Spectral projectors

1 Spectral decomposition background

1.1 Linear operators and invariant subspaces

Let \(V\) be a vector space (often finite-dimensional), and let \(T:V\to V\) be a linear operator. A subspace \(W\subseteq V\) is called *\(T\)-invariant* if \(T(W)\subseteq W\). Invariant subspaces organize how vectors transform under repeated application of \(T\); they are the natural stage on which “spectral parts” of an operator can be separated.

When \(V\) admits a direct-sum decomposition into invariant subspaces, \(T\) can be represented in block form, and functions of \(T\) act blockwise. Spectral projectors formalize this separation for subspaces associated with selected spectral values.

1.2 Eigenvalues, eigenspaces, and diagonalization

An eigenvalue \(\lambda\) of \(T\) is a scalar for which there exists a nonzero vector \(v\) such that \(Tv=\lambda v\). The associated eigenspace \[ E_\lambda=\ker(T-\lambda I) \] collects all eigenvectors with the same eigenvalue. If the direct sum of eigenspaces spans \(V\), then \(T\) is diagonalizable: there exists a basis of \(V\) consisting of eigenvectors, and \(T\) has a diagonal representation.

Spectral projectors are operators whose ranges coincide with particular sums of eigenspaces (or related invariant subspaces). In a diagonalizable setting, they become especially transparent.

1.3 Minimal polynomials and primary decomposition

The *minimal polynomial* \(m_T(x)\) is the monic polynomial of least degree such that \(m_T(T)=0\). It controls how \(T\) behaves on \(V\), particularly through the factorization of \(m_T\) into relatively prime components: \[ m_T(x)=\prod_i (x-\lambda_i)^{k_i}. \] This enables *primary decomposition*, where \(V\) splits into \(T\)-invariant subspaces associated with the powers of each \((x-\lambda_i)\). Even when \(T\) is not diagonalizable, these primary components support projector constructions tied to each eigenvalue’s algebraic contribution.

1.4 Normal and diagonalizable operators

In inner product spaces, an operator \(T\) is *normal* if \(T^*T=TT^*\). Normal operators are diagonalizable via an orthonormal basis and their eigenspaces for distinct eigenvalues are orthogonal. For normal operators, spectral projectors corresponding to different spectral parts are not only idempotent but also orthogonal with respect to the inner product.

Diagonalizable operators need not be normal, but their spectral projectors defined from eigenspace decompositions still yield invariant-subspace isolation. In contrast, non-diagonalizable operators require broader notions using generalized eigenspaces or contour-integral definitions.

2 Definition of spectral projectors

2.1 Projectors from eigenspace decomposition

2.1.1 Idempotence and range/kernel characterization

A *projector* (or projection operator) \(P\) on \(V\) is a linear map satisfying \(P^2=P\). Equivalently, \(P\) acts as the identity on its range and annihilates its complementary null space. For any projector \(P\), \[ \operatorname{Ran}(P)\oplus\ker(P)=V \] and \(\operatorname{Ran}(P)\) is invariant under operators that commute with \(P\).

For a diagonalizable operator \(T\) with eigenvalues \(\{\lambda_i\}\), one defines spectral projectors by selecting a subset \(S\) of eigenvalues and setting \[ P_S:\ V\to V \] to be the projector onto \(\bigoplus_{\lambda\in S}E_\lambda\) along \(\bigoplus_{\lambda\notin S}E_\lambda\). These \(P_S\) are spectral projectors because they extract precisely the components supported on the chosen part of the spectrum.

2.1.2 Orthogonality in the normal/operator-adjoint setting

If \(T\) is normal, eigenspaces for distinct eigenvalues are orthogonal. Consequently, the spectral projectors onto sums of eigenspaces become *orthogonal projectors*. In that case, the projector \(P_S\) is self-adjoint: \[ P_S^*=P_S, \] and different spectral projectors satisfy \[ P_S P_{S'}=P_{S\cap S'} \] with orthogonality when the sets are disjoint.

For general diagonalizable (non-normal) operators, projectors are typically not orthogonal, though they remain idempotent and decompose the space.

