Assume we know that square matrices $A$ and $(I-AD)^{-1}$ are invertible and also $D$ is a diagonal matrix. Also assume that $A$ is a symmetric matrix. My question is when we can express $(I-AD)^{-1}$ as a function of $A, D, A^{-1}, D^{-1}, (I-A)^{-1}$ (without terms $(A-D)^{-1}$ or $(A^{-1}-D)^{-1}$) ?

For example when matrix A is rank 1, then we have:

$(I-AD)^{-1}=I+\frac{1}{1-tr(AD)} AD$. As you can see if A is rank 1 then we can do this easily.

The only related paper I found is a paper by Kenneth S. Miller, but it is not useful for higher rank matrices.

I know it might be very hard for general matrix $A$ but can it be done for special cases where for instance matrix A is positive semidefinite? Any comment is highly appreciated.

  • $\begingroup$ en.wikipedia.org/wiki/Binomial_inverse_theorem ? $\endgroup$ – Inquest Feb 13 '13 at 23:34
  • $\begingroup$ Thanks for pointing this out, but I believe by using this theorem still term $(I-AD)^{-1}$ will show up. $\endgroup$ – user54626 Feb 14 '13 at 0:23
  • $\begingroup$ It's not clear to me what you mean by "a function of $A,D,A^{-1},...$ etc, because $(I - AD)^{-1}$ to me is already written as a function of those things. $\endgroup$ – Christopher A. Wong Feb 14 '13 at 0:36
  • $\begingroup$ For example, we can rewrite this as $(A^{-1} - D)^{-1}A^{-1}$, but that doesn't exactly simplify anything in particular. $\endgroup$ – Christopher A. Wong Feb 14 '13 at 0:37
  • 3
    $\begingroup$ I'm not sure what your context is, since you are mentioning $(I-A)^{-1}$ while the invertibility of this matrix does not follow from your assumption. Would you consider something along the line of $(I - AD)^{-1} = I + (AD) + (AD)^2 + \cdots$ whenever $||AD|| < 1$ as a satisfactory answer? $\endgroup$ – user27126 Feb 16 '13 at 5:40

I have an answer to the following question : Is there a four-variable (not necessarily commutative) polynomial $f$ such that the identity

$$ (I-AD)^{-1}=f((I-A)^{-1},A,D,A^{-1},D^{-1}) \tag{1} $$

holds, whenever $A$ is symmetric positive definite, and $D$ is invertible and diagonal ?

The answer is NO. Indeed, this is already impossible when $n=2$ and $$D=\left(\begin{matrix} 2 & 0 \\ 0 & 3 \end{matrix}\right).$$ If we write

$$ A=\left(\begin{matrix} a & b \\ b & a \end{matrix}\right) $$

with $a\gt 0$ and $a\gt b$, then

$$ \det(A)=a^2-b^2, \det(I-A)=a^2-2a-b^2+1, \det(I-AD)=6a^2-5a-6b^2+1 $$

Thus the RHS in (1) will always have a denominator of the form

$$ (a^2-b^2)^p (a^2-2a-b^2+1)^q, $$

where $p$ and $q$ are integers. Now the LHS in (1) will always a denominator of the form $(6a^2-5a-6b^2+1)^r$. Since we have three distinct irreducible polynomials in ${\mathbb Q}[a,b]$ here, the denominators will never coincide.

  • $\begingroup$ Very nice, I'm curious if it's possible to see the infinite series solution 1/(1-x) = 1 + x + ... Somehow in determinant polynomial space? $\endgroup$ – kbb Feb 19 '13 at 22:29

In his book Some Eclectic Matrix Theory, Kenneth S. Miller presents the following generalization of your rank-$1$ result. This result is handy if your diagonal matrix $D$ is low rank.

Theorem (Miller, p. 14): Let $H$ be an $n\times n$ real matrix or rank $r$, and let $\sigma_1, \ldots, \sigma_n$ be the elementary symmetric functions of its eigenvalues. If $I+H$ is nonsingular then $a_r \ne 0$ and $$(I+H)^{-1} = I-\frac{1}{a_r}\big(a_{r-1}H - a_{r-2}H^2 + \cdots + (-1)^{r-1}a_0 H^r\big)$$ where $$a_m = 1 + \sum_{1 \le k \le m} \sigma_k, \quad m = 0,1,\ldots,r.$$

As for the elementary symmetric functions of the eigenvalues: They can be computed (via the Newton–Girard identities) from the power sums of the eigenvalues, which themselves are simply the traces of the powers of $H$ (as can be seen from writing $H$ in Jordan form): $$\operatorname{tr}(H^m) = \sum_{i=1}^n \lambda_i^m, \quad m \ge 0.$$ The upshot is that $(I+H)^{-1}$ is a degree-$\operatorname{rk}(H)$ polynomial in $H$ whose coefficients are rational functions in the traces of the powers of $H$ (Miller, Some Eclectic Matrix Theory, pp. 14-15).


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.