Skip to main content
4 of 7
link to Mathematica note book
Carlo Beenakker
  • 188.1k
  • 18
  • 448
  • 651

Uniquely reconstruct a matrix $M$ from its inverse $M^{-1}$ if $n$ elements of $M^{-1}$ are unknown and $n$ elements of $M$ are given

This question was motivated by a recent MO post. You know $n$ elements of the $N\times N$ matrix $M$ and you do not know $n$ elements of the inverse $M^{-1}$ (but you know the other $N^2-n$ elements of $M^{-1}$). Equating $(M^{-1})^{-1}=M$ gives $n$ nonlinear equations in $n$ unknowns, which in general will have multiple solutions. Under which additional condition can one reconstruct the matrix $M$ uniquely? Does it matter where in the matrix are the $n$ elements located?

Conjecture: For $n=N$ elements on the diagonal the reconstruction is unique if $M$ is positive definite. (For $N=2$ it is true,$^\ast$ and some experimentation$^{\ast\ast}$ for larger $N$ suggests it is true for all $N$.)


$^\ast$ For $N=2$ one has $M = \begin{pmatrix}a & b \\ b & c \end{pmatrix}$, $M^{-1} = \frac{1}{a c - b^2} \begin{pmatrix}c & -b \\ -b & a \end{pmatrix}$, we know $a,c$ and we know $\beta=b/(ac−b^2)$. There are two solutions for the unknown $b$, $b_\pm=(\pm\sqrt{4ac\beta^2+1}−1)/2\beta$, only $b_+$ gives a positive definite $M$.

$^{\ast\ast}$ Mathematica test for $N=3$, when there are 5 solutions for the three unknown matrix elements, but only one of these gives a positive definite $M$.

Carlo Beenakker
  • 188.1k
  • 18
  • 448
  • 651