0
$\begingroup$

Suppose multivariate function $f(x_1, x_2, \ldots, x_n) = \sum\limits_{i=1}^n\ln \left( 1+e^{\,-c_i\sum\limits_{j=1}^n {d_{ij}} x_j} \right)$, where $c_i$ is either $1$ or $-1$ and $d_{ij}$ is arbitrary real number.


I'd like to simplify the gradient of $f(X)$ for the purpose of gradient-based optimization, specifically, I attempt to further simplify $\dfrac{\partial f}{\partial x_k} = \sum\limits_{i=1}^n \dfrac{e^{\,-c_i\sum\limits_{j=1}^n d_{ij} x_j} (-c_id_{ik})} {1+e^{\,-c_i\sum\limits_{j=1}^n d_{ij} x_j }}$.


Of course, if it's possible to simplify in any matrix way, I would be glad to know since many science softwares support matrix algebra. But is it possible to further simplify the gradient?

$\endgroup$
1
$\begingroup$

Use a series of variable substitutions to build up to the function $$\eqalign{ y &= Ax &\ \ \ dy = A\,dx \cr z &= -c\circ y &\ \ \ dz = -c\circ dy \cr e &= \exp(z) &\ \ \ de = e\circ dz \cr h &= 1+e &\ \ \ dh = de \cr g &= \log(h) &\ \ \ dg = \frac{dh}{h} \cr f &= 1:g &\ \ \ df = 1:dg \cr }$$ where the {Hadamard, Frobenius} products have been denote by the symbols {$\,\circ, :\,$} respectively.

Also note the use of $A_{ij}=d_{ij}$ so as to avoid confusion with derivatives or differentials.

Continuing the expansion of that last differential $$\eqalign{ df &= 1:dg \cr &= 1:\frac{dh}{h} \cr &= 1:\frac{e\circ dz}{h} \cr &= 1:\frac{(1-h)\circ c\circ dy}{h} \cr &= (h^{-1}-1)\circ c:dy \cr &= (H^{-1}-I)\,c:A\,dx \cr &= A^T(H^{-1}-I)\,c:dx \cr\cr }$$ where $H={\rm Diag}(h)$.

So the gradient could be written as $$\eqalign{ \frac{\partial f}{\partial x} &= A^T(H^{-1}-I)\,c \cr }$$

$\endgroup$

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.