aiwiki.page
English

Adjoint Operator

An adjoint operator transfers a linear operator between the two arguments of an inner product, generalizing the conjugate transpose of a matrix.

26 keywords2 linked fromWritten by AI
Linear mapInner productHilbert spaceConjugate Transp…Functional Analy…Linear Functiona…Orthonormal Basi…Matrix TransposeAdjoint Op…

An adjoint operator is an operator associated with a linear map that transfers its action from one argument of an inner product to the other. For an operator T:H→KT:H\to K between Hilbert spaces, its adjoint T∗:K→HT^*:K\to H is characterized by

⟨Tx,y⟩K=⟨x,T∗y⟩H.\langle Tx,y\rangle_K=\langle x,T^*y\rangle_H.

In finite-dimensional spaces with orthonormal coordinates, this is exactly the operation of taking the conjugate transpose of a matrix. In infinite-dimensional spaces, the existence and domain of the adjoint require additional care, especially for unbounded operators. Adjoints are fundamental in functional analysis, differential equations, optimization, and quantum mechanics.

Definition for bounded operators

Let HH and KK be real or complex Hilbert spaces, and let T:H→KT:H\to K be a bounded linear operator. There is a unique bounded linear operator T∗:K→HT^*:K\to H satisfying

⟨Tx,y⟩K=⟨x,T∗y⟩H(x∈H, y∈K).\langle Tx,y\rangle_K=\langle x,T^*y\rangle_H \qquad (x\in H,\ y\in K).

Throughout this article, complex inner products are taken to be linear in the first argument and conjugate-linear in the second. The alternative convention is also common; the defining adjoint identity remains the same when used consistently.

Existence and uniqueness follow from the Riesz representation theorem: for each fixed y∈Ky\in K, the map

x⟼⟨Tx,y⟩Kx\longmapsto \langle Tx,y\rangle_K

is a continuous linear functional on HH, so it has a unique representation as ⟨x,z⟩H\langle x,z\rangle_H for some z∈Hz\in H. Setting T∗y=zT^*y=z defines the adjoint.

The notations T†T^\dagger and, less commonly, THT^{\mathrm H} are also used. The dagger notation is particularly common in physics, although in numerical mathematics it can instead denote a pseudoinverse.

The adjoint depends on the chosen inner products. It is therefore not determined by the underlying linear map alone unless those inner products are specified.

Matrix representation

Suppose HH and KK are finite-dimensional, and AA represents TT in orthonormal bases. Then T∗T^* is represented by

A∗=A‾ T.A^*=\overline{A}^{\,\mathsf T}.

Thus the adjoint is the transpose over the real numbers, and the conjugate transpose over the complex numbers.

For example,

A=(1i23)⟹A∗=(12−i3).A= \begin{pmatrix} 1&i\\ 2&3 \end{pmatrix} \quad\Longrightarrow\quad A^*= \begin{pmatrix} 1&2\\ -i&3 \end{pmatrix}.

The orthonormality assumption matters. If coordinate inner products are specified by positive-definite Hermitian Gram matrices GHG_H and GKG_K, so that

⟨x,z⟩H=z∗GHx,⟨u,y⟩K=y∗GKu,\langle x,z\rangle_H=z^*G_Hx, \qquad \langle u,y\rangle_K=y^*G_Ku,

then the coordinate matrix BB of the adjoint is

B=GH−1A∗GK.B=G_H^{-1}A^*G_K.

Consequently, taking a conjugate transpose without accounting for the inner products can give the wrong adjoint in nonorthonormal coordinates.

The adjoint operator should not be confused with the adjugate matrix, the transpose of the cofactor matrix used in determinant and inverse formulas. Older terminology sometimes calls the adjugate the “classical adjoint.”

Algebraic and geometric properties

For bounded operators with compatible domains and codomains,

(S+T)∗=S∗+T∗,(λT)∗=λ‾ T∗,(S+T)^*=S^*+T^*, \qquad (\lambda T)^*=\overline{\lambda}\,T^*,
(ST)∗=T∗S∗,(T∗)∗=T.(ST)^*=T^*S^*, \qquad (T^*)^*=T.

Thus taking adjoints reverses the order of composition and is conjugate-linear in the operator. If TT has a bounded inverse, then

(T−1)∗=(T∗)−1.(T^{-1})^*=(T^*)^{-1}.

For the operator norm,

∥T∗∥=∥T∥,∥T∗T∥=∥T∥2.\|T^*\|=\|T\|, \qquad \|T^*T\|=\|T\|^2.

Moreover,

⟨x,T∗Tx⟩=∥Tx∥2≥0.\langle x,T^*Tx\rangle=\|Tx\|^2\geq 0.

The operator T∗TT^*T is therefore self-adjoint and positive.

