Assume $A$ is a square matrix defined as follow: $$A=\sum_{i} u_{i}u_{i}^T$$ where for each $i$, $u_i$ is a non-negative column vector.

Do the matrices of these forms have any special name?

  • $\begingroup$ So $A$ is $U\cdot U^T$ where $U$ is a nonnegative square matrix. $\endgroup$ – Louis Nov 15 '13 at 0:09
  • $\begingroup$ No it's not. Note that in the definition I have $u_{i}u_{i}^T$ not $u_{i}u_{j}^T$. $\endgroup$ – user54626 Nov 15 '13 at 0:12
  • $\begingroup$ In general, if $a_i$ are the column vectors of $A$ and $b_i$ are the column vectors of $B$, then $A\cdot B^T = \sum a_i \cdot b_i^T$ $\endgroup$ – Louis Nov 15 '13 at 0:17
  • 3
    $\begingroup$ If you drop the condition that $u$ is non-negative, the name is "symmetric positive-definite". The extra condition is certainly important. Maybe the answer is "symmetric positive-definite non-negative"? Do you have an example that disproves this guess? $\endgroup$ – apt1002 Nov 15 '13 at 0:22
  • $\begingroup$ Not entirely relevant, but a matrix of the form $u v^T$ is called a dyad. $\endgroup$ – copper.hat Nov 15 '13 at 0:31

Yes, these are the completely positive matrices.


Let's assumes there are $n$ such $u_i$'s and there are exactly $k$ linearly independent vectors among them, then matrix $A$ would be a rank-k matrix.
Each of the $u_iu_i^t$ is a rank-1 matrix and they would sum up to a rank-k matrix if exactly k out of them are linearly independent.
As $x^t(u_iu_i^t)x=(x^tu_i)(x^tu_i)^t \gt0$ , each of the $u_iu_i^t$ is a symmetric positive definite matrix, thus A is also a symmetric positive definite matrix.


There is a problem with the term 'non-negative column vector'. The notion of non-negative is defined for square matrices.

We can require $u_i$'s such that $M=[u_{ij}]$ is positive (non-negative) that is such that $M=UU^*$ for some square matrix $U$.

There is also a different notion of positivity (again for square matrices) which requires the entries to be non-negative.

One vaguely related idea is the following. If $M$ is a symmetric matrix that is orthogonally diagonalised by

$P=[u_1\space{} u_2\space{}\dots u_n]$

and if $\lambda_i$'s are eigenvectors of $M$ corresponding to the unit eigenvectors $u_i$'s, then

$M=\lambda_1 u_1u_1^T+\dots +\lambda_nu_nu_n^T$

is called the spectral decomposition of $M$.

  • $\begingroup$ I'm assuming the components of the column vectors are nonnegative. Perhaps the OP should clarify. $\endgroup$ – Stefan Smith Nov 15 '13 at 1:00
  • $\begingroup$ I don't know what do you mean by "$A$, as it is written, denotes a single positive number". I mentioned that each $u_i$ is a column vector so $A$ is a matrix. $\endgroup$ – user54626 Nov 15 '13 at 3:22
  • $\begingroup$ Fixed that. Will look it up. $\endgroup$ – AnyAD Nov 15 '13 at 4:20

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.