I am looking for the tangent space of $SymSL(n,\mathbb{R}) = \{A\in\mathbb{R}^{n\times n} | \,A^{\rm T} = A,\; \det{A}=1 \}$ (actually $n=3$) at an arbitrary $M\in SymSL(n,\mathbb{R})$.

(Actually I am interested in $SymSL^+(n,\mathbb{R}) = \{A\in\mathbb{R}^{n\times n} | \,A^{\rm T} = A,\; \det{A}=1,\; A \text{ is positive definite} \}$, but I have the feeling it boils down to the title of this question.)

Related questions:

Tangent Space of SL(n,ℝ) at arbitrary point, e.g. not at $\mathbb{1}$

Tangent space of Sym(n,ℝ)

The first question's answer doesn't work here because $Sym(n)$ is not a group, the second question's answer is not applicable because $SL(n)$ is not a vector space. Hence $SymSL(n,\mathbb{R})$ is neither a group nor a vector space.

I would very much appreciate any comments.

edit #2: $SL(n,\mathbb{R})$ is connected and smooth. So $SymSL(n,\mathbb{R})$ is the intersection of a connected smooth manifold with a vector space. This makes me believe we can talk about a tangent space in this case. Some more possibly usefull facts: the set $SymSL^+(n,\mathbb{R})$ I am actually interested in is a connected, simply connected and complete Riemannian manifold.

  • $\begingroup$ Does it matter that $SymSL(n,\mathbb{R})$ isn't a group here? $\endgroup$ – Santana Afton Apr 13 '17 at 7:21
  • $\begingroup$ Yes, the first related question's answer makes use of SL being a group. Then you can just move the tangent space from the identity element to any other by group action. $\endgroup$ – Everyday Astronaut Apr 13 '17 at 7:23
  • $\begingroup$ So, what about $SymSL(n,\mathbb{R})$? It isn't a group. $\endgroup$ – Santana Afton Apr 13 '17 at 7:25
  • $\begingroup$ Thats why I don't have an answer to the question of this topic. I also emphasized this fact in the edit. In case you wonder about the solvability of this problem, see the second edit (will be there in 1 min). $\endgroup$ – Everyday Astronaut Apr 13 '17 at 7:44

Let $\mathcal{S}$ be the vector space (of dimension $k=n(n+1)/2$) of the real symmetric matrices. We consider the algebraic set $V=\{X\in\mathcal{S}| \det(X)=1\}$. Let $A\in V$. Then the tangent space of $V$ in $A$ is $T_A(V)=\{H\in \mathcal{S}|trace(HA^{-1})=0\}$.

Proposition. $V$ has dimension $k-1$.

Proof. This is equivalent to show that $dim(T_A(V))=k-1$ or that the linear form $f:H\in\mathcal{S}\rightarrow trace(HA^{-1})$ is not identically zero; this last point is clear because $f(A^{-1})=trace(A^{-2})>0$.

Remark. Note that $W=\{X\in V|X>0\}$ is open in $V$. Then, the tangent space of $W$ in $A\in W$ is the same as the tangent space of $V$ in $A$. Of course, $dim(W)=k-1$.

EDIT. Answer to @ Dr Doolittle . Let $k\leq n$, $f:\mathbb{R}^n\rightarrow \mathbb{R}^k\in C^1$ , $V=\{x\in\mathbb{R}^n|f(x)=0\}$ and $a\in V$. If $Df_a$ has full rank, then, in a neighborhood of $a$, $V$ is a variety of dimension $n-k$ and the tangent space of $V$ in $a$ is $T_a(V)=\{h\in\mathbb{R}^n|Df_a(h)=0\}$.

Here $f(X)=\det(X)-1$ and $Df_A(H)=trace(HA^{-1})$.

  • $\begingroup$ I would be grateful if you could prove that $T_A(V)$ is the tangent space of $V$ in $A$, or give a reference to a proof. Then the question would be wholly answered. $\endgroup$ – Everyday Astronaut May 4 '17 at 8:29
  • $\begingroup$ I see it now. Actually this answer has been given in the opening post of the related question math.stackexchange.com/questions/2231893/… with a different notation. Thank you! $\endgroup$ – Everyday Astronaut May 4 '17 at 9:26
  • $\begingroup$ Addendum: the proof of $T_A(V)$ being the tangent space of $V$ in $A$ simply by deriving the formula $\mathrm{trace}(HA^{-1})=0$ does not necessitate that $T_A(V) \subset S$, i.e. one may think the tangents could be non-symmetric. However they must be symmetric because $V$ is an algebraic variety in the vector space $S$, so all tangent objects must come from $S$. $\endgroup$ – Everyday Astronaut Jun 20 '17 at 9:32

The following is a rather trivial fact that seems to immediately provide a partial solution. Furthermore, this partial solution may be in fact a complete one in many cases.

