Difference between revisions of "Proofs of AM-GM"

Revision as of 15:54, 29 December 2021

This pages lists some proofs of the weighted AM-GM Inequality. The inequality's statement is as follows: for all nonnegative reals $a_1, \dotsc, a_n$ and nonnegative reals $\lambda_1, \dotsc, \lambda_n$ such that $\sum_{i=1}^n \lambda_i = 1$ , then $\sum_{i=1}^n \lambda_i a_i \ge \prod_{i=1}^n a_i^{\lambda_i},$ with equality if and only if $a_i = a_j$ for all $i,j$ such that $\lambda_i, \lambda_j \neq 0$ .

We first note that we may disregard any $a_i$ for which $\lambda_i= 0$ , as they contribute to neither side of the desired inequality. We also note that if $a_i= 0$ and $\lambda_i \neq 0$ , for some $i$ , then the right-hand side of the inequality is zero and the left hand of the inequality is greater or equal to zero, with equality if and only if $a_j = 0 = a_i$ whenever $\lambda_j\neq 0$ . Thus we may henceforth assume that all $a_i$ and $\lambda_i$ are strictly positive.

Proofs of Unweighted AM-GM

These proofs use the assumption that $\lambda_i = 1/n$ , for all integers $1 \le i \le n$ .

Proof by Cauchy Induction

We use Cauchy Induction, a variant of induction in which one proves a result for $2$ , all powers of $2$ , and then that $n$ implies $n-1$ .

Base Case: The smallest nontrivial case of AM-GM is in two variables. By the properties of perfect squares (or by the Trivial Inequality), $(x-y)^2 > 0,$ with equality if and only if $x-y=0$ , or $x=y$ . Then because $x$ and $y$ are nonnegative, we can perform the following manipulations: $x^2 - 2xy + y^2 > 0$ $x^2 + 2xy + y^2 > 4xy$ $\frac{(x+y)^2}{4} \geq xy$ $\frac{x+y}{2} \geq \sqrt{xy},$ with equality if and only if $x=y$ , just as before. This completes the proof of the base case.

Powers of Two: We use induction. Suppose that AM-GM is true for $n$ variables; we will then prove that the inequality is true for $2n$ . Let $x_1, x_2, \ldots x_{2n}$ be any list of nonnegative reals. Then, because the two lists $x_1, x_2 \ldots x_n$ and $x_{n+1}, x_{n+2}, \ldots x_{2n}$ , have $n$ variables each, $\frac{x_1 + x_2 + \cdots + x_n}{n} \geq \sqrt[n]{x_1 x_2 \cdots x_n} \textrm{ and } \frac{x_{n+1} + x_{n+2} + \cdots + x_{2n}}{n} \geq \sqrt[n]{x_{n+1} x_{n+2} \cdots x_{2n}}.$ Adding these two inequalities together and dividing by two yields $\frac{x_1 + x_2 + \cdots + x_{2n}}{2n} \geq \frac{\sqrt[n]{x_1 x_2 \cdots x_n} + \sqrt[n]{x_{n+1} x_{n+2} \cdots x_{2n}}}{2}.$ Then by AM-GM in two variables, $\frac{\sqrt[n]{x_1 x_2 \cdots x_n} + \sqrt[n]{x_{n+1} x_{n+2} \cdots x_{2n}}}{2} \geq \sqrt[2n]{x_1 x_2 \ldots x_{2n}}.$ Combining this inequality with the previous one yields unweighted AM-GM in $2n$ variables, with one exception — equality.

For equality, note that every inequality mentioned must have equality as well; thus, inequality holds if and only if all the numbers in $x_1, x_2, \ldots x_n$ are the same, all the numbers in $x_{n+1}, x_{n+2}, \ldots x_{2n}$ are the same, and $\sqrt[n]{x_1 x_2 \cdots x_n} = \sqrt[n]{x_{n+1} x_{n+2} \cdots x_{2n}}$ . From here, it is trivial to show that this implies $x_1 = x_2 = \cdots x_{2n}$ , which is the equality condition for unweighted AM-GM in $2n$ variables.

This completes the induction and implies that unweighted AM-GM holds for all powers of two.

