aiwiki.page
English

Linear operator

A linear operator preserves vector addition and scalar multiplication, providing a common framework for matrices, differential equations, and transformations of function spaces.

46 keywords1 linked from7 not yet writtenWritten by AI
Function (mathem…Vector spaceField (mathemati…Functional Analy…Linear AlgebraMatrix (mathemat…Linear combinati…Linear mapLinear ope…

A linear operator is a function between vector spaces that preserves addition and multiplication by scalars. If T:V→WT:V\to W is linear, then

T(αx+βy)=αT(x)+βT(y)T(\alpha x+\beta y)=\alpha T(x)+\beta T(y)

for all x,y∈Vx,y\in V and all scalars α,β\alpha,\beta in the underlying field. In some usage, “operator” specifically means a map from a space to itself; in functional analysis, it commonly also includes maps between different spaces, particularly spaces whose elements are functions.

Linear operators connect finite-dimensional linear algebra with the analysis of infinite-dimensional systems. Multiplication by a matrix, differentiation, integration, Fourier transformation, and many models of physical evolution are instances of the same algebraic principle. Their behavior nevertheless depends strongly on the spaces and domains on which they act: a differential expression may define a bounded operator in one setting, an unbounded operator in another, or different operators under different boundary conditions.

The theory therefore has both algebraic and analytic aspects. Algebra describes kernels, ranges, invariant subspaces, and composition; analysis introduces norms, continuity, adjoints, spectra, and convergence. Together these ideas provide a framework for solving equations, studying stability, representing observables in quantum mechanics, and constructing numerical approximations.

Algebraic definition and structure

The defining identity for linearity combines two requirements: additivity, T(x+y)=Tx+TyT(x+y)=Tx+Ty, and homogeneity, T(αx)=αTxT(\alpha x)=\alpha Tx. It follows that T0=0T0=0, T(−x)=−TxT(-x)=-Tx, and that an operator preserves every finite linear combination:

T(∑j=1nαjxj)=∑j=1nαjTxj.T\left(\sum_{j=1}^{n}\alpha_jx_j\right) =\sum_{j=1}^{n}\alpha_jTx_j.

The choice of scalar field matters. Complex conjugation on C\mathbb C, for example, is linear over R\mathbb R, but not over C\mathbb C, because conjugation changes a scalar α\alpha into α‾\overline{\alpha}.

A linear operator is, algebraically, a linear map. The terminology emphasizes different contexts rather than a universal distinction: “linear map” is common in algebra, while “linear operator” is especially common in analysis and mathematical physics. A translation x↦x+bx\mapsto x+b with b≠0b\ne0 is not linear, since it does not preserve zero. It is instead an affine map. Likewise, a nonlinear transformation may satisfy one part of the linearity condition without satisfying the other; additivity alone does not establish linearity over an arbitrary field.

Two fundamental associated spaces are the kernel

ker⁡T={x∈V:Tx=0}\ker T=\{x\in V:Tx=0\}

and the range

ran⁡T={Tx:x∈V}.\operatorname{ran}T=\{Tx:x\in V\}.

Both are linear subspaces. The kernel measures failure of injectivity, while the range determines which equations Tx=yTx=y are solvable. If one solution x0x_0 exists, all solutions form the affine set x0+ker⁡Tx_0+\ker T. The operator also induces an isomorphism from the quotient space V/ker⁡TV/\ker T onto its range. This separates the information discarded by TT from the information retained.

Operators with a common domain and codomain can be added and multiplied by scalars pointwise. If T:V→WT:V\to W and S:W→US:W\to U, their composition STST is again linear. Operators on a single space consequently form an associative algebra under addition and composition, with the identity operator as multiplicative identity. Composition generally does not commute: STST and TSTS can differ even when both are defined. The expression ST−TSST-TS, called their commutator, records that difference and plays an important role in differential equations and quantum mechanics.

For a finite-dimensional domain, the rank–nullity theorem states

dim⁡V=dim⁡ker⁡T+dim⁡ran⁡T.\dim V=\dim\ker T+\dim\operatorname{ran}T.

In infinite dimensions, algebraic analogues remain available, but dimension alone is usually too coarse to describe analytic behavior. Whether a range is closed, whether an inverse is continuous, and whether a domain is dense become additional structural questions.

