Suppose we have the initial-value problem on $(0,L)$:

$$ \frac{d u(x)}{d x} = f(x) u(x),\, \qquad x\in\Omega,\,~~ u(0) = u_0, $$

I am reading a claim that says if we multiply the ODE by $u$ and integrate over $(0,L)$, we have

$$ \frac{1}{2}u^2(L) - \frac{1}{2} u^2_0 = \int_{0}^{L} f(x)u^2(x) \,dx $$

"from which the $L^\infty$-stability of the solution follows." I agree that the equation is correct, but:

  1. Why does this guarantee stability?
  2. What exactly is meant by $L^\infty$-stability? I interpret stability in the context that the numerical solution will remain bounded as the the step size is reduced, but here, we do not have a discretization yet...

This discussion is given in the context of discontinuous Galerkin methods.

  • $\begingroup$ Can you provide the reference? $\endgroup$ – nicoguaro Jun 8 '19 at 16:33
  • 1
    $\begingroup$ The $L^\infty$ norm of a function is its maximum absolute pointwise value, as in $\|f\|_\infty = \max_{x\in\Omega} |f(x)|$. Stability here means that if $\|f\|_\infty < \infty$, then $\|u\|_\infty < \infty$. The reasoning is a bit circular to me, because you can't just multiply pointwise by $u$ if $\|u\|_\infty = \infty$, but I can't immediately think of a counterexample. $\endgroup$ – Richard Zhang Jun 8 '19 at 16:49

Expanding on my comment. The $L^\infty$ norm of a function $f:\Omega\to\mathbb{R}$ is the maximum absolute pointwise value over its range, as in $$ \|f\|_\infty = \max_{x\in\Omega} |f(x)|.$$ Using this norm, we can define the function space $L^\infty$ for all functions with bounded $L^\infty$ norm, as in $$ f \in L^\infty \iff \|f\|_\infty < \infty.$$

Now, the ODE is said to be "stable in $L^\infty$" if every right-hand side $f\in L^\infty$ yields a solution $u\in L^\infty$. Loosely, the ODE is stable if a bounded $f$ yields a bounded $u$. Bounded input gives bounded output.

The claim in the prompt is definitely correct, but I'm not totally sold on the proof. With a minor change of notation, we get the one-dimensional linear time-varying (LTV) equation $$ \frac{\mathrm{d}x(t)}{\mathrm{d}t} = f(t)x(t), \qquad t\in(0,L),$$ and this has a well-known closed-form solution $$ x(t) = x_0\exp \left(\int_0^t f(t') \mathrm{d} t' \right), \qquad t \in (0,L).$$ (This is a special case of the Magnus series, but as an exercise, take the ansatz $x(t)=C \exp(g(t))$ and observe that $\mathrm{d}g/\mathrm{d}t = f$.) We see the claim is true because every $f\in L^\infty$ is integrable, and that the exponential map is analytic.

  • $\begingroup$ Interesting. So the algebra given is not enough to actually make the claim? The author's using additional information about this particular ODE? I had thought the idea was something to do with the derivative of energy $\frac{d}{dt}||u^2_h|| \leq 0$, but it was unclear to me why the RHS had to be less than or equal to zero, since no conditions are given on the forcing function $f(x)$. $\endgroup$ – user3482876 Jun 8 '19 at 18:11
  • $\begingroup$ So by asserting $\int_0^y u (du/dx) dx = (1/2)u(y)^2 - (1/2)u(0)^2$, you're already implicitly assuming that $u (du/dx)$ is integrable over $[0,y]$, i.e. it already excludes $u(x) = 1/x$ from consideration. It's easier to be stable when the possible choices of $u$ have already been significantly narrowed down. $\endgroup$ – Richard Zhang Jun 8 '19 at 18:26

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.