02.11.05 · analysis / functional-analysis

Compact operators

shipped3 tiersLean: partial

Anchor (Master): Reed-Simon Vol. I §VI; Conway §II–§VII; Schaefer §III

Intuition [Beginner]

A compact linear operator is one that is "almost finite-dimensional" in a specific topological sense: it sends bounded sequences to sequences with convergent subsequences. On finite-dimensional spaces, every bounded operator is compact (Bolzano-Weierstrass). On infinite-dimensional spaces, the identity operator is never compact, and neither is most of the algebra.

Compact operators are the well-behaved corner of operator theory: they have point spectra, they admit eigenvector expansions, and (after Atkinson) they form the ideal modulo which Fredholm operators live.

The simplest examples: integral operators with continuous kernels, finite-rank operators (which form a dense subset of the compacts in any Hilbert space), and operators that are uniform limits of finite-rank operators.

Visual [Beginner]

The unit ball in an infinite-dimensional space is not compact — sequences in it can fail to have convergent subsequences. A compact operator squeezes the unit ball down to something whose closure is compact: a relatively compact image.

So while bounded operators preserve "bounded," compact operators upgrade to "preserves bounded into relatively compact." The upgrade is what makes them "almost finite-dimensional."

Worked example [Beginner]

Take the diagonal operator $T : ℓ^{2} \to ℓ^{2}$ defined by $T (x_{1}, x_{2}, x_{3}, \dots) = (x_{1}, x_{2} /2, x_{3} /3, \dots)$ — multiply the $n$ -th coordinate by $1/ n$ .

This $T$ is bounded (operator norm $1$ , achieved on the first coordinate). Is it compact?

Compute: a bounded sequence ${y^{(k)}}$ in $ℓ^{2}$ has $∥ y^{(k)} ∥ \leq M$ for some $M$ . Apply $T$ : each component of $T y^{(k)}$ is multiplied by $1/ n$ , so the tail of $T y^{(k)}$ is small uniformly in $k$ . Concretely, for any $ε > 0$ , pick $N$ with $1/ N < ε / M$ ; then the tail beyond index $N$ contributes at most $ε$ . The first $N$ coordinates are bounded in $R^{N}$ (a finite-dimensional space), so by Bolzano-Weierstrass they have a convergent subsequence. Combine: ${T y^{(k)}}$ has a Cauchy subsequence, hence convergent (since $ℓ^{2}$ is complete).

So $T$ is compact. The compactness comes from the diagonal entries decaying to zero. The non-compact diagonal operator $S (x_{n}) = (x_{1}, x_{2}, x_{3}, \dots)$ (the identity) has constant diagonal $1$ and fails Bolzano-Weierstrass on the standard basis sequence.

Check your understanding [Beginner]

Formal definition [Intermediate+]

Let $X, Y$ be Banach spaces. A bounded linear operator $T \in B (X, Y)$ 02.11.01 is compact if it sends every bounded set to a relatively compact set — a set whose closure is compact.

Equivalently:

$T (B_{X})$ has compact closure in $Y$ , where $B_{X}$ is the unit ball of $X$ .
For every bounded sequence ${x_{n}} \subset X$ , the sequence ${T x_{n}}$ has a Cauchy (hence convergent) subsequence in $Y$ .

Write $K (X, Y)$ for the set of compact operators $X \to Y$ , and $K (X) := K (X, X)$ .

Theorem (basic structure).

$K (X, Y)$ is a closed vector subspace of $B (X, Y)$ in the operator norm.
The composition of a compact operator with any bounded operator (on either side) is compact: $K (X) \subset B (X)$ is a two-sided ideal.
Schauder's theorem: $T \in B (X, Y)$ is compact iff its adjoint $T^{*} \in B (Y^{*}, X^{*})$ is compact.
Every finite-rank operator is compact, and on a Hilbert space the finite-rank operators are dense in $K (X)$ .

(Proofs in §"Full proof set".)

The Calkin algebra $B (X) / K (X)$ is the quotient of the bounded operators by the compact ideal — the natural setting for Fredholm theory 03.09.06: an operator is Fredholm iff its image in the Calkin algebra is invertible (Atkinson).

Key theorem with proof [Intermediate+]

Theorem (Compact operators form a closed two-sided ideal in $B (X)$ ). Let $X$ be a Banach space. The set $K (X) \subseteq B (X)$ of compact operators is:

(i) a vector subspace, closed under sums and scalar multiples;

(ii) closed in operator norm;

(iii) closed under left and right composition by bounded operators: $A \in B (X), K \in K (X) \Rightarrow A K, K A \in K (X)$ .

Proof.

