Why a ring should be a space at all
Guide 1 of this rung built locally ringed spaces in the abstract: a topological space with a sheaf of rings whose stalks are local. Now we manufacture them by the truckload from pure algebra. The lesson of classical algebraic geometry was that a variety over an algebraically closed field k is captured by its coordinate ring — polynomials modulo the equations of the variety. The Nullstellensatz makes the dictionary precise for that case: points of the variety correspond to maximal ideals of the coordinate ring. Grothendieck's move was to drop k, drop reducedness, drop finite generation, and ask the question for ANY commutative ring R: what space has R as its functions?
The first instinct — use maximal ideals as points — almost works but breaks under maps. If phi: R -> S is a ring homomorphism, you would like it to pull maximal ideals of S back to maximal ideals of R, so that a map of rings induces a map of spaces in the opposite direction. It does not: the preimage of a maximal ideal need not be maximal. The honest example is the inclusion Z -> Q; the maximal ideal (0) of Q pulls back to (0) in Z, which is prime but NOT maximal. Prime ideals, by contrast, pull back to prime ideals under any homomorphism. That single robustness is why we change the question from maximal to prime.
Spec R as a topological space
Define the spectrum of R, written Spec R, to be the set of all prime ideals p of R. We topologize it with the Zariski topology, specified by its closed sets: for any ideal I of R, the set V(I) = { p in Spec R : I is contained in p } is declared closed, and these are exactly the closed sets. Equivalently, for a single element f the basic OPEN set D(f) = { p : f is not in p } is where f is 'nonzero', and the D(f) form a basis for the topology. Think of V(I) as 'the common zero locus of the functions in I' — it is the spectrum's version of a vanishing set, just with primes standing in for points.
Here is where the picture becomes vivid. The elements of R behave like functions on Spec R: given f in R and a point p, the 'value of f at p' is the image of f in the residue field kappa(p) = Frac(R/p), the field of fractions of the integral domain R/p. So f 'vanishes at p' precisely when f lies in p — exactly matching the D(f) above. Two warnings come free with this picture. First, the value of f lives in a field that CHANGES from point to point; there is no single target. Second, an element can be nonzero yet vanish at every point: in R = k[x]/(x^2), the function x is not zero but lies in the only prime ideal (x), so it vanishes everywhere. Nilpotents are invisible to the topology but alive in the ring.
Two small worked spectra anchor everything. Spec Z has one point for each prime number — the maximal ideals (2), (3), (5), ... — plus the single non-closed point (0), the generic point, whose closure is all of Spec Z. Spec C[x] looks like the complex line: one closed point (x - a) for each a in C, again plus a generic point (0). And the residue field tells you what kind of point you are at: at (x - a) it is C (a classical point), at (0) it is C(x), the field of rational functions (the generic point seeing the whole line at once). One space, two flavors of point, all bookkept by primes.
The structure sheaf: gluing functions onto the spectrum
A space alone is not a geometric object; we need its functions, organized into a sheaf as Guide 1 demanded. The structure sheaf O_{Spec R} is the sheaf of rings that assigns to each open set the functions defined there. The cleanest way to pin it down is by its values on the basic opens D(f): we set O(D(f)) = R_f, the localization of R inverting f. Localizing at f is exactly 'allowing 1/f as a function', which is legitimate precisely on the set D(f) where f does not vanish. On the whole space D(1) = Spec R the sections are R itself, recovering the slogan 'the ring R is the global functions on Spec R'.
- On each basic open D(f), declare the sections to be the localization R_f = R[1/f]. This says: on the region where f is invertible, the legal functions are ratios with powers of f allowed in the denominator.
- Check the overlaps are consistent: if D(g) is contained in D(f), then g becomes invertible once f is, giving a natural restriction map R_f -> R_g. These maps agree on triple overlaps, so the data is a genuine presheaf on the basis.
- Extend to ALL open sets by the sheaf condition: a section over a general open U is a compatible family of sections over a covering by basic opens, glued. This is forced — the sheaf axioms determine O(U) uniquely from its values on the D(f).
