Let $A$ be positive-definite and $C$ diagonal positive-definite, consider the problem of solving the following equation for $\bf x$ $$A{\bf x}+C\begin{bmatrix} e^{x_1} \\ \vdots \\ e^{x_n} \end{bmatrix}=b.$$

I believe nonlinear conjugate gradient can solve this, however the nonlinear part is pretty benign as far as nonlinearities go, is there a trick to put this into a form where conjugate gradient could solve it?


The trick is to view your equation as the first-order optimality conditions of an unconstrained optimization problem $\min_x f(x)$. Recall that the first-order optimality conditions are just $\nabla f(x) = 0$. So in your case, you only need to set $$ f(x) = -b^T x + \tfrac{1}{2} x^T A x + \sum_{i=1}^n C_{ii} e^{x_i}, $$ where I'm assuming $A$ is symmetric (you didn't specify) and $C_{ii}$ is the $i$-th diagonal element of $C$.

  • $\begingroup$ right, basically I was just wondering if I could somehow linearize the above equation by introducing more variables or constraints or something. For instance setting $y_i=e^{x_i}$ and then having this be a constraint on the problem, or something like that. $\endgroup$ – Thoth Mar 19 '16 at 1:36
  • $\begingroup$ The nonlinear conjugate method doesn't solve constrained problems. $\endgroup$ – Dominique Mar 20 '16 at 18:36

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.