Spectral Theory of Compact Operators

⚡ Key Takeaways

Compact operators on a Banach (or Hilbert) space have spectra that behave almost like those of matrices: the nonzero spectrum consists of at most countably many eigenvalues of finite multiplicity, accumulating only at $0$ . The central result is the Riesz–Schauder theory and, for self-adjoint compact operators, the spectral theorem giving an orthonormal eigenbasis. Applies throughout integral equations, PDE eigenvalue problems, and quantum mechanics. Developed by Fredholm, Hilbert, Schmidt, F. Riesz, and Schauder (1900–1930).

💡 The Core Insight

Compactness forces the operator to "almost factor through finite dimensions." The Fredholm alternative for $T - \lambda I$ ( $\lambda \neq 0$ ) then holds verbatim as in linear algebra: either $T - \lambda I$ is invertible, or $\lambda$ is an eigenvalue with finite-dimensional eigenspace and finite Riesz index. The infinitude of $X$ is squeezed into a single accumulation point at the origin.

DefinitionCompact Operator

Let $X, Y$ be Banach spaces. A linear map $T: X \to Y$ is compact if $\overline{T(B_X)}$ is compact in $Y$ , where $B_X = \{x : \|x\| \le 1\}$ . Equivalently, $T$ is compact iff for every bounded sequence $(x_n) \subset X$ , the sequence $(Tx_n)$ has a convergent subsequence. We write $T \in \mathcal{K}(X,Y)$ .

DefinitionSpectrum and Its Parts

For a bounded operator $T$ on a complex Banach space $X$ , the resolvent set is

1

\quad \rho(T) = \{\lambda \in \mathbb{C} : (T - \lambda I)^{-1} \in \mathcal{B}(X)\},

Derivation hint: $\lambda \in \rho(T)$ iff $T-\lambda I$ is a bijection of $X$ ; boundedness of the inverse is then automatic by the open mapping theorem. and the spectrum is $\sigma(T) = \mathbb{C} \setminus \rho(T)$ . It decomposes as

2

\quad \sigma(T) = \sigma_p(T) \,\cup\, \sigma_c(T) \,\cup\, \sigma_r(T),

Derivation hint: classify $T-\lambda I$ by injectivity and density/closedness of range — point spectrum (not injective), continuous spectrum (injective, dense non-closed range), residual spectrum (injective, non-dense range). where $\sigma_p$ (point/eigenvalues), $\sigma_c$ (continuous), $\sigma_r$ (residual).

TheoremRiesz–Schauder Spectral Theorem for Compact Operators

Let $X$ be an infinite-dimensional complex Banach space and $T \in \mathcal{K}(X)$ . Then: 1. $0 \in \sigma(T)$ . 2. $\sigma(T) \setminus \{0\}$ consists entirely of eigenvalues. 3. Each nonzero $\lambda \in \sigma(T)$ has finite algebraic and geometric multiplicity:

3

\quad \dim \ker(T - \lambda I)^k < \infty \quad \text{for all } k \ge 1,

Derivation hint: $\ker(T-\lambda I)^k$ is a closed subspace on which $T$ acts as $\lambda I$ + nilpotent; compactness restricted to it forces finite dimension since the identity on it must be compact. 4. $\sigma(T)$ is at most countable, and its only possible accumulation point is $0$ :

4

\quad \forall \varepsilon > 0,\quad \#\{\lambda \in \sigma(T) : |\lambda| \ge \varepsilon\} < \infty.

Derivation hint: an infinite sequence of distinct eigenvalues bounded away from $0$ produces, via near-orthonormal eigenvectors and Riesz's lemma, a bounded sequence whose images have no convergent subsequence — contradicting compactness.

TheoremFredholm Alternative

For $T \in \mathcal{K}(X)$ and $\lambda \neq 0$ , exactly one holds: (a) $T - \lambda I$ is bijective (so $\lambda \in \rho(T)$ and the equation $(T-\lambda I)u = f$ has a unique solution for every $f$ ); (b) $\ker(T - \lambda I) \neq \{0\}$ , with

5

\quad \dim \ker(T - \lambda I) = \dim \ker(T^* - \bar\lambda I) = \operatorname{codim}\,\operatorname{ran}(T - \lambda I) < \infty.

