-SEMI-SHEAVES, DISTRUCTURES AND
SCHWARTZ DISTRIBUTIONS
by Andrée
C. EHRESMANN
Université de Picardie Jules Verne, Amiens
ehres@u-picardie.fr
2008
0. Introduction
In the early 50's Charles Ehresmann is
interested in mathematical species of structures (a concept also discussed
during the Bourbaki meetings to which he participated); in particular he gives
the definition of a species of local
structures in connection with his work on locally trivial fibred spaces and
foliated manifolds ([7], Part I). To such a species he associates the pseudogroup
of its local automorphisms and, as soon as 1952, remarks that it is a groupoid
in the sense of Brandt, which plays the same role as an operator group on a
homogeneous space. But it is only in his important 1957 paper « Gattungen
von lokalen Strukturen » [10] that he did make the link with the categories
Eilenberg and
In this paper, he introduces the
notion of a species of structures over a category C and of the associated of action of the category C on the set S of fibers of this species, and he
shows that it is equivalent to a functor F from C to the category Set sending
an object e of C on the fiber Se (hence, in modern terms, to
a presheaf on Cop). He also gives the construction (often
erroneously attributed to Grothendieck) of the associated discrete fibration p:
C*S → C, which
it calls "hypermorphism category", and characterizes the functors so
obtained.
This very substantial 1957 paper has been at
the root of many later works, all the more since Charles also
'internalizes" the above notions in the category of complete sup-lattices
by defining what he calls local
categories, their actions and the corresponding local species of structures. Moreover he introduces two main
constructions: 1. given a species of structures on C and a category C'
containing C, he extends the action of C to C' by the "enlargement
process"; and, in the local case, he gives a "completion
process" of a local species which, in modern terms, corresponds to the
formation of the associated sheaf.
This paper was followed by his 1959
paper « Catégories topologiques, catégories différentiables » [11]
where he introduces internal categories, and their actions, in the category Top
of topological spaces and Diff or differentiable manifolds (well before the
notion of an internal category had been considered). There he characterizes the
locally trivial groupoids and proves
that their theory is equivalent to the theory of principal fibred spaces, while
their actions give the associated locally trivial fibred spaces. In the brief
abstract of the lecture he gave at the Amiens Conference in 1975 7, Part I), he
claims that differential geometry reduces to the study of the differentiable
sub-categories of the category of jets, and of their actions, and this subsumes
his approach to his foundation of differential geometry,
In the first times of our relation,
Charles was correcting the proofs of his 1957 paper and writing the 1959 paper
(which I typed for him). These papers were my initiation to category theory
and, for a long time, remained all what I knew of this theory. I was then
working in functional analysis (under the direction of Choquet) and was
beginning to write my thesis (from 1959 to 1961) [1] where I propose a notion
of distructure to unify different
notions of "generalized functions" and, in particular, to extend the
(vector-valued) Schwartz distributions to infinite dimensional vector spaces. Thinking
of the local characterization of distributions as formal derivatives of
continuous functions, which are then glued together, I was naturally led to compare
this with the enlargement process of a category action, and then its
completion.
However Charles' definitions were
not well adapted for this. Indeed, in the enlargement process, there is
initially an action of a category C on a set S which is extended to a larger
category; but the monoid A of differential operators acts only partially on
continuous functions, and it is this partial action which has to be extended;
moreover the fibers are not sets but locally convex spaces whose structure has
to be preserved. And it is the same for the completion process in which not
only the locally convex structure of the fibers has to be preserved, but also
the action of A on these fibers. So I had to modify and extend Charles' constructions
to the case of a 'partial action' (here I'll use the term of 'semi-action') of
a category, which is at the root of the theory of distructures given in the second part of my thesis.
The concept of a partial action is
also used in my later works, first in my 1963-66 study of control and optimization
problems [3] in the frame of control systems which use partial topological or
differentiable actions; then in the hierarchical evolutive systems [8-9] which
I introduced (with J.-P. Vanbremeersch) to model living systems which
correspond to enriched partial actions, the fibers being hierarchical
categories (a hierarchical category being a category in which the objects are
partitioned into levels, with an object of a level being the colimit of at
least one diagram with values in the strictly lower levels).
However In my former papers, the
notion of an enriched or internalized partial action is applied in various
particular cases, not in a general setting. Later on, to unify these various
examples, Charles has introduced the notions of a system of structures and of a dominated
system of structures [13], and in a
Note of 1966 [14] he gives existence theorems for its extension into a dominated species of structures (in
other words, extension of the partial action into a global one), but under strong conditions on the category
in which the 'domination' rests.
Here I propose a general setting to
study and extend enriched partial actions of a category, and define distructures
which are generated by presheaves of such partial actions. We know that an
action of a category C is equivalent to a presheaf on Cop; likewise
a semi-action (which is a partial
action with a transitivity action stronger than in the Ehresmann's systems of
structures) can be defined by what I call a semi-sheaf.
The Associated Presheaf Theorem
asserts that a (K, M)-semi-sheaf, where K is a category with colimits and M an
adequate class of monomorphisms, can be extended into a presheaf with values in
K, and this presheaf is explicitly constructed. A similar theorem and
construction are valid for internal semi-actions and will be given in a later
paper. Then a (K, M)-generator of distructures
on H is defined as a presheaf on H with values on the category of (K,
M)-semi-sheaves; it generates a K-distructures presheaf and, if H is equipped
with a Grothendieck topology, a K-distructures sheaf. As an application, given
two locally convex topological spaces E and E', I construct the sheaf of E'-valued
distributions on E; if E is finite dimensional these distributions correspond
to Schwartz E'-valued distributions.
1. Semi-actions of categories and semi-sheaves of sets
To
simplify, I'll consider only semi-actions of categories and semi-sheaves on a
category, though everything would extend if the categories are replaced by
semi-categories in the sense of Schröder & Herrlich [18]. (I recall that a
semi-category satisfies the same axioms as a category, except that the
composite of two consecutive arrows is not always defined, so that the associativity
becomes: if the composites yx and zy are defined, then (zy)x
is defined if and only if z(yx) is defined, and then both are equal.)
a) Semi-sheaves
of sets
We know
that the concept of an action of a category is in 1-1 correspondence with that
of a presheaf. There will be an analog relation between semi-actions and
semi-sheaves. Depending on what we want to do, one or the other notion will be
more practical: semi-sheaves are easily 'enriched', while semi-actions are
easily 'internalized'. To make the definition more concrete, we'll begin with
the notion of a semi-sheaf of sets.
Definition. A semi-sheaf (of sets) on the category C
is defined by a map F from C to Set satisfying the
following conditions:
·
For
each object e of C, F(e) is the
identity of a set Fe,
called the fiber of F on e.
·
For
x: e' → e, F(x)
is a map from a subset Fx
of Fe to Fe'. The image F(x)(s) of an element s of Fx will
be denoted by sx and called the composite of (x, s).
·
(Transitivity) If the composites xx' in C and sx are defined, then the composite (sx)x'
is defined if and only if s(xx') is defined, and then we have (sx)x'
= s(xx').
Let S be the disjoint union of the
fibers Fe and p0 the map from S to C
sending an element of Fe
to e. We denote by C*S the set
of pairs (x, s) such that s is in Fx and by k' the map from C*S to S associating to
(x, s) its composite sx. Then we say that (C, S, p0, k') defines a (right) semi-action
of C on S. It is an action of C
on S if the fibers are non void and if Fx = Fe
for each x: e → e'.
Remark. This can be compared to the notion
of a system of structures (C.
Ehresmann [13]) on the opposite category Cop. However the
transitivity axiom we impose is more restrictive and more symmetrical since, in
a system of structures, the composite s(xx') is defined
as soon as (sx)x' is defined and nothing is said in the other way (cf. Section 3)..
Example. The definition extends to the case where C is
a semi-category. In this case, we may define a semi-sheaf F on C whose fiber on
e is the set of arrows whose domain
is e and Fx, for an x: e' → e in C is the set of z such that zx is defined
in C, the composite being zx. The
transitivity is then equivalent to the associativity axiom of a semi-category.
The next proposition will be useful
to define the transitivity in the 'enriched' case where Set is replaced by
another category K (cf. next section).
Proposition. The
transitivity means that, if xx' is defined in C, the inverse image of Fx by F(x) is equal to the intersection
Ixx' of Fx and Fxx'.
Proof. If s is in Ixx'
the composites sx et s(xx') are defined, and the transitivity
implies than (sx)x' must be defined, so that sx
= F(x)(s) must be in Fx’
hence s is in the inverse image of Fx by F(x). Conversely, if (sx)x' is defined, s(xx')
must be defined, so that s is in the
intersection Ixx' of
Fx and Fxx'.
b) Semi-fibration associated to a
semi-sheaf
To a presheaf on C, or rather to the
corresponding action of C on the set S of its fibers, is naturally associated a
discrete fibration over C. This extends to the case of a semi-sheaf F and to
the semi-action (C, S, p0,
k') it defines. Indeed, we obtain a
category by equipping the set C*S of the composition:
(x, s)(x', s') = (xx', s) if and only if xx'
is defined in C, s’ = sx and s' is in Fx'.
The transitivity ensures that this law
defines a category whose objects are the pairs (e, s) such that s is in Fe. The map sending (x, s) to x defines a functor p
from this category to C, which satisfies the condition:
(SF) Each morphism is cartesian and there is at
most one morphism with a given codomain and image by p.
Let us recall that h is cartesian if, for any morphism h" having the same codomain and
such that p(h') = p(h)x', there exists one and only
one h' such that h" = hh' and p(h') = x'.
(In Charles' terminology [13], a cartesian morphism is called a p-injection, and SF implies the p is 'well faithful'.)
Definition. A functor
satisfying the condition (SF) is called a discrete
semi-fibration over C. The
functor p: C*S → C is called the semi-fibration associated to F (or to
the semi-action defined by F).
Proposition. There is a 1-1 correspondence between the
semi-sheaves on C and the discrete semi-fibrations over C.