(i) Subspace: Let $K_{1}, K_{2} \in K (X)$ and $λ \in K$ . For a bounded sequence ${x_{n}}$ , extract subsequences: from ${K_{1} x_{n}}$ a convergent subsequence, then from the corresponding ${K_{2} x_{n}}$ on that subsequence another convergent subsequence. The combined subsequence makes both $K_{1} x_{n}$ and $K_{2} x_{n}$ converge, hence $(K_{1} + λ K_{2}) x_{n}$ converges. So $K_{1} + λ K_{2}$ is compact.

(ii) Closed: Let $K_{n} \to K$ in operator norm with $K_{n} \in K (X)$ . Show $K$ is compact. Take a bounded sequence ${x_{m}}$ , $∥ x_{m} ∥ \leq M$ . By a diagonal argument, extract a subsequence ${x_{m_{k}}}$ on which every $K_{n} x_{m_{k}}$ converges (extract along $K_{1}$ , then a sub-subsequence on which $K_{2}$ converges, etc.; the diagonal subsequence works for every $K_{n}$ ). For this subsequence, estimate

∥ K x_{m_{k}} - K x_{m_{l}} ∥ \leq ∥ K - K_{n} ∥ \cdot ∥ x_{m_{k}} - x_{m_{l}} ∥ + ∥ K_{n} x_{m_{k}} - K_{n} x_{m_{l}} ∥.

Given $ε > 0$ , choose $n$ with $∥ K - K_{n} ∥ < ε / (4 M)$ , then choose $k_{0}$ with $∥ K_{n} x_{m_{k}} - K_{n} x_{m_{l}} ∥ < ε /2$ for $k, l \geq k_{0}$ . Then $∥ K x_{m_{k}} - K x_{m_{l}} ∥ < 2 M \cdot ε / (4 M) + ε /2 = ε$ . So ${K x_{m_{k}}}$ is Cauchy. By completeness it converges, and $K$ is compact.

(iii) Two-sided ideal: Let $A \in B (X)$ , $K \in K (X)$ . For a bounded sequence ${x_{n}}$ , $K x_{n}$ has a convergent subsequence $K x_{n_{k}} \to y$ . Then $A K x_{n_{k}} \to A y$ by continuity of $A$ . So $A K$ is compact. Similarly, $K A x_{n}$ : ${A x_{n}}$ is bounded (since $A$ is bounded), and applying $K$ to it has a convergent subsequence, so $K A$ is compact. $□$

Bridge. The construction here builds toward 03.09.06 (fredholm operators), where the same data is upgraded, and the symmetry side is taken up in 03.09.10 (atiyah-singer index theorem). The defining pattern appears again in those units in a sharpened form, where the local data is glued or quotiented. Putting these together, the foundational insight is that the data of this unit gives the structural signature that the rest of the strand reads off.

Exercises [Intermediate+]

Exercise 5 (medium, conceptual).

Why does the spectrum of a compact operator on an infinite-dimensional space consist of $0$ together with at most countably many isolated eigenvalues that can only accumulate at $0$ ?

Hint

The non-zero spectrum is purely point spectrum; eigenvalues bounded away from zero are finite in number.

Answer

For a compact $T$ , the resolvent equation $(T - λ I) x = y$ for $λ \neq = 0$ becomes Fredholm: $T - λ I = - λ (I - T / λ)$ , and $T / λ$ is compact. Riesz's spectral theorem says any non-zero spectral value of $T$ is an isolated eigenvalue with finite-dimensional generalised eigenspace, and the non-zero eigenvalues form an at-most-countable set whose only possible accumulation point is $0$ . The point $0$ itself is in the spectrum (since $T$ is non-invertible on infinite-dimensional spaces, see Exercise 2), but may or may not be an eigenvalue.

Exercise 6 (hard, proof).

Show that on a Hilbert space, the finite-rank operators are dense in $K (H)$ .

Hint

Use orthonormal bases and approximate by truncating.

Answer

Let $K \in K (H)$ . Choose an orthonormal basis ${e_{n}}$ for $H$ . Let $P_{N}$ be the orthogonal projection onto $span (e_{1}, \dots, e_{N})$ — a finite-rank operator with $∥ P_{N} ∥ = 1$ . Define $K_{N} := P_{N} K$ , a finite-rank operator (rank $\leq N$ ). Show $K_{N} \to K$ in operator norm.

