Talk:Divisor (algebraic geometry)

Weil divisor

When you say that a Weil divisor is a linear combination, do you mean you take sums with integral or complex coeficients?

16:23, 28 August 2008 (UTC)

Indeed, this should be specified. It's integral coefficients. I have added this. Thanks. RobHar (talk) 17:05, 28 August 2008 (UTC)

I'm no expert on this, but I think the phrase "In general Cartier behave better than Weil divisors when the variety has singular points" should be expanded or clarified. Isn't it true that if all singularities occur in codimension greater than 1, then Weil divisors will "behave" better, since they won't detect these singularities? Of course, this was probably intended to mean that Weil divisors are meaningless and cannot even be defined unless you have regularity in codimension 1. Hilbertthm90 (talk) 19:19, 4 April 2010 (UTC)

Divisors in complex (analytic geometry)

I think it would be really nice if this article presented a broader view of the notion, with maybe a definition in terms less algebraic. I added a definition for the case of Riemann surfaces, but I am no expert in complex manifolds. Examples in this language should be provided, with for instance a definition of the notion of ampleness in this context. --

Marsupilamov — Preceding unsigned comment added by Marsupilamov (talk • contribs) 15:14, 13 March 2011 (UTC)

Cartier divisors and fraction field

This article is not the correct place to discuss in too much detail the subtleties of rational or meromorphic sections. The proper article to edit to that end is Function field of an algebraic variety or even Function field (scheme theory).

Anyway this article should probably focus on divisors in algebraic varieties, and I doubt the issues Kleiman mentions arise in this context (not that I am an expert, but Hartshorne goes rapidly over the definition)

The section about Weil divisors should be made more precise and the injection of CaCl to WeCl described more clearly.

Marsupilamov (talk) 07:27, 17 April 2011 (UTC)

You're right that this is not the right place to discuss the fraction field in too much detail. Consequently I've simplified the article; it discusses the simplest way that is also technically correct.

It is still very important to understand Cartier divisors on singular (even highly singular) varieties. This is necessary for moduli problems. Kleiman's examples are pretty simple, and they show that these apparent pathologies can arise even on nice spaces. I suggest reading his article (it's linked in the present article through its DOI).

I agree that there is a lot of work to do here. Ozob (talk) 13:13, 17 April 2011 (UTC)

I feel that the description of the actual sheaf of rational functions is not immediately relevant to the notion of a Cartier divisor.

Is it not true that we could define directly the sheaf $K*/O*$ without defining the sheaf K ? This would spare one sheafification, right ?

Also, you accidentally (I presume) reverted the section about linear systems, and I took the liberty to unrevert.

Best,

Marsupilamov (talk) 12:37, 18 April 2011 (UTC)

We could say that a Cartier divisor is a section of K_X^*/O^* and nothing more, but I think that does not provide the reader enough context. What we have in the present article duplicates some of the material at the function field article, but I think that the article is much clearer this way. It might be better still, however, not to define K_X but instead to explain why it is related to the more elementary description of Cartier divisors given in the first paragraph of the section. Right now the article just asserts that the two definitions are the same; it does not explain why.

It is true that we could spare one sheafification. Because sheaving is exact, it carries the exact sequence 0 → O^* → K′^* → K′^*/O^* → 0 of presheaves to the exact sequence 0 → O^* → K_X^* → K_X^*/O^* → 0 of sheaves. (In fancy language, it is part of a geometric morphism of topoi.) But I don't see why we should do it that way. The presheaf K′ is of almost no interest, but K_X is very interesting. I would rather we focus our readers' attention on interesting objects.

Also, my removal of content was entirely unintentional. Thank you for reverting. Ozob (talk) 22:02, 18 April 2011 (UTC)

Line bundle notation

Is the notation ${\mathcal {L}}(D)$ really standard (as opposed to ${\mathcal {O}}_{X}(D)$)? I know Hartshorne uses it but I haven't seen it anywhere else. -- Taku (talk) 11:36, 12 May 2014 (UTC)

In fact, I think I'm going to make the notational change L(D) -> O(D), unless there is an objection. -- Taku (talk) 21:57, 8 January 2015 (UTC)

