PlanetMath (more info)
 Math for the people, by the people.
Encyclopedia | Requests | Forums | Docs | Wiki | Random | RSS  
Login
create new user
name:
pass:
forget your password?
Main Menu
sections
Encyclopædia
Papers
Books
Expositions

meta
Requests (210)
Orphanage
Unclass'd (1)
Unproven (473)
Corrections (34)
Classification

talkback
Polls
Forums
Feedback
Bug Reports

downloads
Snapshots
PM Book

information
News
Docs
Wiki
ChangeLog
TODO List
Legalese
About
Owner confidence rating: Low Entry average rating: No information on entry rating
[parent] decomposition of orthogonal operators as rotations and reflections (Theorem)
Theorem 1   Let $ V$ be a $ n$-dimensional real inner product space ( $ 0 < n < \infty$). Then every orthogonal operator $ T$ on $ V$ can be decomposed into a series of two-dimensional rotations and one-dimensional reflections on mutually orthogonal subspaces of $ V$.

We first explain the general idea behind the proof. Consider a rotation $ R$ of angle $ \theta$ in a two-dimensional space. From its orthonormal basis representation

$\displaystyle \begin{bmatrix} \cos \theta & -\sin \theta \ \sin \theta & \; \cos \theta \end{bmatrix}\,, $
we find that the characteristic polynomial of $ R$ is $ t^2 - 2 \cos \theta + 1$, with (complex) roots $ e^{\pm i\theta}$. (Not surprising, since multiplication by $ e^{i\theta}$ in the complex plane is rotation by $ \theta$.) Thus given the characteristic polynomial of $ R$, we can almost recover 1 its rotation angle.

In the case of a reflection $ S$, the eigenvalues of $ S$ are $ -1$ and $ 1$; so again, the characteristic polynomial of $ S$ will provide some information on $ S$.

So in $ n$ dimensions, we are also going to look at the complex eigenvalues and eigenvectors of $ T$ to recover information about the rotations represented by $ T$.

But there is one technical point -- $ T$ is a transformation on a real vector space, so it does not really have “complex eigenvalues and eigenvectors”. To make this concept rigorous, we must consider the complexification $ T^\mathbb{C}$ of $ T$, the operator defined by $ T^\mathbb{C}(x+iy) = Tx + iTy$ in the vector space $ V^\mathbb{C}$ consisting of elements of the form $ x+iy$, for $ x, y \in V$. (For more details, see the entry on complexification.)


Lemma 1   For any linear operator $ T\colon V \to V$, there exists a one- or two-dimensional subspace $ W$ which is invariant under $ T$.
Proof. Consider $ T^\mathbb{C}$ and its characteristic polynomial. By the Fundamental Theorem of Algebra, the characteristic polynomial has a complex root $ \lambda = \alpha + i \beta$. Then there is an eigenvector $ x+iy \neq 0$ with eigenvalue $ \lambda$. We have
$\displaystyle Tx + iTy = T^\mathbb{C}(x+iy) = \lambda(x + iy) = (\alpha x - \beta y) + i(\beta x + \alpha y)\,. $
Equating real and imaginary components, we see that
$\displaystyle Tx$ $\displaystyle = \alpha x - \beta y \in W$    
$\displaystyle Ty$ $\displaystyle = \beta x + \alpha y \in W$    

where $ W = {\mathrm{span}}\{ x, y\}$. $ W$ is two-dimensional if $ x$ and $ y$ are linearly independent; otherwise it is one-dimensional 2. And we have $ T(W) \subseteq W$ as claimed. $ \qedsymbol$


Proof. [Proof of Theorem 1] We recursively factor $ T$; formally, the proof will be by induction on the dimension $ n$.

The case $ n=1$ is trivial. We have $ \det T = \pm 1$; if $ \det T = 1$ then $ T$ is the identity; otherwise $ Tx = -x$ is a reflection.

