To prove that if φ(x) satisfies equation (1), then we have
4φ(4x1+x2+x3+x4)⩽2φ(2x1+x2)+2φ(2x3+x4)⩽φ(x1)+φ(x2)+φ(x3)+φ(x4)
and so on. Thus, we have proved that for a special sequence of n, i.e., n=2m, we have
φ(nx1+x2+⋯+xn)⩽nφ(x1)+φ(x2)+⋯+φ(xn)
To prove that equation (4) holds universally, it suffices to show that if it holds for n, then it also holds for n−1. (Here we use reverse induction. For a more direct proof following Cauchy's method, refer to Jensen's related articles.) Thus, assume that equation (4) holds for n numbers, and now consider n−1 numbers x1,x2,⋯,xn−1. Let xn be the arithmetic mean of these n−1 numbers (with equal weights), and apply equation (4), we get
φ(A)=φ(n(n−1)A+A)=φ(nx1+x2+⋯+xn−1+A)⩽nφ(x1)+φ(x2)+⋯+φ(xn−1)+φ(A)
Thus, we have
φ(A)⩽n−1φ(x1)+φ(x2)+⋯+φ(xn−1).
Next, for non-negative rational numbers r1,r2,⋯,rn,r1+r2+⋯+rn=1, there exists a natural number m and non-negative integers p1,p2,⋯,pn, such that m=p1+p2+⋯+pn, and ri=mpi (i=1,2,⋯,n). Now, according to equation (4), we have
φ(mp1x1+p2x2+⋯+pnxn)⩽mp1φ(x1)+p2φ(x2)+⋯+pnφ(xn),
Thus, φ(∑rixi)⩽∑riφ(xi).
Finally, substituting the rational approximations ri for each pi and taking the limit, we obtain equation (3).