This post is a followup on this series of posts. Basically, I in some sense forgot the punchline of the whole story and just now it came back to me.

Let’s consider the case of the Quot functor for example. First let’s consider it in a reasonable level of generality: assume we have a proper morphism X —> S with S Noetherian and a coherent O_X-module F flat over S. Then given a Noetherian S-algebra A and a short exact sequence

0 —> E —> F_A —> G —> 0

on X_A = X ×_S Spec(A) with G flat over A the obstructions to deforming this to a thickening A’ of A lie in Ext^1_{O_X}(E, G ⊗ I) and if the obstruction vanishes the set of deformations is principal homogeneous under Ext^0_{O_X}(E, G ⊗ I).

In order to apply Artin’s method to get openness of versality exactly as in Section 07YV we need to prove there exists a perfect complex of A-modules K such that we have

Ext^1_{O_X}(E, G ⊗ M) = H^2(K ⊗_A M)

and

Ext^0_{O_X}(E, G ⊗ M) = H^1(K ⊗_A M)

functorially in the A-module M and agreeing with boundary maps. What is important here is that you only need to have this for low cohomology groups.

Guess: There exists a bounded below complex of finite projective A-modules P and functorial isomorphisms Ext^i_{O_X}(E, G ⊗ M) = H^{i + 1}(P ⊗_A M).

Firstly, if the guess is true, then we get what we want by just taking K to be a stupid truncation of P!

Secondly, if X —> S is projective, then the guess is correct. Namely, in this case you can find a resolution

…—> E_1 —> E_0 —> E —> 0

of E where each E_i is a finite direct sum O_X(-n) where n ≫ 0. Then the complex

P = (f_*Hom(E_0, G) —> f_*Hom(E_1, G) —> … )

works because each of the flat sheaves Hom(E_i, G) will have only cohomology in degree 0 and hence f_*Hom(E_i, G) is finite locally free on Spec(A). Moreover, the complex will compute the correct cohomology after tensoring with M. (Observation: you do not need E to be flat over A for this argument to work; the key is that G is flat over A.)

Thirdly, in general, if we can find a perfect complex E’ on X and a map E’ –> E which is a quasi-isomorphism in degrees > -5 (for example), then in order to compute Ext^i(E, G ⊗ M) for i < 4 (or something) we can replace E by E’. This leads us to something like Rf_*Hom(E’, G) which is a perfect complex on Spec(A) by standard arguments. Grothendieck uses this argument to study Ext^0 in EGA III, Cor 7.7.8. Anyhow, this should allow us to handle the case where you have an ample family of invertible sheaves on X (that’s not much better than the projective case of course). [Edit: Jack just emailed that the existence of E’ (in the case of schemes) is in a paper by Lipman and Neeman entitled “QUASI-PERFECT SCHEME-MAPS AND BOUNDEDNESS OF THE TWISTED INVERSE IMAGE FUNCTOR”.]

Fourthly, it may well be that the guess follows from Jack Hall’s paper “Coherence results for algebraic stacks”. It is obviously very close and it may just be a translation, but I haven’t tried to think about it. I’d like to know if this is so. [Edit: Jack just explained to me that this is true for example if the base is of finite type over Z which we can reduce to I think.]

Fifthly, if F is not assumed flat over S, then we should probably do something like look at distinguished triangles

E —> F ⊗^L A —> G —> E[1]

where now E is an object of D(X_A) but G is a usual O_{X_A}-module (placed in degree 0) flat over A and hopefully the obstructions lie in Ext^1(E, G ⊗ I), etc, and we can try imitating the above. This is just pure speculation; I’ll have to check with my friends to see what would be the correct thing to do. Just leave a comment if you know what to do.

It is standard practice to construct the Hilbert scheme as a special case of the Quot scheme. Often you can construct the Quot scheme out of a Hilbert scheme too.

Namely, suppose you have X —> S a flat, proper morphism of finite presentation and suppose that F is a finitely presented O_X-module. Then you can consider

Y = Spec(O_X[F]) —> X

where O_X[F] = O_X ⊕ F is the O_X-algebra where F is a square zero ideal. We have a section σ : X —> Y. Then we can consider the closed subscheme Q of Hilb_{Y/S} parametrizing families of closed subschemes of Y which contain σ. If I am not mistaken, then Q = Quot_{F/X/S}.

This only works because we assumed X —> S is proper and flat!

Let A —> B be a finitely presented ring map. Let M be a finitely presented B-module flat over A. Then the projective dimension of M as an A-module is at most 1.

This follows from a result of Jensen which you can find as Theorem 3.2 in Lazard’s paper “Autour de la platitude” which says that a countably presented flat A-module has projective dimension at most 1.

