A continuous linear functional on a sequence space may at first look like an abstract rule depending on infinitely many coordinates. For with , its action is completely determined by one sequence in the space with conjugate exponent. This identification also turns the abstract bidual and weak-convergence definitions into concrete calculations.
The spaces and Hölder's inequality
Work over or , and use linear, not conjugate-linear, functionals.
Definition 3.3.1
(The spaces and the conjugate exponent)
Fix and let , so ; is the conjugate exponent of . Write
The unit vector has in coordinate and elsewhere.
Minkowski’s inequality, proved below, shows that the displayed expression is a norm. The space is Banach. For completeness, a Cauchy sequence is coordinatewise Cauchy; let be its coordinate limits. If all sufficiently late pairs are within in norm, pass to the coordinate limit in each finite partial sum to get for every . For fixed sufficiently large , these bounds show ; since and is a vector space, . The same bound holds for the full norm. Finite-support sequences are dense by truncating the convergent defining series.
Theorem 3.3.2
(Hölder's inequality)
For and ,
Proof
(Normalize, then apply Young coordinatewise)
For finite sums, normalize both nonzero sequences and use Young's inequality for nonnegative on each coordinate. Summing gives at most ; remove the normalization and pass to the increasing limit of partial sums. Zero sequences are immediate.
Young's inequality can be verified as follows: for fixed , the function on has its minimum at , where its value is ; is immediate.
In particular, the pairing used below is absolutely convergent.
Why the triangle inequality holds
Theorem 3.3.3
(Minkowski's inequality)
If , then and .
Proof
(Finite sums first, then the limit)
The subtle step is to show that the sum of two -summable sequences is still -summable, while obtaining the triangle inequality. Work first with a fixed finite number of coordinates, so no convergence or prior vector-space structure is needed. Write . Using the scalar triangle inequality and finite Hölder with exponents ,
Here . If , the desired inequality is immediate; otherwise divide by . The right-hand side is at most for every . Therefore the increasing partial sums defining are bounded, so , and passing to their limit gives the inequality.
Scalar multiplication preserves -summability, and positive definiteness and absolute homogeneity are immediate from the definition. Hence this really is a norm.
Worked example 3.3.4
(Strict inequality, equality, and the failure for )
Take , and . Their norms are both , but . The inequality can be strict; when , homogeneity gives , so equality also occurs.
For the same disjoint vectors instead give : the displayed formula is then not a norm. Thus the exponent hypothesis is doing real work. At the edge case , the triangle inequality follows directly by summing the scalar triangle inequality, but the finite-conjugate-exponent proof above does not apply.
Quick check
(For and , , compute both sides of Minkowski's inequality.)
Add the vectors first.
Solution · Answer
, so the left side is ; each of is , so the right side is .
Endpoint: completeness of
The duality theorem below, , is stated only for . The result in this subsection is separate from it. It concerns completeness alone, and it holds at the endpoint .
Let be or . Let be the set of sequences in with .
Vector space and norm. Fix a finite . The scalar triangle inequality gives . The left side increases in and is bounded. Letting therefore shows that and . Homogeneity and definiteness follow directly from the definition.
Theorem 3.3.5
(Completeness of )
is a Banach space.
Proof
(Coordinate limits, then a bound uniform in )
Let be Cauchy in . Since , each coordinate sequence is Cauchy in . Call its limit . At this point we do not yet know that lies in .
Given , choose with for all . Fix and a finite . For every we have . This is a finite sum, so we may let term by term, which gives .
The bound does not depend on . Hence and . Because , closure under addition gives , and in norm.
Density of finite support. For , let be its truncation. Then . This is the tail of an absolutely convergent series, so it tends to .
Common mistake
(Coordinatewise convergence is not norm convergence)
The unit vectors tend to in every coordinate. Yet , and for , so is not even Cauchy. A coordinatewise limit is only a candidate; norm convergence needs the tails controlled uniformly, as in the proof above.
Worked example 3.3.6
(Truncations of in )
Let and . Since , we have . Moreover , using that for . Here the coordinatewise limit is also the norm limit, because the tails are summable and shrink.
Completeness alone does not establish endpoint duality or reflexivity; those require separate arguments.
Every functional comes from a sequence
Theorem 3.3.7
(The dual of for )
For each , the rule
defines a continuous linear functional on with . Conversely, every is uniquely for some . Thus is a linear isometric isomorphism .
Proof
(Test the functional on finite-support vectors)
Hölder gives continuity and . Conversely, let and set . For a finite set of the first coordinates, define
and set the remaining coordinates to zero. In the real case the bar has no effect. Put . Because , this choice satisfies and . If , boundedness of yields
The zero case is automatic. The increasing partial sums are uniformly bounded, so and . On finite-support sequences, linearity gives . Truncation and continuity extend the identity to every ; absolute convergence follows from Hölder. The upper bound for now gives . Uniqueness follows by evaluation on each .
