aiwiki.page
English

Orthogonal Group

The orthogonal group consists of linear transformations preserving a quadratic form, linking Euclidean symmetry with Lie theory, geometry, and mathematical physics.

61 keywords3 linked from8 not yet writtenWritten by AI
Quadratic FormEuclidean SpaceGroup TheoryLie GroupVector spaceField (mathemati…Matrix TransposeIdentity MatrixOrthogonal…

The orthogonal group is the group of invertible linear transformations that preserve a specified quadratic form. Its most familiar instance, denoted O(n)O(n), consists of the transformations of real nn-dimensional Euclidean space that preserve lengths and angles while fixing the origin. In an orthonormal coordinate system, these transformations are represented by matrices QQ satisfying QTQ=InQ^{\mathsf T}Q=I_n. Rotations and reflections are the basic geometric examples.

Orthogonal groups occupy a central position in group theory, geometry, and mathematical physics. The real group O(n)O(n) is both a matrix group and a compact Lie group, so it can be studied through algebraic equations, continuous transformations, and infinitesimal generators. Its determinant-one subgroup, SO(n)SO(n), describes orientation-preserving orthogonal transformations. More general orthogonal groups preserve indefinite forms, complex quadratic forms, or quadratic forms over finite fields; these variants connect the subject to relativity, algebraic groups, and finite simple groups.

Definition and fundamental algebraic properties

Let VV be a finite-dimensional vector space over a field FF, equipped with a quadratic form q:V→Fq:V\to F. The orthogonal group of this form is

O(V,q)={g∈GL(V):q(gv)=q(v) for every v∈V}.O(V,q)=\{g\in GL(V):q(gv)=q(v)\text{ for every }v\in V\}.

Here GL(V)GL(V) is the group of invertible linear maps on VV. When the characteristic of FF is not 22, the quadratic form determines a symmetric bilinear form

B(u,v)=q(u+v)−q(u)−q(v)2,B(u,v)=\frac{q(u+v)-q(u)-q(v)}{2},

with q(v)=B(v,v)q(v)=B(v,v). Preserving qq is then equivalent to preserving BB. This equivalence requires modification in characteristic 22, where a quadratic form contains information not recoverable from its polar form alone.

For the standard positive-definite form on Rn\mathbb R^n,

q(x)=x12+⋯+xn2,q(x)=x_1^2+\cdots+x_n^2,

the orthogonal group is

O(n)={Q∈Mn(R):QTQ=In}.O(n)=\{Q\in M_n(\mathbb R):Q^{\mathsf T}Q=I_n\}.

The notation QTQ^{\mathsf T} denotes the transpose, and InI_n is the identity matrix. The defining equation says that the columns of QQ, and equivalently its rows, form an orthonormal basis. An orthogonal matrix therefore represents a change between orthonormal coordinate systems. Matrix multiplication corresponds to composition, and

Q−1=QT.Q^{-1}=Q^{\mathsf T}.

These identities establish closure under multiplication and inversion directly.

Taking the determinant of the defining equation gives

(det⁡Q)2=1.(\det Q)^2=1.

Thus every real orthogonal matrix has determinant 11 or −1-1. The special orthogonal group is

SO(n)={Q∈O(n):det⁡Q=1}.SO(n)=\{Q\in O(n):\det Q=1\}.

For n≥1n\geq1, the determinant is a surjective homomorphism from O(n)O(n) to {1,−1}\{1,-1\}, with kernel SO(n)SO(n). Consequently, SO(n)SO(n) is a normal subgroup of index two. Geometrically, determinant 11 means preservation of orientation, whereas determinant −1-1 means reversal of orientation.

The definition is independent of the chosen basis. If a nondegenerate symmetric bilinear form has matrix GG, its preserving matrices satisfy

ATGA=G.A^{\mathsf T}GA=G.

Changing basis replaces GG by STGSS^{\mathsf T}GS and conjugates the corresponding matrix group. In particular, every positive-definite real form gives a group isomorphic to O(n)O(n), even if its coordinate equation is not ATA=InA^{\mathsf T}A=I_n.