Principal divisors on Riemann surfaces may have degree =/= 0

The article says: "It follows from the fact that a meromorphic function has as many poles as zeroes, that the degree of a principal divisor is 0." but this is not true in general. As far as I know, this holds only in the case of the Riemann sphere. -- Hiferator (talk) 20:43, 15 March 2015 (UTC)

It is true on any compact Riemann surface. Ozob (talk) 14:34, 16 March 2015 (UTC)

Then compactness should be added as a condition for said statement. -- Hiferator (talk) 19:12, 18 March 2015 (UTC)

Done. Ozob (talk) 23:09, 18 March 2015 (UTC)

Links

"Divisor class group", "divisor class", "linear equivalence of divisors", "linear equivalence", "linearly equivalent", and "linear equivalent" redirect here. BTotaro (talk) 23:01, 11 March 2016 (UTC)

Weil/Cartier versus homology/cohomology

What does it mean by the following sentence? "In topological terms, Weil divisors play the role of homology classes, while Cartier divisors represent cohomology classes." More clarification or reference? -- Y. Huang, 13:19, 27 May 2016

A Cartier divisor is like a line bundle and so one can take its first Chern class, which lives in cohomology (and, ignoring isomorphism/equivalence, these three point views are the same; see Chern class#The Chern class of line bundles). A Weil divisor, on the other hand, is a codimension-one cycle, which lives in homology. In that sentence, it is probably more accurate to talk about Weil divisor classes and Cartier divisor classes, but then I'm being too pedantic. (The point is that passage to homology or cohomology destroys information about how two objects are considered isomorphic/equivalent. If you need that type of information, I think you need to work with complexes more directly.) -- Taku (talk) 20:41, 27 May 2016 (UTC)

With this in mind, I've updated the article. I don't think I really addressed the original question, but I do think the article is a little better. Well, maybe. I was surprised that I couldn't find a reference that really treated divisors in full generality. This was especially true for Weil divisors. What one really wants is, for any morphism X → S, there should be a definition of a relative cycle for X over S, and the codimension one cycles should be the relative Weil divisors. One wants a simple definition like being a formal sum of integral subschemes flat over the base; but if the base isn't integral, then this is clearly too much to hope for. E.g., if S is two copies of P¹ joined at a point, and if X is P²_S, then a relative divisor ought to be two pencils which share a common member, but in particular if we just fix a single line in P² and take the constant family, the family fails to be integral over the singular point of the base. Maybe we should instead require that a relative divisor is flat, reduced and that each fiber is irreducible? I don't know if that's enough to guarantee a good theory. Surely someone has worked this out but at the moment I can't find any references.

The literature on Cartier divisors is better because at least there's EGA, which (despite having mistakes here, as noted famously by Kleiman) tries to develop the whole theory in great generality. Since then there have been some other expositions of related material, e.g., the Stacks Project and the Bosch, Lütkebohmert, Raynaud Neron Models book. While I don't think I've left the article in perfect condition here, it's an improvement.

I get the impression that there are actually four theories to worry about: Weil divisors, Cartier divisors, reflexive sheaves, and invertible sheaves. In good cases these all agree with each other, and in bad cases they don't. But I don't know quite what's going on.

Ideally, once we had a general theory worked out, the relevant functorialities would be clear, and (returning to the original question) we would find that Weil divisors could be pushed forward, making them like homology classes, while Cartier divisors could be pulled back, making them like cohomology classes. Well, up to linear equivalence, which is all we can hope for. It is, as Taku says, the difference between working with classes in homology and cohomology groups versus working with cycles in the complex defining those groups. One expects the homology and cohomology groups to be better behaved but carry less information. (And if they aren't good enough for the task at hand, one ought to work with some complex in the derived category, e.g., with the complex ${\mathcal {O}}_{X}^{\times }\to {\mathcal {K}}_{X}^{\times }$ whose homology is the group of Cartier divisors.) Ozob (talk) 00:31, 31 May 2016 (UTC)

