Not mathematically, at the very least. Perplexity is a translation of the best measure we have for informing us how a model is doing empirically over a test dat

Not mathematically, at the very least. Perplexity is a translation of the best measure we have for informing us how a model is doing empirically over a test dataset (both pre and post). It is enough to be, usably, at least the final word on how different quantization methods perform. Subjective ratings are different, but for compression things are quite well defined.