Where we are on the ladder
The previous rung, Algebraic Geometry I, left you fluent in varieties: an affine variety is the common zero locus of polynomials, its Zariski topology has closed sets cut out by equations, and its geometry is dictated by its coordinate ring. You also met regular and rational functions — the polynomial-like functions that live on open sets. This rung rebuilds all of that on a far more flexible foundation, the scheme, which will let nilpotents, non-closed points, and arbitrary base rings into the picture. But schemes are not the first idea; they are the payoff. The first idea, the one that makes everything else possible, is a careful theory of how locally-defined functions glue together. That theory is the subject of this guide.
An honesty note before we begin. These are graduate topics; we assume comfort with point-set topology, commutative algebra (rings, ideals, localization, modules), and category-theoretic reflexes (objects, morphisms, functors). We will say 'open set', 'ring', 'localization', 'functor' and lean on what you already know rather than re-deriving it. The abstraction in this guide can feel ceremonial at first — why all this machinery for 'functions on a space'? Resist the temptation to read it as mystification. The sheaf axiom is doing one concrete job, gluing, and we will keep showing you the job rather than the ceremony. The reward, two guides from now, is that the SAME definitions that govern smooth functions on a manifold will govern algebraic functions on a scheme, with no change of language.
Presheaves: bookkeeping local data
Start with the most modest idea. Fix a topological space X. A presheaf F (of, say, rings) assigns to every open set U an object F(U) — think 'the functions you are allowed to write down on U' — and to every inclusion of opens V inside U a restriction map F(U) -> F(V), written s -> s|_V, which simply forgets the behavior outside V. Two bookkeeping rules are demanded: restricting to U itself does nothing, and restricting from U down to W in two steps (through an intermediate V) agrees with restricting in one step. That is the entire definition of a presheaf: data on each open, plus consistent restriction.
An element of F(U) is called a section over U, and an element of F(X) a global section — the name comes from the picture of a section of a bundle, a choice of value over each point, here packaged open-set by open-set. The cleanest example to hold in mind: let F(U) be the ring of continuous real functions on U, with restriction the literal restriction of functions. Or, the example this whole rung is built toward, let F(U) be the regular functions on an open subset of a variety. In both cases F really is a presheaf — restriction obviously composes correctly. The category-theoretic one-liner, worth absorbing because it makes later proofs automatic: a presheaf is exactly a contravariant functor from the poset of opens to the category of rings. 'Contravariant' because the arrow reverses — a bigger open restricts TO a smaller one.
The sheaf axiom: local determines global
A sheaf is a presheaf that glues. Make that precise with two conditions on every open U and every open cover U = union of U_i. Locality (separation): if a section s in F(U) restricts to zero on every U_i, then s = 0 — a section is determined by its restrictions, nothing hides between the pieces. Gluing: if you are given sections s_i in F(U_i) that agree on every overlap, s_i and s_j matching on U_i intersect U_j, then there exists a section s in F(U) restricting to each s_i. Together, locality says the glued section is unique when it exists, and gluing says it does exist. A sheaf is therefore a presheaf in which 'consistent local data' and 'a global object' are the same thing.
Sheaf axiom, for every open U with open cover U = union of U_i :
data: sections s_i in F(U_i), one per piece
match: s_i |_(U_i cap U_j) = s_j |_(U_i cap U_j) on every overlap
----------------------------------------------------------------------
glue: there EXISTS s in F(U) with s |_(U_i) = s_i (existence)
sep: such an s is UNIQUE (separation)
Equivalently, the following is exact (an equalizer):
F(U) -> prod_i F(U_i) =====> prod_(i,j) F(U_i cap U_j)
(restrict to i,j) vs (restrict to j,i)Now re-examine the examples. Continuous functions DO form a sheaf: continuity is a local condition, so functions agreeing on overlaps patch to a single continuous function, and a function that is zero on every piece of a cover is zero. Smooth functions on a manifold, holomorphic functions on a complex manifold, regular functions on a variety — all sheaves, for the same reason: each defining property is checkable in arbitrarily small neighborhoods. The bounded-functions presheaf from the callout is NOT a sheaf, because boundedness is a GLOBAL condition, not a local one — and that is the precise diagnosis of its failure. The sheaf axiom is, in one slogan, the formalization of 'a property that can be checked locally yields objects that can be built locally.'
Stalks: zooming in on a single point
Sections live on open sets, but geometry often happens AT a point. How do you extract the behavior of a sheaf infinitely close to a fixed p in X, without committing to any particular neighborhood? The answer is the stalk, F_p, built as a colimit (a 'direct limit') over all open neighborhoods of p. Concretely, an element of the stalk is a germ: a pair (U, s) of an open U containing p and a section s over U, where two germs (U, s) and (V, t) are declared equal if s and t already agree on some smaller open neighborhood W of p inside both. A stalk thus throws away everything except the infinitesimally-local behavior at p — two functions that look identical near p define the same germ even if they differ far away.
