}}(x),\exp^{M}_{\tilde{x}}(x^{\prime}))\in[c_{2}\epsilon,c_{3}\epsilon]. Meanwhile, ||Rf||_{L^{\infty}}\lesssim||f||_{L^{2}} and the kernel of R is supported in \{(x,x^{\prime}):d_{g}(\exp^{M}_{\tilde{x}}(x),\exp^{M}_{\tilde{x}}(x^{\prime}))\leq\epsilon\} . The implicit constants and \epsilon are chosen to be independent of j by the bounded geometry condition. We may cover A_{j} by balls of size \epsilon . Since A_{j}\subset B_{j} , which are geodesic balls of uniformly bounded volume, the number of balls of size \epsilon needed to cover A_{j} is uniformly bounded. In addition, the intersection of each ball of size \epsilon and A_{j} contains a unique connected component, by condition 2 of Definition 2.2 . Thus, it suffices to consider the operator \mathcal{T} , such that for |x|\leq c\epsilon , (24) \displaystyle\mathcal{T}f(x)=\int_{\mathbb{R}^{n}}e^{-i\lambda d_{g}(\exp^{M}_{\tilde{x}}(x),\exp^{M}_{\tilde{x}}(x^{\prime}))}a(x,x^{\prime})f(x^{\prime})dx^{\prime}. Let \exp^{\Sigma}_{\tilde{x}}:(-c_{4}\epsilon,c_{4}\epsilon)^{k}\to\Sigma , the exponential map at \tilde{x} in \Sigma , be a coordinate chart on \Sigma. For |z|\leq c_{4}\epsilon, let x(z):=(\exp^{M}_{\tilde{x}})^{-1}\circ\iota\circ\exp^{\Sigma}_{\tilde{x}}(z). We denote x=x(z) and x^{\prime}=x(z^{\prime}). We define Tf(z)=\mathcal{T}f(x(z)) , and denote the kernel of T as K. Define \theta to be an even bump function with \theta\equiv 1 on [-1,1] and is supported in [-2,2], and define (25) \displaystyle\theta_{m}(\tau):=(\theta(2^{m}\tau)-\theta(2^{m+1}\tau)). Choose \theta , such that for \tau\leq\epsilon , (26) \displaystyle 1=\theta(\lambda\tau)+\sum_{m=\log\epsilon^{-1}/\log 2}^{\log\lambda/\log 2}\theta_{m}(\tau). Let K_{0}(z,z^{\prime})=\theta(\lambda|z-z^{\prime}|)K(z,z^{\prime}) and let (TT^{*})_{m} be the operator with kernel (27) \displaystyle K_{m}(z,z^{\prime})=\theta_{m}(|z-z^{\prime}|)K(z,z^{\prime}). By the support property of \theta , K_{0} is supported on |z-z^{\prime}|\leq\lambda^{-1} , and K_{0} is bounded by (28) \displaystyle|K_{0}(z,z^{\prime})|\leq C_{0}(1+\lambda|z-z^{\prime}|)^{-\frac{n-1}{2}}. Then, (29) \displaystyle\sup_{z}||K_{0}(z,z^{\prime})||_{L_{z^{\prime}}^{q/2}(\mathbb{R}^{k})}\leq C_{0}\lambda^{-\frac{2k}{q}}. Next, show as in Proposition 6.3 of [ 5 ] that we may choose an \epsilon>0 , such that for all \lambda^{-1}<2^{-m}\leq\epsilon, (30) \displaystyle||(TT^{*})_{m}||_{L^{1}\to L^{\infty}}\leq C_{1}\left(\frac{2^{m}}{\lambda}\right)^{\frac{n-1}{2}} (31) \displaystyle||(TT^{*})_{m}||_{L^{2}\to L^{2}}\leq C_{2}\left(\frac{2^{m}}{\lambda}\right)^{\frac{n-1}{2}+\frac{k-1}{2}}, which implies ( 22 ) and ( 21 ) by (6.7) of [ 5 ] . We now choose \epsilon>0 , such that for |z-z^{\prime}|\lesssim\epsilon , d_{\Sigma}((\exp^{M}_{\tilde{x}})^{-1}x,(\exp^{M}_{\tilde{x}})^{-1}x^{\prime})) is close enough to |x-x^{\prime}| in the C^{\infty}(\mathbb{R}^{n}\times\mathbb{R}^{n}) topology. By the bounded geometry assumption, we have uniform control of the metrics of M and \Sigma and all their derivatives. Thus, the choice of \epsilon is independent of j. Since M and \Sigma are both manifolds with bounded geometry, and the volume of A_{j} is uniformly bounded, the constants C_{0}, C_{1} and C_{2} are taken to be independent of j . We define the vector valued operator \mathcal{A}:L^{q}(\Sigma)\to(\ell^{q}_{j},L^{q}(A_{j})) , \mathcal{A}:=(\phi_{1},\phi_{2},...). We also define the operator \mathcal{B}:(\ell^{2}_{j},L^{2}(B_{j}))\to L^{2}(M) , (32) \displaystyle(f_{1},f_{2},...)\mapsto\sum_{j}\psi_{j}f_{j}. By the uniform locally finiteness of \{A_{j}\}_{j\geq 1} and \{B_{j}\}_{j\geq 1} , we have for all q>2, (33) \displaystyle||\mathcal{A}||_{L^{q}(\Sigma)\to(\ell^{q}_{j},L^{q}(A_{j}))}=O(1)\text{ and }||\mathcal{B}||_{(\ell^{2}_{j},L^{2}(B_{j}))\to L^{2}(M)}=O(1). For any f\in L^{2}(M) , by ( 21 ) and ( 33 ), (34) \displaystyle\begin{split}||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+1]}(P)f||_{L^{q}(\Sigma)}&\lesssim||R_{\Sigma}\sigma_{\lambda}f||_{L^{q}(\Sigma)}\\ &\lesssim||\mathcal{A}\sigma_{\lambda}f||_{(\ell^{q}_{j},L^{q}(A_{j}))}\\ &\lesssim\log(\lambda)^{1/2}\lambda^{\mu(q)}\left\|||\psi_{j}f||_{L^{2}(B_{j})}\right\|_{\ell_{j}^{q}(\mathbb{N})}\\ &\lesssim\log(\lambda)^{1/2}\lambda^{\mu(q)}\sqrt{\sum_{j}||\psi_{j}f||^{2}_{L^{2}(B_{j})}}\\ &\lesssim\log(\lambda)^{1/2}\lambda^{\mu(q)}||f||_{L^{2}(M)},\end{split} if k=n-2 and q=2 , or k=n-1 and q=\frac{2n}{n-1} . Similarly, by ( 22 ) and ( 33 ), (35) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+1]}(P)||_{L^{2}(M)\to L^{q}(\Sigma)}\lesssim||R_{\Sigma}\sigma_{\lambda}||_{L^{2}(M)\to L^{q}(\Sigma)}\lesssim\lambda^{\mu(q)}, if q\geq 2 otherwise. ∎ 2.3. Proof of Theorem 1 Now, we use the unit band estimate to prove the \log -scale estimate. For T\sim\log(\lambda) , define (36) \displaystyle\rho_{\lambda}:=\rho\left(\frac{\lambda-P}{T}\right). By an L^{2} orthogonality argument, we notice (37) \displaystyle||R_{\Sigma}\mathbf{1}_{[\lambda,\lambda+\log\lambda^{-1}]}||_{L^{2}(M)\to L^{q}(\Sigma)}\sim||R_{\Sigma}\rho_{\lambda}||_{L^{2}(M)\to L^{q}(\Sigma)}. By a TT^{*} argument, to prove Theorem 1 , it suffices to prove (38) \displaystyle||R_{\Sigma}(\rho_{\lambda}\rho_{\lambda}^{*})R_{\Sigma}^{*}||_{L^{q^{\prime}}(\Sigma)\to L^{q}(\Sigma)}\lesssim\lambda^{2\mu(q)}/\log\lambda. Set \Psi=\rho^{2} . For f\in L^{q^{\prime}}(\Sigma) , (39) \displaystyle R_{\Sigma}(\rho_{\lambda}\rho_{\lambda}^{*})R_{\Sigma}^{*}(f)(x)=\int_{\Sigma}\int\frac{1}{T}\cos(tP)(x,y)\hat{\Psi}\left(\frac{t}{T}\right)e^{i\lambda t}f(y)dtdy. Let \Phi(t) be a function supported in |t|\leq 1 , and equals to 1 on |t|\leq\frac{1}{2}. For f\in L^{q^{\prime}}(\Sigma), define the local and global part of \rho_{\lambda}\rho_{\lambda}^{*} respectively. (40) \displaystyle L_{\lambda}(f):=\frac{1}{T}\int\cos(tP)\hat{\Psi}\left(\frac{t}{T}\right)\Phi(t)e^{i\lambda t}fdt (41) \displaystyle G_{\lambda}(f):=\int\frac{1}{T}\cos(tP)\hat{\Psi}\left(\frac{t}{T}\right)(1-\Phi(t))e^{i\lambda t}fdt Since \rho_{\lambda}\rho_{\lambda}^{*}=L_{\lambda}+G_{\lambda} , it suffices to show ||R_{\Sigma}G_{\lambda}R^{*}_{\Sigma}||_{L^{q^{\prime}}(\Sigma)\to L^{q}(\Sigma)}=O(\lambda^{2\mu(q)}T^{-1}) and ||R_{\Sigma}L_{\lambda}R^{*}_{\Sigma}||_{L^{q^{\prime}}(\Sigma)\to L^{q}(\Sigma)}=O(\lambda^{2\mu(q)}T^{-1}). 