Compute $∥ K - K_{N} ∥ = ∥ (I - P_{N}) K ∥$ . This is the norm of $K$ projected onto the orthogonal complement of $span (e_{1}, \dots, e_{N})$ . As $N \to \infty$ , this complement shrinks. By compactness of $K$ , $K (B_{H})$ is relatively compact, hence totally bounded; for any $ε > 0$ , finitely many $ε$ -balls cover $K (B_{H})$ . Their centres span a finite-dimensional subspace, captured by $P_{N}$ for $N$ large enough. Concretely: for sufficiently large $N$ , $∥ (I - P_{N}) K x ∥ < ε$ uniformly for $∥ x ∥ \leq 1$ . So $∥ K - K_{N} ∥ \to 0$ .

Exercise 7 (hard, proof).

State and verify Schauder's theorem: $T \in B (X, Y)$ is compact iff its adjoint $T^{*} : Y^{*} \to X^{*}$ is compact.

Hint

Compactness of $T$ corresponds to relative compactness of $T (B_{X})$ ; via Banach-Alaoglu and equicontinuity arguments, this transfers to $T^{*} (B_{Y^{*}})$ .

Answer

Sketch: given ${φ_{n}} \subset Y^{*}$ bounded, restrict each $φ_{n}$ to the relatively compact set $T (B_{X})$ . This restricted family is bounded and equicontinuous (as $T (B_{X})$ is totally bounded), so by Arzelà-Ascoli it has a uniformly convergent subsequence $φ_{n_{k}}$ on $T (B_{X})$ . Then $T^{*} φ_{n_{k}}$ converges in operator norm in $X^{*}$ (the supremum over $B_{X}$ is the same as the supremum of $∣ φ_{n_{k}} (T x) ∣$ over $∥ x ∥ \leq 1$ , which is precisely the uniform-on- $T (B_{X})$ convergence). Hence $T^{*}$ is compact. The converse follows by the same argument applied to $T^{**}$ in place of $T$ , using that $T \to T^{**}$ embeds $T$ into the bidual.

Lean formalization [Intermediate+]

lean_status: partial — Mathlib has IsCompactOperator predicate, the closed-ideal property, and Schauder's theorem.

[object Promise]

The Codex companion module records the ideal structure and the limit-of-finite-rank theorem (the latter requires Hilbert-space hypotheses for the cleanest statement).

Advanced results [Master]

Spectral theory for compact operators. The Riesz-Schauder theorem: for compact $T \in B (X)$ on an infinite-dimensional Banach space $X$ ,

$σ (T) = {0} \cup {λ_{n}}$ where ${λ_{n}}$ is at most a countable set of non-zero eigenvalues with no accumulation points other than possibly $0$ ;
each non-zero $λ_{n}$ is an eigenvalue with finite-dimensional generalised eigenspace;
$λ = 0$ is in the spectrum (since $T$ is not invertible on infinite-dimensional space) and may or may not be an eigenvalue.

This generalises the finite-dimensional spectral theorem: compact operators are the class of bounded operators with a discrete-eigenvalue structure modelled on the matrix case.

Compact self-adjoint operators on Hilbert spaces. The spectral theorem in its sharpest form: if $T \in K (H)$ is self-adjoint, then $H$ has an orthonormal basis of eigenvectors of $T$ , and

T = n \sum λ_{n} ⟨ e_{n}, \cdot ⟩ e_{n},

where the $λ_{n}$ are real and tend to $0$ . Diagonal-matrix-like behaviour, but on infinite-dimensional spaces.

Hilbert-Schmidt and trace-class operators. Inside $K (H)$ are two further ideals. Hilbert-Schmidt: operators with $\sum ∥ T e_{n} ∥^{2} < \infty$ for some (equivalently, every) orthonormal basis ${e_{n}}$ . Trace-class: operators with $\sum ∥ T e_{n} ∥_{H} < \infty$ for an orthonormal basis (equivalently, $T$ is a finite product of Hilbert-Schmidt operators). For trace-class $T$ , $tr (T) := \sum_{n} ⟨ e_{n}, T e_{n} ⟩$ is well-defined and basis-independent. These are Schatten ideals $S^{p}$ for $p = 2$ and $p = 1$ respectively.

Atkinson's theorem. $T \in B (X)$ is Fredholm iff there exists $S \in B (X)$ with $S T - I, T S - I \in K (X)$ — i.e., iff the image of $T$ in the Calkin algebra $B (X) / K (X)$ is invertible. This is the operator-algebraic content of Fredholm theory 03.09.06 and the input to the Atiyah-Singer index theorem 03.09.10.

Approximation property. A Banach space $X$ has the approximation property (AP) if the identity is a limit of finite-rank operators in the strong operator topology. On Hilbert spaces and $L^{p}$ -spaces (for $1 \leq p < \infty$ ), AP holds; on general Banach spaces it can fail (Enflo, 1973). When AP holds, finite-rank operators are dense in compact operators in operator norm.