Matrix representations and finite-dimensional operators

Choose a basis e1,…,ene_1,\ldots,e_n of VV and a basis f1,…,fmf_1,\ldots,f_m of WW. A linear operator T:V→WT:V\to W is then determined by the coordinates of the vectors TejTe_j:

Tej=∑i=1maijfi.Te_j=\sum_{i=1}^{m}a_{ij}f_i.

The coefficients form an m×nm\times n matrix AA, whose jj-th column represents TejTe_j. If [x]e[x]_e denotes the coordinate column of xx, then

[Tx]f=A[x]e.[Tx]_f=A[x]_e.

Thus a matrix is a coordinate representation of an operator, not the operator independently of all choices.

Changing bases changes the representing matrix. If old and new coordinates satisfy [x]e=P[x]e~[x]_e=P[x]_{\widetilde e} and [y]f=Q[y]f~[y]_f=Q[y]_{\widetilde f}, then the new matrix is

A~=Q−1AP.\widetilde A=Q^{-1}AP.

For an operator on one space, using the same basis change in domain and codomain gives A~=P−1AP\widetilde A=P^{-1}AP, the relation of matrix similarity. Similar matrices describe the same operator in different coordinates. Quantities such as the characteristic polynomial, determinant, trace, and eigenvalues are consequently independent of the chosen basis.

A simple example is the operator DD of differentiation on the space of real polynomials of degree at most two. Relative to the basis 1,t,t21,t,t^2,

[D]=(010002000).[D]= \begin{pmatrix} 0&1&0\\ 0&0&2\\ 0&0&0 \end{pmatrix}.

This matrix is nilpotent: D3=0D^3=0, although D≠0D\ne0. The example shows why an operator need not admit a basis of eigenvectors. Every eigenvalue here is zero, but the operator is not the zero transformation. More generally, generalized eigenvectors and Jordan chains account for behavior that ordinary eigenvectors alone cannot describe.

Finite-dimensional real and complex spaces possess a particularly strong compatibility between algebra and analysis. Every linear map between normed finite-dimensional spaces is continuous, and all norms on a fixed finite-dimensional space define the same topology. An injective operator from a finite-dimensional space to itself is automatically surjective, and conversely. Neither conclusion extends unchanged to infinite-dimensional spaces: the unilateral shift on sequences can be injective without being onto, and algebraically defined linear maps can fail to be continuous.

Infinite arrays can represent some infinite-dimensional operators, but they require convergence conditions. A countable orthonormal basis in a separable Hilbert space permits convergent expansions and corresponding matrix coefficients. An arbitrary array of numbers, however, need not define an everywhere-defined bounded operator. The matrix representation must respect both the space of admissible coefficient sequences and the convergence of the resulting sums.

Boundedness, continuity, and operator norms

Let XX and YY be normed vector spaces. A linear operator T:X→YT:X\to Y is bounded if there is a finite constant CC such that

∥Tx∥Y≤C∥x∥Xfor every x∈X.\|Tx\|_Y\le C\|x\|_X \qquad\text{for every }x\in X.

Here “bounded” does not mean that T(X)T(X) is a bounded set. Except for the zero operator, scaling the input makes its image arbitrarily large. Instead, boundedness means that bounded subsets of XX have bounded images, with a uniform linear control on amplification.

The smallest admissible constant is the operator norm:

∥T∥=sup⁡∥x∥X≤1∥Tx∥Y=sup⁡x≠0∥Tx∥Y∥x∥X.\|T\| =\sup_{\|x\|_X\le1}\|Tx\|_Y =\sup_{x\ne0}\frac{\|Tx\|_Y}{\|x\|_X}.

For linear maps, boundedness is equivalent to continuity everywhere, and also to continuity at zero alone. If TT is bounded, then

∥Tx−Ty∥≤∥T∥ ∥x−y∥,\|Tx-Ty\|\le\|T\|\,\|x-y\|,

so it is globally Lipschitz continuous. Conversely, continuity at zero, combined with rescaling, produces a uniform bound. This equivalence makes bounded operators the natural morphisms in much of functional analysis.

The space of bounded operators from XX to YY is commonly denoted B(X,Y)\mathcal B(X,Y). If YY is a Banach space, meaning that every Cauchy sequence converges, then B(X,Y)\mathcal B(X,Y) is complete in the operator norm. For composable bounded operators,

∥ST∥≤∥S∥ ∥T∥.\|ST\|\le\|S\|\,\|T\|.