Adjoints also express the relationship between an operator’s kernel and range. Writing ran⁡T\operatorname{ran}T for the range,

ker⁡T∗=(ran⁡T)⊥,ker⁡T=(ran⁡T∗)⊥,\ker T^*=(\operatorname{ran}T)^\perp, \qquad \ker T=(\operatorname{ran}T^*)^\perp,

where ⊥{}^\perp denotes the orthogonal complement. Equivalently,

ran⁡T‾=(ker⁡T∗)⊥,ran⁡T∗‾=(ker⁡T)⊥.\overline{\operatorname{ran}T}=(\ker T^*)^\perp, \qquad \overline{\operatorname{ran}T^*}=(\ker T)^\perp.

The closures cannot generally be omitted in infinite dimensions. For example, TT has dense range exactly when T∗T^* is injective.

Important classes of operators

Several major classes of operators are defined through their adjoints.

  • A self-adjoint operator satisfies T=T∗T=T^*.
  • A skew-adjoint operator satisfies T∗=−TT^*=-T.
  • A normal operator satisfies T∗T=TT∗T^*T=TT^*.
  • A unitary operator U:H→KU:H\to K satisfies U∗U=IHU^*U=I_H and UU∗=IKUU^*=I_K.
  • An isometry satisfies T∗T=IHT^*T=I_H, but need not be onto.
  • An orthogonal projection satisfies P2=P=P∗P^2=P=P^*.

For real Hilbert spaces, surjective inner-product-preserving operators are usually called orthogonal rather than unitary.

These identities connect adjoints to the spectral theorem. In finite-dimensional complex inner-product spaces, normal operators admit an orthonormal basis of eigenvectors. Self-adjoint operators have real eigenvalues, while unitary operators have eigenvalues of absolute value one.

For a general matrix AA, the singular value decomposition is governed by the positive operators A∗AA^*A and AA∗AA^*. Their nonzero eigenvalues agree, and their square roots are the nonzero singular values of AA.

Examples

Multiplication operators

On the space L2(X,μ)L^2(X,\mu), with

⟨f,g⟩=∫Xf g‾ dμ,\langle f,g\rangle=\int_X f\,\overline g\,d\mu,

let Mmf=mfM_mf=mf, where mm is an essentially bounded measurable function. Then

Mm∗=Mm‾.M_m^*=M_{\overline m}.

In particular, MmM_m is self-adjoint exactly when mm is real-valued almost everywhere.

Integral operators

Suppose XX and YY are finite measure spaces and k∈L2(Y×X)k\in L^2(Y\times X). The integral operator

(Tf)(y)=∫Xk(y,x)f(x) dμ(x)(Tf)(y)=\int_X k(y,x)f(x)\,d\mu(x)

defines a bounded map from L2(X,μ)L^2(X,\mu) to L2(Y,ν)L^2(Y,\nu). Its adjoint is

(T∗g)(x)=∫Yk(y,x)‾ g(y) dν(y).(T^*g)(x)=\int_Y\overline{k(y,x)}\,g(y)\,d\nu(y).

Thus taking an adjoint exchanges the input and output variables and conjugates the kernel, paralleling the conjugate transpose of a matrix.

The unilateral shift

On the Hilbert space ℓ2(N0)\ell^2(\mathbb N_0), define

S(x0,x1,x2,…)=(0,x0,x1,…).S(x_0,x_1,x_2,\ldots)=(0,x_0,x_1,\ldots).

Then

S∗(x0,x1,x2,…)=(x1,x2,x3,…).S^*(x_0,x_1,x_2,\ldots)=(x_1,x_2,x_3,\ldots).

Here S∗S=IS^*S=I, but SS∗≠ISS^*\neq I: the latter removes the zeroth coordinate. This illustrates why an isometry need not be unitary in infinite dimensions.

Unbounded operators and their domains

Many operators arising from differential equations are unbounded. Such an operator is specified not merely by a rule but also by a linear domain:

T:D(T)⊆H⟶K.T:D(T)\subseteq H\longrightarrow K.

Assume D(T)D(T) is dense in HH. The domain of the adjoint is

D(T∗)={y∈K:there exists z∈H such that ⟨Tx,y⟩K=⟨x,z⟩H for all x∈D(T)}.D(T^*)= \left\{ y\in K: \text{there exists }z\in H \text{ such that } \langle Tx,y\rangle_K=\langle x,z\rangle_H \text{ for all }x\in D(T) \right\}.

Density makes zz unique, and T∗yT^*y is defined to be this vector.

Equivalently, y∈D(T∗)y\in D(T^*) precisely when

x⟼⟨Tx,y⟩Kx\longmapsto\langle Tx,y\rangle_K

is bounded on D(T)D(T) with respect to the norm of HH.

Unlike the bounded case, D(T∗)D(T^*) need not be all of KK, or even dense. The adjoint is always a closed operator. A densely defined operator TT is closable exactly when D(T∗)D(T^*) is dense; in that case,

T∗∗=T‾,T^{**}=\overline T,

where T‾\overline T denotes its closure.

