Sometimes two vectors should count as the same because their difference is invisible to a problem. Solutions of a linear equation, for example, differ by vectors in the kernel. A quotient makes that identification precise. To do analysis on the result, we need a distance between equivalence classes, and we must check that Cauchy sequences still have limits. The same construction also reveals which continuous linear measurements ignore the identified directions.
Throughout, are real or complex normed spaces and is a linear subspace of . The notation means the continuous dual. Quotients and their functionals below use the same scalar field as .
The norm of a coset
The quotient vector space consists of cosets , with precisely when .
Definition 3.2.1
(Quotient norm and quotient map)
For put
The quotient map is , .
The infimum need not be attained. The expression is independent of the representative: replacing by simply reindexes the infimum. It is homogeneous, and its triangle inequality follows by choosing near-minimizers and using , then letting the approximation errors tend to zero.
Proposition 3.2.2
(When the quotient expression is a norm)
The expression is a norm on exactly when is closed.
Proof
(Zero distance means membership in the closure)
It is zero exactly when lies in . If is closed, that means , so only the zero coset has norm zero. Conversely, if is not closed, choose . Its nonzero coset has zero value, so the expression is merely a seminorm.
Counterexample mode
A dense subspace gives only a seminorm
Claim: the distance to any subspace defines a norm on the quotient. In , the finite-support sequences form a proper subspace: is in but not in , so is a nonzero coset. Truncating, let . Then
so . The repaired statement assumes closed.
The quotient map is linear and satisfies , by taking . It is onto by construction.
Common mistake
(The quotient norm is not the norm of a representative)
A quotient can collapse a large subspace to zero without making every direction disappear. The relevant number is the distance from a representative to , not the original norm of whichever representative was written down; different representatives of one coset generally have different norms, and the quotient norm is the infimum over all of them.
Worked example 3.2.3
(One subspace, two quotient norms)
Let , so the coset of consists of the vectors , .
With the norm , we have , with equality exactly at the midpoint . Hence , attained by a unique representative.
With the norm , we have , with equality for every between and . Hence , and when a whole segment of representatives attains it.
The edge case gives the zero coset in both norms. The quotient norm thus depends on the norm of , not only on , and its minimizers, when they exist, need not be unique.
Completeness of a Banach quotient
Theorem 3.2.4
(Quotients of Banach spaces are Banach)
If is Banach and is closed, then with the quotient norm is Banach.
Proof
(Lift a fast Cauchy subsequence)
Let be Cauchy in . Choose an increasing subsequence with . For each difference choose a representative with and . This is possible by the definition of the infimum; it does not require a nearest representative.
Choose with and put . The series of norms is summable, so is Cauchy and converges to some . Telescoping gives , while is continuous, so . A Cauchy sequence with a convergent subsequence has the same limit; thus .
Proof X-ray
Why near-minimizers suffice
Each hypothesis has one job. Closedness of is used only so that is a normed space in which a limit is unique; the construction itself never mentions it. Completeness of is used once, to make the lifted sequence converge.
The key move is the strict inequality . The quotient norm is an infimum, and an infimum strictly below guarantees some representative with norm below , even if no representative attains the quotient norm. Without passing to a fast subsequence, representatives of the differences would only have norms close to , whose sum may diverge, and the lift need not be Cauchy. The argument closes with two general facts: is continuous, and a Cauchy sequence with a convergent subsequence converges to the same limit.
If is not closed, the algebraic quotient still exists, but the distance expression cannot make it a normed space. Replacing by identifies additional vectors and restores a norm; it changes the equivalence relation.
Factoring an operator through its kernel
Let be bounded and linear. Its kernel is closed, being the preimage of under a continuous map.
Proposition 3.2.5
(Factorization through the kernel)
The map
is a well-defined bounded linear bijection with and . If are Banach and is closed in , then has a bounded inverse.
Proof
(Infimum over the kernel, then the bounded inverse theorem)
If , then , so is well defined; it is linear, onto , and injective because means is the zero coset. For every , , so taking the infimum yields . The reverse inequality follows from and : .
If is Banach, so is by the previous theorem. If is closed in the Banach space , it is Banach in the inherited norm, and the bounded inverse theorem applies to the bounded bijection .
When is onto, the range is all of and the hypothesis is automatic. Under these hypotheses the factorization is a topological isomorphism, not necessarily an isometry.
Worked example 3.2.6
(A factorization that does not preserve distances)
Let , , with Euclidean norms. The kernel is . The quotient norm of is , attained at , whereas its image has norm . Hence , and has norm . A bounded bijection induced by a quotient need not preserve distances.
The closed-range hypothesis cannot be dropped.
Counterexample mode
An injective operator with unbounded inverse on its range
Claim: a bounded injective operator between Banach spaces has a bounded inverse on its range. On , let . It is bounded with , and its kernel is zero. The inverse on the range sends to , so it multiplies the norm of by and is unbounded. Consistently, the range is not closed: finite-support sequences lie in the range and approach , whose only candidate preimage is not in . With its inherited norm the range is incomplete, so the bounded inverse theorem cannot be invoked.
Duals and annihilators
Definition 3.2.7
(Annihilator)
The annihilator of is
Common mistake
(An annihilator lives in the dual)
The notation here denotes a subspace of , not an inner-product orthogonal complement in . In a general normed space there is no inner product, and even in the two objects lie in different spaces; they are identified only after the dual has been identified with .
Theorem 3.2.8
(Dual of a quotient)
For closed , the map is an isometric isomorphism from onto .
Proof
(Pull back and push down)
Given , its pullback belongs to and , since . Conversely, if , define . It is well-defined and linear. Since for every ,
so is bounded with . The constructions are inverse, and the two inequalities give equality of norms.
For closed , the real or complex normed-space Hahn–Banach theorem yields the useful distance formula
The upper bound follows from . For a nonzero coset, Hahn–Banach supplies a norm-one functional on attaining its norm at that coset; pull it back using the isometry above. For the zero coset both sides vanish. Thus the quotient norm can be tested by all bounded functionals that vanish on ; the choice of a single attaining functional may depend on .
An explicit quotient and its measurements
Worked example 3.2.9
(Evaluation at zero identifies a quotient of )
Take of real- or complex-valued continuous functions, with the supremum norm and . Evaluation at zero is continuous, so is closed. For every ,
The lower bound follows because for . For equality choose , subtracting the constant function. Hence evaluation identifies the quotient isometrically with the scalar field. Its dual is one-dimensional: every bounded functional vanishing on is with , because and . This explicit representative exists here, but the quotient construction and completeness proof did not require one in general.
Practice with solutions
Quick check
(Show that if is closed, . When is injective?)
Use , and compute the kernel of .
Solution · Answer
The first equality is the definition; choose for the inequality. The kernel of is , so it is injective exactly when .
Quick check
(For on Euclidean , find its kernel, induced map and the norm of the induced inverse.)
Compute the quotient norm first.
Solution · Answer
The kernel is ; the induced map sends the coset of to . The quotient norm is , so the induced inverse sends to a coset of norm and has norm .
Quick check
(Describe every continuous functional on that vanishes on the functions satisfying .)
Split into a function vanishing at and a constant.
Solution · Answer
Each is a scalar multiple of evaluation at zero: if vanishes on , then . Conversely every is continuous and vanishes on .
Quick check
(Why does with the distance expression above fail to be a normed space?)
Recall that is dense in .
Solution · Answer
The dense proper subspace gives zero distance for every coset; in particular the nonzero coset of has seminorm zero.