Proof. If p:
H → C is a
semi-fibration, it is associated to the semi-sheaf constructed as follows: the
fiber Fe is the inverse image
of e by p; for x: e'→ e, Fx is the set of objects s of Fe such
that there exists a (necessarily unique) h
with s as its codomain and x as its image by p; and F(x) sends s to the domain sx of h. For the
transitivity, if sx and xx' are defined, then s(xx') is defined iff there exists h" with codomain s
and image xx'; as h is cartesian it means that there
exists h' with codomain sx and image x', hence that (sx)x' is defined.
c) The category of semi-sheaves
The semi-sheaves are the objects of
the category SS whose morphisms are defined as follows: if F is a semi-sheaf on
C and F* a semi-sheaf on C*, a morphism from F to F* is a pair (λ, µ) of a functor λ from
C to C* and of a map µ from C to Set sending x to
the map μ(x) from Fx to F*λ(x) such that:
µ(x)(s)λ(x) = µ(e')(sx)
for x: e' → e and
s in Fx.
In terms
of the associated semi-fibrations p
and p*, it corresponds to a pair of
functors (λ: C →
C', Λ: C*S →
C**S*) such that p*Λ = λp.
The
category SS admits the category PS of presheaves as a full sub-category. In the next section, we'll show that
PS is a reflective sub-category of SS as a corollary of the same assertion for
'enriched' semi-sheaves and we'll give an explicit construction of the presheaf
associated to a semi-sheaf. For the semi-sheaves of sets considered in this
section, this result can also be deduced from the theorem of Street and Walters
[21] which decomposes any functor p:
H → C as the
composite p'° j of a
fibration p' on C and a final functor
j. If p is the semi-fibration associated to a semi-sheaf F, then p' is the fibration corresponding to the
presheaf F' associated to F. Let us recall that the union S' of the fibers of p' is constructed by these authors as
the colimit of the presheaf A on H defined as follows:
A(s) = HomC(p(s), -) and A(h), for h: s → s', is
the map y’ |→ y’p(h) from A(s’)
to A(s).
It follows that the
fiber F'e on e is the class of the connected
components of the comma category p↓e.
2. (K, M)-semi-sheaves and their associated
K-presheaves
In the
applications to analysis or modeling that I have developed the fibers of the semi-sheaves
are enriched (or 'dominated' in Charles' terminology) by supplementary
structures: locally convex spaces, categories, or even (in the case of
distructures) by enriched semi-sheaves. The concept of (K, M)-semi-sheaf will
encompass these different cases. In the Set case studied above we had the
choice between 3 equivalent definitions: semi-sheaves, semi-actions or
semi-fibrations, However to enrich the fibers, we'll have to take the
semi-sheaf approach.
If we try
to transpose the definition of a Set-semi-sheaf F on C to a category K, by what
will we replace the inclusion of Fx in the fiber Fe for a morphism x:
e' → e of C? If K is a concrete category it seems natural to ask that
F(x) be a morphism from a
sub-structure Fx of Fe to Fe' (for instance a functor from a sub-category if K is
Cat). However this will not be adapted for the application to distributions
where the fibers will be a space of continuous maps equipped with the
compact-open topology and the semi-action is that of the differential operators
which are not continuous for the topology it induces on the differentiable
functions. On the other hand the definition would be too lax if we took for Fx any
sub-object of Fe. Thus we
will select a particular class of sub-objects. Moreover to transpose the
transitivity axiom, we need pullbacks (to define the analog of the intersection
Ixx' of Fx and Fxx'). Thus we are led to the following setting.
a) The
category of (K, M)-semi-sheaves
Let K be
a category (generally it will admit pullbacks), and M an ss-admissible class of monomorphisms of K, meaning that it
satisfies the following conditions:
• M
contains the identities and there is at most one element of M between two
objects.
• M is
stable by pullbacks: if a pair (m, f)
of an element m of M and a morphism f admits a pullback I, its projection m' opposite to m is in M:
Examples. If there
is a faithful functor p: K →
Set, then M could be the class of the cartesian morphisms
which define a p-sub-structure (i.e.,
whose image by p is an inclusion),
for instance K may be the category Cat of categories and M the class of inclusions
of a sub-category. To define distributions [2,3], I
took for K the category Lcs of locally convex topological vector spaces and for
M the class of inclusion of a vector sub-space, but with a finer locally convex
topology. In my study of Control Systems [3], K was the category Top of
topological spaces, and M the class of inclusions of an open sub-space. These
examples are particular cases of an ordered category K (i.e., internal to an
adequate category of posets) which is completely regular, with M being the
class of morphisms ('pseudoproducts') s's:
s → s' such that s < s's
< s'; this unifying setting is used in Charles' 1966 Note [14] to define
dominated systems of structures (cf. Section 3).
Definition. A (K, M)-semi-sheaf
on C is a map F from C to K satisfying the following conditions:
·
F(e) for an object e of C is the identity of an object Fe of K, called the fiber
of F on e.
·
For
each x: e' → e in C, the morphism F(x): Fx
→ Fe', called composition by x has for its domain Fx an object such that there
exists a (necessarily unique) mx : Fx → Fe in M.
·
(Transitivity)
If xx' is defined in C, there exists
a pullback Ixx' of (F(x), mx’) which is also the intersection
of (mx, mxx') (hence
their pullback); moreover we have :
F(x’)pxx' = F(xx')p'xx'
where pxx'
and p'xx' are the
projections from Ixx' respectively
to Fx and Fxx'. (Since M is stable by
pullbacks and has only one element between two objects, the other projection m' is in M and is the same for the two
pullbacks.)
A (K, M)-semi-sheaf F such that Fx = Fe
for each x with codomain e is just a functor from Cop
to K, hence a K-presheaf, and M has no more rôle. Indeed, Ixx' is equal to Fe
since we have Fxx' = Fe = Fx so that pxx
= F(x), and the transitivity condition
becomes: F(x’)F(x) = F(xx').
The (K, M)-semi-sheaves are the
objects of the category SS(K, M) whose morphisms are
defined as follows. If F is (K, M)-semi-sheaf on C and F* a
(K, M)-semi-sheaf on C*, we define a morphism
(λ, µ) : F → F' by:
·
A map µ from C to K which sends x: e' → e to a morphism µ(x): Fx
→ F*λ(x) so that:
m*λ(x) µ(x) = µ(e’) mx and
F*(λ(x))µ(x) = µ(e’)F(x)
There is a functor Base: SS(K, M)
→ Cat
mapping (λ, μ) : F → F* on λ: C → C*.
.
b) The Associated Presheaf Theorem
The category PS(K)
of K-presheaves is identified to a full sub-category of the category SS(K, M)
of (K, M)-semi-sheaves. We are going to prove that it is a reflective
sub-category. It means that a (K, M)-semi-sheaf F on C can be 'universally'
extended into a K-presheaf F' on C; equivalently, it will extend a semi-action
of C into an action of C. This construction was essential in my thesis to
extend differential operators to continuous functions.
Theorem (K-presheaf associated to a (K, M)-semi-sheaf). If K admits pullbacks and colimits, the category SS(K, M)
admits the category PS(K) as a reflective sub-category, and the
reflection F' of a (K, M)-semi-sheaf F is explicitly constructed.
Proof. Let F be a (K, M)-semi-sheaf on C.
• We first define the fiber F'ê on ê of the associated K-presheaf F'. It
will be the colimit of a functor Gê: Sub(ê↓C) → K where Sub(ê↓C) is the subdivision category
of the category ê↓C of objects
under ê. The category ê↓C has for objects the morphisms y with domain ê and for morphisms the commutative triangles t = y,x,y';
its subdivision category has for objects all the elements of ê↓C, and it has one morphism it: t → y and one
morphism ct: t → y';
roughly each morphism t of ê↓C is replaced by a span (it, ct). The
functor Gê is
defined by:
Gê(it) = mx: Fx → Fe and Gê(ct) = F(x).
• Let z: ê' → ê
be a morphism in C. It defines a functor z↓C: ê↓C → ê'↓C
sending t = y,x,y'
onto tz = yz,x,y'z, hence a functor
γz: Sub(ê↓C) → Sub(ê'↓C) which sends it and ct
on itz and ctz. Since
Gê'°γz = Gê there exists a morphism F'(z) from the colimit F'ê of Gê to the colimit F'ê'' of Gê'' such that F'(z)jy = jyz for each y: ê → e, where jy
is the injection of Fe = Gê(y) into the colimit F'ê
of Gê and jyz the injection of F(e) = Gê'(yz) into the colimit F'ê' of Gê'. We have
thus constructed the K-presheaf F'.
• We
define a morphism (IdC, ρ): F → F' by taking for ρ(ê) the
injection jê of F(ê) into the colimit of Gê and for ρ(z)
the composite ρ(ê)mz.
Indeed, for z: ê' → ê we have
F'(z)ρ(z) = F'(z)ρ(ê)mz = F'(z)jêmz
= jzmz = jê'F(z) = ρ(ê')F(z).
• It
remains to prove that (IdC, ρ) defines F' as a reflection of F
into the sub-category PS(K) of K-presheaves. Indeed,
let (λ, μ): F → F* be a morphism
from F to a K-presheaf F* on a category C*. We must find a morphism (λ', μ'):
F' → F* whose composite with (IdC, ρ) gives (λ, μ). For λ' we
take λ itself. For an object ê
of C, we have a natural transformation n
from Gê to F*ê such that:
n(y)
= F*(λ(y))μ(e)
and n(t)
= F*(λ(y))μ(e)mx for y
= xy' and t = y,x,y'
It is a
natural transformation since
n(y)mx = n(t) = F*(λ(y))μ(e)mx
= F*(λ(y'))F*(λ(x))μ(x) = F*(λ(y'))μ(e')F(x)
= n(y')F(x)
This
natural transformation glues into the morphism μ'(ê) from the colimit F'ê
of Gê to F*λ(ê)
such that n(y) = μ'(ê)jy.
In particular, for y = ê, we have
μ(ê) = n(ê) = μ'(ê)jê =
μ'(ê)ρ(ê).