In Raynaud-Gruson they prove that “locally” M is actually projective. Namely, suppose that q is a prime of B lying over a prime p of A. Then there exists an \’etale ring map (A, p) —> (A’, p’) inducing a trivial residue field extension k(p) = k(p’) and an element h ∈ B’ = B ⊗_A A’ not contained in the unique prime q’ ⊂ B’ lying over q and p’ such that M’_h is a projective A’-module, where M’ = M ⊗_A A’.

See Lemma 05ME. The key case is where A and B are of finite type over Z.

Challenge: Find a simple proof of Lemma 05ME.

Naively, you might think there is a chance as we only need to reduce the projective dimension from 1 to 0… Of course, it probably is hopeless to find an elementary proof, perhaps even more hopeless than the challenge in the previous blog post.

Let f : X —> B be a morphism of algebraic spaces. Let u : F —> G be a map of quasi-coherent O_X-modules. Consider the functor

F : (Sch/B) —> (Sets), T |—> singleton if u_T is zero and empty else

This functor always satisfies the sheaf property for the fpqc topology (Lemma 083H). It turns out that if f is locally of finite presentation, G is locally of finite presentation, G is flat over B, and the support of G is proper over B, then F is an algebraic space and F —> B is a closed immersion. This is Lemma 083M and the proof uses the Raynaud-Gruson techniques.

Challenge: Give a simple proof of Lemma 083M.

A while back I tried to do this. First, some reductions: you can reduce to the case where B is an affine scheme. You can reduce to the case where f is proper and locally of finite presentation (replace F by image of F in G and replace X by suitable closed subspace supporting G). I think you can also reduce to the case where F is of finite presentation (by a limit argument). Hence, if you like, you can reduce to the case where B is the spectrum of a Noetherian ring and everything is of finite type.

In the case B is affine there is a simple argument that shows: if u_T = 0 for some quasi-compact T over B, then there is a closed subscheme Z ⊂ B such that u_Z = 0 and such that T —> B factors through Z (Lemma 083K). The proof only uses that G is flat over B.

The problem left over is somehow: What if we have infinitely many closed subschemes Z_1, Z_2, Z_3,… ⊂ B such that u_{Z_i} is zero. Why is it true that u_Z = 0 where Z is the scheme theoretic closure of ⋃ Z_n? E.g., what if B = Spec(Q[x, y]) and Z_n is cut out by the ideal (x^n, y – 1 – x – x^2/2 – x^3/6 – … – x^n/n!).

If F is globally generated then you can reduce to the case F = O_X and you can use that Rf_*G is (universally) computed by a perfect complex. This is related to Jack Hall’s paper “Coherence results for algebraic stacks”. Note that Lemma 083M is a consequence of the results there. Jack’s paper uses relative duality which we do not have available in the Stacks project.

If f is projective, you can reduce to the case F = O_X and G such that Rf_*G is universally computed by a finite locally free sheaf, whence the result. This case is straightforward using only standard results.

If the support of G is finite over B then the result is elementary. So you could try to argue by induction on the relative dimension. Alas, I’m having trouble producing enough quotients of G which are flat over B.

I still think something simple might work in general. But I don’t see it. Do you?

Suppose that X is a separated algebraic space of finite type over a ring A. Let W be an affine scheme and let f : W —> X be a surjective \’etale morphism. There exists an integer d such that all geometric fibres of f have ≤ d points. Picking d minimal we get a nonempty open U ⊂ X such that f^{-1}(U) —> U is finite etale of degree d. Let

V ⊂ W x_X W x_X … x_X W (d factors)

be the complement of all the diagonals. Choose an open immersion W ⊂ Y with Y projective over A (this is possible as W is of finite type over A). Let

Z ⊂ Y x_A Y x_A … x_A Y (d factors)

be the scheme theoretic closure of V. We obtain d morphisms g_i : Z —> Y. Then we consider

X’ = ⋃ g_i^{-1}(W) ⊂ Z

Claim: The morphism X’ —> X (coming from the g_i and W —> X) is projective.

The image of X’ —> X is closed and contains the open U. Replace X by X ∖ U and W by the complement of the inverse image of U; this decreases the integer d, so we can use induction. In this way we obtain the following weak version of Chow’s lemma: For X separated and of finite type over a ring A there exists a proper surjective morphism X’ —> X with X’ a quasi-projective scheme over A.

But this post is really about the proof of the claim. This claim comes up in the proof of Chow’s lemma in Mumford’s red book as well. I’ve never been able to see clearly why it holds, but now I think I have a good way to think about it.