2.3.1. Global estimate We obtain the global estimate via interpolation. We first obtain an L^{2} estimate following the argument of Theorem 5.1 in [ 7 ] . We sketch the proof for completeness. First, define \Upsilon\in\mathcal{S}(\mathbb{R}) to be the Fourier transform of (1-\Phi(\cdot))\hat{\Psi}\left(\frac{\cdot}{T}\right) . Notice that |\Upsilon(\tau)|\leq T(1+|\tau|)^{-N} for any N\in\mathbb{N} . By the unit band estimate ( 34 ), ( 35 ) and an almost orthogonality argument, we obtain the desired L^{2} estimates, (42) \displaystyle||R_{\Sigma}G_{\lambda}R_{\Sigma}^{*}||_{L^{2}(\Sigma)\to L^{2}(\Sigma)}\lesssim\begin{cases}\lambda^{2\mu(2)},&\text{if }k\neq n-2,\\ \lambda^{2\mu(2)}\log(\lambda),&\text{if }k=n-2.\end{cases} From Lemma 3.6 in [ 6 ] , we know (43) \displaystyle|G_{\lambda}(x,y)|\lesssim{\lambda^{\frac{n-1}{2}}e^{c_{M}T}} for some c_{M}>0 only depending on M. Thus, (44) \displaystyle||R_{\Sigma}G_{\lambda}R_{\Sigma}^{*}||_{L^{1}(\Sigma)\to L^{\infty}(\Sigma)}\lesssim{\lambda^{\frac{n-1}{2}}e^{c_{M}T}}. Now, we interpolate ( 44 ) and ( 42 ). If k\leq n-2 , then \mu(2)=\frac{n-1-k}{2} . Then (45) \displaystyle||R_{\Sigma}G_{\lambda}R_{\Sigma}^{*}||_{L^{q^{\prime}}(\Sigma)\to L^{q}(\Sigma)}\lesssim{\lambda^{\frac{n-1}{2}+\frac{n-1-2k}{q}}}e^{c_{M}T(1-\frac{2}{q})}. Notice that \frac{n-1}{2}+\frac{n-1-2k}{q}<2\mu(q) for all q>2 . So, we may find some 0\frac{2n}{n-1} , we have \frac{n-1}{2}-\frac{n-2}{q}<2\mu(q) . Similarly, we choose T=c^{*}\log(\lambda) with c^{*}=\frac{b}{c_{M}(1-\frac{2}{q})} for some 02 or k=n-1 , q>\frac{2n}{n-1} . Meanwhile, \Phi(\cdot)\hat{\Psi}\left(\frac{\cdot}{T}\right) is a compactly supported smooth function and |\Phi(\cdot)\hat{\Psi}\left(\frac{\cdot}{T}\right)| is bounded independently of T . If we define \Xi to be the Fourier transform of \Phi(\cdot)\hat{\Psi}\left(\frac{\cdot}{T}\right) , then \Xi is a Schwartz function with \Xi(\tau)\lesssim(1+|\tau|)^{-N} for any N\in\mathbb{N}. We may write L_{\lambda}=\frac{1}{T}(\Xi(\lambda-P)+\Xi(\lambda+P)). Thus, if k\leq n-2 and p>2 or if k=n-1 and p>\frac{2n}{n-1} , we may use an orthogonality argument and Proposition 2.2 to obtain (49) \displaystyle\begin{split}||R_{\Sigma}L_{\lambda}(R_{\Sigma})^{*}||_{L^{q^{\prime}}(\Sigma)\to L^{q}(\Sigma)}\lesssim\frac{1}{T}\lambda^{2\mu(q)}.\end{split} Thus, (50) \displaystyle\begin{split}||R_{\Sigma}\rho_{\lambda}\rho^{*}_{\lambda}(P)R_{\Sigma}^{*}f||_{L^{q}(\Sigma)}&\lesssim||R_{\Sigma}G_{\lambda}R_{\Sigma}^{*}f||_{L^{q}(\Sigma)}+||R_{\Sigma}L_{\lambda}R_{\Sigma}^{*}f||_{L^{q}(\Sigma)}\lesssim\lambda^{2\mu(q)}/\log\lambda||f||_{L^{q^{\prime}}(\Sigma)}.\end{split} This completes the proof of Theorem 1 . 3. Even asymptotically hyperbolic surface Let (M,g) be an even asymptotically hyperbolic surface with bounded geometry and curvature pinched below 0. We follow [ 6 ] to decompose (51) \displaystyle M=M_{tr}\cup M_{\infty}, such that M_{tr} is compact, and M_{\infty} asymptotically agrees with a background manifold (\tilde{M},\tilde{g}) , which satisfies favorable spectral projection estimates. Let S^{*}M be the cosphere bundle of M and denote the principle symbol of P by p(x,\xi) . Let (x(t),\xi(t))=e^{tH_{p}}(x,\xi) , where e^{tH_{p}} denote the geodesic flow on the cotangent bundle. Define (52) \displaystyle\Gamma_{\pm}:=\{(x,\xi)\in S^{*}M:x(t)\not\to\infty\text{ as }t\to\pm\infty\}. Define \pi:S^{*}M\to M with \pi(x,\xi)=x. The trapped set of M is (53) \displaystyle\pi(\Gamma_{+}\cap\Gamma_{-}). We require M_{tr} to be a compact subset of M that contains a neighborhood of the trapped set. We could construct an asymptotically hyperbolic, simply connected manifold with negative curvature and no resonance at the bottom of the spectrum, \tilde{M} , as in [ 6 ] , such that if M_{\infty} is appropriately defined, then the metric, g , in M_{\infty} agrees with the metric, \tilde{g} , in \tilde{M}. The metric of an asymptotically hyperbolic surface near the boundary is given by (54) \displaystyle 4\frac{dr^{2}+s(r,\theta)d\theta^{2}}{(1-r^{2})^{2}}, where s\in C^{\infty} and s(1,\theta)=1. Let \chi\in C^{\infty}_{0}((-1,1)) with \chi=1 in (-1/2,1/2) , then we can define the metric on \tilde{M} as (55) \displaystyle 4\frac{dr^{2}+r^{2}d\theta^{2}}{(1-r^{2})^{2}}+\chi(R(1-r))4\frac{(s(r,\theta)-r^{2})d\theta^{2}}{(1-r^{2})^{2}}, where R is a large enough constant. Then, the metric of \tilde{M} agrees with the metric of M when r\geq 1-(2R)^{-1} . Furthermore, note that |s(r,\theta)-r^{2}|\leq R^{-1} in the support of \chi(R(1-r)) . By choosing R sufficiently large, the Gaussian curvature of (\tilde{M},\tilde{g}) is bounded between -3/2 and -1/2 . Hence, \tilde{M} is a simply connected manifold with curvature pinched below 0 and no conjugate points. Finally, \tilde{M} has no resonance at the bottom of the spectrum by Lemma 2.3 of [ 6 ] . We denote \tilde{P}=\sqrt{-\tilde{\Delta}} , where \tilde{\Delta} is the Laplacian operator on \tilde{M}. Notice that being a geodesic is a local property. Thus, if we denote a connected component of \gamma in M_{\infty} by \gamma_{1} , there exists a geodesic \tilde{\gamma}_{1}\in\tilde{M} , such that \gamma_{1} agrees with \tilde{\gamma}_{1} whenever M agrees with \tilde{M} . In this section, we first prove a spectral projection estimate of P with an arbitrarily small spectral window on (\tilde{M},\tilde{g}) following [ 8 ] . Then, we prove a subcritical \log -scale spectral projection restricted to compact geodesic of M following [ 2 ] and finally use these two estimates to prove Theorem 2 . 3.1. Asymptotically hyperbolic and simply connected surface We prove that the background manifold (\tilde{M},\tilde{g}) satisfies a sharp spectral projection estimate for an arbitrarily small spectral window. Proposition 3.1 . Let (\tilde{M},\tilde{g}) be an asymptotically hyperbolic and simply connected surface with curvature pinched below 0. For \lambda\geq 1 , \eta\in(0,1] , q\geq 2 and \tilde{\gamma}\subset\tilde{M} being a geodesic, (56) \displaystyle||R_{\tilde{\gamma}}\mathbf{1}_{[\lambda,\lambda+\eta]}(\tilde{P})||_{L^{2}(\tilde{M})\to L^{q}(\tilde{\gamma})}\lesssim\lambda^{\mu(q)}\eta^{1/2}. Proof. Let {\mathbf{P}}_{\lambda}=\delta_{\lambda}(\tilde{P}) , where \delta_{\lambda} denotes the Dirac-Delta function. Notice that the spectral measure of \tilde{P} is {\mathbf{P}}_{\lambda}d\lambda . Then, (57) \displaystyle R_{\tilde{\gamma}}\mathbf{1}_{[\lambda,\lambda+\eta]}(\tilde{P})(R_{\tilde{\gamma}})^{*}=\int_{\lambda}^{\lambda+\eta}R_{\tilde{\gamma}}\mathbf{P}_{\kappa}(R_{\tilde{\gamma}})^{*}d\kappa. We aim to show (58) \displaystyle||R_{\tilde{\gamma}}\mathbf{P}_{\lambda}(R_{\tilde{\gamma}})^{*}||_{L^{q^{\prime}}({\tilde{\gamma}})\to L^{q}({\tilde{\gamma}})}\lesssim\lambda^{2\mu(q)}. We recall the kernel estimate by Chen and Hassell. By Theorem 5 of [ 8 ] , \mathbf{P}_{\lambda} can be represented by a convolution operator with kernel p_{\lambda}(x,y) , such that (59) \displaystyle|{p}_{\lambda}(x,y)|\lesssim\begin{cases}\lambda{(1+\lambda d_{\tilde{g}}(x,y))^{-1/2}},&\text{if }d_{\tilde{g}}(x,y)<1,\\ \lambda^{1/2}e^{-d_{\tilde{g}}(x,y)/2},&\text{if }d_{\tilde{g}}(x,y)\geq 1.