In Euclidean space, every bounded sequence has a convergent subsequence. In an infinite-dimensional normed space, boundedness alone never gives that guarantee. Riesz's lemma supplies a concrete obstruction: from any proper closed subspace, one can find a unit vector that stays a prescribed positive distance away. Repeating the construction gives a separated sequence in the unit ball.
All normed spaces here are over or . For a nonempty set and a vector , write
The infimum need not be attained. This is why the next result asks for a distance greater than a number strictly below one, rather than asserting the existence of a best approximation.
A vector away from a closed subspace
Lemma 1.2.1
(Riesz's lemma)
Let be a proper closed linear subspace of a normed space . For each , there exists with
Proof
(Normalize an almost-nearest difference)
Choose . Since is closed, : otherwise a sequence of points of would converge to , and would lie in . Since , the definition of infimum gives such that
Set . For every , the vector is in , so
Taking the infimum over gives . Completeness of was not used.
Proof X-ray
Why closed and why strictly below one?
The proof has two inputs. Closedness of is used once, to make ; without it the ratio could be zero. The strict inequality is used once, to make strictly larger than the infimum , so that an almost-nearest exists without assuming a nearest point exists. The key move is invariance of under the affine change : it turns the distance from to into a rescaled distance from to , which is bounded below by . The argument closes because the lower bound does not depend on . In a Hilbert space, orthogonal projection gives a unit vector at distance exactly 1 from a proper closed subspace; the counterexample below shows this stronger conclusion can fail for general norms.
The value is excluded for a reason, and a concrete subspace of shows that it cannot be restored.
Counterexample mode
The endpoint $\alpha$ equal to one can fail
Claim: every proper closed subspace admits a unit vector with . Let , the real sequences tending to zero with the supremum norm, and define . Then , and : the vector with ones has norm one and . Let , a proper closed subspace.
For any , every gives . Conversely, is a point of at distance from ; letting shows . If , then all , and because at least one is strictly smaller than 1; hence . Thus no unit vector has distance 1 from this . The correct statement is Riesz's lemma with , which still gives distances arbitrarily close to 1.
Closedness matters too. The proper subspace of finitely supported sequences is dense in , so every vector has distance zero from it; no works if closedness is dropped.
Worked example 1.2.2
(Distance exactly one in a Hilbert space)
Let be a proper closed subspace of a Hilbert space . Choose and let be the orthogonal projection onto . Then is nonzero and orthogonal to ; set . For every , the Pythagorean identity gives
with equality at . Hence and the infimum is attained. This is the borderline case of Riesz's lemma; the example shows that it depends on the inner product.
Compactness of the closed unit ball
Theorem 1.2.3
(Compact balls characterize finite dimension)
For a normed space , the closed unit ball is compact in the norm metric if and only if is finite-dimensional.
Proof
(A separated sequence, and Heine–Borel in coordinates)
Infinite dimension. Fix and choose a unit vector . Suppose have been chosen. Their span is finite-dimensional and therefore closed, by the norm equivalence recalled in the next paragraph; it is proper because is infinite-dimensional. Riesz's lemma gives a unit vector at distance greater than from that span. Consequently,
Any two distinct terms of the resulting sequence in are more than apart, so it has no Cauchy subsequence and therefore no convergent subsequence. Metric compactness would give a convergent subsequence, so is not compact.
Finite dimension. The zero-dimensional case is immediate. Otherwise choose a basis of and identify it with or . The given norm on is equivalent to the Euclidean norm of the coordinate vector : by Cauchy–Schwarz, . This upper bound makes the given norm continuous in Euclidean coordinates; its positive minimum on the Euclidean unit sphere gives the lower bound. The same equivalence makes any finite-dimensional subspace complete, hence closed in its ambient normed space. Thus is closed and bounded in Euclidean coordinates, and Heine–Borel makes it compact.
A single explicit sequence often replaces the inductive construction.
Worked example 1.2.4
(The standard vectors of )
In , the standard vectors all have norm one, and for ,
No subsequence is Cauchy, so none converges, and the closed unit ball of is not compact. The sequence nevertheless converges weakly to zero: for each , . Weak convergence does not imply norm convergence.
