Among the constructions a functor can perform, one pairing recurs so often across the curriculum
that it deserves a name and a theory of its own: the situation in which two functors run in
opposite directions and stand in a precise reciprocal relation, each the best possible
approximation to an inverse of the other. The free vector space on a set and the underlying set
of a vector space; the abelianization of a group and the inclusion of abelian groups among all
groups; the product and the exponential of sets. In each case the two functors are not inverse
— they change the objects too drastically for that — yet a single equation binds them, and that
equation turns out to govern an astonishing range of mathematics. The relation is called
adjunction.
Consider two functors in opposite directions,
\(F : \mathscr{A} \to \mathscr{B}\) and \(G : \mathscr{B} \to \mathscr{A}\). Roughly, \(F\) is
adjoint to \(G\) when, for every object \(A\) of \(\mathscr{A}\) and every object \(B\) of
\(\mathscr{B}\), the morphisms \(F(A) \to B\) in \(\mathscr{B}\) are the same thing as
the morphisms \(A \to G(B)\) in \(\mathscr{A}\) — not merely equal in number, but matched by a
correspondence that respects every morphism in sight. Making "the same thing" precise is the
whole content of the definition.
Definition: Adjunction
Let \(F : \mathscr{A} \to \mathscr{B}\) and \(G : \mathscr{B} \to \mathscr{A}\) be functors.
We say \(F\) is left adjoint to \(G\), and \(G\) is right
adjoint to \(F\), written \(F \dashv G\), if there is a bijection
\[
\mathscr{B}\bigl(F(A), B\bigr) \;\cong\; \mathscr{A}\bigl(A, G(B)\bigr)
\]
for each object \(A\) of \(\mathscr{A}\) and each object \(B\) of \(\mathscr{B}\), and this
bijection is natural in \(A\) and \(B\) in the sense made precise below. A
choice of such a natural bijection is an adjunction between \(F\) and \(G\).
The bijection has a standard notation that makes its symmetry visible. Given a morphism
\(g : F(A) \to B\), its image under the bijection is written \(\bar{g} : A \to G(B)\) and called
the transpose of \(g\); given \(f : A \to G(B)\), its inverse image is
\(\bar{f} : F(A) \to B\). The two operations are mutually inverse, so transposing twice returns
the original morphism: \(\bar{\bar{g}} = g\) and \(\bar{\bar{f}} = f\). One passes freely between
a morphism out of \(F(A)\) and a morphism into \(G(B)\), and the bar is the device for doing so.
Naturality is the requirement that this passage be compatible with composition on both sides.
Stated in full, it has two parts: for all morphisms \(g : F(A) \to B\) and \(q : B \to B'\) in
\(\mathscr{B}\),
\[
\overline{q \circ g} = G(q) \circ \bar{g},
\]
and for all morphisms \(p : A' \to A\) in \(\mathscr{A}\) and \(f : A \to G(B)\),
\[
\overline{f \circ p} = \bar{f} \circ F(p).
\]
The first says that postcomposing in \(\mathscr{B}\) and then transposing agrees with transposing
and then applying \(G\); the second says the analogous thing for precomposition in \(\mathscr{A}\)
through \(F\). Together they pin down the correspondence so tightly that no arbitrary choices
remain — which is exactly what one means, informally, by calling the matching between the two
sets of morphisms natural.
A single consequence shows the naturality axiom at work. Suppose we are given a chain of
morphisms in \(\mathscr{A}\) ending at some object, a single morphism out of its \(F\)-image, and
a chain in \(\mathscr{B}\) leading away:
\[
A_0 \to \cdots \to A_n, \qquad F(A_n) \to B_0, \qquad B_0 \to \cdots \to B_m.