Derivation hint: $T-\lambda I = -\lambda(I - \lambda^{-1}T)$ is a "compact perturbation of the identity"; its index is $0$ by the Riesz theory, equating kernel and cokernel dimensions, with solvability of $(T-\lambda I)u=f$ governed by the orthogonality $\langle f, w\rangle = 0$ for all $w \in \ker(T^*-\bar\lambda I)$ .

TheoremHilbert–Schmidt / Spectral Theorem for Compact Self-Adjoint Operators

Let $H$ be a separable Hilbert space and $T = T^* \in \mathcal{K}(H)$ . Then there exist real eigenvalues $(\lambda_n)$ with $|\lambda_1| \ge |\lambda_2| \ge \cdots \to 0$ and an orthonormal system $(e_n)$ of eigenvectors, $Te_n = \lambda_n e_n$ , such that

6

\quad Tx = \sum_{n} \lambda_n \langle x, e_n\rangle\, e_n \qquad \forall x \in H,

Derivation hint: existence of a largest eigenvalue follows from $\|T\| = \sup_{\|x\|=1}|\langle Tx,x\rangle|$ being attained by a compactness/weak-limit argument; iterate on the orthogonal complement and use that eigenvectors for distinct eigenvalues are orthogonal. and the $(e_n)$ together with an orthonormal basis of $\ker T$ form an orthonormal basis of $H$ . Moreover $\|T\| = \max_n |\lambda_n| = |\lambda_1|$ .

ProofRiesz Theory: discreteness of nonzero spectrum

Let $T \in \mathcal{K}(X)$ , $\lambda \neq 0$ . Set $S = T - \lambda I$ .

Step 1 — Kernel is finite-dimensional. On $N := \ker S = \ker(T - \lambda I)$ we have $Tx = \lambda x$ , so $T|_N = \lambda I_N$ . Since $T$ is compact and $\lambda \neq 0$ , the identity $I_N = \lambda^{-1} T|_N$ is compact, forcing $\dim N < \infty$ (a normed space with compact unit ball is finite-dimensional, by Riesz's lemma).

Step 2 — Range is closed. Write $X = N \oplus M$ with $M$ closed complement (possible since $N$ is finite-dimensional). The restriction $S|_M : M \to \operatorname{ran} S$ is injective; one shows $\|Sx\| \ge c\|x\|$ on $M$ for some $c>0$ , otherwise a unit sequence $x_n \in M$ with $Sx_n \to 0$ gives, via compactness of $T$ , a convergent subsequence $x_{n_k} \to x \in M$ with $Sx = 0$ , contradicting injectivity on $M$ . Hence $\operatorname{ran} S$ is closed.

Step 3 — Finiteness of the spectrum off any disk. Suppose $(\lambda_n)$ are distinct eigenvalues with $|\lambda_n| \ge \varepsilon > 0$ and eigenvectors $u_n$ . The $u_n$ are linearly independent; let $X_n = \operatorname{span}(u_1,\dots,u_n)$ , a strictly increasing chain of closed subspaces. By Riesz's lemma pick $y_n \in X_n$ , $\|y_n\|=1$ , $\operatorname{dist}(y_n, X_{n-1}) \ge \tfrac12$ . For $m < n$ one computes

7

\quad \|Ty_n - Ty_m\| = \|\lambda_n y_n - (\lambda_n y_n - Ty_n + Ty_m)\| \ge \tfrac{|\lambda_n|}{2} \ge \tfrac{\varepsilon}{2},

Derivation hint: $(\lambda_n I - T)y_n$ and $Ty_m$ both lie in $X_{n-1}$ , while $\lambda_n y_n$ projects to $\ge 1/2$ distance from $X_{n-1}$ ; combine with $T$ acting triangularly on the chain. so $(Ty_n)$ has no Cauchy subsequence, contradicting compactness. Thus only finitely many $|\lambda_n| \ge \varepsilon$ .

