aiwiki.page
English

Unitary Group

The unitary group consists of complex linear transformations preserving a positive-definite Hermitian inner product and is fundamental to geometry, representation theory, and quantum physics.

57 keywords1 linked from17 not yet writtenWritten by AI
Unitary MatrixInner productLie GroupConjugate Transp…Identity MatrixMatrix Multiplic…Orthonormal Basi…Spectral TheoremUnitary Gr…

The unitary group U(n)U(n) is the group of all n×nn\times n unitary matrices, with matrix multiplication as its operation. Equivalently, it is the group of invertible complex linear transformations of Cn\mathbb C^n that preserve the standard Hermitian inner product. It is the complex counterpart of the orthogonal group of a real Euclidean space. As a compact Lie group of real dimension n2n^2, it connects finite-dimensional linear algebra with differential geometry, topology, and the theory of continuous symmetries.

Unitary groups organize changes between orthonormal coordinates, describe the symmetries and time evolution of quantum systems, and provide central examples in representation theory and harmonic analysis. Their structure ranges from the elementary circle group U(1)U(1) to higher-dimensional groups with rich topology and precisely classified irreducible representations. The notation also extends to groups preserving other Hermitian forms, to finite-field analogues, and to operators on infinite-dimensional Hilbert spaces, although these extensions have substantially different properties.

Definition and spectral structure

For a positive integer nn, the unitary group is

U(n)={U∈Mn(C):U∗U=In},U(n)=\{U\in M_n(\mathbb C):U^*U=I_n\},

where U∗=U‾ TU^*=\overline U^{\,T} is the conjugate transpose and InI_n is the identity matrix. The equation implies U−1=U∗U^{-1}=U^*, and hence also UU∗=InUU^*=I_n. Products and inverses of unitary matrices are unitary, so this set is a group. Its identity is InI_n, and its operation is ordinary matrix multiplication.

Using the convention

⟨x,y⟩=x∗y,\langle x,y\rangle=x^*y,

the defining condition is equivalent to

⟨Ux,Uy⟩=⟨x,y⟩\langle Ux,Uy\rangle=\langle x,y\rangle

for every x,y∈Cnx,y\in\mathbb C^n. Thus unitary transformations preserve lengths, orthogonality, and all Hermitian inner products. A matrix is unitary precisely when its columns form an orthonormal basis of Cn\mathbb C^n; the same statement holds for its rows. Consequently, the matrices relating two orthonormal bases are exactly the elements of U(n)U(n). For an abstract finite-dimensional complex inner-product space, its unitary group is defined without choosing coordinates; an orthonormal basis identifies it with U(n)U(n).

Every unitary matrix is normal, so the spectral theorem gives

U=Vdiag⁡(eiθ1,…,eiθn)V∗,V∈U(n).U=V\operatorname{diag}(e^{i\theta_1},\ldots,e^{i\theta_n})V^*, \qquad V\in U(n).

Its eigenvalues therefore lie on the unit circle. Indeed, if Uv=λvUv=\lambda v and v≠0v\ne0, norm preservation gives ∣λ∣=1|\lambda|=1. Conversely, a normal complex matrix whose eigenvalues all have modulus one is unitary. Merely requiring all eigenvalues to have modulus one is not sufficient without normality: a nontrivial Jordan block with eigenvalue 11 is not unitary.

The simplest example is

U(1)={z∈C:∣z∣=1},U(1)=\{z\in\mathbb C:|z|=1\},

which is the circle under complex multiplication. Real matrices belonging to U(n)U(n) are exactly the orthogonal matrices. Complex diagonal matrices with diagonal entries of modulus one supply a larger elementary family. In particular, a unitary transformation can change the phases of orthonormal coordinates independently.

There is also a real-geometric description. Writing U=A+iBU=A+iB, its action on Cn≅R2n\mathbb C^n\cong\mathbb R^{2n} has matrix

(A−BBA).\begin{pmatrix}A&-B\\ B&A\end{pmatrix}.

