10
$\begingroup$

Assume $A$ is a square invertible matrix and we have $A^{-1}$. If we know that $I+A$ is also invertible, do we have a close form for $(I+A)^{-1}$ in terms of $A^{-1}$ and $A$?

Does it make it any easier if we know that sum of all rows are equal?

$\endgroup$
  • 2
    $\begingroup$ By substituting $A'=A-I$, you are basically asking for the inverse of all matrices. $\endgroup$ – akkkk Feb 9 '13 at 9:29
  • 1
    $\begingroup$ And of course, it might well be that $A$ is invertible, while $I + A$ is not. $\endgroup$ – Andreas Caranti Feb 9 '13 at 9:33
  • $\begingroup$ equal row sum won't help either. Consider the 2x2 matrix obtained from I by interchanging its rows. $\endgroup$ – Ittay Weiss Feb 9 '13 at 9:37
  • $\begingroup$ What does "convertible" mean? $\endgroup$ – Chris Eagle Feb 9 '13 at 9:39
  • 3
    $\begingroup$ The identity $(A^{-1}+I)^{-1} = A - A \, (I+A)^{-1}A$ (consequence of Woodbury matrix identity) does not answer you question, but perhaps you find it useful. $\endgroup$ – leonbloy Feb 9 '13 at 14:12
6
$\begingroup$

If the series $S = I - A + A^2 - A^3 + \dotsb$ converges, expand $S(I + A)$ to find that it is equal to the identity. See here for details on when $S$ converges.

$S$ clearly converges if $A^k = 0$ for some positive integer $k$ (nilpotency).

As for invertibility, if $A$ is diagonalizable, i.e. $A = PDP^{-1}$ for some diagonal matrix $D = \operatorname{diag}(e_1, e_2, \dotsc, e_n)$, then by Sylvester's Determinant Theorem, $$ \det(I + (PD)P^{-1}) = \det(P^{-1}(PD) + I) = \prod_{i = 1}^n(e_i + 1) $$

Hence, $I + A$ is invertible if no eigenvalue $e_i$ has a value of $-1$.

If $A$ is nilpotent, then its only eigenvalue is $0$, so $I + A$ is invertible.

Is there a closed-form solution?

I believe not. Basically, a closed-form expression of $(I + A)^{-1}$ using $A$ and $A^{-1}$ would amount to a closed-form expression of $(1 + x)^{-1}$ using $x$ and $x^{-1}$, where $x$ is real (or complex). A semi-rigorous articulation of this argument follows:

Proposition: There exists no family of matrices $\{X_{ij}\}_{m \times n}$, where every $X_{ij}$ is either equal to $A$, $A^{-1}$ or a constant dependent on the dimension of $A$, such that $(I + A)^{-1} = \sum_{i = 1}^m(\prod_{j = 1}^nX_{ij})$ for all values of $A$.

Proof:

Assume there exists such a family. Let $A$ be the $1 \times 1$ matrix $x$. Note that $\sum_{i = 1}^m(\prod_{j = 1}^nX_{ij})$ is a polynomial $P(x)$, which apparently equals $(1 + x)^{-1}$ for all values of $x$. Hence, $P(x)$ must be the taylor series $1 - x + x^2 - x^3 + \dotsb$, which contradicts the fact that $P(x)$ has a finite number of terms.

$\endgroup$
  • $\begingroup$ Herng, thanks for your comment. if we know that (I+A) is invertible and also we have $A^{-1}$. Then can we write a close form solution for $(I+A)^{-1}$? $\endgroup$ – user54626 Feb 9 '13 at 10:10
  • 1
    $\begingroup$ I doubt so. How can you express $\displaystyle\frac1{1+x}$ using only $\displaystyle\frac1x$ and $x$..? $\endgroup$ – Berci Feb 9 '13 at 14:18
  • $\begingroup$ haha I had the same idea as @Berci and edited it into the answer. $\endgroup$ – Herng Yi Feb 9 '13 at 14:30
4
$\begingroup$

I know I'am a bit late, but I wanted to add the following for future readers. If you know the Eigendecomposition of your matrix $A = Q \Lambda Q^{-1}$ the result is the following: \begin{equation} \left( A + I\right)^{-1} =\left( Q \Lambda Q^{-1} + I\right)^{-1} =\left( Q \left(\Lambda + I \right) Q^{-1}\right)^{-1} = Q \left(\Lambda + I \right)^{-1} Q^{-1} \end{equation} I know that this is a very trivial calculation and the result is maybe not helpful to many of you, but I wasted countless hours using the Shermann-Morrison-Theorem because I forgot about the Eigendecomposition. Cheers, Lukas

$\endgroup$
  • $\begingroup$ (+1) This answer was exactly what I was looking for. I have an iterative method with $A_i = a_iA_0$. So now inverting the matrix at step $i$ just involves inverting a diagonal matrix! $\endgroup$ – knrumsey Sep 20 '18 at 17:12
1
$\begingroup$

Check this question. The first answer presents a recursive formula to retrieve the inverse of a generic sum of matrices. So yours should be a special case.

$\endgroup$
  • $\begingroup$ In particular to answer the question on how to express $\frac{1}{1+x}$ using only $1,x$ and $1/x$ we have $$ \frac{1}{1+x}=\frac{1}{x}-\frac{1}{1+trace(1/x)} \frac{1}{x^2} $$ $\endgroup$ – Ferpect Feb 9 '13 at 23:58
  • $\begingroup$ Thanks for your answer. I have checked that question, but the first answer suggest recursive formula when rank of one of the matrices is 1. But rank of I is not 1. $\endgroup$ – user54626 Feb 10 '13 at 10:27
  • $\begingroup$ And after that Lemma, the autor states the theorem for rank>1 $\endgroup$ – Ferpect Feb 10 '13 at 15:08

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.