\end{cases} Let d_{{\tilde{\gamma}}} denote the distance on \tilde{\gamma} with induced metric from \tilde{M}. Fix x\in{\tilde{\gamma}} and notice that for any y\in{\tilde{\gamma}} , we have d_{\tilde{g}}(x,y)=d_{{\tilde{\gamma}}}(x,y) . Meanwhile, we define (60) \displaystyle p_{\lambda}(x,y)=p_{\lambda}(d_{\tilde{g}}(x,y)). For f\in L^{q^{\prime}}(\tilde{\gamma}), (61) \displaystyle R_{\tilde{\gamma}}\mathbf{P}_{\kappa}(R_{\tilde{\gamma}})^{*}f=\int_{\tilde{\gamma}}p_{\kappa}(d_{\tilde{g}}(\cdot,y))f(y)dy. By Young’s inequality, to prove ( 58 ), it suffices to show \sup_{x}||p_{\lambda}(x,y)||_{L^{q/2}_{y}({\tilde{\gamma}})}\lesssim\lambda^{\mu(q)}. Assume q\neq 4 , we use ( 59 ) to compute \displaystyle||p_{\lambda}(x,y)||_{L^{q/2}_{y}({\tilde{\gamma}})} \displaystyle\lesssim\left(\int_{-\lambda^{-1}}^{\lambda^{-1}}|p_{\lambda}(r)|^{q/2}dr\right)^{2/q}+\left(\int_{\lambda^{-1}<|r|<1}|p_{\lambda}(r)|^{q/2}dr\right)^{2/q}+\left(\int_{|r|>1}|p_{\lambda}(r)|^{q/2}dr\right)^{2/q} \displaystyle\lesssim\lambda^{1-2/q}+\left(\int_{\lambda^{-1}<|r|<1}\left(\frac{\lambda}{r}\right)^{q/4}\right)^{2/q}+\lambda^{1/2}\left(\int_{|r|>1}\left(e^{-r/2}\right)^{q/2}\right)^{2/q} \displaystyle\lesssim\lambda^{1-2/q}+\left(\left.\lambda^{q/4}r^{-q/4+1}\right|^{\lambda^{-1}}_{1}\right)^{2/q}+\lambda^{1/2} \displaystyle\lesssim\lambda^{1-2/q}+\lambda^{1/2}. Thus, \sup_{x}||p_{\lambda}(x,y)||_{L^{q/2}_{y}({\tilde{\gamma}})}\lesssim\begin{cases}\lambda^{1/2}&\text{if }q<4,\\ \lambda^{1-2/q}&\text{if }q>4.\end{cases} We deal with the q=4 case using the Hardy-Littlewood fractional integral theorem. By ( 59 ) and assuming {\tilde{\gamma}} is parametrized by arc length, we have for all t and s\in\mathbb{R}, |p_{\lambda}({\tilde{\gamma}}(t),{\tilde{\gamma}}(s))|\lesssim\lambda^{1/2}|t-s|^{-\frac{1}{2}}. By the Hardy-Littlewood fractional integral theorem, we obtain (62) \displaystyle\left(\int^{-\infty}_{\infty}\left|\int^{-\infty}_{\infty}|p_{\lambda}({\tilde{\gamma}}(t),{\tilde{\gamma}}(s))h({\tilde{\gamma}}(s))|ds\right|^{4}dt\right)^{\frac{1}{4}}\lesssim\lambda^{1/2}||h||_{L^{4/3}({\tilde{\gamma}})}. Thus, ||R_{\tilde{\gamma}}\mathbf{P}_{\lambda}(R_{\tilde{\gamma}})^{*}||_{L^{4/3}({\tilde{\gamma}})\to L^{4}({\tilde{\gamma}})}\lesssim\lambda^{1/2}. By Young’s inequality, for any q\neq 4, (63) \displaystyle\int_{\lambda}^{\lambda+\eta}\left(\int_{\tilde{\gamma}}|{p}_{\kappa}(x,y)|^{q/2}dy\right)^{2/q}d\kappa\lesssim\lambda^{2\mu(q)}\eta. Thus, (64) \displaystyle||R_{\tilde{\gamma}}\mathbf{1}_{[\lambda,\lambda+\eta]}(\tilde{P})R_{\tilde{\gamma}}^{*}||_{L^{4/3}({\tilde{\gamma}})\to L^{4}({\tilde{\gamma}})}\lesssim\lambda^{2\mu(4)}\eta. ∎ 3.2. Critical \log -scaled estimates Notice that Theorem 1 only implies the sharp \log -scaled spectral projection estimates from L^{2}(M) , where M is a 2-dimensional manifold with bounded geometry and negative curvature, to the L^{q} spaces of its uniformly embedded curves for q>4 . We now prove the \log -scaled estimate from L^{2}(M) to L^{q} spaces of its compact geodesic segments for 2\leq q\leq 4 . We shall see later that, in order to prove Theorem 2 , it suffices to consider compact geodesic segments, since M_{tr} is compact. Proposition 3.2 . Let M be a Riemannian surface with curvature pinched below 0 and bounded geometry. Let \gamma be a fixed compact geodesic segment on M . Then for \lambda\geq 1 and q\geq 2 , (65) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\log\lambda^{-1}]}(P)||_{L^{2}(M)\to L^{q}(\gamma)}\lesssim\lambda^{1/4}\log\lambda^{-1/2}. Proof. Recall T\sim\log\lambda and \rho_{\lambda} as defined in ( 36 ). In addition, (66) \displaystyle R_{\gamma}(\rho_{\lambda}\rho_{\lambda}^{*})R^{*}_{\gamma}=R_{\gamma}L_{\lambda}R^{*}_{\gamma}+R_{\gamma}G_{\lambda}R^{*}_{\gamma} with L_{\lambda} and G_{\lambda} as defined in ( 41 ) and ( 40 ). First, a compact geodesic can be covered by a finite number of uniformly embedded geodesic segments. Therefore, we may construct coverings \{A_{j}\} of \gamma and \{B_{j}\} of a neighborhood of \gamma in M as described in Section 1. Then, we may appeal to the local results of small time wave kernels as in page 8 of [ 9 ] to obtain (67) \displaystyle|L_{\lambda}(x,y)|\lesssim\lambda^{1/2}(\log\lambda)^{-1}|x-y|^{-1/2} for any x,y\in A_{j}. Therefore, we have (68) \displaystyle||R_{\gamma}L_{\lambda}R^{*}_{\gamma}f||_{L^{4}({A_{j}})}\lesssim\lambda^{1/2}(\log\lambda)^{-1}||f||_{L^{4/3}(A_{j})}, for every j, and thus (69) \displaystyle||R_{\gamma}L_{\lambda}R^{*}_{\gamma}f||_{L^{4}({\gamma})}\lesssim\lambda^{1/2}(\log\lambda)^{-1}||f||_{L^{4/3}(\gamma)}. We now slightly modify the proof of Theorem 1.1 in [ 2 ] to deal with the R_{\gamma}G_{\lambda}R^{*}_{\gamma} term. By the Cartan-Hadamard theorem, (M,g) has a universal cover, (M^{\prime},g^{\prime}) , diffeomorephic to \mathbb{R}^{2} by \exp_{x} for any x\in M. We lift \gamma\in M to \gamma^{\prime}\in M^{\prime} . Let \pi be the covering map \pi:M^{\prime}\to M. Let \Gamma denote the set of deck transformations of M^{\prime} , such that (70) \displaystyle\Gamma:=\{\alpha:M^{\prime}\to M^{\prime},\alpha\circ\pi=\pi\}. For \tau>0, define the geodesic tube about \gamma^{\prime} of radius \tau, (71) \displaystyle T_{\tau}(\gamma^{\prime})=\{x\in M^{\prime}:d_{g^{\prime}}(x,\gamma^{\prime})\leq\tau\}. Let D be a fundamental domain of \Gamma in \mathbb{H}. Define \Gamma_{T_{\tau}(\gamma^{\prime})}\subset\Gamma as (72) \displaystyle\Gamma_{T_{\tau}(\gamma^{\prime})}=\{\alpha\in\Gamma:\alpha(D)\cap T_{\tau}(\gamma^{\prime})\neq\emptyset\}. Let P^{\prime}=\sqrt{-\Delta_{g^{\prime}}} , where \Delta_{g^{\prime}} is the Laplacian operator on M^{\prime}. Define G_{\lambda}^{\alpha} as the operator with kernel \displaystyle