Extending without losing a bound
A linear functional may be easy to define on a small space but difficult to describe on all of a larger one. Simply choosing values on additional vectors preserves linearity only if the choices are consistent; it need not preserve a norm estimate. Hahn–Banach supplies such a choice, without giving a general formula for it.
Definition 3.1.1
(Sublinear function and domination)
Let be a real vector space. A function is sublinear if
- for all , and
- for all and all .
A linear functional defined on a subspace is dominated by if for every .
A sublinear function need not be symmetric or positive away from zero. Every nonnegative multiple of a norm is sublinear, but so is every linear functional, which takes negative values.
Theorem 3.1.2
(Real Hahn–Banach theorem)
If is a linear subspace of a real vector space , is sublinear, and is linear with on , then there is a linear extension with on .
Proof
(One new direction, then Zorn's lemma)
First extend across one vector . Defining is unambiguous; domination for positive and negative requires
Here is where the two bounds come from. For , divide the required inequality by and put ; positive homogeneity gives . For , write with and put ; rearranging gives the lower bound . These conditions are also sufficient, by reversing the calculation. The same must satisfy them for every , hence the supremum and infimum.
The interval is nonempty: for , subadditivity gives , and therefore . Fix and let vary: every lower candidate is at most . Fix and let vary: every upper candidate is at least . Both candidate sets are nonempty and consist of real numbers, so their supremum and infimum are real. Choosing in the interval gives the required one-dimensional extension, including the case by the original domination.
Order all dominated linear extensions by inclusion of their graphs. A chain has an upper bound: take the union of its domains and functionals; nestedness makes the union linear, well-defined and dominated. Zorn's lemma supplies a maximal extension. If its domain were not all of , the one-dimensional construction would extend it further, a contradiction.
Proof X-ray
Where sublinearity enters, and why choice is needed
The proof combines two independent ideas. The local step uses only the two axioms of sublinearity. Positive homogeneity reduces the requirement for every to the cases and , and these produce the two bounds on . Subadditivity, applied to the splitting , is exactly what places every lower candidate below every upper candidate. No topology and no norm appears in this step.
The global step is set-theoretic. Zorn's lemma produces a maximal dominated extension, and the local step, used once more, turns maximality into the statement that the domain is all of . When is spanned by together with a sequence , ordinary induction and a union of the increasing domains would suffice. Zorn's lemma is what handles an arbitrary, possibly uncountable, family of new directions.
The theorem proves existence; it does not generally prescribe a unique extension, and the one-dimensional step already shows why: any in the interval works.
Common mistake
(Dropping a sign in the lower bound)
Sublinearity does not require to be nonnegative, so sign errors in the interval are not harmless. On let , let be the -axis, and set and . For the correct lower expression is , and the upper expression is . Hence , and the extension is dominated by with equality. An erroneous lower expression would instead equal and have supremum . It would wrongly rule out an extension that is visibly available.
Norm-preserving extension over either scalar field
The normed version follows by choosing the sublinear function to be a multiple of the norm; the complex case is reduced to the real one through the real part.
Theorem 3.1.3
(Norm-preserving extension)
Let be a subspace of a real or complex normed space , and a bounded linear functional. There exists extending with .
Proof
(Dominate by a multiple of the norm)
In the real case apply the theorem to , which is sublinear and dominates on . Applying domination to both and yields , so . Restriction to gives the opposite norm inequality.
For complex spaces, regard as real and extend to a real-linear dominated by ; this is possible because on . Define . This is complex-linear and extends , since . For any , choose a unit scalar for which . Then . Restriction again proves equality of norms. Neither nor needs to be complete.
A finite-dimensional calculation: extensions need not be unique
Worked example 3.1.4
(Norm-preserving extensions in )
Take with , and let be the -axis. The functional has norm one. Every linear extension has the form . Its norm is : the triangle inequality gives , while evaluation at and gives the lower bound. Therefore exactly the parameters preserve the norm.
The interval from the proof gives the same answer. With and , the lower candidates are for , with supremum ; the upper candidates are , with infimum . Both endpoints are attained (at any ), so the closed interval is the whole set of admissible .
For instance, and are distinct norm-preserving extensions. If instead , the extension remains linear but has norm two. Thus linearity alone does not maintain the estimate.