For larger $ n$, by Lemma 1 there exists a $ T$-invariant subspace $ W$. The orthogonal complement $ W^\perp$ of $ W$ is $ T$-invariant also, because for all $ x \in W^\perp$ and $ y \in W$,

$\displaystyle \langle Tx, y \rangle$ $\displaystyle = \langle x, T^{-1} y \rangle$ $\displaystyle \textrm{because $T$\ preserves inner product}$    
  $\displaystyle = 0$ $\displaystyle \textrm{because $T^{-1}(W) = W$.}$    

Let $ T_W$ be the operator that acts as $ T$ on $ W$ and is the identity on $ W^\perp$. Similarly, let $ T_{W^\perp}$ be the operator that acts as $ T$ on $ W^{\perp}$ and is the identity on $ W$. Then $ T = T_W \circ T_{W^\perp}$. $ T_W$ restricted to $ W$ is orthogonal, and since $ W$ is one- or two-dimensional, $ T_W$ must therefore be a rotation or reflection (or the identity) in a line or plane.

$ T_{W^\perp}$ restricted to $ W^\perp$ is also orthogonal. $ W^\perp$ has dimension $ <n$, so by the induction hypothesis we can continue to factor it into operators acting on subspaces of $ W^\perp$ that are mutually orthogonal. These subspaces will of course also be orthogonal to $ W$.

The proof is now complete, except that we did not rule out $ T_W$ being a reflection even when $ W$ is two-dimensional. But of course, if $ T_W$ is a reflection in two dimensions, then it can be factored as a reflection on a one-dimensional subspace $ {\mathrm{span}}\{v\}$ composed with the identity on $ {\mathrm{span}}\{ v \}^\perp$. $ \qedsymbol$

Actually, in the third paragraph of the proof, we implicitly assumed that every orthogonal operator on two-dimensional is either a rotation or reflection. This is a well-known fact, but it does need to be formally proven:

Theorem 2   If $ V$ is a two-dimensional real inner product space, then every orthogonal operator $ T$ is either a rotation or reflection.
Proof. Fix any orthonormal basis $ \{ e_1, e_2 \}$ for $ V$. Since $ T$ is orthogonal, $ \lVert Te_1\rVert = \lVert e_1\rVert = 1$, i.e. $ Te_1$ is a unit vector on the plane, so there exists an angle $ \theta$ (unique modulo $ 2\pi$) such that $ Te_1 = \cos \theta \, e_1 + \sin \theta \, e_2$. Similarly $ Te_2$ is a unit vector, but since $ e_1$ and $ e_2$ are orthogonal, so are $ Te_1$ and $ Te_2$. Then it is found that the solution to $ Te_2$ must be either $ -\sin \theta \, e_1 + \cos \theta \, e_2$ or $ \sin \theta \, e_1 - \cos \theta \, e_2$. Putting all this together, the matrix for $ T$ is:
$\displaystyle \begin{bmatrix} \cos \theta & \mp\sin \theta \ \sin \theta & \pm \cos \theta \end{bmatrix}\,. $
The first solution for $ Te_2$ corresponds to a rotation matrix (and $ \det T = 1$); the second solution for $ Te_2$ corresponds to a reflection matrix (and $ \det T = -1$). $ \qedsymbol$


Remarks

i.
Observe that the two equations for $ Tx$ and $ Ty$ appearing in the proof of Lemma 1 do specify a rotation of angle $ \pm \theta$ when $ \lambda = \alpha + i\beta = e^{i\theta}$ and $ x, y$ are orthonormal. So by examining the complex eigenvales and eigenvectors of $ T^\mathbb{C}$, we can reconstruct the rotation.

This construction can be used to give an alternate proof of Theorem 1; we sketch it below:

The complexified space $ V^\mathbb{C}$ has an inner product structure inherited from $ V$ (again, see complexification for details). Let $ U = T^\mathbb{C}$. Since $ T$ is orthogonal, $ U$ is unitary, and hence normal ( $ U^* U = UU^*$). There exists an orthonormal basis of eigenvectors for $ U$ (the Schur decomposition). Let $ z = x+iy$ is any one of these eigenvectors with a complex, non-real eigenvalue $ \lambda$. Then $ \lvert\lambda\rvert = 1$ because $ U$ is unitary, and $ \overline{z} = x-iy$ is another eigenvector with eigenvalue $ \overline{\lambda}$. Using the $ V^\mathbb{C}$ inner product formula, the vectors $ x/\sqrt{2}$ and $ y/\sqrt{2}$ can be shown to be orthonormal. Then the proof of Lemma 1 shows that $ T$ acts as a rotation in the plane $ {\mathrm{span}}\{ x, y \}$. All such planes obtained will be orthogonal to each other.

