Want to take part in these discussions? Sign in if you have an account, or apply for one below
Vanilla 1.1.10 is a product of Lussumo. More Information: Documentation, Community Support.
am too tired to do it now, but on occasion of an MO discussion:
remind me to insert at smooth manifold the statement and proof that smooth manifolds are equivalently the locally representable sheaves on CartSp (more precisely: the 𝒢=CartSp-schemes).
Here is a question that I have come up against. Consider the subcategory of CartSp] with the same objects, but only those arrows that take 0 to 0. If you like this is a pointed version of CartSp, which I will denote CartSp0. CartSp0 obviously is not a site in the same way that CartSp is because every open set in a cover has to contain the origin, but if we consider presheaves on CartSp and CartSp0, what do we lose passing to the latter? Every map in CartSp is homotopic to one in CartSp0 (taking as the interval object the real line) by an affine transformation, but I don’t know if this is enough to ensure that the category of manifolds survives under the obvious map Pre(CartSp)→Pre(CartSp0).
I think the canonical functor SmoothManifolds→Sh(CartSp0) is still full and faithful, as the argument for the embedding into Sh(CartSp) still applies verbatim:
if two smooth maps differ then they differ on an open subset, hence on some plot;
if a function between manifolds preserves all plots (already all 1-dimensional plots) then it must be smooth.
For these arguments the nature of the morphisms in the site matters very little. All one needs is that a manifold does represent a presheaf on the site in the natural way.
Thanks, Urs. Of course, I’m interested in the case when CartSp is replaced by the analogous category generated by a line object in a Lawvere theory, as you have been doing/advocating. I just wanted to check I wasn’t being a complete idiot from the word go.
I have wondered myself every now and then about using CartSp0. I am still not really sure what to make of it, though, but I feel that’s just me being dense, there must be a good answer.
Notice for instance that in Moduli Problems and DG-Lie Algebras Jacob Lurie considers something like the derived algebraic analog: the site of small E∞-rings with an augmentation A→k to the ground field (Dually that’s a point in SpecA, of course), The idea there is precisely to study all the sheaves in the “big topos” only close to this global point (see top of page 10).
Of course, I’m interested in the case when CartSp is replaced by the analogous category generated by a line object in a Lawvere theory, as you have been doing/advocating.
Ah, interesting. What is it you are trying to do?
smooth manifolds are equivalently the locally representable sheaves on CartSp
Is there a similar statement for smooth manifolds with boundary ?
remind me to insert at smooth manifold the statement and proof that smooth manifolds are equivalently the locally representable sheaves on CartSp
I have added an (overly) detailed proof of this now. While overly detailed, I think I still need to fine-tune some “locally finitely presentable” and “paracompact”.
I also indicated various further statements along these lines: that manifolds are the locally ringed spaces locally isomorphic to objects in CartSp regarded naturally as locally ringed spaces (pretty obvious, I didn’t write out any genuine proof). And then I stated that in the full context of locally representable structured (infinity,1)-toposes over CartSp smooth manifolds are the 0-localic locally representably CartSp-structured (∞,1)-topos. But here I noticed that I need to do sme more thinking before nailing this down. I tried to indicate that this part remains incomplete.
I’ll let you know in a little bit… ;-)
any comment on 7 ? Urs ?
@Zoran,
The recent paper on the arXiv, without realising it, treats manifolds (finite and infinite dimensional) with corners as sheaves on a site of certain subspaces of euclidean spaces. The author defines ℝnk={x∈ℝn|xi≥0,i=1,…,k}, the considers the category with objects ℝnk for k=0,…,n−1 and n=1,2,… and smooth maps between them. (A map is declared smooth at the boundary if it extends to a smooth map on a neighbourhood of the boundary). I haven’t delved into the details, but I would not be surprised if this answers your question.
any comment on 7 ? Urs ?
Maybe I haven’t thought about it enough, but my inclination would be to say: obviously this works for manifolds with corners, too.
As I think the argument that i wrote up shows, saying that a sheaf is “locally representable” is really equivalent to saying that it is the sheaf obtained by gluing a bunch of representables together, where all gluing is by the maps that exist in the site.