K^{\alpha}(x,y)=\frac{1}{T}\int^{T}_{-T}(1-\beta(t))\hat{\rho}(t/T)e^{-\lambda\tau}(\cos(tP^{\prime}))(x,\alpha y)dt. We may write (73) \displaystyle G_{\lambda}=G^{tube}_{\lambda}+G^{osc}_{\lambda}=\sum_{\alpha\in\Gamma_{T_{\tau}}(\gamma^{\prime})}G^{\alpha}_{\lambda}+\sum_{\alpha\notin G_{T_{\tau}}(\gamma^{\prime})}G^{\alpha}_{\lambda}. When \alpha=Id, by (4.14) of [ 7 ] , we have (74) \displaystyle K^{Id}(x,y)\lesssim T^{-1}\lambda^{1/2}d_{g^{\prime}}(x,y)^{-1/2}. By Hardy-Littlewood fractional integral theorem and since \gamma^{\prime} is a geodesic, we have (75) \displaystyle\left\|\int K^{Id}(\cdot,y)f(y)dy\right\|_{L^{4}(\gamma^{\prime})}\lesssim||f||_{L^{4/3}(\gamma^{\prime})}. Now, assume \alpha\neq Id. Following Lemma 3.1 of [ 9 ] , we can write (76) \displaystyle K^{\alpha}(x,y)=\omega(x,\alpha y)\sum_{\pm}b_{\pm}(T,\lambda,d_{g^{\prime}}(x,\alpha y))e^{\pm i\lambda d_{g^{\prime}}(x,\alpha y)}+R(x,\alpha y), where R=O(e^{c_{0}T}) for some c_{0}>0. Since M^{\prime} has curvature pinched below 0, by the Gunther comparison theorem, the volume of a geodesic ball with radius r in M is larger than e^{cr} for some c>0. Thus, by the line above (2.3.7) of [ 15 ] , we have (77) \displaystyle|\omega(x,\alpha y)|\lesssim e^{-cd_{g^{\prime}}(x,\alpha y)}, for some c>0. Meanwhile, by (3.15) of [ 9 ] , (78) \displaystyle|b_{\pm}(T,\lambda,d_{g^{\prime}}(x,\alpha y))|\lesssim T^{-1}\lambda^{1/2}d_{g^{\prime}}(x,\alpha y)^{-1/2}. By finite speed of propagation, K^{\alpha}(x,y) vanishes if d_{g^{\prime}}(x,\alpha y)>T. So, for any fixed x,y and \tau , (79) \displaystyle\#\{\alpha\in\Gamma:K(x,\alpha y)\neq 0\text{ and }\alpha\in\Gamma_{T_{\tau}}(\gamma^{\prime})\}=O(T). Thus, (80) \displaystyle|\sum_{\alpha\in\Gamma_{T_{\tau}}(\gamma^{\prime}),\alpha\neq Id}K^{\alpha}(x,y)|\lesssim\sum_{0\leq 2^{j}\leq T}T^{-1}\lambda^{1/2}e^{-c2^{j}}2^{j}2^{-j/2}\lesssim T^{-1}\lambda^{1/2}. Since \gamma is compact, this implies (81) \displaystyle||R_{\gamma}G^{tube}_{\lambda}R^{*}_{\gamma}f||_{L^{4}({\gamma})}\lesssim\lambda^{1/2}(\log\lambda)^{-1}||f||_{L^{4/3}(\gamma)}. Finally, by Section 2.1 of [ 2 ] , for some C>0, (82) \displaystyle||R_{\gamma}G^{osc}_{\lambda}R^{*}_{\gamma}f||_{L^{2}({\gamma})}\lesssim\lambda^{1/4}e^{CT}||f||_{L^{2}(\gamma)}. By choosing T=\beta\log\lambda with \beta>0 small enough and interpolating with ( 43 ), we have (83) \displaystyle||R_{\gamma}G^{osc}_{\lambda}R^{*}_{\gamma}f||_{L^{4}({\gamma})}\lesssim\lambda^{1/2-\epsilon}||f||_{L^{4/3}(\gamma)}, for some \epsilon>0. Thus, for f\in L^{2}(M) , (84) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+(\log\lambda)^{-1}]}(P)f||_{L^{4}({\gamma})}\lesssim\lambda^{1/4}(\log\lambda)^{-1/2}||f||_{L^{2}(M)}. Since \gamma is compact, we also have (85) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+(\log\lambda)^{-1}]}(P)f||_{L^{q}({\gamma})}\lesssim\lambda^{1/4}(\log\lambda)^{-1/2}||f||_{L^{2}(M)} for all q\leq 4 by Hölder’s inequality. ∎ 3.3. Proof of Theorem 2 Recall the construction of M_{tr} and M_{\infty} in ( 51 ). Let \psi_{{tr}} be a smooth function, which is compactly supported in a neighborhood of M_{tr}, and define \psi_{\infty}\in C^{\infty}(M) such that (86) \displaystyle\psi_{\infty}:=1-\psi_{tr}. Let \gamma be a nontrapped geodesic in M. Since \gamma is not trapped, \gamma\cap\text{supp}(\psi_{{tr}}) is compact. Therefore, for \lambda\geq 1 and q\geq 2 , by Proposition 3.2 , (87) \displaystyle||R_{\gamma}\psi_{tr}\mathbf{1}_{[\lambda,\lambda+(\log\lambda)^{-1}]}(P)||_{L^{2}(M)\to L^{q}({\gamma})}\lesssim\lambda^{\mu(q)}(\log\lambda)^{-1/2}. For 0<\eta<1, where N\in\mathbb{N} , we aim to show for f\in L^{2}(M), (88) \displaystyle||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\eta]}(P)f||_{L^{q}(\gamma)}\lesssim\lambda^{\mu(q)}\eta^{1/2}||f||_{L^{2}(M)}. Define \beta\in C^{\infty}((1/2,2)) and \beta=1 on (3/4,5/4) . By the definition of \mathbf{1}_{[\lambda,\lambda+\eta]}(P) , we may assume, without loss of generality, in the rest of this subsection f=\beta(P/\lambda)f . Let \rho be defined as in the last section. Showing ( 88 ) is equivalent to showing (89) \displaystyle||R_{\gamma}\rho\left(\frac{\Delta-\lambda^{2}}{\lambda\eta}\right)||_{L^{2}(M)\to L^{q}(\gamma)}\lesssim\lambda^{\mu(q)}\eta^{1/2}. We aim to show (90) \displaystyle\|R_{\gamma}\psi_{tr}\rho(({\Delta-\lambda^{2}})({\lambda\eta})^{-1})||_{L^{2}(M)\to L^{q}(\gamma)}\lesssim\lambda^{\mu(q)}\eta^{1/2}, and (91) \displaystyle\|R_{\gamma}\psi_{\infty}\rho(({\Delta-\lambda^{2}})({\lambda\eta})^{-1})||_{L^{2}(M)\to L^{q}(\gamma)}\lesssim\lambda^{\mu(q)}\eta^{1/2}. We follow the argument in [ 6 ] to simplify the problem to proving a resolvent kernel estimate. Let u=e^{-it\Delta}f . To prove ( 91 ), we let v:=\psi_{\infty}u . Then, v solves the Cauchy problem (92) \begin{cases}(i\partial_{t}-\Delta)v=[\psi_{\infty},\Delta]u\\ v|_{t=0}=\psi_{\infty}f.\end{cases} Since \Delta=\tilde{\Delta} on \text{supp }\psi_{\infty} , v also solves the following Cauchy problem on {\tilde{M}} . (93) \begin{cases}(i\partial_{t}-\tilde{\Delta})v=[\psi_{\infty},\tilde{\Delta}]u\\ v|_{t=0}=\psi_{\infty}f.\end{cases} Thus, (94) v=e^{-it\tilde{\Delta}}(\psi_{\infty}f)+i\int_{0}^{t}e^{-i(t-s)\tilde{\Delta}}\bigl([\Delta,\psi_{\infty}]u(s,\,\cdot\,)\bigr)\,ds. By using inverse Fourier transform, ( 94 ) implies (95) \psi_{\infty}\rho((\lambda\eta)^{-1}(-\Delta-\lambda^{2}))f=\rho((\lambda\eta)^{-1}(-\tilde{\Delta}-\lambda^{2}))(\psi_{\infty}f)\\ +(2\pi)^{-1}i\int_{-\infty}^{\infty}\lambda\eta\,\hat{\rho}\bigl(\lambda\eta t\bigr)\,e^{-it\lambda^{2}}\,\Bigl(\int_{0}^{t}e^{-i(t-s)\tilde{\Delta}}\bigl([\Delta,\psi_{\infty}]u(s,\,\cdot\,))\,ds\Bigr)\,dt. Let \tilde{\gamma} be a geodesic in \tilde{M} , such that \tilde{\gamma} agrees with a connected component of \gamma in M_{\infty} . We may assume without loss of generality that \gamma\cap M_{\infty} has at most two connected components. By the lossless spectral projection estimates from L^{2}({\tilde{M}}) to L^{q}(\tilde{\gamma}) , for \kappa\approx\lambda and f\in L^{2}(\tilde{M}), we have (96) \|R_{\tilde{\gamma}}\mathbf{1}_{[\kappa,\kappa+\eta]}(\tilde{P})f\|_{L^{q}(\tilde{\gamma})}\lesssim\lambda^{\mu(q)}\eta^{\frac{1}{2}}\|f\|_{L^{2}({\tilde{M}})}. Thus, we have the desired bounds for the first term on the right side of ( 95 ). Now, we estimate the second term on the right side of ( 95 ). Set \tilde{\rho}(t)=e^{-t}\hat{\rho}(t) . Then, integrating by parts in