From pointwise information to uniform control
An operator estimate often has the form , with one constant working for every vector. Knowing instead that each fixed vector has a bound is weaker: the bound may initially depend on the vector. Completeness bridges this gap for families of bounded linear maps. The bridge is the Baire category theorem.
Definition 2.1.1
(Nowhere dense set)
A subset of a metric space is nowhere dense if the interior of its closure is empty: .
For example, a closed set with empty interior is nowhere dense, but a set with empty interior need not be: has empty interior in and is dense there, so its closure is all of .
Theorem 2.1.2
(Baire category theorem)
Let be a nonempty complete metric space and let be dense open subsets of . Then is dense in .
Proof
(Nested closed balls with shrinking radii)
To prove density, begin with an arbitrary nonempty open ball . Because is dense and open, is a nonempty open set. Choose a closed ball contained in it, with . Inductively, the open ball meets the dense open set , so choose a closed ball inside with . Here ; choosing the radius small enough makes this closed ball lie in the desired open set, even if it is not the closure of the open ball.
For , nestedness gives , hence . The centers are Cauchy, so completeness supplies . For each fixed , all later centers lie in the closed set , so . Also . Thus every nonempty open ball meets the intersection, proving density.
Taking complements shows that a countable union of closed sets with empty interior has empty interior. In particular, if is a countable union of closed sets, at least one has nonempty interior. Replacing nowhere dense sets by their closures shows that a nonempty complete metric space cannot be a countable union of nowhere dense sets. The conclusion concerns countable unions; it does not say one closed set covers a whole ball because it is large by measure.
Worked example 2.1.3
(The irrationals, and why is no counterexample)
Enumerate the rationals as . Each is open and dense, and their intersection is the set of irrational numbers. The theorem explains why every open interval contains an irrational.
In contrast, with its usual metric, is a countable union of singleton closed sets with empty interior relative to . It is not complete: rational approximations to form a Cauchy sequence with no rational limit. This does not contradict the theorem; it shows that completeness is needed.
Quick check
(Is a dense countable intersection of open sets necessarily open? Can countability be dropped from the Baire category theorem?)
Look again at the irrationals.
Solution · Answer
It need not be open. The irrational numbers above are a dense countable intersection of open sets, and they have empty interior because every interval also contains a rational. Countability cannot be dropped: every is dense and open, but the intersection over all is empty.
Uniform boundedness
Theorem 2.1.4
(Uniform boundedness principle)
Let be Banach, normed, and . If for every , then .
Proof
(A ball inside one level set)
Set . Each is closed as an intersection of closed sets, and the pointwise hypothesis says . By the Baire category theorem, some contains an open ball . If , then both and belong to , so for all . For , apply this to : then . Hence for every .
Proof X-ray
Where completeness turns pointwise bounds into one bound
The closed sets encode bounds on individual vectors. Completeness enters through Baire's theorem, which puts a whole ball inside one . Subtracting the ball's center turns that local bound into a bound near zero; linear scaling then makes the same constant work for every vector. Neither pointwise boundedness alone nor Baire's theorem without linearity gives that final operator-norm estimate.
The theorem does not require the index set to be countable. Its countable sets are the levels , and it is completeness of the domain that is used.
Counterexample mode
Pointwise bounds on an incomplete domain
The tempting claim is that pointwise bounded families of bounded functionals on any normed space are uniformly bounded. Let , the finite-support sequences, with the supremum norm, and define . Each is bounded with , attained at . Each fixed has only finitely many nonzero coordinates, say for , so . The family is pointwise bounded, yet . The missing hypothesis is completeness: is not closed in , and the principle holds once the domain is Banach.
The most frequent use of the principle concerns limits of operators.
Worked example 2.1.5
(Pointwise convergence gives a bounded limit)
If , with Banach, and converges in for every , then each sequence is bounded, so the family is pointwise bounded. Uniform boundedness supplies with for all . The pointwise limit is linear, since limits respect sums and scalar multiples, and letting in gives .
Common mistake
(A pointwise limit is not a norm limit)
The bounded pointwise limit need not satisfy . On , the tail projection tends to zero on each vector, because , but for every .
Constructing a limit from a dense set
The usual dense-set test for strong convergence assumes that the candidate limit is already a bounded operator. Often the limit is not yet known. Completeness of the target lets us construct it.
Theorem 2.1.6
(Strong limit from convergence on a dense set)
Let be normed, Banach and bounded and linear with . If converges in for every in a dense subset , then there is a bounded linear such that for every , and . The set need not be a subspace.
Proof
(Fix the approximant before choosing the index)
Fix . For and any ,
Given , first choose so the first term is less than ; this is possible also when . Keep that fixed and choose so the second term is less than for . Thus is Cauchy in . Completeness gives its limit . Taking limits in the linearity identities proves linear, and proves the bound. The order of choices matters: a different approximant for each would not allow the assumed convergence at a fixed point to be used.
Worked example 2.1.7
(Coordinate truncations on )
Let and let keep the first coordinates. Then , and eventually for every finite-support sequence . These sequences are dense, so the theorem constructs the identity as the limit on all of : indeed .
Both hypotheses have work to do. On , define . These maps converge to zero on the dense set but not at , since ; their norms are , so there is no common bound. In fact is in but . For the role of target completeness, take and with its inherited norm, and regard the same truncations as maps . They are uniformly bounded and converge on , but for the Cauchy sequence has no limit in . Its only possible coordinatewise limit has infinitely many nonzero coordinates.
Quick check
(If are uniformly bounded and converge on a dense subset , why is it insufficient to apply uniform boundedness to the restrictions to ?)
Ask what structure has, and what the example shows.