So if you start with a site of Cartesian spaces with boundary and corner (Cartesian half-spaces, quarter spaces, etc) and take care to define the morphisms between them suitably, then I think it will hold that manifolds with boundary and corners are precisely the locally representables sheaves on that site.
David (11):
The recent paper on the arXiv, without realising it,
Interesting choice of words. I had a quick look at that paper (wondered whether or not there was something to add to Comparative Smootheology) and was at a loss to see what new thing it introduced. As a “category of generalised smooth spaces” it sits between Chen spaces and diffeological spaces and can be treated in exactly the same fashion as those two - at least, that’s my interpretation, the defining condition certainly looks like the sheaf condition but I’m no expert on sheaves. What was odd about the paper was that there was no mention of Chen spaces nor of diffeological spaces, though the author was clearly aware of some of Kriegl and Michor’s work.
Urs: so the term “locally representable” is site-specific, then. I see. Without having thought any deeper, it does feel as though paracompactness and Hausdorff are going to bite you on this one as those are conditions imposed “from outside” so it could be hard to generate them from the inside.
@Andrew
The defining condition (definition 2.4) is exactly that of a (separated) (pre)sheaf, even down the use of the prefix/adjective pre-/separated. Lemma 2.6 describing sheafification is also exactly on the money. Ditto internal hom (example 2.9).
Interesting choice of words.
Well, on second thoughts that is a bit diplomatic. Given the pre-/separated inf. dim. manifold terminology, I don’t think the author hasn’t heard of a sheaf before. There is no mention of underlying sets or local representability for these ’manifolds’, so they aren’t even generalised schemes.
Urs: so the term “locally representable” is site-specific, then. I see.
Oh, yes, certainly. The generalized notion of scheme is something available as soon as we fix a “pre-geometry” 𝒢: it’s those objects in Sh(𝒢) that are locally equivalent to objects in 𝒢.
So in this context of geometry we do not consider the topos Sh(𝒢) on its own, but always equip it with its structure of a “𝒢-ringed topos” given by a product-preserving (= algebra) and cover-preserving (= local) functor 𝒪:𝒢→Sh(𝒢), namely the universal such functor, which is the Yoneda embedding.
(added in edit to clarify: this is in response to David’s previous post)
Ah, I hadn’t spotted the lack of underlying set. So it’s just sheaves, then.
The smallness condition seems a little odd. Surely sheaves on “manifolds with corners” is the same as sheaves on “Quadrants”, which is obviously small. So is there really any need to restrict to the given small subcategory of manifolds with corners? (Thm2.3 on p3 for the construction).
I don’t know how to put my next question diplomatically … is there anything in this that Urs and Konrad don’t know?
@Andrew I suspect not. As far as looking at the smallness condition - to us it may seem obvious, but someone working from a(n old) text on sheaves would strive to ensure that their category of presheaves exists (maybe there’s something in SGA about needing sites to be small) and may feel they needed to convince people not au fait with category theory that all is ok. Definitely the description of the site has a very material (as opposed to structural) feel to it, with talking about embeddings to ensure that the site has coproducts.
I have worked on smooth manifold a bit more, in reaction to the above discussion:
expanded the section The geometry CartSp with more of the general story;
removed in the section Smooth manifolds as locally representable sheaves the paracompact red herring.
Does anyone feel like adding some standard references?
Thank you Urs and David. I do not understand how one gets the Hausdorff condition. I mean sheaf, that is gluing is about passing from local to global. Perfect good charts can glue to nonHausdorff manifold. How can this be prevented by saying that something is a sheaf on cartesian site, locally representable etc. (in algebraic geometry we say locally affine) ?
saying that a sheaf is “locally representable” is really equivalent to saying that it is the sheaf obtained by gluing a bunch of representables together, where all gluing is by the maps that exist in the site.