t yields (97) \displaystyle Q_{\lambda}f \displaystyle=\lambda\eta\int_{-\infty}^{\infty}e^{-it(\tilde{\Delta}+\lambda^{2}+\lambda\eta i)}\tilde{\rho}(\lambda\eta t)\Bigl(\int_{0}^{t}e^{is\tilde{\Delta}}[\Delta,\psi_{\infty}]\bigl(e^{-is\Delta}f\bigr)\,ds\,\Bigr)\,dt \displaystyle=-i(\tilde{\Delta}+\lambda^{2}+\lambda\eta i)^{-1}\lambda\eta\int_{-\infty}^{\infty}e^{-it(\tilde{\Delta}+\lambda^{2}+\lambda\eta i)}\frac{d}{dt}\tilde{\rho}(\lambda\eta t)\Bigl(\int_{0}^{t}e^{is\tilde{\Delta}}[\Delta,\psi_{\infty}]\bigl(e^{-is\Delta}f\bigr)\,ds\,\Bigr)\,dt \displaystyle\quad-i(\tilde{\Delta}+\lambda^{2}+\lambda\eta i)^{-1}\lambda\eta\int_{-\infty}^{\infty}[\Delta,\psi_{\infty}]\hat{\rho}(\lambda\eta t)e^{-it\lambda^{2}}e^{-it\Delta}f\,dt \displaystyle=-i(\tilde{\Delta}+\lambda^{2}+\lambda\eta i)^{-1}\bigl[Q^{\prime}_{\lambda}f+S_{\lambda}f\bigr], where Q^{\prime}_{\lambda} is the analog of Q_{\lambda} with \tilde{\rho}(\lambda\eta t) replaced by its derivative, and where S_{\lambda} is the last integral. To prove ( 91 ) it suffices to show that (98) \|R_{\tilde{\gamma}}Q_{\lambda}f\|_{L^{q}(\tilde{\gamma})}\lesssim\lambda^{\mu(q)}\eta^{1/2}\|f\|_{L^{2}({\tilde{M}})}. By (2.38) of [ 6 ] and ( 96 ), for 22\lambda}\mathbf{1}_{[j,j+1]}(\tilde{P})(\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}\psi||_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}\\ &\lesssim\sum_{j<\lambda/2}j^{\mu(q)}\lambda^{-2}+\sum_{j>2\lambda}j^{\mu(q)-2}\lesssim\lambda^{\mu(q)-1}.\end{split} Therefore, it suffices to consider (102) \displaystyle R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})(\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}\psi. Notice that (103) \displaystyle(\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}=\frac{1}{i(\lambda+i\eta)}\int^{\infty}_{0}e^{it\lambda-t/\log\lambda}\cos(t\tilde{P})dt. Fix \beta\in C_{0}^{\infty}(1/2,2) satisfying \sum_{j=-\infty}^{\infty}\beta(s/2^{j})=1 , and define (104) T_{j}h=\frac{1}{i(\lambda+i\eta)}\int_{0}^{\infty}\beta(2^{-j}t)e^{it\lambda-t\eta}\cos(t\tilde{P})h\,dt. Then, it suffices to obtain the desired bounds for the T_{j} operators. Note that T_{j} satisfies (105) T_{j}(\tau)=\frac{1}{i(\lambda+i\eta)}\int_{0}^{\infty}\beta(2^{-j}t)e^{it\lambda-t\eta}\cos(t\tau)\,dt=O(\lambda^{-1}2^{j}(1+2^{j}|\tau-\lambda|)^{-N}). We split the proof into four cases. (i) 2^{j}\leq\lambda^{-1} . Notice that \sum_{2^{j}\leq\lambda^{-1}}T_{j} satisfies (106) \displaystyle\sum_{2^{j}\leq\lambda^{-1}}T_{j}(\tau)=O(\lambda^{-1}(\lambda+|\tau|)^{-1}). Therefore, \displaystyle||R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})\sum_{2^{j}\leq\lambda^{-1}}T_{j}(\psi h)||_{L^{q}(\tilde{\gamma})} \displaystyle\lesssim\lambda^{\mu(q)+1/2}||\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})\sum_{2^{j}\leq\lambda^{-1}}T_{j}(\psi h)||_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)+1/2}(\lambda^{-2})||h||_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)-1}||h||_{L^{2}({\tilde{M}})}. (ii) 1/\lambda\leq 2^{j}\leq C. By ( 96 ), \displaystyle||R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})T_{j}(\psi h)||_{L^{q}({\tilde{\gamma}})} \displaystyle\lesssim||\sum_{|m|<\lambda 2^{j+1}}R_{\tilde{\gamma}}\mathbf{1}_{[m2^{-j},(m+1)2^{-j}]}(\tilde{P})T_{j}(\psi h)||_{L^{q}({\tilde{\gamma}})} \displaystyle\lesssim\lambda^{\mu(q)}2^{-j/2}||\sum_{|m|<\lambda 2^{j+1}}\mathbf{1}_{[m2^{-j},(m+1)2^{-j}]}(\tilde{P})T_{j}(\psi h)||_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)-1}2^{-j/2}\sum_{|m|<\lambda 2^{j+1}}2^{j}(1+|m|)^{-N}||h||_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)-1}2^{-j/2}2^{j}||h||_{L^{2}({\tilde{M}})} \displaystyle=\lambda^{\mu(q)-1}2^{j/2}||h||_{L^{2}({\tilde{M}})}. (iii) 2^{j}\geq C\log\lambda . Recall ( 105 ), and the spectral projection estimate on \tilde{M} , Proposition 3.1 , for all q\geq 2 , (107) \|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})T_{j}(\psi h)\|_{L^{q}({\tilde{\gamma}})}\lesssim\lambda^{\mu(q)-1}2^{j}\|\psi h\|_{L^{2}({\tilde{M}})}\lesssim\lambda^{\mu(q)-1}2^{j}\|h\|_{L^{2}({\tilde{M}})}. By Lemma 2.8 of [ 6 ] , we also have for every N\in\mathbb{N}, (108) \|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})T_{j}(\psi h)\|_{L^{\infty}({\tilde{\gamma}})}\lesssim\lambda^{-1/2}2^{-Nj}\|h\|_{L^{2}({\tilde{M}})}. By ( 107 ), ( 108 ) and interpolation, we have for any q>2, (109) \|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})T_{j}(\psi h)\|_{L^{q}({\tilde{\gamma}})}\lesssim\lambda^{-3/4}2^{-j}\|h\|_{L^{2}({\tilde{M}})}. Thus, if we choose C large enough (which may depend on q ), we have \|\sum_{2^{j}\geq C\log\lambda}R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})T_{j}\psi\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}=O(\lambda^{\mu(q)-1}). (iv) 1\leq 2^{j}\leq C\log\lambda . To handle the contribution of these terms, we shall first prove that for each fixed j with 2^{j}\geq 1 , we have the uniform bounds (110) \|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}T_{j}\psi h\|_{L^{q}(\tilde{\gamma})}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({{\tilde{M}}})}. To see this, let us define (111) E_{\lambda,j,k}=\mathbf{1}_{[\lambda+2^{-j}k,\lambda+(k+1)2^{-j})}(\tilde{P}). By Lemma 2.5 of [ 6 ] , we have (112) \displaystyle||\mathbf{1}_{[\lambda,\lambda+\eta]}(\tilde{P})\psi h||_{L^{2}({\tilde{M}})}\lesssim\eta^{1/2}||h||_{L^{2}({\tilde{M}})} By using ( 96 ) and ( 112 ) for \eta=2^{-j} , we have (113) \displaystyle\begin{split}\|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})\,&T_{j}\psi h\|_{L^{q}(\tilde{\gamma})}\\ &\leq\sum_{|k|\lesssim\lambda 2^{j}}\|R_{\tilde{\gamma}}\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})\,E_{\lambda,j,k}T_{j}\psi h\|_{L^{q}(\tilde{\gamma})}\\ &\leq\lambda^{\mu(q)}2^{-j/2}\sum_{|k|\lesssim\lambda 2^{j}}\|\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})E_{\lambda,j,k}T_{j}\psi h\|_{L^{2}({\tilde{M}})}\\ &\lesssim\lambda^{\mu(q)}2^{-j/2}\sum_{|k|\lesssim\lambda 2^{j}}(1+|k|)^{-N}\lambda^{-1}2^{j}\|\mathbf{1}_{[\lambda/2,2\lambda]}(\tilde{P})E_{\lambda,j,k}\psi h\|_{L^{2}({\tilde{M}})}\\ &\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})},\end{split} Thus, the proof of ( 110 ) is complete, and it suffices to consider the values of j such that C_{0}\leq 2^{j}\leq c_{0}\log\lambda , where C_{0} is sufficiently large and c_{0} is sufficiently small. We shall specify later the choices of C_{0} and