Angles between flats


The concept of angles between lines in the plane and between pairs of two lines, two planes or a line and a plane in space can be generalized to arbitrary dimension. This generalization was first discussed by Jordan. For any pair of flats in a Euclidean space of arbitrary dimension one can define a set of mutual angles which are invariant under isometric transformation of the Euclidean space. If the flats do not intersect, their shortest distance is one more invariant. These angles are called canonical or principal. The concept of angles can be generalized to pairs of flats in a finite-dimensional inner product space over the complex numbers.

Jordan's definition

Let and be flats of dimensions and in the -dimensional Euclidean space. By definition, a translation of or does not alter their mutual angles. If and do not intersect, they will do so upon any translation of which maps some point in to some point in. It can therefore be assumed without loss of generality that and intersect.
Jordan shows that Cartesian coordinates in can then be defined such that and are described, respectively, by the sets of equations
and
with. Jordan calls these coordinates canonical. By definition, the angles are the angles between and.
The non-negative integers are constrained by
For these equations to determine the five non-negative integers completely, besides the dimensions and and the number of angles, the non-negative integer must be given. This is the number of coordinates, whose corresponding axes are those lying entirely within both and. The integer is thus the dimension of. The set of angles may be supplemented with angles to indicate that has that dimension.
Jordan's proof applies essentially unaltered when is replaced with the -dimensional inner product space over the complex numbers.

Angles between subspaces

Now let and be subspaces of the -dimensional inner product space over the real or complex numbers. Geometrically, and are flats, so Jordan's definition of mutual angles applies. When for any canonical coordinate the symbol denotes the unit vector of the axis, the vectors form an orthonormal basis for and the vectors form an orthonormal basis for, where
Being related to canonical coordinates, these basic vectors may be called canonical.
When denote the canonical basic vectors for and the canonical basic vectors for then the inner product vanishes for any pair of and except the following ones.
With the above ordering of the basic vectors, the matrix of the inner products is thus diagonal. In other words, if and are arbitrary orthonormal bases in and then the real, orthogonal or unitary transformations from the basis to the basis and from the basis to the basis realize a singular value decomposition of the matrix of inner products. The diagonal matrix elements are the singular values of the latter matrix. By the uniqueness of the singular value decomposition, the vectors are then unique up to a real, orthogonal or unitary transformation among them, and the vectors and are unique up to equal real, orthogonal or unitary transformations applied simultaneously to the sets of the vectors associated with a common value of and to the corresponding sets of vectors .
A singular value can be interpreted as corresponding to the angles introduced above and associated with and a singular value can be interpreted as corresponding to right angles between the orthogonal spaces and, where superscript denotes the orthogonal complement.

Variational characterization

The [|variational characterization] of singular values and vectors implies as a special case a variational characterization of the [|angles between subspaces] and their associated canonical vectors. This characterization includes the angles and introduced above and orders the angles by increasing value. It can be given the form of the below alternative definition. In this context, it is customary to talk of principal angles and vectors.

Definition

Let be an inner product space. Given two subspaces with, there exists then a sequence of angles called the principal angles, the first one defined as
where is the inner product and the induced norm. The vectors and are the corresponding principal vectors.
The other principal angles and vectors are then defined recursively via
This means that the principal angles form a set of minimized angles between the two subspaces, and the principal vectors in each subspace are orthogonal to each other.

Examples

Geometric example

Geometrically, subspaces are flats that include the origin, thus any two subspaces intersect at least in the origin. Two two-dimensional subspaces and generate a set of two angles. In a three-dimensional Euclidean space, the subspaces and are either identical, or their intersection forms a line. In the former case, both. In the latter case, only, where vectors and are on the line of the intersection and have the same direction. The angle will be the angle between the subspaces and in the orthogonal complement to. Imagining the angle between two planes in 3D, one intuitively thinks of the largest angle,.

Algebraic example

In 4-dimensional real coordinate space R4, let the two-dimensional subspace be
spanned by and, and let the two-dimensional subspace be
spanned by and with some real and such that. Then and are, in fact, the pair of principal vectors corresponding to the angle with, and and are the principal vectors corresponding to the angle with
To construct a pair of subspaces with any given set of angles in a dimensional Euclidean space, take a subspace with an orthonormal basis and complete it to an orthonormal basis of the Euclidean space, where. Then, an orthonormal basis of the other subspace is, e.g.,

Basic properties

The notion of the angles and some of the variational properties can be naturally extended to arbitrary inner products and subspaces with infinite dimensions.

Computation

Historically, the principal angles and vectors first appear in the context of canonical correlation and were originally computed using SVD of corresponding covariance matrices. However, as first noticed in, the canonical correlation is related to the cosine of the principal angles, which is ill-conditioned for small angles, leading to very inaccurate computation of highly correlated principal vectors in finite precision computer arithmetic. The sine-based algorithm fixes this issue, but creates a new problem of very inaccurate computation of highly uncorrelated principal vectors, since the sine function is ill-conditioned for angles close to /2. To produce accurate principal vectors in computer arithmetic for the full range of the principal angles, the combined technique first compute all principal angles and vectors using the classical cosine-based approach, and then recomputes the principal angles smaller than /4 and the corresponding principal vectors using the sine-based approach. The combined technique is implemented in open-source libraries Octave and SciPy and contributed and
to MATLAB.