Randomly-oriented k-d Trees Adapt to Intrinsic Dimension

The classic k-d tree data structure continues to be widely used in spite of its vulnerability to the so-called curse of dimensionality. Here we provide a rigorous explanation: for randomly rotated data, a k-d tree adapts to the intrinsic dimension of the data and is not affected by the ambient dimension, thus keeping the data structure efficient for objects such as low-dimensional manifolds and sparse data. The main insight of the analysis can be used as an algorithmic pre-processing step to realize the same benefit: rotate the data randomly; then build a k-d tree. Our work can be seen as a refinement of Random Projection trees [Dasgupta 2008], which also adapt to intrinsic dimension but incur higher traversal costs as the resulting cells are polyhedra and not cuboids. Using k-d trees after a random rotation results in cells that are cuboids, thus preserving the traversal efficiency of standard k-d trees.