2
$\begingroup$

Let $a_{ij}$ be the elements of a $n$-dimensional covariance matrix. Can we prove the following?

$$ 1-\sum_{k=1}^n a_{ik} \lambda_k + \sum_{j=1}^n \sum_{k=1}^n \lambda_j a_{jk} \lambda_k > 0, \qquad i \in \{1, \dots, n\} $$

where the $\lambda_k$ are constrained by

$$ \sum_{k=1}^n \lambda_k = 1, \quad \lambda_i > 0 \text { for } i \in \{1, \dots, n\}? $$

Or, in a more general way, what are the conditions that the elements of the covariance matrix should satisfy so that the above set of inequalities hold?

NOTE: in matrix form, if $A$ is a covariance matrix and $a_i$ is a row vector having the $i$-th row of $A$, the question is:

is $1-a_i \lambda + \lambda^T A \lambda >0 $ where $\lambda=[\lambda_1 \ldots \lambda_n]^T$ and $\lambda^T e_n =1$ with $\lambda_k >0$ and $e_n$ a column vector of ones? The question should be formulated as: find the conditions for the elements of matrix $A$ so that the inequality holds for $i=1,\dotsc, n$.

$\endgroup$
4
  • $\begingroup$ I have formulated it in matrix form. thank you for your interest Rodrigo, muito obrigado ! $\endgroup$ Apr 7, 2020 at 14:53
  • $\begingroup$ It would be very long to explain where this problem comes from. It arises in the context of RUM (Random Utility Maximization) model for choice modeling. $\endgroup$ Apr 7, 2020 at 19:03
  • $\begingroup$ Are you acquainted with the spectrahedron? $\endgroup$ Apr 7, 2020 at 20:57
  • $\begingroup$ Sorry for the delay in my answer, I have read the definiton of spectrahedron, but how can this help me? $\endgroup$ Apr 17, 2020 at 9:42

1 Answer 1

0
$\begingroup$

Here is a sufficient condition. The function $f(\lambda) = 1-a_i \lambda + \lambda^T A \lambda >0$ restricted on the simplex can be written, with a change of coordinates $\lambda\rightarrow x$ given by $\lambda_i=x_i^2$, as: $$ f(\lambda) = f(x) = \left(\sum_k x_i^2\right)^2 - \left(\sum_k a_{ik}x^2_k\right)\left(\sum_i x_i^2\right) + \sum_{i,j}A_{i,j}x_i^2x_j^2. $$ Now, if $f(\lambda)\geq 0$ on the simplex, then this implies that $f(x)$ must be positive semidefinite on the unit sphere. Now, a homogeneous quartic PSD on the unit sphere is globally PSD. A sufficient condition for a quartic to be globally PSD is to choose a sum of squares representation, which must have a form $zPz^\top$ with $z=[x_1^2~\cdots~x_n^2]$ and $P\succeq 0$. This translates to a linear matrix inequality on elements of $A$, which is the final sufficient condition.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.