Backward Step: Assume that AM-GM holds for $n$ variables. We will then use a substitution to derive AM-GM for $n-1$ variables. Letting $x_n = \frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}$ , we have that $\frac{x_1 + x_2 + \cdots + x_{n-1} + \frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}}{n} \geq \sqrt[n]{x_1 x_2 \cdots x_{n-1} \left(\frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}\right)}.$ Because we assumed AM-GM in $n$ variables, equality holds if and only if $x_1 = x_2 = \cdots x_{n-1} = \frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}$ . However, note that the last equality is implied if all the numbers of $x_1, x_2 \ldots x_{n-1}$ are the same; thus, equality holds if and only if $x_1 = x_2 = \cdots x_{n-1}$ .

We then simplify the lefthand side. Multiplying both sides of the fraction by $n-1$ and combining like terms, we get that $\frac{x_1 + x_2 + \cdots + x_{n-1} + \frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}}{n} = \frac{nx_1 + nx_2 + \cdots + nx_{n-1}}{n(n-1)} = \frac{x_1 + x_2 + \cdots x_{n-1}}{n-1}.$ Thus, $\frac{x_1 + x_2 + \cdots x_{n-1}}{n-1} \geq \sqrt[n]{x_1 x_2 \cdots x_{n-1} \left(\frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1} \right)}.$ Raising both sides to the $n$ th power yields $\left( \frac{x_1 + x_2 + \cdots x_{n-1}}{n-1}\right)^n \geq x_1 x_2 \cdots x_{n-1}\left(\frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}\right).$ From here, we divide by $\frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1}$ and take the $(n-1)^{\textrm{th}}$ root to get that $\frac{x_1 + x_2 + \cdots + x_{n-1}}{n-1} \geq \sqrt[n-1]{x_1 x_2 \cdots x_{n-1}}.$ Every step we took preserves our earlier equality condition, which proves that AM-GM holds for $n-1$ variables. This completes the backward step, which proves the unweighted AM-GM Inequality. $\square$

Proof by Rearrangement

Define the $n$ sequence $\{ r_{i,j}\}_{i=1}^{n}$ as $r_{i,j} = \sqrt[n]{a_i}$ , for all integers $1 \le i,j \le n$ . Evidently these sequences are similarly sorted. Then by the Rearrangement Inequality, $\sum_i a_i = \sum_i \prod_j r_{i,j} \ge \sum_i \prod_j r_{i+j,j} = n \prod_i \sqrt[n]{a_i} ,$ where we take our indices modulo $n$ , with equality exactly when all the $r_{i,j}$ , and therefore all the $a_i$ , are equal. Dividing both sides by $n$ gives the desired inequality. $\blacksquare$

Proof by Calculus

We will start the proof by considering the function $f(x) = \ln x - x$ . We will now find the maximum of this function. We can do this simply using calculus. We need to find the critical points of $f(x)$ , we can do that by finding $f'(x)$ and setting it equal to $0$ . Using the linearity of the derivative $f'(x) = \frac{1}{x} - 1$ . We need $f'(x) = 0$ $f'(x) = \frac{1}{x} - 1 = 0$ $\frac{1}{x} = 1$ $x = 1$ Note that this is the only critical point of $f(x)$ . We can confirm it is the maximum by finding it's second derivative and making sure it is negative. $f''(x) = -\frac{1}{x^2}$ letting x = 1 we get $f''(1) = -\frac{1}{1} < 0$ . Since the second derivative $f''(x) < 0$ , $f(1)$ is a maximum. $f(1) = \ln 1 - 1 = -1$ . Now that we have that $f(1) = -1$ is a maximum of $f(x)$ , we can safely say that $\ln x - x \leq -1$ or in other words $\ln x \leq x - 1$ . We will now define a few more things and do some manipulations with them. Let $A = \frac{\sum_{i=1}^{n} a_i}{n}$ , with this notice that $nA = \sum_{i=1}^{n} a_i$ . This fact will come into play later. now we can do the following, let $x = \frac{a_i}{A}$ and plug this into $\ln x < x - 1$ , we get $\ln\left(\frac{a_1}{A}\right) \leq \frac{a_1}{A} - 1$ $\ln\left(\frac{a_2}{A}\right) \leq \frac{a_2}{A} - 1$ $\vdotswithin{=}$ $\ln\left(\frac{a_n}{A}\right) \leq \frac{a_n}{A} - 1$ Adding all these results together we get $\ln\left(\frac{a_1}{A}\right) + \ln\left(\frac{a_2}{A}\right) + \dots + \ln\left(\frac{a_n}{A}\right) \leq \frac{a_1}{A} - 1 + \frac{a_2}{A} - 1 + \dots + \frac{a_n}{A} - 1$ $\ln\left(\frac{\prod_{i=1}^{n} a_i}{A^n}\right) \leq \frac{\sum_{i=1}^{n}a_i}{A} - n$ $\ln\left(\frac{\prod_{i=1}^{n} a_i}{A^n}\right) \leq \frac{nA}{A} - n$ $\ln\left(\frac{\prod_{i=1}^{n} a_i}{A^n}\right) \leq n - n$ $\ln\left(\frac{\prod_{i=1}^{n} a_i}{A^n}\right) \leq 0$ Now exponentiating both sides we get $e^{\ln\left(\frac{\prod_{i=1}^{n} a_i}{A^n}\right)} \leq e^{0}$ $\frac{\prod_{i=1}^{n} a_i}{A^n} \leq 1$ $\prod_{i=1}^{n} a_i \leq A^n$ $\sqrt[\leftroot{2}\uproot{3}n]{\prod_{i=1}^{n} a_i} \leq \sqrt[\leftroot{2}\uproot{3}n]{A^n}$ $\sqrt[\leftroot{2}\uproot{3}n]{\prod_{i=1}^{n} a_i} \leq A$ $\sqrt[\leftroot{2}\uproot{3}n]{\prod_{i=1}^{n} a_i} \leq \frac{\sum_{i=1}^{n} a_i}{n}$ This proves the AM-GM inequality. $\blacksquare$

