Let $I$ be a bounded interval of $\mathbb{R}$ and $f$ be a continuous function on $I$.

Assume $I$ is closed, ie: $I=[a,b]$, and $\forall n\geq 0$, $\int_I x^n f(x)dx=0$. Then, one can show with Weierstrass approximation that $f=0$ on $I$.

This can be done because, with this assumption, $I$ is compact so one can apply Weierstrass theorem. But someone told me that the result holds even though $I$ is open, ie: $I=(a,b)$.

How can this be done ?

  • $\begingroup$ This has been addressed many times before. A possible approach is the following: $C^0(I)\subset L^2(I)$, and there is an orthogonal base of polynomials (shifted Legendre polynomials) for $L^2(I)$. By the given constraints, the Fourier-Legendre series of $f(x)$ over $I$ only has zero coefficients. By applying Parseval's identity, it follows that $\int_I f(x)^2\,dx = 0$, hence $f(x)\equiv 0$. $\endgroup$ – Jack D'Aurizio Oct 13 '16 at 14:01
  • $\begingroup$ A similar question: math.stackexchange.com/questions/1373720/… $\endgroup$ – Jack D'Aurizio Oct 13 '16 at 14:03
  • $\begingroup$ Are you assuming $f$ is absolutely integrable on $(a,b)$? Or are you assuming improper Riemann integrals exists on $(a,b)$? $\endgroup$ – DisintegratingByParts Oct 15 '16 at 16:08

If you assume that $f$ is absolutely integrable on $(a,b)$, then you can get what you want. To see this, suppose $g$ is continuous on $[a,b]$. Then the Weierstrass approximation theorem gives you a sequence $\{ g_n \}$ of polynomials on $[a,b]$ that converges uniformly to $g$. Therefore, $$ \left|\int_{a}^{b}g_n f dx - \int_{a}^{b}g fdx\right| \le \int_{a}^{b}|f(x)|dx\cdot\sup_{x\in [a,b]}|g_n(x)-g(x)|\rightarrow 0 \mbox{ as } n\rightarrow\infty. $$ Because $\int_{a}^{b}g_n(x)f(x)dx = 0$ for all $n$, it follows that $$ \int_{a}^{b}g(x)f(x)dx=\lim_n \int_{a}^{b}g_n(x)f(x)dx=0. $$ The above must hold for all $g \in C[a,b]$. Fix a subinterval $[c,d]\subset (a,b)$ and define $g_{\delta}(x)$ to be $f$ on $[c,d]$, to be $0$ on $(a,c-\delta)\cup(d+\delta,b)$, and to be linear and continuous on the remaining intervals $[c-\delta,c]\cup[d,d+\delta]$. Then $$ 0=\lim_{\delta\downarrow 0}\int_{a}^{b}f(x)g_{\delta}(x)dx=\int_{c}^{d}f(x)^2dx. $$ The above holds for all $a < c < d < b$. And $f(x)^2$ is continuous on $(a,b)$. By the Fundamental Theorem of Calculus, $$ 0=\frac{d}{dy}\int_{c}^{y}f(x)^2dx=f(y)^2. $$

  • $\begingroup$ where did you use the integrability hypthesis ? $\endgroup$ – anonymus Oct 16 '16 at 23:44
  • $\begingroup$ @anonymus : I used the absolute integrability of $f$ in the very first equation. Without that integral being finite, the right side is meaningless. $\endgroup$ – DisintegratingByParts Oct 17 '16 at 0:44
  • $\begingroup$ Oh yes of course ! Thank you :) $\endgroup$ – anonymus Oct 17 '16 at 0:56
  • $\begingroup$ @anonymus : You're welcome. It's good that you're being careful. $\endgroup$ – DisintegratingByParts Oct 17 '16 at 1:00

Viewing the problem as if it was about vectors in the space $L^2(X)$, we note that the spaces $L^2[a,b]$ and $L^2(a,b)$ are practicaly the same space ($L^2(X)$ is the space of square integrable, lebesgue measurable functions defined on the measurable space $X$).

This condition is equivalent to $f$ being orthogonal to a spanning set: $\{x^n\} _{n=1}^\infty$, and the only vector (function) that is orthogonal to a dense set is the zeroth vector.

Now, the zeroth vector in such spaces are functions that equal $0$ almost everywhere (up to a null set). Since $f$ is continuous, it must be $0$ everywhere.

  • $\begingroup$ Pretty. But it's a bit too sophisticated for my purpose. $\endgroup$ – anonymus Oct 13 '16 at 14:20
  • 1
    $\begingroup$ Sure, but I like this point of view and I think people can benefit from seeing a problem in a wider prespective. $\endgroup$ – Ranc Oct 13 '16 at 14:27

You can apply Weierstrass on the intervals $[a+\frac1k,b-\frac1k]$ for natural $k>\frac 2{b-a}$.

  • $\begingroup$ Is the following argument correct/rigorous ? Let $I_k :=[a+1/k,b-1/k]$ and $P_k$ a polynomial function s.t. $\|f-P_k \|_{L^\infty(I_k)}<\epsilon$. Since $\int_{I_k} fP_k = 0$, we have $\int_I f^2 = \lim_k \int_{I_k} f^2 = \lim_k \int_k f(f-P_k) \leq (\lim_k \int_k f)\epsilon \leq \lambda(I) \|f\|_{L^\infty (I)} \epsilon$. $\endgroup$ – anonymus Oct 13 '16 at 13:58
  • $\begingroup$ Why do all that? We get $f\equiv 0$ on each $I_k$ This implies $f\equiv 0$ on $(a,b).$ $\endgroup$ – zhw. Oct 15 '16 at 17:44
  • $\begingroup$ yes you're right : $\cup_k I_k =\lim_k I_k = (a,b)$. $\endgroup$ – anonymus Oct 16 '16 at 23:42

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.