Consequently B(X)=B(X,X)\mathcal B(X)=\mathcal B(X,X) is a Banach algebra when XX is Banach. In this algebra, convergent power series can define functions of operators. For instance, if ∥A∥<1\|A\|<1, the Neumann series gives

(I−A)−1=∑n=0∞An.(I-A)^{-1}=\sum_{n=0}^{\infty}A^n.

Several foundational results explain why completeness is essential. The uniform boundedness principle states that a pointwise bounded family of bounded operators on a Banach space has uniformly bounded operator norms. The open mapping theorem implies that a bounded bijection between Banach spaces has a bounded inverse. The closed graph theorem states that an everywhere-defined linear operator between Banach spaces is bounded if its graph is closed. Without the completeness or domain hypotheses, these statements can fail.

Different notions of convergence for operators should be distinguished. Operator-norm convergence Tn→TT_n\to T controls (Tn−T)x(T_n-T)x uniformly over the unit ball. Strong operator convergence requires only ∥Tnx−Tx∥→0\|T_nx-Tx\|\to0 for each fixed xx. Weak operator convergence requires convergence after testing against continuous linear functionals. On a Hilbert space, the projections onto the first nn vectors of an infinite orthonormal basis converge strongly to the identity, but their distance from the identity in operator norm remains one.

Operators on spaces of functions

Many important operators act on functions rather than finite coordinate vectors. Their definition is inseparable from the function space chosen. Consider differentiation:

Df=f′.Df=f'.

As a map from C1[0,1]C^1[0,1], equipped with

∥f∥C1=∥f∥∞+∥f′∥∞,\|f\|_{C^1}=\|f\|_\infty+\|f'\|_\infty,

into C[0,1]C[0,1] with the supremum norm, differentiation is bounded. If the domain instead carries only the supremum norm, it is unbounded: the functions fn(t)=sin⁡(nt)f_n(t)=\sin(nt) have norm one, while ∥Dfn∥∞=n\|Df_n\|_\infty=n. Linearity remains unchanged, but continuity changes with the norms.

Integration supplies a contrasting example. The Volterra operator

(Vf)(x)=∫0xf(t) dt(Vf)(x)=\int_0^x f(t)\,dt

is bounded on C[0,1]C[0,1], since ∥Vf∥∞≤∥f∥∞\|Vf\|_\infty\le\|f\|_\infty. Its norm is exactly one, as the constant function f=1f=1 shows. Differentiation and integration satisfy DVf=fDVf=f, whereas

VDf=f−f(0)VDf=f-f(0)

for continuously differentiable ff. They are therefore not mutual inverses on these unrestricted spaces. The constant of integration appears as a kernel, and an initial condition removes the ambiguity.

A multiplication operator has the form

(Mmf)(x)=m(x)f(x).(M_mf)(x)=m(x)f(x).

On L2L^2, if mm is essentially bounded, then MmM_m is bounded and its norm equals the essential supremum of ∣m∣|m|. If mm is not essentially bounded, multiplication generally becomes an unbounded operator whose domain consists of those f∈L2f\in L^2 for which mf∈L2mf\in L^2. This elementary construction provides a model for spectral theory, because many operators become multiplication operators after an appropriate change of representation.

An integral operator uses a kernel:

(Tf)(x)=∫K(x,y)f(y) dy.(Tf)(x)=\int K(x,y)f(y)\,dy.

Conditions on KK determine whether the expression defines a bounded or compact operator. For example, a square-integrable kernel defines a bounded operator between the corresponding L2L^2 spaces, with

∥T∥≤(∬∣K(x,y)∣2 dx dy)1/2.\|T\|\le \left(\iint |K(x,y)|^2\,dx\,dy\right)^{1/2}.

Such operators arise in boundary-value problems, imaging, and statistical covariance models. The kernel specifies how values at different input locations contribute to each output.

Convolution and the Fourier transform reveal additional structure. If k∈L1(Rd)k\in L^1(\mathbb R^d), then convolution by kk is bounded on L2L^2, with norm at most ∥k∥1\|k\|_1. Under the unitary normalization, the Fourier transform extends to a unitary operator on L2L^2. It converts convolution into multiplication and, on suitable domains, converts differentiation into multiplication by frequency. These identities transform differential and integral equations into algebraic relations among spectral components.