It suffices to prove that X’ —> X is proper. To do this we may use the valuative criterion for properness. Since V is scheme theoretically dense in X’ it suffices to check liftability to X’ for diagrams

Spec(K) -------> V
  |              |
  v              v
Spec(R)------->  X

where R is a valuation ring with fraction field K. Note that the top horizontal map is given by d distinct K-valued points w_1, …, w_d of W and in fact this is a complete set of inverse images of the point x in X(K) coming from the diagram. OK, and now, since W —> X is surjective, we can, after possibly replacing R by an extension of valuation rings, lift the morphism Spec(R) —> X to a morphism w : Spec(R) —> W. Then since w_1, …, w_d is a complete collection of inverse images of x we see that w|_{Spec(K)} is equal to one of them, say w_i. Thus we see that we get a diagram

Spec(K) -------> Z
  |              | g_i
  v              v
Spec(R) --w--->  Y

and we can lift this to z : Spec(R) —> Z as g_i is projective. The image of z is in g_i^{-1}(W) ⊂ X’ and we win.

Update 10/23/12: This has now been added to the Stacks project. See Lemma Tag 089J.

Let me try again to find a tweak to the proof of Grothendieck’s existence theorem.

Let X be a Noetherian scheme, Z ⊂ X a closed subscheme, and U ⊂ X the complement. Denote X’ the completion of X along Z (this is a formal scheme). Suppose that we have a triple (F, G, a) where F is a coherent O_X-module, G is a coherent O_{X’}-module, and a : F’ —> G is a map of the completion F’ of F along Z to G whose kernel and cokernel are annihilated by a power of the ideal sheaf of Z. Then there exists a unique coherent O_X-module H with H’ = G and a map F –> H which produces a on completion along Z. This follows from the results on formal glueing which we discussed in this blog post, although this particular statement is a bit easier to prove.

Suppose now that X —> Spec(R) is a proper morphism of schemes with R a Noetherian ring complete wrt an ideal I. Let G be a coherent module on the completion of X along the ideal IO_X. We want to show that G is the completion of a coherent module on X. By Noetherian induction we may assume this is the case whenever G is supported on a proper closed subscheme of X. Chow’s lemma gives U ⊂ X an dense open subscheme and f : Y —> X a U-admissible blowup such that Y is projective over R. By the projective case (which is “easy”) we know that f^*G is the completion of a coherent module H on Y. Let J ⊂ O_X be a quasi-coherent ideal sheaf with Z = V(J) = X – U. Let X’ be the completion of X along Z. By our induction hypotheses the modules G/J^mG are I-adic completions of coherent O_X/J^m modules E_m. The system E_m gives rise to a coherent module E on X’. Then for some n > 0 we obtain a triple (J^nf_*H, E, a) for some map a (this is actually part of Grothendieck’s proof). Applying the result of the previous paragraph we obtain a coherent module F on X. I think it is pretty clear that the completion of F gives G as desired.

What I like about this argument is that it avoids dealing with extensions of formal modules. Note however, that one of the steps of the proof of formal glueing is an Ext computation, so we are not actually avoiding this issue altogether.

Update 10/12/12: Yesterday I finished adding this material to the stacks project. The proof in the projective case is short and sweet, see Section Tag 087V. The proof for the general case is in Section Tag 0886. The exposition avoids working with formal schemes (because it would take several hundred pages to introduce them) and instead consistently works with certain systems of coherent modules. This also has the advantage that the exact same arguments will work in the setting of algebraic spaces (and possibly algebraic stacks).

One version of Chow’s lemma is that given a finite type, separated morphism of Noetherian schemes X —> Y, there exists a blowing up X’ —> X with nowhere dense center such that X’ —> Y is quasi-projective.

Chow’s lemma also holds if you replace “schemes” with “algebraic spaces”; see Corollary 5.7.13 of the paper by Raynaud and Gruson. To parse this you have to know what it means for a morphism Z —> W of algebraic spaces to be quasi-projective.

No doubt Raynaud and Gruson have in mind a definition a la EGA: we say Z —> W is quasi-projective if it is representable, of finite type, and there exists an invertible sheaf L on Z such that for every S —> W, where S is an affine scheme, the pullback of L to the fibre product S x_W Z (this is a scheme) is an ample invertible sheaf.

I will show by a very simple example that you cannot use Knutson’s definition and expect Chow’s lemma to hold: Let’s say a morphism of algebraic spaces Z —> W is Knutson-quasi-projective if there exists a factorization Z —> P^n_W —> W where the first arrow is an immersion.

The example is the morphism X = A^1 —> Y = A^1/R where R = Δ ∐ {(t, -t) | t not zero}. In this case Chow’s lemma as formulated above just states that X —> Y is quasi-projective. On the other hand, my faithful readers will remember that in this post we showed that there cannot be an immersion X —> A^n_Y. The exact same argument shows there cannot be an immersion into P^n_Y (or you can easily show that if you have an immersion into P^n_Y, then you also have one into A^n_Y perhaps after a Zariski localization on Y).