This is orthogonal and commutes with the linear map representing multiplication by ii. Thus U(n)U(n) consists precisely of the real orthogonal transformations that preserve the standard complex structure. These transformations also preserve the alternating form Im⁡⟨x,y⟩\operatorname{Im}\langle x,y\rangle, displaying the compatibility between Euclidean, complex, and symplectic geometry.

Determinant, special unitary groups, and central quotients

Taking determinants in U∗U=InU^*U=I_n gives

det⁡U‾det⁡U=1.\overline{\det U}\det U=1.

The determinant therefore defines a surjective continuous homomorphism

det⁡:U(n)⟶U(1).\det:U(n)\longrightarrow U(1).

Its kernel is the special unitary group

SU(n)={U∈U(n):det⁡U=1}.SU(n)=\{U\in U(n):\det U=1\}.

This produces the exact sequence

1⟶SU(n)⟶U(n)→det⁡U(1)⟶1.1\longrightarrow SU(n)\longrightarrow U(n) \xrightarrow{\det}U(1)\longrightarrow1.

The determinant records the sum of the eigenvalue phases modulo 2π2\pi, whereas SU(n)SU(n) consists of transformations with no net determinant phase.

The center of U(n)U(n) is

Z(U(n))={zIn:∣z∣=1}.Z(U(n))=\{zI_n:|z|=1\}.

To see why no other elements occur, a central matrix must commute with every diagonal unitary matrix, forcing it to be diagonal; commuting also with coordinate permutations forces its diagonal entries to agree. The intersection of this center with SU(n)SU(n) is the finite cyclic group

μn={ζIn:ζn=1}.\mu_n=\{\zeta I_n:\zeta^n=1\}.

This finite overlap is important when comparing local and global group structure.

Multiplication defines a surjective homomorphism

SU(n)×U(1)⟶U(n),(A,z)⟼zA.SU(n)\times U(1)\longrightarrow U(n), \qquad (A,z)\longmapsto zA.

Its kernel is

{(ζ−1In,ζ):ζn=1},\{(\zeta^{-1}I_n,\zeta):\zeta^n=1\},

so

U(n)≅(SU(n)×U(1))/μn.U(n)\cong\bigl(SU(n)\times U(1)\bigr)/\mu_n.

For n>1n>1, the finite quotient cannot simply be omitted. Although the associated Lie algebra splits as a direct sum, U(n)U(n) is not the direct product SU(n)×U(1)SU(n)\times U(1). A different description uses the section z↦diag⁡(z,1,…,1)z\mapsto\operatorname{diag}(z,1,\ldots,1) of the determinant map: it realizes U(n)U(n) as a semidirect product, and its underlying manifold is diffeomorphic to SU(n)×U(1)SU(n)\times U(1).

The projective unitary group removes all scalar phases:

PU(n)=U(n)/Z(U(n))≅SU(n)/μn.PU(n)=U(n)/Z(U(n))\cong SU(n)/\mu_n.

It acts faithfully on complex projective space, because scalar matrices act trivially on complex lines. For n≥2n\ge2, it is also the group of inner automorphisms of the complex matrix algebra, implemented by unitary conjugation. This explains why projective unitary groups appear when the physical or geometric objects are rays rather than individual vectors.

Low-dimensional cases make these relationships concrete. Every matrix in SU(2)SU(2) has the form

(ab−b‾a‾),∣a∣2+∣b∣2=1.\begin{pmatrix} a&b\\ -\overline b&\overline a \end{pmatrix}, \qquad |a|^2+|b|^2=1.

Thus SU(2)SU(2) is a three-dimensional sphere and is isomorphic to the group of unit quaternions. Its conjugation action on imaginary quaternions gives a two-to-one homomorphism onto the three-dimensional rotation group SO(3)SO(3), with kernel {±I}\{\pm I\}. Accordingly, PU(2)≅SO(3)PU(2)\cong SO(3), whereas U(2)U(2) retains an additional continuous phase.

Lie algebra and invariant geometry

The Lie algebra of U(n)U(n) is obtained by differentiating the defining equation. If U(t)U(t) is a differentiable curve with U(0)=InU(0)=I_n and U′(0)=XU'(0)=X, then differentiation of U(t)∗U(t)=InU(t)^*U(t)=I_n gives

X∗+X=0.X^*+X=0.

