This paper addresses the computational and theoretical limitations of current distribution alignment methods for source-free unsupervised domain adaptation (SFUDA). Specifically, we focus on estimating classification performance and confidence in situations where target labels are absent. To overcome the limitations of existing theoretical frameworks, which generate computationally intractable quantities and fail to adequately reflect the characteristics of the alignment algorithms used, we propose the Optimal Transport (OT) score, a novel theoretically derived confidence metric that leverages the flexibility of the decision boundary induced by the Semi-Discrete Optimal Transport alignment. The proposed OT score is intuitively interpretable, theoretically rigorous, and provides a principled uncertainty estimation for a given set of target pseudolabels. Experimental results demonstrate that the OT score outperforms existing confidence scores, improves SFUDA performance through training-time weighting adjustments, and provides a reliable label-free proxy for model performance.
