4

The inverse of a SPD matrix is also symmetric. But what about the square root? Intuitively, I would say yes. But I'm not sure about it.

2 Answers2

7

Each PSD matrix certainly has a symmetric square root. In fact, PSD matrices have a unique PSD square root. However, not every square root of a PSD matrix is symmetric.

For example, the $2 \times 2$ zero matrix is PSD, but $$ A = \pmatrix{0&1\\0&0} $$ Is not symmetric, even though $A^2$ is PSD.

EDIT: If, like Pavel (below) you need some positivity in your PSD matrix, you should consider $$ A = \pmatrix{1&0&0\\0&0&1\\0&0&0} $$

Ben Grossmann
  • 225,327
3

There exist both positive-definite and non-positive-definite square roots for a Symmetric Positive-Definite matrix.

If A is SPD, then A can be represented as UDU* for some unitary matrix U and diagonal matrix D, with the eigenvalues represented in D, and all eigenvalues positive.

If I let R = sqrt(D) then I claim that URU* is a square root of A. Notice:

(URU*)^2 = URU*URU* = URRU = UDU = A

To find a particular square root, I only need to worry about finding square roots of the eigenvalues of D. There will be both positive and negative values for the square roots and sometimes a mix of both.

However, a matrix is positive-definite only when all the eigenvalues are positive. You can indeed find a square root with all positive values, but you can also find some with both positive and negative eigenvalues.