Step 4 — Index zero / Fredholm. The ascending kernels $\ker S^k$ stabilize at finite index $p$ (the Riesz index), and $X = \ker S^p \oplus \operatorname{ran} S^p$ . On the finite-dimensional $\ker S^p$ , $S$ is nilpotent; on $\operatorname{ran} S^p$ , $S$ is bijective. This yields $\dim\ker S = \operatorname{codim}\operatorname{ran} S$ , the Fredholm alternative. $\blacksquare$

∎

ProofSpectral theorem for compact self-adjoint

T

Step 1 — Largest eigenvalue. For self-adjoint $T$ , $\|T\| = \sup_{\|x\|=1}|\langle Tx,x\rangle|$ . Take $\|x_n\|=1$ with $\langle Tx_n,x_n\rangle \to \lambda_1$ where $|\lambda_1|=\|T\|$ . By compactness $Tx_n \to y$ (subsequence) and one shows $\|Tx_n - \lambda_1 x_n\|^2 = \|Tx_n\|^2 - 2\lambda_1\langle Tx_n,x_n\rangle + \lambda_1^2 \to 0$ , so $x_n \to \lambda_1^{-1} y =: e_1$ is a unit eigenvector with $Te_1 = \lambda_1 e_1$ (assuming $T\neq 0$ ).

Step 2 — Iterate. The orthogonal complement $\{e_1\}^\perp$ is $T$ -invariant (self-adjointness), and $T|_{\{e_1\}^\perp}$ is again compact self-adjoint with norm $|\lambda_2| \le |\lambda_1|$ . Repeat to extract $e_2, e_3, \dots$ .

Step 3 — Convergence to zero. If infinitely many $\lambda_n$ occur, $|\lambda_n| = \|T|_{\{e_1,\dots,e_{n-1}\}^\perp}\| \to 0$ , since otherwise Step 3 of the Riesz proof is violated.

Step 4 — Completeness on $(\ker T)^\perp$ . For $x \in H$ , $x - \sum_n \langle x,e_n\rangle e_n$ lies in $\bigcap_n \{e_n\}^\perp$ where $T$ has norm $\to 0$ , hence $T$ vanishes there: the residual lies in $\ker T$ . Equation [6] follows. $\blacksquare$

∎

DH

David Hilbert

1862–1943

Hilbert's six papers on integral equations (1904–1910) introduced the spectral viewpoint that now bears his name. He treated kernels $K(s,t)$ as infinite "matrices," developed eigenvalue expansions, and coined the language of bilinear and quadratic forms in infinitely many variables. His student Erhard Schmidt then geometrized the theory into the language of orthonormal systems and what we now call Hilbert spaces, making the compact self-adjoint spectral theorem a clean geometric statement.

🧠 Intuition

Think of a compact operator as a matrix whose entries decay fast enough that it is "almost finite." A diagonal operator $T e_n = \lambda_n e_n$ on $\ell^2$ is compact iff $\lambda_n \to 0$ — the eigenvalues must drain toward zero. The accumulation at $0$ is unavoidable: there simply is not "room" in a bounded set's image to keep large eigenvalues spread out, because compactness demands convergent subsequences. The single point $0$ is where all the infinite-dimensional complexity gets dumped; everything else looks like ordinary linear algebra, complete with eigenvalues, multiplicities, and a Jordan-like structure (the Riesz index $p$ ).

Figure

Spectrum of a compact operator: eigenvalues accumulating only at the origin

ExampleDiagonal compact operator on

\ell^2

Given. $T: \ell^2 \to \ell^2$ , $T(x_1, x_2, x_3, \dots) = (x_1, \tfrac{x_2}{2}, \tfrac{x_3}{3}, \dots)$ , i.e. $Te_n = \tfrac1n e_n$ .

Step 1 — Compactness. The finite-rank truncations $T_N(x) = (\tfrac{x_1}{1}, \dots, \tfrac{x_N}{N}, 0, \dots)$ satisfy $\|T - T_N\| = \tfrac{1}{N+1} \to 0$ . A norm-limit of finite-rank operators is compact, so $T \in \mathcal{K}(\ell^2)$ .