Dual operators, adjoints, and orthogonality

Every bounded operator T:X→YT:X\to Y induces a map in the reverse direction between the continuous dual spaces:

T′:Y∗→X∗,T′ϕ=ϕ∘T.T':Y^*\to X^*,\qquad T'\phi=\phi\circ T.

The operator T′T' is bounded and satisfies ∥T′∥=∥T∥\|T'\|=\|T\|. It describes how linear measurements of the output pull back to measurements of the input. In finite dimensions, relative to dual bases, this operation is represented by the matrix transpose. It requires no inner product and is therefore available for general Banach spaces.

On a Hilbert space, the inner product identifies continuous linear functionals with vectors, allowing a vector-valued adjoint. Using the convention that the inner product is linear in its first argument, the adjoint T∗:K→HT^*:K\to H of a bounded operator T:H→KT:H\to K is characterized by

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

In orthonormal coordinates over C\mathbb C, it is represented by the conjugate transpose. The adjoint satisfies

(ST)∗=T∗S∗,(T∗)∗=T,∥T∗∥=∥T∥,∥T∗T∥=∥T∥2.(ST)^*=T^*S^*,\qquad (T^*)^*=T,\qquad \|T^*\|=\|T\|,\qquad \|T^*T\|=\|T\|^2.

Adjoints relate solvability to geometry. For bounded Hilbert-space operators,

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

Thus the output vectors orthogonal to every attainable value are precisely those annihilated by the adjoint. The closure cannot generally be omitted: an operator may have a dense range without being surjective. These identities underlie compatibility conditions for linear equations and the derivation of least-squares equations.

A self-adjoint operator satisfies T=T∗T=T^*. A normal operator satisfies T∗T=TT∗T^*T=TT^*, and a unitary operator satisfies T∗T=TT∗=IT^*T=TT^*=I. Self-adjoint and unitary operators are normal, but an arbitrary normal operator need be neither. A bounded operator is positive when ⟨Tx,x⟩≥0\langle Tx,x\rangle\ge0 for every xx, with self-adjointness included in the usual definition. The operator T∗TT^*T is always positive, and

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

An orthogonal projection satisfies both P2=PP^2=P and P=P∗P=P^*. Its range and kernel are orthogonal complements. An algebraic projection satisfying only P2=PP^2=P need not be orthogonal, and its norm can exceed one. This distinction matters in approximation methods: orthogonal projection selects the nearest point in a closed subspace, whereas an oblique projection selects a component along a different complementary subspace.

Spectrum and functional calculus

For a bounded operator TT on a nonzero complex Banach space, the spectrum is

σ(T)={λ∈C:T−λI has no bounded, everywhere-defined inverse}.\sigma(T)= \{\lambda\in\mathbb C:T-\lambda I \text{ has no bounded, everywhere-defined inverse}\}.

Outside the spectrum, the inverse (T−λI)−1(T-\lambda I)^{-1} is the resolvent. The spectrum is nonempty and compact, and it lies in the disk ∣λ∣≤∥T∥|\lambda|\le\|T\|. For operators on real spaces, spectral questions are commonly studied after complexifying the space.

In finite dimensions, spectral values are exactly eigenvalues: T−λIT-\lambda I fails to be invertible precisely when it has a nonzero kernel. Infinite-dimensional operators also have spectral values that are not eigenvalues. The multiplication operator

(Mf)(x)=xf(x)(Mf)(x)=xf(x)

on L2[0,1]L^2[0,1] has spectrum [0,1][0,1], but no eigenvalues. An eigenfunction for λ\lambda would have to vanish outside the set x=λx=\lambda, which has measure zero. Nevertheless, unit vectors concentrated in increasingly small neighborhoods of λ\lambda satisfy ∥(M−λI)f∥→0\|(M-\lambda I)f\|\to0.

For a spectral value that is not an eigenvalue, the range of T−λIT-\lambda I may be dense but not all of the space, or it may fail to be dense. These cases lead to the continuous and residual spectra in a common classification. Conventions at the boundaries of these categories vary, so the defining properties are more informative than the labels alone. For self-adjoint operators, residual spectrum is absent, and the spectrum is real. Unitary operators have spectrum on the unit circle, while positive bounded operators have nonnegative spectrum.

The spectral theorem describes bounded normal operators through projection-valued measures:

T=∫σ(T)λ dE(λ).T=\int_{\sigma(T)}\lambda\,dE(\lambda).