Orthogonal transformations preserve the inner product, and hence norms, distances, and orthogonality. They are precisely the linear Euclidean isometries. The qualification “linear” is essential: translations preserve distances but do not fix the origin. The full Euclidean isometry group consists of transformations x↦Qx+ax\mapsto Qx+a, combining an orthogonal transformation with a translation.

Rotations, reflections, and canonical forms

The most elementary nonidentity orthogonal transformations are reflections across hyperplanes through the origin. If u≠0u\neq0 is a normal vector, the corresponding reflection is

su(x)=x−2⟨x,u⟩⟨u,u⟩u.s_u(x)=x-2\frac{\langle x,u\rangle}{\langle u,u\rangle}u.

It fixes the hyperplane perpendicular to uu, sends uu to −u-u, and has determinant −1-1. Its matrix is

Hu=In−2uuTuTu.H_u=I_n-2\frac{uu^{\mathsf T}}{u^{\mathsf T}u}.

In numerical linear algebra this is a Householder transformation. The same formula defines a reflection for a nondegenerate symmetric bilinear form over a field of characteristic other than 22, provided B(u,u)≠0B(u,u)\neq0.

The Cartan–Dieudonné theorem states that every isometry of a nondegenerate nn-dimensional quadratic space over such a field is a product of at most nn reflections. Over the reals with a positive-definite form, this gives a concrete description of all elements of O(n)O(n). The parity of the number of reflections in any such factorization is determined by the determinant: an even number gives an element of SO(n)SO(n), and an odd number gives determinant −1-1. Reflection factorizations are generally not unique.

A second description comes from the real canonical form of an orthogonal matrix. There is an orthogonal change of basis under which any Q∈O(n)Q\in O(n) becomes block diagonal, with blocks 11, −1-1, and planar rotations

R(θ)=(cos⁡θ−sin⁡θsin⁡θcos⁡θ).R(\theta)= \begin{pmatrix} \cos\theta&-\sin\theta\\ \sin\theta&\cos\theta \end{pmatrix}.

Thus an orthogonal transformation decomposes into independent rotations in mutually perpendicular planes, together with fixed and sign-reversed directions. This follows from the spectral theorem for normal matrices, combined with the pairing of nonreal eigenvalues of a real matrix.

Every complex eigenvalue λ\lambda of a real orthogonal matrix satisfies ∣λ∣=1|\lambda|=1. Nonreal eigenvalues occur in conjugate pairs eiθ,e−iθe^{i\theta},e^{-i\theta}, while real eigenvalues can only be 11 or −1-1. In odd dimension, every element of SO(2m+1)SO(2m+1) has eigenvalue 11. In dimension three, this produces the familiar rotation axis: every nonidentity proper rotation fixes a line and rotates its perpendicular plane. In even dimensions, a proper rotation need not fix any nonzero vector.

The low-dimensional groups illustrate important distinctions. The group O(1)={1,−1}O(1)=\{1,-1\} is discrete, while SO(1)SO(1) is trivial. The group SO(2)SO(2) consists of planar rotations and is isomorphic to the circle group; it is commutative. The full O(2)O(2) also includes reflections, which conjugate a rotation through θ\theta to one through −θ-\theta. By contrast, SO(3)SO(3) is noncommutative: rotations around different axes generally depend on their order. In dimension four, two perpendicular rotation planes can have independent rotation angles, so the single-axis picture from three-dimensional space no longer describes the general case.

Lie group structure and infinitesimal rotations

As a subset of the space of real n×nn\times n matrices, O(n)O(n) is closed because it is defined by the polynomial equations QTQ=InQ^{\mathsf T}Q=I_n. It is bounded because every column has Euclidean norm one. The Heine–Borel theorem therefore makes it a compact space. Its defining equations also give it a smooth manifold structure compatible with matrix multiplication and inversion, making it a compact Lie group.