2.2 Polynomial and Lagrange interpolation formulas

2.2.1 Simple-eigenvalue case

Suppose \(T\) is diagonalizable with distinct eigenvalues \(\lambda_1,\dots,\lambda_m\). For the projector onto the eigenspace \(E_{\lambda_i}\), one can use Lagrange interpolation polynomials: \[ p_i(x)=\prod_{j\neq i}\frac{x-\lambda_j}{\lambda_i-\lambda_j}. \] Then \[ P_{\lambda_i}=p_i(T). \] The polynomial satisfies \(p_i(\lambda_i)=1\) and \(p_i(\lambda_j)=0\) for \(j\neq i\), so the operator \(p_i(T)\) acts as desired on each eigenspace.

2.2.2 Multiple-eigenvalue and generalized eigenspaces

When eigenvalues repeat or the operator is not diagonalizable, eigenspaces alone may not span the space. One can still construct projectors onto generalized eigenspace components using polynomials derived from the minimal polynomial and the primary decomposition. In that setting, the spectral projector associated with \(\lambda\) (or with a set of eigenvalues) is obtained by choosing a polynomial that equals \(1\) on the relevant primary factor and \(0\) on the others, interpreted through \(m_T(T)=0\).

Practically, this often yields projectors that extract the invariant subspace associated with a cluster of eigenvalues, even when Jordan blocks are present.

2.3 Functional calculus viewpoint

2.3.1 Characteristic functions of spectral sets

A key perspective is to treat projectors as operators corresponding to indicator functions of spectral sets. In the spectral theorem for normal operators, one associates to \(T\) a projection-valued measure \(E(\cdot)\) so that, informally, \[ T=\int \lambda\, dE(\lambda). \] Then for suitable sets \(\Delta\) in the spectrum, the projector onto the part of the space where the spectral parameter lies in \(\Delta\) is \[ P_\Delta = E(\Delta). \] For discrete spectra, this reduces to sums of eigenspace projections.

2.3.2 Continuous and holomorphic functional calculus

For operators where functional calculus is available (e.g., normal operators via continuous functional calculus, or general operators under holomorphic functional calculus), one can define \[ f(T) \] for a broad class of functions \(f\). Spectral projectors emerge by applying functions that approximate or encode characteristic behavior of subsets of the spectrum. In holomorphic settings, they are often obtained from contour integrals (see Riesz projectors below), which avoids direct use of discontinuous indicator functions.

3 Properties and algebraic identities

3.1 Idempotent and complementary projector relations

By construction, spectral projectors satisfy idempotence: \[ P^2=P. \] Often they also come in complementary pairs. If \(P\) projects onto the spectral part associated with a set \(S\), then the complementary projector \[ Q=I-P \] projects onto the remaining spectral part \(S^c\) (in a finite-dimensional diagonalizable context where the chosen spectral parts cover the whole spectrum).

3.2 Sum rules over spectral parts

If the spectrum is partitioned into disjoint subsets \(S_1,\dots,S_k\) and the corresponding projectors exist with ranges forming a direct sum, then \[ P_{S_1}+\cdots+P_{S_k}=I. \] More generally, for nested sets one has monotonic consistency on the corresponding ranges, reflecting how the underlying invariant subspaces assemble.

3.3 Product rules and mutual orthogonality

Projectors associated with disjoint spectral sets satisfy \[ P_{S}P_{S'}=0 \quad \text{when } S\cap S'=\varnothing. \] For arbitrary sets (in settings where the functional calculus viewpoint holds cleanly), one typically has a product rule of the form \[ P_{S}P_{S'}=P_{S\cap S'}. \] For normal operators, these projectors are additionally orthogonal in the inner-product sense when the sets are disjoint.

3.4 Commutation with the underlying operator

Spectral projectors commute with the operator: \[ TP_S = P_S T. \] This follows because the range of \(P_S\) is \(T\)-invariant, and (in the diagonalizable or functional-calculus framework) the projector is constructed as a function of \(T\). Commutation implies that \(T\) preserves the decomposition determined by the projectors.

3.5 Trace and rank connections in finite dimensions

In finite-dimensional spaces, the rank of a spectral projector equals the dimension of its range. For diagonalizable operators, the rank of the projector onto eigenvalue \(\lambda\) equals the geometric multiplicity summed over the chosen eigenvalues: \[ \operatorname{rank}(P_S)=\dim\Big(\bigoplus_{\lambda\in S}E_\lambda\Big). \] If \(T\) is normal, trace identities are particularly clean: for instance, the trace of a spectral projector equals its rank because projectors are diagonalizable with eigenvalues \(0\) and \(1\).

