The Two Components of $\mathrm{GL}_n(\mathbb{R})$ and the Failure of the Exponential
Introduction
The group of invertibles of $M_n(\mathbb{R})$ has exactly two connected components, cut by the sign of the determinant, and the exponential does not reach all of the positive one: the matrix with both diagonal entries $-1$ and a $1$ above the diagonal has positive determinant and is not an exponential.
The article proves that the determinant cuts $\mathrm{GL}_n(\mathbb{R})$ into exactly two components, that the polar decomposition joins every matrix of positive determinant to the identity, and that the exponential image is a proper part of that component. It proves the counterexample in full: the centraliser of the matrix is two-dimensional, the exponential of every element of that centraliser has a positive diagonal, and the matrix has a negative one. It replaces the false guess that the image is dense by the parity obstruction. It assumes The Exponential of a Real Matrix and Its Image, and defers the connectedness of $\mathrm{SO}(n)$ to The Exponential of the Antisymmetric Part: One-Parameter Rotations.
Throughout $n\ge2$, the transpose is ${}^{\mathsf T}$ and the identity is $\mathrm{I}$.
The Determinant Cuts the Group
The determinant is a continuous homomorphism onto $\mathbb{R}^{\times}$. It is a polynomial in the entries, so it is continuous, and $\det(AB)=\det A\det B$, so it is a homomorphism from the group to the multiplicative group of the field. It is onto, because $\det\operatorname{diag}(t,1,\dots,1)=t$ for every $t\ne0$.
Definition. Put
$$ \mathrm{GL}_n(\mathbb{R})^{+}=\{A:\det A>0\},\qquad \mathrm{GL}_n(\mathbb{R})^{-}=\{A:\det A<0\}. $$
These are the preimages of the two components of $\mathbb{R}^{\times}$, they are open, and they are exchanged by multiplication by $-\mathrm{I}$.
The sign of the determinant is locally constant. A continuous function on a connected set has connected image, and the image of $\det$ on a connected set is a connected subset of $\mathbb{R}^{\times}$, that is, a single point of $(0,\infty)$ or of $(-\infty,0)$.
At Least Two Components
Two matrices of opposite determinant sign lie in different components. Since $\det$ is locally constant, a component lies entirely in one of the two halves. The identity has determinant $1$ and $-\mathrm{I}$ has determinant $(-1)^{n}$, so for odd $n$ the two halves are already separated by the identity and its negative, and for even $n$ the matrix $\operatorname{diag}(-1,1,\dots,1)$ separates them. So there are at least two components, and each half is a union of components.
The count is the whole content. The hard part is not the lower bound; it is the statement that each half is connected. That is the next two sections.
The Polar Decomposition
Every invertible matrix is a positive definite matrix times an orthogonal one. Let $A\in\mathrm{GL}_n(\mathbb{R})$. The matrix $AA^{\mathsf T}$ is symmetric and positive definite, so it has a unique positive definite square root $P=(AA^{\mathsf T})^{1/2}$, by the spectral theorem for $\operatorname{Sym}_n$. Then
$$ A=PQ,\qquad Q=P^{-1}A, $$
and $Q$ is orthogonal, because
$$ Q^{\mathsf T}Q=A^{\mathsf T}P^{-2}A=A^{\mathsf T}(AA^{\mathsf T})^{-1}A=\mathrm{I}. $$
The factor $P$ is symmetric positive definite and the factor $Q$ is orthogonal. The decomposition is unique: $P$ is forced to be $(AA^{\mathsf T})^{1/2}$ and $Q$ is then forced.
The sign of the determinant lives in the orthogonal factor. Since $\det P>0$,
$$ \det Q=\frac{\det A}{\det P}, $$
so $Q\in\mathrm{SO}(n)$ exactly when $\det A>0$. On the positive half the polar decomposition is a decomposition into a positive definite stretch and a rotation.
At Most Two Components
On the positive half the exponential of the symmetric part does the stretching. A positive definite $P$ is $\exp S$ for the symmetric $S=\log P$, by the bijection of The Exponential of a Real Matrix and Its Image. Then
$$ A=\exp(S)\,Q,\qquad S\in\operatorname{Sym}_n,\quad Q\in\mathrm{SO}(n), $$
and therefore
$$ \mathrm{GL}_n(\mathbb{R})^{+}=\exp(\operatorname{Sym}_n)\cdot\mathrm{SO}(n). $$
Both factors are connected. The space $\operatorname{Sym}_n$ is a real vector space, hence connected, and $\exp$ is continuous, so $\exp(\operatorname{Sym}_n)=\operatorname{Sym}_n^{++}$ is connected. The group $\mathrm{SO}(n)$ is connected: every rotation is the exponential of an antisymmetric matrix, so it is the continuous image of the connected space $\operatorname{Skew}_n$. The product of two connected spaces is connected, and the multiplication map is continuous.