It is a
natural transformation since
n(y)mx = n(t) = F*(λ(y))μ(e)mx
= F*(λ(y'))F*(λ(x))μ(x) = F*(λ(y'))μ(e')F(x)
= n(y')F(x)
Let us prove that μ' defines a natural
transformation from F' to F*. For z: ê'
→ ê and each y: ê →
e, we have
μ'(ê')F'(z)jy
= μ'(ê')jyz (by definition of F'(z))
= F*(λ(yz))μ(e)
(by definition of μ'(ê')
)
= F*(λ(z))F*(λ(y))μ(e) = F*(λ(z))μ'(ê)jy (by definition of μ'(ê) ).
whence μ'(ê')F'(z) = F*(λ(z))μ'(ê) since the jy are the injections toward the colimit F'ê. Thus (λ, μ')
defines a morphism from F' to F* such that (λ, μ) = (λ, μ')(IdC,
ρ). This morphism is unique because, if (λ, μ") was
another one, for each ê the equality
μ'(ê)ρ(ê) = μ"(ê)ρ(ê) implies:
μ'(ê)jy
= μ'(ê)F'(y)ρ(e) = F*(λ(y))μ'(e)ρ(e) = F*(λ(y))μ"(e)ρ(e) = μ"(ê)F'(y)ρ(e) =
μ"(ê)jy.
for each y, and thus μ'(ê) = μ"(ê).
This completes the proof that (IdC, ρ): F → F' is a
reflector into the sub-category of K-presheaves.
Corollary. The presheaf F' associated to a semi-sheaf of sets F has for fiber on ê the quotient of the
disjoint union Σê of
the family (Fe)y:ê→e by the equivalence relation Rê generated by the relations: (y',
sx) ~ (xy', s) if s is in Fx.
Proof. A semi-sheaf of
sets F is a (Set, M)-semi-sheaf where M is the class of inclusions of a subset into
a set. By the above construction, the fiber F'ê of the associated presheaf of sets is the colimit of
the functor Gê. In Set
this colimit can be constructed as follows: we first take the set Σê consisting of all the pairs
(y, s) where y: ê → e and s
is in Fe. For each triangle
t = y,x,y' we
must have jymx = jy'F(x) so that, for s in Fx we
must identify (y, s) and (y', sx), whence the relation Rê since y = xy'.
3. Comparison with dominated systems of
structures
Since a
semi-sheaf is a particular system of structures, the Associated presheaf
Theorem of Section 2 can be compared to the "Dominated Expansion Theorem"
for systems of structures dominated in a concrete ordered category (K, <)
given by Charles Ehresmann [14]. For that, we must transpose his notion of a dominated
system of structures in our setting.
a) (K, M)-dominated
systems of structures
Given a
category admitting pullbacks and an ss-admissible class M of monomorphisms of
K, we define a (K, M)-dominated
system of structures in the same way as a (K, M)-semi-sheaf, except
that the transitivity is replaced by:
If xx'
is defined in C, there exists a pullback Ixx'
of (F(x), mx’) and a p'xx': Ixx' → Fxx 'in M such that F(x’)pxx'
= F(xx')p'xx' where pxx'
is the projection from Ixx'
to Fx and Fxx'.
The
difference with a semi-sheaf is that here we do not impose that Ixx' be the intersection of Fx and Fxx'. As this stronger condition is not used in the proof
of the Associated Presheaf Theorem, this theorem extends to (K, M)-dominated
systems of structures: Whence
Theorem. If K admits colimits, PS(K) is also a reflective sub-category of the category Dss(K, M) of (K, M)-dominated systems of structures (the
morphisms being defined as for semi-sheaves), and the associated K-presheaf is constructed as for semi-sheaves.
This
theorem is much stronger than the "Dominated Expansion Theorem" of
Charles which supposes that K is a concrete category whose forgetful functor to
Set has strong enough properties. In a comment that I have written to
supplement Charles' Note in "Charles Ehresmann : Oeuvres
complètes et commentées" [7] (Comment 468-1, Part III), I prove the
existence of the associated K-presheaf (but not its construction) under the
above conditions. For this, I give another characterization of a (K, M)-dominated
system of structures and of its associated K-presheaf, when the ss-admissible
class M is moreover stable by composition (so that it is a sub-category of K),
and I deduce the existence of the associated presheaf from an existence theorem
on lax colimits. As it allows another view on these problems, I am going to recall
this characterization.
To (K, M)
I associate the following 2-sub-category KM of the bicategory of spans of K
whose 1-morphisms are the spans (f, m)
where m is in M, their
composition being done via pullbacks. As we have imposed the stability of M by
pullbacks and that there exists at most one element of M between 2 objects, the
identities are the sole invertible elements in M (otherwise the inverse would
also be in M), so that we obtain a usual category for the composition:
(f', m')(f, m) = (f'f", m"m) if m' and f have the same codomain,
where f", m" are the projections of
the pullback of (m', f). A 2-cell m*: (f, m) → (f1, m1) of KM is defined by an m* in
M such that m = m1m* and f = f1m*.
The
2-category KM is the 2-category of objects for a 3-fold category KM* with K as
the category of objects for the 3rd composition; 3-blocks are
degenerated squares of spans, and their composition is deduced from the
'ver1tical' composition of squares.
Proposition. A (K, M)-dominated system of structures can be
defined as a unitary lax 2-functor F from
the trivial 2-category Cop
to KM, hence also as a 2-functor lF
from an appropriate 2-category C2
to KM. And the associated
K-presheaf is a KM*-wise colimit of
the 2-functor lF.
The
category C2 has been constructed by several authors (Gray [17],
Vaugelade [22]). The notion of a KM*-wise colimit of a 2-functor is introduced in
a 1974 paper on multiple categories I have written with Charles. In this paper
we give an existence theorem for such lax colimits. In the comment mentioned
above, I use this theorem to prove that, if K is cocomplete, then lF admits a KM*-wise colimit, so that F
has an associated K-presheaf.
b) Completely
regular ordered categories
In fact
in Charles' Note [14] as well as in my comment on it [7], the setting seems
somewhat different since the 'domination' depends on an appropriate order <
on K rather than on an ss-admissible sub-category M of K. However, the
difference is not essential as I am going to prove.
There it
is supposed that (K, <) is a completely
regular ordered category. It means
that it is a category internal to the category Pos of posets which satisfies
the following conditions:
(i) If f and
f' have the same domain, the same codomain and are both less than f", then f = f'.
(ii) If k < gf, there exists g’ < g and f’ < f such that k = g’f’.
(iii) If s' is an object less than the codomain β(f) of f,
the set of f' such that f' < f and β(f') = β(f) has a maximal element, denoted by s'f and called the pseudoproduct
of (s', f).
(iv) For two objects s'
< s, there exists an ss': s'
→ s, called the pseudoproduct of (s,
s'), such that s' < ss' < s
In such a
category it is said that g and f admit a pseudoproduct gf if the
class of composites g'f' where g' < g and f' < f admits gf as a
largest element. This pseudoproduct extends the composition of the category,
but it does not always exist. Charles has proved [12] that, if the pseudoproducts
(hg)f and h(gf) are both defined, they are equal.
Proposition. Let (K, <) be a completely regular ordered category. Then the class M<
of the pseudoproducts s's of objects such
that s' < s is ss-admissible. Conversely, if K is a category admitting pullbacks and M is an ss-admissible class of monomorphisms of K closed by composition, then there exists an
order < on K for which M = M<.
Proof. 1. M< contains the identities and ss': s' → s is the unique element
of M< between s and s'. To prove that it is stable by pullbacks,
let us show that, for each f: s*
→ s there exists a pullback of
(ss', f) whose projections are s'f and s*s'*: We have (ss')(s'f) = f(s*s'*) both composites having the same domain
and codomain and being less than f. If
(ss')g = fg',
the pseudoproduct s'*g' factors both g and
g'. Indeed, using the associativity of the pseudoproduct, we get:
(s'f)(s'*g') = ((s'f)s'*)g' = (s'f)g' =
s'(fg') = s'((ss')g) = (s'ss')g = g;
it follows
that s'*g' has the same domain as g, whence (s*s'*)(s'*g') = g', both having the same doman and codomain and
being less than g'.
2. Let M be an ss-admissible sub-category of K. We
define an order < on K as follows:
f' < f if there exist m and
m* in M such that fm* = mf'.
The axioms
of an ordered category are easily proved, as well as condition (i) above, using
the fact that there exists at most one monomorphism in M between two objects. And
for each m: s' → s in M, we
have s' < m < s, hence m is the pseudoproduct ss'
To prove
condition (ii) we suppose that k < gf,
so that there exist m' and m* in M with m'k = gfm. We form the pullback of (g, m'), its projections being m in M and g'; since m'g' = gm, we
have g' < g. Since m'k = gfm* by
definition of a pullback, there exists an f'
such that g'f' = k and mf' = fm*, hence f' < f.
It
remains to prove that condition (iii) is satisfied. If f: s* → s and s' < s
we form the pullback of (f, m), where
m = ss'. Its projections are m* in M and f'. As fm* = mf', we see
that f' is an element less than f and with codomain s'. It is the pseudoproduct s'f; indeed, for any h < f with codomain s' we have mh = fm' for some m' in
M, so that there exists an n with m*n = m' and f'n = h; the first equality implies that n is in M, and the second that h
< f'.
In my
thesis, completely regular inductive categories were also used as a setting to
construct the associated sheaf. The following proposition (which I gave in
Comment 143.1, Part II of [7]) explains why it is possible.
A completely regular inductive category is
a category internal to the category of complete join-lattices (the join of a
family (si)i is noted
Visi) which is
also completely regular as an ordered category and moreover satisfies:
(v) If s = Visi for f:
s' → s, then f = Vsif.
Proposition. If K is
a completely regular inductive category, then there is a Grothendieck topology
on K generated by the pretopology
admitting as covering families of s the families (ssi)i
of the pseudoproducts such that s = Vsi.
The axioms of a
pretopology are easily verified, using the fact that sif is a projection of the pullback of (f, ssi) and the axiom (v).
