From vectors to measurements
A vector in a function or sequence space may contain infinitely much information. A continuous linear functional extracts one scalar measurement in a way that is stable under norm convergence. Collecting all such measurements produces another normed space, the dual. An operator can be studied from the other side by asking how it transforms measurements; on Hilbert spaces the Riesz theorem turns that question into an operator on vectors called the adjoint.
Compact operators bring an additional form of control: they send bounded sequences to sequences with convergent subsequences. This is weaker than finite dimensionality but much stronger than boundedness. The distinction matters in equations, where an injective operator may still have an inverse that amplifies small data errors without bound.
All vector spaces below are over or . For Hilbert spaces we use inner products linear in the first argument and conjugate-linear in the second. This convention determines the placement of conjugates in every adjoint formula.
The continuous dual
Definition 2.3.1
(Continuous dual)
For a normed space , its continuous dual is
It contains bounded linear functionals, equivalently continuous linear functionals. Since the scalar field is complete, is Banach even if itself is not: an operator-norm Cauchy sequence converges pointwise, and the uniform operator estimate passes to its limit. Algebraic linear functionals need not be continuous on infinite-dimensional normed spaces; they are not automatically members of .
A functional and its zero set
The kernel of a continuous linear functional is closed: if and , continuity gives . For scalar-valued linear maps the converse is also true. This lets us test stability by studying the geometry of a single hyperplane, even before knowing an operator norm.
Theorem 2.3.2
(Continuity of a linear functional through its kernel)
Let be a real or complex normed space and let be linear. Then is continuous if and only if is closed. If is discontinuous, its kernel is dense in . Completeness of is not needed.
Proof
(A positive distance to the kernel)
The forward implication follows from continuity. Conversely, assume is closed. If , it is continuous. Otherwise choose with . Since and is closed, : an open ball around misses . For any with , the vector belongs to . Therefore
The estimate also holds when , proving boundedness and hence continuity. For the final assertion, if is discontinuous it is unbounded on the unit ball. For each choose with and , and set . Then and . For any , the vectors lie in and tend to . Thus the kernel is dense.
Proof X-ray
Why one scalar equation is special
The normalization identifies as a kernel vector whenever . Closedness gives a positive distance from to that kernel, yielding a bound on . For a map with an infinite-dimensional codomain, a closed kernel by itself does not control all output directions.
Worked example 2.3.3
(Derivative at one point)
Let carry the supremum norm of the function, and let . This is a well-defined linear functional, but the norm does not control derivatives. Put . Then , while : indeed . Thus uniformly although , and is discontinuous. More concretely, for each the functions have derivative zero at and converge uniformly to . Hence the proper hyperplane is dense, not closed. This example uses an incomplete normed space; the theorem itself makes no completeness assumption.
Quick check
(On with the supremum norm, let . Find and determine whether is closed.)
Find an upper bound and a function attaining it.
Solution · Answer
For every , , so . The constant function has norm one and , so . Since is continuous, its kernel is closed, by the forward direction of the theorem.
The theorem uses that takes scalar values. Without that hypothesis the converse direction fails.
Counterexample mode
A closed kernel does not bound an operator
Claim: a linear map between normed spaces is bounded whenever is closed. On with the supremum norm of the function, let , mapping into with its supremum norm. Its kernel is the space of constant functions, a one-dimensional and hence closed subspace. But while , attained at , so is unbounded. The correct statement is the scalar one: there a single normalized vector with and its positive distance to the kernel control every value of .
For concrete sequence spaces the dual can be computed coordinate by coordinate, by testing a functional on the unit vectors.
Theorem 2.3.4
(The continuous dual of )
For , the formula defines an element of , and is a linear bijection of onto with .
Proof
(Coordinates of a functional on the unit vectors)
If , define for . The series converges absolutely and . Conversely, for let . Since , the sequence is bounded. Finite-support sequences are dense in , so continuity gives . The upper bound above and evaluation at show
The supremum need not be attained by any one coordinate. The result is an isometric identification of the continuous dual with ; no inner product on was used.
Every defines a functional on by . The map is linear and . Hahn–Banach supplies a norm-one functional taking value at each nonzero , so in fact . Thus is an isometric embedding. A space is reflexive when this map is onto; being embedded in the bidual alone does not establish surjectivity.
The Hilbert adjoint
Definition 2.3.5
(Hilbert adjoint)
Let be Hilbert spaces and . For fixed , the map is a bounded linear functional on . Riesz representation gives a unique vector, denoted , satisfying
Theorem 2.3.6
(Basic properties of the adjoint)
The map is bounded and linear, , and for composable bounded operators. Also , using the canonical identification of each Hilbert space with itself.
Proof
(Riesz uniqueness and the norm identity)
Uniqueness in Riesz representation and conjugate-linearity in the second inner-product argument show that is linear. For each ,
Thus . Conjugate symmetry gives , so satisfies the defining relation for the adjoint of ; uniqueness yields . Applying the same bound to proves the reverse norm inequality. Finally, , so uniqueness gives the composition rule.
For complex spaces, , not in general. A bounded operator is self-adjoint when . This equation compares two operators on the same Hilbert space; it does not simply mean that their matrices have real entries.