Some comment: I want to mention that Fulton and MacPherson define what is called the bivariant intersection theory (bivariant = covariant + contravariant.) It is discussed in Fulton's "Intersection theory". In this setup, they define the Chow group of an arbitrary morphism X → Y. I don't have a good intuition of this theory so I can't elaborate but maybe this is the general theory of relative Weil divisors? In practical applications, I suppose there is a natural notion of a relative canonical divisor (which is relative dualizing sheaf in the smooth case). It is also natural to consider "relative ampleness" for relative divisors. So, there is no question the concept "relative divisors" exists in nature. There is also the fairly standard notion of the Picard scheme/functor/stack for a relative scheme; so obviously there should be the divisor class group-stack of sort? (But I haven't seen one.) -- Taku (talk) 00:03, 7 June 2016 (UTC)

I don't have an opinion as to the Fulton–MacPherson bivariant intersection theory; I've never looked at it hard. But I would be surprised if it really works well for an arbitrary morphism; I would only expect it to work for a morphism satisfying some conditions (like being a locally complete intersection). The problem (if it is one) with the intersection theory that Fulton defines is that it is too geometric and not homological enough. This is great for his goals, since he can prove many classical geometric results. (And it makes for an excellent book! So much beautiful geometry...) But there are no higher Chow groups; there are no objects in a derived category; and in order for the theory to work in total generality you should need that sort of machinery.

The relative canonical divisor is not the relative dualizing sheaf. The relative canonical divisor is actually a divisor, not a divisor class; it is defined to be the unique divisor supported on the exceptional locus (of a given map) whose class is the relative canonical class. Of course, one needs some conditions in order for such a divisor to exist and be unique; at minimum, I think one needs a morphism which is birational and whose exceptional locus has codimension one, but right now I can't recall if that's sufficient.

Ampleness is really a cohomological phenomenon, so it fits naturally in the framework where one identifies a Weil divisor with a reflexive sheaf and a Cartier divisor with a line bundle. I agree that the concepts of reflexive sheaf and line bundle have obvious generalizations to the relative case, but I don't believe that Weil divisors are always identified with reflexive sheaves; I think you need some conditions like being normal. Nobody really wants to work on non-normal varieties, but sometimes you're forced to. For instance, the Picard functor is defined by its values on arbitrary schemes, and in particular that means you need to consider non-normal varieties. So as long as I'm not convinced that Weil divisors are the same as reflexive sheaves (and I'm not), then I'm not convinced I know what a relative Weil divisor is.

Also, representability of the Picard functor is not obvious. It is not a trivial theorem, just a well-known one. These kinds of representability questions are very subtle; for instance, the Neron models book gives an example of a variety whose Picard functor is representable by a scheme, but that scheme is non-separated! Also, it describes varieties whose Picard functors are not representable by schemes, but which are representable by algebraic spaces. Moreover, the Picard functor does not always represent the obvious functor. There is a Picard presheaf that sends T to Pic(X_T / T), but in general this is not a sheaf. The Picard functor is actually the sheaving of this presheaf with respect to the fppf topology. As an example of when they are different, consider a smooth projective genus one curve C over Q. This curve has a Jacobian J. If C admits a Q-rational point, then C is isomorphic to J. In this case, the Picard presheaf is a sheaf and it is represented by J. But if C has no Q-rational points (there are plenty of such C) then C is not Q-isomorphic to J. The Picard presheaf actually represents the Galois-invariant line bundles over the algebraic closure, and there is an exact sequence

$0\to \operatorname {Pic} C\to J(\mathbf {Q} )=(\operatorname {Pic} {\bar {C}})^{G}\to \operatorname {Br} (\mathbf {Q} )\to \operatorname {Br} (C),$

where G is the absolute Galois group and the superscript denotes invariants. (This is the five-term exact sequence for the Leray spectral sequence of ${\bar {C}}\to C\to \operatorname {Spec} k$.) If you can find a line bundle with non-trivial image in the Brauer group – and this can be done in many cases – then you have proven that the Picard presheaf is not representable. Bjorn Poonen's notes, I think they're called Rational Points on Algebraic Varieties, discuss this sort of question.

One last comment is that there is a functor of relative divisors, but it represents the sheaving of the functor that sends T to relative effective Cartier divisors on X_T / T. (Again, see the Neron Models book.) I don't know of any similar construction for relative effective Weil divisors. Ozob (talk) 02:56, 7 June 2016 (UTC)

O(D) subsheaf of sheaf of regular functions?

The article states "When X is a normal Noetherian scheme and D is an effective divisor, then O(D) is a subsheaf of the sheaf of regular functions, and in particular each of its sections is a regular function."

