Why compactness makes diagonalization possible
A real symmetric or complex Hermitian matrix has an orthonormal eigenbasis. In an infinite-dimensional Hilbert space, self-adjointness alone does not ensure such a discrete list of eigenvectors: the identity already shows that a bounded self-adjoint operator need not be compact. Compactness supplies the missing finite-dimensional behavior. Its nonzero eigenvalues can be enumerated with multiplicity, and only zero can be a limit of an infinite list.
Throughout, is a separable real or complex Hilbert space, and the inner product is linear in its first argument. Let be compact and self-adjoint, so is bounded and . We will prove the eigenvector expansion directly. We use the Hilbert projection theorem, orthonormal-series convergence and Parseval's identity for a closed orthonormal span, and the fact that a compact operator sends bounded sequences to sequences with convergent subsequences. These are the earlier Hilbert and compact-operator results; no general spectral theorem is assumed.
The theorem, including zero
Theorem 5.1.1
(Compact self-adjoint spectral theorem)
There is a finite or countably infinite orthonormal family of eigenvectors with nonzero real eigenvalues , repeated according to multiplicity, such that
The sum converges in norm. If the family is infinite, ; each nonzero eigenvalue has finite multiplicity. For every ,
If , the family is empty and . If the construction stops after finitely many vectors, the sums are finite and their span is the range of . The kernel may be zero, finite-dimensional, or infinite-dimensional. In finite-dimensional the family is necessarily finite; an orthonormal basis of the kernel can be appended to obtain an ordinary matrix eigenbasis. In the separable infinite-dimensional case a finite or countable orthonormal basis of the kernel may likewise be appended, but the displayed formula does not require us to enumerate it. The proof occupies the next two sections: first one eigenvector, then the whole family.
Common mistake
(Expecting zero to be an eigenvalue)
The theorem separates off , but it does not say that . An injective compact operator on an infinite-dimensional space can have nonzero eigenvalues tending to zero while its kernel is zero; the diagonal example below is one. Zero is then a limit of eigenvalues without being an eigenvalue. The theorem also does not claim that every bounded self-adjoint operator has a countable eigenvector expansion of this form.
First step: find an extremal eigenvector
Lemma 5.1.2
(Real eigenvalues and orthogonal eigenspaces)
Every eigenvalue of a self-adjoint operator is real, and eigenvectors belonging to distinct eigenvalues are orthogonal.
Proof
(Move across the inner product)
Self-adjointness makes real. If with , then is real, so is real. If also with , then
using , so .
The key lemma obtains an eigenvector without assuming the unit sphere is compact.
Lemma 5.1.3
(Extremal eigenvector)
If , there is a unit vector and a real such that and .
Proof
(Maximize the quadratic form, then use compactness once)
Put . Clearly . For unit , choose a sign in the real case, or a complex phase multiplying , so that is real and nonnegative. Self-adjointness and the parallelogram identity give
Homogeneity extends the bound defining to nonunit vectors, including zero. Taking the supremum over unit gives , so .
Choose unit vectors with . By passing to a subsequence, the real numbers tend to or . Compactness supplies a further subsequence for which converges. Along it,
Because , converges to some unit vector . Continuity of gives .
Proof X-ray
Where each hypothesis enters the extremal lemma
Self-adjointness is used twice. It makes real, so a maximizing sequence can be pushed toward one sign ; and it turns the polarization identity into the bound , so the quadratic form detects the full operator norm. For a non-self-adjoint operator such as the nilpotent matrix below, the quadratic form is strictly smaller than the norm and no eigenvector of modulus exists.
Compactness is used exactly once, on the image sequence . The non-obvious move is the identity for : it shows the vectors are almost eigenvectors, so convergence of transfers to after dividing by . This division is why is needed. Without compactness the near-eigenvectors need not converge: for the multiplication operator discussed below, comes arbitrarily close to but no eigenvector exists.
Common mistake
(Applying compactness to the unit sphere)
A compact operator does not make the unit sphere of compact; in infinite dimensions it never is. A merely bounded sequence of unit vectors, such as an orthonormal sequence, has no convergent subsequence. The proof extracts a convergent subsequence of only, and recovers convergence of from the almost-eigenvector identity.
Constructing the expansion
The lemma gives one eigenvector; repeating it on orthogonal complements gives all of them.
Proof
(Peel off one eigenvector at a time)
Suppose eigenvectors with nonzero eigenvalues have been chosen orthonormally. Their orthogonal complement is closed and is invariant under : for ,
The restriction is compact and self-adjoint. If it is nonzero, the extremal lemma supplies a unit eigenvector with . Begin in the same way with . The norms of the restrictions do not increase, so the magnitudes of the selected eigenvalues are nonincreasing.
If a restriction becomes zero after choices, then , and . Conversely, a kernel vector is orthogonal to every nonzero-eigenvalue vector by the first lemma, so . This also covers , with .
