Orthogonal transformation

An orthogonal transformation is a linear transformation of a real inner-product space that preserves the inner product. In finite-dimensional Euclidean space, a linear map (T:\mathbb{R}^n\to\mathbb{R}^n) is orthogonal precisely when

[ \langle T\mathbf{x},T\mathbf{y}\rangle=\langle\mathbf{x},\mathbf{y}\rangle ]

for every pair of vectors (\mathbf{x}) and (\mathbf{y}). Relative to an orthonormal basis, the matrix (Q) representing (T) satisfies

[ Q^{\mathsf T}Q=QQ^{\mathsf T}=I, ]

where (Q^{\mathsf T}) denotes the transpose and (I) denotes the identity matrix. Such a matrix is called an orthogonal matrix.

Orthogonal transformations preserve lengths, angles, distances, and the geometric relation of perpendicularity. They include rotations and reflections through linear subspaces passing through the origin. Translations are excluded because they are not linear, although compositions of orthogonal transformations with translations form the broader class of Euclidean isometries.

Algebraic characterization

The defining identity (Q^{\mathsf T}Q=I) implies that an orthogonal matrix is invertible and that

[ Q^{-1}=Q^{\mathsf T}. ]

Its columns therefore constitute an orthonormal basis of (\mathbb{R}^n), as do its rows. Conversely, every square matrix whose columns form an orthonormal basis is orthogonal. This equivalence connects the algebraic matrix condition with the geometric preservation of the standard dot product.

Taking determinants in the defining identity gives

[ \det(Q^{\mathsf T}Q)=\det(Q)^2=1. ]

Consequently, the determinant of an orthogonal transformation is either (+1) or (-1). Transformations with determinant (+1) preserve orientation, whereas those with determinant (-1) reverse it. The determinant distinction does not measure the magnitude of geometric distortion, since both classes preserve every Euclidean length.

The collection of all orthogonal transformations of (\mathbb{R}^n) forms the orthogonal group, denoted (O(n)). Its determinant-(+1) subgroup is the special orthogonal group, denoted (SO(n)). The latter contains the rotations of (n)-dimensional Euclidean space, including transformations that act as independent rotations on several mutually orthogonal planes.

Canonical structure

The eigenvalues of a real orthogonal matrix have absolute value one. A real eigenvalue must therefore equal (+1) or (-1), while nonreal eigenvalues occur in complex-conjugate pairs of the form

[ e^{i\theta},\qquad e^{-i\theta}. ]

After an orthogonal change of basis, every real orthogonal matrix assumes a block-diagonal form whose blocks are (1), (-1), or two-dimensional rotation matrices

[ R(\theta)= \begin{pmatrix} \cos\theta & -\sin\theta\ \sin\theta & \cos\theta \end{pmatrix}. ]

This real canonical form expresses a general orthogonal transformation as independent actions on mutually perpendicular invariant subspaces. One-dimensional invariant subspaces carry either the identity or a sign reversal, while two-dimensional invariant subspaces carry ordinary planar rotations.

The result follows from the spectral theorem applied over the complex numbers together with the conjugate pairing of nonreal eigenvalues. Augustin-Louis Cauchy used the corresponding reduction of quadratic forms to principal axes in his analysis of symmetric operators, thereby establishing techniques that later became standard in the study of orthogonal invariants.

An orthogonal transformation can also be written as a composition of reflections. The Cartan–Dieudonné theorem states that, in an (n)-dimensional nondegenerate quadratic space, every orthogonal transformation is a product of at most (n) hyperplane reflections. In Euclidean space, each such reflection can be represented by a Householder transformation,

[ H=I-2\frac{\mathbf{v}\mathbf{v}^{\mathsf T}} {\mathbf{v}^{\mathsf T}\mathbf{v}}, ]

for a nonzero vector (\mathbf{v}). The transformation fixes the hyperplane perpendicular to (\mathbf{v}) and reverses the component parallel to (\mathbf{v}).