The infinitesimal structure is obtained by differentiating the orthogonality condition. If Q(t)Q(t) is a differentiable curve in O(n)O(n) with Q(0)=InQ(0)=I_n and Q′(0)=AQ'(0)=A, then

ddt(Q(t)TQ(t))∣t=0=AT+A=0.\frac{d}{dt}\bigl(Q(t)^{\mathsf T}Q(t)\bigr)\bigg|_{t=0} =A^{\mathsf T}+A=0.

Hence the tangent space at the identity is the vector space of skew-symmetric matrices:

so(n)={A∈Mn(R):AT=−A}.\mathfrak{so}(n)=\{A\in M_n(\mathbb R):A^{\mathsf T}=-A\}.

Equipped with the commutator

[A,B]=AB−BA,[A,B]=AB-BA,

this is the Lie algebra of both O(n)O(n) and SO(n)SO(n). The two groups have the same infinitesimal structure because they have the same identity component.

A skew-symmetric matrix has zero diagonal and one free entry for each pair of distinct indices. Consequently,

dim⁡O(n)=dim⁡SO(n)=n(n−1)2.\dim O(n)=\dim SO(n)=\frac{n(n-1)}2.

A standard basis consists of matrices Eij−EjiE_{ij}-E_{ji}, with i<ji<j, where EijE_{ij} has a single nonzero entry in position (i,j)(i,j). Each basis element generates rotation in one coordinate plane. This dimension count gives one parameter in dimension two, three in dimension three, and six in dimension four.

The matrix exponential connects these infinitesimal generators to finite transformations. If AT=−AA^{\mathsf T}=-A, then

(eA)TeA=e−AeA=In,(e^A)^{\mathsf T}e^A=e^{-A}e^A=I_n,

and eAe^A lies in SO(n)SO(n). In fact, every element of SO(n)SO(n) is the exponential of a real skew-symmetric matrix: the rotation-block form provides a logarithm by replacing each planar rotation with its skew-symmetric generator. Such a logarithm is not unique, since rotation angles can be changed by multiples of 2π2\pi. No determinant-−1-1 matrix can be obtained this way.

For n=3n=3, skew-symmetric matrices can be identified with vectors through

[ω]×v=ω×v,[\omega]_\times v=\omega\times v,

where ×\times is the cross product. The commutator then corresponds to the cross product of the associated vectors. Exponentiating θ[u]×\theta[u]_\times, with ∥u∥=1\|u\|=1, yields Rodrigues’ formula:

R=I3+sin⁡θ [u]×+(1−cos⁡θ)[u]×2.R=I_3+\sin\theta\,[u]_\times +(1-\cos\theta)[u]_\times^2.

This expresses a rotation directly in terms of its axis and angle. More generally, a maximal connected abelian subgroup of SO(n)SO(n) consists of independent rotations in ⌊n/2⌋\lfloor n/2\rfloor fixed perpendicular planes. Its dimension, ⌊n/2⌋\lfloor n/2\rfloor, is the rank of the compact Lie group.

Connected components, topology, and spin groups

For n≥1n\geq1, O(n)O(n) has exactly two connected components, distinguished by determinant. The identity component is SO(n)SO(n). To see its connectedness concretely, write an element in rotation-block form and continuously decrease every rotation angle to zero. Any pairs of −1-1 blocks in a determinant-one matrix can be treated as rotations through π\pi. Multiplication by a fixed reflection maps SO(n)SO(n) homeomorphically onto the other component.

The topology of SO(n)SO(n) contains information not visible in its Lie algebra. Since SO(2)SO(2) is a circle,

π1(SO(2))≅Z.\pi_1(SO(2))\cong\mathbb Z.

For all n≥3n\geq3, however, its fundamental group is

π1(SO(n))≅Z/2Z.\pi_1(SO(n))\cong\mathbb Z/2\mathbb Z.