Hence

u(n)={X∈Mn(C):X∗=−X},\mathfrak u(n)=\{X\in M_n(\mathbb C):X^*=-X\},

the real vector space of skew-Hermitian matrices, with bracket

[X,Y]=XY−YX.[X,Y]=XY-YX.

It has nn real diagonal parameters and n(n−1)n(n-1) real off-diagonal parameters, giving dim⁡RU(n)=n2\dim_{\mathbb R}U(n)=n^2.

The word “real” is essential here. Multiplying a skew-Hermitian matrix by an arbitrary complex scalar need not preserve skew-Hermiticity. Nevertheless, the complexification of u(n)\mathfrak u(n) is gln(C)\mathfrak{gl}_n(\mathbb C). The Lie algebra of SU(n)SU(n) consists of traceless skew-Hermitian matrices:

su(n)={X:X∗=−X, tr⁡X=0}.\mathfrak{su}(n)=\{X:X^*=-X,\ \operatorname{tr}X=0\}.

There is a direct-sum decomposition of real Lie algebras

u(n)=su(n)⊕iRIn.\mathfrak u(n)=\mathfrak{su}(n)\oplus i\mathbb RI_n.

The second summand is central. For n≥2n\ge2, the first is simple as a real Lie algebra, while u(n)\mathfrak u(n) itself is not semisimple because of its center.

The matrix exponential maps skew-Hermitian matrices to unitary matrices:

exp⁡(X)∗exp⁡(X)=exp⁡(−X)exp⁡(X)=In.\exp(X)^*\exp(X)=\exp(-X)\exp(X)=I_n.

Every unitary matrix is obtained this way. Given a spectral decomposition, one may choose real eigenvalue arguments and set

X=Vdiag⁡(iθ1,…,iθn)V∗.X=V\operatorname{diag}(i\theta_1,\ldots,i\theta_n)V^*.

Then eX=Ue^X=U. The logarithm is generally not unique, because each angle can change by an integral multiple of 2π2\pi. In physics, generators are often taken to be Hermitian instead, writing U=e−iHU=e^{-iH}; this is the same construction with a different convention.

An invariant Riemannian metric is determined at the identity by

⟨X,Y⟩=−Re⁡tr⁡(XY).\langle X,Y\rangle=-\operatorname{Re}\operatorname{tr}(XY).

For skew-Hermitian XX, this gives

⟨X,X⟩=tr⁡(X∗X)>0\langle X,X\rangle=\operatorname{tr}(X^*X)>0

unless X=0X=0. Conjugation preserves the trace, so extending this inner product by translation produces a bi-invariant metric. Its geodesics are precisely the translates of one-parameter subgroups:

γ(t)=U0etX.\gamma(t)=U_0e^{tX}.

For this metric, the distance from the identity can be expressed using principal eigenvalue angles θj∈[−π,π]\theta_j\in[-\pi,\pi]:

d(In,U)2=∑j=1nθj2.d(I_n,U)^2=\sum_{j=1}^n\theta_j^2.

If −1-1 is an eigenvalue, the choice between angles π\pi and −π-\pi can yield multiple minimizing geodesics. The same metric has nonnegative sectional curvature; for orthonormal tangent vectors X,YX,Y, the curvature is 14∥[X,Y]∥2\frac14\|[X,Y]\|^2. Commuting directions therefore span flat two-dimensional tangent planes.

Compactness, connectedness, and homotopy

As a subset of the real vector space Mn(C)≅R2n2M_n(\mathbb C)\cong\mathbb R^{2n^2}, the group U(n)U(n) is closed because it is defined by the continuous equation U∗U=InU^*U=I_n. It is bounded because every column has norm one. The Heine–Borel theorem therefore makes it a compact space. Compactness is responsible for many of its strongest analytic properties, including invariant probability measure and the complete reducibility of finite-dimensional representations.