I don’t quite understand the notion of “locally representable” and maybe this will help. But if you want to obtain manifolds and only manifolds in this way, doesn’t there also have to be some condition about gluing by open maps? The site CartSp contains non-open maps, but if I glue two cartesian spaces along a non-open subspace, I don’t get a manifold, do I?
right. I think I need to add that the maps Ui→X are monos.
More later, need to run now.
Say X∈Sh(CartSp) is locally representable if there is a family of monomorphisms {Ui↪X} with Ui∈CartSp↪Sh(CartSp) such that ∐iUi→X is a regular epi, hence so that
∐i,jUi×XUj→→∐iUi→Xis a coequalizer in Sh(CartSp). It follows that such an X is a diffeological space, whose underlying set is the corresponding coequalizer of the underlying sets.
So each pullback diagram
Ui∩Uj→Uj↓↓Ui→Xmay be computed in diffeological spaces inside Sh(CartSp), where it exhibits Ui×XUj as the intersection of the open Ui,Uj↪X.
So how do you rule out non-Hausdorff manifolds ?
So how do you rule out non-Hausdorff manifolds ?
Let X be as in #23. Let x,y∈X be two points. The pojnt x is in the image of some Ui, the point y in the image of some Uj. If Ui∩UJ is empty, they serve as open subsets that separate x and y. if not, either both x and y sit in the intersection Ui∩Uj, or at least one does not. Say x does not, then we can find an open neighbourhood for x in Ui that does not intersect Ui∩Uj. So again they are separated. If both points are in the interseciton Ui∩Uj, use that this is an open subsets of Ui and hence is Hausdorff.
How does just assuming that each Ui↪X is monic imply that they are open subsets? What if I take two copies of ℝ and glue them together at a point? Then don’t these two copies of ℝ satisfy your conditions, but aren’t open subsets of the glued space?
Say x does not, then we can find an open neighbourhood for x in Ui that does not intersect Ui∩Uj.
Why ? What if x and y are both in the closure of Ui∩Uj while not in Ui∩Uj ? I mean I do not see how this proof works for the usual examples of nonHausdorff manifolds.
Yeah, I am screwing it up. I guess you are right and I will have to require that X is concrete and the Ui→X are open embeddings.
Let’s see, I want a way to say this that does not make use of the by-hand verifiction that a concrete sheaf has canonically an underlying topological space. A more abstract way. Hm…
Edit: this is to address the concern about Hausdorffness, not local representability
Something like a condition on the diagonal X→X×X, I reckon. cf algebraic spaces. Maybe require the diagonal is representable and closed….
I should maybe stop posting to this here while busy with something else. But: what are the regular monos?
Regarding local representability: how about asking that there is a cover ∐Ui→X which is representable (as in, given any map f:ℝn→X, the pullback of the cover along f yields a coproduct of representables)?
the pullback of the cover along f yields a coproduct of representables)?
This cannot quite work: the intersection of two contractible opens is in general not contractible.
But something that sounds similar is that one could ask that each Ui×XUj is either empty representable, which is the beginning of saying that {Ui→X} is a good open cover. But that does not seem to get around having to invoke the topology on X to get a manifold.
When schemes are defined as locally representable sheaves on the Zariski site, there is also the explicit condition that the cover is open (as on slide 3 here). That’s why I said in #1
smooth manifolds are equivalently the locally representable sheaves on CartSp (more precisely: the 𝒢=CartSp-schemes).
only to trick myself into ignoring the “more precisely”-clause later on ;-)
It seems unlikely to me that we’ll be able to avoid equipping the site CartSp with some extra structure, specifying which maps are “open,” if we want to be able to define a notion which reduces to manifolds.
There must be a way, as in def. 2.3.9 of Structured Spaces:
Pass to the little (∞-)topos 𝒳 over X, ask for an effective epi ∐iUi→* in the topos, and ask that each 𝒳/Ui is equivalent, as a ringed topos, to a representable.
If I understand well then the fact that we ask the covering to be a morphism in the little topos makes it étale, so that’s where the topology is taken care of.
@Urs 33
since we are working with a coverage on CartSp (good open covers), then perhaps we don’t ask for the pullback ℝn×X∐Ui, but a weak pullback, as in the definition of coverage.
