Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I have found something about the product rules for matrix-functions in https://ccrma.stanford.edu/~dattorro/matrixcalc.pdf $$ \frac{d(f(x)^Tg(x))}{dx}=\frac{df(x)}{dx}\cdot g(x)+\frac{dg(x)}{dx}\cdot f(x) $$ I verify this in the example list in http://www.psi.toronto.edu/matrix/calculus.html. For example: $$ \frac{d (Ax+b)^TC(Dx+e)}{dx} = A^TC(Dx+e) + D^TC^T(Ax+b) $$ $$ \frac{d (x^TCx)}{dx} = (C+C^T)x $$

But when I met this one, I'm confused... $$ \frac{d (a^TX^TXb)}{dX} = X(ab^T + ba^T) $$ Following the formula in matrixcalc.pdf, I get this $$\begin{align} f(x)&=Xa\\ g(x)&=Xb\\ \frac{d (a^TX^TXb)}{dX} &= \frac{df(x)}{dx}\cdot Xb+\frac{dg(x)}{dx}\cdot Xa \\ &=a^TXb+b^TXa\end{align}$$ which is different with the correct result. I don't know what I'm doing wrong, please help me.. Thanks!

share|improve this question
    
Pretty nearly unreadable. Maybe you want to consult meta.math.stackexchange.com/questions/5020/… and/or meta.math.stackexchange.com/questions/1773/… –  Gerry Myerson Oct 1 '12 at 7:32
    
In the first equation, do you mean: $$\frac{d(f(x)^T g(x))}{dx}=\frac{d(f(x))}{dx}\cdot g(x)+\frac{d(g(x))}{dx}\cdot f(x)$$ Can you please use $^T$ for transpose, since $'$ is confusing when derivatives are also of concern... –  Dennis Gulko Oct 1 '12 at 7:40
    
Gerry Myerson: Yes, in the first equation, I mean that. Dennis Gulko: I'm sorry, I'm being here for the first time, I have corrected that... –  areslp Oct 1 '12 at 7:51
add comment

2 Answers

up vote 2 down vote accepted

Coming back to the definitions and considering $u:X\mapsto a^TX^TXb$, one looks for a linear function $v_X:H\mapsto v_X(H)$ such that $u(X+H)=u(X)+v_X(H)+o(H)$. Now, $$ u(X+H)=u(X)+a^TH^TXb+a^TX^THb+a^TH^THb, $$ hence the gradient of $u$ at $X$ is the linear function $v_X$ defined by $$ v_X(H)=a^TH^TXb+a^TX^THb. $$ Note that this is not $H\mapsto XFH$, for any matrix $F$. However, one can express $v_X(H)$ as the trace of a matrix, as follows.

The matrices $Ha$, $Hb$, $Xa$ and $Xb$ are all column vectors and, for any column vectors $C$and $D$, $C^TD$ is simply a scalar hence $C^TD=D^TC$. In particular, $a^TH^TXb=b^TX^THa$. Using the fact that the trace of a $1\times1$ matrix is its unique coefficient and the fact that $\mathrm{tr}(CD)=\mathrm{tr}(DC)$ and $\mathrm{tr}(C+D)=\mathrm{tr}(D)+\mathrm{tr}(C)$ for every matrices $C$ and $D$ of suitable sizes, one gets $$ v_X(H)=\mathrm{tr}(H^TXba^T+b^TH^TXa)=\mathrm{tr}(H^T(Xba^T+Xab^T)). $$ If one wishes to call derivative of a function $u$ at $X$ any matrix $W_X$ such that, for every $H$, $v_X(H)=\mathrm{tr}(H^TW_X)$ (see more about this in the comments), then, in the case at hand, the derivative at $X$ is $$ W_X=X(ba^T+ab^T). $$

share|improve this answer
    
Why should it be $H\mapsto XFH$? The derivative with respect to $X$ is a matrix $A$ such that $v_X(H)=\operatorname{tr}(AH^T)$, and by the invariance of the trace under cyclic rotation and transposition, $A=X(ab^T+ba^T)$ indeed gives your $v_X(H)$. –  joriki Oct 1 '12 at 8:33
    
@joriki We have a problem of definition here. To me, the differential of any real valued function $u$ at any point $x$ is the unique linear function $v_x$ such that $u(x+h)=u(x)+v_x(h)+o(h)$. In the case at hand, it happens, as you write, that $v_X(H)=\mathrm{tr}(A_XH^T)$ for some matrix $A_X$, but I fail to see why this would make $A_X$ the differential of $u$ at $X$. –  Did Oct 1 '12 at 8:39
    
But the question is about the derivative, not the differential. The derivative is the "constant of proportionality" in the differential, and my trace formula is just a convenient way of expressing that proportionality for the differential of a function of a matrix. See equation $(1725)$ on p. 659 of the PDF file linked to in the question for an equivalent definition of what they call the "gradient" with respect to $X$. –  joriki Oct 1 '12 at 8:47
    
@joriki Thanks for the pointer. I am not sure I subscribe to this terminology nor that it is universal nor that it is useful. I agree that the trace formula you mention is a trick to express concisely the differential. Unfortunately, I feel that this brevity hinders the learning of these notions more than it helps (as exemplified by nearly every question on this site about multidimensional differential calculus), which is why I prefer to come back to the definitions. But hey, one may disagree with such a stance... –  Did Oct 1 '12 at 8:57
    
I see your point and I agree about the confusion apparent in the questions asked; but I think it goes too far to say that the formula is wrong when it's correct according to the definition being used, irrespective of the merits of that definition. –  joriki Oct 1 '12 at 9:37
show 1 more comment

When calculating a derivative of a matrix function, I strongly recommend to use the most basic equation: $\mathrm{d}(f(X)g(X))=\mathrm{d}f(X)g(X)+f(X)\mathrm{d}g(X)$. I don't recommend you to use any high-level equations. On one hand you need to remember them or search them before you can use, on the other hand, you must be very clear about the conditions under which these high-level equations are valid.

First, in the first equation you give, I believe the $x$ should be a vector. But in the problem that confuses you, the $X$ is a matrix. I don't think the first equation can be applied in this case.

Second, $a^TX^TXb$ is a scalar and $X$ is a matrix, the derivative of $a^TX^TXb$ with respect to $X$ should be a matrix instead of a scalar (see eq. 1725 in 'matrix calculus').

Third, I will use the most basic equation I give in the beginning to calculate the derivative of $a^TX^TXb$ with respect to $X$.

$$\mathrm{d}(a^TX^TXb) =a^T\mathrm{d}X^TXb+a^TX^T\mathrm{d}Xb =b^TX^T\mathrm{d}Xa+a^TX^T\mathrm{d}Xb\\ =\mathrm{tr}(ab^TX^T\mathrm{d}X+ba^TX^T\mathrm{d}X) =\mathrm{tr}[(ab^T+ba^T)X^T\mathrm{d}X]$$

Then $$\frac{\mathrm{d}(a^TX^TXb)}{\mathrm{d}X}=X(ba^T+ab^T)$$

share|improve this answer
    
The PDF file linked to in the question does claim on p. 661 that the formula is applicable to matrix variables. –  joriki Oct 1 '12 at 8:34
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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