4. Semi-sheaves with values in concrete
categories
Here we consider the case where K is
a concrete category, thus a category equipped with a faithful functor q to the category Set. We suppose that q preserves pullbacks and that the
ss-admissible class M of monomorphisms of K is sent by q into the class Ins of inclusions of a subset
into a set. Then a (K, M)-semi-sheaf F has an underlying semi-sheaf of sets qF whose fibers are the images by
q of those of F and qF(x)
is the image by q of F(x). The associated presheaf theorem can
be applied both to F, giving the K-presheaf F', and to qF. However the underlying presheaf qF' to F' is generally not the presheaf (qF)' associated to qF, except if q also preserves colimits (for example
if K is the category Top); there is only a non-invertible morphism from (qF)' to qF'.
In this section we consider two such
situations which are extensively used in my former work: semi-sheaves of
categories which I initially called category
of categories [1] and semi-sheaves of locally convex spaces which are at
the root of in my theory of distructures.
a) Locally convex semi-sheaves
To define locally convex
semi-sheaves, we take for the category K the category Lcs
of locally convex topological vector spaces, and for M the class MLcs
of inclusions m: E' → E of a
vector sub-space E' of E, but with a locally convex topology finer than the
topology induced by E. The reason for this choice of M comes from the initial
example which I met to define distributions, where the fibers are spaces of
continuous functions with the compact-open topology, on which differential operators
act on sub-spaces of differentiable functions, the operation being continuous
only if these sub-spaces are equipped with a finer topology, namely the
compact-open topology for the functions and their derivatives.
Definition. A locally
convex semi-sheaf is defined as a (Lcs, MLcs)-semi-sheaf.
Thus a locally convex semi-sheaf F
on C has for fiber Fe a
locally convex space, and F(x), for an
x: e' → e, is a continuous linear map to Fe' from a vector sub-space Fx of Fe equipped with a locally convex topology finer than
that induced by Fe. If we
note F(x)(s) = sx
for s in Fx the transitivity becomes:
If sx
is so defined and if xx' is defined
in C, then s(xx') is defined if and only if (sx)x' is defined, in which case both are
equal.
The Associated Presheaf Theorem
associates to F a presheaf F' of locally convex spaces on C. As we have proved
in Section 2 (of which we take back the notations), its fiber F'ê on the object ê of C is the
colimit of the functor Gê
from the subdivision category of ê↓C to Lcs which, for each commutative
triangle t = y,x,y' sends it
on the inclusion mx:: Fx → Fe and ct
on F(x). To construct this colimit in
Lcs, we first take the direct locally convex sum of the family (Fe)y:ê→e; its elements are linear
combinations of pairs (y, s) where y: ê → e and s
is in Fe. Then F'ê is the locally convex space
quotient of this sum by the vector sub-space generated by the elements (y’x, s) – (y’, sx), where y' has ê for its domain and y'x is defined in C. The topology of F'ê is the final locally convex
topology for the sink (jy),
where jy: Fe →
F'ê maps s on the equivalence class [y, s] of (y, s).
Thus the underlying (Set-)presheaf to F' is 'larger' than the presheaf P associated
to the semi-sheaf underlying F; more specifically, the fibers of F' are
generated (in the sense of vector spaces) by those of P.
b)
Lcs-sheaf associated to a locally convex
semi-sheaf
In the application to distributions
on a locally convex space E, we want not only to construct a presheaf (which
will correspond to finite order distributions), but also to associate a sheaf
to this presheaf, a distribution gluing together finite order distributions. To
define a sheaf, more structure must be given on the category C; in the
classical case of distributions C is the category of open sets of E. In my
thesis [1] I supposed that C was a completely regular inductive category, and
the associated sheaf was defined in terms of completion of an inductive species
of structures (in the terminology of C. Ehresmann [10]). More generally, C can
be equipped with a Grothendieck topology J (why it is more general is explained
at the end of Section 3). Then an Lcs-sheaf on C for J is
an Lcs-presheaf F* on C satisfying the following condition, for each covering
sieve (ci)i of e:
For each pair (i, j) of indices, let eij
be the pullback of (ci, cj)
with its projections prji
and prij.
There exist morphisms pr and pr', from the product of the fibers of
F* on the different ei
to the product of the fibers on the different eij, which glue together respectively the family (F*(prij))i and
(F*(prij))i. Then F*e is the kernel of (pr, pr').
Theorem (Lcs-sheaf associated to a locally convex semi-sheaf). If J is
a Grothendieck topology on C, the category
of locally convex semi-sheaves on C admits
the category Sh(C,J)(Lcs) of Lcs-sheaves as a reflective sub-category. The reflection F* of a locally convex semi-sheaf F is the Lcs-sheaf associated to the Lcs-presheaf
F' associated to F.
Proof. To describe the Lcs-sheaf F*, we use a method similar
to the construction of the complete enlargement of an inductive species of
structure given by C. Ehresmann (to be compared to the construction given by
Farina & Meloni [16] through "locally compatible families with covering
support").
1. Let e an
object of H. The set E underlying F*e will be the set of complete compatible families on e,
where a complete compatible family on u is
defined as a family σ = (σx)x satisfying the following conditions:
• its index
set, denoted Iσ, is a covering sieve of e for J; for each morphism x:
ex → e in Iσ, σx
is an element of the fiber of F' on ex;
• if y has ex for codomain, the 'restriction' σxy = F'(y)( σx)
of σx belongs to σ;
• two elements
σx and σx' of σ are compatible, meaning that σxp = σx'p' where p and p' are the projections of the pullback of (x, x');
• σx belongs to σ if there exists a covering sieve W
of ex such that its
restrictions σxw to the elements
w of W belong to σ.
2. For each x with
codomain e, let Ax be the subset of E formed by the σ for which Iσ contains x, and fx
the map from Ax to the
fiber F'x of F' on ex sending σ on σx. If p has ex for
its codomain, Ax is
included in Axp since
σ contains with σx
all its restrictions.
We equip
Ax with the locally convex
structure inverse image by fx of the locally convex space F'x. For each p the above inclusion becomes a linear
continuous map A(p)
from Ax to Axp
which satisfies the equality: F'(p)fx = fxpA(p). This defines a functor A from the
category e↓Cop to Lcs. The fiber
F*e is the colimit of this functor.
More
explicitly: its vector structure is such that kσ = (kσx)x and σ + σ'
is the complete compatible family generated by the family (σx' + σ'x')x' indexed by the intersection of Iσ and Iσ'.
The topology is the finest topology for which the inclusions of Ax in F*e become continuous. A basis of neighborhoods of 0 is of
the form N((Wy)yεY) where Y
is a covering family of e and Wy is a convex neighborhood of
0 in the fiber of F' on ey
for each y: ey → e
in Y, and N((Wy)yεY) is the
set of σ for which Iσ contains y
and σy is in Wy. for
each y in Y.
3. For z: ê → e,
the continuous linear
map F*(z): F*e → F*ê sends σ on the complete compatible family
generated by the restrictions σwz of the
elements of σ.
The natural
transformation η': F' → F* defining the reflection into Sh(C,J)(Lcs) is defined as follows: η'(e): F'e → F*e
is the continuous linear map sending s
on the family (sx)x of all its restrictions.
Corollaire. The above construction also gives the Top-sheaf for J associated to a Top-presheaf, or to a (Top, Ins)-semi-sheaf.
Proposition. If the fibers Fe
of F are complete locally convex spaces, then F' is a presheaf of complete locally convex spaces and F* a sheaf of complete locally convex spaces.
Indeed, the locally convex sum of a family of
complete locally convex spaces is complete, the
quotient of a complete locally convex space is complete as well as the inverse
image and a colimit.
c) Semi-sheaves of categories
A semi-sheaf of categories is defined as a (Cat, MCat)-semi-sheaf
F on C, where MCat is the class of inclusions of a sub-category in a
category. Thus F associates to each object e
of C a category Fe and to
an x: e' → e a functor from a sub-category Fx of Fe to Fe'
subject to the transitivity axiom: if xx'
is defined in C and b is in Fx then b is in the intersection of Fx
and Fxx' if and only if bx = F(x)(b)
is in Fx' and then (bx)x' = b(xx').
The construction given in Section 2
and the form of colimits in Cat imply that the fiber F'ê of the associated presheaf of categories F' is the
quasi-quotient category of the category L coproduct of the family (Fe)y:ê→e by the equivalence relation R
generated by the relations:
(y, bx) ~ (yx, b) for each y: ê → e and b in Fe.
(For the definition and constructions of a
quasi-quotient category, cf. C. Ehresmann (1965) who shows that the underlying
set of the quasi-quotient category of L by R is generally larger than the
quotient set L/R .).
If J is a Grothendieck topology on
C, the construction given above of the associated sheaf via complete compatible
families is readily transcribed to construct the sheaf of categories
F*associated to F', hence also to the semi-sheaf F. In particular the objects
of the category F*e are the
complete compatible families e whose
terms σx are objects in the corresponding
fibers of F'.
The Evolutive Systems which we have introduced with Jean-Paul
Vanbremeersch [8] to model the evolution of a living system are semi-sheaves of
categories F on a category Time which is a category defining the order > on
an interval or a finite part T of the space R+ of positive real numbers,
so that its morphisms are pairs (t, t')
with t < t'. The idea is that, for
each instant t the category Ft models the configuration of
the system around t: its objects are
the state at t of the components of
the system which then exist, and the morphisms between them represent the interactions
around t. The functor F(t, t'), called transition, models the change of the
system from t to t' and, to take account of the possible loss of some components
and/or relations between t and t', it is defined only on a sub-category
Ftt' of Ft. If at is an object of Ft,
either it models a component that still exists at t', and then F(t, t')(at) is its new state at t'; or it does no more exist at t'
and then it is not in Ftt';
and the same for the interactions represented by the morphisms. If S is the set
consisting of all the objects of the different fibers Ft and p: S → T its projection map, a component of the system can be defined
as a maximal partial section of p consisting
of its various successive states during its 'life'.