Suppose instead the process continues indefinitely. Compactness forces . Otherwise, since the magnitudes are nonincreasing, they stay above some . But for ,
so has no convergent subsequence, contradicting compactness. Let be the closed span of the selected eigenvectors. The orthogonal complement lies in every and is invariant under . Thus for ,
Hence . The reverse inclusion follows from orthogonality of a kernel vector to each nonzero-eigenvalue vector, so . Projection onto and the orthonormal-series theorem now give the two expansions in the statement. Applying bounded to the norm-convergent expansion gives the formula for .
The list contains every nonzero eigenvalue with its full multiplicity. If with , then , and comparing with shows unless ; so the -eigenspace is spanned by the selected with . Finally, that eigenspace is finite-dimensional: an infinite orthonormal sequence in it would have images under separated by the fixed distance , contradicting compactness. An infinite list can therefore accumulate only at zero.
Proof X-ray
What keeps the peeling process honest
Self-adjointness makes each complement invariant, so the extremal lemma can be reapplied to a compact self-adjoint restriction; this is the only place the earlier eigenvectors matter. Beyond the lemma, compactness is used twice, both times on an orthonormal sequence: images that stay a fixed distance apart would have no convergent subsequence, which forces and finite multiplicities. The closing move is the estimate on . Because lies inside every , the shrinking restriction norms squeeze to zero there, identifying the leftover space with without choosing a basis of the kernel.
Examples and limits of the hypotheses
Worked example 5.1.4
(An infinite diagonal operator)
On , let . Its finite-coordinate truncations satisfy , so they converge to in operator norm and is compact; its real diagonal entries make it self-adjoint. The standard vectors satisfy , so the family in the theorem is with , each of multiplicity one. If , every coordinate vanishes, so the kernel is zero and . The nonzero eigenvalues tend to zero; zero is not an eigenvalue here, although the inverses of the nonzero eigenvalues are unbounded.
Worked example 5.1.5
(A finite-rank projection with an infinite-dimensional kernel)
On , let be the constant function . The constant function has norm one, and . Thus is an orthogonal projection, self-adjoint and compact (its range is one-dimensional). Its sole nonzero eigenvalue is 1, with eigenspace spanned by ; its kernel is the infinite-dimensional space of functions with integral zero. The theorem becomes , and Parseval reads . This is the edge case in which the construction stops after one step.
Worked example 5.1.6
(The construction picks the eigenvalue of largest modulus first)
On , let . It has rank two, so it is compact, and its block is real symmetric, so is self-adjoint. For a unit vector the quadratic form is in the complex case, and its largest modulus is , attained at , where . So the lemma selects , even though the other eigenvalue is positive.
On the restriction has norm , attained at with , so . On every vector has , so and the construction stops with . Check the expansion on : , and
Concept lensAnalytic
Eigenvalues as successive maxima of the quadratic form
The construction also describes the eigenvalues without solving an eigenvalue equation. The extremal lemma identifies the norm of a self-adjoint restriction with the supremum of its quadratic form, so
attained at . In the rank-two example, is the maximum of over all unit vectors and its maximum over unit vectors in . Each eigenvalue is an extremum of one scalar function on a sphere, constrained by the eigenvectors already found.
The two hypotheses of the theorem cannot be traded for each other. The identity on an infinite-dimensional Hilbert space is self-adjoint but not compact: an orthonormal sequence has no convergent subsequence. It has an eigenbasis in a separable Hilbert space, but its nonzero eigenvalues do not tend to zero, so the compact conclusion cannot be transferred unchanged. A sharper failure follows.
Counterexample mode
Self-adjoint without compactness: no eigenvectors at all
Claim to test: every bounded self-adjoint operator has an orthonormal family of eigenvectors spanning the orthogonal complement of its kernel. Take the multiplication operator on . It is bounded because and self-adjoint because is real. Its kernel is zero, since almost everywhere forces almost everywhere. Yet it has no eigenvector: any eigenvalue is real by the first lemma, and implies almost everywhere, so vanishes off the single point , a set of measure zero. The claim would require eigenvectors spanning all of , and there are none. The repair is the compactness hypothesis of the theorem; is not compact.
Conversely, the matrix is compact on but not self-adjoint. Its only eigenvalue is zero and its eigenspace is one-dimensional, so its eigenvectors cannot form a basis. Compactness without self-adjointness does not give this diagonalization.
Practice with solutions
Quick check
(On , let . Give its nonzero eigenvalues with multiplicities, its kernel, and its spectral expansion.)
Read the eigenvectors off the standard basis.
Solution · Answer
The nonzero eigenvalues are and , each of multiplicity one, with eigenvectors . The kernel consists of all sequences with first two coordinates zero, and .
Quick check
(Let be compact and self-adjoint with and for nonzero . Prove and give a lower bound for .)
Move across .
Solution · Answer
Self-adjointness gives , so . Since , .
Quick check
(For the projection , find the kernel component and the nonzero-eigenvalue component of .)
Compute .
Solution · Answer
, so the nonzero-eigenvalue component is the constant , while has integral zero and is the kernel component.
Quick check
(Explain why an infinite-dimensional eigenspace for a nonzero eigenvalue contradicts compactness.)
Apply to an orthonormal sequence in that eigenspace.
Solution · Answer
Choose an orthonormal sequence in the eigenspace. Then for , so has no convergent subsequence, contrary to compactness.