3
$\begingroup$

Let $A=A^T\in \mathbb R^{k\times k}$ be a nonzero symmetric matrix and define $F:\mathbb R^k\to\mathbb R$ by $$f(x):=x^TAx$$ Then why $df(x)\xi=2x^TA\xi$ for $x,\xi\in\mathbb R^k$?

$\endgroup$
  • $\begingroup$ to begin to see, try the cases $k=1$, $k=2$ $\endgroup$ – janmarqz Dec 16 '14 at 17:07
  • 1
    $\begingroup$ Possibly duplicate, see here (very similar) $\endgroup$ – alexjo Dec 16 '14 at 17:11
5
$\begingroup$

It's because

\begin{align}df(x)\xi &= \frac{d}{dt}|_{t = 0} f(x + t\xi)\\ &= \frac{d}{dt}|_{t = 0} (x + t\xi)^T A(x + t\xi)\\ &= \frac{d}{dt}|_{t = 0} (x^T + t\xi^T) A(x + t\xi)\\ &= \frac{d}{dt}|_{t = 0} (x^T + t\xi^T)(Ax + tA\xi)\\ &= \frac{d}{dt}|_{t = 0} (x^TAx + t(\xi^T Ax + x^TA\xi) + t^2\xi^TA\xi)\\ &= \xi^T Ax + x^TA\xi\\ &= 2x^TA\xi \end{align}

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Because both terms are real scalars; and $a^T=a$ when $a$ is a scalar. $\endgroup$ – Michael Grant Dec 16 '14 at 17:52
  • $\begingroup$ @pxc3110 -- What inequality? $\endgroup$ – Robin Goodfellow Dec 16 '14 at 19:07
  • $\begingroup$ that's supposed to be equality, sorry :) $\endgroup$ – pxc3110 Dec 16 '14 at 19:13
2
$\begingroup$

There's a nice answer linked by alexjo using coordinates. Here's an answer without coordinates, using the fact that we know the derivative of a linear map:

Consider $h(x,y) = x^T A y$, with $h: \mathbb{R}^k \times \mathbb{R}^k \rightarrow \mathbb{R}$.

We can restrict $h$ to each factor with $h(x,-) : \{x\} \times \mathbb{R}^k \rightarrow \mathbb{R}$ and $h(-,y): \mathbb{R}^k \times \{y\} \rightarrow \mathbb{R}$.

Then $dh_{x,y} (\xi_x\oplus \xi_y) = dh_{x,y}(\xi_x \oplus 0) + dh_{x,y}(0\oplus \xi_y) = d(h(-,y))_x (\xi_x) + d(h(x,-))_y (\xi_y)$


Because $h(-,y)(x) = x^T A y = y^T A x$ is linear in $x$, we get $d(h(-,y))_x(\xi_x) = y^T A \xi_x$.

Similarly, because $h(x,-)(y) = x^T A y$ is linear in $y$, we get $d(h(x,-))_y(\xi_y) = x^T A \xi_y$.

Thus $dh_{x,y}(\xi_x \oplus \xi_y) = y^T A \xi_x + x^T A \xi_y$.


Finally, we have $f(x) = h(x,x) = h \circ \Delta$ for $\Delta: \mathbb{R}^k \rightarrow \mathbb{R}^k \times \mathbb{R}^k$ given by $\Delta(x) = (x,x)$.

We have $d\Delta_x(\xi) = \xi \oplus \xi$

Thus $df_x(\xi) = (dh_{x,x} \circ d\Delta_x)(\xi) = 2 x^T A \xi$.

| cite | improve this answer | |
$\endgroup$

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.