2
$\begingroup$

Let $A$ be a $k \times k$ invertible matrix over complex numbers.

If it possible to write its nth root as an analytic function (i.e. power series in $A$)?

EDIT: Complex coefficients can be functions of $A$.

Notes

If a matrix $A$ has only one eigenvalue $\lambda$, then it is simple. We take

$$B = \exp\left[\tfrac{1}{n} \log (A ) \right]$$

where we have $B^n = A$. Using Jordan decomposition, we can simplify the logarithm to a polynomial in that matrix (as $(A - \lambda \mathbb{I})$ is nilpotent)

$$\log(A) = \log(\lambda) - \sum_{i=1}^{k} \frac{\left(- \tfrac{A}{\lambda} + \mathbb{I}_k \right)^{i}}{i}.$$

$\endgroup$

4 Answers 4

3
$\begingroup$

If I am reading this correctly, you are fine with a power series whose (scalar) coefficients depend on the matrix $A$. In this case, it suffices to take a polynomial $p$ that interpolates $\sqrt[n]{x}$, such that for each eigenvalue $\lambda$ with multiplicity $k_\lambda$, the first $k_\lambda-1$ derivatives of $p$ coincide with those of $\sqrt[n]{x}$ (Hermite interpolant). A degree-$k$ polynomial will always do the job.

$\endgroup$
5
  • $\begingroup$ Yes, I'm fine with coefficients depending on the matrix. I don't know why I overlooked this solution. $\endgroup$ Feb 24, 2013 at 20:19
  • $\begingroup$ However, it is not as simple - I cannot assume that the matrix is diagonalizable; so any function which just maps eigenvalues to their roots won't work. Take as a counterexample $A = [[1, 1], [0, 1]]$ and $f(z)=z$ (sure, another polynomial works for this $A$). $\endgroup$ Feb 24, 2013 at 20:30
  • 2
    $\begingroup$ The multiplicity of the eigenvalue $1$ is $2$, so you need a polynomial that matches $g(z)$ and $g'(z)$ in $z=1$, where $g(x)=\sqrt[n]{x}$. For some more detail on this approach, you can check Higham's Functions of matrices, SIAM Press 2008. $\endgroup$ Feb 24, 2013 at 21:08
  • $\begingroup$ Sorry - I meant to add Chapter 1, but I pressed "enter" too quickly. $\endgroup$ Feb 24, 2013 at 21:11
  • $\begingroup$ @Federico So, you have my thanks in arxiv.org/abs/1305.1506. $\endgroup$ May 8, 2013 at 8:54
2
$\begingroup$

Let $A$ be a $k\times k$ invertible matrix, i.e. in $Gl(k)$. Assume that the segment $[I,A]$ lies in $Gl(k)$. Let us define $$ \text{Log}A=\int_{[1,A]} \frac{d\xi}{\xi}=\int_0^1(I-tI+tA)^{-1}(A-I)dt. $$ It makes sense since $A$ commutes with the denominator inside the integral. The assumption is satisfied in particular whenever $A$ is symmetric invertible with a nonnegative real part. Analytic continuation arguments entail $$ \exp(\text{Log}A)=A\quad \bigl(\exp(\frac{1}{n}\text{Log}A)\bigr)^n=A. $$ Looking at the Jordan canonical form of $A$, it is not difficult to see that the only thing to be avoided for the above method to work is that eigenvalues should not be negative real numbers. Let $z=a+ib$ be an eigenvalue not in $\mathbb R_-$ in a Jordan block $J_N$ of size $N$, with 1 above the diagonal. Considering the segment $[I_N,J_N]$, we find on the diagonal $$ (1-t)+tz\notin \mathbb R_-\text{ since $z\notin \mathbb R_-$}, $$ and above the diagonal $ (1-t)0+t=t. $ The logarithm formula above works.

$\endgroup$
0
$\begingroup$

If all eigenvalues of $A$ are in the right half plane, there is $\alpha > 0$ such that the circle $|z - \alpha| < \alpha$ contains all the eigenvalues, and the principal branch of $f(z) = z^{1/n}$ is analytic in that circle. We then have a convergent binomial series $$ A^{1/n} = \alpha^{1/n} \sum_{k=0}^\infty {{1/n} \choose k} (\alpha^{-1} A - I)^k $$

$\endgroup$
1
  • 1
    $\begingroup$ It was my initial thought, but in my case I cannot make such an assumption. $\endgroup$ Feb 25, 2013 at 0:40
-3
$\begingroup$

It easy to apply a series to a diagonalizable matrix, and diagonalizable matrices are dense. This should answer your question.

$\endgroup$

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.