0
$\begingroup$

I want to take the first and second derivative of the following matrix quadratic form,

$$Q = {\phi}(\beta)'W{\phi}(\beta),$$

where, $\phi(\beta)$ is a $q \times 1$ vector, $\beta$ is a $k \times 1$ vector, and W is a $q \times q$ symmetric matrix, then Q is a scalar. So, I would expect the first derivative Q w.r.t to $\beta$ to be a $k \times 1$ vector $\frac{\partial{Q}}{\partial{\beta}}$, and the second derivative to be a $k \times k$ matrix $\frac{\partial^2{Q}}{\partial{\beta}\partial\beta'}$.

Then, taking the first derivative of Q w.r.t $\beta$ yields,

$$\frac{\partial{Q}}{\partial{\beta}} = 2\frac{\partial{\phi}'}{\partial{\beta}}W\phi(\beta),$$

which is $k \times 1$ as expected. Then, the second derivative yields me,

$$\frac{\partial^2{Q}}{\partial{\beta}\partial\beta'} = 2\left(\frac{\partial^2{\phi(\beta)'}}{\partial{\beta}\partial\beta'}W\phi(\beta) + \frac{\partial{\phi(\beta)'}}{\partial\beta}W\frac{\partial{\phi(\beta)}}{\partial\beta'}\right),$$

which should be $k \times k$. The second term inside the parenthesis is $k \times k$, but the first one isn't? It doesn't quite make sense to me, but it can be of the right size because $W$ and $\phi$ give together $q \times 1$, while $\frac{\partial^2{\phi(\beta)'}}{\partial{\beta}\partial\beta'}$ I don't know what shape it might have ($\frac{\partial{\phi(\beta)'}}{\partial{\beta}}$, is $k \times q$, but differentiating such a matrix w.r.t. a $1 \times k$ vector..., how does that work? I don't see how would it be compatible with the terms that follow it).

In any case, I know it is wrong, because I have a formula outlining what the correct form should be, but I don't understand how to arrive to it or how does it work. According to this formula, the second derivative should be,

$$\frac{\partial^2{Q}}{\partial{\beta}\partial\beta'} = 2\left(({\phi}(\beta)'W\otimes I_k)\frac{\partial vec(\partial\phi(\beta)'/\partial\beta)}{\partial{\beta'}}+\frac{\partial{\phi(\beta)'}}{\partial\beta}W\frac{\partial{\phi(\beta)}}{\partial\beta'}\right).$$ Here, $vec(\partial\phi(\beta)'/\partial\beta)$ is the vectorisation of a $k \times q$ matrix, which would be, to my understanding, a $(k \times q)\times 1$ vector(?). Differentiating this vector w.r.t. a $1 \times k$ vector $\beta'$ would yield a $(k \times q)\times k$ matrix. Furthermore, following my understanding of the Kronecker product, $W \otimes I_k$ would yield a $(k \times q)\times (k \times q)$ matrix. And all this would be alright except the term $\phi(\beta)'$ is still a $1 \times q$ vector and thus incompatible with the other terms following -- unless we do $\phi(\beta)W$ and then follow with $I_k$ yielding a $1 \times q$ vector $\otimes$ a $k \times k$ matrix, but I can't anywhere where it says the Kronecker product works between a vector and a matrix (and if so, what shape would the result then have? How would I compute it?)

I don't know where I am going wrong, and I would appreciate it very much if someone could guide me through this conundrum.

$\endgroup$
  • $\begingroup$ I haven't read all what you wrote but I just would like to remark that you can treat your matrix $W$ to be diagonal because $W = U^TDU$ with a diagonal $D$ and an orthogonal $U$. Hence, $Q = \phi^TW\phi = (U\phi)^TD(U\phi) = \psi^T D\psi = \sum d_i\phi_i^2$, where $\psi = U\phi$. I have $Q'' = 2\sum d_i[(\nabla\psi_i)(\nabla\psi_i)^T + \psi_i\cdot H_i]$, where $H_i$ is the Hessian of $\psi_i$. $\endgroup$ – amsmath Aug 28 '19 at 17:11
  • $\begingroup$ Your question is similar to this: "What is the second derivative of $f(x)$?" Without specifying the function, the best answer you will get is the blindingly obvious: $f''(x)$. So if you expect a useful answer to this question, you'll need to specify the functional dependence of the vector $\phi$ upon the vector $\beta$. $\endgroup$ – greg Aug 29 '19 at 19:50
  • $\begingroup$ @greg $\phi(\beta)$ = $Z'(y-\exp(X\beta))$, where $Z$ is a $n \times q$ matrix, $y$ is a $n \times 1$ vector, $X$ is a $n \times k$ matrix, and $\beta$ is a $k \times 1$ vector. Then, $$Q = (Z'(y - e^{X\beta})'W(Z'(y - e^{X\beta}))$$ $\endgroup$ – porcos Aug 29 '19 at 21:57
1
$\begingroup$

For typing convenience, define the following variables $$\eqalign{ &A = ZWZ^T, \quad &b = \beta, \quad &e = \exp(Xb), \quad &E = {\rm Diag}(e) \\ & &p = (e-y), \quad &r = Ap, \quad &R = {\rm Diag}(r) \\ }$$ And the differentials $$\eqalign{ dp &= de = e\odot(X\,db) = EX\,db \\ dr &= A\,dp = AEX\,db \\ }$$ where the elementwise/Hadamard product $(\odot)$ of vectors was replaced by ordinary matrix multiplication by a diagonal matrix. This is a standard trick, e.g. $$Er = e\odot r = r\odot e = Re$$

Write the quadratic form in terms of these new variables. Then find its differential and gradient. $$\eqalign{ Q &= p^TZWZ^Tp &= p^Tr \\ dQ &= p^Tdr + r^Tdp &= 2r^TEX\,db \\&= (2X^TEr)^Tdb \\ g=\;\frac{\partial Q}{\partial b} &= 2X^TEr &= 2X^TRe \\ }$$ Now find the differential of the gradient, and then the hessian. $$\eqalign{ dg &= 2X^TE\,dr + 2X^TR\,de \\ &= 2X^TEAEX\,db + 2X^TREX\,db \\ &= 2X^T(EA + R)EX\,db \\ H =\; \frac{\partial g}{\partial b} &= 2X^T(EA + R)EX \\ }$$

The hessian is symmetric since the matrices $(A,E,R,W)$ are all symmetric, and the matrices $(E,R)$ are diagonal (and therefore commute with each other).

$\endgroup$
  • $\begingroup$ Thanks so much! That was really helpful!. Just one more thing, for $p = (e - y)$, shouldn't it be $(y - e)$ as I originally wrote? Do you change the order for some reason? I think that writing it $(y-e)$ shouldn't change anything (except maybe the sign?), right? $\endgroup$ – porcos Sep 6 '19 at 11:28
  • $\begingroup$ It avoids carrying around a bunch of minus signs in the derivation, but makes no difference to the quadratic form since $$\eqalign{ Q&=p^TZWZ^Tp \\&= (-p)^TZWZ^T(-p) }$$ $\endgroup$ – greg Sep 6 '19 at 13:50

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.