Theoretical construction supports the fact that large language models can express any labeling of training data