[Home]

Table of contents


Some ineqalities

Some ineqalities

Here we shall discuss the proofs of some inequalities. Actually, we are interested in just one inequality, the last one. We shall need the first two to prove the last one.

Let $p\in[1, \infty)$ be any real number. Consider this function from ${\mathbb R}^n\rightarrow{\mathbb R}$ $$ \|\vec a \|_p= \left(\sum_{i=1}^n |a_i|^p \right)^{1/p}. $$ As you may guess, We are just imitating the definition of Euclidean norm of $\vec a.$ Our hope is that this is also a norm. The first conditions of being a norm are easy to check. The third condition (triangle inequality) is what we shall attack here. In this context triangle inequality is called Minkowski inequality. The proof, I am sorry to admit, is not very intuitive or elegant (in my humble opinion). We start with another inequality first.

Young's inequality

We already have a $p\in[1, \infty).$ Define $q = p/(p-1).$ Clearly, $q\in[1, \infty)$ and $\frac1p+\frac1q=1.$ Young's inequality says: If $a,b>0,$ then $$\frac{a^p}{p}+\frac{b^q}{q} \geq ab.$$ Looks pretty wierd! But the proof is simple. The LHS is a sum and the RHS is a product. So taking logarithm naturally comes to mind. The log of the LHS is $$ \log\left(\frac{a^p}{p}+\frac{b^q}{q}\right). $$ Notice that, since $\frac1p+\frac1q=1,$ so the LHS was just a convex combination of $a^p$ and $b^q.$ Since $\log x$ is a concave function, we must have $$ \log\left(\frac{a^p}{p}+\frac{b^q}{q}\right) \geq \frac{\log a^p}{p} + \frac{\log b^q}{q}=\log a + \log b = \log (ab). $$ Hence the inequality is proved (since $\log x$ is strictly increasing).

Holder's inequality

This is the second step in proving the triangle inequality (or Minkowski inequality). We are still working with the same $p,q$ as above.

Holder's inequality says: If $\vec a,\vec b\in{\mathbb R}^n$ then $$ \|\vec a\|_p\|\vec b\|_q\geq \sum|a_ib_i|. $$ Rather complicated looking! Let us simplify it a bit before trying the prove it. Divide both sides by the LHS (if the LHS is zero, then the result is trivial, anyway) to get $$ 1 \geq \sum A_i B_i, $$ where $A_i = |a_i|/\|\vec a\|_p$ and $B_i = |b_i|/\|\vec b\|_q.$

Just apply Young's inequality to get $$ A_iB_i \leq \frac{A_i^p}{p}+\frac{B_i^q}{q}. $$ Summing both sides over $i,$ the RHS becomes $\frac1p+\frac1q=1,$ completing the proof.

Minkowski inequality

Now at last we in a position to proof Minkowski inequality, which is just the triangle inequality for $\|\cdots\|_p:$ $$ \|\vec a+\vec b\|_p \leq \|\vec a\|_p+\|\vec b\|_p. $$ The proof is somewhat indirect. We start with the LHS raised to the power $p$: $$ \|\vec a+\vec b\|_p^p = \sum |a_i+b_i|^p= \sum |a_i+b_i|\cdot|a_i+b_i|^{p-1}. $$ Now apply triangle inequality to only the $|a_i+b_i|$ part to get $$ \sum |a_i+b_i|\cdot|a_i+b_i|^{p-1} \leq \sum (|a_i|+|b_i|)\cdot|a_i+b_i|^{p-1} = \sum|a_i|\cdot|a_i+b_i|^{p-1} +\sum |b_i|\cdot|a_i+b_i|^{p-1}. $$ Next, we shall apply Holder's inequality to each of these terms. For this we shall introduce $q$ as earlier. Also, it might help to write $c_i = (a_i+b_i)^{p-1}.$

For example, application of Holder's inequality on the first term gives $$ \sum|a_ic_i| \leq \|\vec a\|_p\|\vec c\|_q. $$ Similarly, the second term gives $$ \sum|b_ic_i| \leq \|\vec b\|_p\|\vec c\|_q. $$ So we get $$ \sum|a_i|\cdot|a_i+b_i|^{p-1} +\sum |b_i|\cdot|a_i+b_i|^{p-1} \leq (\|\vec a\|_p+\|\vec b\|_p)\cdot\|\vec c\|_q. $$ Thus we have got $$ \|\vec a+\vec b\|_p^p \leq (\|\vec a\|_p+\|\vec b\|_p)\cdot\|\vec c\|_q. $$ The result now follows on slight rearrangement.

HTML Comment Box is loading comments...