In non-normal or non-diagonalizable cases, care is needed: trace still equals the sum of eigenvalues of the projector (hence the rank in many finite-dimensional contexts), but those ranks correspond to generalized eigenspace dimensions rather than just eigenspace dimensions.

4 Construction methods

4.1 Contour integral (Riesz) projectors

4.1.1 Resolvent-based formulas

For a bounded operator \(T\) on a suitable space, the *resolvent* is \[ R(z;T)=(zI-T)^{-1}, \] defined for \(z\) outside the spectrum. If \(\Gamma\) is a positively oriented simple closed contour enclosing a portion of the spectrum and excluding the rest, then the Riesz projector associated with the enclosed spectral part can be defined by \[ P=\frac{1}{2\pi i}\int_\Gamma (zI-T)^{-1}\,dz. \] This operator extracts the invariant subspace corresponding to eigenvalues inside \(\Gamma\), even when \(T\) is not diagonalizable.

4.1.2 Choice of contours and spectral separation

The contour must separate the targeted eigenvalues from the rest of the spectrum. In finite dimensions, this amounts to selecting a curve in the complex plane that encloses exactly the desired eigenvalues (counted with algebraic multiplicity, in the generalized-eigenspace sense) and avoids others. For infinite-dimensional operators, additional assumptions (e.g., isolated spectral components) ensure the integral is well defined and yields a projector of finite rank in common scenarios.

A key practical consideration is stability: when eigenvalues approach each other, contours that previously separated spectral parts may cease to do so without modification.

4.2 Polynomial projector construction

4.2.1 Using the minimal polynomial

Given the factorization of the minimal polynomial \(m_T(x)=\prod_i (x-\lambda_i)^{k_i}\), one can construct a polynomial \(p(x)\) such that \(p(T)\) acts as the identity on the primary component for selected \(\lambda_i\) and as zero on the remaining primary components. Existence follows from coprimeness between appropriate factors. Then \[ P=p(T) \] produces a spectral projector (in the sense of mapping onto the invariant subspace associated with the chosen eigenvalues).

This approach is algebraic and often computationally efficient when minimal polynomial factors are available.

4.2.2 Lagrange basis polynomials for simple spectra

When eigenvalues are distinct and \(T\) is diagonalizable, the minimal polynomial splits into distinct linear factors. Then Lagrange interpolation polynomials yield projectors directly, as in the simple-eigenvalue case: \[ P_{\lambda_i}=p_i(T) \] with \(p_i(\lambda_i)=1\) and \(p_i(\lambda_j)=0\) for \(j\neq i\). This method produces exact projectors in exact arithmetic.

4.3 Spectral measure approach (operator-theoretic setting)

4.3.1 Projection-valued measures

For normal operators (and more general classes under suitable conditions), there exists a projection-valued measure \(E(\cdot)\) on the spectrum such that spectral projectors correspond to measurable sets: \[ P_\Delta = E(\Delta). \] These projectors satisfy countable additivity in the strong operator topology and integrate naturally into the functional calculus.

4.3.2 Integrating functions against spectral projectors

Once \(E\) is available, one defines \[ f(T)=\int f(\lambda)\, dE(\lambda), \] where the integral is taken in the operator sense. Taking \(f\) as an indicator-like function approximant (or using exact indicators for suitable sets) recovers spectral projectors as special cases. This provides a unified framework linking operator theory, measure theory, and decomposition of spaces.

5 Spectral projectors for families of operators

5.1 Dependence on parameters and stability

Often \(T\) depends on a parameter \(s\), giving a family \(T(s)\). If a spectral portion remains isolated as \(s\) varies, the associated projector \(P(s)\) can be defined continuously (or analytically under stronger hypotheses) using contour integrals with contours that move but continue to separate the spectrum. Stability fails when the spectral gap closes, causing projector definitions to change discontinuously or become ill-conditioned.

