We study exact algorithms for EUCLIDEAN TSP in R d . In the early 1990s algorithms with n O( √ n)running time were presented for the planar case, and some years later an algorithm with n O(n 1−1/d ) running time was presented for any d 2. Despite significant interest in subexponential exact algorithms over the past decade, there has been no progress on EUCLIDEAN TSP, except for a lower bound stating that the problem admits no 2 O(n 1−1/d−ε ) algorithm unless ETH fails. Up to constant factors in the exponent, we settle the complexity of EUCLIDEAN TSP by giving a 2 O(n 1−1/d ) algorithm and by showing that a 2 o(n 1−1/d ) algorithm does not exist unless ETH fails. affirmatively by Arora [1] who provided a PTAS with running time n(log n) O( √ d/ε) d−1 . Independently, Mitchell [18] designed a PTAS in R 2 . The running time was improved to 2 (1/ε) O(d) n + (1/ε) O(d) n log n by Rao and Smith [22]. Hence, the computational complexity of the approximation problem has essentially been settled. Results on exact algorithms for EUCLIDEAN TSP-these are the topic of our paper-are also quite different from those on the general problem. The best known algorithm for the general case runs, as already remarked, in exponential time, and there is no 2 o(n) algorithm under ETH due to classical reductions for HAMILTONIAN CYCLE [5, Theorem 14.6]. EUCLIDEAN TSP, on the other hand, is solvable in subexponential time. For the planar case this has been shown in the early 1990s by Kann [14] and independently by Hwang, Chang and Lee [12], who presented an algorithm with an n O( √ n) running time. Both algorithms use a divide-and-conquer approach that relies on finding a suitable separator. The approach taken by Hwang, Chang and Lee is based on considering a triangulation of the point set such that all segments of the tour appear in the triangulation, and then observing that the resulting planar graph has a separator of size O( √ n). Such a separator can be guessed in n O( √ n) ways, leading to a recursive algorithm with n O( √ n) running time. It seems hard to extend this approach to higher dimensions. Kann obtains his separator in a more geometric way, using the fact that in an optimal tour, there cannot be too many long edges that are relatively close together-see the Packing Property we formulate in Section 2. This makes it possible to compute a separator that is crossed by O( √ n) edges of an optimal tour, which can be guessed in n O( √ n)