The major limitations of decision tree approaches to data analysis that I know of are:
- Provide less information on the relationship between the predictors and the response.
- Biased toward predictors with more variance or levels.
- Can have issues with highly collinear predictors.
- Can have poor prediction accuracy for responses with low sample sizes.
Are there any others? Are they robust to traditional statistical assumptions such as homogeneity, normality, independence?