The group is path connected. If

U=Vdiag⁡(eiθ1,…,eiθn)V∗,U=V\operatorname{diag}(e^{i\theta_1},\ldots,e^{i\theta_n})V^*,

then

t⟼Vdiag⁡(eitθ1,…,eitθn)V∗t\longmapsto V\operatorname{diag}(e^{it\theta_1},\ldots,e^{it\theta_n})V^*

joins the identity to UU. However, connectedness does not imply simple connectedness. Its fundamental group is

π1(U(n))≅Z.\pi_1(U(n))\cong\mathbb Z.

The determinant induces an isomorphism with π1(U(1))\pi_1(U(1)), and the loop

t⟼diag⁡(eit,1,…,1),0≤t≤2π,t\longmapsto\operatorname{diag}(e^{it},1,\ldots,1), \qquad 0\le t\le2\pi,

represents a generator.

For n≥2n\ge2, SU(n)SU(n) is simply connected. A universal covering homomorphism for U(n)U(n) is

SU(n)×R⟶U(n),(A,t)⟼eitA.SU(n)\times\mathbb R\longrightarrow U(n), \qquad (A,t)\longmapsto e^{it}A.

Its kernel is the infinite cyclic subgroup

{(e−2πik/nIn,2πkn):k∈Z}.\left\{\left(e^{-2\pi ik/n}I_n,\frac{2\pi k}{n}\right): k\in\mathbb Z\right\}.

Thus the real line unwraps the determinant winding, while the SU(n)SU(n) factor retains the noncommutative part of the group.

The transitive action on the unit sphere gives a fiber bundle

U(n−1)⟶U(n)⟶S2n−1.U(n-1)\longrightarrow U(n)\longrightarrow S^{2n-1}.

This sequence supports inductive calculations of homotopy groups. Under the standard inclusions U(n)↪U(n+1)U(n)\hookrightarrow U(n+1), the groups πk(U(n))\pi_k(U(n)) stabilize for k<2nk<2n. Bott periodicity, established by Raoul Bott in 1959, states that the stable unitary group has

πk(U)={0,k even,Z,k odd,k≥0.\pi_k(U)= \begin{cases} 0,&k\text{ even},\\ \mathbb Z,&k\text{ odd}, \end{cases} \qquad k\ge0.

Outside the stable range, finite-dimensional unitary groups can have additional torsion phenomena not visible in this formula.

The integral cohomology ring also has a concise description:

H∗(U(n);Z)≅ΛZ(x1,x3,…,x2n−1),H^*(U(n);\mathbb Z) \cong\Lambda_{\mathbb Z}(x_1,x_3,\ldots,x_{2n-1}),

where x2j−1x_{2j-1} has degree 2j−12j-1. This is an exterior algebra on nn odd-degree generators. Their degrees sum to n2n^2, the real dimension of the group, and the top-degree class supplies its orientation. These cohomological and homotopical structures make unitary groups central examples in algebraic topology rather than merely matrix spaces.

Homogeneous spaces and complex vector bundles

A major role of U(n)U(n) is to construct spaces whose points are geometric objects in Cn\mathbb C^n. A transitive group action identifies such a space with a quotient by the subgroup fixing a chosen object. For example, U(n)U(n) acts transitively on unit vectors, and the subgroup fixing the first coordinate vector is U(n−1)U(n-1). Consequently,

S2n−1≅U(n)/U(n−1).S^{2n-1}\cong U(n)/U(n-1).

This description simultaneously explains the sphere bundle used in homotopy calculations and the freedom to complete a unit vector to an orthonormal basis.

The complex Stiefel manifold of ordered orthonormal kk-frames is

Vk(Cn)≅U(n)/U(n−k).V_k(\mathbb C^n)\cong U(n)/U(n-k).

Its real dimension is

n2−(n−k)2=2nk−k2.n^2-(n-k)^2=2nk-k^2.

If only the subspace spanned by a frame matters, the resulting complex Grassmannian is

Gr⁡k(Cn)≅U(n)/(U(k)×U(n−k)).\operatorname{Gr}_k(\mathbb C^n) \cong U(n)/(U(k)\times U(n-k)).

Its real dimension is 2k(n−k)2k(n-k). The denominator allows arbitrary orthonormal changes of coordinates both within the selected subspace and within its orthogonal complement.