- Compute the stalk at a prime p: it is the local ring R_p, the localization of R at the prime p (inverting everything outside p). Its unique maximal ideal is p R_p, so every stalk is local — and that is what makes Spec R a LOCALLY ringed space.
The stalk computation in step 4 is the payoff and deserves a slow look. At a point p the stalk O_p is the local ring R_p, and its maximal ideal is the set of functions that vanish at p, its residue field R_p / p R_p is the kappa(p) from before. So the abstract demand of Guide 1 — every stalk is a local ring — is met automatically, with concrete content: 'germs of functions near p, modulo those vanishing at p, give the value at p'. The structure sheaf is not an extra choice bolted on; it is forced by the algebra of localization, the same localization that already governed which functions are allowed on which open set.
Affine schemes and the great equivalence
Bundle the two halves together. An affine scheme is the locally ringed space (Spec R, O_{Spec R}) for some commutative ring R, and a morphism of affine schemes is a morphism of locally ringed spaces between them. That last clause matters: a morphism is a continuous map f together with a map of sheaves f# pulling functions back, AND f# must respect the local structure — it sends germs vanishing at the target point to germs vanishing at the source point. This locality condition is not automatic, and it is exactly what rules out pathological maps and makes the next theorem clean.
Now the theorem that makes the whole construction worth the trouble: the functor R |-> Spec R is a contravariant equivalence between commutative rings and affine schemes. Concretely, the set of morphisms of affine schemes from Spec S to Spec R is in natural bijection with the set of ring homomorphisms from R to S — arrows reverse, and nothing is lost or invented. A ring map phi: R -> S literally IS a map of spaces Spec S -> Spec R (pulling back primes), and the recovered global functions on each side reproduce R and S exactly. Algebra and geometry are not merely analogous here; they are the same category read in two directions.
The affine dictionary (R, S commutative rings):
ALGEBRA GEOMETRY
---------------------------------- --------------------------------------
commutative ring R <-> affine scheme Spec R
prime ideal p of R <-> point of Spec R
maximal ideal <-> closed point
the zero ideal (if R a domain) <-> generic point
ideal I <-> closed subset V(I)
localization R_f <-> sections over the open D(f)
localization R_p <-> stalk at the point p
nilpotent element <-> function invisible to topology
ring map phi : R -> S <-> morphism Spec S -> Spec R (REVERSED)
tiny examples:
Spec Z : one point per prime (2),(3),(5),... + generic point (0)
Spec k : a single point (k any field)
Spec k[x] : a line of closed points (x-a) + generic point (0)
Spec k[e]/(e^2) : one point, but a FAT one (a tangent direction's worth of data)What the abstraction buys, and where it goes next
It is fair to ask whether all this machinery earns its keep, since over an algebraically closed field a variety already does a lot of work. It earns it in three concrete ways. First, NILPOTENTS become legitimate: Spec k[e]/(e^2), the dual numbers, is a single point carrying a tangent direction's worth of data, so 'infinitesimal thickenings' and tangent spaces become honest geometric objects rather than hand-waving. Second, ARITHMETIC joins geometry: Spec Z is a genuine space, and a scheme over Z lets you do geometry over the integers, the bridge to arithmetic geometry. Third, BASE FIELDS are no longer required, so reduction mod p, mixed characteristic, and families over a base all live in one framework.
Be honest about the limits of what we built. An affine scheme is only the LOCAL model — exactly as a single chart is the local model of a manifold, but no compact manifold is a single chart. Projective space CP^n is the cleanest example that is decidedly not affine: its only global functions are the constants, so its ring of global sections is just the base field and cannot possibly remember the whole space. The way forward is the one Vol I's charts and atlases foreshadowed: glue affine schemes along open sets to build general schemes, whose locally ringed structure makes the gluing automatic.
That gluing is precisely Guide 3 of this rung, where schemes and morphisms get their full definition and the functor-of-points viewpoint reframes everything once more. For now hold the core picture: a commutative ring R becomes a space Spec R of prime ideals under the Zariski topology, the localizations R_f and R_p assemble into a structure sheaf whose stalks are local rings, and the resulting affine scheme stores R losslessly. Every scheme to come is built by gluing these pieces, exactly as every manifold was built by gluing Euclidean charts — same instinct, new category.