c_{0} . Furthermore, |T_{j}(x,y)|=O(\lambda^{-N}) if d_{\tilde{g}}(x,y)\notin[2^{j-2},2^{j+2}] . We may assume that \tilde{\psi} is supported in a small neighborhood of some point y_{0} . Then, (114) \displaystyle\left\|R_{\tilde{\gamma}\cap\left\{x\in{\tilde{M}}:d_{\tilde{g}}(x,y_{0})\notin\left[\frac{C_{0}}{4},4c_{0}\log\lambda\right]\right\}}T_{j}\psi\right\|_{L^{2}({\tilde{M}})\to L^{\infty}(\tilde{\gamma})}\lesssim\left\|R_{\tilde{\gamma}\cap\left\{x\in{\tilde{M}}:d_{\tilde{g}}(x,y_{0})\notin\left[\frac{C_{0}}{4},4c_{0}\log\lambda\right]\right\}}T_{j}\psi\right\|_{L^{1}({\tilde{M}})\to L^{\infty}(\tilde{\gamma})}\lesssim\lambda^{-N}. By interpolation with ( 113 ), for any q>2 , we have (115) \displaystyle\left\|R_{\tilde{\gamma}\cap\left\{x\in{\tilde{M}}:d_{\tilde{g}}(x,y_{0})\notin\left[\frac{C_{0}}{4},4c_{0}\log\lambda\right]\right\}}T_{j}\psi\right\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}\lesssim\lambda^{-N}. Hence, it suffices to show that (116) \left\|R_{\tilde{\gamma}}\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}T_{j}(\tilde{\psi}h)\right\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})},\\ \text{where }\,S=\tilde{\gamma}\cap\left\{x\in{\tilde{M}}:\frac{C_{0}}{4}\leq d_{\tilde{g}}(x,y_{0})\leq 4c_{0}\log\lambda\right\}. By ( 105 ), if we fix \beta\in C^{\infty}_{0}((1/4,4)) with \beta=1 on (1/2,2), it suffices to show (117) \left\|R_{\tilde{\gamma}}\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\beta({\tilde{P}}/\lambda)T_{j}(\tilde{\psi}h)\right\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})}. To prove ( 117 ), we need to introduce microlocal cutoffs involving pseudodifferential operators. Since {\tilde{M}} has bounded geometry, we can cover the set S by a partition of unity \{\psi_{k}\} , which satisfies (118) 1=\sum_{k}\psi_{k}(x),\quad\text{supp}\,\psi_{k}\subset B(x_{k},{\delta_{0}}), with \delta_{0}>0 is a small fixed constant and |\partial_{x}^{j}\psi|\lesssim 1 uniformly in the normal coordinates around x_{k} for different k . Here B(x_{k},{\delta_{0}}) denotes geodesic balls of radius \delta_{0} \text{with }d_{\tilde{g}}(x_{k},x_{\ell})\geq{\delta_{0}}\,\,\text{if }\,k\neq\ell , and the balls B(x_{k},2{\delta_{0}}) have finite overlap. Using a volume counting argument, the number of values of k for which \text{supp}\,\psi_{k}\cap S\neq\emptyset is O(\lambda^{Cc_{0}}) for some fixed constant C . If we extend \beta\in C_{0}^{\infty}((1/4,4)) to an even function by letting \beta(s)=\beta(|s|) , then we can choose an even function \rho\in C_{0}^{\infty}(\mathbb{R}) satisfying \rho(t)=1 , |t|\leq\delta_{0}/4 and \rho(t)=0 , |t|\geq\delta_{0}/2 such that (119) \displaystyle\beta(\tilde{P}/\lambda)= \displaystyle(2\pi)^{-1}\int_{\mathbb{R}}\lambda\hat{\beta}(\lambda t)\cos t\tilde{P}dt \displaystyle= \displaystyle(2\pi)^{-1}\int\rho(t)\lambda\hat{\beta}(\lambda t)\cos t\tilde{P}dt+(2\pi)^{-1}\int(1-\rho(t))\lambda\hat{\beta}(\lambda t)\cos t\tilde{P}dt. \displaystyle= \displaystyle B+R The symbol of the operator R is O((1+|\tau|+\lambda)^{-N}) . Therefore, by the spectral projection theorem, we have (120) \displaystyle\|R_{\tilde{\gamma}}R\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}\lesssim_{N}\lambda^{-N}. On the other hand, by using the finite propagation speed property of the wave propagator, we may argue as in the compact manifold case to show that B is a pseudodifferential operator with principal symbol \beta(p(x,\xi)) , with p(x,\xi) here being the principal symbol of \tilde{P} . Choose \tilde{\psi}_{k}\in C^{\infty}_{0}(\tilde{M}) with \tilde{\psi}_{k}(y)=1 for y\in B(x_{k},\tfrac{5}{4}\delta_{0}) and \tilde{\psi}_{k}(y)=0 for y\notin B(x_{k},\tfrac{3}{2}\delta_{0}) . We may also assume that the \tilde{\psi}_{k} have bounded derivatives in the normal coordinates about x_{k} by taking \delta_{0}>0 small enough, given that {\tilde{M}} is of bounded geometry. Then, if B(x,y) is the kernel of B , we have \psi_{k}(x)B(x,y)=\psi_{k}(x)B(x,y)\tilde{\psi}_{k}(y)+O(\lambda^{-N}) , and so (121) \displaystyle\psi_{k}(x)B(x,y) \displaystyle=(2\pi)^{-2}\lambda^{2}\int e^{i\lambda\langle x-y,\xi\rangle}\psi_{k}(x){\beta(p(x,\xi))}\tilde{\psi}_{k}(y)d\xi+R_{k}(x,y) \displaystyle=A_{k}(x,y)+R_{k}(x,y). R_{k} is a lower order pseudodifferential operator which satisfies (122) \|R_{k}\|_{L^{2}({\tilde{M}})\to L^{\infty}({\tilde{M}})}=O(1). Since the x -support of R_{k} is compact and {\tilde{\gamma}} is uniformly embedded, we have for any q, (123) \|R_{\tilde{\gamma}}R_{k}\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}=O(1). By ( 123 ), the support property of R_{k} , and ( 107 ), we have (124) \displaystyle\| \displaystyle\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\sum_{k}R_{\tilde{\gamma}}R_{k}T_{j}(\tilde{\psi}h)\|_{L^{q}(S)} \displaystyle\lesssim\lambda^{Cc_{0}}\|\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}T_{j}(\tilde{\psi}h)\|_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{Cc_{0}}\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\lambda^{-1}2^{j}\|\tilde{\psi}h\|_{L^{2}({\tilde{M}})}. Note that \mu(q)\geq\frac{1}{4} for all q\geq 2 . Therefore, by choosing c_{0} sufficiently small, the bound in ( 124 ) is better than the estimate in ( 117 ). Moreover, A_{k}(x,y)=0,\,\,\,\text{if }\,x\notin B(x_{j},{\delta_{0}})\,\,\text{or }\,\,y\notin B(x_{j},3{\delta_{0}}/2). For each x_{k} , let \omega_{k} be the unit covector such that e^{-tH_{p}}(x_{k},\omega_{k})=(y_{0},\delta_{0}) for some \delta_{0} and t=d_{\tilde{g}}(x_{k},y_{0}) , with y_{0} as in ( 116 ). We define a_{k}(x,\xi)\in C^{\infty} such that in the normal coordinate around x_{k} , (125) a_{k}(x,\xi)=0\,\,\text{if }\bigl|\tfrac{\xi}{|\xi|_{\tilde{g}(x)}}-\omega_{k}\bigr|\geq 2\delta_{1},\,\,\,\text{and}\,\,\,a_{k}(x,\xi)=1\,\,\text{if }\bigl|\tfrac{\xi}{|\xi|_{\tilde{g}(x)}}-\omega_{k}\bigr|\leq\delta_{1}. Here |\xi|_{g(x)}=p(x,\xi) , and \delta_{1} is a fixed small constant that will be chosen later. By the bounded geometry assumption, we may assume that \partial^{\alpha}_{x}\partial^{\gamma}_{\xi}a_{k}=O(1) if p(x,\xi)=1 , independent of k , with \partial_{x} denoting derivatives in the normal coordinate system about x_{k} . We finally define the kernel of the microlocal cutoffs A_{k,0} and A_{k,1} as (126) \displaystyle A_{k}(x,y) \displaystyle=A_{k,0}(x,y)+A_{k,1}(x,y) \displaystyle=(2\pi)^{-2}\lambda^{2}\int e^{i\lambda\langle