Do these three kinds of vector spaces, those with an inner-product, those with a norm and those with a metric, are the same sets of vector spaces? At least for finite dimensional vector spaces all of these coincide?

It would be great to know finite dimensional counter-examples if they exist any and if anyone can link to some lecture notes explaining this point.

Like are there examples of inner product spaces which cannot have a metric or metric spaces which cannot have a norm or various other such possible conflicts between these 3 properties.

  • $\begingroup$ You have : inner-product $\implies$ norm $\implies$ metric, but the converses are wrong in general. $\endgroup$ – Watson Feb 12 '16 at 13:36
  • $\begingroup$ Yes, but with finite dimensional vector spaces will the converses also work? $\endgroup$ – guest Feb 12 '16 at 13:38
  • $\begingroup$ Not even in finite-dimensional vector spaces. $\endgroup$ – John B Feb 12 '16 at 13:42
  • $\begingroup$ in the finite dimensional case, all the topologies are equivalent, that is if a sequence $(\phi_n)_n$ is convergent $\to \phi$ for a certain norm, it will also converge $\to \phi$ for all the other possible norms on the vector space. this is because a norm asks for $||a u|| = |a| \, ||u||$ for any constant $a$ and $||u+v|| \le ||u||+||v||$ $\endgroup$ – reuns Feb 12 '16 at 14:04

Even in finite dimension, the converses are not true.

For instance, the $p$-norm ($p≠2$) on $\Bbb R^n$ doesn't come from an inner product. This can be proven by contradiction, using the parallelogram law or the polarization identity. The polarization identity $$2 \|x\|^2+2\|y\|^2=\|x+y\|^2+\|x-y\|^2 \qquad \forall x,y \in V$$ precisely gives a necessary and sufficient condition for a normed vector space $V$ to have an inner product that induces that norm.

The discrete metric on $\Bbb R$ doesn't come from a norm, because it takes only $0$ and $1$ as values.

If $d$ is a metric such that there is a norm $\|\cdot\|$ on $V$ with $d(x,y)=\|x-y\|, \; \forall x,y \in V$, then two necessary conditions are that $$d(ax,0)=|a|d(x,0) \tag 1 $$ $$d(x,y)=d(x+z,y+z) \tag 2$$ for every $x,y,z \in V$ and $a \in \Bbb R$. Conversely, if the metric $d$ satisfies $(1)$ and $(2)$, then $\|x\|:=d(x,0)$ yields a norm on $V$.

You can also see this question.

  • $\begingroup$ Thanks! Is there any reference for the proofs for the tests you said for the converses to work? Typically I would think of a metric on a vector space to be coming from a positive-semidefinite symmetric bilinear map/form (but I guess not all metrics come like this?) and these particular class of metrics will always then I guess induce a norm and such norms that they infact induce induce inner-products? $\endgroup$ – guest Feb 12 '16 at 14:17
  • $\begingroup$ For inner products/norms you can see here. For norms/metrics, you can try to prove by yourself, this is not difficult. A metric is already symmetric and "definite-positive". $\endgroup$ – Watson Feb 12 '16 at 14:22
  • $\begingroup$ @guest : Yes, if $(.,.)$ is a positive-definite symmetric bilinear form, then $d(x,y)=(x-y,x-y)$ is a metric, as I said before. All metrics don't come like this, as I said before (e.g. discrete metric). $\endgroup$ – Watson Feb 12 '16 at 14:25
  • $\begingroup$ but these psd bilinear form metrics will necessarily induce norms which will necessarily induce an inner product? $\endgroup$ – guest Feb 12 '16 at 14:28
  • $\begingroup$ Oops it was $d(x,y)=\sqrt{(x-y,x-y)}$ in my previous comment. $\endgroup$ – Watson Feb 12 '16 at 14:32