Living systems have a whole hierarchy of
components of various complexity levels. To model this hierarchy we have
enriched the evolutive systems and defined the Hierarchical Evolutive Systems. The idea is that a complex
component a of level n+1 can be modeled as the colimit of a
diagram modeling its lower level internal structure, that is its components of
levels less than, or equal to, n and
their interactions concerning a. Thus
we define a hierarchical category as
a category whose objects are partitioned into a finite number of 'levels', say
0, 1, ..., m, so that an object of
level n+1 be the colimit of at least
one diagram with values in the full sub-category whose objects are of levels
less than, or equal to, n. A hierarchical functor is a functor
between hierarchical categories which respects the levels (but it may or may
not preserve colimits). Let HCat be the category of hierarchical categories.
A Hierarchical Evolutive System is a (HCat, MCat)-semi-sheaf
on a category Time. Otherwise, it is an evolutive system whose fibers are
hierarchical categories, the transition functors respecting the levels [8].
The theory of Memory Evolutive
Systems developed with Jean-Paul Vanbremeersch (cf. our book [8], summarized on
our internet site [9]) is based on such hierarchical evolutive systems.
5. (K, M)-distructures
The definition of distributions will
be made 'locally' on an open set U of a locally convex space E, by extending
the differential operators to continuous functions defined on U (using the Associated
Presheaf Theorem); and then we'll have to connect the distributions so defined
on various U. Thus we have to consider not just one semi-sheaf on U, but a
'presheaf of semi-sheaves' on E. It is this situation which the notion of a
distructure generalizes.
a) Generators of distructures
We suppose that K is a category
admitting pullbacks and colimits and M an ss-admissible class of monomorphisms
of K. Then the category SS(K, M) of (K,
M)-semi-sheaves has a functor Base to
Cat and an 'associated presheaf' functor Ps toward the category Ps(K) of
K-presheaves.
Definition. An SS(K, M)-presheaf
D on a category H is called a (K, M)-generator of distructures. The composite
of D with Base is a presheaf of
categories Γ on H called the base of D,
and the composite D' of D with Psh: SS(K, M) → PS(K) is called the K-distructures
presheaf generated by D.
Let D be a generator of
(K, M)-distructures on H, For each object u of H, the fiber Du of D is a (K, M)-semi-sheaf on the category Γu
which is the fiber on u of the
presheaf of categories Γ base of D. For a morphism v: u' → u of H, D(v) is a morphism of semi-sheaves:
D(v) = (Γ(v),
μv): Du → Du'. The
K-distructures presheaf D' generated by D is a Ps(K)-presheaf
on H with the same base Γ, whose fiber on u is the
K-presheaf associated to Du.
If H is reduced to a unique object,
then a (K, M)-generator of distructures reduces to a (K, M)-semi-sheaf,
its base being reduced to one category.
We denote by GDisH(K,
M) the category of (K, M)-generators of distructures on H which is a full
sub-category of PS(SS(K, M)). A morphism from D to E
is a natural transformation. Let DisH(K) be
its sub-category of K-distructures presheaves. Since PS(K)
is reflective in SS(K, M), we have:
Proposition. The category DisH(K) is a reflective sub-category of GDisH(K, M). The fiber D'u of the
reflection D' of D is the K-presheaf associated to Du.
b) Distructures as connecting two semi-sheaves
We are going to give another
characterization of a (K, M)-generator of distructures with base Γ by considering the fibration Fib: F(Γ) → H associated to Γ. It will explain in
which sense we can speak of 'di'structures. Let us recall that the class of
objects of F(Γ) is the (disjoint) union of the classes of objects of
the various Γu; a
morphism from e' in Γu' to e in Γu is
defined by a pair (v, x) where
v: u' → u is in H and x: e' → ev = Γ(v)(e) is in Γu';
and its composite with
(v', x'): e" → e', where v':
u" → u', is (vv', xv'.x'):
e" → e (the dot denotes the composition in Γu"). The functor Fib maps (v, x) on v.
The category F(Γ)
has 2 distinguished sub-categories:
• the category Vert of 'vertical morphisms' (those mapped by
Fib on an object) which is the coproduct of the categories Γu; for y in Γu;
we identify (u, y) with y if no confusion is possible;
• the sub-category Cart of cartesian morphisms which are of
the form (u, e) where e is an object of Γu.
These
sub-categories generate F(Γ) since we have: (v, x) = (v, ev)x for x in Γu'.
Theorem. There is
an isomorphism φ from the category of (K, M)-generators of distructures with base Γ onto the category of (K, M)-semi-sheaves on F(Γ)
whose restriction to Cart is a K-presheaf. It sends the presheaf of K-distructures generated by
B to the K-presheaf associated to φB. If D'
is a presheaf of K-distructures, then
φD' is a presheaf;
Proof. 1. If D is a (K, M)-generator of distructures,
we define a (K, M)-semi-sheaf φD on F(Γ) as
follows: Its fiber φDe
on the object e of Γu is the fiber (Du)e of the (K, M)-semi-sheaf Du on e. For a
morphism (v, x): e' → e, in the semi-sheaf Du we have mx: (Du')x → (Du')ev in M and we take the
pullback φDv,x of (mx, μv(e)), where D(v) =
(Γ(v), μv); the projections in the
corresponding pullback square PBx are denoted qv,x and mv,x
in M (since M is stable by pullbacks), and we define:
φD(v, x) = Du'(x)qv,x: φDv,x → (Du')ev.
To prove that φD so defined is a (K, M)-semi-sheaf, we must
prove the transitivity. For this we also consider (v', x'): e" → e',
and its φD(v',
x') = Du"(x')qv',x'.:
φDv',x' →
(Du")e". To construct the
image by φD of the composite c = (vv', xv'.x') we
first construct the image by Du" of the composite xv'.x'. Since Du" is a semi-sheaf, the domain of Du"(xv'.x') is the codomain of the projection p'xv'.x: I → (Du")xv'.x
in the pullback square PBm
(drawn in red in the diagram), and I is also the pullback of (mx', Du"(xv')), with corresponding pullback square
PB1. Then we have φD(c)
= Du"(xv'.x)qc
where qc is the projection
of the pullback of (mxv'.x, μv'(ev)μv(e)), whose corresponding pullback square
is PBxv'.x.
The
transitivity of φD means that the pullback I' of (mv',x', φD(v, x)) (with pullback square PB2)
is also the pullback of (mv,x,
mc). To show this, we
consider the composite pullback square PBx''°PB2
and the adjacent pullback square PB1. Their bases form a commutative
square with μv'(x)qv,x as its 4th side, so that there
exists a morphism q from I' to I
closing a pullback square PB3 (with I' and (Du')xv' as opposite vertices) which satisfies
PB1'°PB3 = PBx''°PB2.
Similarly,
considering the composite pullback square PBm°PB3
and the adjacent square PBxv'.x'
connected by mv,x we
complete by p' a pullback square PB4
(in red on the diagram) which defines I' as the pullback of (mv,x, mc).
For a cartesian morphism (v, e),
the above construction implies φD(v, e) = μv(e) and the restriction of φD to Cart is reduced to a
K-presheaf.
If D takes its values in the
category of K-presheaves, so that it is already a presheaf of K-distructures,
then mx for x in Γu' is
an identity so that
mv,x = Id and
φD(v, x) = Du'(x)μv(e).
Thus φD is a K-presheaf (and not only a (K,
M)-semi-sheaf).
In this case, there is another
construction of φD (which,
for K = Set, I gave in Comment 490-1, p. 808, Part III [7]). It relies on the
fact that the fibration F(Γ) is a lax colimit of Γ looked at as a 2-functor from Hop
to the 2–category CAT, the
canonical lax cocone with vertex F(Γ)op associating to v the
2-cell (Γ(v), (v,
-v)) from the inclusion Insu:
Γuop → F(Γ)op to Insu'. Now D can be seen
as defining a lax cocone with basis Γ and vertex K associating to v the 2-cell from Du to Du' defined by D(v) = (Γ(v),
μv). This lax cocone
factors through φD.
2. Conversely, let B be a
(K, M)-semi-sheaf on F(Γ) whose restriction to Cart is a
K-presheaf. We define a (K, M)-generator D of distructures as follows. The (K,
M)-semi-sheaf Du will be a
restriction of B, namely the composite of B with the 'inclusion' Insu: Γuop → F(Γ)op.
Since B becomes a presheaf on Cart, for
v: u' → u and e
an object of Γu, B(v, ev) is a
morphism from Be = Du(e) to Bev = Du'(ev) which we take for μv(e). To construct μv(y) for an y: ê → e in Γu, we take into account the two decompositions:
y(v, ê) = (v, y) = (v, ev)yv.
Since the restriction of B to Cart
is a presheaf, the first decomposition implies (by transitivity) that there
exists a p'(v,ê),y
such that:
B(v, ê)B(y) = B(v, y)p'(v,ê),y and p'(v,ê),ymv,y = my.
From the second decomposition we deduce
that Bv,y
is the pullback of (myv,B(v, e)), the projections being mv,y and a qv,y such that B(yv)qv,y
= B(v, y). Then we define μv(y) = qv,y p'(v,ê),y. The above diagram being
commutative, this entirely defines a morphism (Γ(v), μv): Du → Du', whence a (K, M)-generator of distructures D, and we
have φD = B,
both K-presheaves having the same restrictions to the generating sub-categories
Vert and Cart of F(Γ).
3. The bijection φ extends into an isomorphism from the category
of (K, M)-generators of distructures with base Γ, onto the category of (K, M)-semi-sheaves
on F(Γ) whose restriction to Cart are presheaves. Indeed, let γ: D → E be a natural transformation; for
each object u of H, γ(u) is a morphism
of (K, M)-semi-sheaves from Du
to Eu and we take φγ(e)
= γ(u)(e)
for an object e in Γu.
To define φγ(v, x) for v: u' → u and x: e' → ev, we come back to the definition of φD(v, x) and φE(v,x) via the formation of the
pullbacks PBx and PB'x.