5.2 Perturbation theory basics

Perturbation theory studies how eigenvalues and invariant subspaces change under small changes in the operator. Spectral projectors are central because they encode the subspace rather than only individual eigenvectors, which may be unstable under perturbations. In many settings, controlling \(P(s)\) yields robust information about the spectral component.

5.3 Tracking eigenvalue clusters via projectors

When multiple eigenvalues are close, a projector onto the entire cluster can be more stable than projectors onto individual eigenvalues. The contour method naturally targets clusters by enclosing them together. This “bundle projection” viewpoint is used in applications where eigenvalue ordering is not reliable.

5.4 Differentiation of projector-valued maps

If \(P(s)\) is defined via an analytic functional calculus or via an integral representation, \[ P(s)=\frac{1}{2\pi i}\int_\Gamma (zI-T(s))^{-1}\,dz, \] then differentiation can be performed by differentiating under the integral sign, yielding formulas involving the derivative \(T'(s)\) and the resolvent. Such identities are used to derive first-order sensitivity of invariant subspaces.

5.5 Numerical approximations in parameter regimes

In computation, projectors are approximated using finite-dimensional models or iterative methods. Sensible numerical schemes often rely on resolving the invariant subspace via subspace iteration, rational filters, or contour-integral quadrature. Near spectral transitions, approximation errors can grow, especially if the targeted part of the spectrum is not well separated.

6 Applications

6.1 Decomposing vectors and solving linear systems

Given a decomposition \(V=\operatorname{Ran}(P_S)\oplus\operatorname{Ran}(I-P_S)\), any vector \(v\) splits into components supported on chosen spectral parts: \[ v=P_S v + (I-P_S)v. \] This decomposition can simplify linear solves for \((T-\alpha I)x=b\) or for related operator equations by reducing the problem to invariant subspaces where the operator acts more simply.

6.2 Block-diagonalization and invariant subspace isolation

When projectors commute with \(T\), they induce a block structure: the operator acts independently on the ranges of the projectors. This enables block-diagonalization strategies that isolate modes or decouple dynamics. Such reductions are common in control, vibration analysis, and model reduction.

6.3 Modal analysis and system identification

In systems governed by linear dynamics, the spectrum corresponds to natural modes. Spectral projectors separate modal contributions, including those associated with repeated or clustered eigenvalues. This is useful when identifying which parts of the response originate from specific eigenvalues or frequency bands.

6.4 Reduced models using projector truncation

Projector truncation replaces the full state space with a lower-dimensional invariant subspace: \[ x \approx P_S x. \] The resulting reduced-order model captures dominant behavior when the ignored spectral components have limited influence. The projector viewpoint clarifies what is preserved: the reduction retains dynamics associated with the selected portion of the spectrum.

6.5 Connections to resolvent identities and Green’s functions

The resolvent \((zI-T)^{-1}\) is closely related to Green’s functions in differential-equation settings. Since spectral projectors can be obtained from contour integrals of the resolvent, they connect directly to residues and expansions of Green’s functions. This underpins many analytic methods for understanding how forcing and boundary conditions excite different spectral components.

7 Examples and worked computations

7.1 2×2 and 3×3 explicit eigen-decomposition

Consider a \(2\times 2\) matrix \(T\) with two distinct eigenvalues \(\lambda_1,\lambda_2\). If \(T\) is diagonalizable, then the projectors \(P_{\lambda_1}\) and \(P_{\lambda_2}\) satisfy \[ P_{\lambda_1}+P_{\lambda_2}=I,\quad P_{\lambda_1}P_{\lambda_2}=0, \] and \[ T=\lambda_1 P_{\lambda_1}+\lambda_2 P_{\lambda_2}. \] Given eigenvectors \(v_1,v_2\), one can form \(P_{\lambda_i}\) via \(P_{\lambda_i}=\frac{v_i w_i^*}{w_i^* v_i}\) for left/right eigenvectors \(w_i\) in the non-normal case, or via orthogonal projection in the normal case.

For \(3\times 3\) with three distinct eigenvalues, the same principle extends: each eigenvalue projector is a polynomial in \(T\) of degree at most two, determined by Lagrange interpolation.