Should O(D) be instead O(-D)? My understanding was that for an effective divisor, sections of O(D) are rational functions with some poles allowed, while O(-D) are rational functions with no pole allowed and some zero imposed. — Preceding unsigned comment added by 140.105.42.135 (talk) 11:20, 3 August 2016 (UTC)

Yes, I think the sign was wrong. I tried to fix it; let me know if you spot any more errors (or edit the article yourself! New editors are always welcome). Ozob (talk) 12:37, 3 August 2016 (UTC)

I agree the sentence in a question is problematic. I think for this type of stuff, we really need to careful about the distinction between Weil divisors and Cratier divisors. I've made an edit to address this. -- Taku (talk) 20:31, 3 August 2016 (UTC)

I think what we need something like:

To each Cartier divisor D, there are the associated subsheaf L(D) of the sheaf of rational functions and a (non-unique) nonzero rational section s_D such that $\operatorname {div} (s_{D})=D$.Then there is an isomorphism (depending on a choice of s_D):

$O_{X}(D)\to L(D),\,f\mapsto fs_{D}$.

(see also #Global sections of line bundles and linear systems.) The point is that even though one can view O(D) as a subsheaf, this should be done via a choice of s_D.

(Sorry about mess, It's my morning and I'm not thinking clearly.) -- Taku (talk) 21:23, 3 August 2016 (UTC)

What really determines the Cartier divisor is the sheaf O(D) together with its embedding in the sheaf M_X of rational functions. I've tried to clarify this in the article. Let me know what you think. Ozob (talk) 01:56, 4 August 2016 (UTC)

No, no, I'm not disagreeing with that. It's more of a matter of convention: when people talk about the sections of O(K), K canonical divisor class, they don't mean rational functions but rational differential forms. By a line bundle, I was thinking of the canonical map from the space of Cartier divisors to Pic(X). It is necessary to mention that the relationship between O(D) and the image of D under this map. I added an example to explain what I meant. I get th clarification looked strange but I think it's an important point; I will give another shot. -- Taku (talk) 02:36, 4 August 2016 (UTC)

Taku, some of what you wrote does not make sense. If D is a Cartier divisor, what is the difference between O_X(D) and L(D)? Also, why is this so important that it must be stated before the article has defined Cartier divisors? Ozob (talk) 03:03, 4 August 2016 (UTC)

The point of the note is that they have the same isomorphism class, but not literally the same. It's the difference between O(K) and Omega^n (see the canonical divisor class section). Without the clarification, the definition in the section may not make a complete sense. Are sections of O(K) rational functions or rational differential forms? The point is that when one talks about the sections of O(D), there can be a potential confusion; for example, given an effective divisor D, 1 in O(D) never vanish on D but a section s_D corresponding to it vanishes precisely along D. This is why you see the phrasing the section corresponding to "1" (not just for a principal divisor but an effective divisor); fixing s_D fixes what one means by the sections of O(D). The note may have not been written artfully but still an important point. If it helps, I learned this matter from Griffirthsa-Harris (and didn't make up). -- Taku (talk) 03:23, 4 August 2016 (UTC)

Think about a case of an elliptic curve X and a point P on it. What are O(P) and O(3P), more precisely, what are their sections? I don't view their sections as rational functions but I view their sections are sections which vanish at P with order 1 or 3, which correspond to the rational functions with poles along P of order 1 or 3. That is, it matters whether one is thinking the sections as sections or rational functions. (We probably should add this type of an example in the article.) This is precisely why we might not want to stress O(D) as a subsheaf of the sheaf of rational functions (better to think of it as a sheaf). -- Taku (talk) 04:35, 4 August 2016 (UTC)

Thinking this way might illuminate what I'm saying. Given a section s of the sheaf of rational functions, the meaning of div(s) depends whether one is thinking s as a section or as a rational function. In fact, when 1 corresponds to s, div(1) is empty (zero??) divisor while div(s) = D. Similarly, the zero-locus of 1 and the zero-locus of s differ; the former empty, the latter precisely D. So, we can but we don't want to think O(D) as a subsheaf of the sheaf of rational functions. Maybe this sort of the stuff that needs to be mentioned in the article. -- Taku (talk) 05:55, 4 August 2016 (UTC)