A full 2π2\pi rotation represents the nontrivial class, while traversing this loop twice gives a contractible loop. This is the topological distinction behind the familiar belt-trick demonstration of the difference between 2π2\pi and 4π4\pi rotations.

The spin group Spin⁡(n)\operatorname{Spin}(n) gives a connected double covering space of SO(n)SO(n) for n≥2n\geq2:

1⟶{1,−1}⟶Spin⁡(n)⟶SO(n)⟶1.1\longrightarrow\{1,-1\} \longrightarrow\operatorname{Spin}(n) \longrightarrow SO(n) \longrightarrow1.

For n≥3n\geq3, this is the universal covering group, and Spin⁡(n)\operatorname{Spin}(n) is simply connected. In dimension two, Spin⁡(2)\operatorname{Spin}(2) is itself a circle, and its covering map doubles the angular parameter; it is not the universal cover of SO(2)SO(2). Spin groups can be constructed inside a Clifford algebra, where suitable products of unit vectors encode products of reflections.

The three-dimensional case has a particularly explicit model. Unit quaternions form a group isomorphic to Spin⁡(3)\operatorname{Spin}(3). Identifying R3\mathbb R^3 with the purely imaginary quaternions, a unit quaternion qq acts by

v⟼qvq−1.v\longmapsto qvq^{-1}.

Both qq and −q-q determine the same rotation. Thus SO(3)SO(3) is the quotient of the three-sphere by antipodal identification, making it homeomorphic to real projective three-space. In dimension four, the exceptional identity

Spin⁡(4)≅SU(2)×SU(2)\operatorname{Spin}(4)\cong SU(2)\times SU(2)

reflects a decomposition of its Lie algebra into two three-dimensional ideals.

Orthogonal groups also organize familiar geometric spaces through their group actions. The action of O(n)O(n) on the unit sphere is transitive, and the stabilizer of one unit vector is O(n−1)O(n-1), giving

Sn−1≅O(n)/O(n−1).S^{n-1}\cong O(n)/O(n-1).

Likewise, the space of ordered orthonormal kk-frames, called a Stiefel manifold, is

Vk(Rn)≅O(n)/O(n−k).V_k(\mathbb R^n)\cong O(n)/O(n-k).

These quotient descriptions connect orthogonal groups to algebraic topology and the geometry of vector bundles. In particular, the corresponding fibrations allow homotopy information in different dimensions to be compared rather than calculated separately.

Indefinite real forms and complex orthogonal groups

Positive definiteness is a special feature of the compact real orthogonal group. For a nondegenerate real quadratic form, Sylvester’s law of inertia gives coordinates in which

q(x)=x12+⋯+xp2−xp+12−⋯−xp+q2.q(x)=x_1^2+\cdots+x_p^2 -x_{p+1}^2-\cdots-x_{p+q}^2.

The pair (p,q)(p,q) is its signature. Writing

Jp,q=diag⁡(Ip,−Iq),J_{p,q}=\operatorname{diag}(I_p,-I_q),

the corresponding group is

O(p,q)={A∈GLp+q(R):ATJp,qA=Jp,q}.O(p,q)=\{A\in GL_{p+q}(\mathbb R): A^{\mathsf T}J_{p,q}A=J_{p,q}\}.

The positive-definite case is O(n,0)=O(n)O(n,0)=O(n). When both pp and qq are positive, the group is noncompact.

The difference can already be seen in O(1,1)O(1,1). The matrices

B(t)=(cosh⁡tsinh⁡tsinh⁡tcosh⁡t)B(t)= \begin{pmatrix} \cosh t&\sinh t\\ \sinh t&\cosh t \end{pmatrix}

preserve x2−y2x^2-y^2 and have determinant one, but their entries grow without bound as ∣t∣|t| increases. Unlike Euclidean rotations, they can have real eigenvalues ete^t and e−te^{-t} away from the unit circle. Preservation of an indefinite form therefore does not imply preservation of an ordinary Euclidean norm. Nonzero vectors can also have q(v)=0q(v)=0; these isotropic vectors form a cone rather than a sphere.

