Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Let $f$ and $g$ be locally integrable, say on $R^n$ (for arbitrary open domains, just extend trivially). Suppose

$\forall \phi \in C_c^\infty : \int f \phi dx = \int g \phi dx$.

Let $K = supp(\phi)$. If $f,g \in L^2(K)$, we see by Hilbert space theory $f = g$ almost everywhere, as the $C_c^\infty(K)$ is dense in $L^2(K)$.

So the theorem holds for $f,g \in L^2_{loc}$. But $L^\infty_{loc} \subset L^2_{loc}$ is dense in $L^1_{loc}$ (in the respective topology), hence we have $f = g$ almost everywhere even for $L^1_{loc}$.

I haven't seen this proof in literature, and I wonder whether there is a gap. If that is the case, do you know a better proof?

share|improve this question
1  
I don't quite see this. $f$ is approximated by a sequence $(f_n)$ in $L^2$ and so is $g$ (by $(g_n)$). How do you ensure that $\int f_n \phi=\int g_n \phi$? Anyway, the result is called the "fundamental lemma of calculus of variations", and you will find proofs in lots of places (for example any introduction to distributions must include this). –  Florian Mar 24 '11 at 9:39
    
Where do you see a gap? This is completely standard (and there's no need to pass to $L^2$). You can find a proof e.g. in Yosida's book on functional analysis (pp. 48f). –  t.b. Mar 24 '11 at 9:46

2 Answers 2

up vote 4 down vote accepted

Like Theo said, this is a standard question whose standard answer (as far as I know) uses smoothing by convolution: take an arbitrary compact $K \subset \mathbb{R}^n$, regularize its characteristic function $\chi_K$ by means of a mollifier sequence and use the hypothesis to show that $\int_K \big(f(x)-g(x)\big)\ dx=0$.

I'd like to provide another proof which doesn't rely on convolution (at least not explicitly). The idea behind it is precisely the same, though.

We need two lemmas.

Lemma 1 If a Borel measurable function $f \colon \mathbb{R}^n \to \mathbb{R}$ is such that, for every (compact) rectangle $R$,

$$\int_R f(x)\,dx=0,$$

then $f=0$ almost everywhere.

Proof Start by recording that the hypotheses implicitly tell us $f \in L^1_{\rm{loc}}(\mathbb{R}^n)$ (click).

It is a known fact (see, e.g., Rudin's Real and complex analysis, §2.19) that every open subset of $\mathbb{R}^n$ can be written as the union of a countable collection of rectangles intersecting only at the boundary. So the $\sigma$-algebra generated by the collection of the rectangles is the whole Borel $\sigma$-algebra.

Fix $M >0$ and consider the rectangle $R_M=[-M, M] \times \ldots \times [-M, M]$. The family

$$\mathcal{A}_M=\left\{ B \subset R_M \mid \int_B f(x)\,dx\ \text{makes sense and vanishes}\right\}$$

is a $\lambda$-system (click) of subsets of $R_M$, because

  • $R_M \in \mathcal{A}_M$;
  • for every $A, B \in \mathcal{A}_M\ \mathrm{s.t.}\ A \subset B$, we have $$\int_{B-A}f(x)\,dx=\int_B f(x)\, dx-\int_A f(x)\, dx=0;$$
  • if $A_1 \subset A_2 \subset \ldots \in \mathcal{A}_M$ then, once put $A=\cup_k A_k$, we have $$\int_A f(x)\, dx= \lim_{k \to \infty} \int_{A_k} f(x)\, dx=0$$ by dominated convergence: in fact $\lvert f\chi_{A_k} \rvert \le \lvert f \chi_{R_M}\rvert \in L^1(\mathbb{R}^n)$ (recall that $f \in L^1_{\rm{loc}}$).

