3
$\begingroup$

Simplified question*:

Given $f(t)$ that satisfies $f'(t)>0$, $f'(t)=\omega\left(t^{-1}\right)$, $\log\left(f'(t)\right)=o\left(f(t)\right)$ we denote $F=\exp\left(f\left(t\right)\right)$. Let $H(t)$ be a solution of $$ \dot{H}=F $$ Can we approximate H by F?

Specifically, I want to show that $$ \lim_{s\to\infty}\frac{H^{-1}(s)}{F^{-1}(s+C)}=1 $$ for some constant C. $F^{-1}$ and $H^{-1}$ are the inverse functions of $F$ and $H$, respectively.

If necessary I can add additional assumptions on $f(t)$, though I would like to keep it as general as possible.

* thanks to Willie Wong helpful suggestion.


The original question (to make the motivation more clear):

I have the following differential equation: $$\dot{g}(t)=\exp\left(-f\left(g\left(t\right)\right)\right)$$ and I know that $f(t)$ satisfies $f'(t)>0$, $f'(t)=\omega\left(t^{-1}\right)$, $\log\left(f'(t)\right)=o\left(f(t)\right)$.

I would like to show that the solution to this equation is $$g(t)=f^{-1}(\log(t+C))+h(t)$$ where $h(t)=o\left( f^{-1}(\log(t)) \right)$, $f^{-1}(t)$ is the inverse function of $f(t)$ and C is some constant.

The motivation behind this solution is $$ \exp\left(f\left(g(t)\right)\right)\approx\exp\left(f\left(g\left(t\right)\right)+\log\left[ f'\left(g\left(t\right)\right)\right]\right) $$ since $\log\left(f'(t)\right)=o\left(f(t)\right)$, and therefore we can solve $$ \exp\left(f\left(g\left(t\right)\right)+\log\left[ f'\left(g(t)\right)\right]\right)\dot{g}(t)=1 \Rightarrow \frac{d}{dt}\exp\left(f(g\left(t\right)\right)=1 \Rightarrow f(g(t))=\log(t+C)$$ for some constatnt C.

Can I rigorously show this result? If necessary I can add additional assumptions on $f(t)$, though I would like to keep it as general as possible.

$\endgroup$
  • $\begingroup$ What is your $\omega(t^{-1})$? $\endgroup$ – user64494 Aug 28 '18 at 11:56
  • $\begingroup$ $f(t)=\omega(g(t))\Rightarrow$f dominates g asymptotically en.wikipedia.org/wiki/Big_O_notation $\endgroup$ – Mor Aug 28 '18 at 12:52
  • 2
    $\begingroup$ May be you can simply your question a little bit: I don't think $g$ plays an important role here. The main question seems to be: given $f(t)$ with the specified properties, denote by $F = e^f$ and let $H$ be a solution of $\dot{H} = F$. Can we approximate $H$ by $F$. Or more precisely, can we approximate $H^{-1}(s)$ by $F^{-1}(s + C)$ as $s \to \infty$. The mode of approximation you want seems to be requiring $\lim_{s\to \infty} H^{-1}(s) / F^{-1}(s+C) = 1$. $\endgroup$ – Willie Wong Aug 30 '18 at 16:41
  • $\begingroup$ That is a really good suggestion. Thank you. I updated the question. $\endgroup$ – Mor Aug 30 '18 at 18:48
2
$\begingroup$

The simplest cases are the linear cases where $f(t) = \alpha t$, and $H = \alpha^{-1}F$. Modelling on those cases, I will show that if asymptotically $f'(t)$ is bounded below (so if $f$ is asymptotically superlinear) than the desired conclusion hold.


The first step is to prove that the conclusion holds if $H$ is eventually $< F$, using merely a convexity argument.

Suppose there exists some $t_0$ such that for every $t > t_0$ we have that $H(t) < F(t)$. Notice that $\dot{H} = F$ and $\ddot{H} = \dot{f} F > 0$ we have that $H$ is convex. We are interested in the quantity

$$ H^{-1}(s) - F^{-1}(s) $$

which is positive for all sufficiently large $s$ when $H$ is eventually $< F$.

By convexity, we have that for any $r < s$

$$ H^{-1}(s) < H^{-1}(r) + (H^{-1})'(r) \cdot (s-r) $$

Since $H < F$ we can shoose $r = H\circ F^{-1}(s)$ and get

$$ H^{-1}(s) < F^{-1}(s) + (H^{-1})'\circ H \circ F^{-1}(s) \cdot (s - H\circ F^{-1}(s)) $$

Now

$$ (H^{-1})' = \frac{1}{H'\circ H^{-1}} $$

so we get

$$ H^{-1}(s) < F^{-1}(s) + \frac{1}{s} \cdot (s - H\circ F^{-1}(s)) $$

So in particular, you have that $H^{-1}(s) - F^{-1}(s)$ is bounded, since $H\circ F^{-1}$ is increasing.

Using that $F^{-1}$ grows unboundedly this implies

$$ H^{-1} - F^{-1} = o(F^{-1})$$

as desired.


Notice that in this argument I have not explicitly used the conditions $f' = \omega(t^{-1})$ and $\log(f') = o(f)$; however, the condition that $H(t) < F(t)$ can be derived from strengthened versions of your assumption.

For example, a sufficient condition to guarantee that $H(t) < F(t)$ eventually for every solution $H$ is that $f'(t) > 1 + \epsilon$ for all sufficiently large $t$.


This can be generalized further. Suppose that $f'(t)$ is lower-bounded by $\delta > 0$.

Define the function $\tilde{f}(\tau) = f(2 \delta^{-1} \tau)$. Define $\tilde{F}$ and $\tilde{H}$ analogously. We have that $\tilde{F} = F(2 \delta^{-1}\tau)$ and $\tilde{H} = \frac{\delta}{2} H(2\delta^{-1}\tau)$

Applying the above argument we get

$$ H^{-1}(2 \delta^{-1} s) - F^{-1}(s) < \frac{2}{\delta} - \frac{1}{s} H\circ F^{-1}(s) $$

Now

$$ F^{-1}(s) = f^{-1} \ln(s) $$

so

$$ F^{-1}(2 \delta^{-1} s) = f^{-1} ( \ln(s) + \ln 2 - \ln \delta) $$

Using that $f'$ is bounded below by $\delta$, we have that

$$ |F^{-1}(2 \delta^{-1} s) - F^{-1}(s) | < \delta^{-1} \ln(2 \delta^{-1}) $$

is also bounded, so that we conclude

$$ H^{-1}(s) - F^{-1}(s)$$

is bounded, and the unbounded growth of $F^{-1}$ takes care of the rest.


The remaining case is when $\liminf_{t\to\infty} f'(t) = 0$, which is probably the case you are really interested in to boot, but unfortunately I don't yet see a proof for.

$\endgroup$
  • $\begingroup$ Thank you! Nice proof. One question - why do you need $r<s$? $\endgroup$ – Mor Sep 1 '18 at 11:06
  • $\begingroup$ @Mor: just my personal laziness. (I didn't want to have to type $\leq$ everywhere, and the function involved in strictly convex, so I can get away without $\leq$ if I rule out $r = s$. So instead of $r \neq s$ I typed $r < s$, which happens to be true for the case we need.) $\endgroup$ – Willie Wong Sep 4 '18 at 14:50
  • $\begingroup$ @WillieWong: Sorry if I'm missing something obvious: where and why do we need to use $r<s$ exactly? Wouldn't the next lines (e.g., the first equation stated after the condition) also hold if $r>s$? $\endgroup$ – Daniel Soudry Sep 8 '18 at 12:39
  • $\begingroup$ @DanielSoudry: see my previous comment? (I don't need $r < s$; it is that for the construction only the $r < s$ case is needed.) $\endgroup$ – Willie Wong Sep 10 '18 at 1:08
  • $\begingroup$ Ah, yes. Sorry, I misunderstood your previous comment (I thought it explained why you used $r < s$ instead of $r \leq s$). $\endgroup$ – Daniel Soudry Sep 10 '18 at 6:54

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.