Proof of Weighted AM-GM

Proof by Convexity

We note that the function $x \mapsto \ln x$ is strictly concave. Then by Jensen's Inequality, $\ln \sum_i \lambda_i a_i \ge \sum_i \lambda_i \ln a_i = \ln \prod_i a_i^{\lambda_i} ,$ with equality if and only if all the $a_i$ are equal. Since $x \mapsto \ln x$ is a strictly increasing function, it then follows that $\sum_i \lambda_i a_i \ge \prod_i a_i^{\lambda_i},$ with equality if and only if all the $a_i$ are equal, as desired. $\blacksquare$

Alternate Proof by Convexity

This proof is due to G. Pólya.

Note that the function $f:x \mapsto e^x$ is strictly convex. Let $g(x)$ be the line tangent to $f$ at $(0,1)$ ; then $g(x) = x+1$ . Since $f$ is also a continuous, differentiable function, it follows that $f(x) \ge g(x)$ for all $x$ , with equality exactly when $x=0$ , i.e., $1+x \le e^x,$ with equality exactly when $x=0$ .

Now, set $r_i = a_i/\biggl( \sum_{j=1}^n \lambda_j a_j \biggr) - 1,$ for all integers $1\le i \le n$ . Our earlier bound tells us that $r_i +1 \le \exp(r_i),$ so $(r_i +1)^{\lambda_i} \le \exp(\lambda_i r_i) .$ Multiplying $n$ such inequalities gives us

$\prod_{i=1}^n (r_i + 1)^{\lambda_{i}} \le \prod_{i=1}^n \exp \lambda_i r_i$

Evaluating the left hand side:

$\prod_{i=1}^n (r_i + 1)^{\lambda_{i}} = \frac{\prod_{i=1}^n a_i^{\lambda_i} }{ (\sum_{j=1}^n \lambda_j a_j)^{\sum_{j=1}^n \lambda_i} } = \frac{\prod_{i=1}^n a_i^{\lambda_i} }{ \sum_{j=1}^n \lambda_j a_j } ,$

for

$\sum_{j=1}^n \lambda_i = 1 .$

Evaluating the right hand side:

$\prod_{i=1}^n \exp \lambda_i r_i = \prod_{i=1}^n \exp \left(\frac{a_i \lambda_i}{ \sum_{i=1}^n \lambda_j a_j} - \lambda_i\right) = \exp \left(\frac{\sum_{i=1}^n a_i \lambda_i}{ \sum_{i=1}^n \lambda_j a_j} - \sum_{i=1}^n \lambda_i\right) = \exp 0 = 1 .$

Substituting the results for the left and right sides:

$\frac{\prod_{i=1}^n a_i^{\lambda_i} }{ \sum_{j=1}^n \lambda_j a_j } \le 1$

$\prod_{i=1}^n a_i^{\lambda_i} \le \sum_{j=1}^n \lambda_j a_j = \sum_{i=1}^n \lambda_i a_i ,$

as desired. $\blacksquare$

@@ Line 141: / Line 141: @@
 as desired. <math>\blacksquare</math>
-[[Category:Inequality]]
+[[Category:Algbera]]
+[[Category:Inequalities]]

Page

Toolbox

Search