In the proof of the theorem which states that the Lipschitz condition guarantees well posed-ness of an initial value problem $y'=f(x,y)$, $y(x_0)=y_0$, I came across this

Let the perturbed problem be

$$ z'=f(x,z) + \delta(t) \\ z(x_0)=y_0 + \epsilon_0 $$

Let $\epsilon(x)=z(x)-y(x)$ and let $f$ satisfy the Lipschitz condition. Then we have

$$ \epsilon '=z'-y'=f(x,z)-f(x,y)+\delta(x)\\ \implies |\epsilon'(x)| \leq |f(x,z)-f(x,y)| + |\delta(x)|\\ $$

Let $\delta(x)$ have the maximum value $\epsilon$. Therefore, we get,

$$ |\epsilon'(x)| \leq |f(x,z)-f(x,y)| + \epsilon\\ \implies |\epsilon'(x)| \leq L|\epsilon(x)| + \epsilon $$

I don't understand the next step i.e. how is the above inequality integrated to give

$$ |\epsilon(x)| \leq \frac{\epsilon}{L} \left[ (L+1)e^{Lt}-1\right] $$

How is the inequality integrated?

  • $\begingroup$ Is this in a homework? Or did you come up with it yourself? $\endgroup$
    – Ellya
    Apr 2 '14 at 17:45
  • $\begingroup$ @ellya this is not homework...came across this in a textbook... Gear(1971)...i'll update the exact title later $\endgroup$
    – yaska
    Apr 2 '14 at 18:43

It is more convenient to use the integral form: $$ y(x)=y_0+\int_0^x f\big(s,y(s)\big)\,ds $$ and $$ z(x)=y_0+\varepsilon_0+\int_0^x \big(\,f\big(s,z(s)\big)+\delta(s)\big)\,ds $$ and hence $$ z(x)-y(x)=\varepsilon_0+\int_0^x\delta(s)\,ds+\int_0^x\big(f\big(s,z(s)\big)-f\big(s,y(s)\big)\big)\,ds, $$ which implies that (for $x>0$) $$ \lvert z(x)-y(x)\rvert \le \lvert \varepsilon_0\rvert +\varepsilon x+ L\int_0^x\big|\,z(s)-y(s)\big|\,ds. \tag{1} $$ Now, let $w(x)=\int_0^x \big|\,z(s)-y(s)\big|\,ds$, so the above becomes $$ w'(x) \le \lvert \varepsilon_0\rvert +\varepsilon x+ L w(x) $$ or $$ \left(\mathrm{e}^{-Lx}w(x)\right)'=\mathrm{e}^{-Lx}\big(w'(x)-Lw(x)\big)\le \mathrm{e}^{-Lx}\big(\lvert\varepsilon_0\rvert +\varepsilon x\big). $$ Integrating the above in $[0,x]$ we obtain $$ \mathrm{e}^{-Lx}w(x)-w(0)=\frac{\lvert\varepsilon_0\rvert}{L}(1-\mathrm{e}^{-Lx}) +\varepsilon\mathrm{e}^{-Lx}\left(-\frac{x}{L}+\frac{1}{L^2}\right)-\frac{\varepsilon}{L^2} $$ or $$ \int_0^x\lvert z(s)-y(s)\rvert\,ds\le \frac{\lvert\varepsilon_0\rvert}{L}(\mathrm{e}^{Lx}-1) +\varepsilon\left(-\frac{x}{L}+\frac{1}{L^2}\right)-\frac{\varepsilon \mathrm{e}^{Lx}}{L^2}. $$ Using $(1)$ we obtain $$ \lvert z(x)-y(x)\rvert \le \lvert\varepsilon_0\rvert(\mathrm{e}^{Lx}-1) +\varepsilon\left(-x+\frac{1}{L}\right)-\frac{\varepsilon \mathrm{e}^{Lx}}{L}+\lvert\varepsilon_0\rvert +\varepsilon x =\cdots. $$


Multiply through by $|\epsilon (x )|$ to get:

$|\epsilon'(x )\epsilon (x )|\leq L|\epsilon (x )|^2+\epsilon|\epsilon (x )|$

$\frac{1}{2}|(\epsilon^2(x))'|\leq L |\epsilon^2(x)|+\epsilon|\epsilon (x)|\leq L|\epsilon(x)^2|+\epsilon^2$

$\Rightarrow \frac {1}{2}\frac {d}{dx}(|\epsilon ^ 2( x ) | e ^ { -2Lx })\leq \epsilon^2e^{-2Lx}\Rightarrow\frac {1}{2}|\epsilon ^ 2( x ) | e ^ { -2Lx}\leq \frac{1}{2}|\epsilon ^ 2 ( 0) |+\frac{1}{2L}(\epsilon ^ 2)(1-e^{-2Lx}) $

$\Rightarrow |\epsilon ^ 2( x ) |\leq|\epsilon ^ 2( 0) |e^{2Lx}+\frac{1}{L}(\epsilon^2)(e^{2Lx}-1)\leq\epsilon^2e^{2Lx}+\frac{1}{L}(\epsilon^2)(e^{2Lx}-1)=\frac{\epsilon^2}{L^2}(L^2+Le^{2Lx}-L)$

still working on this, but it's quite close to the desired estimate.


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.