Perhaps you can just ignore my above long post. I have just made an edit to implement the point I wanted to make. Please make further edits if needed. About O_X(D) and L(D); logically speaking there is a difference. In the section, O_X(D) is defined for a Weil divisor D and there is also the invertible sheaf associated to a Cartier; namely, the image of the connecting homomorphism from the group of Cartier divisors (i.e., space of sections of a certain sheaf) to Pic(X). A priori, thus, there is a distinction. -- Taku (talk) 23:51, 4 August 2016 (UTC)

What you wrote was extremely confusing because you didn't clearly state that you were using O(D) to mean a Weil divisor and L(D) to mean a Cartier divisor. And it was still using Cartier divisors before the article had defined them. I have replaced it with something better. Ozob (talk) 03:19, 5 August 2016 (UTC)

I think the explanation of a canonical sectio looks very good (obviously more detailed and address the issue I was fussing about more clearly). I will also try to write more clearly in the future (but the current version looks ok). -- Taku (talk) 23:08, 6 August 2016 (UTC)

I think a few more simple examples to reinforce this matter are needed; I'm sufficiently convinced some readers find this matter confusing. When H is a hyperplane in a projective space (or a restriction of it to a polarized projective variety), what are the sections of O(H)? For me, they are precisely those of O(1); that is, degree-one elements of the homogeneous coordinate ring. (I will add such an example if I'm confident that I'm not damaging the article.) -- Taku (talk) 23:25, 6 August 2016 (UTC)

I agree that this point is confusing. The sections of O(H) are, in some sense, the same as the sections of O(1) because the two are isomorphic as line bundles. However, if we consider sections up to isomorphism of line bundles, then we cannot identify the sections with rational functions. In order to identify sections as rational functions, the line bundle must be embedded in the sheaf of rational functions, and most isomorphisms of line bundles do not respect this embedding. For example, if H₁ = {X = 0} and H₂ = {Y = 0} are hyperplanes in P², then O(H₁) and O(H₂) are isomorphic as line bundles, but they are different as subsheaves of the sheaf of rational functions: The rational function 1 / X is in $\Gamma (\mathbf {P} ^{2},{\mathcal {O}}(H_{1}))$ but not in $\Gamma (\mathbf {P} ^{2},{\mathcal {O}}(H_{2}))$. This is precisely the difference between equality and linear equivalence of divisors.

Here's another interesting example. Let E be an elliptic curve (over, say, C), and let O be a flex. Fix an embedding of E in P². The hyperplane sections of E are the effective degree 3 divisors P₁ + P₂ + P₃ that are linearly equivalent to 3O. There are other degree three divisors on E, of course: Any three random points are almost surely not linearly equivalent to 3O. If D is one such divisor, then there is no choice of rational function f on P² whose divisor is precisely D. Indeed, f cannot have poles, so it's a hypersurface section; it must be a degree 1 section by Bezout's formula; but we assumed that D didn't come from a hyperplane section. Despite this, since D is an effective degree three divisor, its complete linear system embeds E into P², and with respect to that embedding, D is a hyperplane section.

The only way to see D in its original embedding is to intersect E with some conic (or higher degree) section C of P². There will be some residual divisor D′ = D − C. The linear system of conics through the residual divisor D′ yields the same embedding into P² as the linear system O(D) except that the former has a base locus and the latter doesn't. When we embed O(D) into the sheaf of rational functions in the way given by the embedding of E that we chose, its global sections correspond to those rational functions which have poles at most along D. If q is the equation of D′, then 1 / q is not a global section of O(D) because it has poles along the residual divisor. However, there are plenty of functions we can multiply it by that get rid of those poles. But in the embedding given by the conics through D′, O(D) embeds in the sheaf of rational functions as O(H) for some hyperplane H. So it's just the rational functions with a fixed linear denominator! Ozob (talk) 00:58, 7 August 2016 (UTC)

