An infinite-dimensional unit ball is bounded but need not be compact in norm. The coordinate vectors in stay a fixed distance apart, so no subsequence converges in norm. Yet every fixed continuous linear measurement of those vectors tends to zero. Weak convergence records this second, less demanding form of convergence.
The distinction is useful in minimization. A bounded minimizing sequence may have no norm-convergent subsequence. We will extract a weakly convergent one, prove that closed convex constraints survive its limit, and identify energy hypotheses that prevent the limit from having a larger energy. Each step has a separate role; boundedness by itself does not solve an arbitrary minimization problem.
Mathematical prerequisites and weak convergence
Let be a real or complex Hilbert space, with inner product linear in its first variable. We use these mathematical facts: the Baire theorem for a complete metric space; orthogonal projection onto closed subspaces and onto nonempty closed convex sets; Riesz representation of bounded linear functionals; and orthonormal-coordinate expansion in a separable Hilbert space. In particular the convex projection satisfies for every . An orthonormal basis means an orthonormal family whose finite linear combinations are dense. For subsequence extraction we need basis existence and norm reconstruction only in the separable closed span of the given sequence. These facts are the mathematical inputs to the arguments below; the new compactness and minimization conclusions will be proved here.
Definition 2.5.1
(Weak convergence)
We say converges weakly to , and write , when
By Riesz representation this is equivalent to convergence under every bounded linear functional. The test vector is fixed before taking the limit. One cannot replace it by a different vector depending on . A weak limit is unique: testing the difference of two proposed limits against itself gives zero norm. Norm convergence implies weak convergence by Cauchy–Schwarz; the converse will fail in our first example.
Common mistake
(The test vector must be fixed)
In the unit vectors satisfy : for a fixed , because . Testing against the moving vector gives for every , which does not contradict weak convergence. Weak convergence is a statement about each fixed measurement, with the measurement chosen before the limit.
Theorem 2.5.2
(Boundedness and coordinate tests)
A weakly convergent sequence is bounded in norm. If is an orthonormal basis, then
The basis may be uncountable. The uniform norm bound cannot be discarded.
Proof
(Baire for boundedness, finite approximation for convergence)
First suppose . For each , the scalar sequence is bounded. Define closed sets
Their union is . Baire gives one containing a ball with . For , both and belong to that ball. Subtracting the two scalar evaluations gives
The supremum over unit-ball vectors is , by Cauchy–Schwarz and testing in its own direction. Thus . Coordinate convergence is immediate from the definition of weak convergence.
Conversely suppose the norms are bounded by and the coefficients converge to those of . Given a fixed and , choose a finite basis combination so that . Then
The first term tends to zero because only finitely many coordinates occur. Letting proves weak convergence. This uses finite approximations to each test vector, not an enumeration of a possibly uncountable basis.
Extracting a weakly convergent subsequence
Theorem 2.5.3
(Bounded Hilbert sequences have weakly convergent subsequences)
Every bounded sequence in a Hilbert space has a weakly convergent subsequence. No separability assumption on the whole space is needed. If , the extracted weak limit has norm at most .
Proof
(Diagonal extraction in the span of the sequence)
Let be the closed linear span of . This space is separable: finite combinations of these vectors with rational coefficients, or rational real and imaginary parts, give a countable dense subset. Take a finite or countable orthonormal basis of . The zero-space case is immediate; the finite case uses only finitely many extractions. In the countably infinite case write the basis as .
Each coordinate lies in a bounded scalar disk. Successive subsequence extraction and a diagonal choice give a subsequence whose every coordinate converges, say to . For each finite , Bessel's inequality and a finite limit give
Thus and orthonormal reconstruction gives with these coordinates and . The bounded coordinate criterion proves weak convergence in . To test it against , decompose . All and lie in , so their inner products with the second term vanish. Hence convergence also holds against every test in .
Proof X-ray
The sequence creates its own separable space
A nonseparable Hilbert space can have an uncountable orthonormal basis. There is no diagonal sequence that processes all of those coordinates one by one. The proof instead passes to the closed span of the given sequence, which is always separable. After extracting there, orthogonal projection brings every ambient test vector back into that smaller space. This is a proof for sequences; it does not silently identify sequential compactness with topological compactness in every weak topology.
Which constraints and energies survive the limit?