We can restate a map (of topological spaces or manifolds) X→Y being étale as being existence of covers ∐Ui=:U→X,U→Y making the obvious triangle commute (or better, draw it as a square with an equality on one side). Surely we can state the topos theoretic condition as some sort of existence theorem about covers etc.?
Urs, I wasn’t quite able to convince myself some months back that the regular monos in the category of smooth Hausdorff manifolds are precisely the closed embeddings, but I still believe that’s what they’ll probably wind up being.
There must be a way, as in def. 2.3.9 of Structured Spaces
But doesn’t the “admissibility structure” there essentially encode the same data about which maps are “open”?
But doesn’t the “admissibility structure” there essentially encode the same data about which maps are “open”?
Yes. (It is effectively the coverage data, I think.)
I had thought this does not crucially play a role in def 2.3.9, because it doesn’t in saying that ∐iUi→*𝒳 is a regular epi and not in saying that we have equivalences of structured topoi SpecUi≃𝒳/Ui (I think, because an equivalence of structure sheaves just as left exact functors should automatically be a morphism that respects the admissibility structure and hence be an equivalence as structure sheaves). Clearly I should think more about this.
Maybe the issue is more vivid in terms of the classifying topos description: 𝒢-structures are classified by geometric morphisms into Sh(𝒢). That does not involve the admissibility structure. Instead, the admissibility structure on 𝒢 is equivalently a natural factorization system on Topos(−,Sh(𝒢)), hence affects only the morphisms. The morphisms of 𝒢-structured toposes are those landing in the right half of this factorization system. But every equivalence is in there, so it does not affect the notion of equivalence. I’d think.
Urs, do you have a way out of the “gap” in your Hausdorffness proof I seem to point out in 27 ? I am still not seeing it.
Zoran, I think you are right, I was wrong. I need to come back this whole issue here and do things right. I should maybe not have started this without the time to concentrate on it. On the other hand, except that I am making a fool of myself, I am enjoying that we are discussing this! :-)
Here is another thought that I won’t have time to look at in detail right now, but which I’ll mention anyway:
as a slight variant of the definition in Structured Spaces with more an emphasis on the big topos Sh(CartSp): maybe it makes sense to say that for X∈Sh(CartSp) and Ui∈CartSp↪Sh(CartSp) we have an effective epi ∐iUi→X plus the condition that the corresponding morphisms in Topos/Sh(CartSp)
Sh(CartSp)/Ui→Sh(CartSp)/X↘⇙↙Sh(CartSp)sit in the right half of a suitable factorization system. But I still need to understand better why the analogous condition is not needed in def. 2.3.9 of Structured Spaces .
It is effectively the coverage data, I think
I feel like the “class of open maps” (a term which I much prefer to “admissibility structure,” ugh) is slightly more than giving a coverage (or pretopology) which generates the topology — it’s more like a class of maps which could appear in a generating covering family, but which don’t necessarily do so. There is an abstract notion of “class of open maps” which has been studied by I think Joyal and Moerdijk, among others.
Going back to the points raised in #2 and #3, I read in MacLane&Moerdijk just recently an exercise that says that the usual open cover topology on the category of manifolds (presumably separable ones, and also finite dimensional) is the smallest one containing the covering families
U1={(−∞,1),(−1,∞)}
U2={(−n,n)|n≥1}
Now these are covering families of ℝ in the category CartSp0 of cartesian spaces and maps preserving zero. They generate covering families on all ℝn by taking products of these.
I’m not sure what this means, I’m just putting ’on paper’ as it were (and it’s late here, for me). Any thoughts?
As the reference is only alluded above here it is (about the “class of open maps”) for the record, in full:
In the cold light of day, perhaps the key point alluded to in the exercise I quote in #44 is that the open covers on manifolds can be generated by an explicitly given class of open covers. This might be helpful to pin down what it means for a sheaf on CartSp to be locally representable.
Added link to category SmoothManifolds.
have brushed-up the referencing of and to Lawvere 1973 and Zack 2014.
1 to 50 of 50