Question about Attention

[SRDdev]

I some basic questions as a student.
I have implemented Transformers multiple times but still learning new things about them. So these are the questions

As seen in the attention maps , only a few values contribute to the final output the maximum.

1. Isn't this like a long tail distribution where only a few values have very high values and the rest are very low ?
2. If that is the case then , can we remove some parts on the input query randomly (say 25%) and still achive same results ?

Thank you !