$\mathcal{A}_M$ contains all rectangles of $\mathbb{R}^n$ contained in $R_M$ and those obviously form a $\pi$-class (a class of subsets closed for finite intersections): now by the $\pi$-$\lambda$ theorem (click) we can conclude that $\mathcal{A}_M$ contains all Borelian subsets of $R_M$ and thus $f$ vanishes almost everywhere on it. Since $M$ was arbitrary, we infer that $f$ vanishes almost everywhere on $\mathbb{R}^n$. $\square$

The other lemma we need is the existence of bump functions:

Lemma 2 Let $R=[a_1, b_1] \times \ldots \times [a_n, b_n], \varepsilon > 0, R_{\varepsilon}=[a_1-\varepsilon, b_1+\varepsilon] \times \ldots \times [a_n - \varepsilon, b_n + \varepsilon]$. Then there exists a function $b\in C^\infty_c(\mathbb{R}^n)$ s.t. $0\le b \le 1,\ b \equiv 1$ on $R$ and $\mathrm{supp}(b)\subset R_\varepsilon$.

Proof For each $j$ let $$\varphi^{-}_j(t)=\begin{cases} e^{\frac{1}{(a_j-\varepsilon - t) (t-a_j)}} & t \in (a_j-\varepsilon, a_j) \\ 0 & \text{otherwise} \end{cases}.$$ Then put $\psi^{-}(s)=C^{-}_j\int_{a_j-\varepsilon}^s\varphi^{-}(t)\,dt$ where $C_j$ is a costant chosen so that $\psi\le 1$. Now do the same for $b_j$: let $$\varphi^{+}_j(t)=\begin{cases}e^{\frac{1}{(b_j- t) (t-b_j-\varepsilon)}} & t \in (b_j, b_j+\varepsilon) \\ 0 & \text{otherwise} \end{cases}$$ and $\psi^{+}_j(s)=C^{+}_j\int_{b_j+\varepsilon}^s \varphi^{+}(t)\,dt.$

The function $b_j = \psi_j^-\psi_j^+$ is $C^\infty$ and satisfies $0 \le b_j \le 1, b_j \equiv 1$ on $[a_j, b_j]$ and $\mathrm{supp}(b_j) \subset [a_j-\varepsilon, b_j + \varepsilon]$. So $b(x_1 \ldots x_n)=\prod_j b_j(x_j)$ is what we were looking for. $\square$

With those lemmas it is pretty easy to prove the original claim. In fact, let $f, g$ as in the hypothesis and put $F=f-g$. Take a rectangle $R$ and, for every $n\in \mathbb{N}$, a bump function $b_n$ separating $R$ from $R_{n^{-1}}$. Then $b_n \to \chi_{R}$ pointwise, so that $Fb_n \to F\chi_R$ pointwise; also $\lvert Fb_n \rvert \le \lvert F \chi_{R_1}\rvert \in L^1(\mathbb{R}^n)$. By the dominated convergence theorem, we get $$\int_R F(x)\,dx=0.$$ By Lemma 1, $F\equiv 0$ almost everywhere.

share|improve this answer
    
Nice answer! ${}$ –  t.b. Mar 24 '11 at 13:53
    
Thank you Theo! I've edited a bit, there were some mistake here and there. This question spanned another one... How to prove Lemma 1 without that additional assumption? I'll open another thread for this. –  Giuseppe Negro Mar 24 '11 at 15:11
    
Well, in fact that "additional assumption" was already hard-coded into the hypotheses. Edited. –  Giuseppe Negro Mar 24 '11 at 19:10

Another way to do this is to let $\phi(x)$ be a smooth cutoff function equal to $1$ on some ball $B(a,r)$. Then your condition implies that the Fourier coefficient $\int_{R^n}(f(x) - g(x))\phi(x)e^{i\xi \cdot x}\,dx$ is zero for all $\xi \in R^n$. By the uniqueness theorem for the Fourier transform, $(f(x) - g(x))\phi(x) = 0$ a.e, so $f(x) = g(x)$ a.e. on the ball $B(a,r)$. This ball was arbitrary so $f(x) = g(x)$ a.e.

share|improve this answer

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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