# $\frac{d}{dx}(b^TAx)$ where $b, x \in R^{n\times 1}$ and $A \in R^{n\times n}$

How do you differentiate the following expressions with respect to the vector $x$.

I think I might be a little conceptually confused on what happens when you take the derivative with respect to a vector. What dimensions should you end with? For the problem, I am trying to solve, I think I should end up for each of these derivatives with the results in the derivative to be $R^{1\times1}$.

1. $\frac{d}{dx}(b^TAx)$

2. $\frac{d}{dx}(x^TAb)$

3. $\frac{d}{dx}(x^TAx)$

where $b, x \in R^{n\times 1}$ and $A \in R^{n\times n}$

Also, if it helps A for this case is symmetric.

Update:

Thanks to the extra motivation by Mike and joriki, I think I now have them solved.

1. $\frac{d}{dx}(b^TAx) = Ab$

2. $\frac{d}{dx}(x^TAb) = Ab$

3. $\frac{d}{dx}(x^TAx) = \textbf{A}\textbf{x} + \textbf{A}^T\textbf{x}$

But if anyone, would like to double check it that would be great.

-
Wikipedia's article on Matrix calculus may help with the definition of vector derivatives. – Mike Wierzbicki Oct 19 '11 at 19:30

This is to advocate simultaneously for some more rigorous notations and for an easy way to solve these questions by using as much as possible the wonders of linear algebra.

First, if $U$ is a function defined on $\mathbb R^n$ with values in $\mathbb R$, the notation $\frac{d}{dx}U(x)$ is strange when $n\geqslant2$ because $\frac{d}{dx}$ usually denotes the derivative operator applied to a function from $\mathbb R$ to $\mathbb R$. For example, if $u$ is defined on $\mathbb R$ by $u(x)=\mathrm e^{x^2}$, I know that $\frac{d}{dx}u(x)=2x\mathrm e^{x^2}$ but if $U$ is defined on $\mathbb R^n$ by $U(x)=\mathrm e^{\|x\|^2}$, I do not know the meaning of the notation $\frac{d}{dx}U(x)$ when $n\geqslant2$.

Here, one looks for the gradient $\nabla U(x)$ of $U$ at $x$. In relaxed terms, $\nabla U(x)$ is a vector in $\mathbb R^n$ but in fact $\nabla U(x)$ is a linear form defined on $\mathbb R^n$ (this $\mathbb R^n$ is the tangent vector space of the manifold $\mathbb R^n$ at $x$). Rigorously speaking, $\nabla U(x):\mathbb R^n\to\mathbb R$ is defined by the fact that, for every vector $v$ in $\mathbb R^n$, $$\nabla U(x)(v)=\lim\limits_{h\to0}\frac1h(U(x+hv)-U(x)),$$ if the limit exists. This defines a linear function $\nabla U(x):\mathbb R^n\to\mathbb R$ and the identification of this function with an element $w$ of $\mathbb R^n$ comes through the identification of the tangent space of the manifold $\mathbb R^n$ at $x$ with the vector space $\mathbb R^n$ itself through the choice of a vector basis. This basis $B$ defines a scalar product on $\mathbb R^n$ by $(v_1,v_2)\mapsto v_1^Tv_2$ thanks to the decompositions of $v_1$ and $v_2$ in $B$, and one gets the relation $$\nabla U(x)(v)=w^Tv.$$ Such a vector $w$ is often denoted $w=\mathrm{grad}\ U(x)$ and both $\nabla$ and $\mathrm{grad}$ are pronounced gradient. Thus writing $v$ and $w$ in the basis $B$ as $v=(v_i)_i$ and $w=\left(\frac{\partial U}{\partial x_i}(x)\right)_i$, one gets $$\nabla U(x)(v)=w^Tv=\sum\limits_iw_iv_i=\sum\limits_i\frac{\partial U}{\partial x_i}(x)\ v_i.$$ To sum up everything above:

• To compute $\nabla U(x)=w$ is to write $U(x+hv)=U(x)+hw^Tv+o(h)$ for every $v$ in $\mathbb R^n$ when $h$ in $\mathbb R$ goes to $0$.
• The fact that $\nabla U(x)=w$ is equivalent to the fact that $\nabla U(x)(v)=w^Tv$ for every vector $v$, which is equivalent to the fact that $\dfrac{\partial U}{\partial x_i}(x)=w_i$ for every $i$.

Let us now compute the gradient of your examples. We will make a heavy use of the fact that for every matrices $C$ and $D$ of suitable dimensions, $(CD)^T=D^TC^T$ and of the fact that $z^T=z$ for every $1\times1$ matrix (also known as a number), but of pretty much nothing else. Here we go.

1. If $U(x)=b^TAx$, $U(x+hv)-U(x)=h(b^TAv)=h(A^Tb)^Tv$ hence $\nabla U(x)=A^Tb$.

2. If $U(x)=x^TAb$, $U(x+hv)-U(x)=h(v^TAb)=h(Ab)^Tv$ hence $\nabla U(x)=Ab$.

3. If $U(x)=x^TAx$, $U(x+hv)-U(x)=h(v^TAx+x^TAv)+h^2v^TAv$ hence $\nabla U(x)(v)=v^TAx+x^TAv=(Ax)^Tv+(A^Tx)^Tv$ hence $\nabla U(x)=Ax+A^Tx=(A+A^T)x$.

To transform these considerations into some concrete formulas, let us compute the coordinates of the gradient in cases 1. and 3. In case 1., one gets $$\frac{\partial U}{\partial x_i}(x)=(A^Tb)_i=\sum\limits_j(A^T)_{ij}b_j=\sum\limits_jA_{ji}b_j,$$ and in case 3., $$\frac{\partial U}{\partial x_i}(x)=((A+A^T)x)_i=\sum\limits_j(A+A^T)_{ij}x_j=\sum\limits_j(A_{ij}+A_{ji})x_j.$$

-
To sum up, your answers to 2. and 3. are correct and your answer to 1. almost correct... – Did Oct 20 '11 at 16:59
A real tour-de-France*, er, I mean, tour-de-force. – gary Oct 20 '11 at 17:04

A good way to get a feel for these things (and make sure you get the right answer) is to start out the pedestrian way, write out the expressions in coordinate form with summations over indices, differentiate, and then rewrite the result in matrix form. That makes the rules less mysterious, and when you've done it a couple of times it becomes obvious how to do it without writing it out.

[Update:]

In your solutions, numbers 2 and 3 are correct (except the use of boldface should be consistent on both sides, including the derivative).

You can see that 1 and 2 can't both be correct because $b^TAx$, which is symmetric (being just a number), is the same as $x^TA^Tb$, so 2 implies that that its derivative must be $A^Tb$. I suggest to go back to the coordinate form for 1 and check whether you've correctly transformed the summation in the derivative into a matrix product. You'll probably find that the summation was over the wrong index and you need to take the transpose of $A$ to make the indices come out right for a matrix multiplication.

-
Thanks joriki, I am trying that right now. – tryingsomemaths Oct 19 '11 at 19:49
@tryingsomemaths: I updated my answer in response to your update. – joriki Oct 20 '11 at 15:32