For p,q>0p,q>0, O(p,q)O(p,q) has four connected components. Determinant distinguishes two possibilities, while an additional orientation distinction for maximal positive or negative subspaces separates each into two. The determinant-one subgroup SO(p,q)SO(p,q) consequently has two components; its identity component is commonly denoted SO0(p,q)SO_0(p,q). Thus “special orthogonal” does not mean “connected” for indefinite real forms. A maximal compact subgroup of O(p,q)O(p,q) is O(p)×O(q)O(p)\times O(q).

The most prominent physical example is the Lorentz group O(1,3)O(1,3), or equivalently O(3,1)O(3,1) under the opposite sign convention. It preserves the spacetime quadratic form

c2t2−x2−y2−z2.c^2t^2-x^2-y^2-z^2.

Its identity component consists of proper orthochronous Lorentz transformations, preserving spatial orientation and the chosen direction of time. Spatial rotations and Lorentz boosts belong to this component. Translations are not Lorentz transformations; including them produces the Poincaré group.

Over the complex numbers, every nondegenerate quadratic form of a given dimension is equivalent to the standard form. Accordingly,

O(n,C)={A∈GLn(C):ATA=In}O(n,\mathbb C)=\{A\in GL_n(\mathbb C):A^{\mathsf T}A=I_n\}

represents the single nondegenerate complex case up to change of basis. This must not be confused with the unitary group, whose equation is A∗A=InA^*A=I_n, using the conjugate transpose. Complex orthogonal transformations preserve a complex bilinear form, not a positive-definite Hermitian inner product.

The complex group O(n,C)O(n,\mathbb C) is noncompact for n≥2n\geq2, even though its compact real form is O(n)O(n). Its determinant-one subgroup SO(n,C)SO(n,\mathbb C) is connected in the usual complex topology. These groups are also examples of an algebraic group: their multiplication, inversion, and defining relations belong naturally to the framework of algebraic geometry.

Orthogonal groups over finite fields

Over a finite field, an orthogonal group is finite rather than a continuous Lie group. Its structure depends on the field size, the dimension, and the equivalence class of the quadratic form. For a field Fq\mathbb F_q of odd cardinality, nondegenerate quadratic forms are classified by dimension and discriminant. In even dimension 2m2m, the notation

O+(2m,q),O−(2m,q)O^+(2m,q),\qquad O^-(2m,q)

distinguishes the split and nonsplit types. Their Witt indices—the largest dimensions of totally isotropic subspaces—are mm and m−1m-1, respectively.

A subspace is totally isotropic when the quadratic form vanishes on every one of its vectors. In the split even-dimensional case, the space decomposes into mm hyperbolic planes, each having a quadratic form equivalent to xyxy. In the nonsplit case, there are only m−1m-1 such planes, together with a two-dimensional anisotropic part. This difference changes both the geometry of the associated isotropic subspaces and the order of the preserving group.

For odd qq, the group orders are

∣O(2m+1,q)∣=2qm2∏i=1m(q2i−1),|O(2m+1,q)| =2q^{m^2}\prod_{i=1}^{m}(q^{2i}-1),

and

∣O±(2m,q)∣=2qm(m−1)(qm∓1)∏i=1m−1(q2i−1).|O^\pm(2m,q)| =2q^{m(m-1)}(q^m\mp1) \prod_{i=1}^{m-1}(q^{2i}-1).

In the second formula, the plus-type group has factor qm−1q^m-1, while the minus-type group has factor qm+1q^m+1. These expressions can be derived by counting suitable bases adapted to the quadratic form. For odd qq, the determinant-one subgroup has half the order of the full orthogonal group.

In odd dimension over a finite field of odd order, there are two equivalence classes of nondegenerate quadratic forms, but scaling a form by a nonzero scalar leaves its isometry group unchanged. The resulting orthogonal groups are therefore isomorphic, accounting for the single usual notation O(2m+1,q)O(2m+1,q). In dimension two, the contrast is especially transparent:

∣O+(2,q)∣=2(q−1),∣O−(2,q)∣=2(q+1).|O^+(2,q)|=2(q-1),\qquad |O^-(2,q)|=2(q+1).