The same sequence refutes the most tempting transfer of a finite-dimensional theorem.
Counterexample mode
Closed and bounded is not always compact
Claim: in a Banach space, a closed bounded set is compact. The closed unit ball of is closed and bounded, and it is complete, yet the sequence inside it has all pairwise distances equal to , so it has no convergent subsequence. The Heine–Borel theorem applies to finite-dimensional normed spaces; by the theorem above, it holds for the closed unit ball exactly when the space is finite-dimensional.
Common mistake
(Separating consecutive terms is not enough)
To show that a bounded sequence has no convergent subsequence, it is not enough that stays large. The sequence has for every , yet its even-indexed terms are all equal to and converge. The construction above separates from every earlier vector, which is why Riesz's lemma is applied to the whole span of .
Local compactness and compact operators
Definition 1.2.5
(Locally compact normed space)
A normed space is locally compact if every point has a compact neighborhood.
Corollary 1.2.6
(Local compactness characterizes finite dimension)
A normed space is locally compact if and only if it is finite-dimensional.
Proof
(Shrink, then rescale a compact ball)
If is locally compact, translation by is an isometry, so zero also has a compact neighborhood . It contains an open ball about zero, and therefore contains a smaller closed ball with . That closed ball is a closed subset of a compact set, hence compact; the continuous map makes the closed unit ball compact, so is finite-dimensional. Conversely, finite-dimensional closed balls are compact, so every point has a compact neighborhood.
The theorem also limits what compact operators can do.
Definition 1.2.7
(Compact operator)
A bounded linear operator between normed spaces is compact if it maps the closed unit ball of to a relatively compact subset of , that is, a set whose closure is compact.
If is infinite-dimensional, the identity is not compact by the theorem. Thus, for infinite-dimensional , no compact operator can be a bounded linear isomorphism onto a normed space : if were bounded, then would be compact. The composition is compact because lies in the image under the continuous map of the compact closure of , and that image is compact.
Worked example 1.2.8
(An injective compact diagonal operator)
On , define by , that is, . It is injective, since forces every . Let keep the first coordinates of and set the rest to zero. Then has finite rank and
with equality at , so . An operator-norm limit of finite-rank operators into a Banach space is compact, so is compact. Its inverse on its range is unbounded, because while .
Common mistake
(Compact does not mean a nontrivial kernel)
The corollary says that a compact operator on an infinite-dimensional space has no bounded inverse; it does not say the operator must kill some nonzero vector. The diagonal operator above is compact and injective. What fails is the boundedness of the inverse on the range, and, as the last practice item shows, the closedness of the range.
Practice with solutions
Quick check
(If is a proper closed subspace, can Riesz's lemma guarantee a unit vector whose distance from exceeds ?)
Read off the allowed values of .
Solution · Answer
Yes: apply the lemma with , which satisfies .
Quick check
(Why is the span of finitely many previously chosen vectors eligible for the lemma in an infinite-dimensional space?)
Two hypotheses must be checked: closed and proper.
Solution · Answer
The span is finite-dimensional, hence complete in the inherited norm and therefore closed. It is proper because the whole space is infinite-dimensional, so it cannot be spanned by finitely many vectors.
Quick check
(Is the closed unit ball of compact under an arbitrary norm?)
Compare the norm with the Euclidean norm.
Solution · Answer
Yes. Every norm on a finite-dimensional space is equivalent to the Euclidean norm, so the ball is closed and bounded in Euclidean coordinates, and Heine–Borel applies.
Quick check
(Why does the sequence in have no norm-convergent subsequence?)
Compute the distance between two distinct terms.
Solution · Answer
Distinct terms are at distance from each other, so no subsequence is Cauchy, and a convergent subsequence would have to be Cauchy.
Quick check
(Can an injective compact operator from an infinite-dimensional Banach space into a Banach space have closed range? Explain.)
Use the bounded inverse theorem on the range.
Solution · Answer
No. If its range were closed, it would be a Banach space, and the operator would be a bounded bijection from the Banach domain onto that Banach range; the bounded inverse theorem would make its inverse bounded. The identity of the domain would then be compact, contradicting infinite dimension.