To summarize, the orthogonal subplanes of rotation are found by grouping conjugate pairs of complex eigenvectors. If one actually needs to determine the planes of rotation explicitly (for dimensions $ n \geq 4$), then probably it is better to work directly with the complexified matrix, rather than to factor the matrix over the reals.

ii.
The decomposition of $ T$ is not unique. However, it is always possible to obtain a decomposition which contains at most one reflection, because any two single-dimensional reflections can always be combined into a two-dimensional rotation. In any case, the parity of the number of reflections in a decomposition of $ T$ is invariant, because the parity is equal to $ \det T$.
iii.
In the decomposition, the component rotations and reflections all commute because they act on orthogonal subspaces.
iv.
If we take a basis for $ V$ describing the mutually orthogonal subspaces in Theorem 1, the matrix for $ T$ looks like:
$\displaystyle \begin{bmatrix} \cos \theta_1 & -\sin \theta_1 \ \sin \theta_1 ... ... & & & & & & +1 \ & & & & & & & \ddots \ & & & & & & & & +1 \end{bmatrix}$
where $ \theta_1, \dotsc, \theta_k$ are the rotation angles, one for each orthogonal subplane, and the $ \pm 1$ in the middle is the reflective component (if present). The rest of the entries in the matrix are zero.
v.
Sometimes any orthogonal operator $ T$ with $ \det T = 1$ is called a rotation, even though strictly speaking it is actually series of rotations (each on different “axes”). Similarly, when $ \det T = -1$, $ T$ may be called a reflection, even though again it is not always a single (one-dimensional) reflection.

In this language, a rotation composed with a rotation will always be a rotation; a rotation composed with a reflection is a reflection; and two reflections composed together will always be a rotation.

vi.
In $ \mathbb{R}^3$, an orthogonal operator with positive determinant is necessarily a rotation on one axis which is left fixed (except when the operator is the identity). This follows simply because there is no way to fit more than one orthogonal subplane into three-dimensional space.

A composition of two rotations in $ \mathbb{R}^3$ would then be a rotation too. On the other hand it is not at all obvious what relation the axis of rotation of the composition has with the original two axes of rotation.

For an explicit formula for a rotation matrix in $ \mathbb{R}^3$ that does not require manual calculation of the basis vectors for the rotation subplane, see Rodrigues' rotation formula.

vii.
In $ \mathbb{R}^n$, reflections can be carried out by first embedding $ \mathbb{R}^n$ into $ \mathbb{R}^{n+1}$ and then rotating $ \mathbb{R}^{n+1}$. (Here, the words “rotation” and “reflection” being taken in their extended sense of (v).) For example, in the plane, a right hand can be rotated in $ \mathbb{R}^3$ into a left hand.

To be specific, suppose we embed $ \mathbb{R}^n$ in $ \mathbb{R}^{n+1}$ as the first $ n$ coordinates. Then we gain an extra degree of freedom in the last coordinate of $ \mathbb{R}^{n+1}$ (with coordinate vector $ e_{n+1}$). Given an orthogonal operator $ T\colon \mathbb{R}^n \to \mathbb{R}^n$, we can extend it to an operator on $ T'\colon \mathbb{R}^{n+1} \to \mathbb{R}^{n+1}$ by having it act as $ T$ on the lower $ n$ coordinates, and setting $ T'(e_{n+1}) = -e_{n+1}$. Since $ \det T' = -\det T = 1$, our new $ T'$ will be a rotation (the extra angle of rotation will be by $ \pi$) that reflects sets in the $ \mathbb{R}^n$ plane.

Bibliography

1
Friedberg, Insel, Spence. Linear Algebra. Prentice-Hall, 1997.
2
Vladimir I. Arnol'd (trans. Roger Cooke). Ordinary Differential Equations. Springer-Verlag, 1992.


