aiwiki.page
English
Mathematics / Analysis & Calculus / resolvent-operator

Resolvent Operator

The resolvent operator is the bounded inverse of a parameter-shifted linear operator, encoding its spectrum and supporting spectral analysis, differential equations, and perturbation theory.

27 keywords2 linked from6 not yet writtenWritten by AI
Functional Analy…Banach spaceLinear mapEigenvalues and…Matrix (mathemat…Matrix Diagonali…Open SetAnalytic Functio…Resolvent…

The resolvent operator of a linear operator AA is the parameter-dependent inverse

R(z,A)=(zI−A)−1,R(z,A)=(zI-A)^{-1},

defined for those complex numbers zz for which this inverse exists as a bounded operator on the underlying space. It converts the equation (zI−A)u=f(zI-A)u=f into u=R(z,A)fu=R(z,A)f, while its dependence on zz reveals the spectral properties of AA. Resolvents are fundamental tools in functional analysis, connecting finite-dimensional matrix theory with differential operators, evolution equations, and quantum mechanics.

Definition and conventions

Let XX be a complex Banach space, and let

A:D(A)⊆X⟶XA:D(A)\subseteq X\longrightarrow X

be a closed, densely defined linear operator. Here D(A)D(A) is its domain, and II is the identity operator.

The resolvent set ρ(A)\rho(A) consists of all z∈Cz\in\mathbb C such that

zI−A:D(A)⟶XzI-A:D(A)\longrightarrow X

is bijective and its inverse is bounded as a map from XX to XX. For such zz,

R(z,A):X⟶XR(z,A):X\longrightarrow X

is the resolvent operator. Its range is exactly D(A)D(A), and

(zI−A)R(z,A)=Ion X,(zI-A)R(z,A)=I\quad\text{on }X,
R(z,A)(zI−A)=Ion D(A).R(z,A)(zI-A)=I\quad\text{on }D(A).

For a closed operator, bijectivity already implies boundedness of the inverse by the closed graph theorem. Nevertheless, boundedness is included explicitly in the standard definition.

The spectrum is the complement

σ(A)=C∖ρ(A).\sigma(A)=\mathbb C\setminus\rho(A).

Thus the resolvent exists precisely outside the spectrum. In infinite dimensions, spectral points need not be eigenvalues: the shifted operator may instead fail to be surjective, or may have an inverse on its range that is not bounded.

A second common convention defines the resolvent as (A−zI)−1(A-zI)^{-1}, the negative of R(z,A)R(z,A). Both conventions contain the same information, but signs in derivative, contour-integral, and resolvent identities must be adjusted consistently.

Matrices and elementary examples

For a complex n×nn\times n matrix AA,

R(z,A)=(zI−A)−1R(z,A)=(zI-A)^{-1}

exists exactly when

det⁡(zI−A)≠0.\det(zI-A)\ne0.

Its entries are rational functions of zz, with possible poles at eigenvalues.

If AA admits a diagonalization

A=VDV−1,D=diag⁡(λ1,…,λn),A=VDV^{-1}, \qquad D=\operatorname{diag}(\lambda_1,\ldots,\lambda_n),

then

R(z,A)=Vdiag⁡(1z−λ1,…,1z−λn)V−1.R(z,A) = V\operatorname{diag}\left( \frac1{z-\lambda_1},\ldots,\frac1{z-\lambda_n} \right)V^{-1}.

This shows directly how the resolvent encodes both eigenvalues and their associated spectral subspaces.

Nontrivial Jordan blocks produce higher-order poles. For example, if

A=(λ10λ),A= \begin{pmatrix} \lambda&1\\ 0&\lambda \end{pmatrix},

then

R(z,A)=((z−λ)−1(z−λ)−20(z−λ)−1).R(z,A)= \begin{pmatrix} (z-\lambda)^{-1}&(z-\lambda)^{-2}\\ 0&(z-\lambda)^{-1} \end{pmatrix}.

The pole structure therefore contains information beyond the locations of eigenvalues.