First of all, thank you for the detailed explanation of the issue. I need to think about what you wrote myself but I'm convinced we should include this type of the discussion in the article. (I think for a linear system one usually work with Cartier divisors and so this stuff is easier since the definition of a canonical section.) My problem is more of the convention. It seems it is a typical attitude to think O(H) = O(1) in literature; that is, one can and does define O(D) as a subsheaf of the constant sheaf of rational functions. But then you start talk about the sections of O(H); in particular the canonical section is the section corresponding to 1, an element of $\Gamma (P^{n},O(H))$. In fact, even within this article, we write O(K) for the canonical sheaf, which is not really a subsheaf of the sheaf of rational functions. I think you and I agree with math (geometric part like linear equivalence and embedding into a projective space); the only issue is that the typical abuse of identifying the isomorphism class of line bundles with a line bundle. I will see if I can come up with a cleaner exposition (I find the discussion in Griffiths-Harris very clean, except they don't distinguish the iso class of a line bundle and a line bundle :).) -- Taku (talk) 03:42, 8 August 2016 (UTC)

I think the key point that is not emphasized enough in the article is that one needs to be specific about whether one is treating the sections of O(D) as rational functions or as sections of O(D) as an abstract sheaf. This is precisely the difference between 1 and the canonical section. Let $H=\{X=0\}\subset \mathbb {P} ^{n}$ be the hyperplane. Then $(1)+H=H\geq 0$ and so $1\in \Gamma (\mathbb {P} ^{n},{\mathcal {O}}(H))$. The zero-locus of 1 is, of course, empty. But we can look at the zero locus of 1 viewed as a section of ${\mathcal {O}}(H)$ and then its H, since 1 corresponds to the canonical section, which is X in this case. Here, by "correspondence", I mean under the isomorphism ${\mathcal {O}}(H)\simeq {\mathcal {O}}(1)$. In practice, perhaps abusively but commonly, one often "identifies" ${\mathcal {O}}(H)$ and ${\mathcal {O}}(1)$; so 1 and X correspond to each other. The case of the canonical sheaf O(K) goes in the similar way: for a smooth variety, a global/rational section of O(K) should be a global/rational differential form of top degree (which corresponds to a rational function). (For a non-smooth variety, they are extensions of such forms??) Note in this case, the canonical section, denoted by ω in the article, is really the canonical "rational" section; this should not worry you since the canonical divisor is not effective (so a canonical section is not globally defined.) As I said, this is a matter of convention, but this attitude is more natural for me and is consistent with what I've been reading. If you disagree or if I'm plainly wrong, please tell me. All I'm saying that, when doing an explicit calculation with sections, one needs to be aware of how such an identification is made (and this needs to be clearly noted in the article.) -- Taku (talk) 04:17, 8 August 2016 (UTC)

It's very unusual to work with sections as rational functions. Usually one is only interested in the locus of zeroes and poles; the actual rational function usually doesn't matter. This is why it's common in the literature to identify O(H) and O(1). One computes $\Gamma (X,{\mathcal {O}}(H))$ as $\Gamma (X,{\mathcal {O}}(1))$, and one computes that group using the long exact sequence in sheaf cohomology,

$0\to \Gamma (X,{\mathcal {I}}_{X}(1))\to \Gamma (\mathbf {P} ^{n},{\mathcal {O}}_{\mathbf {P} ^{n}}(1))\to \Gamma (X,{\mathcal {O}}(1))\to H^{1}(X,{\mathcal {I}}_{X}(1))\to \cdots .$

One hopes that I_X(1) has some tractable description. For example, perhaps X is a smooth divisor. Perhaps one can apply Kodaira vanishing to show that the H¹ is zero. It depends on the situation, though, since not all varieties of interest are smooth and not all fields are C.

Usually one gets at the canonical divisor, and hence the canonical sheaf, using the adjunction formula or something like that. But this is not always true; when using cohomological tools it has other descriptions (that apply when X is Cohen–Macaulay; see below), and when working analytically one can actually try to integrate things (or get L² bounds, that sort of stuff).