\]
From this data there is exactly one morphism \(A_0 \to G(B_m)\) that the adjunction produces:
compose the \(\mathscr{B}\)-chain onto the middle morphism, transpose the result across the
bijection, and precompose with the \(\mathscr{A}\)-chain. The two naturality equations are
precisely the guarantee that the order of these operations does not matter — transposing first
and then composing (which sends the \(\mathscr{B}\)-chain through \(G\), since a transpose lands
in \(G(B_0)\) and only \(G(q)\) can be composed onto it), or composing first and then transposing
(which sends the \(\mathscr{A}\)-chain through \(F\)), yield the same arrow \(A_0 \to G(B_m)\).
Naturality is what makes "the morphism obtained from this data" a well-defined phrase rather than
an ambiguous recipe.
Naturality as a coming attraction
The naturality requirement is the categorical meaning of "natural" we met for transformations
between functors: a construction defined without arbitrary choices. This is no analogy. The
full force of that earlier notion surfaces again once we recognize the adjunction bijection
itself as a
natural isomorphism
between two functors built from the hom-sets — at which point the two naturality equations
above become a single statement that one natural transformation is invertible.
Free Constructions as Left Adjoints
Once one knows to look for them, adjoint functors turn up everywhere. The reliable signal is a
pair of functors running in opposite directions between two categories: whenever such a pair
presents itself, there is an excellent chance that one is left adjoint to the other. The
phenomenon is common enough to serve as a working heuristic. Told that some construction turns
every object of one kind into an object of another, and that a construction in the reverse
direction also exists, one should suspect at once that the two are adjoint — and the suspicion is
usually correct. A mathematician describing, say, a way to turn any Lie algebra into an
associative algebra and a way back again is, whether or not the word is used, describing an
adjunction; recognizing it spares one the labor of learning either construction in detail, since
each determines the other.
The most transparent family pairs a forgetful functor with a free
construction. A forgetful functor discards structure — sending a group, a vector space,
or a topological space to its underlying set, and a structure-preserving map to the underlying
function. Its left adjoint, when it exists, builds the most economical structured object on a
given set: the one that imposes no relations beyond those forced by the axioms. The adjunction
equation expresses precisely the sense in which the free object is universal among all maps into
structured objects.
The Free Vector Space
Let \(k\) be a field. The forgetful functor
\(U : \mathbf{Vect}_k \to \mathbf{Set}\) sends a
vector space
to its underlying set of vectors and a
linear map
to its underlying function. Its left adjoint \(F : \mathbf{Set} \to \mathbf{Vect}_k\) sends a set
\(S\) to the vector space \(F(S)\) with basis \(S\) — the space of formal finite linear
combinations \(\sum_{s \in S} \lambda_s\, s\) with coefficients in \(k\). We claim
\(F \dashv U\), and unlike the slogan this is a claim one can check by hand.
Proposition: The Free-Forgetful Adjunction on Vector Spaces
For each set \(S\) and vector space \(V\) there is a bijection
\[
\mathbf{Vect}_k\bigl(F(S), V\bigr) \;\cong\; \mathbf{Set}\bigl(S, U(V)\bigr),
\]
natural in \(S\) and \(V\); that is, \(F \dashv U\).
Proof:
We exhibit the two transpose operations and verify they are mutually inverse. Given a linear
map \(g : F(S) \to V\), define a function \(\bar{g} : S \to U(V)\) by restricting \(g\) to
the basis: \(\bar{g}(s) = g(s)\) for each \(s \in S\). This is a well-defined map of sets.
Conversely, given a function \(f : S \to U(V)\), define a linear map
\(\bar{f} : F(S) \to V\) by extending \(f\) linearly off the basis:
\[
\bar{f}\Bigl(\textstyle\sum_{s \in S} \lambda_s\, s\Bigr) = \sum_{s \in S} \lambda_s\, f(s).
\]
Because \(S\) is a basis of \(F(S)\), every element of \(F(S)\) is a unique finite linear
combination of basis vectors, so \(\bar{f}\) is well defined and linear, and it is the unique
linear map agreeing with \(f\) on \(S\).
These two operations are mutually inverse. Starting from a linear map \(g\), the map
\(\bar{\bar{g}}\) extends \(\bar{g} = g|_S\) linearly; but \(g\) is itself linear and agrees
with \(\bar{g}\) on the basis, so by uniqueness of linear extension \(\bar{\bar{g}} = g\).
