Let $A$ and $B$ be orthogonal matrices, and $x_1$, $x_2$ be vectors. What are:

$$ \dfrac{\partial}{\partial x_1}\left( x_1^TAB^Tx_2 \right)$$

$$ \dfrac{\partial}{\partial x_2}\left( x_1^TAB^Tx_2 \right)$$

Are they

$$\dfrac{\partial}{\partial x_1}\left( x_1^TAB^Tx_2 \right) = AB^Tx_2 $$

$$\dfrac{\partial}{\partial x_2}\left( x_1^TAB^Tx_2 \right) = B A^Tx_1 $$

  • $\begingroup$ Do $A$ and $B$ depend on $x_1,x_2$? If not then it looks good! Just take care with whether or not you need the derivative to be a row or a column. $\endgroup$ – vb628 May 4 at 7:49
  • 1
    $\begingroup$ These are gradients, not quite partial derivatives. $\endgroup$ – Rodrigo de Azevedo May 4 at 11:17


$$f ({\rm x}, {\rm y}) := {\rm x}^\top {\rm A} \, {\rm B}^\top {\rm y} = \langle {\rm x} , {\rm A} \,{\rm B}^\top {\rm y} \rangle = \langle {\rm B} \,{\rm A}^\top {\rm x}, {\rm y} \rangle$$


$$\begin{aligned} \nabla_{{\rm x}} f ({\rm x}, {\rm y}) &= \color{blue}{{\rm A} \,{\rm B}^\top {\rm y}} \\ \nabla_{{\rm y}} f ({\rm x}, {\rm y}) &= \color{blue}{{\rm B} \,{\rm A}^\top {\rm x}} \end{aligned}$$


Yes, with the convention that all results are column vectors, then what you have written is true. In fact, for any constant matrix $A$ and constant column vector $b$,

$$ \frac{\partial}{\partial x} x^T A b = Ab $$

Thus, we have

$$ \frac{\partial}{\partial x_1} x_1^T AB^T x_2 = \frac{\partial}{\partial x_1} x_1^T (AB^T) x_2 = AB^T x_2 $$

Since $ x_1^T AB^T x_2 $ is a scalar, $$ x_1^T AB^T x_2 = (x_1^T AB^T x_2)^T = x_2^T BA^T x_1 $$

and so

$$ \frac{\partial}{\partial x_2} x_1^T AB^T x_2 = \frac{\partial}{\partial x_2} x_2^T BA^T x_1 = BA^T x_1 $$

Whether $A, B$ are orthogonal is irrelevant.


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.