The document presents a new approach called TSCAN for temporally summarizing topics from a collection of documents. TSCAN first derives the major themes of a topic from the eigenvectors of a temporal block association matrix. It then extracts significant events and their summaries for each theme by examining the eigenvectors. Finally, it associates the extracted events based on their temporal closeness and context similarity to form an evolution graph of the topic. Experiments on the TDT4 corpus show that temporal summaries generated by TSCAN present topics in a comprehensible form and are superior to existing summarization methods based on human references.