Starting from a function \(f : S \to U(V)\), the function \(\bar{\bar{f}}\) restricts
\(\bar{f}\) to the basis, and \(\bar{f}(s) = f(s)\) there, so \(\bar{\bar{f}} = f\). Hence the
two operations are inverse bijections between \(\mathbf{Vect}_k(F(S), V)\) and
\(\mathbf{Set}(S, U(V))\) for each \(S\) and \(V\).
Naturality is the statement that these bijections commute with precomposition by functions
\(S' \to S\) and postcomposition by linear maps \(V \to V'\). For postcomposition by a linear
map \(q : V \to V'\): transposing \(q \circ g\) restricts it to the basis, giving
\(s \mapsto q(g(s)) = q(\bar{g}(s))\), which is \(U(q) \circ \bar{g}\) since \(U(q)\) is the
underlying function of \(q\). The precomposition identity is verified the same way, using
that \(F\) sends a function \(p : S' \to S\) to the linear map extending it on bases. Both
naturality equations therefore hold, completing the proof.
The proof is worth dwelling on because it exposes the general mechanism. The bijection rests on a
single fact about \(F(S)\): a linear map out of it is determined by, and may be freely prescribed
by, its values on \(S\). This is the universal property of the free object, and it is
what every free construction has in common. We isolate that property in the next section, where
it acquires its categorical name.
Free Groups and the Pattern in General
The same adjunction holds with
groups
in place of vector spaces: the forgetful functor \(\mathbf{Grp} \to \mathbf{Set}\) has a left
adjoint sending a set to the free group on it, the group of reduced words in the elements of the
set and their formal inverses. The free group is harder to construct explicitly than the free
vector space — there is no basis to extend along so transparently — but the adjunction
characterizes it completely all the same: a homomorphism out of the free group on \(S\) is the
same thing as a function from \(S\) into the underlying set of the target. Forgetful functors
between categories of algebraic structures almost always admit such left adjoints, a fact that
can be established once and for all rather than rediscovered case by case.
A left adjoint, when it exists, is unique up to isomorphism — we prove this in the next section
— so the phrase "the free group on \(S\)" is unambiguous even before any construction is given.
Knowing that a functor is left adjoint to the forgetful functor characterizes it completely,
which is often all one needs.
Monoids, and a Forgetful Functor with Two Adjoints
Groups are not the only structure built from an associative binary operation. Dropping the
requirement of inverses, while keeping associativity and an identity element, gives the notion of
a monoid — a structure we will need repeatedly, since a monoid is the simplest interesting
example of a category with a single object.
Definition: Monoid
A monoid is a set \(M\) equipped with an associative binary operation
\(M \times M \to M\), written multiplicatively, and an identity element
\(e \in M\) satisfying \(e\,m = m = m\,e\) for all \(m \in M\). A monoid
homomorphism is a function preserving the operation and the identity. Monoids and
their homomorphisms form a category \(\mathbf{Mon}\).
Every group is a monoid — one in which each element happens to have an inverse — so there is an
inclusion functor \(U : \mathbf{Grp} \to \mathbf{Mon}\) viewing a group as a monoid and a
homomorphism as itself. Forgetting that inverses exist loses no data; it simply declines to
record a property. The morphisms are genuinely unchanged, not merely relabeled: a monoid
homomorphism between two groups — a map preserving the product and the identity — automatically
preserves inverses, since \(\phi(g)\,\phi(g^{-1}) = \phi(g\,g^{-1}) = \phi(e) = e\) forces
\(\phi(g^{-1}) = \phi(g)^{-1}\), so it is already a group homomorphism. Thus
\(\mathbf{Grp}(G, H) = \mathbf{Mon}(UG, UH)\), exhibiting \(\mathbf{Grp}\) as a
full subcategory
of \(\mathbf{Mon}\). What makes this inclusion remarkable is that it has adjoints on
both sides:
\[
F \dashv U \dashv R,
\]
a configuration we have not seen before.