Worked example 2.3.7
(Adjoint of a complex matrix)
With the usual inner product , a matrix acts by and its adjoint is . For ,
The conjugation is essential: transposing without conjugating fails the inner-product identity over .
Common mistake
(Transposing without conjugating)
Over the adjoint of a matrix is its conjugate transpose, not its transpose. For the matrix above, . With and , , but . The conjugate transpose gives , as required. For the same reason .
Worked example 2.3.8
(An integral operator and its adjoint)
On , set . Cauchy–Schwarz gives , hence . For , Fubini's theorem applies to the absolutely integrable product on the triangle, giving
Thus is not self-adjoint: for instance, but . These formulas are understood as equalities of equivalence classes.
Compactness and approximation
Definition 2.3.9
(Compact operator)
A bounded linear operator between normed spaces is compact if the image of every bounded sequence in has a subsequence converging in . Equivalently, the image of the unit ball has compact closure in . Convergence must be in the codomain norm; pointwise convergence of functions alone is not enough.
Proposition 2.3.10
(Finite rank and norm limits)
Every bounded finite-rank operator is compact. If is Banach and compact operators converge to in operator norm, then is compact.
Proof
(Total boundedness passes to norm limits)
A bounded sequence in a finite-dimensional range has a convergent subsequence, proving the first claim. For the second, write for the unit ball of and take a sequence in it. Given , choose with . The image of the unit ball is totally bounded, so it has a finite -net whose centers lie in . The same points form an -net for . Hence that image is totally bounded. Its closure is complete because it is a closed subset of Banach , and a complete totally bounded metric space is compact. Rescaling handles any bounded sequence.
The completeness hypothesis on is used at the last step. For an operator between Hilbert spaces it is automatic. The converse to the first statement is false: on define
Let keep only the first output coordinates. Then has finite rank and , so is compact by the proposition. Its range is infinite-dimensional because it contains every standard unit vector after rescaling its input.
Theorem 2.3.11
(Compact adjoints on Hilbert spaces)
If is compact, then is compact.
Proof
(Finite-rank approximation through a projection)
Write for the unit ball of , let , and take a finite -net for . Let be the orthogonal projection onto the span of the net points. Because is the nearest-point projection, . Thus is finite-rank. Since , its adjoint is finite-rank, and the adjoint norm identity gives . Taking a sequence of such projections with and applying the norm-limit proposition proves compactness. This argument uses the Hilbert projection; it is not a proof for arbitrary Banach-space codomains.
Proof X-ray
Compactness compressed into one finite-rank operator
The key move turns compactness of into a single estimate. Total boundedness of supplies finitely many net points, and for the vector lies within of one of them; since is the nearest point of their span, . Passing to adjoints then costs nothing: makes finite-rank, and transfers the error bound. Completeness of is used by the norm-limit proposition. The Hilbert structure enters twice, through the nearest-point projection and through .
A compact operator need not have a stable inverse
On an infinite-dimensional Hilbert space, the identity is not compact. Construct an orthonormal sequence inductively; its terms lie in the unit ball but satisfy when , so no subsequence converges. Consequently an invertible compact operator between infinite-dimensional Hilbert spaces cannot have a bounded inverse: composing that inverse with the compact operator would make the identity compact. A bounded operator composed on either side of a compact operator is compact, as follows directly from the sequence definition.
The diagonal operator above is injective, but its inverse on its range is unbounded. Indeed , so . The vector lies in but is not in the range: its only coordinatewise preimage would be , which is not square-summable. Small output perturbations can therefore correspond to large changes in input, and surjectivity onto also fails. Compactness is not a blanket guarantee that an inverse problem is stable.
The same orthonormal sequence explains a frequent error in checking compactness.
Common mistake
(Testing compactness with coordinatewise limits)
In the unit vectors satisfy coordinate by coordinate, yet and no subsequence converges in norm. If convergence of coordinates were accepted, the identity would wrongly appear compact. Compactness requires a subsequence converging in the norm of the codomain.
Practice with solutions
Quick check
(For , define on . Find and evaluate .)
Use the isometric identification of with .
Solution · Answer
The sequence has supremum norm one, hence . Directly, , and shows the bound is attained. Also .
Quick check
(On , find the adjoint of and decide whether is self-adjoint.)
Conjugate and transpose.
Solution · Answer
The transpose is , so . This differs from , so is not self-adjoint.
Quick check
(On , let with . Prove that is compact and bound the norm error of its first--coordinate truncation .)
Compare with the diagonal example above.
Solution · Answer
A convergent scalar sequence is bounded, so and is bounded. The truncation has finite rank, and
Hence , and the norm-limit proposition shows that is compact.
Quick check
(Explain why the image of the closed unit ball under the identity on an infinite-dimensional Hilbert space is not compact.)
Look for a sequence in the ball with no convergent subsequence.
Solution · Answer
An orthonormal sequence lies in the unit ball, and distinct terms are distance apart. It therefore has no Cauchy subsequence and no convergent subsequence. A compact set would contain a convergent subsequence of every sequence in it, so the ball is not compact.