The process of gradient descent involves moving towards smaller error in an 
error function.

The first order linear approximation of the function is:

For a vector u,
E(u+delta_u)=E(u)+Sum(i,dE/du_i*delta_u_i)

With the constraint that |delta_u|=C (a constant)

=> delta_u_i=1/C*dE/du_i