The morphism X —> Y above can be “compactified” by embedding X = A^1 into the affine with 0 doubled which is finite etale over Y. So you can find an open immersion of X into an algebraic space finite over Y (this is a general property of quasi-finite separated morphisms). You just cannot find an immersion into the product of P^n and Y.

In the stacks project we don’t yet have defined the notions: relatively ample invertible sheaf, relatively very ample invertible sheaf, quasi-projective morphism, projective morphism for morphisms of algebraic spaces. I think a weaker version of Chow’s lemma that avoids introducing these notions, and is still is somewhat useful, is the following: given a finite type, separated morphism X —> Y with Y Noetherian (say) there exists a blowing up X’ —> X with nowhere dense center and an open immersion of X’ into an algebraic space representable and proper over Y. If Y is a scheme (which is the most important case in applications) you can then use Chow’s lemma for schemes to bootstrap to the statement above.

Knutson proves a version of Chow’s lemma with X’ —> Y Knutson-quasi-projective and with X’ –> X Knutson-projective and birational when both X and Y are separated. As mentioned in the other blog post, I think the problem pointed out above cannot happen if the base algebraic space Y is locally separated. Thus I think it may be possible to generalize Knutson’s version of Chow’s lemma to the case where Y is locally separated.

Surely, you’re not still reading this are you?

So I am gearing up to write a bit about Grothendieck’s existence theorem.

Let R be a Noetherian ring complete with respect to an ideal I. Let X be a proper scheme over R. Let O_n = O_X/I^nO_X. Consider an inverse system (F_n) of sheaves on X, such that F_n is a coherent O_n-module and such that the maps F_{n + 1} —> F_n induce isomorphisms F_n = F_{n + 1} ⊗_{O_{n + 1}} O_n. The statement of the theorem is that given any such system there exists a coherent O_X-module F such that F_n ≅ F/I^nF (compatible with transition maps and module structure).

Mike Artin told me Grothendieck was proud of this result.

Because it is all the rage, let’s try to construct F directly from the system via category theory. So consider the functor

G |—-> lim_n Hom_{O_X}(G, F_n)

on QCoh(O_X). Since QCoh(O_X) is a Grothendieck abelian category (see Akhil Mathew’s post) and since this functor transforms colimits into limits, we can apply the folklore result Lemma Tag 07D7. Thus there exists a quasi-coherent sheaf F such that

Hom_{O_X}(G, F) = lim_n Hom_{O_X}(G, F_n)

The existence of F comes for free. (A formula for F is F = Q(lim F_n) where Q is the coherator as in Lemma Tag 077P).

Of course, now the real problem is to show that F is coherent and that F/I^nF = F_n, and I don’t see how proving this is any easier than attacking the original problem. Do you?

In accordance with the promise I made at the end of this post today I FedExed the first T-shirt to the author of this comment pointing out a mistake in the formulation of a lemma. There are still 1 large and 3 medium T-shirts left…

September 25, 2012: Two more errors found. For the first, see this comment. For the second see this change.

Let X be a scheme and let U be an open subscheme. The scheme theoretic closure of U in X is the smallest closed subscheme Z of X such that j : U —> X factors through Z. We say that U is scheme theoretically dense in X if the scheme theoretic closure of U ∩ V in V equals V for every open V of X. See Definition Tag 01RB. Then U is scheme theoretically dense in X if and only if O_X —> j_*O_U to be injective, see Lemma Tag 01RE.

If X is locally Noetherian, then U is scheme theoretically dense in X if and only if U is dense in X and contains all embedded points of X (Lemma Tag 083P).

For general schemes the situation isn’t as nice. For example, there exists a scheme with 1 point but no associated point (Lemma Tag 05AI). As a replacement for associated points, we sometimes use weakly associated primes (Definition Tag 0547) and the corresponding notion for schemes. This notion agrees with associated point for locally Noetherian schemes. There are enough weakly associated points: if U contains all the weakly associated points, then U is scheme theoretically dense (result not yet in the stacks project). But in some sense there are too many: there is an example of a scheme theoretically dense open subscheme U of a scheme X which does not contain all weakly associated points of X (Section Tag 084J).

We have the following result from Raynaud-Gruson: If X —> Y is an etale morphism and x ∈ X with image y ∈ Y then x is a weakly associated point of X if and only if y is a weakly associated point of Y (Lemma Tag 05FP).

What about scheme theoretic density? Given an etale morphism of schemes g : X’ —> X and a scheme theoretically dense open U ⊂ X the inverse image g^{-1}U is a scheme theoretically dense in X’ (Lemma Tag 0832). This was added recently in order to show that scheme theoretic density defined as above (and as in EGA IV 11.10.2) makes sense in the setting of algebraic spaces.

If you have trouble falling asleep tonight, try proving some of the results above.