In my text book, Cauchy-Schwarz Inequality is extended to a positive definite matrix.

But I neither understand what the relationship between Cauchy-Schwarz Inequality and a positive definite matrix nor the sentence underlined in red, I am not a strong linear algebra person, I have just started studying it.

I'd like post the page here so that I might get a better explanation. Thank you.

enter image description here


As a complement to the other answers, I just wanted to note that reasoning a bit more abstractly you can get this result from the ordinary Cauchy-Schwarz inequality for free, without any computation.

The Cauchy-Schwarz inequality is valid in any inner product space (Euclidean vector space if you are working over the real numbers). In a given an inner product space (such as $\Bbb R^n$ with the standard inner product), what a positive definite symmetric linear operator $B$ (given by a positive definite symmetric matrix in the case of $\Bbb R^n$) gives you is precisely another inner product on the same space, defined by $(x\mid y)_B\overset{\rm def}=(x\mid By)=( Bx\mid y)$. Now the Cauchy-Schwarz inequality holds for this inner product as well: $$ (x\mid y)_B^2 \leq (x\mid x)_B ~~(y\mid y)_B \qquad\text{with equality only if $x \parallel y$} $$ (the notation $x \parallel y$ here means the vectors $x,y$ are linearly dependent, one being a scalar multiple of the other). Now taking $\def\b{\mathbf b}x=\b$ and $\def\d{\mathbf d}y=B^{-1}\d$, this becomes $$ (\b\mid B^{-1}\d)_B^2 \leq (\b\mid \b)_B ~~(B^{-1}\d\mid B^{-1}\d)_B \qquad\text{with equality only if $\b \parallel B^{-1}\d$.} $$ which expanding the definition of $(\mid )_B$ simplifies to $$ (\b\mid\d)^2 \leq (\b\mid B\b) ~~(\d\mid B^{-1}\d) \qquad\text{with equality only if $\b \parallel B^{-1}\d$,} $$ which is your "extended" Cauchy-Schwarz inequality.


The sentence you underlined in red abbreviates these steps:

$$(\mathbf{B}^{1/2} \mathbf{b} \cdot \mathbf{B}^{-1/2} \mathbf{d})^2 \leq (\mathbf{B}^{1/2} \mathbf{b} \cdot \mathbf{B}^{1/2} \mathbf{b})(\mathbf{B}^{-1/2} \mathbf{d} \cdot \mathbf{B}^{-1/2} \mathbf{d}) $$

This is the Cauchy-Schwarz inequality applied as requested. Now we use the transpose-based definition of the dot product $(\mathbf{a} \cdot \mathbf{b}) = \mathbf{a}^T \mathbf{b}$:

$$(\mathbf{b}^T (\mathbf{B}^{1/2})^T \mathbf{B}^{-1/2} \mathbf{d})^2 \leq (\mathbf{b}^T (\mathbf{B}^{1/2})^T \mathbf{B}^{1/2} \mathbf{b}) (\mathbf{d}^T (\mathbf{B}^{-1/2})^T \mathbf{B}^{-1/2} \mathbf{d})$$

Finally we take advantage of the the following observations:

  • $\mathbf{B}$ is positive definite and therefore symmetric
  • $\mathbf{B}^{1/2}$ is also positive definite and symmetric
  • The inverse of a positive definite matrix is positive definite

As a result:

$$(\mathbf{B}^{1/2})^T \mathbf{B}^{-1/2} = \mathbf{B}^{1/2} \mathbf{B}^{-1/2} = \mathbf{I}$$

$$(\mathbf{B}^{1/2})^T \mathbf{B}^{1/2} = \mathbf{B}^{1/2} \mathbf{B}^{1/2} = \mathbf{B}$$

$$(\mathbf{B}^{-1/2})^T \mathbf{B}^{-1/2} = \mathbf{B}^{-1/2} \mathbf{B}^{-1/2} = \mathbf{B}^{-1}$$

Completing the proof.

You can see how this relates to the regular Cauchy-Schwarz inequality (for the dot product) by considering what it states when $\mathbf{B} = \mathbf{I}$ (Since $\mathbf{I}$ is a positive definite matrix). In this case we have

$$(\mathbf{b} \cdot \mathbf{d})^2 \leq (\mathbf{b} \cdot \mathbf{b}) (\mathbf{d} \cdot \mathbf{d})$$

Which is precisely the usual Cauchy-Schwarz inequality for the dot product.


Isn't CSI $x'y\leq \sqrt{x'\cdot x}\sqrt{y'\cdot y}$? Square it to get $$(x'\cdot y)^2\leq (x'\cdot x)(y'\cdot y)$$ Now apply this inequality by letting the vectors $x=B^{1/2} b$ and $y=B^{-1/2}d.$ You get $$b'\cdot d \leq (b'Bb)(d'B^{-1}d)$$ as required.


If you put $B = I$ then you recover the usual Cauchy–Schwarz inequality, so this is a generalization. You can write the Cauchy–Schwarz inequality in the form $$ (x'y)^2 \leq (x'x)(y'y). $$ Taking $x = B^{1/2} b$ and $y = B^{-1/2} d$, we get $$ (b'd)^2 = ((B^{1/2} b)'(B^{-1/2} d))^2 \leq ((B^{1/2}b)'B^{1/2}b)((B^{-1/2}d)'D^{-1/2}d) = \\ (b' B^{1/2} B^{1/2} b) (d'B^{-1/2} B^{-1/2} d) = (b' B b) (d' B^{-1} d). $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.