The generative theory for relevance and its operational manifestation-the relevance model-are based on the premise that a single query is used to represent an information need for retrieval. In this work, we extend the theory and devise novel techniques for relevance modeling using as set of query variations representing the same information need. Our new approach is based on fusion at the term level, the model level, or the document-list level.We theoretically analyze the connections between these methods and provide empirical support of their equivalence using TREC datasets. Specifically, our new approach of inducing relevance models from multiple query variations substantially outperforms relevance model induction from a single query which is the standard practice. Our approach also outperforms fusion over multiple query variations, which is currently one of the best known baselines for several commonly used test collections.