In the standard recipe for Unsupervised Domain Adaptation (UDA), we have a labeled source dataset (e.g., photos of cars) and an unlabeled target dataset (e.g., sketches of cars). The goal is to learn a model that works on the target domain. The intuitive solution is distribution alignment: transform the features so that the source and target distributions look identical. If the model can't tell the difference between a photo and a sketch, it should theoretically apply its knowledge of "cars" to both.
However, a growing body of research shows that this intuition is dangerously incomplete. Even if you perfectly align two distributions, you have no guarantee that you have mapped the correct labels to the correct features. This is the identifiability problem.
The Alignment Trap
The fundamental issue is that UDA is "ill-posed." As Tyler Lu, Teresa Luu, and Dávid Pál proved in their impossibility theorems, it is theoretically impossible to guarantee success using only labeled source data and unlabeled target data.
Why? Because there are infinitely many ways to transform a feature space. You might find a transformation that makes the target data look exactly like the source data, but it might "swap" the classes. For example, your model might align the distribution of "cats" in the source to "dogs" in the target. Because you have no labels in the target domain, the algorithm has no way to realize it has made a semantically catastrophic error.
Ishaan Gulrajani and Tatsunori Hashimoto highlight this "trust issue" in their work on identifiability conditions. They point out that aligning $P(X)$ (the features) does not mean you have aligned $P(Y|X)$ (the relationship between features and labels). Without additional constraints, the model is free to choose any "spurious map" that satisfies the alignment criteria.
Measuring the Hardness of Ambiguity
If the problem is fundamentally ambiguous, can we at least measure how ambiguous it is? Zhiyi Dong and Yongyi Mao propose a metric called Posterior Target Label Uncertainty (PTLU) in their information-theoretic perspective on UDA.
PTLU quantifies the "hardness" of a domain adaptation task by looking at how many conflicting but consistent labeling functions exist for the target data. If the available data allows for multiple valid-looking mappings that produce different predictions, the task has high PTLU. This serves as a provable lower bound for risk—if the PTLU is high, no algorithm, no matter how sophisticated, can guarantee a correct answer.
Seeking Identifiability through Structure
To move past these impossibility results, researchers are looking for specific structural assumptions that make the mapping "identifiable" (unique and correct).
One promising direction is partial disentanglement. Lingjing Kong, Shaoan Xie, and colleagues introduced the iMSDA framework, which assumes the data is generated by a combination of invariant "content" (the object) and domain-specific "style" (the environment). By mathematically proving that these latent components can be separated, they show that we can "un-distort" the domain shift. This structural assumption provides the missing information needed to distinguish the correct mapping from the infinite sea of incorrect ones.
Another approach involves co-regularization. In Co-DA, Abhishek Kumar and Prasanna Sattigeri use multiple classifiers that must "agree" on the unlabeled target data. The intuition is that while a single model might easily find a spurious mapping, it is much harder for multiple diverse models to all converge on the same incorrect mapping. This consensus acts as a proxy for the missing labels.
From Generative Models to Semi-Supervision
The identifiability problem also haunts generative models. Aditya Grover and Stefano Ermon demonstrated with AlignFlow that even with exact cycle consistency and maximum likelihood objectives, models can still converge on semantically meaningless correspondences.
When these unsupervised methods reach their theoretical limits, the only remaining solution is to introduce a small amount of supervision. Bo Li and Dongsheng Li argue in their work on LIRR that semi-supervised domain adaptation—using just a few target labels—is often necessary to resolve the ambiguities that purely unsupervised methods cannot touch. These few labels act as "anchors" that fix the mapping and prevent the model from drifting into semantically incorrect alignments.