Solution · Answer
need not even be linear or complete, so the principle does not apply to it. Moreover, pointwise boundedness only on does not imply pointwise boundedness on , as the example shows. The common bound must be established independently before using the approximation estimate.
Open mapping
Definition 2.1.8
(Open map)
A map between metric spaces is open if is open in for every open set .
This property is distinct from continuity and concerns the direction of images rather than inverse images.
Theorem 2.1.9
(Open mapping theorem)
A bounded linear surjection between Banach spaces is open.
Proof
(A ball in the closure, then removal of the closure)
Let denote the open unit ball. Surjectivity gives , a countable union of closed sets, so by the Baire category theorem some contains a ball. Write and suppose . For , both and lie in . The set is symmetric and convex, so the negative of the latter is in , and their midpoint is in . Thus with , and by scaling for every .
The closure must now be removed. Given with , choose with . Inductively, the residual after terms has norm less than , so it lies in ; choose so that the residual after terms has norm less than . The series converges absolutely in the Banach space to some with , and continuity gives . Hence , so contains the ball . If is open and with , linearity gives , so is open.
Proof X-ray
Two completeness hypotheses, two separate jobs
The proof has two stages, and each uses completeness of a different space. Completeness of enters through the Baire category theorem: surjectivity writes as the countable union of the closed sets , so one of them contains a ball. Symmetry and convexity of , both inherited from linearity, move that ball to the origin. At this point the information is only approximate: small vectors of are limits of images of small vectors. Completeness of removes the closure. Each correction halves the residual using a vector of geometrically shrinking norm, so the corrections sum absolutely in , and continuity of passes through the sum. The halving costs only a constant factor in the radius.
The conclusion means that there is a constant such that every has some preimage with .
Common mistake
(Openness does not bound every preimage)
The open mapping theorem does not assert that all preimages are bounded by . If is nonzero, arbitrary multiples of a kernel vector may be added to a preimage, so the set of preimages is unbounded.
Worked example 2.1.10
(A controlled preimage need not be unique)
On with , define . Cauchy–Schwarz gives ; equality at for gives . The map is onto: . A smaller preimage is , with norm . This is minimal because the operator bound forces every preimage to satisfy . Adding any preserves the image and can make the preimage arbitrarily large. Openness promises a controlled choice, not a bound on all choices.
Completeness of the target is as essential as completeness of the domain.
Counterexample mode
An identity onto an incomplete target
Claim to test: a bounded linear surjection from a Banach space onto a normed space is open. On the same vector space of continuous functions, let and , where , and let be the identity. The inequality makes bounded, and it is onto. The image of the open unit ball of consists of continuous functions with supremum norm less than . For , the continuous spike has supremum norm , so it is outside that image, but . Every -ball around zero therefore contains a function outside the image, although zero is in the image, so the image is not open. The repair is that must be complete; here it is not, because continuous approximations to a step function are Cauchy in the integral norm without converging to a continuous function in that norm.
Combined with injectivity, the open mapping theorem gives the most frequently quoted consequence.
Corollary 2.1.11
(Bounded inverse theorem)
If a bounded linear map between Banach spaces is bijective, its inverse is bounded.
Proof
(Openness is continuity of the inverse)
For an open set , the preimage of under is , which is open by the open mapping theorem. Thus is continuous; it is linear, so continuity at zero gives boundedness.
Worked example 2.1.12
(The Neumann series gives an explicit inverse)
Let be a bounded operator on a Banach space with . The series converges in operator norm because is Banach and converges. Since and , passing to the limit yields and . Here the explicit estimate provides more information than the bounded inverse theorem alone.
Theorem 2.1.13
(Closed graph theorem)
For Banach spaces , a linear map is bounded if and only if its graph is closed in , equipped with .
Proof
(Invert the first-coordinate projection)
The forward implication follows from continuity: if and , then , so . Conversely, a closed subspace of the Banach product is Banach. The first-coordinate projection , , is a bounded bijection, so its inverse is bounded by the bounded inverse theorem. Composing that inverse with the second-coordinate projection gives , which is therefore bounded.
Worked example 2.1.14
(A closed, unbounded differentiation operator)
Let carry the supremum norm of the function, let carry its supremum norm, and define . The graph of is closed in , yet is unbounded. This does not contradict the theorem: is not complete in the chosen norm.
To check the graph, suppose uniformly with , and uniformly with . The fundamental theorem of calculus gives . Passing to the limit yields for every . Since is continuous, ; thus lies in the graph. For unboundedness take . We have , but . Closedness of a graph is a statement about limits that already exist in the product; it does not supply the missing completeness of the domain.
Quick check
(Give the graph norm . Is bounded, and is complete in this norm?)
Use the same integral identity for a graph-norm Cauchy sequence.
Solution · Answer
Yes to both. , so is bounded with norm at most one. A graph-norm Cauchy sequence has uniform limits for the functions and derivatives, respectively. Passing to the same integral identity gives and , so convergence is in graph norm. The operator has not changed as an algebraic map; the norm on its domain has.
Practice with solutions
Quick check
(Suppose bounded linear functionals on a Banach space converge pointwise to . Prove is bounded. Must converge to in norm?)
Pointwise convergence gives pointwise boundedness.
Solution · Answer
Pointwise convergence gives pointwise boundedness, so uniform boundedness gives , and . Norm convergence need not follow: on take . Each , but .
Quick check
(Let be a bounded bijection, where is Banach and is merely normed. Give a counterexample to the assertion that must be bounded.)
Compare two norms on .
Solution · Answer
Take the identity from with the norm onto the same vector space with . It is a bounded bijection, while its inverse is unbounded: has supremum norm one and integral norm . The target is not complete in that norm.
Quick check
(For a bounded with , bound the error after replacing by .)
The error is the tail of the Neumann series.
Solution · Answer
The error is , whose norm is at most .