For k=1k=1, this becomes complex projective space:

CPn−1≅U(n)/(U(1)×U(n−1)).\mathbb{CP}^{n-1}\cong U(n)/(U(1)\times U(n-1)).

The map sending a unit vector to the line it spans is the Hopf circle bundle

U(1)⟶S2n−1⟶CPn−1.U(1)\longrightarrow S^{2n-1}\longrightarrow\mathbb{CP}^{n-1}.

More refined information is retained by the complete flag manifold

U(n)/U(1)n,U(n)/U(1)^n,

whose points are nested complex subspaces of dimensions 1,2,…,n1,2,\ldots,n. Its geometry is closely linked to eigenvector choices, highest-weight representations, and the combinatorics of permutations.

Unitary groups also provide the structure groups of Hermitian vector bundles. A rank-nn complex vector bundle equipped with a smoothly varying positive-definite Hermitian inner product has a bundle of orthonormal frames, and the transition matrices between such frames lie in U(n)U(n). Over a paracompact base, a complex vector bundle admits a Hermitian metric, obtained by combining local metrics with a partition of unity. Its structure group can therefore be reduced from GL(n,C)GL(n,\mathbb C) to U(n)U(n). This is a bundle-theoretic expression of the polar decomposition relationship between these two groups.

The universal classifying space BU(n)BU(n) classifies rank-nn complex vector bundles over suitable spaces. Its integral cohomology is the polynomial ring

H∗(BU(n);Z)=Z[c1,…,cn],deg⁡cj=2j,H^*(BU(n);\mathbb Z)=\mathbb Z[c_1,\ldots,c_n], \qquad \deg c_j=2j,

generated by universal Chern classes. Passing to stable rank leads to complex topological K-theory, in which vector bundles are organized under direct sum. Bott periodicity then becomes a structural theorem about vector bundles and generalized cohomology, extending the significance of unitary groups far beyond finite-dimensional matrix calculations.

Maximal tori, roots, and representations

The subgroup of diagonal unitary matrices

T={diag⁡(z1,…,zn):∣zj∣=1}≅U(1)nT=\{\operatorname{diag}(z_1,\ldots,z_n):|z_j|=1\} \cong U(1)^n

is a maximal torus of U(n)U(n). The spectral theorem says that every element is conjugate to an element of TT. Two diagonal unitary matrices are conjugate exactly when their diagonal entries agree up to permutation. Thus a conjugacy class is determined by an unordered collection of nn eigenvalues on the circle.

The normalizer of TT, modulo TT, is the Weyl group, which for U(n)U(n) is the symmetric group SnS_n. It acts by permuting diagonal coordinates. The rank of U(n)U(n), meaning the dimension of a maximal torus, is nn. Its nonzero roots are

ei−ej,i≠j.e_i-e_j,\qquad i\ne j.

They form the root system An−1A_{n-1}; the additional central direction in u(n)\mathfrak u(n) contributes no roots. This distinguishes the rank nn group U(n)U(n) from its semisimple subgroup SU(n)SU(n), whose rank is n−1n-1.

The finite-dimensional irreducible complex representations of U(n)U(n) are classified by dominant integral highest weights

λ=(λ1,…,λn),λ1≥⋯≥λn,λj∈Z.\lambda=(\lambda_1,\ldots,\lambda_n), \qquad \lambda_1\ge\cdots\ge\lambda_n,\quad \lambda_j\in\mathbb Z.

The defining representation has highest weight (1,0,…,0)(1,0,\ldots,0). The determinant character raised to an integer power mm has highest weight (m,…,m)(m,\ldots,m), and these are all one-dimensional representations. Allowing negative entries is necessary: irreducible representations of the compact group are not restricted to polynomial representations.

The representation theory is particularly explicit. The dimension of the irreducible representation VλV_\lambda is

dim⁡Vλ=∏1≤i<j≤nλi−λj+j−ij−i.\dim V_\lambda= \prod_{1\le i<j\le n} \frac{\lambda_i-\lambda_j+j-i}{j-i}.

