A continuous linear functional on a sequence space can often be described by another sequence. But the answer changes with the norm and with the space on which the functional acts. The identification is a useful first example: it explains why coordinate tests, together with a uniform bound, detect weak convergence in , and why is not reflexive.
The space and its finite sequences
Definition 3.4.1
(The space and finitely supported sequences)
Work over or . Let
and let denote the sequence with a 1 in coordinate and 0 elsewhere. Write for the sequences with only finitely many nonzero coordinates.
We use linear functionals, also over . The finite-support sequences are dense in : for , the truncation obeys
The space is Banach. The proof of completeness of the ambient and closedness of appears in Completion of a normed space, under “Why the supremum-norm argument works”. The point to carry forward is the order of the estimates: for a sup-norm Cauchy sequence, one Cauchy index controls all coordinates at once. After taking each coordinate limit, that same bound gives convergence in the sup norm. Coordinatewise convergence on its own does not give such a bound.
To see why the limit remains in , choose one sufficiently late term within of the limit in sup norm. Since on a sufficiently late tail, there. Hence is closed in the Banach space and is itself Banach. This completeness matters later when uniform boundedness is applied to .
Every continuous functional is a summable sequence
For define
The series is absolutely convergent and , so is continuous. In the complex case the coefficients in (2) are not conjugated; this convention keeps linear in .
Common mistake
(Conjugating the coefficients in the complex case)
The Hilbert-space habit of writing produces a map with , which is conjugate-linear and therefore not an element of . The representation in (2) pairs with without conjugation; conjugation appears only in the auxiliary phase vector of the proof.
Theorem 3.4.2
(The dual of )
Every has a unique representation with , and . Thus is a linear isometric isomorphism from onto .
Proof
(Testing finitely many phases)
Set . For each , choose with when and , and zero elsewhere. This formula also works over the reals. Since ,
The displayed value is real and nonnegative, even when the scalar field is complex. Its partial sums are bounded, so and . For any finitely supported , linearity gives . Given , apply this to its truncations and pass to the limit using (1), continuity of , and absolute convergence of (2). Hence . The estimate before the theorem and (3), now applied to , give . Finally proves uniqueness.
Proof X-ray
Why finite phase choices suffice
The vector used to test the first coefficients is finitely supported, so it is in even if an infinite sequence of phases would fail to tend to zero. The same bound holds for every ; only after this finite test do we take the limit and obtain summability.
Continuity of is used twice: once for the bound (3) and once to pass from the truncations of to itself, which density (1) makes possible. The isometry closes by combining (3), which gives , with the elementary estimate in the other direction.
Worked example 3.4.3
(A norm that is approached but not attained)
Let . Its norm is . The truncations have norm one and , so they approach the norm. No single with attains it: equality would force at every coordinate, contradicting . A functional can have a well-defined norm without attaining that norm on the unit ball.
In fact, a nonzero attains its norm in absolute value on the unit ball of if and only if has finite support. Finite support permits the phase choice above to be made for every nonzero coefficient at once. If support is infinite, choose with for all . Some lies in that tail, so
Worked example 3.4.4
(A finitely supported functional attains its norm)
Take , so . The phase choice gives with and . Any in the unit ball with , works, whatever its later coordinates, because they meet zero coefficients. This is the finite-support half of the criterion; the previous example shows the infinite-support half.
The bidual and a missing limit
We also recall, and briefly prove, with the same linear pairing . Bounded clearly defines a continuous functional. Conversely, for put ; then . Density of in gives , and testing unit vectors gives .
Concept lensStructural
The canonical embedding as an inclusion
Consequently is isometrically identified with . The canonical embedding is defined by . Under the identifications above, is the same coordinate sequence : its value on is . The bounded constant sequence is in but not in , so is not onto. Thus is not reflexive.
