Assume that $R(k_1,k_2)$ is a $3 \times 3$ positive definite matrix for $k_1 = 0,1,k_2 =0,1$, $L(k_1,k_2)$ is a $3 \times 1$ vector for $k_1 = 0,1,k_2 =0,1$, $z$ is a scalar, and $\alpha_i(k) \in (0,1)$ for $i=1,2$, $k=0,1$. We want to find x and functions $f_1$ and $f_2$ that solve the following optimization problem:

\begin{align} \min_{x, f_1, f_2} \sum_{y_1 \in \{0,1\}} \sum_{y_2 \in \{0,1\}} \alpha_1(y_1) \alpha_2(y_2) \Big( \begin{bmatrix} x &f_1(y_1) & f_2(y_2) \end{bmatrix} R(y_1,y_2) \begin{bmatrix} x \\f_1(y_1) \\ f_2(y_2) \end{bmatrix} + \begin{bmatrix} x &f_1(y_1) & f_2(y_2) \end{bmatrix} L(y_1,y_2) z \Big). \end{align}

The minimizer $x$, $f_1$, and $f_2$ should be based on parameter $z$. Any idea for solving this problem?

The function we are optimizing over is the sum of four functions each of them is strictly convex. So, the sum of them is also strictly convex. If we define $s_1^0 = f_1(0)$, $s_1^1 = f_1(1)$, $s_2^0 = f_2(0)$, and $s_2^1 = f_2(1)$. Then we can write the optimization problem as optimizing over $x,s_1^0,s_1^1, s_2^0, s_2^1$ as follows which is a optimization over a strictly convex function of $x,s_1^0,s_1^1, s_2^0, s_2^1$. Right?

\begin{align} \min_{x,s_1^0,s_1^1, s_2^0, s_2^1} \Bigg[ &\alpha_1(0) \alpha_2(0) \Big( \begin{bmatrix} x &s_1^0 & s_2^0 \end{bmatrix} R(0,0) \begin{bmatrix} x \\s_1^0 \\ s_2^0 \end{bmatrix} + \begin{bmatrix} x &s_1^0 & s_2^0 \end{bmatrix} L(0,0) z \Big) \\ + &\alpha_1(0) \alpha_2(1) \Big( \begin{bmatrix} x &s_1^0 & s_2^1 \end{bmatrix} R(0,1) \begin{bmatrix} x \\s_1^0 \\ s_2^1 \end{bmatrix} + \begin{bmatrix} x &s_1^0 & s_2^1 \end{bmatrix} L(0,1) z \Big) \\ + &\alpha_1(1) \alpha_2(0) \Big( \begin{bmatrix} x &s_1^1 & s_2^0 \end{bmatrix} R(1,0) \begin{bmatrix} x \\s_1^1 \\ s_2^0 \end{bmatrix} + \begin{bmatrix} x &s_1^1 & s_2^0 \end{bmatrix} L(1,0) z \Big) \\ + &\alpha_1(1) \alpha_2(1) \Big( \begin{bmatrix} x &s_1^1 & s_2^1 \end{bmatrix} R(1,1) \begin{bmatrix} x \\s_1^1 \\ s_2^1 \end{bmatrix} + \begin{bmatrix} x &s_1^1 & s_2^1 \end{bmatrix} L(1,1) z \Big) \Bigg] \end{align}

This optimization can be written as

\begin{align} \min_{x,s_1^0,s_1^1, s_2^0, s_2^1} \Big( \begin{bmatrix} x &s_1^0 &s_1^1 &s_2^0 &s_2^1 \end{bmatrix} Q \begin{bmatrix} x \\ s_1^0 \\ s_1^1 \\ s_2^0 \\ s_2^1 \end{bmatrix} + \begin{bmatrix} x &s_1^0 &s_1^1 &s_2^0 &s_2^1 \end{bmatrix} P z \Big) \end{align} where $Q$ is a $5 \times 5$ matrix and $P$ is a $5 \times 1$ vector. Since the second optimizing optimization problem is a strictly convex function of $x,s_1^0,s_1^1, s_2^0, s_2^1$, this one should also be. Right? Does this mean that $Q$ is PD and invertible?

  • $\begingroup$ @OGC I added my idea to the question. I think that idea may work, but I wanted to know whether it works or there is any better approach? $\endgroup$ – m0_as May 14 '17 at 4:17
  • $\begingroup$ Just assemble Q and check its eigenvalues. $\endgroup$ – Michael Grant May 14 '17 at 23:27
  • $\begingroup$ Even if Q isn't positive semidefinite, it's an unconstrained quadratic, and can therefore be solved analytically. $\endgroup$ – Michael Grant May 14 '17 at 23:28
  • $\begingroup$ @MichaelGrant My problem is not numerical. I know that $R(\cdot,\cdot)$'s are PD, and $\alpha$'s belong to $(0,1)$. So, based on this I want to know whether $Q$ is PD. $\endgroup$ – m0_as May 15 '17 at 19:37
  • $\begingroup$ @MichaelGrant How can we solve a quadratic optimization problem when we don't know whether $Q$ is positive definite? If $Q$ is not PD (PSD) nor ND (NSD), then it may have lots of local mins and maxs. $\endgroup$ – m0_as May 15 '17 at 19:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.