Since γ is a
natural transformation, we have
μv(e)(γ(u')(ev)) = (γ(u)(e))μ'v(e)
= φγ(e)μ'v(e)
where (Γ(v), μ'v):
Eu → Eu'; whence a commutative square whose composite with PBx must factor through the pullback
square PB'x to give a cube
with other edges γ(u)(x) and the wanted φγ(v, x):
φD(v, x) → φE(v, x). The commutativity of the diagram ensures that we have thus
defined a morphism (Id, φγ):
φD → φE. From the definition, it follows that the map sending
γ to φγ defines the required isomorphism.
4.
The isomorphism φ sending the reflective sub-category
of presheaves of K-distructures with base Γ onto the reflective sub-category of
K-presheaves on F(Γ), it preserves the reflections.
Hence the presheaf of K-distructures D' associated to D has for image by φ the K-presheaf associated to φD.
This theorem shows how a (K,
M)-generator of distructures D determines 'two structures' (whence the name distructure)
of semi-sheaves with the same fibers:
• a 'vertical' (K, M)-semi-sheaf on the
category Vert which is the 'union' of the semi-sheaves Du,
• a 'horizontal' K-presheaf on the
category Cart, deduced from the morphisms between the Du.
These two structures are inter-connected since
they both are restrictions of the 'global' (K, M)-semi-sheaf on F(Γ).
6. Locally convex distructures
The preceding double structure can
be made more explicit in the case where K is the category Set, or more
generally a concrete category..
a) Distructures valued in a
concrete category
If K = Set and M is the class of inclusions
of a subset in a set, then a (K, M)-generator of distructures D is simply
called a generator of distructures
and its associated presheaf of K-distructures D' a distructures presheaf.
The double structure can then be
described as follows. Let Su
be the (disjoint) union of the fibers (Du)e of
the semi-sheaf Du and Σ the (disjoint) union of the sets Su for the various objects u of H. If Γ is the base of D and F(Γ) its associated fibration, the preceding theorem defines a semi-action
of F(Γ) on Σ (corresponding to the presheaf φD). It has for restrictions:
• a 'vertical' semi-action of the sub-category Vert
on Σ (gluing
the semi-actions of the various Γu),
with the fibers of D as 'horizontal' fibers; and
• a
'horizontal' action of the sub-category Cart with the same fibers.
Gluing the maps images by φD of the cartesian morphisms over a
given v, we obtain a global 'horizontal'
action of H on S; it corresponds to the presheaf S on H having for 'vertical' fibers
the various Su and such
that H(v) glues the various maps μv(e),
where D(v) = (Γ(v),
μv). The figure shows what we mean by 'vertical'
and 'horizontal'.
For the distructures presheaf D'
generated by D, the semi-action of Vert is also an action, so that the union Σ' of all its fibers is equipped with
two actions: the vertical action of Vert with horizontal fibers (D'u)e, and the horizontal
action of H with vertical fibers S'u
uniting the horizontal fibers over u.
More generally, let K be a category
with a faithful functor q to Set which preserves pullbacks, and M an ss-admissible class
whose image by q consists of inclusions.
If D is a (K, M)-generator of distructures, it admits an underlying generator
of distructures, denoted by qD, whose
fibers are the semi-sheaves qDu underlying the fibers of D (cf. Section
4). The vertical structure of qD (defined
by the semi-action of Vert) underlies the vertical structure of D, defined by
the (K, M)-semi-sheaf gluing the semi-sheaves Du. However the horizontal structure of qD is more precise than that of D, since
the presheaf S on H is not underlying a K-presheaf, except if K admits
coproducts which are preserved by q.
In this special case, the K-presheaf on H has for fiber on u the coproduct of the fibers (Du)e for the various objects e of Γu.
For instance, if K = Top, the
presheaf S lifts to a presheaf of topological spaces; or if K is the category
Cat, it lifts to a presheaf of categories.
b) Locally convex distructures
Here we examine the case where K is
the category Lcs of locally convex topological vector
spaces, and M the class MLcs of inclusions m: E' → E of a vector sub-space E'
of E with a locally convex topology finer than the topology induced by E.
Definition. A (Lcs, MLcs)-generator
of distructures L is called a generator
of locally convex distructures (abbreviated in generator of lc-distructures), and the presheaf of distructures L'
it generates a locally convex distructures
presheaf (abbreviated in lc-distructures
presheaf).
Thus a
generator of lc-distructures L on H is a presheaf of locally convex semi-sheaves.
Its composite with the base functor gives a presheaf of categories Γ. The
fiber Lu of L on an object
u of H is a locally convex semi-sheaf
on Γu and L(v) for v: u' → u is a morphism (Γ(v),
μv): Lu
→ Lu'. From the preceding
theorem, it follows that L 'extends' into the locally convex semi-sheaf φL on the fibration F(Γ) associated to Γ, thus
interconnecting the two structures of L: the 'vertical' one corresponding to
the locally convex semi-sheaf (on Vert) gluing the Lu's; and the 'horizontal' one consisting of an Lcs-presheaf
on the cartesian morphisms of F(Γ).
The
forgetful functor qLT: Lcs → Top preserves the limits and sends MLcs
into the class MTop
of continuous inclusions of a sub-space with a finer topology. It
follows that L has an underlying (Top, MTop)-generator of distructures T with base Γ:
the (Top, MTop)-semi-sheaf Tu
is the composite qLTLu. Hence the fiber (Tu)e
is the topology of the fiber (Lu)e of Lu, and, for y in
Γu from ê to e,
the continuous map Tu(y): (Tu)y
→ (Tu)ê maps s on Lu(x)(s).
By a
construction similar to that of the presheaf S above we obtain the following
presheaf of topologies Θ on H:
• Its fiber on u is the topological space Θu coproduct in Top of the topologies of
the fibers of Lu;
• Θ sends v:
u' → u to the
continuous map Θ(v):
Θu → Θu' gluing together the continuous maps µv(e)
for the various objects e in Γu, so that Θ(v)(e,
s) = (ev, µv(e)(s)
for each s in the fiber (Lu)e.
We can apply the same construction to the lc-distructures presheaf L'
generated by L; it leads to the Top-distructures presheaf T' whose fiber on u is the Top-presheaf qL'u;
and to the Top-presheaf Θ' whose fiber Θ'u
is the topology coproduct of the topologies on the fibers of L'u. As the functor qLT does not preserve
colimits, T' is not the Top-presheaf
associated to T, nor Θ' the Top-presheaf associated to Θ.
If a Grothendieck topology J is given on H, L
also generates a Ps(Lcs)-sheaf L*, called the lc-distructures sheaf on H generated by
L, which is the Ps(Lcs)-sheaf associated to L'. It is constructed hereafter. It
has an underlying Top-sheaf Θ* and the following figure shows how the fibers increse
from L to L' and to L*.
c) Construction of the lc-distructures
presheaf and sheaf generated by L
Let L be a generator of lc-distructures with basis Γ. Using the description of the
associated locally convex presheaf given in section 4, we can describe as
follows the lc-distructures presheaf L' generated by L.
• For an object ê of the fiber Γu
of Γ on u, the fiber (L'u)ê
is the locally convex space constructed as follows. We take the locally convex
direct sum Vê of the
family of locally convex spaces (Lu)e indexed
by the morphisms y: ê → e in Γu; its elements are linear combinations of pairs
(y, s) where s is an element of (Lu)e. The relations
(y, σ) ~ (y’, σx) if y =
y'x in Γu
and σ is in (Lu)x
generate a congruence Rê on the vector space Vê. Then (L'u)ê is the locally convex space
quotient of Vê by Rê. Its topology is the final
locally convex topology for the sink (jy),
where jy: (Lu)e
→ (L'u)ê
maps s on the equivalence class [y,
s] of (y, s). Let us note that it
is linearly generated by the classes [y,
s].
• For z:
ê' → ê in Γu, the
continuous linear map L'u(z): (L'u)ê
→ L'u(ê') is entirely defined by the fact that it sends [y, s] on [yz, s]. Thus we have well defined the Lcs-presheaf L'u.
• The morphism L'(v) image by L' of a v: u' → u in H is L'(v) = (Γ(v), μ'v) where μ'v(ê) is the unique continuous linear map from (L'u)ê
to (L'u')ê such that
μ'v(ê)([y,
s]) = [yv, μv(e)(s)].
• The
natural transformation η from L to L' defining L' as a reflection of L in
the sub-category DisH(Lcs) of lc-distructures pesheaves is defined
as follows: η(u)(e): (Lu)e
→ (L'u)e is the continuous linear
map sending s on [e, s].
Proposition. If H is equipped with a Grothendieck
topology J and if Γ is
a sheaf of categories, there
is an lc-distructure sheaf generated by L, namely the PS(Lcs)-sheaf L* for
J associated to L'; it hash Γ as its base and its underlying Top-sheaf Θ* is the Top-sheaf associated to Θ'.
Proof. The hypothesis that Γ is a sheaf of categories
for J could be omitted, but then the base of L* would not be Γ but its asociated
Cat-sheaf. To construct L* we first construct the Top-sheaf Θ* associated
to Θ', by the method indicated in Section 4b, Corollary.
1. The elements of
the fiber Θ*u are the
complete compatible families on u; such
a family is of the form (ev, σv)vεV
where V is a covering sieve on u and (ev, σv) is an element
in the fiber of Θ' on the domain dv of
v, meaning that σv is an element of the fiber
of L'dv on the object ev of Γdv. Moreover the σv must be compatible and the
family 'maximal'. These conditions implies that the corresponding family (ev)v
forms a complete compatible family of objects of Γ on u and, since Γ is a sheaf for J, it determines an object e of Γu admitting ev
for image ev by Γ(v) for each v in V; thus the family (ev,
σv)vεV can be written (e, σ), where σ = (σv)vεV, and the
fiber (L*u)e has these families σ for elements. The topology is such that
a basis of neighborhoods of 0 is formed by the sets Ω(Wv)vεV formed
by the σ containing a σv
in Wv, where V is a
covering sieve on u and Wv is a convex neighborhood of
0 in (L'dv)ev
for each v in V. For an x: e' → e in Γu the linear continuous map
L*u(x) from (L*u)e to (L*u)e' sends σ on
the family (σvxv)v, = σx, where σvxv = L'(xv)(σv).