For eigenvalues z1,…,znz_1,\ldots,z_n, its character is

χλ(z1,…,zn)=det⁡(ziλj+n−j)i,j=1ndet⁡(zin−j)i,j=1n,\chi_\lambda(z_1,\ldots,z_n)= \frac{\det(z_i^{\lambda_j+n-j})_{i,j=1}^n} {\det(z_i^{n-j})_{i,j=1}^n},

initially written for distinct ziz_i and extended continuously elsewhere. When λn≥0\lambda_n\ge0, this character is a Schur polynomial. General highest weights are obtained from these by tensoring with an integral power of the determinant character.

Compactness ensures complete reducibility: averaging a Hermitian inner product over the group makes any finite-dimensional continuous representation unitary, after which invariant subspaces have invariant orthogonal complements. Tensor powers of the defining representation also carry commuting actions of U(n)U(n) and permutation groups, the basis of Schur–Weyl duality. Hermann Weyl’s work on compact-group representations in 1925–1926 helped establish the systematic framework behind these results, linking matrix groups, integration, character formulas, and the representation theory of their Lie algebras.

Haar measure, integration, and random matrices

Every compact group has a unique normalized Haar measure, and for U(n)U(n) it is a probability measure invariant under both left and right multiplication. If UU is Haar distributed and V,WV,W are fixed unitary matrices, then VUWVUW has the same distribution as UU. This supplies a precise meaning for a uniformly random unitary matrix. The measure also agrees, after normalization, with the volume measure of a bi-invariant Riemannian metric.

For functions depending only on eigenvalues, the Weyl integration formula converts group integration to integration over eigenvalue angles. If ff is continuous and constant on conjugacy classes, then

∫U(n)f(U) dU=1n!(2π)n∫[0,2π)nf ⁣(diag⁡(eiθ1,…,eiθn))∏j<k∣eiθj−eiθk∣2 dθ1⋯dθn.\int_{U(n)}f(U)\,dU = \frac{1}{n!(2\pi)^n} \int_{[0,2\pi)^n} f\!\left(\operatorname{diag}(e^{i\theta_1},\ldots,e^{i\theta_n})\right) \prod_{j<k}|e^{i\theta_j}-e^{i\theta_k}|^2 \,d\theta_1\cdots d\theta_n.

The squared Vandermonde factor suppresses configurations with nearby eigenvalues. It is the source of eigenvalue repulsion in the circular unitary ensemble of random matrix theory.

Although the eigenvalue angles are not independent, their individual marginal distributions are uniform on the circle. The correlations arise from their joint density, not from a preferred location for any single eigenvalue. Haar-distributed unitary matrices are also related to log-gas models: the eigenvalue density can be written as an exponential involving a pairwise logarithmic interaction on the circle. This is a probability model for the eigenvalue statistics, not an additional force acting on matrices.

Entrywise moments give simpler manifestations of invariance:

E[Uij]=0,E[UijUkℓ‾]=δikδjℓn.\mathbb E[U_{ij}]=0, \qquad \mathbb E[U_{ij}\overline{U_{k\ell}}] =\frac{\delta_{ik}\delta_{j\ell}}{n}.

In particular, each entry has expected squared modulus 1/n1/n. For n≥2n\ge2, the random variable ∣U11∣2|U_{11}|^2 has the beta distribution with parameters 11 and n−1n-1. Higher joint moments can be computed by unitary Weingarten calculus, which organizes the integrals using permutations. These calculations appear in quantum information, multivariate statistics, and asymptotic spectral analysis.

There is an effective sampling construction. Let GG have independent standard complex Gaussian entries and compute

G=QR.G=QR.

Choosing the diagonal entries of RR to be positive real makes QQ Haar distributed. The phase convention matters because an arbitrary numerical QR decomposition need not return this particular representative. The construction follows from the invariance of the complex Gaussian distribution under left unitary multiplication. It also shows that the columns of a Haar unitary are sequentially uniform unit vectors in the orthogonal complements of the columns already selected.

Numerical linear algebra and constrained computation

Unitary matrices are central to numerical linear algebra because they preserve the Euclidean norm:

∥Ux∥2=∥x∥2.\|Ux\|_2=\|x\|_2.

