The terms in Vaughan's identity are not random sums: they are convolutions of simpler and/or shorter range arithmetic functions. Convolution allows the separation of variables in the sum, and hence it allows a more efficient estimation of the sum. A classical and prime example is [Dirichlet's hyperbola method][1]. 


  [1]: http://planetmath.org/dirichlethyperbolamethod