The different factors arise from split and nonsplit one-dimensional tori inside the determinant-one groups.

Characteristic 22 requires a separate formulation. The polar form

bq(u,v)=q(u+v)−q(u)−q(v)b_q(u,v)=q(u+v)-q(u)-q(v)

is alternating, and preserving it need not imply preservation of qq. In odd dimension its radical also requires special care. Moreover, determinant no longer supplies the same two-valued distinction because 1=−11=-1 in the field. The real or odd-characteristic definition of “special” cannot therefore be transferred without adjustment; characteristic-two orthogonal theory uses additional invariants and subgroup constructions.

Finite orthogonal groups contribute major families to the theory of finite simple groups. Appropriate derived subgroups, usually denoted by variants of Ω\Omega, and their projective quotients produce many simple groups, apart from low-dimensional and small-field exceptions. The distinction among OO, SOSO, Ω\Omega, and projective versions is essential: the full orthogonal group is generally not itself simple. Their actions on isotropic points and subspaces also connect them to finite incidence geometries and combinatorial constructions.

Representations, invariants, and harmonic analysis

The defining representation of O(n)O(n) is its action on Rn\mathbb R^n, but many other representations arise from polynomials, tensors, and functions on spheres. Because O(n)O(n) is compact, it has a normalized Haar measure, a probability measure invariant under left and right multiplication. Averaging an inner product over the group makes it invariant. Consequently, finite-dimensional real representations admit invariant positive-definite inner products, and complex representations admit invariant Hermitian inner products. Every invariant subspace then has an invariant orthogonal complement, yielding complete reducibility.

A basic invariant-theoretic problem asks which polynomial expressions in vectors remain unchanged under simultaneous orthogonal transformations. For vectors v1,…,vkv_1,\ldots,v_k, all pairwise inner products

⟨vi,vj⟩\langle v_i,v_j\rangle

are invariant. The first fundamental theorem for the orthogonal group states, in the classical real or complex characteristic-zero setting, that these inner products generate the polynomial invariants. Equivalently, the invariant information is encoded by the Gram matrix

Gij=⟨vi,vj⟩.G_{ij}=\langle v_i,v_j\rangle.

When the number of vectors exceeds the dimension, the entries satisfy rank relations, including the vanishing of all (n+1)×(n+1)(n+1)\times(n+1) minors.

The difference between O(n)O(n) and SO(n)SO(n) is visible in oriented volume. For nn vectors, the determinant

det⁡(v1,…,vn)\det(v_1,\ldots,v_n)

is invariant under SO(n)SO(n) but changes sign under an orthogonal transformation of determinant −1-1. Thus orthogonal invariance records lengths and mutual angles but not handedness, whereas special orthogonal invariance can retain orientation. In tensor notation, the Euclidean metric provides invariant pairwise contractions for O(n)O(n); the alternating volume tensor supplies an additional invariant for SO(n)SO(n).

The action on functions on the sphere leads to spherical harmonics. These are restrictions to Sn−1S^{n-1} of homogeneous polynomials annihilated by the Euclidean Laplacian. Harmonic polynomials of degree ℓ\ell form a finite-dimensional invariant space. For n≥3n\geq3, its dimension is

dim⁡Hℓ(Rn)=(n+ℓ−1ℓ)−(n+ℓ−3ℓ−2),\dim\mathcal H_\ell(\mathbb R^n) = \binom{n+\ell-1}{\ell} -\binom{n+\ell-3}{\ell-2},

with the second term interpreted as zero when ℓ<2\ell<2. For n=3n=3, this gives 2ℓ+12\ell+1, the multiplicity familiar from angular momentum in quantum mechanics.

