We address the problem of learning similarities between documents written in different languages for language pairs where little or no direct supervision (in the form of a comparable or parallel corpus) is available. To make up for the lack of direct supervision, our approach takes advantage of the fact that they may be linked indirectly by a hub language. That is, correspondences exist between each of the languages and a third, hub language. The main goal of our paper is to explore the viability of cross-lingual learning under such conditions. We propose a method that extracts a set of multilingual topics that facilitate a common representation of documents in different languages. The method is suitable for a comparable multilingual corpus with missing documents. We evaluate the approach in a truly multi-lingual setting, performing document retrieval across eight Wikipedia languages.