Defaults to 0.0--1.0 and 0.05. But after init probing run, user will get an idea of a sub-range that looks interesting: they will then rerun the algorithm but drill down on a narrower range at smaller steps to find a roughly optimal perplexity value.
The 2017 pre-print on arXiv: Automatic Selection of t-SNE Perplexity Have two parameters:
Defaults to 0.0--1.0 and 0.05. But after init probing run, user will get an idea of a sub-range that looks interesting: they will then rerun the algorithm but drill down on a narrower range at smaller steps to find a roughly optimal perplexity value.
Some explanatory text of perplexity: An Introduction to t-SNE with Python Example