x-y,\xi\rangle}\psi_{k}(x)a_{k}(x,\xi)\beta((p(x,\xi))\tilde{\psi}_{k}(y)d\xi \displaystyle+(2\pi)^{-2}\lambda^{2}\int e^{i\lambda\langle x-y,\xi\rangle}\psi_{k}(x)(1-a_{k}(x,\xi))\beta((p(x,\xi))\tilde{\psi}_{k}(y)d\xi. Notice that (127) \|R_{\tilde{\gamma}}A_{k,\ell}\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})}\lesssim\|R_{\tilde{\gamma}}\|_{L^{2}({\tilde{M}})\to L^{q}(\tilde{\gamma})},\,\,2\leq q\leq\infty,\,\,\ell=0,1. Note that the support of A_{k,\ell} are finitely overlapping. Thus, ( 127 ) implies that (128) \|R_{\tilde{\gamma}}\sum_{k}A_{k,\ell}h\|_{L^{q}(\tilde{\gamma})}\lesssim\|R_{\tilde{\gamma}}h\|_{L^{q}(\tilde{\gamma})},\,\,2\leq p\leq\infty,\,\,\ell=0,1. By ( 118 ), ( 121 ) and ( 126 ), to prove ( 117 ), it suffices to show (129) \|\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\sum_{k}R_{\tilde{\gamma}}A_{k,0}T_{j}(\tilde{\psi}h)\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})}, as well as (130) \|\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\sum_{k}R_{\tilde{\gamma}}A_{k,1}T_{j}(\tilde{\psi}h)\|_{L^{q}(S)}\lesssim_{N}\lambda^{-N}\|h\|_{L^{2}({\tilde{M}})}. Now we shall give the proof of ( 130 ). It suffices to show (131) \|R_{\tilde{\gamma}}A_{k,1}T_{j}(\tilde{\psi}h)\|_{L^{q}(S)}\lesssim_{N}\lambda^{-N}\|h\|_{L^{2}({\tilde{M}})},\,\,C_{0}\leq 2^{j}\leq c_{0}\log\lambda. Note that S is a uniformly embedded geodesic segment in a ball of radius O(\lambda^{Cc_{0}}) . so the volume of the set S is O(\lambda^{Cc_{0}}) . To prove ( 131 ), it suffices to show the following pointwise bound (132) \int_{0}^{\infty}\beta(2^{-j}t)e^{it\lambda-t\eta}(A_{k,1}\circ\cos(t\tilde{P}))(x,y)\tilde{\psi}(y)\,dt\lesssim_{N}\lambda^{-N}. But ( 132 ) is (2.87) of [ 6 ] , so the proof of ( 130 ) is complete. Now we give the proof of ( 129 ). By ( 128 ) and our previous results for the operators T_{j}\tilde{\psi} when 2^{j}\leq C_{0} and 2^{j}\geq c_{0}\log\lambda , proving ( 129 ) is equivalent to showing that (133) \|\sum_{k}R_{\tilde{\gamma}}A_{k,0}(\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}({\tilde{\psi}}h)\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})}. To prove ( 133 ), it suffices to show (134) \|\sum_{k}R_{\tilde{\gamma}}A_{k,0}(\tilde{\Delta}+(\lambda-i\eta)^{2})^{-1}({\tilde{\psi}}h)\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})} and (135) \|\sum_{k}R_{\tilde{\gamma}}A_{k,0}\big((\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}-(\tilde{\Delta}+(\lambda-i\eta)^{2})^{-1}\big)({\tilde{\psi}}h)\|_{L^{q}(S)}\lesssim\lambda^{\mu(q)-1}\|h\|_{L^{2}({\tilde{M}})}. Note that if we define E_{\lambda,m}=\mathbf{1}_{[\lambda+m\eta,\lambda+(m+1)\eta)}(\tilde{P}) , then the symbol of the operator E_{\lambda,m}\big((\tilde{\Delta}+(\lambda+i\eta)^{2})^{-1}-(\tilde{\Delta}+(\lambda-i\eta)^{2})^{-1}\big) is O\left((\lambda\eta)^{-1}(1+|m|)^{-2}\right) . Thus ( 135 ) can be proved using the same arguments as in the proof of ( 110 ). \displaystyle||R_{\tilde{\gamma}}\sum_{k}A_{k,0}\beta(\tilde{P}/\lambda)\left((\tilde{\Delta}+\lambda^{2}+i\lambda\eta)^{-1}-(\tilde{\Delta}+\lambda^{2}-i\lambda\eta)^{-1}\right)\tilde{\psi}h||_{L^{q}(\tilde{\gamma})} \displaystyle\lesssim\left\|R_{\tilde{\gamma}}\psi\sum_{|m|\lesssim\lambda\eta^{-1}}E_{\lambda,m}\left(\left(\tilde{\Delta}+\lambda^{2}+{i\lambda}\eta\right)^{-1}-\left(\tilde{\Delta}+\lambda^{2}-{i\lambda}\eta\right)^{-1}\right)\tilde{\psi}h\right\|_{L^{q}(\tilde{\gamma})} \displaystyle\lesssim\left\|R_{\tilde{\gamma}}\psi\sum_{|m|\lesssim\lambda\eta^{-1}}\left({(1+m^{2})\lambda\eta}\right)^{-1}E_{\lambda,m}\tilde{\psi}h\right\|_{L^{q}(\tilde{\gamma})} \displaystyle\lesssim\lambda^{\mu(q)}\eta^{1/2}\left\|\sum_{|m|<\lambda\eta^{-1}}\left({(1+m^{2})\lambda\eta}\right)^{-1}E_{\lambda,m}(P)\tilde{\psi}h\right\|_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)}\eta^{1/2}\sum_{|m|<\lambda\eta^{-1}}\left({(1+m^{2})\lambda\eta}\right)^{-1}\eta^{1/2}||h||_{L^{2}({\tilde{M}})} \displaystyle\lesssim\lambda^{\mu(q)-1}||h||_{L^{2}({\tilde{M}})}. To prove ( 134 ), note that (136) \left(\tilde{\Delta}+(\lambda-i\eta)^{2}\right)^{-1}=\frac{i}{(\lambda-i\eta)}\int_{0}^{\infty}e^{-it\lambda-t\eta}\cos(t\tilde{P})\,dt. As in ( 104 ), if we define (137) \bar{T}_{j}h=\frac{i}{(\lambda-i\eta)}\int_{0}^{\infty}\beta(2^{-j}t)e^{-it\lambda-t\eta}\cos(t\tilde{P})h\,dt, then the above arguments implies that the analog of ( 134 ), involving the operators \bar{T}_{j}\tilde{\psi} for 2^{j}\leq C_{0} and 2^{j}\geq c_{0}\log\lambda , satisfies the desired bound. By (2.102) of [ 6 ] and the fact that S has length \lesssim\log\lambda , we have (138) \|R_{\tilde{\gamma}}\sum_{\{j:C_{0}\leq 2^{j}\leq c_{0}\log\lambda\}}\sum_{k}A_{k,0}\bar{T}_{j}({\tilde{\psi}}h)\|_{L^{q}(S)}\lesssim_{N}\lambda^{-N}\|h\|_{L^{2}({\tilde{M}})}, which completes the proof. ∎ Since S_{\lambda}f is compactly supported in {\tilde{M}} , by Proposition 3.3 , (139) \displaystyle\|R_{\tilde{\gamma}}(\tilde{\Delta}+\lambda^{2}+i\eta\lambda)^{-1}S_{\lambda}f\|_{L^{q}({\tilde{\gamma}})}\lesssim\lambda^{\mu(q)-1}\|S_{\lambda}f\|_{L^{2}({\tilde{M}})}. Now, we estimate ||S_{\lambda}f||_{L^{2}({\tilde{M}})} . By the sentence below (2.41) of [ 6 ] , we have (140) \displaystyle||S_{\lambda}f||_{L^{2}({\tilde{M}})}\lesssim\lambda\eta^{1/2}||f||_{L^{2}(M)}. So, \|R_{\tilde{\gamma}}(\tilde{\Delta}+\lambda^{2}+i\eta\lambda)^{-1}S_{\lambda}f\|_{L^{q}({\tilde{\gamma}})} satisfies the desired bound in ( 98 ). Now, we aim to obtain ( 90 ). Following Section 2 of [ 6 ] , we choose \alpha\in C^{\infty}_{0}((-1,1)) that satisfies \sum_{j}\alpha(t-j)=1 , for any t\in{\mathbb{R}} . Then let \alpha_{j}(t)=\alpha((\lambda/\log\lambda)t-j), to obtain a smooth partition of unity associated with \log\lambda/\lambda -intervals. Let u_{j}=\alpha_{j}(t)\psi_{tr}e^{-it\Delta}f. Then, (i\partial_{t}-\Delta)u_{j}=v_{j}+w_{j}, where v_{j}=i\frac{\lambda}{\log\lambda}\alpha^{\prime}\left(t\frac{\lambda}{\log\lambda}-j\right)\psi_{tr}u and w_{j}=-\alpha\left(t\frac{\lambda}{\log\lambda}-j\right)[\Delta_{x},\psi_{tr}]u. Then, if \rho is as above then \hat{\rho}(\lambda\eta t)u_{j}(t,x)=-i\hat{\rho}(\lambda\eta t)\int^{t}_{0}e^{-i(t-s)\Delta}v_{j}(s,x)ds-i\hat{\rho}(\lambda\eta t)\int_{0}^{t}e^{-i(t-s)\Delta}w_{j}(s,x)ds. Let I_{j}=[(j-1)\lambda^{-1}\log\lambda,(j+1)\lambda^{-1}\log\lambda], we follow [ 6 ] to observe \displaystyle\int\lambda\eta\hat{\rho}(\lambda\eta