The left adjoint \(F : \mathbf{Mon} \to \mathbf{Grp}\) freely adjoins inverses: it sends a monoid
\(M\) to the group obtained by formally throwing in an inverse for every element, subject only to
the relations already holding in \(M\). When \(M\) is the additive monoid of natural numbers, for
instance, \(F(M)\) is the group of integers — the natural numbers with subtraction made possible.
This is the group completion of \(M\), and adjointness \(F \dashv U\) says exactly that a monoid
homomorphism \(M \to U(H)\) into (the monoid underlying) a group \(H\) is the same thing as a
group homomorphism \(F(M) \to H\): maps out of the completion correspond to maps out of the
original monoid.
The right adjoint \(R : \mathbf{Mon} \to \mathbf{Grp}\) goes the other way, extracting rather than
adjoining. It sends a monoid \(M\) to its group of units — the submonoid of
elements that do possess an inverse, which forms a group. Adjointness \(U \dashv R\) says that a
group homomorphism \(H \to R(M)\) is the same thing as a monoid homomorphism \(U(H) \to M\); since
every element of a group is invertible, any monoid homomorphism out of a group must land among the
invertible elements of \(M\), which is to say in \(R(M)\). The two adjoints capture the two
canonical ways to mediate between monoids and groups: complete a monoid into a group, or carve out
the largest group sitting inside it.
A two-sided rarity
Forgetful functors routinely have left adjoints — the free constructions of this section are
exactly those — but they only rarely have right adjoints as well. The inclusion
\(\mathbf{Grp} \hookrightarrow \mathbf{Mon}\) is the unusual case where both exist, which in
the standard terminology makes \(\mathbf{Grp}\) simultaneously a reflective
and a coreflective subcategory of \(\mathbf{Mon}\): a subcategory whose
inclusion admits a left adjoint is reflective, one whose inclusion admits a right adjoint is
coreflective, and here both hold at once. We will meet this two-sided pattern
\(F \dashv U \dashv R\) again in the next section.
Universal Properties and Initial Objects
The free constructions above all rested on a property of the form: a map out of the constructed
object is freely determined by less data. Such characterizations — "the unique object through
which every map of a certain kind factors" — pervade mathematics under the name
universal property, and the curriculum has met many already, each proved in its
own setting by its own hands. The categorical viewpoint reveals them as instances of one
structure, definable purely in terms of arrows.
Abelianization as a Left Adjoint
A free construction need not start from a bare set. The inclusion functor
\(U : \mathbf{Ab} \to \mathbf{Grp}\), viewing an abelian group as a group, has a left adjoint
\(F : \mathbf{Grp} \to \mathbf{Ab}\) that forces a group to become abelian in the most economical
way. It sends a group \(G\) to its abelianization \(G^{\mathrm{ab}}\), the
quotient of \(G\) by the smallest
normal subgroup
containing all commutators
\(xyx^{-1}y^{-1}\). Together with the quotient homomorphism \(\eta : G \to G^{\mathrm{ab}}\), this
construction has a universal property that is the very shape of the adjunction.
The universal property of abelianization. Any homomorphism \(\varphi\) from \(G\) to an
abelian group \(A\) (purple) factors as \(\varphi = \bar{\varphi} \circ \eta\) through the
quotient \(\eta : G \to G^{\mathrm{ab}}\) (blue), and the factoring map \(\bar{\varphi}\)
(dashed) is unique.
