If $Y\subset X^*$ is a closed subspace (where $X$ is a separable Banach space), the preannihilator of $Y$ in $X$ is $Y_{\perp}:=\{x\in X : y^*(x)=0, \forall y^*\in Y \}$. If $Y$ is a proper subspace of $X^*$, and $X$ is reflexive then it can be proved that $Y_{\perp}\neq\{0\}$. On the other hand if $X=l_1$ and $Y=c_0\subset l_{\infty}$, then $Y_{\perp}=\{0\}$.

  • 1) If $X^*$ is separable, is it still possible that $Y_{\perp}=\{0\}$?
  • 2) Can $c_0$ be a subspace of a separable dual? This question is probably unrelated, and most likely well known, but I realized I do not know an answer to it.

For (1), take $X = c_0$ so that $X^* = \ell^1$, which is separable. Take $Y = \{y \in \ell^1 : \sum_i y(i) = 0\}$ which is a proper closed subspace. Since $Y$ contains all the elements of the form $e_i - e_{i+1}$, where $e_i$ are the usual basis functions, you can see that any $x \in Y_\perp$ satisfies $x(1) = x(2) = x(3) = \dots$. The only such element of $c_0$ is $0$.

The intuitive idea is that by Hahn-Banach, $Y$ will certainly have a nontrivial annihilator living in $X^{**}$. In this case it is the constants in $X^{**} = \ell^\infty$ which are not in $X$.

More generally, whenever $X$ is non-reflexive, you should be able to choose any $f \in X^{**} \setminus X$ and take $Y$ to be its kernel. Then if $x \in Y_\perp$, it is also in the annihilator $Y^\perp$ when considered as an element $f_x$ of $X^{**}$. The kernel of $f_x$ contains the kernel of $f$, so $f_x$ is a scalar multiple of $f$. Since $f_x \in X$, this is only possible if $f_x = 0$, i.e. $x=0$.


The answer to Q2 is negative. One way to see this (which might not be the first, and might not be the easiest) is to combine the following results.

  1. A separable dual (Banach) space has the Radon-Nikodym property.

  2. The RNP is inherited by closed subspaces

  3. $c_0$ does not have the RNP.

(The statement that $c_0$ does not embed as a closed subspace of a separable dual space can be found as Theorem 6.3.7 in the book of Albiac-Kalton; the argument above is sketched by the same authors after Corollary 6.3.9.)


A negative answer to the second question can be obtained by a norming technique. Assume $0<a\leq b<\infty$ and $(x^*_n)_{n=1}^\infty\subset X^*$ are such that $$1/a\leq \inf \{\|\sum_{i=1}^n a_ix_i^*\|: n\in\mathbb{N}, \max_i |a_i|=1\} \leqslant \sup \{\|\sum_{i=1}^n a_ix_i^*\|: n\in\mathbb{N}, \max_i |a_i|=1\}\leq b.$$ In particular, $(x^*_n)_{n=1}^\infty$ is a basic sequence equivalent to the $c_0$ basis (and it must therefore be weakly null). Now let $(\phi_n)_{n=1}^\infty$ be the coordinate functionals to the $(x^*_n)_{n=1}^\infty$ and let $x^{**}_n$ be a Hahn-Banach extension of $\phi_n$. Then $x^{**}_n(x_m)=1$ if $m=n$ and $0$ otherwise.

Our assumptions imply that $\|x^{**}_n\|\leq a$ for all $n\in \mathbb{N}$. By Helly's Lemma (or Goldstine's theorem + perturbation, or local reflexivity if you don't mind overkill), we may find a sequence $(x_n)_{n=1}^\infty\subset 2a B_X$ such that $x^*_n(x_n)=1$ and $x^*_m(x_n)=0$ for $1\leq m<n$ (note that we do not control $x^*_m(x_n)$ for $m>n$).

Now fix a sequence $(\epsilon_n)_{n=1}^\infty$ of positive numbers such that $\sum_{n=1}^\infty \epsilon_n=1/2$. Since the sequence $(x^*_n)_{n=1}^\infty$ is weakly null, we may recursively select $n_1<n_2<\ldots$ such that, with $y^*_i=x^*_{n_i}$ and $y_i=x_{n_i}$, $|y^*_i(y_j)|<\epsilon_i$ for $i>j$. Note that these sequences retain the properties that $y^*_i(y_i)=1$ and $y^*_i(y_j)=0$ for $i<j$. Thus we have extracted subsequences $(y_i)_{i=1}^\infty$ and $(y^*_i)_{i=1}^\infty$ which are ``almost biorthogonal.'' Then the $c_0$ behavior of the $(y^*_i)_{i=1}^\infty$ sequence allows us to show $\ell_1$ behavior of $(y_i)_{i=1}^\infty$, which we do next.

Now fix $n\in \mathbb{N}$ and scalars $(b_j)_{j=1}^n$. For each $1\leq j\leq n$, fix $a_j$ with $|a_j|=1$ and $a_jb_j=|b_j|$. Let $x=\sum_{j=1}^n b_j y_j$ and $x^*=\sum_{j=1}^n a_j y^*_j$. Note that $\|x^*\|\leq b$, so \begin{align*} b\|x\| & \geq |x^*(x)| = \sum_{j=1}^n a_jb_jy^*_j(y_j)-\sum_{j=1}^n |b_j| \sum_{i=j+1}^n |a_jy^*_i(y_j)| \\ & =\sum_{j=1}^n |b_j| - \sum_{j=1}^n |b_j|\frac{1}{2} = \frac{1}{2}\sum_{j=1}^n |b_j|. \end{align*}

By the triangle inequality, $\|x\|\leq 2a\sum_{j=1}^n |b_j|$, so $(y_j)_{j=1}^\infty$ is a basis for a closed subspace $Y$ of $X$ isomorphic to $\ell_1$. This means $X^*$ has a quotient isomorphic to $\ell_\infty$, and $X^*$ is not separable.

By working slightly harder, one may also choose the $(y_j)_{j=1}^\infty$ to have $Y$ complemented in $X$, so $X^*$ actually has not just a quotient, but a complemented subspace isomorphic to $\ell_\infty$. Using results of James and sharper constants above, for $\epsilon>0$, we may have $Y$ $1+\epsilon$-isomorphic to $\ell_1$ and the range of a projection of norm at most $1+\epsilon$.

  • $\begingroup$ This is much more informative than the answer I found by "combining black boxes" :) $\endgroup$ – Yemon Choi Aug 16 '18 at 23:12
  • 1
    $\begingroup$ As my PhD advisor likes to point out to me repeatedly, I have an attitude of "Why say less when you can go on and on and on..." $\endgroup$ – user114263 Aug 17 '18 at 21:25

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.