For unbounded operators, equality includes equality of domains. Merely showing

⟨Tx,y⟩=⟨x,Ty⟩(x,y∈D(T))\langle Tx,y\rangle=\langle x,Ty\rangle \qquad(x,y\in D(T))

establishes that TT is symmetric, meaning T⊆T∗T\subseteq T^*. Self-adjointness requires the stronger statement

D(T)=D(T∗),Tx=T∗x(x∈D(T)).D(T)=D(T^*), \qquad Tx=T^*x\quad(x\in D(T)).

This distinction is essential in infinite-dimensional spectral theory.

Differentiation and boundary conditions

Consider

Tf=f′Tf=f'

on L2(0,1)L^2(0,1), with

D(T)=H01(0,1),D(T)=H_0^1(0,1),

the Sobolev space of H1H^1 functions whose endpoint traces vanish. By integration by parts,

∫01f′(x)g(x)‾ dx=[f(x)g(x)‾]01−∫01f(x)g′(x)‾ dx.\int_0^1 f'(x)\overline{g(x)}\,dx = [f(x)\overline{g(x)}]_0^1 - \int_0^1 f(x)\overline{g'(x)}\,dx.

The boundary term vanishes for every f∈D(T)f\in D(T), yielding

T∗g=−g′,D(T∗)=H1(0,1).T^*g=-g', \qquad D(T^*)=H^1(0,1).

The adjoint therefore has a different domain from the original operator. If instead f′f' is defined on all of H1(0,1)H^1(0,1), its adjoint is −g′-g' on H01(0,1)H_0^1(0,1).

A formal adjoint is the differential expression obtained by integration by parts before domains and boundary conditions are fully specified. It need not be the complete Hilbert-space adjoint: the latter includes its precise domain.

Relation to dual-space adjoints

There is a related construction that does not require an inner product. For a bounded linear map T:X→YT:X\to Y between Banach spaces, its transpose, also called its Banach-space adjoint, is

T′:Y′→X′,T′φ=φ∘T,T':Y'\to X', \qquad T'\varphi=\varphi\circ T,

where X′X' and Y′Y' are the continuous dual spaces.

For complex spaces, this transpose is linear in TT:

(λT)′=λT′,(\lambda T)'=\lambda T',

whereas the Hilbert-space adjoint satisfies

(λT)∗=λ‾ T∗.(\lambda T)^*=\overline{\lambda}\,T^*.

The two constructions are connected by the Riesz identifications. Under the inner-product convention used here, define the conjugate-linear maps

JHx=⟨ ⋅ ,x⟩H,JKy=⟨ ⋅ ,y⟩K.J_Hx=\langle\,\cdot\,,x\rangle_H, \qquad J_Ky=\langle\,\cdot\,,y\rangle_K.

Then

T∗=JH−1T′JK.T^*=J_H^{-1}T'J_K.

The conjugate-linearity of these identifications explains why the Hilbert-space adjoint introduces complex conjugation.

Applications

Least squares and optimization

For a bounded operator T:H→KT:H\to K and data b∈Kb\in K, consider the least-squares objective

F(x)=12∥Tx−b∥2.F(x)=\frac12\|Tx-b\|^2.

Its gradient is

∇F(x)=T∗(Tx−b).\nabla F(x)=T^*(Tx-b).

In a complex Hilbert space, this statement uses the underlying real inner product Re⁡⟨⋅,⋅⟩\operatorname{Re}\langle\cdot,\cdot\rangle.

Any minimizer satisfies the normal equation

T∗Tx=T∗b.T^*Tx=T^*b.

Conversely, every solution of this equation minimizes FF. In infinite dimensions, a minimizer need not exist if the range of TT is not closed.

Adjoint-state methods extend the same principle to optimization constrained by partial differential equations. They transfer derivatives through the governing equations, allowing gradients of scalar objectives to be computed without separately differentiating the full state with respect to every parameter.

Quantum mechanics

In quantum mechanics, observables are represented by self-adjoint operators, often unbounded ones. Domain conditions are part of their mathematical specification: a formally symmetric differential expression alone does not determine a self-adjoint observable. Unitary evolution also uses the adjoint through U∗=U−1U^*=U^{-1}.

Numerical computation

Algorithms for large-scale inverse problems often require only the actions x↦Txx\mapsto Tx and y↦T∗yy\mapsto T^*y, rather than an explicitly stored matrix. An implementation can be checked using the adjoint identity

⟨Tx,y⟩K=⟨x,T∗y⟩H\langle Tx,y\rangle_K=\langle x,T^*y\rangle_H

on test vectors, allowing for floating-point error. If discretized spaces use weighted inner products, those weights must be included; an ordinary matrix transpose may not represent the intended adjoint.