4
$\begingroup$

I had this question when reading section 3.5.3 on page 170 of "Pattern Recognition and Machine Learning" written by Christopher M. Bishop:

enter image description here

Here $\mathbf\Phi$ represents the design matrix and $\beta$ is precision which is positive.

My question is: why is $\mathbf\Phi^\top\mathbf\Phi$ a positive definite matrix?

I tried to prove it by showing that the error function defined in equation (3.80) of the same book has a local minimum at $\mathbf m_N$. But I don't have the luck (I can prove $\nabla E$ is zero at $\mathbf m_N$, but I cannot show further that $\mathbf m_N$ is necessarily a local minimum) and, even if it is the case, I cannot establish positive definiteness of $\mathbf\Phi^\top\mathbf\Phi$ from that of $\mathbf A$. So, can you please help me prove that $\mathbf\Phi^\top\mathbf\Phi$ is positive definite? Thanks a lot.

Edit: The columns of the design matrix $\mathbf\Phi$ are linearly independent almost surely, because the elements of this matrix are (functions of) random variables. This condition is essential in excluding positive semi-definiteness. As an aside, it is usually assumed that the number of samples $N$ is larger than the number of basis functions $M$.

$\endgroup$
2
  • 1
    $\begingroup$ Can you say more about $\Phi$? In general, for a real-valued matrix $X$, we have that $X^T X$ is positive semi-definite: $$ z^T X^T Xz = \| X z \|_2^2 \ge 0 $$ for some vector $z \neq 0$. The intuition is that the sums of squares of real numbers must be non-negative. So there must be some additional property to $\Phi$. As a hint, the key is to work out what property $\Phi$ has to make the claim true. $\endgroup$
    – Sycorax
    Commented Oct 15, 2022 at 14:34
  • 3
    $\begingroup$ The product of X transpose and X is a gram matrix en.m.wikipedia.org/wiki/Gram_matrix. You can prove these matrices are positive definite. $\endgroup$ Commented Oct 15, 2022 at 14:55

2 Answers 2

8
$\begingroup$

Note we have $$ v^\top \left(\Phi^\top\Phi\right) v = \left(v^\top \Phi^\top\right) \left(\Phi v\right) = \left(\Phi v \right)^\top \left(\Phi v\right) = \|\Phi v \|_2^2 \geq 0 $$ for all $\Phi \in \mathbb{R}^{N \times M}, v \in \mathbb{R}^M; M, N \in \mathbb{N}_{>0}$. This establishes the general positive-semidefiniteness of $\Phi^\top\Phi$.
Moreover, $$ \|\Phi v \|_2^2 = 0 \iff \|\Phi v \|_2 = 0 \iff \Phi v \ = 0_{\mathbb{R}^N}, $$ which means that $\Phi^\top\Phi$ and hence $\beta\Phi^\top\Phi$ (with $\beta \in \mathbb{R_{>0}}$) is positive-definite if (and only if) the columns of $\Phi$ are linearly independent.

$\endgroup$
11
  • 1
    $\begingroup$ semi-definite... $\endgroup$
    – usεr11852
    Commented Oct 15, 2022 at 15:16
  • $\begingroup$ @usεr11852 No, positive definite. If $\Phi$ has full column rank then $\Phi v \neq 0_{ \mathbb{R}^N}$ and thus $v^\top \left(\Phi^\top\Phi\right) v = \|\Phi v \|_2^2 > 0$ for all $v \in \mathbb{R}^M \setminus \left\{ 0_{ \mathbb{R}^M} \right\}$ $\endgroup$
    – statmerkur
    Commented Oct 15, 2022 at 15:26
  • 2
    $\begingroup$ You don't say that about $\Phi$ in your answer nor is mentioned in the OP's question. $\endgroup$
    – usεr11852
    Commented Oct 15, 2022 at 15:38
  • 1
    $\begingroup$ @usεr11852 I do say that in my answer: "is positive definite if the columns of $\Phi$ are linearly independent" $\endgroup$
    – statmerkur
    Commented Oct 15, 2022 at 15:42
  • 2
    $\begingroup$ This answer might be clearer if you state up front that the claim in the question is not true in general, unless we make an additional assumption about $\Phi$. $\endgroup$
    – Sycorax
    Commented Oct 15, 2022 at 15:44
8
$\begingroup$

Observation $1.$ A symmetric matrix $\mathbf A^\mathsf T\mathbf A$ is positive semi-definite.

Note that

$$\mathbf v^\mathsf T\mathbf A^\mathsf T\mathbf A\mathbf v= (\mathbf A\mathbf v)^\mathsf T\cdot ( \mathbf A\mathbf v)\geq 0.\tag 1$$

Also, $\mathbf A\mathbf v=\mathbf 0$ implies $\mathbf A^\mathsf T\mathbf A$ being positive definite when $\mathcal N(\mathbf A) =\{\mathbf 0\}$, i.e., when $\mathbf A$ has linearly independent columns.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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