7.2 Projectors for diagonal matrices

If \(T=\mathrm{diag}(\lambda_1,\dots,\lambda_n)\), then the spectral projector onto a subset \(S\) is simply the diagonal matrix \[ (P_S)_{ii}=\begin{cases} 1,& \lambda_i\in S,\\ 0,& \lambda_i\notin S. \end{cases} \] This example shows the defining property directly: \(P_S\) selects precisely the coordinates corresponding to eigenvalues in \(S\).

7.3 Projectors for a single eigenvalue with multiplicity

Let \(T\) be diagonalizable with eigenvalue \(\lambda\) of multiplicity \(m\) (geometric multiplicity, in this setting, equals \(m\)). Then \[ P_\lambda \] projects onto the \(m\)-dimensional eigenspace \(E_\lambda\). In an eigenbasis where \(T\) is diagonal, \(P_\lambda\) is a block diagonal matrix with an \(m\times m\) identity block and a zero block elsewhere.

For normal operators, this projector is orthogonal and can be computed by selecting the eigenvectors spanning \(E_\lambda\) and forming the orthonormal basis projection.

7.4 Riesz projector example via a resolvent contour integral

Suppose \(T\) is a \(2\times 2\) matrix with eigenvalues \(\lambda_1,\lambda_2\). Choose \(\Gamma\) as a circle enclosing \(\lambda_1\) but not \(\lambda_2\). The Riesz projector \[ P=\frac{1}{2\pi i}\int_\Gamma (zI-T)^{-1}\,dz \] equals the projector onto the invariant subspace associated with \(\lambda_1\). In practice, one may compute \((zI-T)^{-1}\) explicitly, integrate term-by-term using residue calculus, and recover \(P\) as the residue at \(z=\lambda_1\).

This illustrates the method’s power: it works without constructing eigenvectors and remains valid even in non-diagonalizable cases (with appropriate spectral isolation).

7.5 Projectors in a generalized eigenvalue (Jordan) setting

Consider a Jordan block \(J\) associated with eigenvalue \(\lambda\). If the operator consists of a single Jordan block, the generalized eigenspace equals the full space, and the projector onto that eigenvalue part is the identity. If \(T\) has several Jordan blocks for the same eigenvalue cluster, the generalized eigenspace projector still extracts the entire Jordan-associated invariant subspace.

When Jordan structure is present, eigenvectors alone do not provide a complete basis, but the Riesz projector constructed from the resolvent isolates the correct invariant subspace corresponding to the eigenvalue, including vectors needed to represent generalized eigenvectors.

8 Common pitfalls and edge cases

8.1 Non-diagonalizable operators and generalized eigenspaces

A common mistake is to assume that spectral projectors correspond only to ordinary eigenspaces. For non-diagonalizable operators, the invariant subspace associated with an eigenvalue is generally the generalized eigenspace. Projectors defined from contour integrals or from primary decomposition target this larger invariant space.

8.2 Spectral overlap and contour selection issues

Contour-integral projectors require spectral separation. If the chosen contour encloses eigenvalues that are intended to be excluded, the resulting projector targets a larger invariant subspace. Conversely, if the contour fails to isolate the desired eigenvalues (e.g., because eigenvalues cross the contour under parameter variation), the projector definition may change abruptly or become undefined.

8.3 Non-normal operators and loss of orthogonality

For non-normal operators, projectors need not be orthogonal and may be highly non-unitarily conditioned. Even though projectors remain idempotent and commute with \(T\) (when constructed properly), numerical computations can suffer from large sensitivity due to non-orthogonal eigenvectors and non-normal transient behavior.

8.4 Domain/boundedness concerns in infinite-dimensional cases

In infinite-dimensional settings, one must account for whether \(T\) is bounded, whether the resolvent exists on the contour, and whether the integral defines a bounded operator. Spectral projectors can be unbounded or fail to exist in the naive form if the operator’s spectral assumptions do not hold (for example, if relevant spectral components are not isolated).

8.5 Misinterpreting rank, trace, and multiplicity relations

Rank and trace are determined by the eigenvalues of the projector itself (typically \(0\) and \(1\)), but the relationship to operator multiplicities can be subtle. In particular, algebraic multiplicity, geometric multiplicity, and generalized eigenspace dimension are distinct notions in non-diagonalizable cases. Correct interpretation requires matching the projector to the spectral set and the invariant subspace that projector targets.