The theory predicts that expected risk and expected return should be positively related. But no one has convincingly proved this. The results are very sensitive to how you determine the expectations of risk and return and the timeline you use. Many have shown that there isn't a positive trade-off between risk and return in a CAPM-framework (i.e., Fama and French 1992 showed that $\beta$ has no predictive power in the cross section of returns -- although in 1993 they "saved" $\beta$ by adding two more factors).
What is odd about this Baker et al research is that they're buying the 20% with the highest vol and holding them for only one year. They're are constantly investing in the most uncertain stocks without ever really allowing the uncertainty to get resolved. I would be interested in how these results change with a two to five year holding period (e.g., I imagine in the mid 90s both amazon.com and pets.com where volatile; if you held both for one year, then you may or may not come out ahead; but if you held both for five years, pets.com delists and amazon.com skyrockets; but this is anecdotal and far from convincing.)