3
$\begingroup$

I'm looking to prove the following lemma:

Let $A$ be a matrix in $\mathbb{R}^{n\times n}$. Then for any $\lambda^* > \max_{\lambda} \; \mathrm{Re} \; (\lambda)$ such that $ \lambda \in\sigma (A)$, there exists $\beta > 0$ such that \begin{eqnarray*} ||e^{At}|| \leq \beta e^{\lambda^* t}\qquad \forall t\geq 0 . \end{eqnarray*}

This is what I have

Proof: We know there exists a transformation $T$ such that $e^{At}=Te^{Jt}T^{-1}$, where \begin{eqnarray*} T^{-1}AT&=&\begin{bmatrix} J_1 & \: & 0 \\ \: & \ddots & \: \\ 0 & \: & J_k \end{bmatrix} ,\; J_i=\begin{bmatrix} \lambda_i & 1 & 0 & \cdots & 0 \\ 0 & \lambda_i & 1 & \ddots & \vdots \\ \vdots & \ddots & \ddots & \ddots & 0 \\ \vdots & \: & \ddots & \ddots & 1 \\ 0 & ... & 0 & \lambda_i \end{bmatrix}\text{ for $i=1,\dots,k$,} \\ e^{Jt}&=&\begin{bmatrix} e^{J_1 t} & \: & 0 \\ \: & \ddots & \: \\ 0 & \: & e^{J_k t} \end{bmatrix} \text{and,}\; e^{J_i t}=e^{\lambda_i t}\begin{bmatrix} 1 & t & \cdots & \frac{t^{q_i -1}}{(q_i -1)!} \\ 0 & 1 & \ddots & \vdots \\ \vdots & \ddots & \ddots & t \\ 0 & \cdots & 0 & 1 \end{bmatrix} \end{eqnarray*} Therefore \begin{eqnarray*} ||e^{At}||&=&||Te^{Jt}T^{-1}|| \\ &\leq &||T||\;||T^{-1}||\;||e^{Jt}|| \\ &=& \beta_1 ||e^{Jt}|| \end{eqnarray*} Let $\epsilon = \lambda^* - \mathrm{max} \; \lambda\in\sigma (A)$, then $\epsilon > 0$ and \begin{eqnarray*} ||e^{Jt}||&\leq & \left| \left| e^{(\lambda^* - \epsilon)t} \begin{bmatrix} 1 & t & \cdots & \frac{t^{n -1}}{(n -1)!} \\ 0 & 1 & \ddots & \vdots \\ \vdots & \ddots & \ddots & t \\ 0 & \cdots & 0 & 1 \end{bmatrix} \right| \right| \\ &\leq & e^{\lambda^* t} \left| \left| e^{ - \epsilon t} \begin{bmatrix} 1 & t & \cdots & \frac{t^{n -1}}{(n -1)!} \\ 0 & 1 & \ddots & \vdots \\ \vdots & \ddots & \ddots & t \\ 0 & \cdots & 0 & 1 \end{bmatrix} \right| \right| \end{eqnarray*} Then $b^*_t$, the largest term at time $t$ as \begin{eqnarray*} b^*_t = \max_i \left( e^{-\epsilon t} \frac{t^{i}}{(i-1)!} \right) \end{eqnarray*} Since $\epsilon>0$, the $\exists$ $b^*$ such that \begin{eqnarray*} b^* = \max_t \; b^*_t \end{eqnarray*} Then \begin{eqnarray*} ||e^{Jt}||&\leq & e^{\lambda^* t} \left| \left| \begin{bmatrix} b^* & \cdots & b^* \\ & \ddots & \vdots \\ 0 & & b^* \end{bmatrix} \right| \right| \\ &=& \beta_2 e^{\lambda^* t} \end{eqnarray*} Putting it all together we get that \begin{eqnarray*} ||e^{At}||\leq \beta e^{\lambda^* t} \qquad \forall t\geq 0 \end{eqnarray*}

My problem is that I'm not sure if I'm allowed to do my jump from \begin{eqnarray*} ||e^{Jt}||&\leq & \left| \left| e^{(\lambda^* - \epsilon)t} \begin{bmatrix} 1 & t & \cdots & \frac{t^{q_n -1}}{(q_n -1)!} \\ 0 & 1 & \ddots & \vdots \\ \vdots & \ddots & \ddots & t \\ 0 & \cdots & 0 & 1 \end{bmatrix} \right| \right| \end{eqnarray*}

Any advice? Alternatively if you could provide a reference to this lemma it would be greatly appreciated! :)

Thanks is advance

$\endgroup$
2
$\begingroup$

Your approach is reasonable, but can be simplified a bit.

