A while back, we had mentioned the AM-GM inequality:

a1a2anna1++ann.\sqrt[n]{a_1a_2\cdots a_n}\leq \frac{a_1 + \cdots +a_n}{n}.

I had said something like, there are many great resources out there where you can find a proof of the inequality. And so, I skipped the proof! But this will stand no longer! I hope that akickinthediscovery.com will be one of those “great” resources. In order to achive this, we will prove a more general version of the AM-GM inequality:

Theorem (Generalized AM-GM): Let a1,a2,,an,p1,p2,,pn0a_1,a_2,\cdots,a_n,p_1,p_2,\cdots,p_n \in \R_{\geq0} such that p1+p2++pn=1.p_1+p_2+\cdots+p_n=1. Then,

a1p1a2p2anpnp1a1+p2a2+pnan.a_1^{p_1}a_2^{p_2}\cdots a_n^{p_n}\leq p_1a_1 + p_2a_2\cdots +p_na_n.

Remark: Note that when p1=p2==pn=1/np_1=p_2 = \cdots = p_n = 1/n, we recover the standard AM-GM inequality.

George Pólya’s Proof

The proof we are going to give is one that the great mathematician George Pólya discovered. He said that this proof came to him in a dream, which is very impressive! This proof is like a good movie; there is suspense and a twist at the end. I hope you enjoy it!

This proof relies on the following observation/proposition:

Proposition: For any x,x\in \R, we have 1+xex.1+x \leq e^x.

However, we will require a quick lemma before we prove the above proposition.

Lemma (Bernoulli’s Inequality): For all x[1,)x\in [-1,\infty) and nn\in \N we have

1+nx(1+x)n.1+nx \leq (1+x)^n.

Proof of Lemma: (Click in the Discovery)

Let x[1,)x\in [-1,\infty). We will prove Bernoulli’s inequality by induction on n.n. Our base case is immediate: 1+x(1+x)1.1+x \leq (1+x)^1. Now assume that 1+nx(1+x)n1+nx \leq (1+x)^n for some nn\in \N and consider 1+(n+1)x.1+(n+1)x. This induction step is one of those cases where we discover what we need to do by working backwards until we get what we want, and then when we type out the proof, we write it forwards, obscuring how we might have come up with the argument. However, I hope that you will allow me to simply give the tidy proof without the justification (in fact, it might be a fun challenge to see if you can figure out how you might come up with it!)

Observe the following steps:

1+(n+1)x1+(n+1)x+nx2becausenx20=1+nx+x+nx2=(1+nx)(1+x)(1+x)n(1+x)=(1+x)n+1.1+(n+1)x \leq 1+(n+1)x +nx^2\qquad \mathrm{because}\; nx^2\geq0 \\ \qquad\qquad\;\;\;\;\;= 1+nx+x +nx^2 \\ \qquad\qquad\;\;\;\;\;=(1+nx)(1+x)\\ \qquad\qquad\;\;\;\;\;\leq(1+x)^n(1+x)\\ \qquad\qquad\;\;\;\;\;=(1+x)^{n+1}.

Closing the induction.

*Where did we use the fact that x[1,)x\in [-1,\infty)? We used it when we said: (1+x)n(1+x)(1+nx)(1+x).(1+x)^n (1+x) \geq (1+nx)(1+x). This is because x[1,)x\in [-1,\infty) implies 1+x01+x\geq 0.

\square

Proof Idea of Proposition: (Click in the Discovery)

Let x.x\in \R. Recall that exe^x (Euler’s number) is given by the following limit:

ex=limn(1+xn)n.e^x= \lim_{n\rightarrow \infty}{(1+\frac{x}{n})^n}.

Letting x/nxx/n \mapsto x in Bernoulli’s inequality gives the desired result.

\square

Proof of the AM-GM

Now onto the fun part!

Theorem (Generalized AM-GM): Let a1,a2,,an,p1,p2,,pn0a_1,a_2,\cdots,a_n,p_1,p_2,\cdots,p_n \in \R_{\geq0} such that p1+p2++pn=1.p_1+p_2+\cdots+p_n=1. Then,

a1p1a2p2anpnp1a1+p2a2+pnan.a_1^{p_1}a_2^{p_2}\cdots a_n^{p_n}\leq p_1a_1 + p_2a_2\cdots +p_na_n.

Proof: (Click in the Discovery)

Let a1,a2,,an,p1,p2,,pn0a_1,a_2,\cdots,a_n,p_1,p_2,\cdots,p_n \in \R_{\geq0} such that p1+p2++pn=1.p_1+p_2+\cdots+p_n=1.

First, notice that xex1x\leq e^{x-1} by our proposition. Furthermore, xk(ex1)k=ekxk.x^k\leq (e^{x-1})^k = e^{kx-k}. Hence,

a1p1a2p2anpnep1a1p1ep2a2p2epnanpn=ep1a1+p2a2+pnan(p1+p2+pn)=ep1a1+p2a2+pnan1sincep1+p2++pn=1.=e(i=1npiai)1a_1^{p_1}a_2^{p_2}\cdots a_n^{p_n}\leq e^{p_1a_1-p_1}e^{p_2a_2-p_2}\cdots e^{p_na_n-p_n}\\\qquad \qquad \;\;\;\;\;\,=e^{p_1a_1+p_2a_2 + \cdots p_na_n-(p_1+p_2+ \cdots p_n)}\\ \qquad \qquad \;\;\;\;\;\,=e^{p_1a_1+p_2a_2 + \cdots p_na_n-1} \qquad \mathrm{since}\;p_1+p_2+\cdots+p_n=1.\\\qquad \qquad \;\;\;\;\;\,=e^{\Big(\sum_{i=1}^np_ia_i\Big)-1}

We now have an upperbound for a1p1a2p2anpn.a_1^{p_1}a_2^{p_2}\cdots a_n^{p_n}. How might we go on from here? This is were out twise comes in! We will normalize our sequence a1,,ana_1,\cdots,a_n to obtain the desired result.

Let’s now consider a sequence b1,b2,,bn,0b_1,b_2,\cdots,b_n, \in \R_{\geq0} where b1=a1/Ab_1 =a_1 /A and A=p1a1+p2a2+pnan. A = p_1a_1 + p_2a_2\cdots +p_na_n. Using what we’ve already established,

(a1A)p1(a2A)p2(anA)pn=b1p1b2p2bnpne(i=1npibi)1=e(i=1npiai)/A1=e11=e0=1.\Big(\frac{a_1}{A}\Big)\,^{p_1}\cdot \Big(\frac{a_2}{A}\Big)\;^{p_2}\cdots \,\Big(\frac{a_n}{A}\Big)\,^{p_n}={b}_1^{p_1}b_2^{p_2}\cdots b_n^{p_n}\leq e^{\Big(\sum_{i=1}^np_ib_i\Big)-1}=e^{\Big(\sum_{i=1}^np_ia_i\Big)/A-1} = e^{1-1} = e^0 = 1.

It follows that,

a1p1a2p2anpnAp1++pn1.\frac{a_1^{p_1}\cdot a_2^{p_2}\cdots a_n^{p_n}}{A^{p_1+\cdots + p_n}}\leq 1.

Recalling p1+p2++pn=1,p_1+p_2+\cdots+p_n=1,we deduce

a1p1a2p2anpnAp1++pn1.\frac{a_1^{p_1}\cdot a_2^{p_2}\cdots a_n^{p_n}}{A^{p_1+\cdots + p_n}}\leq 1.

concluding the proof.

\square

How Wonderful!

I find Pólya’s proof to be absolutely wonderful. It has a great step when we think we are stuck, and yet, with some clever thinking, we are able to push through!

I hope you had some fun with today’s proof! There is also a wonderful geometric proof of the AM-GM inequality: ab12(a+b).\sqrt{ab}\leq \frac{1}{2}(a+b). But that’s for another day!

Be Kind. Be Curious. Be Compassionate. Be Creative.

And Have Fun!

Leave a comment