This does not mean that the constant sequence is a norm limit of vectors in . The partial sums satisfy
but for . The evaluations in (4) have limits for every functional, yet no vector of represents all these limits: testing would force every coordinate to equal 1. Thus is weakly Cauchy without a weak limit in ; its canonical images have a weak-star limit in the bidual, specifically in . The topology and the ambient space both matter.
Common mistake
(Treating a weakly Cauchy sequence as weakly convergent)
The partial sums are weakly Cauchy, because every converges, but the only candidate limit, the coordinate sequence , is not in . Weak Cauchy sequences in therefore need not converge weakly in . A limit of the numbers for each yields a weak limit only after a vector of the space representing all these limits has been found.
Coordinate tests and a uniform bound
Coordinatewise convergence becomes weak convergence once a uniform bound is added.
Proposition 3.4.5
(Coordinate criterion for weak convergence in )
Let be a sequence in and . Then converges weakly to if and only if
Proof
(Uniform boundedness and a tail estimate)
For necessity, each coordinate is the continuous functional , and uniform boundedness applied to the functionals on gives the common norm bound. The canonical embedding is isometric by Hahn–Banach; alternatively finite coordinate tests show its norm is .
For sufficiency, put . Given , choose with . For this fixed , coordinate convergence makes for large . The tail is bounded by , proving .
The hypothesis cannot be dropped when asserting a weak limit in : the sequence from (4) has bounded norm and coordinatewise limit , which lies outside the space.
Counterexample mode
Coordinatewise convergence without a bound
Let . Each fixed coordinate is eventually zero, so converges coordinatewise to zero. But is in , and does tend to zero; that one test is not enough to decide weakness. Instead take and otherwise. This belongs to , while . Thus the sequence is not weakly convergent, as also follows from its unbounded norms.
For a sequence , weak-star convergence to relative to has the analogous criterion: a common bound plus convergence of each coordinate to . Necessity uses uniform boundedness on the Banach space . Conversely, coordinatewise limits under a common bound themselves form an element of : every finite partial sum is at most . For sufficiency, approximate a test by a finite-support sequence; the tail is at most . In particular in , but does not converge weakly in : coordinate tests force any weak limit to be zero, while the bounded coefficient sequence defines a functional in taking value 1 on every . Weak and weak-star are different topologies here.
Worked example 3.4.6
(Averages of unit vectors in )
Put . Then and each coordinate equals once , so it tends to zero. The weak-star criterion gives ; directly, is the average of the first terms of a null sequence, which tends to zero. Against , however, every takes the value , so does not converge weakly to zero, and by the coordinate tests it has no other possible weak limit. Spreading the mass over more coordinates does not help.
Do not continue the dual chain by analogy
The two identifications above do not identify with by the natural sequence pairing. To see a continuous functional outside the sequence pairing, define on . The expression is unique, and , so is bounded with norm one. Hahn–Banach over the same scalar field extends it to all of without increasing the norm. The extension vanishes on and equals 1 on . If it were for some , testing every would force , contradicting its value at . The extension is an existence argument, not an explicit summation formula.
Practice with solutions
Quick check
(For on , compute and decide whether it is attained on the unit ball.)
Solution · Answer
The coefficients have norm . Their support is infinite, so no vector in the unit ball of attains the norm.
Quick check
(Show that weakly in but not in norm.)
Solution · Answer
Each coordinate tends to zero, and the vectors have common sup norm 1. Criterion (5) gives weak convergence. Their norms remain 1, so norm convergence fails.
Quick check
(Put . Find its coordinatewise limit and decide, with a reason, whether it is weakly convergent in .)
Solution · Answer
Every fixed coordinate is eventually zero, so the coordinatewise limit is zero. The norms equal and are unbounded, so the sequence cannot converge weakly by uniform boundedness; the counterexample above also gives a specific witnessing functional.
Quick check
(For , evaluate on an arbitrary , and compare weak-star and weak convergence.)
Solution · Answer
The pairing is , so relative to . The functional on represented by takes value 1 on every , so weak convergence to zero fails. Coordinate tests make zero the only possible weak limit, so there is no weak limit in .