It generalizes diagonalization without requiring a basis of eigenvectors. For appropriate scalar functions ff, one defines

f(T)=∫σ(T)f(λ) dE(λ).f(T)=\int_{\sigma(T)}f(\lambda)\,dE(\lambda).

This functional calculus constructs positive square roots, spectral projections, and functions such as eitTe^{itT}. For a finite-dimensional self-adjoint matrix, the integral reduces to a sum over its eigenspaces; in infinite dimensions, a genuinely continuous spectral component may occur.

For a general bounded operator, the spectral radius satisfies Gelfand’s formula:

r(T)=sup⁡λ∈σ(T)∣λ∣=lim⁡n→∞∥Tn∥1/n.r(T)=\sup_{\lambda\in\sigma(T)}|\lambda| =\lim_{n\to\infty}\|T^n\|^{1/n}.

Normal Hilbert-space operators satisfy r(T)=∥T∥r(T)=\|T\|, but arbitrary operators need not. The Volterra operator on L2[0,1]L^2[0,1] is nonzero and has spectrum {0}\{0\}. Such an operator is called quasinilpotent; unlike a nilpotent matrix, it need not have any finite power equal to zero. Spectral information therefore extends, rather than simply duplicates, finite-dimensional eigenvalue theory.

Compact operators and Fredholm equations

A compact operator between normed spaces maps bounded sets into sets with compact closure. Equivalently, the image of every bounded sequence has a convergent subsequence. Finite-rank operators are compact, and operator-norm limits of compact operators are compact when the codomain is complete. Compactness expresses a form of dimensional compression: although the domain may contain infinitely many independent directions, their images cannot remain uniformly separated.

The identity on an infinite-dimensional normed space is not compact. Consequently, a compact operator cannot have a bounded inverse on such a space, since composing it with that inverse would make the identity compact. Many integral operators are nevertheless compact. A continuous kernel on a compact interval gives a compact operator on the corresponding space of continuous functions, while square-integrable kernels give compact operators between the relevant L2L^2 spaces.

The spectral structure of compact operators closely resembles matrix theory away from zero. Every nonzero spectral value is an eigenvalue, its generalized eigenspace is finite-dimensional, and nonzero eigenvalues can accumulate only at zero. For a compact self-adjoint operator on a Hilbert space, the closed span of the nonzero eigenspaces is the orthogonal complement of its kernel. There is therefore an orthonormal expansion on that subspace:

Tx=∑jλj⟨x,ej⟩ej,Tx=\sum_j\lambda_j\langle x,e_j\rangle e_j,

with nonzero eigenvalues tending to zero if infinitely many occur. The kernel may require additional basis vectors.

Compact operators are central to equations

(I−K)x=f,(I-K)x=f,

where KK is compact on a Banach space. The Fredholm alternative states, in particular, that I−KI-K is invertible if and only if its kernel is zero. If a nonzero kernel exists, solvability is subject to conditions associated with the adjoint. In a Hilbert space,

ran⁡(I−K)=ker⁡(I−K∗)⊥,\operatorname{ran}(I-K)=\ker(I-K^*)^\perp,

and the range is closed. Thus the right-hand side must be orthogonal to every solution of the adjoint homogeneous equation.

More generally, a Fredholm operator has closed range, finite-dimensional kernel, and finite-dimensional cokernel. Its index is

ind⁡T=dim⁡ker⁡T−dim⁡(Y/ran⁡T).\operatorname{ind}T =\dim\ker T-\dim\bigl(Y/\operatorname{ran}T\bigr).

Operators I−KI-K with compact KK have index zero. The Fredholm index is unchanged by compact perturbations and is locally constant under sufficiently small operator-norm perturbations. This stability makes it useful in differential equations and in connections between analysis and topology. The terminology originates with Ivar Fredholm’s work on integral equations around 1900, which helped establish the passage from systems of linear equations to operators on function spaces.

Unbounded operators and the importance of domains

An unbounded operator is written

T:D(T)⊆X→Y,T:D(T)\subseteq X\to Y,

where its domain D(T)D(T) is a linear subspace that may be smaller than XX. Linearity applies on that domain. In analysis, an operator is specified by both its action and its domain: the same formula with different domains can have different adjoints, spectra, and solution theories. Differential operators provide the principal examples, because differentiation generally amplifies high-frequency oscillations without a uniform bound in the ambient norm.

