In Nash & Sen p.162, they show that the space of all positive definite symmetric matrices $C$, while not a vector space itself, is homeomorphic to the space of all symmetric matrices $S$ (which is a vector space), via the map

$$ s \rightarrow e^s \in C, s \in S $$

My question: can I not make $C$ into a vector space by defining addition of two elements as first adding the elements in $S$ and then exponentiating, i.e. $ e^a + e^b \equiv e^{a + b}$? Note this is not the usual multiplication of matrices. It seems to me that we then have an inverse ($e^{-a}$) and an identity element ($e^0$), that we inherit commutativity etc. from $S$, and we can define scalar multiplication in the usual way.

More generally, why is something homeomorphic to a vector space not itself a vector space?

Can someone tell me where I'm going wrong?

  • 4
    $\begingroup$ The article about transport of structure might be of interest to you: en.wikipedia.org/wiki/Transport_of_structure $\endgroup$ – Vercassivelaunos Apr 21 at 7:49
  • 8
    $\begingroup$ The issue is that the already existing operations on $C$ do not make it into a vector space. It is not even closed under taking inverses. However as you point out, you can 'borrow' the vector space structure of $S$ using the bijection $s\mapsto e^s$. This works in general, but again the vector space structure you obtain may have little to do with any already existing one. $\endgroup$ – Alejandro Epelde Apr 21 at 7:50
  • $\begingroup$ @AlejandroEpelde Ah! They use the proof in the context of showing that C is contractible (because vector spaces are). Does that mean that, given any set, it is contractible if I can make it into a vector space (even if the addition rule or whatever is not the usual one)? $\endgroup$ – quixot Apr 21 at 7:53
  • 2
    $\begingroup$ That seems reasonable, so long as the resulting vector space operations are continuous with respect to some topology on your set. Which in this case they are since $s\mapsto e^s$ is a homeomorphism. $\endgroup$ – Alejandro Epelde Apr 21 at 7:57
  • $\begingroup$ @AlejandroEpelde Excellent, that makes sense. Thank you $\endgroup$ – quixot Apr 21 at 8:00

Here is a simpler example: $(-1,1)$ and $\Bbb R$ are homeomorphic; just consider the map$$\begin{array}{rccc}f\colon&(-1,1)&\longrightarrow&\Bbb R\\&x&\mapsto&\tan\left(\frac\pi2x\right).\end{array}$$However, $(-1,1)$ isn't automatically a vector space. It becomes a vector space if you define the addition by $x+y=f^{-1}\bigl(f(x)+f(y)\bigr)$ and if you define the multiplication by a scalar by $\lambda x=f^{-1}\bigl(\lambda f(x)\bigr)$. This also works in that situation that you described indeed.

  • $\begingroup$ Okay great thank you! So as long as I can create rules which make a set into a vector space (equivalently, find a homeomorphism to a vector space) I can use the properties of vector spaces? In the Nash & Sen example, showing that a space is contractible because vector spaces are. $\endgroup$ – quixot Apr 21 at 7:56
  • $\begingroup$ @quixot This is because this set is homeomorphic to a vector space. In general, for cardinality reason, one can find a bijection from some set to some vector space (for example, $S^1$ and $\mathbb{R}$ are in bijection), which cannot lead to a homeomorphism. $\endgroup$ – Didier Apr 21 at 8:42

The set of positive semi-definite matrices has a structure on its own: it is a cone (The space of positive semidefinite $n \times n$ matrices is a cone). In particular, it is convex. See also the deep article here with its geometrical point of view.


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.