Every matrix of positive determinant is joined to the identity. Let $A=\exp(S)Q$ and let $t\mapsto Q_{t}$ be a path in $\mathrm{SO}(n)$ from $Q$ to $\mathrm{I}$. Then
$$ t\longmapsto \exp(tS)\,Q_{t} $$
is a path in $\mathrm{GL}_n(\mathbb{R})^{+}$ from $A$ to $\mathrm{I}$. So the positive half is path-connected, hence connected.
The negative half is a translate. Multiplication by $-\mathrm{I}$ is a homeomorphism carrying $\mathrm{GL}_n(\mathbb{R})^{+}$ onto $\mathrm{GL}_n(\mathbb{R})^{-}$, so the negative half is path-connected too. The group has exactly two components, and both are path-connected.
The two components are not the two halves of the transpose. The symmetric and the antisymmetric matrices both meet each half, since $\mathrm{I}$ and $-\mathrm{I}$ are symmetric. The cut by the sign of the determinant is topological.
The Counterexample
A matrix of positive determinant that is not an exponential. Put
$$ A=\begin{pmatrix}-1&1\\0&-1\end{pmatrix}. $$
Then $\det A=1>0$ and $\operatorname{tr}A=-2$. The matrix lies in $\mathrm{GL}_2(\mathbb{R})^{+}$, and it is not $\exp X$ for any real $X$.
Every candidate commutes with $A$. Suppose that $\exp X=A$ for some $X\in M_2(\mathbb{R})$. The matrix $X$ commutes with its own exponential, because $X$ commutes with every power of $X$ and the series converges:
$$ XA=X\exp X=\exp X\,X=AX . $$
So the problem is confined to the centraliser of $A$.
The Centraliser
The centraliser is two-dimensional. Let $X=\begin{pmatrix}a&b\\c&d\end{pmatrix}$. The condition $XA=AX$ reads
$$ \begin{pmatrix}-a&a-b\\-c&c-d\end{pmatrix}=\begin{pmatrix}-a+c&d-b\\-c&-d\end{pmatrix}, $$
and comparing entries gives $c=0$ and $d=a$. So
$$ X=a\mathrm{I}+bN,\qquad N=\begin{pmatrix}0&1\\0&0\end{pmatrix},\qquad N^{2}=0 . $$
Conversely every matrix of that form commutes with $A$, so the centraliser is the two-dimensional space $\mathbb{R}\mathrm{I}\oplus\mathbb{R}N$. A numerical check of the linear system $XA-AX=0$ in the four entries returns rank two and the same two-dimensional kernel.
The centraliser is the algebra generated by $\mathrm{I}$ and a square-zero matrix. The two generators commute, $\mathrm{I}$ with everything and $N$ with itself, so the exponential factorises on the centraliser.
The Exponential on the Centraliser
The exponential of a centraliser element is the scalar exponential times a unipotent factor. Since $a\mathrm{I}$ and $bN$ commute and $N^{2}=0$,
$$ \exp(a\mathrm{I}+bN)=e^{a}\big(\mathrm{I}+bN\big)=\begin{pmatrix}e^{a}&e^{a}b\\0&e^{a}\end{pmatrix}. $$
The formula is checked numerically on $20$ random pairs $(a,b)$ with maximum error $3.8\cdot10^{-15}$.
No value of the parameters gives $A$. The two diagonal entries of $\exp(a\mathrm{I}+bN)$ are equal to $e^{a}$, while those of $A$ are $-1$. So equality would require
$$ e^{a}=-1, $$
which has no real solution, because $e^{a}>0$ for every real $a$. Equivalently, $\det\exp(a\mathrm{I}+bN)=e^{2a}>0$ is consistent with $\det A=1$, and the trace $2e^{a}>0$ is not consistent with $\operatorname{tr}A=-2$.
The obstruction survives the whole centraliser, not just the diagonal. Both diagonal entries are $e^{a}$, so the matrix $\exp X$ is always a scalar multiple of a unipotent matrix, and its diagonal is positive. The matrix $A$ is a negative scalar multiple of a unipotent matrix. The two classes are disjoint.