Let $N$ be a smooth manifold, and let $L,M\subset N$ be submanifolds. Let $p\in L\cap M$ such that around $p$ the intersection $L\cap M$ is smooth. Then$$T_p(L\cap M)\subset \left(T_pL\right)\cap\left(T_pM\right).$$If $L$ and $M$ intersect transversally, the above $\subset$ is an equality.

Edit: A bit more on transversality: We say $L$ and $M$ intersect each other transversally if for every $p\in L\cap M$ we have$$T_pN=T_pL+T_pM.$$The most important fact about transversality is that whenever $L$ and $M$ intersect each other transversally, it automatically follows that $L\cap M$ is a smooth submanifold. Additionally, we also have $$T_p(L\cap M)=(T_pL)\cap(T_pM)$$for any $p\in L\cap M$. In particular, we know the dimension of $L\cap M$.

  • $\begingroup$ Thank you for this contribution. What are criteria for the transversality of intersections? If not generally, then in the case $L=Sym(n)$, $M=SL(n)$, $N=\mathbb{R}^{n \times n}$. $\endgroup$ – Everyday Astronaut Apr 25 '17 at 11:05

I am interested in seeing an answer to this question, I haven't been able to get an answer but this is what I do have to help prompt a conclusive answer.

First as noted in the OP the space of positive symmetric matrices is a Riemannian manifold with tangent space at a particular point $\mathbf{P}$ given by $\{ \mathbf{P} \} \times \mbox{Sym}(n,\mathbb{R})$. We are interested in the variety defined by $\{ \mathbf{P} \; : \; \mbox{det}(\mathbf{P}) = 1 \}$, using the point in Amitai's answer leads to $$ T_{\mathbf{P}} \mathcal{M} \subset \mbox{Sym}(n,\mathbf{R}) \cap \{ \mathbf{M} \; : \; \mbox{tr}(M) = 0\}. $$ Proving one way or another that subset is proper or not I haven't been able to do, but here is a motivating example in $\mathbb{R}^{2\times 2}$, I also define the symmetric, trace zero matrices $$ \mathbf{E}_1 = \begin{bmatrix} 1 & 0\\0 &-1 \end{bmatrix}, \qquad \mathbf{E}_2 = \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}. $$ Now we want $$ 0 = \frac{d}{dt}\mbox{det}\,\mathbf{A}(t) = \mbox{tr}\left(\mbox{adj}\, \mathbf{A}(t) \frac{d A}{dt} \right), $$ or $$ \mbox{adj} \, A(t) \frac{dA}{dt} = c_1 \mathbf{E}_1 + c_2\mathbf{E}_2, $$ for some constants $c_1$ and $c_2$, although I guess $c_1, c_2$ could be time dependent also? Anyway working with this I get \begin{align} \frac{dA}{dt} = \frac{1}{\mbox{det}(\mbox{A}) } \mbox{A} \left(c_1\mathbf{E}_1 + c_2 \mathbf{E}_2 \right). \end{align} In particular we should have $\mathbf{P} = \mathbf{A}(0)$ and therefore we should be able to find constants $c_1$ and $c_2$ such that $$ \mathbf{P} \left(c_1 \mathbf{E}_1 + c_2 \mathbf{E}_2 \right) = \mathbf{E}_i, \qquad i = 1,2 $$ a particular instance of this when $i=1$ implies $$ \begin{bmatrix} c_1 & c_2 \\ c_2 & -c_1 \end{bmatrix} = \begin{bmatrix} p^{-1}_{11} & -p^{-1}_{12} \\ p^{-1}_{12} & -p^{-1}_{22} \end{bmatrix} $$ (where by $p_{ij}^{-1}$ I mean the $i,j$th entry of the inverse matrix) which jointly seem to suggest that $\mathbf{P} = \mathbf{I}$. On the otherhand keeping $\mathbf{P}$ fixed then $c_2$ is determined as soon as $c_1$ is which suggests the tangent space is only one dimensional.

Would be very interested in comments or a solution for the general case.

  • $\begingroup$ I'm unable to follow the argument from "Now we want". Why do we want that? The only way I can think of the trace coming into play is if $\mathbf{A}(0)=\mathbf{P}=\mathbf{I}$ because only there the derivative of the $\mathrm{det}$ function is the trace. $\endgroup$ – Everyday Astronaut May 4 '17 at 8:17

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.