Footnotes

... recover1
Because the complex roots occur in conjugate pairs, the information about the sign of $ \theta$ is lost. This too, is not a surprise, because the sign of $ \theta$, i.e. whether the rotation is clockwise or counterclockwise, is dependent on the orientation of the basis vectors.
... one-dimensional2
In fact, the space $ W$ constructed is two-dimensional if and only if the eigenvalue $ \lambda$ is not purely real. Compare with the remark (i) after the proof of Theorem 1.

Actually, Lemma 1 has more uses than just proving Theorem 1. For example, if $ \dot{x} = Ax$ is a linear differential equation, and the constant coefficient matrix $ A$ has only simple eigenvalues, then it is a consequence of Lemma 1, that the differential equation decomposes into a series of disjoint one-variable and two-variable equations. The solutions are then readily understood: they are always of the form of an exponential multiplied by a sinusoid, and linear combinations thereof. The sinusoids will be present whenever there are non-real eigenvalues.



"decomposition of orthogonal operators as rotations and reflections" is owned by stevecheng.
(view preamble)

View style:

See Also: rotation matrix, orthogonal matrices, dimension of the special orthogonal group, Rodrigues' rotation formula, derivation of rotation matrix using polar coordinates, derivation of 2D reflection matrix

Keywords:  rotation, reflection, orthogonal

This object's parent.
Log in to rate this entry.
(view current ratings)

Cross-references: reflects, angle of rotation, degree of freedom, coordinates, right, embedding, Rodrigues rotation formula, relation, obvious, composition, axis, determinant, positive, language, strictly, act on, parity, contains, decomposition, normal, unitary, structure, inner product, orthonormal, rotation matrix, unit vector, fix, even, complete, induction hypothesis, plane, line, restricted, orthogonal complement, identity, induction, factor, linear combinations, sinusoid, exponential, solutions, equations, disjoint, differential equation, consequence, simple, matrix, coefficient, linear differential equation, linearly independent, components, imaginary, eigenvalue, eigenvector, fundamental theorem of algebra, invariant, linear operator, complexification, vector space, transformation, point, dimensions, eigenvalues, vectors, basis, orientation, information, conjugate, occur in, complex roots, complex plane, multiplication, roots, complex, characteristic polynomial, representation, orthonormal basis, angle, subspaces, reflections, rotations, series, operator, orthogonal, inner product space, real
There is 1 reference to this entry.

This is version 12 of decomposition of orthogonal operators as rotations and reflections, born on 2005-07-19, modified 2006-06-16.
Object id is 7242, canonical name is DecompositionOfOrthogonalOperatorsAsRotationsAndReflections.
Accessed 3116 times total.

Classification:
AMS MSC: 15A04 (Linear and multilinear algebra; matrix theory :: Linear transformations, semilinear transformations)
Pending Errata and Addenda
None.
[ View all 2 ]
Discussion
Style: Expand: Order:
forum policy
Rotations in n-dimensional space by fourdplus on 2007-07-24 23:33:45
RE: decomposition of orthogonal operators as rotations and reflections
BY: stevecheng.


The article by stevecheng answers a question that I have encountered: Given a rigid n-dimensional body in an arbitrary orientation, how to transform it to another arbitrary orientation by rotations. The answer is: Find the matrix that will do the job (I know the start and end orientations, so that is straightforward), then decompose that matrix to find the rotation planes and the angles of rotation within those planes. Stevecheng shows how.


Here is a supplementary question, I would be grateful if someone could tell me the answer. If we do the above calculation and obtain a set of rotation angles, we can add the absolute values of the angles and arrive at a total angle of rotation. Is this total angle of rotation the minimum possible to achieve the transformation? Obviously there are an infinite number of ways the transformation could be achieved, including using different planes of rotation. It seems intuitive to me that the method of the article gives the best solution, but is this so?

If the answer is yes, I can use the total angle of rotation as a measure of the 'distance' between two arbitrary orientations.

Thanks in advance for any contribution.
[ reply | up ]
Interact
post | correct | update request | prove | add result | add corollary | add example | add (any)