I am following the proof of the following Poincaré's inequality from the Lieb and Loss textbook

8.12 Let $Ω∈\mathbb{R}^n$ be a bounded, connected open set that has the cone property for some $θ$ and $r.$ Let $1≤p≤∞$, and let $g$ be a function in $L^{p′}(Ω)$, s.t. $\int_{Ω}g=1.$ Let $1 < q < np/(n - p)$ when $p < n,$ $q <\infty$, when $p = n,$ and $1 < q <\infty$ when $p > n.$ Then there is a finite number $S>0,$ which depends on $Ω,g,p,q$ such that \begin{equation}\lvert|f−\int_Ωfg\rvert|_{L^q(\Omega)}≤S||∇f||_{L^p(\Omega)},\quad \quad\quad \quad (1)\end{equation} $∀f∈W^{1,p}(Ω).$

They begin the proof by assuming $q\geq p$. Then they assume, by contradiction, that the Poincaré inequality $(1)$ is false for every $S>0$.

I don't understand the next part: Then there is a sequence of functions $f_j$, such that the left side of the inequality $(1)$ equals to $1$, for all $j$, while the right side tends to zero as $j\rightarrow \infty$.

Next they let $h_j=f_j+\int_{\Omega}gf_j$, such that the gradient of $f_j$ equals the gradient of $h_j$. (Why?) Then the sequence $h_j$ is bounded in $W^{1,p}(\Omega).$

The rest of the proof I think makes sense and follows from Rellich-Kondrachov theorem, but I was wondering if someone can please ellaborate on the bolded steps?


1 Answer 1


If the Poincaré inequality is false for every $S$, then for each $j$, we may take $S=j$ to see that there exists a function $f_j$ for which $$\begin{equation}\lvert|f_j−\int_Ωf_j g\rvert|_{L^p(\Omega)}> j ||∇f_j||_{L^p(\Omega)},\quad \quad\quad \quad (1)\end{equation}$$ Both sides are homogeneous in $f$, so by replacing $f_j$ with $c_j f_j$ for some constant $c_j$, we can assume the left side equals 1. Therefore we have $||∇f_j||_{L^p(\Omega)} < 1/j$ which tends to 0 as $j \to \infty$.

Replacing $f_j$ with $h_j$ by adding the constant $\int g f_j$ is just a convenience. The left side now equals $1 = \|h_j\|_{L^p}$ and the right side still equals $\|\nabla f_j\|_{L^p} = \|\nabla h_j\|_{L^p} < 1/j$. Then the $W^{1,p}$ norm of $h_j$ is comparable to $\|h_j\|_{L^p} + \|\nabla h_j\|_{L^p} = 1 + 1/j \le 2$.

Some authors would summarize this step just by saying: suppose without loss of generality that $\int g f_j = 0$.

  • $\begingroup$ Thank you that is much clearer now! I have one more question; they state next that $\int_{\Omega} hg=0$ which seem to follow from the fact that $\nabla h=0$, and therefore $h$ is a constant function. Are you able to clarify? $\endgroup$
    – kirkos73
    Dec 12, 2022 at 17:19
  • $\begingroup$ No, the equality $\int hg = 0$ is nothing to do with $\nabla f$, and at this point there is no reason for $h$ to be constant. It's just simple algebra. Plug in $h(x) = f(x) - \int f(y)g(y)\,dy$, and distribute over the multiplication by $g(x)$. The first term is $\int f(x) g(x)\,dx$. The second term is now $\int \int f(y) g(y) dy \, g(x) dx$. The inner integral does not depend on $x$, so this equals $\left(\int f(y)g(y)\,dy\right) \left(\int g(x)\,dx\right)$. But $\int g = 1$ so this just equals $\int f(y)g(y)\,dy$, which exactly cancels the first term. $\endgroup$ Dec 15, 2022 at 2:11

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .