I have the heat equation $$u' - \Delta u = f$$ as equality in $L^2(0,T;V')$,i.e., $$(u',v) + (\nabla u, \nabla v) = (f,v)$$ for all $v \in L^2(0,T;V)$, where I used the same brackets for duality pairing and inner product for succinctness.

Let $w_j$ be the basis in $V$ and $H$. I read in a book that the finite dimensional (Galerkin) approximations to the PDE $$(u_n',w_j) + (\nabla u_n, \nabla w_j) = (f, w_j), \quad\text{for $j=1,...,n$}$$ can be written as $$\frac{d}{dt}u_n - \Delta u_n = P_nf\tag{1}$$ where $P_n$ is a projection operator.

In what sense is this latter equation an equality? Presumable not an equality in $L^2(0,T;V')$ because the weak form I wrote above only holds for the basis functions. However, the author later takes an inner product of (1) with an element on $L^2(0,T;V)$, how can he do that?


Well, it is indeed an abuse of notation though, by no means you could say (1) is an equality in $L^2(0,T;V')$, because (1) only holds in a finite dimensional subspace. Multiply (1) by the basis of the finite dimensional subspace $$ (u_n',w_j) - (\Delta u_n,w_j) = (P_n f,w_j). $$ Proper assumption about the boundary condition and continuity leads to: $$ (u_n',w_j) - (\Delta u_n,w_j) = (u_n',w_j) + (\nabla u_n, \nabla w_j). $$ Hence: $$ (P_n f,w_j) = (f,w_j), \quad \text{for } j=1,\ldots,n, $$ which is to say, $P_n f$ is the $L^2$-projection of $f$ onto this finite dimensional subspace.

I didn't see the paper, but I guess the reason to multiply a function in $L^2(0,T;V)$ is to get some a priori error estimate for this approximation by exploiting the fact that $$ (u'-u_n',w_j) + \big(\nabla (u-u_n), \nabla w_j\big) = (f-P_n f,w_j). $$

EDIT: question in the comment, instead of claiming (1) holds, rather what we can do is the following $$ \begin{aligned} &(u_n',\phi) + a(u_n,\phi) \\ =& (u_n',P_n\phi) + a(u_n,P_n\phi) + (u_n',\phi - P_n\phi) + a(u_n,\phi-P_n\phi) \\ =& (P_n f,P_n \phi) + (u_n',\phi - P_n\phi) + a(u_n,\phi-P_n\phi) \\ =& (P_n f,\phi) + \color{blue}{(u_n',\phi - P_n\phi)} + \color{red}{a(u_n,\phi-P_n\phi) }. \end{aligned} $$ Blue term could vanish if we integrate w.r.t. time and integration by parts to move the derivative to $\phi-P_n \phi$. However for the red term to vanish, we must have something like $\Delta u_n = 0$...

EDIT2: If we say $f = g$ in $L^2(0,T;V')$, I don't know what the author did, but according to the Evans' book (see Section 7.1), it is $$\newcommand{\lsub}[2]{{\vphantom{#2}}_{#1}{#2}} \lsub{V'}{\langle f(t,\cdot),v \rangle}_V = \lsub{V'}{\langle g(t,\cdot),v \rangle}_V, $$ for any $v\in V$, at a.e. time $t\in [0,T]$. This is the same with what you wrote in the comment though.

  • $\begingroup$ Thanks for your reply. A thought occurred to me while reading it: since $(u_n', v) = (P_n u_n', v) = (P_n u_n', P_n v)$ by properties of the projection,and likewise with the other terms, maybe it does make sense as an equality in $L^2(0,T;V')$? Because we can take the pairing with any $v \in L^2(0,T;V)$ and use the calculation I just wrote. What do you think? $\endgroup$ – workl Jun 12 '13 at 15:30
  • $\begingroup$ Once we do the projection, say $$(u_n',P_n v) + (\nabla u_n, \nabla (P_n v)) = (P_n f, P_n v).$$ Still $\{P_n v\}$ is finite dimensional even for all test function $v$, we still can't say this equation holds in $L^2(0,T;V')$. $\endgroup$ – Shuhao Cao Jun 12 '13 at 15:46
  • $\begingroup$ I agree it's finite dimension but is that the definition of "equality?" As i understand, the equality holds in $L^2(0,T;V')$ if we can take the dual pairing with $v \in L^2(0,T;V)$ and that dual pairing makes sense. Because eg. $(P_nf, v) = (P_nf, P_nv)$, the equation above (1) tells us that $(P_nf, v)$ makes sense, and similarly for the other terms. $\endgroup$ – workl Jun 12 '13 at 15:58
  • $\begingroup$ So I guess it depends on defn. of equality?? $\endgroup$ – workl Jun 12 '13 at 15:59
  • $\begingroup$ @workl An equation $A u = f$ being valid in $L^2(0,T;V')$ is equivalent to say $\langle A u,v \rangle = \langle f,v \rangle$, for any $v\in V$. Notice $A u$ and $f$ are in $V'$. In the finite dimensional approximation, $(P_n f,v)$ certainly makes sense, and you can view $P_n f$ in $V'$ as well. But it is definitely not correct to state "(1) holds in $V'$": because $$(u_n',v) + (\nabla u_n,\nabla v) \neq (P_n f,v)$$ for all $v\in V$. Because $$(\nabla u_n,\nabla v)\neq (\nabla u_n,\nabla (P_n v))$$ in general. Did the author claim that? $\endgroup$ – Shuhao Cao Jun 12 '13 at 19:44

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.