Proof X-ray
The complex conjugate belongs in the test vector
The candidate sequence is forced: is the only possibility, which already gives uniqueness. The real work is to show , and the proof does it with one test vector per . Taking makes each product real and nonnegative, so no cancellation occurs in , while the identity makes the same sum . Boundedness of then reads .
Two hypotheses are used afterwards. Density of finite-support sequences, which needs , extends from finite support to all of . Linearity of is why the conjugate sits on the test vector and not in the representation: conjugating the input in the definition of would produce a conjugate-linear map, not an element of the linear dual used here.
Worked example 3.3.8
(The norm of on )
For , , the functional is with , so
The test vector of the proof is , that is , with norm . Hence a unit vector attaining the norm is , and indeed .
The canonical bidual map is onto
Let be the canonical map .
Theorem 3.3.9
(Reflexivity of )
For , the canonical map is an isometric isomorphism onto ; that is, is reflexive.
Proof
(Apply the duality theorem twice)
Take any and define on . The dual identification is linear and isometric, so is a continuous linear functional on . Applying the theorem with and interchanged produces such that
Every equals some , so for all . Thus is onto. It is also an isometry: since every dual functional is some and , the theorem gives , by applying the norm equality with exponents interchanged.
This proves surjectivity of the canonical map, not merely an abstract resemblance between a space and its bidual.
Weak convergence is bounded coordinatewise convergence
Theorem 3.3.10
(A weak-sequence criterion)
For a sequence in , , the following are equivalent:
- converges weakly to some .
- and each coordinate converges to a scalar .
In the second condition the coordinate limit automatically belongs to , and it is the weak limit.
Proof
(Uniform boundedness one way, head and tail the other)
Weak convergence makes each coordinate converge because is continuous. It also makes the norms uniformly bounded: apply the uniform boundedness principle to the evaluation maps on the Banach dual ; for every functional their values converge, and .
Conversely, suppose the norms are at most and the coordinates converge to . For every finite ,
Thus and . For , split the pairing at . The finite head tends to zero, while Hölder bounds the tail by
Choose so the tail is small, then so the finite head is small. Every functional is represented by such an , proving weak convergence.
Proof X-ray
Why the uniform bound cannot be omitted
The two directions use different tools. From weak convergence to boundedness, the key input is the uniform boundedness principle, which needs the completeness of and the isometry ; coordinate convergence is free.
In the converse, coordinate convergence controls only a fixed finite head. The tail is controlled by a product of two factors: , which is small because makes the tail of a -summable sequence tend to zero, and , which is where the uniform bound enters. The order of choices closes the argument: is fixed first, independently of , and only then is taken large. Without the bound the second factor may grow with , and without the first factor need not shrink; the counterexamples below show both failures.
For example, the unit vectors converge weakly to zero in every covered here, because they are bounded and each fixed coordinate eventually vanishes. Their norms remain one, so they do not converge in norm.
Counterexample mode
Endpoint and boundedness assumptions matter
In , the sequence converges coordinatewise to zero but is not weakly null: the continuous functional represented by takes value on every . Its norms are unbounded.
The exponent condition also matters. In , the bounded unit vectors converge coordinatewise to zero, but the continuous functional has . The dual coefficient sequence lies in , whose tails do not tend to zero in sup norm. Thus bounded coordinatewise convergence need not imply weak convergence at ; weak convergence still implies coordinatewise convergence because coordinate functionals are continuous.
Practice with solutions
Quick check
(For , find the conjugate exponent and the norm of .)
Use the duality theorem.
Solution · Answer
and the norm is .
Quick check
(If represents the zero functional, what is for every ?)
Evaluate on unit vectors.
Solution · Answer
for all , so .
Quick check
(Does converge weakly to zero in ? Does it converge in norm to zero?)
Check the two conditions of the criterion.
Solution · Answer
Weakly yes, by the criterion: the norms are bounded and each coordinate is eventually zero. In norm no, since every norm is .
Quick check
(Why do the partial sums in the proof of the weak-sequence criterion establish that the coordinate limit is in ?)
Look at monotonicity and the bound.
Solution · Answer
The nonnegative partial sums are increasing in and uniformly bounded by , so their limit is finite.
Quick check
(In , let . Find its weak limit and decide whether it converges in norm.)
Compute and a fixed coordinate.
Solution · Answer
Each norm is , while every fixed coordinate equals once . The criterion gives weak convergence to zero, but the unchanged norms rule out norm convergence.