The Numerical Check
A hundred samples of the centraliser never reach $A$. With $a$ and $b$ drawn independently from the normal law with standard deviation $2$, and $X=a\mathrm{I}+bN$, the smallest value of $\lVert\exp X-A\rVert$ over the sample is
$$ 1.26 . $$
The distance is bounded below, as it must be: the quantity $e^{a}$ is positive, and the diagonal entry $-1$ is at distance $1+e^{a}\ge1$ from it.
The exponent of the series is not the issue. The truncated sums converge to $\exp X$ to machine precision, so the failure is not numerical. It is the sign of the diagonal: no truncation of a series of matrices with a positive diagonal converges to a matrix with a negative one.
What the Image Is
The exponential is not surjective onto the positive component. The counterexample gives $\exp(M_2(\mathbb{R}))\ne\mathrm{GL}_2(\mathbb{R})^{+}$, and the block embedding $X\mapsto\operatorname{diag}(X,1,\dots,1)$ carries the same failure into every dimension $n\ge2$.
The image is not open and not dense. The matrices with two simple negative eigenvalues form an open set, they have no real logarithm by the parity argument of The Exponential of a Real Matrix and Its Image, and $\operatorname{diag}(-1,-2)$ is an interior point of that set. So an open set misses the image, and the image is neither dense nor the whole component. The claim that the image is dense in $\mathrm{GL}_n(\mathbb{R})^{+}$ is false, and the example at $n=2$ refutes it.
The exact criterion is Culver's. A real invertible matrix $A$ is $\exp X$ for a real $X$ if and only if, for every negative real eigenvalue $\lambda$ of $A$ and every size $d$, the number of Jordan blocks of size $d$ for $\lambda$ is even. For the counterexample the eigenvalue $-1$ carries a single block of size $2$, an odd number, and the criterion fails. The parity of the multiplicity alone is not enough: the multiplicity of $-1$ is two, and the block count is one.
The image sits strictly between two extremes. It contains a neighbourhood of $\mathrm{I}$, because the exponential is a local diffeomorphism at the origin, and it misses an open set, so it is neither a neighbourhood of the whole component nor a thin subset of it.
What the two components say about the exponential. The exponential is a map whose image lies in a connected component, misses part of it, and generates it as a group. That is the sharp picture: the local statement is true, the global one is false, and the two components are the exact arena in which the difference is visible.
Summary
The determinant is a continuous surjection of $\mathrm{GL}_n(\mathbb{R})$ onto $\mathbb{R}^{\times}$, so the group has at least two components, one of positive and one of negative determinant. The polar decomposition $A=\exp(S)Q$ with $S$ symmetric and $Q$ orthogonal makes both halves connected, so the components are exactly two. The matrix $\begin{pmatrix}-1&1\\0&-1\end{pmatrix}$ has determinant $1$ and is not an exponential: any $X$ with $\exp X=A$ commutes with $A$, the centraliser is $\{a\mathrm{I}+bN\}$ with $N^{2}=0$, and $\exp(a\mathrm{I}+bN)=e^{a}(\mathrm{I}+bN)$ has positive diagonal entries while $A$ has $-1$ there. Over $100$ samples of the centraliser the distance to $A$ never falls below $1.26$. So the exponential misses part of the positive component, is not dense there, and its exact image is described by Culver's criterion on the Jordan blocks at the negative eigenvalues.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $\mathrm{GL}_n(\mathbb{R})^{+}$, $\mathrm{GL}_n(\mathbb{R})^{-}$ | the matrices of positive and of negative determinant |
| $\det$ | the determinant homomorphism |
| $P=(AA^{\mathsf T})^{1/2}$ | the positive definite polar factor |
| $Q=P^{-1}A$ | the orthogonal polar factor |
| $A$ | the counterexample $\begin{pmatrix}-1&1\\0&-1\end{pmatrix}$ |
| $N$ | the nilpotent $\begin{pmatrix}0&1\\0&0\end{pmatrix}$ |
| $\operatorname{Sym}_n$ | the symmetric matrices |
| $\mathrm{SO}(n)$ | the rotations, of determinant one |
Further Reading
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for the polar decomposition and the sign of the determinant.
- W. J. Culver, On the existence and uniqueness of the real logarithm of a matrix, Proceedings of the American Mathematical Society (1966), for the criterion of §What the Image Is.
- The Exponential of a Real Matrix and Its Image, for the series, the determinant identity and the local diffeomorphism.
- The Exponential of the Antisymmetric Part: One-Parameter Rotations, for the connectedness of $\mathrm{SO}(n)$.
- The Matrix Algebra $M_n(\mathbb{R})$, for the group of invertibles and the singular set.
- The Minimal Polynomial over $\mathbb{R}$ and the Real Jordan Form, for the Jordan blocks at a negative eigenvalue.