Every
homomorphism
\(\varphi : G \to A\) into an abelian group \(A\) factors uniquely as
\(\varphi = \bar{\varphi} \circ \eta\) through a homomorphism
\(\bar{\varphi} : G^{\mathrm{ab}} \to A\). The reason is that \(A\), being abelian, sends every
commutator to the identity, so the
kernel
of \(\varphi\) contains the commutator subgroup;
\(\varphi\) therefore descends to the quotient, and the descended map is forced. This factoring
is precisely the adjunction bijection
\[
\mathbf{Ab}\bigl(G^{\mathrm{ab}}, A\bigr) \;\cong\; \mathbf{Grp}\bigl(G, U(A)\bigr),
\]
sending \(\bar{\varphi}\) on the left to \(U(\bar{\varphi}) \circ \eta\) on the right. The functor
\(U\) appears because \(\bar{\varphi}\) is a morphism of \(\mathbf{Ab}\) while \(\eta\) is a
morphism of \(\mathbf{Grp}\); to compose them one first views \(\bar{\varphi}\) as a group
homomorphism, which is what \(U\) does. Since \(U\) is the inclusion and changes nothing, it is
routinely suppressed, and one writes simply \(\bar{\varphi} \circ \eta\). That
\(F = (-)^{\mathrm{ab}}\) is left adjoint to \(U\) is exactly the statement that this factoring
exists and is unique for every \(G\) and \(A\). The quotient map \(\eta\), which receives the
universal homomorphism, is the structural heart of the construction — and it is an instance of a
phenomenon we now name.
Definition: Initial and Terminal Object
Let \(\mathscr{C}\) be a category. An object \(I\) of \(\mathscr{C}\) is
initial if for every object \(A\) there is exactly one morphism
\(I \to A\). An object \(T\) is terminal if for every object \(A\) there is
exactly one morphism \(A \to T\).
The empty set is initial in \(\mathbf{Set}\) and a one-element set is terminal; the trivial group
is both initial and terminal in \(\mathbf{Grp}\). The defining clause — exactly one morphism —
is a universal property in its barest form, and it is rigid enough to determine the object
completely.
Lemma: Uniqueness of Initial Objects
Let \(I\) and \(I'\) be initial objects of a category. Then there is a unique isomorphism
\(I \to I'\); in particular \(I \cong I'\). The same holds for terminal objects.
Proof:
Since \(I\) is initial there is a unique morphism \(f : I \to I'\), and since \(I'\) is
initial there is a unique morphism \(f' : I' \to I\). The composite \(f' \circ f\) is a
morphism \(I \to I\); but \(I\) is initial, so there is exactly one such morphism, and the
identity \(1_I\) is one. Hence \(f' \circ f = 1_I\). Symmetrically \(f \circ f' = 1_{I'}\),
so \(f\) is an
isomorphism.
Its uniqueness as a morphism \(I \to I'\) is built in, since initiality of \(I\) allows only
one morphism to \(I'\) at all. The statement for terminal objects follows by reversing every
arrow.
This is the categorical source of a phrase used freely throughout the curriculum: "the" object
with a given universal property. Whenever an object is pinned down by a universal property, that
property exhibits it as initial or terminal in a suitable category of "objects equipped with the
relevant data," and the lemma then guarantees uniqueness up to a single canonical isomorphism.
The same argument settles the uniqueness of left adjoints promised earlier: a left adjoint to
\(G\) assigns to each object \(A\) an initial object in an associated category — the category
whose objects are the morphisms \(A \to G(B)\) out of \(A\) into a \(G\)-image — so any two left
adjoints are canonically isomorphic. (The value \(F(A)\) is the source of that initial morphism;
its initiality is precisely the universal property defining the free object.)
The universal properties already proved across the curriculum
The constructions whose universal properties were established earlier are exactly initial or
terminal objects in disguise. The
universal property of the tensor product
states that bilinear maps out of \(V \times W\) correspond to linear maps out of a single
object \(V \otimes W\) — that object is initial among "vector spaces receiving a bilinear map
from \(V \times W\)." The
universal property of the quotient topology
and the
universal property of the product topology
have the identical shape, one initial and one terminal. The abelianization just discussed is
the same phenomenon once more: \(G^{\mathrm{ab}}\) is initial among abelian groups receiving a
homomorphism from \(G\). Each was proved separately, in its own chapter; the categorical
definition names the one pattern they all instantiate.