Take $X=\mathbb R^n$ and equip it with the topology induced by the the distance function $$ {\rm d}(x,y) = \begin{cases} 0 &\text{if } x=y\\1&\text{if }x\neq y\end{cases} $$ Then $X$ is a metric space, specifically the discrete metric space (in which all points are isolated) However, $(X,{\rm d})$ fails to be a normed space in that ${\rm d}$ does not induce any norm, as you can easily verify: $$ {\rm d}(\lambda x,\lambda y) \neq |\lambda|{\rm d}(x,y) \quad\text{whenever $x\neq y$ and $\lambda\neq\pm1$} $$

Let now $$ \|x\|_p = \left(\sum_{i=1}^n|x_i|^p\right)^{\frac1p} \quad\text{for some $p>0$} $$ Then you can easily verify that $\|{}\cdot{}\|_p$ is a norm for any $p>0$, hence $(X,\|{}\cdot{}\|_p)$ is a normed space, but for $p\neq 2$ it is not induced by a scalar product. Therefore, for $p\neq 2$ $(X,\|{}\cdot{}\|_p)$ is a normed space which is not induced by an inner product.


I don't know if this helps, but a metric space is not necessarily a vector space.

"The set of positive real numbers, ℝ+=(0,∞), with the metric given by d(x,y):=|x−y| is a metric space, but it is not a linear space, since it contains neither an additive identity (0) nor additive inverses (−x)."

Source: https://wiki.math.ntnu.no/users/ehrnstro/teaching/linearmethods/metricspaces


In the plane the ordinary Euclidean metric comes from the norm that comes from the usual inner product.

In the taxicab metric the distance from $(a,b)$ to $(c,d)$ is $|a-c| + |b-d|$. That metric does come from a norm, but the norm doesn't come from an inner product.

The metric for which the distance between any two distinct points is $1$ doesn't come from a norm.

  • $\begingroup$ On vector spaces is there an "iff" test which one can apply to a metric to know if it can come from a norm? Similarly an "iff" test one can do on a norm to test if it comes from an inner product? $\endgroup$ – guest Feb 12 '16 at 13:56
  • 1
    $\begingroup$ @guest : you can see my answer : the second one is solved thanks to the polarization identity, whereas the first one is just a homogeneity condition on the metric… $\endgroup$ – Watson Feb 12 '16 at 13:58
  • $\begingroup$ @Watson I was just about to point the OP to your fine answer. $\endgroup$ – Ethan Bolker Feb 12 '16 at 13:59

You can always use an inner product to define a norm and you can always use a norm to define a metric. Specifically, if $<a,b>$ is the inner product of $a$ and $b$, then $\sqrt{<a,a>}$ is a norm, i.e. you can say $\Vert a\Vert = \sqrt{<a,a>}$ and it has all the properties required of a norm. If you have a norm $\Vert \cdot \Vert$, then you can define a metric by $d(a, b)=\Vert a-b\Vert$ and that metric has all the properties required.

However, as pointed out in the comments, the reverse is not true. I.e. having a metric does not imply that you have an inner product.

  • $\begingroup$ But on finite dimensional vector spaces will the converses also always work? (..like trivially an "unit ball" is a metric space that is not a vecctor space and hence it makes no sense to talk of an inner-product or a norm structure on it anyway!..) And if one goes beyond vector spaces then do we know of criteria for the converses to work? $\endgroup$ – guest Feb 12 '16 at 13:43
  • $\begingroup$ @guest, What I'm saying is that any time you have an inner product you automatically have a norm and a metric. $\endgroup$ – TravisJ Feb 12 '16 at 13:44
  • $\begingroup$ I am asking about conditions when the converse would work.. $\endgroup$ – guest Feb 12 '16 at 13:45
  • $\begingroup$ @guest : the polarization identity gives a condition for the "norm $\implies$ inner product" (you can see my answer). $\endgroup$ – Watson Feb 12 '16 at 13:49
  • $\begingroup$ Just a side note: the norm should be $\sqrt{ <a,a> }$, I think. $\endgroup$ – Watson Feb 12 '16 at 14:33

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.