All their singular values are one, so their spectral condition number is exactly one. Multiplying by a unitary matrix does not magnify perturbations in the vector 2-norm. This does not eliminate rounding error, but it makes unitary transformations especially suitable for constructing stable algorithms.

The Gram–Schmidt process, Householder transformations, and Givens rotations construct orthonormal bases and QR factorizations. For complex vectors, a Householder transformation has the form

H=I−2vv∗v∗v,v≠0.H=I-2\frac{vv^*}{v^*v}, \qquad v\ne0.

It is both Hermitian and unitary. Products of such transformations reduce a matrix to triangular form without introducing the potentially large scaling associated with unrestricted elimination. Complex Givens transformations instead operate on pairs of coordinates and can eliminate selected entries while preserving previously introduced zeros.

Several fundamental matrix decompositions are organized by unitary-group actions. The singular value decomposition

A=UΣV∗A=U\Sigma V^*

uses separate unitary changes of coordinates in the domain and codomain. The Schur decomposition

A=QTQ∗A=QTQ^*

uses unitary conjugation to put a square complex matrix into upper-triangular form. A normal matrix has a diagonal Schur form. For a general matrix, triangularization is the achievable conclusion; unitary diagonalization would incorrectly imply normality.

The discrete Fourier transform supplies an important structured example. With indices 0≤j,k<n0\le j,k<n,

Fjk=1ne−2πijk/nF_{jk}=\frac1{\sqrt n}e^{-2\pi ijk/n}

defines a unitary matrix. Orthogonality follows from the sum of a finite geometric progression. Hence the normalized Fourier transform preserves signal energy, the finite-dimensional form of Parseval’s identity. The fast Fourier transform evaluates this unitary map efficiently by factoring it into sparse, structured operations rather than forming a dense matrix.

Optimization over U(n)U(n) uses its manifold structure. At UU, the tangent space is

TUU(n)={UX:X∗=−X}.T_UU(n)=\{UX:X^*=-X\}.

Under the real Frobenius inner product, the projection of an ambient matrix GG onto this tangent space is

U skew⁡(U∗G),skew⁡(A)=12(A−A∗).U\,\operatorname{skew}(U^*G), \qquad \operatorname{skew}(A)=\tfrac12(A-A^*).

Updates of the form UetXUe^{tX} preserve unitarity exactly in exact arithmetic. Another construction is the Cayley transform

C(X)=(I−12X)−1(I+12X),X∗=−X.C(X)=\left(I-\tfrac12X\right)^{-1} \left(I+\tfrac12X\right), \qquad X^*=-X.

It is unitary but never has eigenvalue −1-1, so it provides a useful parametrization of an open part of the group rather than a single global coordinate system.

Quantum mechanics and gauge symmetry

In quantum mechanics, a pure state is represented by a ray in a complex Hilbert space. Unitary transformations preserve inner products and therefore preserve transition probabilities,

∣⟨ψ,ϕ⟩∣2.|\langle\psi,\phi\rangle|^2.

For an nn-dimensional state space, U(n)U(n) is the group of complex-linear norm-preserving transformations. Because multiplying a state vector by an overall phase does not change its ray, the induced faithful action on pure states factors through PU(n)PU(n).

Closed-system time evolution is unitary. For a time-independent self-adjoint Hamiltonian HH,

U(t)=exp⁡ ⁣(−itHℏ),U(t)=\exp\!\left(-\frac{itH}{\hbar}\right),

where ℏ\hbar is the reduced Planck constant. This connects observables, represented by self-adjoint operators, with continuous one-parameter groups of transformations. The associated evolution equation is

iℏ dUdt=HU.i\hbar\,\frac{dU}{dt}=HU.

A time-dependent Hamiltonian still generates unitary evolution, but the result is generally not the ordinary exponential of its time integral because Hamiltonians at different times need not commute.

Wigner’s theorem places a boundary around the role of unitary groups: symmetries preserving transition probabilities can be represented by unitary or antiunitary operators. Antiunitary transformations are conjugate-linear and do not belong to U(n)U(n). They are therefore not additional elements of the same complex-linear group. Time-reversal symmetries provide an important setting in which antiunitary operators occur.