Synthesis. This construction generalises the pattern fixed in 02.11.01 (bounded linear operators), with the symmetric data replaced by its skew or twisted analogue. Read in the opposite direction, the construction is dual to the metric story: complements and orthogonality are taken with respect to the bilinear datum of this unit, not a metric, and the resulting category of subobjects is the one the rest of the strand classifies. The central insight is that this datum identifies algebra with geometry: functions become vector fields, subspaces become quotients, and invariants become cohomology classes — and that identification is the engine driving every theorem downstream.

Full proof set [Master]

$K (X)$ is a closed two-sided ideal. Proved in §"Key theorem".

Schauder's theorem. Sketched in Exercise 7; full proof uses Banach-Alaoglu + Arzelà-Ascoli on the relatively compact image $T (B_{X})$ .

Riesz-Schauder spectral theorem. Let $λ \neq = 0$ be in $σ (T)$ for $T$ compact. Then $T - λ I$ is not invertible. Write $T - λ I = - λ (I - λ^{- 1} T)$ . Riesz's lemma: if $∥ S ∥$ is sufficiently controlled (which $λ^{- 1} T$ is, modulo finite-dimensional subspaces), then $I + S$ is "Fredholm of index 0" — surjectivity is equivalent to injectivity. So $λ I - T$ has a nonzero kernel iff it has a nonzero cokernel, both finite-dimensional. The eigenvalue $λ$ thus has a finite-dimensional generalised eigenspace. Discreteness of the non-zero spectrum follows from the fact that compact operators have "isolated points away from $0$ " — concretely, if $λ_{n} \to λ \neq = 0$ in $σ (T)$ , the corresponding eigenvectors $x_{n}$ form an orthonormal-like sequence, but compactness forces them to converge to zero, contradiction.

Density of finite-rank operators in $K (H)$ . Proved in Exercise 6. The argument uses that $H$ has a countable orthonormal basis (Hilbert-space separability) and that compact images are totally bounded.

Atkinson's theorem. Proved in unit 03.09.06 §"Key theorem" (Fredholm operators). The Calkin algebra is the natural quotient encoding compact-modulo-bounded structure.

Connections [Master]

Bounded linear operators 02.11.01 — $K (X) \subseteq B (X)$ as a closed two-sided ideal.
Banach space 02.11.04 — domain and codomain of the operators.
Vector space 01.01.03 — underlying linear structure.
Hilbert space 02.11.08 — where the spectral theorem and Hilbert-Schmidt / trace-class refinements are sharpest.
Fredholm operators 03.09.06 — Fredholm $⟺$ invertible modulo $K$ in $B (X) / K (X)$ (Atkinson).
Atiyah-Singer index theorem 03.09.10 — index theory of elliptic operators uses compact resolvents on Sobolev pairs.

Historical & philosophical context [Master]

Hilbert's 1906 work on integral equations introduced what we now call compact operators (in the special case of integral operators with continuous kernels). Riesz's 1918 spectral theory abstracted Hilbert's framework to operators on arbitrary Banach spaces, identifying the discrete-spectrum / finite-eigenspace structure that characterises compact operators. Schauder's 1930 theorem cemented compactness as a self-adjoint property under operator-theoretic duality.

The realisation that compact operators form a two-sided ideal — and that the Calkin algebra $B / K$ is the natural setting for index theory — was Atkinson's 1951 contribution ^{[Reed-Simon §VI]}. Schatten and von Neumann developed the trace-class / Hilbert-Schmidt refinements in the 1940s, opening the door to the modern theory of operator algebras and noncommutative geometry.

In quantum mechanics, compact operators are the bounded observables with discrete spectra: the Hamiltonian of a confined system, density matrices (which are positive trace-class operators of trace one), and projection onto finite-energy subspaces. The infinite-dimensional generalisations of matrix-mechanics — eigenvalue expansions, perturbation theory — all live in the compact-operator universe.

Bibliography [Master]

Reed, M. & Simon, B., Methods of Modern Mathematical Physics, Vol. I, Academic Press, 1980. §VI.
Conway, J. B., A Course in Functional Analysis, 2nd ed., Springer, 1990. §II.4.
Schaefer, H. H., Topological Vector Spaces, Springer, 1971. §III.
Riesz, F., "Über lineare Funktionalgleichungen", Acta Mathematica 41 (1918), 71–98.
Atkinson, F. V., "The normal solvability of linear equations in normed spaces", Mat. Sb. 28 (1951), 3–14.

Wave 2 Phase 2.3 unit #6. Compact operators — the closed ideal in $B (X)$ underlying Fredholm theory and the spectral theory of "almost finite-dimensional" operators. Final unit of the Phase 2.3 catch-up batch.