You need $\lambda^* > \max_{\lambda \in \sigma(A)} \operatorname{re} \lambda $, remember that the spectrum have complex eigenvalues.

However, I think it would be easier to show that $\|e^{-\lambda^* t} e^{At} \|$ is bounded for $t \ge 0$.

In particular, this reduces to showing that each entry in $e^{-\lambda^* t} e^{At}$ (or rather, $e^{-\lambda^* t} e^{Jt}$) is bounded.

Since $\lim_{t \to \infty} e^{(\lambda_i-\lambda^*) t} \frac{t^{k}}{k!} = 0$ for all $k$, we see that every entry in $e^{-\lambda^* t} e^{At} $ goes to zero, hence it is bounded.

So, a slightly tighter result would be $\lim_{t \to \infty} e^{-\lambda^* t} e^{At} = 0$.

Addendum: Since $\lim_{t \to \infty} e^{-\lambda^* t} e^{At} = 0$ and $t \mapsto e^{-\lambda^* t} e^{At}$ is continuous, we see that $\beta = \sup_{t \ge 0} \| e^{-\lambda^* t} e^{At}\| < \infty$. hence we have $\| e^{-\lambda^* t} e^{At}\| \le \beta$ for all $t \ge 0$, from which it follows that $\| e^{At}\| \le \beta e^{\lambda^* t}$ for $t \ge 0$.

$\endgroup$
  • $\begingroup$ Yeah thanks for the comment on the complex eigenvalues. Unfortunately for my research I need the lemma as stated. I know I've seen people use this result before, just can't seem to find a reference for it. Do you see a flaw in my current proof? If so what is it? Thanks $\endgroup$ – Kroeze Oct 3 '13 at 16:23
  • $\begingroup$ I'm not sure what you mean by need the lemma as stated. If $\lim_{t \to \infty} e^{-\lambda^* t} e^{At} = 0$, then, in particular, $\|e^{-\lambda^* t} e^{At}\|$ is bounded, say by $\beta$. Then you have $\|e^{At}\| \le \beta e^{\lambda^* t} $, which is what you were looking for. Also, you can't just replace the $e^{\lambda_i t}$ term by $e^{(\lambda^*-\epsilon)t}$, because in general it is not true. It is easier to work termwise instead of on the matrix norm in this case. $\endgroup$ – copper.hat Oct 3 '13 at 16:47
  • $\begingroup$ The essence of the result is that $\lim_{t \to \infty} e^{\alpha t} \frac{t^{k}}{k!} = 0$, whenever $\alpha$ has a strictly negative real part. I'm not sure it is a 'standard lemma' as such, you could look for results on exponential stability and the like. It is a fairly basic result on control theory, for example. Maybe try "Nineteen Dubious Ways to Compute the Exponential of a Matrix" by Moler & Van Loan. $\endgroup$ – copper.hat Oct 3 '13 at 16:56
  • $\begingroup$ In my lemma it states that I need $||e^{At}||\leq \beta e^{\lambda^* t}\; \forall t\geq 0$, not in the limit. Also I'm not just 'replacing' $e^{\lambda^i t}$ by $e^{\lambda_{max}t}$, I'm saying the right hand side is greater than or equal to the left hand side. A worst case scenario if you want to think of it that way. Thanks for the reference by Moler and Van Loan, unfortunately they do not have such a result since they are looking for tight approximations, which mine is not. $\endgroup$ – Kroeze Oct 3 '13 at 17:33
  • $\begingroup$ Perhaps you could try using the fact that $\|e^{\alpha t} M \| = e^{(\operatorname{re} \alpha) t} \|M\|$. Also, you jump from $\|e^Jt\|$ to a norm on the individual Jordan blocks. This leaves another step that needs to be justified. It is much easier to work termwise. $\endgroup$ – copper.hat Oct 3 '13 at 17:46

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.