Definition 2.5.4
(Lower semicontinuity and convexity)
A function is norm lower semicontinuous if whenever in norm. Equivalently its sublevel sets are norm closed for every real . It is convex when
with the usual extended-value interpretation. Weak sequential lower semicontinuity uses the same inequality for weakly convergent sequences. The word sequential specifies exactly which conclusion is being proved.
Theorem 2.5.5
(Weak limits preserve closed convex constraints and convex lower bounds)
If is norm closed and convex, and converges weakly to , then . Also
If is convex and norm lower semicontinuous, it is weakly sequentially lower semicontinuous. Separately, weak convergence together with implies in norm.
Proof
(Separate by projection, then use sublevels)
The empty-constraint case is vacuous. For nonempty , suppose , let , and put . The projection inequality gives for all . But weak convergence would imply
a contradiction. Thus . This is the precise separating measurement supplied by the projection geometry.
For the norm bound, if test against and use . The scalar left side tends to , proving the lower bound. If , it is immediate.
For the energy statement, each sublevel is closed and convex. Write . If there is nothing to prove. If is finite, choose a subsequence whose values tend to . For every , a tail belongs to and still converges weakly to , so . Let . If , a subsequence eventually belongs to each , giving for every real , impossible for the stated codomain. Thus that case cannot occur along a weakly convergent sequence. Finally,
proves the last assertion when the norms converge.
Proof X-ray
A closed convex set is cut out by fixed measurements
The decisive step separates a point outside by one fixed test vector : every has , while gives . Weak convergence controls exactly such fixed tests, so the limit cannot leave . Convexity enters only through the projection inequality, and completeness only through the existence of . The energy statement needs no new idea: a convex lower semicontinuous energy has closed convex sublevels, and the constraint result is applied to each of them along a subsequence realizing the lower limit.
Common mistake
(Norm closedness and norm continuity are not enough)
Norm closedness alone is not enough for a constraint: the unit sphere in infinite-dimensional contains every coordinate vector but not their weak limit zero. Likewise, norm continuity alone is not enough for an energy: is continuous, yet . Convexity is doing substantive work in the theorem.
A direct existence proof for minimizers
Definition 2.5.6
(Proper functions and bounded sublevels)
A function is proper if it is finite somewhere and never equals . We say its sublevels are bounded if is bounded for every real ; on an unbounded effective domain this is the usual coercivity condition as .
Theorem 2.5.7
(A Hilbert direct method)
Let be proper, convex and norm lower semicontinuous. Assume and every sublevel is bounded. Then attains its infimum. If is strictly convex on its effective domain , the minimizer is unique.
Proof
(A minimizing sequence and its weak limit)
Properness and the lower bound make a finite number. Choose a minimizing sequence with finite values. A tail lies in , so it is bounded. Extract a weakly convergent subsequence . The previous theorem gives , so . If two distinct minimizers existed, strict convexity at their midpoint would give a smaller value, a contradiction.
For a nonempty closed convex constraint , let be convex and relatively norm lower semicontinuous, and extend it by on and outside . The extension is convex and norm lower semicontinuous. To apply the existence theorem, also require a point of with finite energy, , and bounded sets for every real . Closedness and convexity of the constraint alone do not supply these extra conditions.
Concept lensStructural
Three separate reasons the limit is a minimizer
Bounded sublevels supply a weakly convergent subsequence. Closed convex sublevels keep its energy from increasing at the limit. Strict convexity, when available, rules out a second minimizer. Existence, preservation of the constraint and uniqueness therefore come from distinct assumptions. A proof that checks only boundedness has left the central limit-passage question unanswered.
Worked examples
Worked example 2.5.8
(Unit vectors disappear under fixed measurements)
In , let be the coordinate unit vectors. For a fixed , its coordinates tend to zero, so . Thus . However and for , so no subsequence converges in norm. Testing with the moving vector would give one; that calculation does not contradict weak convergence because its test is not fixed.
Worked example 2.5.9
(An unattained least-squares infimum, repaired by regularization)
Let on , and let . The energy is continuous, convex and nonnegative. Its infimum is zero: take with its first coordinates equal to one and the rest zero. But zero cannot be attained, because would force every , which is not square-summable. The minimizing sequence has norm , and bounded sublevels fail.
For define . Its sublevels are bounded, and it is continuous and strictly convex. The direct method gives a unique minimizer. In fact
This vector belongs to , since its squared coordinates are bounded by . Its coordinate equation is . Expanding the energy and canceling the cross terms gives, for every ,
The cross term is by the displayed coordinate equation; its series converge by Cauchy–Schwarz. This identity verifies the minimizer and quantifies stability directly. The added term supplies the missing bound; it does not claim the unregularized equation has acquired an exact solution.
