Unsupervised Disentanglement Without Compromises : How Functional Orthogonality Enforces Identifiability

This paper presents a novel approach to unsupervised disentangled representation learning by introducing a functional orthogonality constraint on the Jacobian of generative mappings. It demonstrates that this constraint ensures identifiability of nonlinear generative models without the need for statistical independence or causal assumptions, validated through experiments with orthogonality-regularized normalizing flows. This work challenges existing limitations in unsupervised disentanglement and offers a new theoretical foundation that could enhance model performance in applications requiring reliable factor recovery.