Common mistake
(Speaking of “the” Hahn–Banach extension)
The theorem selects no particular extension. In the example above a whole interval of norm-preserving extensions exists, so a proof may not use properties of “the” extension beyond the two that are guaranteed: it agrees with on and has norm .
Functionals detect norms and distance
Corollary 3.1.5
(Norming functionals)
Let be a normed space and with . There is with and . Consequently, for every ,
and the supremum is attained.
Proof
(Extend from the line through )
For , define on the one-dimensional span of . Since , its norm is one, so the extension theorem gives with and . Every with satisfies , so the supremum equals and is attained. For both sides are zero, attained by .
Thus continuous linear functionals separate distinct points, even when no inner product is available: if , a norming functional for takes different values at and . The supremum is attained by a suitable functional, although a given functional need not attain its own norm on the unit sphere.
A stronger construction separates a point from a closed subspace.
Corollary 3.1.6
(Separating a point from a closed subspace)
Let be a closed subspace of a normed space , let , and put . Then , and there is with , and .
Proof
(A functional on with norm )
If , then is a limit of points of , and closedness gives ; hence . On define . The representation is unique because , and for ,
Thus ; taking vectors with approaching shows equality, since . Extend without increasing its norm. The result vanishes on , takes value one at , and has norm .
Scaling the functional by gives a norm-one functional that vanishes on and takes value at . This yields a second description of the distance itself.
Concept lensStructural
Distance read from annihilating functionals
For a closed subspace and any ,
If is admissible, then for every , which gives the inequality from right to left. If , the scaled separating functional attains the value ; if , both sides are zero. A metric quantity, the distance to , is therefore the largest value of a family of linear quantities. With this is the norming formula above.
Closedness enters the proof only to guarantee . That is not a technicality.
Counterexample mode
A dense proper subspace cannot be separated
Claim: for every proper subspace and every , some vanishes on and satisfies . Take with the supremum norm, the subspace of polynomials, and . A polynomial of degree has vanishing st derivative, while every derivative of is , so . By the Weierstrass approximation theorem there are polynomials uniformly. If vanishes on , continuity gives , so no admissible has . The repaired statement requires closed, or more generally .
When the subspace is closed, an explicit separator can sometimes be written down.
Worked example 3.1.7
(An explicit separator in )
In with the supremum norm, let , which is closed because uniform convergence implies convergence at , and let . Every satisfies , and attains that distance, so . The functional vanishes on , satisfies with equality at the constant , so , and . This example gives an explicit separator; the theorem also works when no explicit formula is apparent.
A dual-norm calculation: moving point evaluations
The dual norm takes a supremum over all unit vectors at each stage. A concrete calculation shows why convergence on every fixed vector need not imply convergence in that norm.
Worked example 3.1.8
(Distances between point evaluations)
Let over or , with the supremum norm, and define for . Evaluation is linear and . The constant function attains equality, so .
If , the triangle inequality gives . To obtain the reverse inequality, choose a continuous real-valued function with , and . For instance, use the straight line joining these two values between and , and extend it constantly on either side. Then . Thus
The edge case is the only one in which the distance is not , however close and are.
Common mistake
(Convergence on each vector is not convergence in norm)
Suppose with for every . Every fixed is continuous, so ; in the language of duality this is weak-star convergence in . It is tempting to conclude in , but the calculation above gives for every . The function witnessing the norm changes with and becomes steeper as approaches . Pointwise convergence controls one fixed test function at a time, while the norm takes the worst test function afresh at every stage.
Quick check
(For , compute in the dual of .)
Treat and separately.
Solution · Answer
If , the functional is , so the norm is . If , the triangle inequality gives at most ; the function above gives value , so the norm is . In particular, closeness of and does not change this answer as long as they are distinct.
Practice with solutions
Quick check
(If vanishes on a dense subspace , prove that .)
Use continuity of .
Solution · Answer
For any , choose with . Continuity gives . Hence .
Quick check
(In with the norm, find a norm-one functional taking value at .)
Match the signs of the coordinates.
Solution · Answer
Take . Since and equality holds at , its norm is one; .
Quick check
(Let be a proper closed subspace of a normed space . Show that some nonzero continuous functional vanishes on .)
Pick a vector outside .
Solution · Answer
Since , choose . The separation corollary yields with and ; in particular .