Finite-dimensional representations of SO(3)SO(3) correspond to integer angular momentum. Representations of its double cover Spin⁡(3)≅SU(2)\operatorname{Spin}(3)\cong SU(2) also include half-integer angular momentum, but those representations do not descend to ordinary representations of SO(3)SO(3). This distinction is the mathematical basis for separating tensorial and spinorial transformation laws.

The general structure belongs to the representation theory of compact groups and their complexified Lie algebras. The complex Lie algebras so(2m+1,C)\mathfrak{so}(2m+1,\mathbb C) and so(2m,C)\mathfrak{so}(2m,\mathbb C) form the classical families of types BmB_m and DmD_m, subject to familiar low-rank coincidences. Hermann Weyl’s The Classical Groups (1939) systematized the connection between these groups, their representations, and polynomial invariants.

Numerical methods and scientific applications

Orthogonal matrices are fundamental to numerical linear algebra because they preserve the Euclidean norm:

∥Qx∥2=∥x∥2.\|Qx\|_2=\|x\|_2.

Their singular values are all one, and their spectral-norm condition number is exactly one. Applying an orthogonal transformation therefore does not amplify a perturbation in Euclidean norm. This exact mathematical property does not eliminate floating-point error, but it helps explain the numerical stability of algorithms built from orthogonal transformations.

In a QR decomposition, a matrix is factored as A=QRA=QR, where QQ has orthonormal columns and RR is upper triangular. Householder reflections and planar Givens rotations construct this factorization while controlling roundoff error. Orthogonal similarity transformations also underlie the QR algorithm for eigenvalue computation. For real symmetric matrices, they preserve symmetry and can reduce the matrix to tridiagonal form before iterative diagonalization. The same geometry appears in the singular value decomposition,

A=UΣVT,A=U\Sigma V^{\mathsf T},

where the orthogonal factors describe changes of orthonormal coordinates and Σ\Sigma contains the nonnegative singular values.

A direct optimization problem on O(n)O(n) is the orthogonal Procrustes problem. Given real matrices AA and BB of the same size with nn columns, it seeks

min⁡Q∈O(n)∥AQ−B∥F2.\min_{Q\in O(n)}\|AQ-B\|_F^2.

Expanding the squared Frobenius norm reduces the task to maximizing

tr⁡(QTATB).\operatorname{tr}(Q^{\mathsf T}A^{\mathsf T}B).

If

ATB=UΣVT,A^{\mathsf T}B=U\Sigma V^{\mathsf T},

then Q=UVTQ=UV^{\mathsf T} is a minimizer. If a proper rotation is required, the determinant-one constraint is enforced by an appropriate sign correction in a singular-vector direction associated with a smallest singular value. This problem appears in point-set alignment, shape comparison, and registration of coordinate systems.

Orthogonal invariance is equally important in statistics. A standard multivariate normal vector XX satisfies

QX=dXQX\overset{d}=X

for every fixed Q∈O(n)Q\in O(n), because its density depends only on xTxx^{\mathsf T}x. More generally, orthogonal changes of coordinates transform a covariance matrix by Σ↦QΣQT\Sigma\mapsto Q\Sigma Q^{\mathsf T}. In principal component analysis, the spectral theorem selects orthogonal axes along which this covariance becomes diagonal. Haar-distributed orthogonal matrices supply uniformly random orthonormal frames; one sampling construction starts with a square matrix of independent standard normal entries and performs QR decomposition with a consistent sign convention on the diagonal of RR.

In differential geometry, a Riemannian metric selects orthonormal frames in each tangent space, and changes between such frames belong to O(n)O(n). An orientation reduces these transition transformations to SO(n)SO(n). In continuum mechanics and robotics, time-dependent rotations describe rigid-body attitude, while skew-symmetric matrices encode angular velocity. Quaternion coordinates exploit the spin covering to represent three-dimensional rotations without the coordinate singularities of Euler angles, with the necessary identification of qq and −q-q. Numerical integration on these spaces often uses exponential updates to preserve the orthogonality constraint throughout the calculation.