Why is this the right notion of 'value plus all derivatives at p'? Because germs remember more than a single value. For the sheaf of smooth functions, the stalk at p is the ring of germs of smooth functions — it knows every Taylor coefficient, since two functions with the same germ have the same value, the same gradient, the same Hessian, all of it. This is the algebraic shadow of 'infinitely close.' And the stalk is not just a set: it inherits the ring structure, and for the sheaves we care about it is a LOCAL ring — the germs vanishing at p form the unique maximal ideal, while a germ nonzero at p is invertible nearby (a continuous or regular function nonzero at p stays nonzero on a small neighborhood, so 1/s is a legitimate germ there). That localness is the seed of the next section.
Sheafification: forcing a presheaf to glue
Many natural constructions produce a presheaf that is not yet a sheaf — the bounded-functions example, but also the presheaf you get by taking 'constant functions with value in a fixed group' (locally constant is the sheaf you actually want), or the image presheaf of a map of sheaves. We need a canonical way to repair a presheaf into the closest sheaf. That repair is sheafification: from a presheaf F we build a sheaf F^+ together with a map F -> F^+ that is universal — any map from F to a sheaf factors uniquely through F^+. The slogan is that sheafification is the best possible sheaf approximation to F, changing the data as little as possible.
- Compute the stalks F_p of the presheaf — these are already correct, because stalks are colimits over neighborhoods and never see the gluing failure. The stalks survive sheafification untouched.
- Define F^+(U) as the set of functions s assigning to each point p in U an element s(p) of the stalk F_p, subject to a LOCAL compatibility: near every point the chosen germs must come from one honest section of F on a common neighborhood.
- This F^+ is automatically a sheaf — both axioms hold because the data is, by construction, 'a coherent choice of germ at every point,' which is exactly what gluing demands.
- The natural map F -> F^+ sends a section to its tuple of germs; it is an isomorphism on every stalk, and it is already an isomorphism whenever F was a sheaf to begin with — so sheafification leaves genuine sheaves alone.
Notice the role the stalks played: sheafification rebuilds the sheaf entirely out of its stalks plus a local-compatibility rule. That is why germs are not a side curiosity but the structural heart — they are the indestructible local data, and a sheaf is precisely 'a compatible collection of germs.' Honest caveat: this construction, like the equalizer description of the axiom, is one of several equivalent definitions, and books differ in whether they build F^+ via germs, via a two-step plus-construction, or via the étalé space. They produce the same sheaf up to canonical isomorphism, but if you read two sources side by side, expect the recipes to look unalike. Pick one construction, learn it cold, and trust that the others land in the same place.
Ringed spaces: a space that carries its functions
Now assemble the payload. A ringed space is a pair (X, O_X) of a topological space X and a sheaf of rings O_X on it, called the structure sheaf — the official ring of functions on each open set. This single package abstracts every geometry you know: a smooth manifold is its underlying space together with the sheaf of smooth functions; a complex manifold is the space with the sheaf of holomorphic functions; a variety is its Zariski space with the sheaf of regular functions, the structure sheaf this rung will define for schemes. The genius of the abstraction is that it makes 'a space and the functions allowed on it' into a single mathematical object you can map between.
A morphism of ringed spaces is therefore richer than a continuous map. It is a continuous map f: X -> Y PLUS a comparison of structure sheaves going the OTHER way — a rule pulling functions on Y back to functions on X (because if g is a function near a point f(x) in Y, then g composed with f is a function near x). Spelled out, this is a map of sheaves O_Y -> f_* O_X, the algebra following the geometry backwards exactly as restriction reversed inclusions. This 'space plus a sheaf of rings, with sheaf-respecting maps' is the genuine category in which schemes will live, and morphisms of schemes will be morphisms of ringed spaces, no more and no less.
One refinement makes the abstraction match real geometry: a locally ringed space demands that every stalk O_(X,p) be a local ring, and that morphisms send the maximal ideal into the maximal ideal — pullback of a function vanishing at f(x) must vanish at x. This sounds technical but encodes something obvious: the maximal ideal of the stalk is 'functions vanishing at the point,' so the condition just says a morphism respects WHERE functions vanish. A locally ringed space is the correct setting, because honest geometric spaces have this property and pathological gluings of rings do not. With this in hand, the road forward is clear and we will follow it next guide: take any commutative ring R, manufacture from it a topological space and a structure sheaf — the spectrum and its structure sheaf — and the resulting locally ringed space is the affine scheme Spec R, the atom from which all schemes are glued.