Let $V$ be a vector space. Assume $H_1$ and $H_2$ are subspaces of $V$, and that both $H_1$ and $H_2$ are Hilbert spaces with inner-products $\langle \cdot, \cdot\rangle_1$ and $\langle \cdot,\cdot\rangle_2$ respectively. Let $x\in H_1$, and let $\left\{h_n,\,n\in\mathbb{N}\right\}$ be an orthonormal set in $H_1$ (not necessarily a basis) such that $$\big\Vert x - \sum_{k=1}^n\langle x, h_k\rangle_1 h_k\big\Vert_1 \underset{n\rightarrow\infty}\longrightarrow 0,$$ that is, $x$ lies in the closed linear span of $\left\{h_n\right\}$.

Assume now that $x\in H_2$ and $\left\{h_n\right\}\subset H_2$. Notice that since $\left\{h_n\right\}$ is orthonormal in $H_1$, it is linearly independent in $H_1$ and hence in $V$ and $H_2$, but not necessarily orthogonal in $H_2$. Let $\left\{e_n,\,n\in\mathbb{N}\right\}$ be an orthonormal set in $H_2$ obtained by a Gram-Schmidt process on $\left\{h_n\right\}$, that is, such that $\mbox{span}\left\{e_1,\dots,e_n\right\} = \mbox{span}\left\{h_1,\dots,h_n\right\}$ for each $n$. Is it true that $$\big\Vert x - \sum_{k=1}^n\langle x, e_k\rangle_2 e_k\big\Vert_2 \underset{n\rightarrow\infty}\longrightarrow 0?$$


In response to user gerw: let me be more specific and maybe we can relate the two norms. Given two probability measures $\mu$ and $\nu$ on $(\mathbb{R},\mathcal{B})$, both equivalent to Lebesgue measure, let $V$ be the vector space of $\mu$-equivalence classes of measurable functions $f:\mathbb{R}\rightarrow\mathbb{R}$ (which is of course equal to the set of $\nu$-equivalence classes of such functions), let $H_1 = L^2(\mu)$ and $H_2 = L^2(\nu)$.

Let $T_\mu:L^2(\mu)\rightarrow L^2(\mu)$ be the positive Hilbert-Schmidt operator defined by $$T_\mu f(x) = \int c(x,y) f(y) d\mu(y),\qquad f\in L^2(\mu)$$ where $c$ is a bounded, measurable kernel. Define $T_\nu$ similarly.

I am given a certain bounded measurable function $\varphi$ such that both $$\big\Vert \varphi - \sum_{k=1}^n\langle \varphi, h^\mu_k\rangle_1 h^\mu_k\big\Vert_1 \underset{n\rightarrow\infty}\longrightarrow 0\quad \mbox{and} \quad\big\Vert \varphi - \sum_{k=1}^n\langle \varphi, h^\nu_k\rangle_2 h^\nu_k\big\Vert_2 \underset{n\rightarrow\infty}\longrightarrow 0$$ hold, where $\left\{h^\mu_n\right\}$ is the orthonormal set of eigenfunctions of $T_\mu$, and similarly for $\left\{h^\nu_n\right\}$. I'd like to show that $\varphi$ is the $\Vert\cdot\Vert_2$-limit of linear combinations of the $h^\mu_n$.

My question can be rephrased as follows: is the $\Vert\cdot\Vert_2$-closure of $\mbox{span}\left\{h_n^\mu\right\}$ equal to the $\Vert\cdot\Vert_2$-closure of $\mbox{span}\left\{h_n^\nu\right\}$?

  • 2
    $\begingroup$ Effectively, you ask whether $x$ lies in the closure (w.r.t. $H_2$) of the span of $\{h_n\}$. Since the norms of $H_1$ and $H_2$ cannot be compared without further assumptions, I suspect that the answer will be "False" in general. However, I do not have a counterexample at hand. $\endgroup$ – gerw Jan 10 '16 at 18:43

Let $H_1 = L^2[0,\pi]$, and let let $H_2$ be the subspace of absolutely continuous functions on $[0,2\pi]$ with first derivative in $L^2$ and with Sobolev norm $$ \|f\|_2 = \sqrt{\|f\|_{L^2}^{2}+\|f'\|_{L^2}^{2}}. $$ Define $h_n(x) = \sin(nx)$ for $n=1,2,3,\cdots$. Then $\{ h_n \} \subset H_2$ as well. The constant function $1$ is in $H_1, H_2$ also. The set $\{ \sqrt{2}\sin(nx)\}_{n=1}^{\infty}$ is a complete orthonormal basis of $H_1$. Hence, $$ \lim_N\left\|1 - 2\sum_{n=1}^{N}(1,\sin(nx))_1\sin(nx)\right\|_{L^2} = 0. $$ The functions $\{ \sin(nx) \}_{n=1}^{\infty}$ are in $H_2$, and $1\in H_2$. However $1$ is not in the closure of the linear span of $\{ \sin(nx)\}_{n=1}^{\infty}$ in $H_2$ because convergence in $H_2$ implies uniform convergence on $[0,\pi]$; to see why, \begin{align} f(x) & =xf(x)+(1-x)f(x) \\ & =\int_{0}^{x}\frac{d}{dt}(tf(t))dt -\int_{x}^{1}\frac{d}{dt}((1-t)f(t))dt \\ & =\int_{0}^{x}\{f(t)+tf'(t)\}dt-\int_{x}^{1}\{-f(t)+(1-t)f'(t)\}dt,\\ |f(x)| & \le \|f\|_{L^2}\|1\|_{L^2}+\|f'\|_{L^2}(\|t\|_{L^2}+\|1-t\|_{L^2}) \\ & \le C\sqrt{\|f\|_{L^2}^2+\|f_{L^2}'\|_1^2} \\ & = C\|f\|_{2}. \end{align}


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.