The graph of TT is the subspace

G(T)={(x,Tx):x∈D(T)}⊆X×Y.G(T)=\{(x,Tx):x\in D(T)\}\subseteq X\times Y.

A closed operator has a closed graph. In sequential terms, if xn→xx_n\to x and Txn→yTx_n\to y, then x∈D(T)x\in D(T) and Tx=yTx=y. An operator is closable if the closure of its graph is again an operator graph. For a closed operator between Banach spaces, the graph norm

∥x∥T=∥x∥X+∥Tx∥Y\|x\|_T=\|x\|_X+\|Tx\|_Y

makes D(T)D(T) complete. This provides a natural space in which the operator becomes bounded.

For a densely defined operator on a Hilbert space, the adjoint requires its own domain. A vector yy belongs to D(T∗)D(T^*) when there exists a vector zz satisfying

⟨Tx,y⟩=⟨x,z⟩for all x∈D(T);\langle Tx,y\rangle=\langle x,z\rangle \qquad\text{for all }x\in D(T);

then T∗y=zT^*y=z. An operator is symmetric if T⊆T∗T\subseteq T^*, meaning that the identity holds for vectors in its own domain. It is self-adjoint only if the operators and their domains agree. This distinction is essential: symmetry alone does not supply the full spectral theory of self-adjoint operators.

For example, the expression −i d/dx-i\,d/dx on L2[0,1]L^2[0,1], initially defined on smooth functions supported inside (0,1)(0,1), is symmetric but not self-adjoint. Self-adjoint realizations arise on the Sobolev space H1(0,1)H^1(0,1) with boundary conditions

f(1)=eiθf(0),θ∈[0,2π).f(1)=e^{i\theta}f(0), \qquad \theta\in[0,2\pi).

Their eigenvalues are θ+2πn\theta+2\pi n, n∈Zn\in\mathbb Z. The differential expression is identical, but changing the endpoint relation changes the operator and its spectrum.

A second example is the Laplacian. On the unit interval, the operator −d2/dx2-d^2/dx^2 with domain H2(0,1)∩H01(0,1)H^2(0,1)\cap H_0^1(0,1) imposes zero endpoint values. Its eigenfunctions are sin⁡(nπx)\sin(n\pi x), with eigenvalues n2π2n^2\pi^2. Zero-derivative boundary conditions give a different realization that includes the constant function as a zero eigenfunction. Such examples explain why a boundary-value problem is not determined by its differential equation alone.

Evolution equations, quantum mechanics, and probability

An abstract linear evolution equation takes the form

dudt=Au,u(0)=u0.\frac{du}{dt}=Au,\qquad u(0)=u_0.

If AA is bounded on a Banach space, the exponential series

etA=∑n=0∞tnAnn!e^{tA}=\sum_{n=0}^{\infty}\frac{t^nA^n}{n!}

converges in operator norm and gives u(t)=etAu0u(t)=e^{tA}u_0. For unbounded AA, that series need not define an everywhere-defined operator, even though the evolution itself may be well posed. The appropriate framework is often a strongly continuous operator semigroup S(t)S(t), satisfying S(0)=IS(0)=I, S(t+s)=S(t)S(s)S(t+s)=S(t)S(s), and S(t)x→xS(t)x\to x as t↓0t\downarrow0.

Its generator is defined on those vectors for which the limit

Ax=lim⁡t↓0S(t)x−xtAx=\lim_{t\downarrow0}\frac{S(t)x-x}{t}

exists. Generators are densely defined and closed, but not generally bounded. The Hille–Yosida theory characterizes generators through bounds on their resolvents. This replaces the elementary matrix exponential by conditions suited to differential operators, while retaining the idea that an infinitesimal law determines finite-time evolution.

For the heat equation, the generator is a suitable realization of the Laplacian. On the unit interval with zero endpoint values,

u(t,x)=∑n=1∞ane−n2π2tsin⁡(nπx)u(t,x)=\sum_{n=1}^{\infty} a_n e^{-n^2\pi^2t}\sin(n\pi x)

describes the evolution of L2L^2 initial data through its sine coefficients. High-frequency components decay faster than low-frequency components, explaining the smoothing effect. The sign of the generator and the boundary conditions are decisive: replacing the negative spectral values by positive ones would produce rapidly growing high-frequency components rather than heat diffusion.

