Is there a topological vector space such that, for every $x\in X$, there is a proper neighbourhood $V$ of $x$ in $X$ which is convex, but the whole space is not locally convex (i.e. $X$ has a local base consisting of convex sets in each of its point)?

  • 6
    $\begingroup$ I think you need to narrow it down a little bit, because the whole space is obviously a convex neighbourhood of each of its points, so any non-locally convex space will satisfy your query. $\endgroup$ – kahen Aug 27 '15 at 22:15
  • 1
    $\begingroup$ for a topological vector space $V$ is a neighbourhood of $x$ iff $V-x$ is a neighbourhood of $0$, so the existence of a neighbourhood for all $x$ is equivalent to the existence of a neighbourhood of $0$, $\endgroup$ – Hamza Aug 27 '15 at 23:36

For $0 < p < 1$, let

$$\ell^p(\mathbb{N}) = \Biggl\{ x \colon \mathbb{N}\to \mathbb{C} : \sum_{n = 0}^\infty \lvert x(n)\rvert^p < +\infty\Biggr\}.$$

We endow it with the $p$-seminorm

$$s_p(x) = \sum_{n = 0}^\infty \lvert x(n)\rvert^p$$

and the metric $d_p(x,y) = s_p(x-y)$ derived from it.

Note the difference from the case $p \geqslant 1$. If $p > 1$, one needs to take the $p^{\text{th}}$ root to have the triangle inequality, and one obtains a normed, hence locally convex, space. For $p < 1$, taking the $p^{\text{th}}$ root would destroy the triangle inequality and thus we obtain a space that isn't locally convex.

$s_p$ is similar to a norm, but it satisfies $s_p(\lambda x) = \lvert\lambda\rvert^p\cdot s_p(x)$ instead of the homogeneity a norm has.

The triangle inequality follows from

$$a^p + b^p \leqslant (a+b)^p\tag{1}$$

for $a,b \geqslant 0$ and $0 < p < 1$. For $a = 0$ or $b = 0$ that is immediate, and for $a,b > 0$ we have $\frac{a}{a+b}, \frac{b}{a+b} \in (0,1)$, so

$$1 = \frac{a}{a+b} + \frac{b}{a+b} < \biggl(\frac{a}{a+b}\biggr)^p + \biggl(\frac{b}{a+b}\biggr)^p,$$

whence $(1)$. The translation-invariance of $d_p$ and the $p$-homogeneity ensure that the induced topology is a vector space topology:

$$d_p(x+y,x_0+y_0) = s_p\bigl((x-x_0)+(y-y_0)\bigr) \leqslant s_p(x-x_0) + s_p(y-y_0) = d_p(x,x_0) + d_p(y,y_0)$$

shows that addition is uniformly continuous, and

\begin{align} d_p(\lambda x,\lambda_0 x_0) &= s_p(\lambda x - \lambda_0 x_0) \\ &= s_p\bigl((\lambda-\lambda_0)x_0 + \lambda_0(x-x_0) + (\lambda - \lambda_0)(x-x_0)\bigr)\\ &\leqslant s_p\bigl((\lambda-\lambda_0)x_0\bigr) + s_p\bigl(\lambda_0(x-x_0)\bigr) + s_p\bigl((\lambda-\lambda_0)(x-x_0)\bigr)\\ &= \lvert\lambda -\lambda_0\rvert^p\cdot s_p(x_0) + \lvert\lambda_0\rvert^p\cdot d_p(x,x_0) + \lvert\lambda-\lambda_0\rvert^p\cdot d_p(x,x_0) \end{align}

shows the continuity of scalar multiplication at $(\lambda_0,x_0)\in \mathbb{C}\times \ell^p(\mathbb{N})$.

These spaces are not locally convex: Define the "standard unit vectors" $e_k$ by

$$e_k(n) = \begin{cases} 1 &, n = k\\ 0 &, n \neq k.\end{cases}$$

Then $s_p(e_k) = 1$ for all $k$, and by the $p$-homogeneity $y_k := (\varepsilon/2)^{1/p}\cdot e_k \in B_\varepsilon(0)$ for $\varepsilon > 0$. But with $x_N = \frac{1}{N}\sum_{k = 0}^{N-1} y_k$, we have

$$s_p(x_N) = N\cdot\frac{\varepsilon}{2N^p} = N^{1-p}\frac{\varepsilon}{2} \xrightarrow{N\to\infty} +\infty,$$

so no $d_p$-ball contains the convex hull of any other $d_p$-ball.

Yet, there are plenty of convex open sets, since the inclusion $\iota_p \colon \ell^p(\mathbb{N}) \hookrightarrow \ell^1(\mathbb{N})$ is continuous.

| cite | improve this answer | |

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.