Linear Algebra I MT22, Cauchy-Schwarz Inequality


Flashcards

Can you state the Cauchy-Schwarz Inequality in full?

For $v, w$ in an inner product space $V$, then

\[ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert \]

The Cauchy-Schwarz inequality states

For $v, w$ in an inner product space $V$, then $ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $.

When does equality hold?

When $v, w$ are linearly dependent, i.e. $v + t _ 0 w = 0$.

In the proof of the Cauchy-Schwarz inequality

For $v, w$ in an inner product space $V$, then $ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $.

Why can you assume $w \ne 0$?

Because if $w = 0$ then the result is immediate.

In the proof of the Cauchy-Schwarz inequality

For $v, w$ in an inner product space $V$, then $ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $.

What statement do you start with that the whole proof follows from?

\[0 \le \vert \vert v+tw \vert \vert ^2\]

In the proof of the Cauchy-Schwarz inequality

For $v, w$ in an inner product space $V$, then $ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $.

You get to a stage where you have

\[0 \le \langle v,w \rangle + 2t\langle v,w \rangle + t^2\langle w,w \rangle\]

What do you do next?

Use the fact the discriminant $b^2 - 4ac \le 0$.

In the proof of the Cauchy-Schwarz inequality

For $v, w$ in an inner product space $V$, then $ \vert \langle v, w\rangle \vert \le \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $.

You get to a stage where you have

\[\begin{aligned} 0 &\le \vert \vert v+tw \vert \vert ^2 \\\\ &=\langle v,v \rangle + 2t\langle v,w \rangle + t^2\langle w,w \rangle \end{aligned}\]

If equality holds in Cauchy-Schwarz, i.e. $ \vert \langle v, w\rangle \vert = \vert \vert v \vert \vert \text{ } \vert \vert w \vert \vert $, why are $v, w$ linearly dependent? (Assume $w \ne 0$.)

Equality in Cauchy-Schwarz means the discriminant of the quadratic in $t$ is zero, so there is a (repeated) real root $t = t _ 0$. At this $t _ 0$, $ \vert \vert v + t _ 0 w \vert \vert ^2 = 0$, hence $v + t _ 0 w = 0$ (norms are positive definite), i.e. $v = -t _ 0 w$. So $v, w$ are linearly dependent. (If $w = 0$, dependence is immediate via $0 \cdot v + 1 \cdot w = 0$.)

Bite-sized

Cauchy-Schwarz inequality: $ \vert \langle v, w \rangle \vert \le $ $\ \vert v\ \vert \cdot \ \vert w\ \vert $. Equality holds iff $v, w$ are linearly dependent. Proof uses $0 \le \ \vert v + tw\ \vert ^2$ as a quadratic in $t$ with non-positive discriminant.

Source: Basicblock cards above in this entry; NLA MT25 §1.5 of the lecture notes.

@bite~

Two key consequences of Cauchy-Schwarz in NLA?

(1) Triangle inequality: $\|v + w\| \le \|v\| + \|w\|$ follows by squaring and using Cauchy-Schwarz on the cross term $\langle v, w \rangle$. (2) $x^\top y \le \|x\| _ 2 \|y\| _ 2$ (vector dot product) is the standard inner-product specialisation, used throughout NLA: bounds for matrix-vector products, residual bounds in LS and CG, etc. The matrix version $\|AB\| _ F \le \|A\| _ F \|B\| _ F$ also has Cauchy-Schwarz flavour.

Source: NLA MT25 §1.5 of the lecture notes (Cauchy-Schwarz proof).

@bite~