What's the gain for the multi:softmax in the xgboost R package?

I’m studying in xgboost recently. There are kinds of gain in multi:softmax:
Gain 1 - maximum default gain over all classes : max {G^2/H+lamda} over all classes
Gain 2 - maximum leaf score over all classes : max {-G/H+lamda} over all classes
Gain 3 - sum of defaults gains over all classes : sum {G^2/H+lamda} for all classes

There are different pros and cons in each kinds of gain form in multi:softmax. I would like to know which kinds of gain is in xgboost R package. And the reason why you choose it to be the gain form?

Thank you!

You can look at my master’s thesis to help your study: https://drive.google.com/file/d/0B0c0MbnP6Nn-eUNRRkVOOGpkbFk/view?usp=sharing&resourcekey=0-nVw3WhovKW5FPvPM5GPHfg