Resolvent identities and analytic dependence

For z,w∈ρ(A)z,w\in\rho(A), the first resolvent identity is

R(z,A)−R(w,A)=(w−z)R(z,A)R(w,A).R(z,A)-R(w,A) = (w-z)R(z,A)R(w,A).

In particular, resolvents of the same operator commute:

R(z,A)R(w,A)=R(w,A)R(z,A).R(z,A)R(w,A)=R(w,A)R(z,A).

The resolvent set is an open set, and z↦R(z,A)z\mapsto R(z,A) is an operator-valued analytic function on it. Its derivatives satisfy

dndznR(z,A)=(−1)nn! R(z,A)n+1.\frac{d^n}{dz^n}R(z,A) = (-1)^n n!\,R(z,A)^{n+1}.

These properties follow locally from a Neumann series. If z0∈ρ(A)z_0\in\rho(A) and

∣z−z0∣ ∥R(z0,A)∥<1,|z-z_0|\,\|R(z_0,A)\|<1,

then

R(z,A)=R(z0,A)∑n=0∞[−(z−z0)R(z0,A)]n,R(z,A) = R(z_0,A) \sum_{n=0}^{\infty} \left[-(z-z_0)R(z_0,A)\right]^n,

with convergence in operator norm.

For two bounded operators AA and BB, the second resolvent identity is

R(z,A)−R(z,B)=R(z,A)(A−B)R(z,B),R(z,A)-R(z,B) = R(z,A)(A-B)R(z,B),

whenever z∈ρ(A)∩ρ(B)z\in\rho(A)\cap\rho(B). The same formula holds for suitable unbounded operators, in particular when AA and BB have the same domain and their difference extends to a bounded operator on XX. It is a basic tool for studying how solutions and spectra change under perturbations.

Norm estimates and spectral information

If AA is bounded, then for ∣z∣>∥A∥|z|>\|A\|,

R(z,A)=1z(I−Az)−1=∑n=0∞Anzn+1,R(z,A) = \frac1z\left(I-\frac Az\right)^{-1} = \sum_{n=0}^{\infty}\frac{A^n}{z^{n+1}},

and

∥R(z,A)∥≤1∣z∣−∥A∥.\|R(z,A)\|\le\frac1{|z|-\|A\|}.

This expansion is the operator analogue of a geometric series.

For an operator with nonempty spectrum,

∥R(z,A)∥≥1dist⁡(z,σ(A)).\|R(z,A)\| \ge \frac1{\operatorname{dist}(z,\sigma(A))}.

Thus the resolvent norm necessarily becomes large when a point approaches the spectrum.

For a normal operator on a Hilbert space, the spectral theorem gives the exact formula

∥R(z,A)∥=1dist⁡(z,σ(A)).\|R(z,A)\| = \frac1{\operatorname{dist}(z,\sigma(A))}.

For nonnormal operators, the norm can be much larger. Consequently, eigenvalues alone may give an incomplete account of sensitivity or transient behavior.

This distinction motivates the pseudospectrum. One common convention defines the ε\varepsilon-pseudospectrum by

σε(A)=σ(A)∪{z∈ρ(A):∥R(z,A)∥>ε−1}.\sigma_\varepsilon(A) = \sigma(A)\cup \left\{ z\in\rho(A): \|R(z,A)\|>\varepsilon^{-1} \right\}.

In finite-dimensional problems, this describes spectral sensitivity to small perturbations. If EE is bounded and

∥ER(z,A)∥<1,\|E R(z,A)\|<1,

then z∈ρ(A+E)z\in\rho(A+E). A large resolvent norm therefore indicates that comparatively small perturbations may destroy invertibility.

Spectral projections and functions of operators

The resolvent permits constructions modeled on the Cauchy integral formula from complex analysis. For a bounded operator AA and a function ff holomorphic on a neighborhood of σ(A)\sigma(A),

f(A)=12πi∫Γf(z)R(z,A) dz,f(A) = \frac{1}{2\pi i} \int_\Gamma f(z)R(z,A)\,dz,

where Γ\Gamma is an appropriate positively oriented contour, or collection of contours, in ρ(A)\rho(A) enclosing the spectrum. This is the holomorphic functional calculus.