t)v_{j}(t)\,e^{-it\lambda^{2}}\,dt=-i(2\pi)^{-1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}\bigl[R^{\prime}_{j,v,\lambda}f+S_{j,v,\lambda}f\bigr], with R^{\prime}_{j,v,\lambda}f=\lambda\eta\int_{I_{j}}e^{-it(\Delta+\lambda^{2}+i\lambda/\log\lambda)}\frac{d}{dt}\bigl(e^{-t\lambda/\log\lambda}\hat{\rho}(\lambda\eta t)\bigr)\Bigl(\int_{0}^{t}\bigl(e^{is\Delta}[\partial_{s},\alpha_{j}]\,\psi_{tr}e^{-is\Delta}f\bigr)\,ds\,\Bigr)\,dt, and S_{j,v,\lambda}f=\lambda\eta\int_{I_{j}}e^{-it\lambda^{2}}\hat{\rho}(\lambda\eta t)[\partial_{t},\alpha_{j}]\psi_{tr}e^{-it\Delta}f\,dt. Similarly, set \int\lambda\eta\hat{\rho}(\lambda\eta t)w_{j}(t)\,e^{it\lambda^{2}}\,dt=(2\pi)^{-1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}\bigl[R^{\prime}_{j,w,\lambda}f+S_{j,w,\lambda}f\bigr], where (141) \displaystyle\begin{split}&R^{\prime}_{j,w,\lambda}f\\ &=\lambda\eta\int_{I_{j}}e^{-it(\Delta+\lambda^{2}+i\lambda/\log\lambda)}\frac{d}{dt}\bigl(e^{-t\lambda/\log\lambda}\hat{\rho}(\lambda\eta t)\bigr)\Bigl(\int_{0}^{t}\bigl(e^{is\Delta}\alpha_{j}(s)[\Delta,\psi_{tr}]\,e^{-is\Delta}f\bigr)\,ds\,\Bigr)\,dt,\end{split} and S_{j,w,\lambda}f=\lambda\eta\int_{I_{j}}e^{-it\lambda^{2}}\alpha_{j}(t)\hat{\rho}(\lambda\eta t)[\Delta,\psi_{tr}]e^{-it\Delta}f\,dt. Let \psi_{1}\in C^{\infty}_{0}(M) with \psi_{1}=1 on M_{tr} . We have following analog of ( 99 ). For any q>2, (142) \|R_{\gamma}\psi_{1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}h\|_{L^{q}(\gamma)}\lesssim\lambda^{\mu(q)-1}(\log\lambda)^{1/2}\|h\|_{L^{2}(M)}. This follows from the Cauchy-Schwarz inequality, L^{2} orthogonality, and the sharp spectral projection estimates, Proposition 3.2 . By ( 142 ), as well as the arguments on page 20 and 21 of [ 6 ] , we have (143) \Bigl(\,\sum_{j}\|R_{\gamma}\psi_{1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}R^{\prime}_{j,v,\lambda}f\|^{2}_{L^{q}(\gamma)}\,\Bigr)^{1/2}\\ +\Bigl(\,\sum_{j}\|R_{\gamma}\psi_{1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}S_{j,v,\lambda}f\|^{2}_{L^{q}(\gamma)}\,\Bigr)^{1/2}\lesssim\lambda^{\mu(q)}\eta(\log\lambda)^{1/2}\,\,\|f\|_{L^{2}(M)}, and (144) \displaystyle\left(\sum_{j}\|R_{\gamma}\psi_{1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}R^{\prime}_{j,w,\lambda}f\|^{2}_{L^{q}(\gamma)}\right)^{1/2}\lesssim\lambda^{\mu(q)}\eta(\log\lambda)^{1/2}\|f\|_{L^{2}(M)}. Now, it suffices to estimate (145) \displaystyle\Bigl(\,\sum_{j}\|R_{\gamma}\psi_{1}(\Delta+\lambda^{2}+i\lambda/\log\lambda)^{-1}S_{j,w,\lambda}f\|^{2}_{L^{q}(\gamma)}\Bigr)^{1/2}. We need the following proposition. Proposition 3.4 . Let M be an even asymptotically hyperbolic surface with curvature pinched below 0, \gamma be a geodesic in M. Let \psi_{1}\in C^{\infty}_{0}(M) with \psi_{1}=1 on M_{tr} , and \tilde{\psi_{1}}\in C^{\infty}_{0}(M_{\infty}) supported away from the trapped set, then, for 21, (162) \displaystyle||f||_{L^{2}(\mathbb{H})}\sim\int^{\infty}_{0}\left|\frac{\Gamma(\frac{1}{2}+i\kappa)}{\Gamma(i\kappa)}\tilde{f}(\kappa)\right|^{2}\sim\sqrt{\lambda\eta}. The spherical function \varphi_{\kappa} is defined as (163) \displaystyle\varphi_{\kappa}=\frac{\sqrt{2}}{\pi}\int^{r}_{0}\cos(\kappa s)(\cosh r-\cosh s)^{-1/2}ds. For \kappa\in[\lambda,\lambda+\eta] with \lambda>1 and r<\frac{1}{\lambda+\eta/2}, (164) \displaystyle\varphi_{\kappa}(r)\sim\varphi_{0}(r)\sim e^{-r/2}. By the spherical Fourier inversion formula, f(r)\sim\int^{\infty}_{0}\tilde{f}(\kappa)\varphi_{\kappa}(r)\left|\frac{\Gamma(\frac{1}{2}+i\kappa)}{\Gamma(i\kappa)}\right|^{2}d\kappa\sim\lambda\eta. Hence (165) \displaystyle||f||_{L^{q}(\gamma)}\gtrsim\left(\int_{|r|<\frac{1}{\lambda+\eta/2}}(\lambda\eta)^{q}dr\right)^{1/q}\gtrsim\lambda^{1-1/q}\eta. Combining ( 162 ) and ( 165 ), ||R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\eta]}(P)||_{L^{2}(\mathbb{H})\to L^{q}(\gamma)}\gtrsim\eta^{1/2}\lambda^{1/2-1/q}. Now we present an example to illustrate why we require M to be a surface with bounded geometry. Example 4.3 (Hyperbolic surface with cusp) . Consider the upper half-plane model. Let x=x_{1}+ix_{2}\in\mathbb{C}^{+} for x_{1},x_{2}\in\mathbb{R}. A parabolic cylinder X=\mathbb{H}/\langle h_{\alpha}\rangle , for \alpha\in\mathbb{R} and h_{\alpha}(x):=x+\alpha. Notice that the injectivity radius of a parabolic cylinder is not a positive number. Thus, a parabolic cylinder is not a surface of bounded geometry, and Theorem 1 does not apply to it. Recall that \Delta_{{\mathbb{H}}}-\frac{1}{4}=-x_{2}^{2}(\partial_{1}^{2}+\partial_{2}^{2})-x_{2}\partial_{2}-\frac{1}{4} , so g(x)=x_{2}^{{1}/2-i\xi} is a generalized eigenfunction of P_{\mathbb{H}} of the eigenvalue \xi. Note that \varsigma_{\lambda} is independent of x_{1} . Thus, g is also a generalized eigenfunction of P_{X}=\sqrt{\Delta_{X}-\frac{1}{4}} of the same eigenvalue. Consider \varsigma_{\lambda}(x)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}\phi(\eta^{-1}(\lambda-\xi)))\,x_{2}^{1/2-i\xi}\,d\xi, where \phi is supported in [-1/10,1/10] . Then, the P_{X} spectrum of \varsigma_{\lambda} is in [\lambda-\eta,\lambda+\eta] if \lambda is large and \eta\in(0,1] . Furthermore, \varsigma_{\lambda}(x)=\eta\,x_{2}^{1/2-i\lambda}\hat{\phi}(\eta\log x_{2}). Using the change of coordinates \omega=\log x_{2} we see that \|\varsigma_{\lambda}\|_{L^{2}(X)}=\eta\Bigl(\int_{0}^{\infty}x_{2}\,|\hat{\phi}(\eta\log x_{2})|^{2}\frac{dx_{2}}{x_{2}^{2}}\,\Bigr)^{1/2}=\eta\Bigl(\int_{-\infty}^{\infty}|\hat{\phi}(\eta\omega)|^{2}\,d\omega\Bigr)^{1/2}. Without loss of generality, let X=\{x_{1}+ix_{2}|x_{1}\in[-\alpha/2,\alpha/2),x_{2}>0\} . Notice that the vertical line \gamma(t)=it is a geodesic in X. We compute the L^{q} norm of \varsigma_{\lambda} restricted to \gamma. \displaystyle\|R_{\gamma}\varsigma_{\lambda}\|_{L^{q}(\gamma)} \displaystyle=\eta\Bigl(\,\int_{0}^{\infty}x_{2}^{\frac{q}{2}}\,|\hat{\phi}(\eta\log x_{2})|^{q}\,\frac{dx_{2}}{x_{2}}\,\Bigr)^{1/q} \displaystyle=\eta\Bigl(\int_{-\infty}^{\infty}e^{\frac{q}{2}\omega}\,|\hat{\phi}(\eta\omega)|^{q}\,d\omega\Bigr)^{1/q}. If we take \phi(s)=a(s)\cdot\mathbf{1}_{[0,1]}(s) where a\in C^{\infty}_{0}((-1/10,1/10)) satisfies a(0)=1 , then |\hat{\phi}(\tau)|\approx|\tau|^{-1} for large |\tau| . In this case, \varsigma_{\lambda}\in L^{2}(X) but R_{\gamma}\varsigma_{\lambda}\notin L^{q}(\gamma) for any q\in(2,\infty] . Thus, R_{\gamma}\mathbf{1}_{[\lambda,\lambda+\eta]}(P_{X}) are unbounded between L^{2}(X) and L^{q}(\gamma) .