Let $\mathbf{A}$ be a ${n\times J}$ matrix with $A_{ij}\geq0$ (and $A_{ij}>0$ for most $ij$, there cannot be any rows or columns that consist only of $0$s), $Q=\left\{\mathbf{q}\mid \mathbf{q}\in\Bbb R^n_+\land\sum\limits_{i=1}^n q_i=1\right\}$ and an $0<\alpha<1$. Abusing notation, let $\mathbf{q}^{\frac{1}{\alpha}}=\bigl(q_1^\frac{1}{{\alpha}},\dots,q_n^\frac{1}{{\alpha}}\bigr).$

Consider the set $K=\{\mathbf{k}\in\Bbb R^J_+\mid \mathbf{k}=\mathbf{A}^{\top}\mathbf{q}^\frac{1}{{\alpha}}\land \mathbf{q}\in Q\}.$ There exist $J$ members of $K$ such that the $j$-th component of $\underline{\mathbf{k}}^j$ is lower than the $j$-th component of any other $\mathbf{k}\in K$. These can be easily identified using calculus.

The $\mathbf{k}\in K$ that I'm interested in can be found as follows

  1. Choose some $\beta \in\Bbb R^J_+$ such that $\sum\limits_{j=1}^J b_j=1,$
  2. Compute $\widehat{\mathbf{k}} =\sum\limits_{j=1}^J\beta_j\underline{\mathbf{k}}^j,$
  3. Determine the proportions $\mathbf{r}=\frac{1}{\hat k_1}\widehat{\mathbf{k}}$
  4. Let $k_j=\sum\limits_{i=1}^nA_{ij}q_i^{\frac{1}{\alpha}}$ and minimize $k_1$ subject to $\mathbf{k}\in K$ and $k_j=r_j k_1 \quad \forall j=\{2,\dots,J\}.$ If $k^*_1$ is the minimum of this problem, then $\mathbf{k}^*=(k^*_1,r_2 k^*_1,\dots,r_J k^*_1)$ lies on the frontier.

@MikeY suggested that I could solve instead the following problem:

$$\begin{align} \min\limits_{\{q_i\}}&\quad \delta^{\top}\mathbf{A}^{\top}\mathbf{q}^\frac{1}{{\alpha}}\\ \text{s.t.}&\quad\begin{cases} \sum\limits_{i=1}^n q_i=1 \\ q_i \geq 0 & i=\{1,\dots,n\} \\ \end{cases} \end{align}$$

for some $\delta \in \Bbb R_+^J$ such that $\sum\limits_{j=1}^J \delta_j=1$. If $\mathbf{q}^*$ is the minimand of this problem, then $\mathbf{k}^*=\mathbf{A}^{\top}\left(\mathbf{q}^*\right)^\frac{1}{{\alpha}}$ lies on the frontier. I've tried this approach, that is computationally much more efficient, and it works fine in the sense that it results in a frontier that is indistinguishable from the one obtained with my approach.

My questions are:

  1. Why is this second approach equivalent to mine? I can see that if $\delta^j$ is a vector of zeroes except for the $j-th$ component that is a $1$, @MikeY's approach yields $\underline{\mathbf{k}}^j$, but I don't see why this is also the case for other points in the frontier.
  2. What is the relationship between $\beta$ and $\delta$? For $J=2$, @MikeY suggests that $\delta=(\beta,1-\beta)$. My exploration of a $J=3$ case, indicates that the relationship is not so obvious.
  3. Ideally, I'd like to be able to determine the $\delta$ that corresponds to any $\widehat{\mathbf{k}}$ that is a convex combination of the $\underline{\mathbf{k}}^j$s.

Edit: @MikeY's suggestion is the third answer to my question in Stack Exchange Mathematica site.

Edit 2: We can write the Lagrangian of the first problem to ressemble that of the second (ommiting the non-negativity constraints is of no consequence): $$\begin{align}\mathcal L &=k_1+\sum\limits_{j=2}^J \lambda_j (k_j -k_1 r_j) +\gamma \left(1-\sum\limits_{i=1}^nq_i\right)\\ & =\left(1-\sum\limits_{j=2}^J \lambda_j r_j\right)k_1 + \sum\limits_{j=2}^J \lambda_j k_j+\gamma \left(1-\sum\limits_{i=1}^nq_i\right)\end{align}.$$ One could set $\rho_1=1-\sum\limits_{j=2}^J \lambda_j r_j$ and $\rho_j=\lambda_j \quad j=\{2,\dots,J\},$ observe that $$\sum\limits_{j=1}^J\rho_j=1-\sum\limits_{j=2}^J\lambda_j(r_j-1),$$ and write $\delta_j=\frac{\rho_j}{\sum\limits_{j=1}^J\rho_j}$ which would result in an equivalent lagrangian (in the sense that the minimand is the same)

$$ \mathcal L' =\sum\limits_{j=1}^J \delta_j k_j+\gamma' \left(1-\sum\limits_{i=1}^nq_i\right)$$ The problem is that, although these deltas do add up to one, $\delta_1$ can be negative, in which case (some) $\delta_j$s can be larger than one. Further, I'd like to be able to obtain $\delta$ without having to compute the $\lambda_j$s first.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.