1: \begin{definition}[Relative robustness] 

2: Let $T$ (given by any representation), $\lambda_i$ and $\mathcal{I}$ be as

3: in the previous definition. Furthermore, let $\mathcal{Z}_{\mathcal{I}}$ be

4: the invariant subspace associated with $\mathcal{I}$. We say 

5:   that the representation of $T$ is relatively robust for $\mathcal{I}$

6:   if for all element-wise relative perturbation in the data bounded by

7:   $\xi \ll 1$ and for all $i \in \mathcal{I}$, we have 

8:   \begin{align*}

9:     |\tilde{\lambda}_i - \lambda_i| &\leq k_{rr} n \xi |\lambda_i| \,,  \\

10:     \sin \angle (\tilde{\mathcal{Z}}_{\mathcal{I}},\mathcal{Z}_{\mathcal{I}}) &\leq \frac{k_{rr} n \xi}{\mbox{\em relgap}(\mathcal{I})} \,,

11:   \end{align*}

12: where $\tilde{\lambda}_i$ and $\tilde{\mathcal{Z}}_{\mathcal{I}}$ denote the

13: eigenvalues and the corresponding invariant subspaces of the perturbed matrices,

14: respectively; $\angle(\tilde{\mathcal{Z}}_{\mathcal{I}},\mathcal{Z}_{\mathcal{I}})$

15: denotes the largest principle angle; 

16: and $k_{rr}$ is moderate constant, say about 10.\footnote{According to

17:   \cite{Willems:Diss,Willems:framework}, the requirement on the eigenvalues

18:   can be removed entirely: for $\mathcal{I}=\{i\}$, by

19:   Theorem~\ref{thm:gapthm} stated below, the second condition implies 

20:   the first up to a small constant provided $gap(\tilde{\lambda}_i) \approx

21:   gap(\{i\})$.  Similarly, if $\mathcal{I} = \{p,\ldots,q\}$ is not a

22:   singleton, the second term implies that $\tilde{\lambda}_i \in [\lambda_p

23:   - k_{rr} n \xi |\lambda_p|, \lambda_q - k_{rr} n \xi |\lambda_q|]$ for all $i \in \mathcal{I}$. 

24: }

25: \label{def:RRR}

26: \end{definition}

27: