Question: Unstability of the eigenvectors of some general matrix


 Good morning all.
Consider $A=[a_{i,j}]$, a $18\times 18$ matrix (for example); the integer entries are randomly chosen in $[[-5,5]]$ (for example). In general, $A$ has distinct eigenvalues.

I seek the eigen-elements of $B=Transpose(A^{-1})A$ with $10$ significand digits (for example); in general, $B$ is diagonalizable ($P^{-1}BP=D$, a diagonal complex matrix). I use the command $evalf(Eigenvectors(?))$.

Many randomized tests require working with hundreds of Digits. The worst one requires $629$ digits!! Moreover (in this test), when $Digits:=400$, the condition number of $P$ is $10^{118}$ and , with $Digits:=619$, the condition number drops to $376$.

I am surprised by this instability. In particular, this method seems to be unusable when $n=100$.
Does there exist a method (using maple) which allows to solve the problem without dragging behind me a multitude of digits ? (perhaps with iterations...)
Thanks in advance.


Please Wait...