For a non-smooth X, the canonical sheaf is a tricky object. When X is not smooth, it's not the top exterior power of the cotangent sheaf. When X is Gorenstein, it's a line bundle; when X is Cohen–Macaulay but not Gorenstein, it's a sheaf but not a line bundle; in both cases it's defined cohomologically instead of geometrically. When X isn't even Cohen–Macaulay, there is no canonical sheaf. Ozob (talk) 13:00, 8 August 2016 (UTC)

Effective divisors as subschemes

It is standard (e.g., Hartshorne) that any effective Cartier divisor can be viewed as a subscheme (since locally such a divisor is just the zero-locus of a regular function). Is this true for any effective Weil divisor? On a smooth variety, yes, but, in general, this seems unclear. (Probably true for reduced divisors or prime divisors, but probably not so for an effective divisor in general). At least I cannot see how this can be done. -- Taku (talk) 05:19, 18 August 2016 (UTC)

How is this not obvious? A prime divisor D has an ideal sheaf I_D. If E is a sum n₁D₁ + ... + n_kD_k, then it has an ideal sheaf $I_{D_{1}}^{n_{1}}\cdots I_{D_{k}}^{n_{k}}$. This determines a subscheme. Ozob (talk) 13:27, 18 August 2016 (UTC)

Because I don't think the argument is right (it's already false in the level of commutative algebra). For example, O(-2D) is already not necessary O(-D)^2; without some smoothness. (It works for a reduced divisor, i.e., the coefficients are all 1) Ok. If you replace products with the reflexive hull of tensor products, then it probably works but that may be original research. -- Taku (talk) 01:52, 20 August 2016 (UTC)

More crucially, I cannot find any reliable source telling us how to view an effective Weil divisor as a subscheme, which makes me very uncomfortable seeing the statement to the effect (not to mention the violation of the policy.) -- Taku (talk) 01:56, 20 August 2016 (UTC)

Maybe it would help for me to explain in greater detail what I have in mind in more detail, especially since you caught me in a mistake (sorry, I was rushed). Let X be a normal integral Noetherian scheme (I've always assumed that was part of the context, but maybe that's been unclear). Let D be a Weil divisor. By definition,

${\mathcal {O}}(-D)(U)=\{0\}\cup \{f\in k(X):-D+\operatorname {div} f\geq 0\}.$

The non-zero sections are assumed to be rational functions. We will prove that if D is effective, then they are regular. To see this, it suffices to work on open affine sets U. Recall the following fact from commutative algebra: If A is a Noetherian integrally closed domain, then it is the intersection of its localizations $A_{\mathfrak {p}}$, where ${\mathfrak {p}}$ varies over the height one prime ideals of A. We are given f is in k(A) such that for each height one prime ideal ${\mathfrak {p}}$, $v(-D)+v(f)\geq 0$, where v is the valuation associated with ${\mathfrak {p}}$. (Since A is normal, there is such a valuation.) Because D is effective, v(−D) is non-positive. The only way that the sum with v(f) can be non-negative is if v(f) is non-negative. In particular, f must lie in the local ring $A_{\mathfrak {p}}$. This is true for all ${\mathfrak {p}}$, so f lies in $\cap _{\mathfrak {p}}A_{\mathfrak {p}}$. As noted earlier, this intersection equals A, so f is regular, not just rational.

It follows that ${\mathcal {O}}(-D)$ is an ideal sheaf. Above I erred and said that it was a product of ideal sheaves, but as you pointed out, it's not. Nevertheless, it is certainly an ideal sheaf. The closed subscheme defined by this ideal sheaf is the closed subscheme associated to D.

After much searching, I've found a partial reference. In EGA IV₄, 21.7.1, Grothendieck defines the closed subscheme associated to an effective Weil divisor (in his terminology, a positive 1-codimensional cycle). It's not the definition I gave above; it doesn't need a normality hypothesis. I suspect the two are equivalent for normal integral locally Noetherian schemes, but I haven't tried to check it yet. Ozob (talk) 05:12, 20 August 2016 (UTC)

I also skimmed EGA but couldn't find anything relevant (but I didn't read it carefully). I do agree that O(-D) is an ideal sheaf for a Noetherian normal integral scheme (perhaps more generally for a scheme whose local rings are Krull domains, but let's not get fancy here.) The problem I had with the proposed definition is that the exact sequence mentioned in the article is tautology; in this approach, O_D is defined by that sequence. One can get an effective divisor from a pure-codimension-one subscheme (namely by taking the fundamental cycle of the subscheme). In that case, the exact sequence is nontrivial; hence, my proposed change of the wording. (I don't have a strong opinion on the presentation, except it needs to be correct :)) - Taku (talk) 03:02, 21 August 2016 (UTC)