Worked example 2.5.10
(A norm-closed bounded constraint can lose the limit)
On the unit sphere , minimize
This energy is norm continuous, nonnegative and convex on the ambient space: it is the square of the norm of the bounded diagonal map . Its infimum over is zero because . Yet for every , so the infimum is not attained. The sequence is bounded and has weak limit zero, but zero is inadmissible. The sphere is norm closed but not convex or weakly sequentially closed; boundedness alone cannot replace the constraint hypothesis. On the closed unit ball instead, the admissible point zero attains the minimum zero.
Counterexample mode
A bound does not give weak subsequences in every Banach space
In , the unit vectors have norm one. Consider any proposed subsequence . Define a bounded scalar sequence by and zero elsewhere. The functional is bounded on , with . Its values on the proposed subsequence alternate, so that subsequence cannot converge weakly. Because this construction works for every subsequence, there is no weakly convergent subsequence. The Hilbert extraction theorem is not a theorem about all Banach spaces.
Practice
Quick check
(In , let . Prove every fixed coordinate tends to zero but the sequence does not converge weakly. Give one fixed square-summable test vector witnessing failure.)
Solution · Answer
Each coordinate of is eventually zero. But for the fixed vector , we have for every . Any weak limit would have every coordinate zero and hence equal zero, contradicted by this test. The norms also equal , violating the boundedness necessary for weak convergence.
Quick check
(If and is a finite-dimensional subspace, prove in norm. Does the same conclusion hold for every closed infinite-dimensional ?)
Solution · Answer
Take an orthonormal basis of . The squared norm of is , which tends to zero as a finite sum. For any closed infinite-dimensional , choose an orthonormal sequence in , inductively taking a vector outside each preceding finite span and normalizing its residual. Bessel gives for every , so . But has norm one, so no norm convergence to zero occurs. The conclusion depends on the finite number of tests, not simply on closedness.
Quick check
(Let be a bounded linear map between Hilbert spaces. Prove weak convergence implies . If is compact, prove that implies in norm.)
Solution · Answer
For , the functional is bounded and linear on . By Riesz it is an inner-product test, so . This proves weak continuity. If is compact, weak boundedness makes relatively norm compact. If norm convergence to failed, a subsequence would remain at distance at least some . Compactness would give a norm-convergent further subsequence with limit . Norm convergence implies weak convergence, and the already proved weak continuity identifies , a contradiction. Thus the whole image sequence converges in norm.
Quick check
(For the least-squares example, show that the unregularized energy is weakly sequentially lower semicontinuous despite having no minimizer. Identify precisely which hypothesis of the direct method is missing, and prove as .)
Solution · Answer
The unregularized energy is finite everywhere, convex and norm continuous because is bounded and the squared norm is continuous and convex. Thus it is proper, bounded below by zero and weakly sequentially lower semicontinuous. The missing existence hypothesis is bounded sublevels; strict convexity is not needed for existence. The truncations have unbounded norms while their energies tend to zero, so one fixed positive sublevel contains an unbounded tail. For any fixed positive integer , each of the first coordinates of tends to one as , giving . Since is arbitrary, the norms tend to infinity. This does not contradict the boundedness obtained for each fixed positive regularization parameter.
Quick check
(Suppose attains a finite minimum at and, for some , satisfies for and finite-energy . Prove for finite-energy . Deduce norm convergence of every sequence whose energies tend to the minimum. Which constant works for above?)
Solution · Answer
Use , . Minimality bounds the left side below by . Rearranging and dividing by gives . Let . Thus a sequence of finite energies tending to the minimum must converge in norm to . The same bound makes the minimizer unique, since a second minimizer would have zero distance from . For , expansion of squared norms gives the convexity gap , so works. The distance bound then agrees with the explicit energy difference in the example.
Exact boundaries
The compactness result proved here is weak subsequence extraction for bounded Hilbert sequences. We have not proved general weak topological compactness, the corresponding theory in arbitrary reflexive Banach spaces, weak-star compactness or a spectral theorem. The minimization result needs the stated convexity, lower-semicontinuity and bounded-sublevel hypotheses; it does not guarantee convergence of a numerical algorithm or existence for nonconvex constraints. These limitations are visible in the counterexamples rather than concealed in a compactness slogan.