This defines the Lcs-presheaf L*u.
2. Let w: u° → u be a
morphism in H. To define the morphism (Γ(w), μ*w): L*u → L*u°, for each v of the covering sieve V on u we consider the pullback of (v, w) with its projections v°, w°. The v° generate a covering sieve on u°. Then μ*w(e) is the continuous linear map from (L*u)e
to (L*u°)ew sending the element σ
= (σv)v of (L*u)e
on the family (σv°)v°, where σv° = μ'w°(e)(σv).
The natural
transformation η*: L'
→ L* is such that η*(u)(e) sends an element s of (L'u)e on the family (μ'w(e)(s))w of L*u(e) indexed by all the w with codomain u.
Proposition. If the locally convex spaces (Lu)e are complete, so will be the locally
convex spaces (L'u)e (resp. (L*u)e if J is a Grothendicek topology on H)
so that the lc-distructures presheaf L'
(resp. sheaf L*) generated by L is a complete lc-distructures presheaf (resp.
sheaf).
d) Locally convex distructures
with base a presheaf of monoids
In many applications, in particular
to define distributions as we will do in the next section,, we have a generator
of lc-distructures L on H, whose base Γ is a sheaf of monoids, Thus each category Γu has
a unique unit e, so that the
Lcs-semi-sheaf Lu has a
unique fiber (Lu)e; let us denote it by Lu. These locally convex
spaces Lu are the
fibers of a locally convex presheaf L on H: for v: u' → u in H, the
linear continuous map L(u): Lu → Lu' is equal to μv(e), where L(v) = (Γ(v), μv). We
call L the Lcs-presheaf underlying L.
Then, the Top-presheaf underlying L
(obtained by 'forgetting' the vector structure on the fiber, keeping only their
topology) is identical to the Top-presheaf Θ constructed in Section (b); indeed
this presheaf had for fiber Θu the topology coproduct of the topologies of
the fibers of Lu and now
there is only one fiber.
In this case, a generator of lc-distructures on H can be defined as
follows:
• a presheaf of monoids Γ
and a presheaf of
locally convex spaces L, both on H;
• for each object u of H a semi-action of Γu on the fiber Lu these semi-actions being preserved by the 'change of
fiber' L(v).
Similarly, from the lc-distructures
presheaf L' generated by L, we construct the underlying Lcs-presheaf L'
on H whose fiber on u is the unique
fiber of L'u, and the
action of Γu on L'u, these actions being preserved by the 'change of
fiber'. And, if H is equipped with a Grothendieck topology J, we also construct
the Lcs-sheaf L* underlying the lc-distructures sheaf L* generated by L';
from the construction of L* we see that L* is the Lcs-sheaf associated
to L'.
7. Distributions
The situation which has motivated
the introduction of distructures is the example of Schwartz distributions [19,20]. As already said, in my thesis, my aim was to define such
distributions on infinite dimensional spaces by 'extending' the differential
operators to continuous functions from a locally convex space E to a locally
convex space E'. In fact Schwartz mentions in his book [19] that this was the
initial idea, and he proves that, in the finite dimensional case he considers,
each distribution glues together finite order distributions defined on open sub-spaces,
a finite order distribution reducing to a higher derivative (in the
distribution sense) of a continuous function.
To translate this in terms of
distructures, the rough idea is to consider the monoid of 'derivations' and its
semi-action on the continuous functions on an open set U of E. This leads to a
generator of lc-distructures on the category H of open sets of E. Distributions
correspond to the associated lc-distructures sheaf. In the case of
finite-dimensional E, they give back the Schwartz vector-valued distributions [20].
a) The generator of distributions
In the usual definition of
distributions defined through the dual of a space of infinitely differentiable
functions, the choice of the concept of differentiability is important. Our
definition is less dependent on this concept (this will be made more precise
later on). What will be essential is the concept (more or less underlying the
various differentiability concepts) of successive partial directional derivatives
which are at the basis of the construction of any differential operator.
Let E and E' be two locally convex
spaces, and f a continuous function
from U to E', where U is an open set of E (with the induced topology). If αn is an element (vector) of E, we say
that f has a derivative in the direction
of αn on U if,
for each a in U, there exists
∂f(a)/dαn = limk→0((f(a+
and if the function ∂f/dαn so defined is continuous from U to
E'. By iteration, we define higher order directional derivatives.
To fix simple enough notations for
such derivatives, we use the notion of a multiset
α of
vectors of E. It is an element of the free commutative monoid on the set of
vectors of E, hence a collection of n
possibly repetitive vectors α = <α1, α2, …αn>; the angular brackets indicate that the
order is not taken into account and to be repetitive means that we may have αi = αj for two different indices;
we call n the order of α.
We say
that the function f from U to E' has an
α-derivative, or more explicitly
an n-th partial derivative with respect
to α, denoted by f.α if, for each a in U, the restriction of f to the affine sub-space V = a+ΣiRαi is n-differentiable in a,
with (f.α)(a) as its partial derivative with respect to the n-multiset
<α1, α2, …αn>, and if f.α is continuous on U. The differentiability of f ensures that it has an n-linear
differential Dnf on U
which, because of the continuity of the partial derivatives, is symmetric.
Now we define the generator of
distributions on E valued in E'.
• It is a generator of lc-distructures D
on the category H (defining the order on the set) of open sets of E; we denote
by (U', U) the morphism from the subset U' of U to U. This category is equipped
with the usual Grothendick topology in which the covering sieves for U
correspond to the open covers of U
• The base of D will be reduced to the
constant functor on the free commutative monoid A on the set of vectors of E.
Its elements are the multisets α, its unit is the multiset void, denoted o, and the composition merges
the multisets. In fact, to have a smaller monoid, we may select an algebraic
base of E and take only multisets on the vectors of this base; as the n-th differential of a function is
supposed to be n-linear, we may only
take n-th partial derivatives with
respect to the vectors of this basis, the others being obtained by linear
combinations; and the result will be independent of the choice of the basis.
• For an open set U of E,
the locally convex semi-sheaf DU is defined as follows: its unique fiber (since A is a monoid) is the locally
convex space C(U) of continuous functions from U to E', with the compact-open
topology. For each multiset α in A, (DU)α is the vector sub-space of C(U)
consisting of the functions f admitting
an α-derivative on U, equipped with the compact-open topology for the
functions and their successive derivatives up to the order n of α; this topology being finer than that induced by C(U),
the inclusion (Du)α
→ C(U) belongs to the ss-admissible class MLcs. And DU(α): (DU)α
→ C(U) is the map sending a function f
which has an α-derivative
on this derivative f α. This
defines a locally convex semi-sheaf DU since f has an αα'-derivative if and
only if its α-derivative f. α has an α'-derivative.
• If U' is an
open subset of U, the morphism D(U', U) from DU to DU' reduces
to the map from C(U) to C(U') sending a continuous function g on U on its restriction g/U' to U'. This completes the definition
of a generator of lc-distructures; the Lcs-presheaf underlying D (cf. Section
6) is the presheaf C of continuous functions from E to E'.
Definition. The
generator of lc-distructures D defined above is called the generator of E'-valued
distributions on E, and the (pre)sheaf of lc-distructures it generates is
called the (pre)sheaf of (finite order) E'-valued distributions on E.
b) The
(pre)sheaf of distributions
The
general construction of the generated presheaf of distructures can be applied
to D, where the fact that A has a unique unit simplifies the description. Thus in
the presheaf of distributions D' generated by D, the presheaf D'U
has a unique fiber, denoted by D'U and obtained as follows:
we take the direct sum Σ of A copies of the vector space C(U); its
elements are linear combinations of pairs (α, g) where α is in A and g
a continuous function from U to E'. Then D'U is the
quotient of Σ by the congruence R generated by the relations r:
(α'α,
f)
~ (α', f.α) if f has an α-derivative f.α on U.
It means that an element of D'U
is an equivalence class modulo R of such linear combinations. The topology is
the finest locally convex topology for which the injections [α, -]: g |→ [α, g] become continuous from C(U) to D'U.
If U' is an open subset of U, D'(U', U) is determined
by the linear continuous 'restriction map' D'(U', U) from D'U
to D'U' which sends the equivalence class d = [α, g] on its restriction [α, g/U'] to U' (denoted d/U'), and the equivalence class of a
linear combination on the linear combination of their restrictions. This
describes the Lcs-presheaf D' underlying D'.
The monoid A acts on D'U as
follows:
D'U(β)([α, g]) = [α, g]β = [αβ, g]
for each β in A.
Remark. The equivalence relation depends on the
selected notion of differentiability. Here we define (DU)α as the sub-space of
functions f whose restriction to the
affine sub-space generated by α is differentiable. We might take only those f which are n-differentiable
in a stricter sense, for instance in the sense I have defined in [1,4]. Then the relation R becomes stricter. The advantage of
the laxer definition used here is that it allows a more explicit description of
R which the following theorem gives.
From now on, we suppose that E' is complete.
Theorem. If E' is complete, D'U identifies to the quotient of AxC(U) by the
equivalence relation R' defined by:
(α, g) R' (α', g') if there exist β, β' in A and
f, f' such that
g = f.β, g' = f'.β' , αβ = α'β' and (f
– f').αβ = 0.
Denoting [α, g] the equivalence class of (α, g), the vector structure is determined
by:
[α, g]
+ [α, h] = [α, g+h]
and κ[α, g] = [α,
κg] for a real κ.
Its topology is the finest locally
convex topology which makes continuous the maps [α, -] from C(U) to D'U. The continuous linear map β-derivation D'U(β)
from D'U to D'U maps [α, g] on [αβ, g]. For U' included in U, D'(U', U) is the continuous linear map sending
[α, g] on the class [α, g/U'].
If E is metrizable, then D'U is complete.