Step 2 — Eigenvalues. $\sigma_p(T) = \{1, \tfrac12, \tfrac13, \dots\}$ , each simple, with eigenvector $e_n$ .

Step 3 — Spectrum. The accumulation point $0$ must be included: $\sigma(T) = \{\tfrac1n : n\ge1\} \cup \{0\}$ . Note $0$ is not an eigenvalue ( $T$ is injective) but lies in $\sigma_c(T)$ since $\operatorname{ran} T$ is dense but not closed.

Answer.

\boxed{\sigma(T) = \{1, \tfrac12, \tfrac13, \dots\}\cup\{0\}, \quad 0 \in \sigma_c(T),\ \|T\| = 1.}

ExampleSolving an integral equation via the Fredholm alternative

Given. The Fredholm equation $u(s) - \mu \int_0^1 st\, u(t)\,dt = f(s)$ on $L^2[0,1]$ , with kernel $K(s,t) = st$ .

Step 1 — Identify the operator. $T u(s) = \int_0^1 st\,u(t)\,dt = s\langle u, t\rangle_{L^2}$ . This is rank one (range $= \operatorname{span}\{s\}$ ), hence compact and self-adjoint.

Step 2 — Eigenvalue. $Tu = \lambda u$ forces $u(s) = c\,s$ . Then $T(s) = s\int_0^1 t\cdot t\,dt = s\cdot \tfrac13$ . So the single nonzero eigenvalue is $\lambda_1 = \tfrac13$ with eigenvector $e_1(s) = \sqrt{3}\,s$ .

Step 3 — Apply Fredholm alternative for $\mu T u - u$ . We solve $(I - \mu T)u = f$ . This is uniquely solvable iff $\tfrac1\mu \notin \sigma_p(T)\setminus\{0\}$ , i.e. iff $\mu \neq \tfrac{1}{\lambda_1} = 3$ .

Step 4 — Explicit solution for $\mu \neq 3$ . Write $u = f + \mu s\,c$ with $c = \langle u, t\rangle$ . Then $c = \langle f,t\rangle + \mu c\langle s,t\rangle = \langle f,t\rangle + \tfrac{\mu}{3}c$ , giving $c = \dfrac{\langle f,t\rangle}{1 - \mu/3}$ .

Answer.

\boxed{u(s) = f(s) + \frac{\mu\,s}{1 - \mu/3}\int_0^1 t\,f(t)\,dt, \qquad \mu \neq 3;}

for $\mu = 3$ the equation is solvable iff $\int_0^1 t f(t)\,dt = 0$ (Fredholm orthogonality).

⚠️ Common mistake

" $0$ is always an eigenvalue of a compact operator." False. $0$ always lies in the spectrum (in infinite dimensions), but it need not be an eigenvalue: the diagonal operator $Te_n = \tfrac1n e_n$ is injective, so $\ker T = \{0\}$ , yet $0 \in \sigma_c(T)$ . The statement " $0 \in \sigma(T)$ " holds because a compact operator on an infinite-dimensional space cannot be invertible — its inverse would make the identity compact.

⚠️ Common mistake

"Compact operators always have at least one eigenvalue." False in general Banach (and even Hilbert) spaces if $T$ is not self-adjoint or normal. The nonzero spectrum can be empty — see the counterexample below. The spectral theorem's guaranteed eigenbasis requires self-adjointness (or normality, via $\sigma_p$ on $\mathbb{C}$ ).