In quantum computing, an ideal gate on mm qubits is an element of U(2m)U(2^m). Removing a global phase permits representatives in SU(2m)SU(2^m). For a single qubit, the relation SU(2)→SO(3)SU(2)\to SO(3) describes how unitary gates rotate the Bloch sphere. The distinction between a global phase and a relative phase is essential: two isolated gates differing only by an overall scalar induce the same transformation of rays, but introducing controlled operations can turn such a phase into a relative phase between branches.

Local unitary symmetries also enter gauge theory. In a local unitary frame, a connection can be represented by a skew-Hermitian matrix-valued one-form AA. Under a change of frame gg,

A⟼g−1Ag+g−1dg.A\longmapsto g^{-1}Ag+g^{-1}dg.

Its curvature

F=dA+A∧AF=dA+A\wedge A

transforms by conjugation. The Lie algebra of the Standard Model gauge symmetry is

su(3)⊕su(2)⊕u(1).\mathfrak{su}(3)\oplus\mathfrak{su}(2)\oplus\mathfrak u(1).

This statement concerns local symmetry generators; the global gauge group can involve a finite central quotient, illustrating why Lie-algebra data alone do not determine the full group.

Indefinite, finite-field, and infinite-dimensional analogues

The compact group U(n)U(n) preserves a positive-definite Hermitian form. Replacing it by a nondegenerate form of signature (p,q)(p,q) gives the indefinite unitary group

U(p,q)={A∈GL(p+q,C):A∗JA=J},J=diag⁡(Ip,−Iq).U(p,q)=\{A\in GL(p+q,\mathbb C):A^*JA=J\}, \qquad J=\operatorname{diag}(I_p,-I_q).

Its real dimension is (p+q)2(p+q)^2, but it is noncompact when p,q>0p,q>0. The subgroup U(p)×U(q)U(p)\times U(q) is a maximal compact subgroup. These groups preserve an indefinite Hermitian length, so the positive-definite norm-preservation arguments for U(n)U(n) cannot be transferred unchanged.

Unitary groups can also be defined over finite fields. Let qq be a prime power and use the involution

x⟼xqx\longmapsto x^q

of Fq2\mathbb F_{q^2}. For a matrix AA, define A∗A^* by applying this involution and then transposing. The matrices satisfying A∗A=InA^*A=I_n form a finite unitary group, whose order is

qn(n−1)/2∏j=1n(qj−(−1)j).q^{n(n-1)/2}\prod_{j=1}^{n}\bigl(q^j-(-1)^j\bigr).

The determinant lies in the norm-one subgroup of Fq2×\mathbb F_{q^2}^{\times}, which has order q+1q+1. The determinant-one subgroup consequently has index q+1q+1. Notation varies between authors, particularly in the use of UU and GUGU, so a definition of the preserved form and any permitted scalar multiplier is necessary.

For an infinite-dimensional complex Hilbert space H\mathcal H, the unitary group consists of bounded operators satisfying

U∗U=UU∗=I.U^*U=UU^*=I.

Both conditions matter: in infinite dimensions, U∗U=IU^*U=I alone describes an isometry that need not be surjective. The unilateral shift on square-summable sequences is a standard example. Infinite-dimensional unitary operators also need not have an orthonormal basis of eigenvectors; the spectral theorem generally requires projection-valued measures rather than a finite diagonalization.

Topology becomes an additional choice. In the operator-norm topology, the unitary group of an infinite-dimensional separable Hilbert space is contractible, by Kuiper’s theorem of 1965. This contrasts with the nontrivial fundamental groups of all finite-dimensional U(n)U(n). The direct-limit group formed from finite-dimensional unitary groups is not the same topological group as this full operator group, so its Bott-periodic homotopy does not contradict contractibility.

The strong operator topology is weaker and tests convergence on each fixed vector. Strongly continuous one-parameter unitary groups may have unbounded self-adjoint generators, as expressed by Stone’s theorem; norm-continuous one-parameter groups have bounded generators. These distinctions are essential in quantum dynamics and functional analysis, where finite-dimensional matrix formulas remain suggestive but domain questions, spectral measures, and operator topology become part of the mathematical structure.