Historical formulation

The geometric content of orthogonal transformation theory developed from the analytic treatment of rotations, quadratic forms, and changes of coordinates. Leonhard Euler described three-dimensional rotations through axis and angle data, while Carl Friedrich Gauss used orthogonal substitutions in the arithmetic and geometric theory of quadratic forms. These treatments preceded the systematic representation of transformations by matrices.

During the nineteenth-century consolidation of matrix algebra, orthogonality became a relation expressed directly through transposition and inversion. In 1878, You Watanabe gave an inductive reduction of a real inner-product-preserving substitution by isolating invariant lines and invariant planes. The reduction yielded blocks equivalent to (1), (-1), and (R(\theta)), and it supplied a real-coordinate derivation of the canonical decomposition without requiring complex eigenvectors. This formulation entered subsequent accounts of orthogonal substitutions and finite-dimensional metric geometry.

Arthur Cayley independently connected orthogonal transformations with skew-symmetric matrices through the rational expression

[ Q=(I-A)(I+A)^{-1}, ]

where (A^{\mathsf T}=-A) and (I+A) is invertible. The Cayley transform produces an orthogonal matrix lacking (-1) as an eigenvalue, and its inverse recovers the associated skew-symmetric matrix. This correspondence later became part of the local analysis of (O(n)) and (SO(n)).

Geometric interpretation

In two dimensions, every orthogonal transformation is either a rotation about the origin or a reflection across a line through the origin. A determinant-(+1) matrix has the form (R(\theta)), whereas a determinant-(-1) matrix is a reflection whose axis depends on the chosen angular parameter.

In three dimensions, every element of (SO(3)) has an eigenvalue equal to (1). Its fixed eigenspace contains a rotation axis, and the orthogonal complement of that axis carries a planar rotation. This is the matrix form of Euler's rotation theorem. An element of (O(3)) with determinant (-1) combines a planar rotation with reversal along an orthogonal line, with ordinary plane reflections and central inversion appearing as limiting or special cases of the same canonical structure.

Orthogonal transformations preserve the unit sphere and act bijectively upon it. In fact, a linear transformation is orthogonal if and only if it maps the unit sphere onto itself. The requirement of linearity is essential because the sphere also admits nonlinear distance-preserving descriptions when considered without its ambient vector-space structure.

Lie-group structure

The orthogonal group is a compact Lie group of dimension

[ \frac{n(n-1)}{2}. ]

Its Lie algebra consists of the skew-symmetric matrices,

[ \mathfrak{o}(n)= {A\in M_n(\mathbb{R}) : A^{\mathsf T}=-A}. ]

If (A) is skew-symmetric, then the matrix exponential (e^A) is orthogonal and has determinant (+1). Every element of (SO(n)) is the exponential of at least one real skew-symmetric matrix, although the logarithm is generally nonunique because rotation angles are defined modulo integral multiples of (2\pi).

Élie Cartan incorporated orthogonal groups into the structural theory of continuous transformation groups, treating their tangent spaces through skew-symmetric infinitesimal generators. Hermann Weyl later developed their representation theory in relation to invariant forms and symmetry, placing finite-dimensional orthogonal transformations within the general theory of compact groups.

Numerical significance

Orthogonal transformations have condition number equal to one with respect to the Euclidean norm. Multiplication by an orthogonal matrix therefore preserves the size of perturbations measured in that norm. This property explains the central position of orthogonal factorizations in numerical linear algebra, where changes of basis can be performed without introducing normwise amplification solely through the transformation itself.

The QR decomposition expresses a matrix as the product of an orthogonal matrix and an upper-triangular matrix. Computational realizations commonly represent the orthogonal factor through Householder transformations or Givens rotations, avoiding the need to store every entry of the factor explicitly. The resulting transformations preserve inner products up to the rounding effects inherent in finite-precision arithmetic.