Motivation
A quotient and a kernel do not determine the middle module of a short exact sequence. Both and can sit between a copy of and a quotient . A useful classification must also remember how the kernel is embedded and how the quotient is identified. First Ext records exactly that extension data.
The computation starts with a free presentation. Lift its generators into the proposed middle module, then measure how its relations land in the kernel. Changing the lifts changes that relation map by a map already defined on the free module. The resulting quotient of Hom groups classifies extensions, and a concrete construction proves this rather than treating Ext as unexplained notation.
Let be a unital ring, possibly noncommutative, and use left modules. Hom sets below are additive abelian groups; no commutative scalar-module structure is silently assumed. We use the free/projective and splitting results. Free covers and choices of lifts use the usual axiom of choice. This chapter concerns module extensions, not arbitrary possibly nonabelian group extensions.
Definitions
Definition 9.2.1
(Equivalence and splitting of extensions)
An extension of by is a short exact sequence . Two extensions are equivalent if there is an isomorphism of their middle modules commuting with both maps while acting as identity on the fixed end modules .
An extension splits if has an -linear section. By the proved splitting theorem, this is equivalent to a retraction of , and identifies the sequence with the standard one having middle module . An abstract isomorphism of middle modules alone is weaker than equivalence of extensions.
Choose a free presentation
Identify with its image in . It need not itself be free or projective. Restriction defines . The cokernel of this map will turn out to be independent of the chosen presentation because it represents the intrinsic extension classes.
Definition 9.2.2
(Injective modules)
A module is injective if every homomorphism extends across every submodule inclusion to a map . Projectivity instead lifts maps from a module across surjections. These are different directions and different positions in an exact sequence.
Theorems
Theorem 9.2.3
(First Ext classifies extensions)
The equivalence classes of extensions of by form an abelian group, denoted , under the Baer addition described below. There is an isomorphism
The zero class consists exactly of split extensions. In particular this quotient, with its extension interpretation, is independent of the chosen free presentation.
Theorem 9.2.4
(Projectivity and injectivity detected by extensions)
A module is projective exactly when for every . A module is injective exactly when for every .
Over , injective modules are exactly the divisible abelian groups: multiplication by every positive integer is onto. This last characterization uses the usual choice principle through Zorn's lemma.
A fixed extension also measures which maps into its quotient can be lifted.
Proposition 9.2.5
(The lifting obstruction of an extension)
Fix an extension and a module . For , let be the class of the pullback extension , where . Then is additive, lifts to exactly when , and the sequence
is exact.
Proof sketch or proof idea
Relations produce an extension, not just an obstruction symbol
Proof
(Classification by relation maps)
Given , put
The denominator is a submodule because it is the image of an -linear map. Define and . The first map is injective: if , then . The second is well-defined and onto. If its value is zero, , and . Thus its kernel is precisely the image of , proving exactness.
Conversely, take an extension . Projectivity of the free module gives a lift of . On , its image is in , so it determines by . The map kills and induces . It is onto because subtracting a suitable lift leaves an element of the kernel. If its value is zero, then and , precisely a defining relation. It is therefore an isomorphism commuting with the end maps.
Changing the lift to for changes to . Conversely, if , the map is an extension equivalence, with the opposite correction giving its inverse. Any equivalence transports a chosen lift, so no additional identifications occur. This proves the claimed bijection with the Hom cokernel. At , the middle module is . The zero-coset statement and the lift-change argument prove that precisely the split extensions represent zero.
Proof X-ray
A relation map depends on lifts; its class does not
Freeness of is used exactly once, to produce a lift of ; nothing about is needed, which is why may be any submodule. Exactness of then gives , so , and injectivity of turns into a well-defined linear . The non-obvious move is the quotient : it glues to along , and the map shows that this glued module is already . The ambiguity is closed by the lift change: two lifts differ by with , so the relation maps differ by , the subgroup the cokernel removes. The zero class is read off the same way: if , the corrected lift vanishes on , so it factors through and is a section.
The group operation is intrinsic
Proof
(Baer sum and addition of classes)
For extensions of the same by the same , form the submodule
Its kernel over is . Quotient by the antidiagonal submodule . The new kernel is identified with by addition of its two coordinates. This is an extension of by , called the Baer sum. Equivalent inputs give equivalent outputs by their induced pair and quotient maps.
Choose lifts from the same free into the two extensions. Their pair is a lift into , and on its kernel coordinates are . After the quotient they become . Thus the Baer sum corresponds exactly to addition of the two Hom-cokernel classes. It follows that the operation is associative and commutative, has the split class as identity and has additive inverses. Because the construction of this operation and equivalence uses only the extensions, not the presentation, the isomorphism is one of abelian groups independently of the free cover.
Vanishing detects both lifting and extension properties
Proof
(Ext vanishing detects projective and injective modules)
If is projective, every surjection splits, so all its extension classes vanish. Conversely, if every vanishes, apply it with to a free cover . That sequence splits; is a direct summand of a free module and hence projective by the earlier characterization.
If is injective, its identity map extends across to a retraction, splitting every extension. Conversely, suppose all extensions by split. For an inclusion and a map , form
The same exactness check as before gives an extension of by ; it did not require that the larger module be free. A splitting supplies a retraction fixing the included copy of . Then extends , because . Thus is injective. This proves the two quantifier statements in the theorem and shows why the variable positions matter.
First Ext measures a specific failed lift
Proof
(The pullback obstruction)
Fix an extension and any module . For a map , form its pullback
Projection to , with kernel , makes this an extension of by . Denote its class by . It is zero exactly when lifts to : a lift gives the section , and every section has this form.