CounterexampleCompact operator with no nonzero eigenvalues

The Volterra operator on $L^2[0,1]$ ,

8

\quad (Vu)(s) = \int_0^s u(t)\,dt,

Derivation hint: $V$ is Hilbert–Schmidt with kernel $K(s,t)=\mathbf 1_{\{t\le s\}}$ , hence compact; a suspected eigenfunction $Vu = \lambda u$ with $\lambda\neq0$ would be differentiable with $u' = \lambda^{-1}u$ , $u(0)=0$ , forcing $u\equiv 0$ . is compact but quasinilpotent: $\sigma(V) = \{0\}$ and it has no eigenvalues at all. This shows the Riesz theory guarantees only that the nonzero spectrum consists of eigenvalues — it does not guarantee any exist. ( $V$ is non-self-adjoint; $V + V^*$ is the rank-one projection-like operator with full real eigenvalue structure.)

connection

The compact spectral theorem sits inside a hierarchy. It is the cleanest case of the Spectral Theorem for Bounded Self-Adjoint Operators, which for non-compact operators replaces the discrete sum [6] by an integral against a Projection-Valued Measure. Compact operators form a closed two-sided ideal in $\mathcal{B}(X)$ ; quotienting yields the Calkin Algebra, where invertibility encodes Fredholm Operators of arbitrary index — the Riesz theory is exactly the index-zero, compact-perturbation case. On Hilbert space the compact operators are the closure of finite-rank operators, and the refined classes (trace class, Hilbert–Schmidt Operators, Schatten $p$ -classes) are defined by summability of the singular values $s_n(T)$ — the eigenvalues of $|T| = (T^*T)^{1/2}$ from the Singular Value Decomposition.

🔬 Where this appears

⏳ History

- 1900–1903 — Ivar Fredholm solves the integral equation

u - \lambda Ku = f

using infinite determinants, establishing the "Fredholm alternative" for continuous kernels.

- 1904–1910 — David Hilbert publishes his series on integral equations, introducing eigenvalue expansions and the theory of infinitely many variables (quadratic forms, "vollstetig" = completely continuous operators).

- 1907 — Erhard Schmidt recasts Hilbert's theory geometrically with orthonormal systems, proving the eigenfunction expansion (the Hilbert–Schmidt expansion).

- 1909 — Frigyes Riesz and Maurice Fréchet independently prove the

L^2

representation theorem, anchoring Hilbert-space methods.

- 1916 — Frigyes Riesz develops the abstract theory of completely continuous (compact) operators on

C[a,b]

, proving discreteness of the spectrum and finite multiplicity — the Riesz theory in essentially modern form.

- 1918 — Riesz proves what is now Riesz's lemma, the key compactness tool for the spectral discreteness argument.

- 1930 — Juliusz Schauder extends the Fredholm alternative to general Banach spaces and proves compactness of the adjoint (Schauder's theorem:

T

compact

\iff T^*

compact).

- 1932 — John von Neumann's Mathematical Foundations of Quantum Mechanics and Stone's work place the spectral theorem on rigorous Hilbert-space footing, distinguishing compact (discrete) from continuous spectra.

- 1940s–1950s — Development of Schatten–von Neumann classes (trace class, Hilbert–Schmidt) quantifies compactness by singular-value summability.

open-problem

Where the clean theory ends. Several genuinely hard or open issues lurk:

Invariant subspace problem. Whether every bounded operator on a separable infinite-dimensional Hilbert space has a nontrivial closed invariant subspace remains open. For compact operators the answer is famously yes — Aronszajn–Smith (1954), extended by Lomonosov (1973) to operators commuting with a nonzero compact operator — making compactness a rare island of resolution within an open ocean.
Eigenvalue vs. singular-value asymptotics. The relation between the eigenvalues $\lambda_n(T)$ and singular values $s_n(T)$ is subtle: Weyl's inequality $\sum|\lambda_n|^p \le \sum s_n^p$ holds, but precise asymptotics of eigenvalues for non-self-adjoint compact operators (Volterra-type, pseudodifferential) remain delicate and active.
Completeness of eigenvectors for non-self-adjoint compact operators is not guaranteed; root-vector completeness (Keldysh-type theorems) requires extra structure and is sharp only under decay conditions on $s_n$ .
Non-quasi-analytic perturbations. Small non-compact perturbations can dissolve the discrete spectrum into continuous spectrum, and pinpointing the essential spectrum under such perturbations connects to deep Fredholm-theoretic stability questions.

Formal Statement

Why It Holds

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also