A normed space supplies a way to measure errors, but not every sequence with vanishing tail error has a limit inside the space. Series make that distinction useful: if the total size of a sequence of corrections is finite, does the corrected answer exist in the same space? The answer characterizes completeness and explains why infinite sums of bounded operators can be used to construct inverses.
Completeness and summable corrections
Let be a real or complex normed space.
Definition 1.3.1
(Convergent and absolutely convergent series)
A series in converges in when its partial sums converge in norm to an element of . It is absolutely convergent when .
Absolute convergence is a statement about real numbers only; convergence asks for a limit inside .
Theorem 1.3.2
(Completeness through absolute series)
is complete if and only if every absolutely convergent series in converges in .
Proof
(Cauchy partial sums one way, a fast subsequence the other)
If is complete and , then for the triangle inequality gives
The partial sums are Cauchy and therefore converge in .
Conversely, let be Cauchy. Select indices such that for every . This is possible by the Cauchy property. The series of successive selected differences converges by assumption, since its norms have a geometric bound. Its partial sums telescope:
The selected subsequence thus converges to some . The entire Cauchy sequence converges to : choose one sufficiently late selected term close to , then use the Cauchy condition to bring every sufficiently late close to it. This proves completeness.
Proof X-ray
Why a subsequence is enough
The forward direction uses completeness exactly once, after the triangle inequality has turned a tail of real numbers into a Cauchy estimate for partial sums. The converse has to manufacture an absolutely convergent series from a Cauchy sequence, and this is the non-obvious move: an arbitrary Cauchy sequence need not have summable consecutive differences (for instance, the convergent partial sums of the alternating harmonic series have consecutive differences of size ), so the proof passes to a subsequence with differences below .
The hypothesis is then applied to that one series, and telescoping converts its sum into a limit of the subsequence. The argument closes with a general fact about Cauchy sequences: a convergent subsequence forces the whole sequence to the same limit. No limit in a larger completion is assumed at any stage.
The hypothesis concerns every absolutely convergent series, and a single failure is enough to destroy completeness.
Counterexample mode
Summable sizes without a limit
Claim: an absolutely convergent series converges in any normed space. In with the norm, put . The sum of the norms is , but the partial sums have no limit in : norm convergence implies coordinatewise convergence, so any limit would have coordinate equal to for every and hence would not have finite support. The same series converges in the complete space . The repaired statement assumes complete.
The space of bounded operators
Let be normed and let be Banach, over the same scalar field.
Definition 1.3.3
(Operator norm)
is the space of bounded linear maps , with operator norm . We write .
Theorem 1.3.4
(Operator spaces with complete target are complete)
If is Banach, then is Banach, even if is not complete.
Proof
(Pointwise limit, then a uniform estimate)
Take an operator-norm Cauchy sequence . For each , the sequence is Cauchy in , since , so define . Passing limits through finite linear combinations makes linear. Operator-norm Cauchy sequences are bounded, say ; taking limits gives . Thus is bounded. Given , choose so that for . Letting pointwise yields for every , hence .
Common mistake
(A pointwise limit is not yet a norm limit)
Pointwise existence of was not the conclusion; the uniform estimate in the last step is what proves norm convergence. Without the operator-norm Cauchy hypothesis the two notions differ. On , let keep the first coordinates and set the rest to zero. For every fixed , , so pointwise; yet gives for every .
Inverting a small perturbation
Theorem 1.3.5
(Neumann series)
Let be Banach and satisfy . Then is invertible in , , and
Proof
(Sum the geometric series in operator norm)
Since and is Banach, the operator series converges absolutely in operator norm. Multiplying its finite partial sum gives
The remainder satisfies . Passing to the limit on both sides proves , so . The two bounds follow by summing over and over .
The same series controls perturbations of any invertible operator.
Corollary 1.3.6
(Inverse perturbation bound)
Suppose is a bounded bijection between Banach spaces with bounded inverse, and is bounded. If , then is invertible and
Proof
(Factor out )
Write with and . By the Neumann series, is invertible, so . Subtracting the identity from the series gives , and hence .
Common mistake
(Reading as necessary)
The bound is a sufficient criterion, not a necessary one: on , has norm , yet is invertible. The geometric series simply does not converge in this case, so it cannot be used to produce the inverse.
Worked example 1.3.7
(Inverting for the right shift)
On , let , an isometry with , and let . Then , so with norm at most . Applied to ,
Check: , so . Also , within the bound .
Worked example 1.3.8
(A perturbation of a diagonal matrix)
On Euclidean , let and . Then , whose singular values are and , so . With the corollary predicts
Directly, , so and . This symmetric matrix has eigenvalues , so its norm is , comfortably below the bound.
Worked example 1.3.9
(An edge case: a nilpotent with )
On Euclidean , let . Here , so the theorem does not apply. Nevertheless , the series stops after two terms, and . Hence exactly. The hypothesis guarantees convergence of the series; other mechanisms, such as nilpotency, can make it terminate.
Practice and solutions
Quick check
(Let be Banach and . Bound the norm of the tail after terms of .)
Compare with a geometric series.
Solution · Answer
The series is absolutely convergent, so the sum exists, and the tail norm is at most .
Quick check
(On , let . Find and the exact error after the zeroth partial sum .)
Here is a scalar multiple of .
Solution · Answer
, so the inverse is and the error is . The Neumann estimate with gives , so here it is exact.
Quick check
(Let , with the norm. Explain why is not complete.)
A linear map on is determined by one value.
Solution · Answer
A bounded map from is determined by , and its operator norm is . This identifies isometrically with incomplete ; the partial sums of give a Cauchy sequence without a limit there. Completeness of the target cannot be dropped.