Self-supervised contrastive learning has achieved strong representation learning performance by pulling positive samples together and pushing negative samples apart. However, most existing methods rely on pointwise feature comparisons, and even multi-view extensions often use simple aggregation of view-level similarities, which limits their ability to capture joint multi-view structure. We propose a Divergence-Based Similarity Function (DSF) for multi-view contrastive learning. DSF represents multiple augmented views as a von Mises--Fisher distribution on the unit hypersphere and defines similarity through the negative KL divergence between distributions. This allows DSF to capture both the central direction and concentration structure of multiple views. Experiments show that DSF outperforms standard contrastive methods and existing multi-view baselines under comparable computational budgets, while improving transfer performance and robustness. We also establish a connection to cosine-similarity-based InfoNCE and show that DSF works effectively without temperature tuning.