Initial and terminal objects are themselves adjoints, which closes the circle. There is exactly
one functor from any category \(\mathscr{C}\) to the terminal category \(\mathbf{1}\) — the
category with one object and only its identity morphism. A functor from \(\mathbf{1}\) to
\(\mathscr{C}\) is the same thing as a choice of object of \(\mathscr{C}\). Viewing objects of
\(\mathscr{C}\) this way, a left adjoint to the unique functor
\(\mathscr{C} \to \mathbf{1}\) is precisely an initial object of \(\mathscr{C}\), and a right
adjoint is precisely a terminal object. The two most basic universal properties are thus the two
simplest adjunctions.
Adjunctions Beyond Algebra
The free–forgetful pattern might suggest that adjunctions are a feature of algebra in particular.
They are not. The same relation organizes topology and the elementary set theory of functions,
and two further examples show its range while introducing structures we will need later.
Topology: A Forgetful Functor Between Two Adjoints
The forgetful functor \(U : \mathbf{Top} \to \mathbf{Set}\), sending a
topological space
to its underlying set of points, sits between adjoints on both sides:
\[
D \dashv U \dashv I.
\]
The left adjoint \(D : \mathbf{Set} \to \mathbf{Top}\) equips a set with the discrete
topology, in which every subset is open; the right adjoint
\(I : \mathbf{Set} \to \mathbf{Top}\) equips it with the indiscrete topology, in
which only the empty set and the whole space are open. The adjunctions express two familiar
facts about continuity. Adjointness \(D \dashv U\) says that a continuous map out of a discrete
space \(D(S)\) is the same thing as an arbitrary function out of \(S\): from a discrete space
every function is continuous, so prescribing a continuous map \(D(S) \to X\) is prescribing a
bare function \(S \to U(X)\). Dually, \(U \dashv I\) says that a continuous map into an
indiscrete space \(I(S)\) is the same thing as an arbitrary function into \(S\): into an
indiscrete space every function is continuous. The discrete and indiscrete topologies are the
finest and coarsest topologies on a set, and adjointness is the precise expression of their
extremal character.
This is the same shape \(F \dashv U \dashv R\) met for groups inside monoids: a single
structure-forgetting functor with a free construction on its left and a cofree one on its right.
The pattern will appear once more, decisively, when categories of functors take the place of
categories of spaces.
Sets: The Adjunction Behind Currying
Fix a set \(B\). Forming the product with \(B\) is a functor
\(- \times B : \mathbf{Set} \to \mathbf{Set}\), sending \(A\) to \(A \times B\). Forming the set
of functions into a fixed target is another: writing \(C^B\) for the set of functions
\(B \to C\), the assignment \(C \mapsto C^B\) is a functor \((-)^B : \mathbf{Set} \to
\mathbf{Set}\). These two are adjoint, \(- \times B \dashv (-)^B\), and the bijection is one used
constantly, usually without naming it.
Proposition: The Currying Adjunction
For all sets \(A\) and \(C\) there is a bijection
\[
\mathbf{Set}\bigl(A \times B,\, C\bigr) \;\cong\; \mathbf{Set}\bigl(A,\, C^B\bigr),
\]
natural in \(A\) and \(C\); that is, \(- \times B \dashv (-)^B\).
Proof:
Given a function \(g : A \times B \to C\) of two arguments, define
\(\bar{g} : A \to C^B\) by holding the first argument fixed: \(\bar{g}(a)\) is the function
\(b \mapsto g(a, b)\). Conversely, given \(f : A \to C^B\), define
\(\bar{f} : A \times B \to C\) by \(\bar{f}(a, b) = \bigl(f(a)\bigr)(b)\). The two operations
are mutually inverse: applied in succession they return
\(\bar{\bar{g}}(a, b) = \bigl(\bar{g}(a)\bigr)(b) = g(a, b)\) and
\(\bigl(\bar{\bar{f}}(a)\bigr)(b) = \bar{f}(a, b) = \bigl(f(a)\bigr)(b)\), recovering \(g\)
and \(f\) respectively. Naturality in \(A\) and \(C\) follows by tracking the two
constructions through precomposition by a function \(A' \to A\) and postcomposition by a
function \(C \to C'\), each of which commutes with fixing and releasing the argument \(b\).