The assignment is additive. To check this, choose one free cover and lift composed with that cover to . Its restriction to the relation module lands in and computes the pullback's class by the preceding classification. For two maps, add their chosen lifts; the relation maps add too, so , independently of those choices. Thus we obtain an exact initial obstruction sequence
The Hom terms use -linear maps. Exactness at the first two nonzero terms follows from the original kernel property, while exactness at is precisely the lift-if-and-only-if-zero statement. The last Ext term has no asserted surjectivity. This makes the earlier failure of Hom surjectivity concrete: a particular map fails to lift because its pullback extension is nonsplit. Vanishing of the entire Ext group is sufficient to remove every such obstruction.
Concept lensStructural
An extension as the obstruction to lifting the identity
Take and in the lifting proposition. The pullback is isomorphic to by , compatibly with both end maps, so . For a general , is by definition the pullback of along . Thus the class of is exactly the obstruction to a section of , and every other lifting obstruction is obtained from it by pulling back.
For abelian groups, divisibility is exactly enough
Proof
(Injective abelian groups are the divisible ones)
If is injective, extend the map sending to an arbitrary across . The image of one satisfies . Thus is divisible.
For the converse, let be divisible and start with , where are abelian groups. Order all extensions of this map to intermediate subgroups by further extension. Unions along chains give upper bounds, so Zorn provides a maximal extension . If , consider the ideal , . If , choose any . If , divisibility supplies with .
Define on . To check well-definedness, equality of two such expressions gives . If it forces . Otherwise , so , exactly the needed equality. This extends to a strictly larger subgroup, a contradiction. Therefore and the original map extends. The use of choice and the scope to abelian groups are explicit.
Worked examples
Worked example 9.2.6
(Cyclic quotients give a simple Ext computation)
The free presentation identifies each Hom group with , using the image of in the first and of one in the second. Restriction is multiplication by , so . Taking and gives two classes. The zero class has middle group ; the nonzero class has generators with , , hence middle group . These are module extensions, so all middle groups are abelian.
Worked example 9.2.7
(Distinct extension classes can have the same middle group)
For , , the extension group is . A representative has middle group generated by with , where is the prescribed kernel generator. Integer Smith form gives , with an order-one summand omitted. In particular and both give an abstract infinite cyclic middle group, but their classes modulo six are different. No isomorphism fixing both end identifications can equate the two extensions. Classifying middle modules without the maps would lose this distinction.
Worked example 9.2.8
(Divisibility and a nonsplit sequence)
Both and are divisible: divide a rational representative by the specified positive integer. They are therefore injective abelian groups. The group is not divisible. The extension does not split. Indeed, every homomorphism is zero: the image of any rational is divisible by every positive integer, since the rational can itself be divided by each integer. A retraction onto is impossible. Thus has a nonzero class, whereas injectivity gives for every .
Common mistakes
Common mistake
(Swapping the two variables of Ext)
The kernel module is the second variable in Ext, and the quotient is the first. Projectivity kills all extensions with that first variable, injectivity all extensions with that second variable. Nonzero Ext means some nonsplit extensions exist; the split class is always present as well.
The quantifier in the projectivity criterion is just as important as the variable position.
Counterexample mode
Vanishing against one module does not give projectivity
Claim: if for some nonzero , then is projective. Take and . Since is divisible, it is injective, so . But is not projective: a projective module is a direct summand of a free abelian group , and a nonzero summand of has some coordinate projection that is nonzero on it, giving a nonzero homomorphism . Every such homomorphism is zero, because its value at equals times its value at for every positive integer . The theorem needs vanishing for every ; in the proof only , the relation module of a free cover, is used.
The second typical error concerns what counts as the same extension.
Common mistake
(An isomorphic middle module is not an equivalent extension)
An extension equivalence preserves its two end maps, a stronger demand than middle-module isomorphism; the classes and in the second worked example have isomorphic middle groups and are still different. The free presentation's relation module need not be projective.
This construction computes first Ext and its extension meaning; higher Ext, projective-resolution invariance and the full long exact sequence are not being assumed or claimed.
Summary
Lifting free generators turns an extension into a relation map modulo changes of lift. A quotient construction recovers the extension, and the intrinsic Baer sum matches addition of these classes. Vanishing detects projective and injective modules in their different variables. Over the integers, a proved divisibility criterion makes this concrete.
Exercises
Quick check
(Compute for positive . When does every extension of by in abelian groups split?)
Use the cyclic-quotient formula .
Solution · Answer
The formula is , since multiplication by has image of size . Thus every abelian-group extension splits exactly when the gcd is one. This says nothing about extensions with nonabelian middle groups, which are outside the module category.
Quick check
(List the extension classes of by , and determine their middle groups. Explain why there are more extension classes than middle-group isomorphism types.)
A class gives generators with and .
Solution · Answer
The Ext group is , with representatives . Relations are , . At zero the middle group is . For either nonzero , the relation expresses as a multiple of , and has order nine, giving . The two nonzero classes remain inequivalent when the fixed kernel and quotient identifications must be preserved, even though their middle groups are isomorphic.
Quick check
(Let now be commutative and let be a nonzero non-zero-divisor. Prove .)
Find a free presentation of of length one.
Solution · Answer
The sequence is exact because multiplication by is injective. Identifying with , the restriction map is multiplication by . Its cokernel is , proving the formula. The non-zero-divisor condition is what makes this length-one free presentation valid.
Quick check
(Prove is injective but not projective as a -module. Contrast it with the projective but noninjective module .)
Combine divisibility with the fact that every homomorphism is zero.
Solution · Answer
Divisibility proves injectivity. If were projective, a free cover would have a section. Composing such a section with every integer coordinate projection of the free abelian group gives a map , necessarily zero by the example. All coordinates of the section are zero, so the section itself is zero, impossible for a right inverse onto nonzero . Finally is free and hence projective, but is not divisible by two, hence is not injective.