There are $n$ multiplication coefficients for which optimized values are searched. There is a restriction that the multiplication coefficients are not allowed to be negative and must have a sum of $1.0$:

$\sum_{i =0}^n c_i = 1.0$

The coefficients are multiplied by an individual vector and summed up to a new vector:

$\pmb v^{sum} = \sum_{i =0}^n c_i \cdot \pmb v_i$

Finally, the mean squared error between $\pmb v^{sum}$ and a vector $\pmb u$ is formed:

$mse = \frac{1}{m}\sum_{i = 0}^m (u_i - v^{sum}_i)^2$

So I want to find optimized $c_i$ values for which the $mse$ is minimized.

What's the best way to do that?

  • $\begingroup$ Are the $c_n$ required to be nonnegative? $\endgroup$ – Rob Pratt Nov 20 '19 at 13:28
  • $\begingroup$ @RobPratt Yes, good hint. I have edited my text. $\endgroup$ – Anne Bierhoff Nov 20 '19 at 13:47

Geometrically, this is a projection onto the convex hull of the $\mathbf{v}_i$. You can solve it as a quadratic programming problem.

  • $\begingroup$ Thanks for your answer. I saw that ordinary least squares are a part of quadratic programming. Would ordinary least squares be a possible way to solve the problem? $\endgroup$ – Anne Bierhoff Nov 20 '19 at 18:14
  • $\begingroup$ This is ordinary least squares with some side constraints on the regression coefficients. $\endgroup$ – Rob Pratt Nov 20 '19 at 18:26
  • $\begingroup$ Currently, I have a hard time trying to translate the problem into a quadratic programming problem. So I created a new question: math.stackexchange.com/questions/3454546/… $\endgroup$ – Anne Bierhoff Nov 28 '19 at 13:58

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.