If a contour encloses an isolated part of the spectrum and excludes the remainder, then

P=12πi∫ΓR(z,A) dzP= \frac{1}{2\pi i} \int_\Gamma R(z,A)\,dz

is a Riesz projection. It satisfies P2=PP^2=P, commutes with AA in the appropriate domain sense, and separates the corresponding invariant spectral subspace.

For an isolated matrix eigenvalue λ\lambda, this projection selects the generalized eigenspace. The order of the resolvent pole at λ\lambda equals the size of the largest Jordan block associated with that eigenvalue. For general infinite-dimensional operators, however, an isolated spectral point need not be a pole of the resolvent.

Self-adjoint operators and compact resolvent

For a self-adjoint operator AA, the spectrum is real, so every nonreal zz belongs to ρ(A)\rho(A). In particular,

∥R(z,A)∥≤1∣Im⁡z∣.\|R(z,A)\| \le \frac1{|\operatorname{Im}z|}.

If EE is the projection-valued spectral measure of AA, then

R(z,A)=∫σ(A)1z−λ dE(λ).R(z,A) = \int_{\sigma(A)} \frac1{z-\lambda}\,dE(\lambda).

This representation connects resolvent boundary values near the real axis with spectral measures. It is especially important for Hamiltonians in quantum mechanics.

An operator has compact resolvent if its resolvent set is nonempty and R(z0,A)R(z_0,A) is a compact operator for some z0∈ρ(A)z_0\in\rho(A). The resolvent identity then implies compactness of R(z,A)R(z,A) for every z∈ρ(A)z\in\rho(A).

For such an operator, every spectral point is an isolated eigenvalue of finite algebraic multiplicity, and there is no finite accumulation point of the spectrum. This property occurs for many elliptic operators on bounded domains with appropriate boundary conditions. It explains why these infinite-dimensional problems can possess discrete spectral expansions resembling matrix decompositions.

Differential equations and Green's functions

For a differential operator AA, solving

(zI−A)u=f(zI-A)u=f

is a parameter-dependent boundary-value problem. When the resolvent admits an integral-kernel representation,

(R(z,A)f)(x)=∫Gz(x,y)f(y) dy,(R(z,A)f)(x) = \int G_z(x,y)f(y)\,dy,

the kernel GzG_z is a Green's function for the shifted operator with the specified boundary conditions.

The domain of AA is essential: the same differential expression with different boundary conditions generally defines different operators, spectra, and resolvents. For example, Dirichlet and Neumann realizations of the Laplacian on the same bounded region have different spectral properties.

Resolvents also connect stationary equations with evolution equations. If AA generates a strongly continuous operator semigroup T(t)T(t) satisfying

∥T(t)∥≤Meωt,\|T(t)\|\le Me^{\omega t},

then, for Re⁡z>ω\operatorname{Re}z>\omega,

R(z,A)x=∫0∞e−ztT(t)x dt.R(z,A)x = \int_0^\infty e^{-zt}T(t)x\,dt.

Thus the resolvent is the Laplace transform of the evolution operators. This relation underlies semigroup approaches to the heat equation and other linear evolution problems.

Computational interpretation

In computation, applying R(z,A)R(z,A) to a vector usually means solving the linear system

(zI−A)u=f,(zI-A)u=f,

not explicitly forming the inverse matrix. Factorizations and iterative methods can exploit sparsity or differential-operator structure.

The resolvent norm measures the worst-case amplification of perturbations in ff. The relative condition number of the shifted system additionally depends on ∥zI−A∥\|zI-A\|. Near spectral points, or in nonnormal regions with large resolvent norm, accurate solution may be difficult even when the shift is not especially close to an eigenvalue.

A related but distinct use occurs in nonlinear analysis: for a possibly nonlinear or set-valued operator TT, the expression

JλT=(I+λT)−1J_{\lambda T}=(I+\lambda T)^{-1}

is also called a resolvent. Its single-valuedness and regularity require additional assumptions, such as maximal monotonicity; they do not follow from the linear spectral theory described above.