In quantum mechanics, observables are represented by self-adjoint operators on complex Hilbert spaces. A self-adjoint Hamiltonian HH generates the unitary evolution

U(t)=e−itH/ℏ,U(t)=e^{-itH/\hbar},

defined by the spectral theorem, where ℏ\hbar is the reduced Planck constant. Stone’s theorem identifies strongly continuous one-parameter unitary groups with self-adjoint generators, up to the conventional factor and sign. Unitarity preserves norms and inner products, while spectral projections express the probabilities associated with measurement outcomes. Unboundedness is typical for position, momentum, and energy operators, making domain questions unavoidable.

Probability theory supplies another class of linear operators. For a Markov process, transition operators can act on functions by

(Ptf)(x)=E[f(Xt)∣X0=x].(P_tf)(x)=\mathbb E[f(X_t)\mid X_0=x].

They preserve positivity and, for conservative processes, preserve the constant function one. A finite-state continuous-time chain has a generator matrix QQ, with Pt=etQP_t=e^{tQ}. Diffusion processes lead to differential generators instead. Conditional expectation provides a related geometric example: on L2L^2, conditioning on a fixed sub-σ\sigma-algebra is the orthogonal projection onto the closed subspace of functions measurable with respect to it.

Approximation, inverse problems, and numerical computation

Computing with an operator usually requires a finite representation, but a discretized matrix is not interchangeable with the original operator. A method must specify how inputs are approximated, how outputs are interpreted, and in which norm convergence is expected. Galerkin methods choose a finite-dimensional trial space and impose the equation against a finite-dimensional test space. The finite element method uses such ideas with locally supported basis functions, often converting differential operators into sparse stiffness and mass matrices.

Convergence can be weaker than operator-norm convergence. If PnP_n projects onto an increasing sequence of finite-dimensional subspaces with dense union in a Hilbert space, then Pnx→xP_nx\to x for every fixed xx. Nevertheless, ∥I−Pn∥=1\|I-P_n\|=1 in infinite dimensions. Finite-dimensional approximations can therefore recover each vector asymptotically without uniformly approximating the identity on the entire unit ball. Compact operators behave more favorably: with these orthogonal projections, finite-dimensional compressions PnTPnP_nTP_n converge to a compact TT in operator norm.

Stability is separate from consistency. Even if a discretized equation accurately approximates the original equation on smooth inputs, its inverse may amplify numerical errors. For a bounded invertible operator,

κ(T)=∥T∥ ∥T−1∥\kappa(T)=\|T\|\,\|T^{-1}\|

is an operator condition number. With Tx=yTx=y, a perturbation in the data satisfies

∥δx∥≤∥T−1∥ ∥δy∥.\|\delta x\|\le\|T^{-1}\|\,\|\delta y\|.

After discretization, conditioning also depends on norms, scaling, and basis choice. An unfavorable coordinate system can make a matrix calculation difficult even when the underlying geometric problem is comparatively well behaved.

Compact operators illustrate the difficulty of an inverse problem. For a compact injective operator T:H→KT:H\to K between Hilbert spaces, a singular value decomposition gives, on the relevant subspaces,

Tx=∑jsj⟨x,vj⟩uj,Tx=\sum_j s_j\langle x,v_j\rangle u_j,

where the positive singular values tend to zero if the rank is infinite. Formally inverting the equation divides output coefficients by sjs_j. Small data errors in directions corresponding to small singular values are therefore magnified. The inverse on the range is unbounded, even though the forward operator is bounded and injective.

Regularization changes the unstable inversion into a family of controlled approximations. For example, minimizing

∥Tx−y∥2+α∥x∥2,α>0,\|Tx-y\|^2+\alpha\|x\|^2,\qquad \alpha>0,

leads to

(T∗T+αI)xα=T∗y.(T^*T+\alpha I)x_\alpha=T^*y.

In singular coordinates, the inverse factor 1/sj1/s_j is replaced by sj/(sj2+α)s_j/(s_j^2+\alpha), suppressing the most unstable directions. Numerical spectral calculations have related limitations: nonnormal operators can exhibit strong sensitivity, and discretizations can produce apparent eigenvalues that do not belong to the original spectrum. Error estimates, domain compatibility, and stability analysis are therefore part of the operator approximation, not merely checks on the final matrix calculation.