Given the following matrix $$A = \begin{bmatrix} 8 & 1 & 0\\ 1 & 4 & \epsilon\\ 0 & \epsilon & 1\\ \end{bmatrix}, |\epsilon| < 1.$$ Gerschgorin's theorem states that each of the $\lambda_i$ eigenvalues of A will be placed inside a circular disk with center $a_{i,i}$ and radius $\sum^n_{j = 1, j\neq i} |a_{i,j}|$. As a direct consequence, the lowest eigenvalue $\lambda_3$ of A is such that $|\lambda_3 -1| < \epsilon$.
However I'm now asked to prove the tighter bound $|\lambda_3-1| < \epsilon^2$ using diagonal similarity transformations. My initial idea was to use the first step of the QR algorithm for Hessemberg diagonalization as usual to obtain a similar matrix $A_1 = Q^*AQ$ (with $Q$ computed by Householder transformations) preserving the eigenvalues and hopefully the tighter bound appears using Gerschgorin's theorem once again. I've made some numerical tests using Octave for some particular $\epsilon$ and it seems to be indeed the case that $A_1$ satisfies the bound. The problem is that the exact computation of $Q$ and $A_1$ have been absolutely painful, specially when dealing with the unknown symbol $\epsilon$. I also tried to use a shift of $\rho = 1$ with no better results. So now I'm suspicious there is a simpler similarity transformation that could lead to the desired result. Any suggestions would be highly appreciated.