The bijection is the formal content of currying: a function of two arguments is
the same thing as a function of the first argument whose values are functions of the second.
Reading a map \(A \times B \to C\) as assigning to each point of \(A\) a map \(B \to C\) is the
set-theoretic shadow of a structure — the exponential object \(C^B\), the object
that internalizes the morphisms \(B \to C\) — that a category may or may not possess. A category
with finite products in which every such exponential exists, with the currying adjunction holding
throughout, is called cartesian closed. We do not develop the notion here, but
it is the categorical home of function types and of the composition of processes, and the setting
in which the categorical account of learning systems is later staged.
When a Left Adjoint Fails to Exist
Adjunction is not automatic, and the exceptions are as informative as the rules. Let
\(\mathbf{Field}\) be the category of
fields
with
ring homomorphisms
as morphisms. The forgetful functor
\(\mathbf{Field} \to \mathbf{Set}\) has no left adjoint: there is no "free field" on a
set.
Why fields admit no free construction
The obstruction is exactly the feature distinguishing a field from a group or a vector space.
For groups, vector spaces, and monoids, the defining data are operations that are
everywhere defined on the underlying set, subject to equations that hold
everywhere: every element of a group has an inverse, and \(x \cdot x^{-1} = e\) holds
for all \(x\). A structure of this kind — call it the model of an algebraic
theory — always admits a free construction, because one can generate freely and then
impose only the universal equations. A field violates the condition: the inverse operation
\(x \mapsto x^{-1}\) is defined only for \(x \neq 0\), a partial operation rather than a total
one. The theory of fields is therefore not algebraic in this sense, and the existence of a
left adjoint, guaranteed for algebraic theories, fails. That adjunction can fail is what
makes its presence meaningful: when two functors are adjoint, the relation is genuine
information about how two kinds of structure interlock, not a formality available for the
asking.
Duality and Composition
Two structural features of adjunctions deserve recording before we put the relation to work, both
of which amplify its reach. The first is that the entire theory comes in mirror-image pairs; the
second is that adjunctions chain together.
Duality
The notions of this page are organized by the
opposite category
into dual pairs. Terminal is dual to initial: an object is terminal in \(\mathscr{C}\) exactly
when it is initial in \(\mathscr{C}^{\mathrm{op}}\). Right adjoint is dual to left adjoint in the
same way, since reversing all arrows in the defining bijection
\(\mathscr{B}(F(A), B) \cong \mathscr{A}(A, G(B))\) interchanges the roles of \(F\) and \(G\).
Consequently every theorem about initial objects or left adjoints yields, for free, a theorem
about terminal objects or right adjoints — the uniqueness lemma above was stated once and
applied to both cases by exactly this principle. Duality halves the labor of the subject.
Composition of Adjunctions
Adjunctions compose along a chain of categories. Suppose \(F \dashv G\) between \(\mathscr{A}\)
and \(\mathscr{B}\), and \(F' \dashv G'\) between \(\mathscr{B}\) and \(\mathscr{C}\). Then the
composite functors satisfy \(F' \circ F \dashv G \circ G'\), because for objects \(A\) of
\(\mathscr{A}\) and \(C\) of \(\mathscr{C}\) the two adjunction bijections compose:
\[
\mathscr{C}\bigl(F'(F(A)), C\bigr) \;\cong\; \mathscr{B}\bigl(F(A), G'(C)\bigr) \;\cong\;
\mathscr{A}\bigl(A, G(G'(C))\bigr),
\]
naturally in \(A\) and \(C\). The free-forgetful adjunctions stack in exactly this way, and the
pattern of composable structured maps will reappear, one level up, when adjunctions between
categories of functors become the organizing principle of the categorical approach to
learning systems.
With adjunction in hand, the universal properties that organize quotients, products, tensor
products, and free objects are revealed as special cases of one relation — and duality doubles
every result while composition lets these relations be assembled into the layered structures from
which the categorical view of geometry and of learning is later built.