Proof. 1. First we prove that R' is a relation which
contains r and is contained in the
congruence R defined (before the theorem) on the direct sum Σ (it is
possible since AxC(U) is
a part of Σ).
At this end, we use the fact that, since E' is
complete, each continuous function g to E' admits a β-anti-derivative,
denoted by β-1g, for each multiset β =
<β1, β2,..., βn> in A, meaning that it is of the form f.β for
at least one function f having a β-derivative on U. Indeed, as E' is complete, each
continuous g admits on U an E'-valued
integral [6]
β-1g = ∫g dβ1dβ2...dβn
(defined up to a function b having 0 for β-derivative) which has g for its β-derivative.
It follows that, for each β in A we have
(α', g) r (α'β, β-1g).
In particular, if (αα', f) r (α', f.α) and if we take g = f = g', β = o, β' = α, we get
αα'β = αα' = α'β', f =
g.o, f.α = g'.α and (g' - g).αα' = 0,
thus R'
contains the relation r.
On the other hand, if (α, g) R' (α', g'), using the β, β', f, f' which define R', we find:
(α, g) =
(α, f.β) r (αβ, f), (α', g') = (α', f'.β') r (α'β', f'),
thus, in terms of the congruence R we have defined
before the theorem on the sum Σ,
((α, g) - (α', g')) R ((αβ, f) – (αβ, f')) R (αβ, f – f') R 0,
hence (α, g) R (α', g').
2. R' is
clearly reflexive and symmetric. To prove that it is an
equivalence, it remains to prove that it is transitive. For this, let us
suppose that
(α, g) R' (α', g') and (α', g') R' (α", g");
there exist β, β', f and
f' such that g = f.β, g' = f'.β', αβ = α'β' and (f – f').αβ = 0, and also γ, γ', h and h' tels que g' = h.γ, g" = h'.γ', α'γ = α"γ' and (h - h').α'γ = 0. If
we write ρ = γβ, ρ' = β'γ' we have
αρ = αγβ = α'β'γ = α"β'γ' = α"ρ'.
On the
other hand, let us select
k = γ-1f, whence f
= k.γ and g = k.βγ = k.ρ;
and
k' = β'-1h',
whence h' = k'.β' and g" = k'.γ'β' = k'.ρ'.
Now h.γ = g' = f'.β', and we must prove that k –k' has a ρρ'-derivative
0. We may write
k – k' = γ-1f – γ-1f' + γ-1f' - β'-1h'
= γ-1(f – f') + b + γ-1β'-1g' – β'-1h'
+ b'
where b.γ = 0 = b'.β'. However γ-1g' = h + b" with b".γ = 0. Hence
k – k ' = γ-1(f – f') + β'-1(h
– h') + b + b' + β'-1b" + b"' with
b"'.β' = 0.
f – f' has a derivative
with respect to αβ = α'β' and h – h' for α'γ = α"γ', thus the two first
terms have a derivative for αβγ = α'β'γ, their derivative
being 0, and the other terms have also a derivative since they have 0 for
derivative for γ or
for β'. It
follows that (k - k').αβγ = 0, as wanted.
3. Let us define the
vector structure on the quotient of AxC(U) by R'. For each real κ, we define κ[α, g] = [α, κg];
it is independent on the choice of the representative (α, g). Let us note that AxC(U),
looked at as a sub-space of the direct sum Σ of the family
of A exemplars of C(U), is
closed under the scalar multiplication, but only partially under the addition:
it is not a vector sub-space of Σ though the addition of Σ induces a
partial addition defined for elements with the same first term. However given two elements (α, g) and (β, h) of AxC(U)
they are equivalent modulo R' (hence also modulo R as proved in the first part
of the proof) to the elements (αβ, β-1g) and (αβ, α-1h) which can be added, the result being (αβ, β-1g + α-1h), in
AxC(U). This
allows to define directly in the quotient (by R' or by
R):
[α, g] + [β, h] = [αβ, β-1g] + [αβ, α-1h] = [αβ, β-1g + α-1h].
It does not depend on the choice of the
anti-derivatives: if we replace β-1g and α-1h by other anti-derivatives b and b' the added term
[αβ, b + b'] is 0 since b.β = 0 = b'.α. As R' contains r and is included in the congruence R, we
deduce that the result is also independent of the choice of representatives, and
that AxC(U)/R' equipped with this addition
becomes a vector space which can be identified to Σ/R, hence to D'U.
4. The
topology of D'U is the final locally convex topology for the linear
maps [α, -] from C(U) to D'U. If
E' is complete and E metrizable, then C(U) is complete
for the compact-open topology, so that D'U is also complete.
A basis of neighborhoods of 0 is obtained as follows: for each α in A, let
Kα be a compact included in U and Oα a
neighborhood of 0 in E'; let Ω(Kα,
Oα)α be the convex hull of the set of elements d of D'U admitting a
representative of the form (α, g)
for some g sending Kα
into Oα. When (Kα, Oα)α vary, these sets form a basis of
neighborhoods of 0 in D'U.
5. To achieve
the description of the Lcs-presheaf D', let U' be an open subset of U. If we
have (α, g) R' (α', g'), taking the restrictions of g and g' to U', we have also (α, g/U')
R" (α', g'/U') where
R" is the equivalence corresponding to U' (defined as R'). So we define,
by passage to the quotients, a continuous linear map 'restriction' D'(U',
U) from D'U to D'U' sending d = [α, g] on d/U' = [α, g/U].
Definition. An element
of D'U is called a finite
order E'-valued distribution on
U. Its order is the smallest n such
that it admits a representative [α, g] where the order of α is n. For each d = [α, g] in D'U the distribution D'U(β)(d) = [αβ, g] is called the β-derivative
of d, denoted d.β.
A continuous function g on U can be identified to the
distribution [o, g], but the topology
induced by D'U on C(U) is finer than
that of C(U). In particular a function f which
has an α-derivative is thus identified to the distribution [o, f] of order 0, and its α-derivative f.α is identified to [o, f.α] = [α, f], which is also its α-derivative as a distribution, according to the above definition. Moreover,
the topology induced by D'U on the locally convex
space of α-derivable functions (DU)α is the compact-open topology for
the functions and their derivatives, which is the topology we had initially considered
on this space.
c) The
sheaf of distributions
We denote by D* the lc-distructures sheaf generated
by D. It is associated to D'. As the category H on which it is defined is the
category of open sets of E, we can use the classical construction of the sheaf associated
to a Lcs-presheaf on a topological space, through
sections of the etale space of its germs (or local jets in Ehresmann's terminology).
Let us develop this construction.
To the Lcs-presheaf D' on the topological
space E, we first associate an etale space p:
JD' → E whose stalk on an element a of E is defined as follows. Let
H/a be the full sub-category of H having
for objects the open sets U of E containing a,
and D'/a the functor from H/a to Lcs restriction of D'. The
colimit of D'/a is
a complete locally convex space JD'a.
An element of JD'a is an
equivalence class <U, d> of
pairs (U, d) where d is a finite order distribution on an U containing a, and
(U, d) ~ (U', d') if d/V
= d'/V for some open subset V of U
and U' containing a.
Definition. The
equivalence class <U, d> in JD'a is called the local jet of d at a, denoted jλad.
(This terminology is that of Charles
Ehresmann [7]; a local jet of d is
often called a germ of d.)
Let JD' be the union of the spaces JD'a for the different elements a of E and p: JD' → E the map sending an element of JD'a on a. For each d in D'U,
let jλd: U → JD' the map sending a to the locl jet of d at a.
We equip JD' of the finest topology for which the maps jλd become
continuous. It admits the images of these maps as a basis of its open sets.
With this topology, JD' is an etale space on E through p. The monoid A acts fiberwise on JD' via the action jAα: (jλad) |→ jλa(d.α) for each α in A.
We associate to this etale space the
sheaf Sec(p) of its sections, whose fiber on U is the
set of continuous sections d of p on U.
Definition. A
continuous section d of p on U is called
a (general) distribution on U.
The sheaf D* of distributions is
obtained by enriching the fibers of Sec(p) with a structure of locally convex
spaces and with an action of A.
• For an open set U of E', we denote by D*U
the locally convex space defined as follows: Its elements are the distributions
d
on U; the sum d + d' is the section 'sum' which sends a to d(a) + d'(a). The topology is such that a
family of distributions di
converges to a distribution d if the local jets di(a) converge to d(a) in JD'a
for each a in U, and
this uniformly on each compact of E. An explicit description of its open sets
will be given later on (cf. also my
• The monoid A acts on D*U
via the linear continuous maps D*U(α) sending d to jAα(d).
• For U' an open subset of U, D*(U', U) is the
restriction map sending d to d/U'.
Proposition. D' is
identified to a sub-lc-distructures presheaf of D*by identifying d in D'U
to the section jλd: U
→ JD'. If E is metrizable and E' complete, D*
is a sheaf of complete locally convex spaces.
The
preceding definition of a distribution makes explicit use of the 'points' of E,
via the local jets. We have given this definition because local jets of distributions
have an intrinsic interest; in particular they are used in my
To explicit it, let
us say that two finite order distributions d
in D'U and d' in
D'U' are compatible if they have the same
restriction to the intersection of U and U'. Then a complete compatible family of finite order distributions on U reduces
to a maximal family Δ = (di)i of
finite order distributions di on
Ui which are compatible
and such that U be the union of the Ui.
The 'maximality' implies that it contains with a distribution its restriction
and that a distribution on U' whose restrictions to a cover of U' belong to
Δ also belongs to Δ.
Theorem (Local structure of a distribution). A
distribution on U can be identified
to a complete compatible family Δ = (di)i
of finite order distributions di
on U. If E is finite dimensional, the restriction of a distribution to a bounded
open set of E is of finite order.
Proof. 1. Given such
a family Δ = (di)i the compatibility ensures
that all the di for which Ui
contains a have the same local jet at
a, and the map d sending a on this local jet is a section of the
etale map p: JD' → E. Let us
prove that this section is continuous for the etale topology; indeed, let jλd' (U') be a
neighborhood of d(a); it implies that jλad' = d(a)
= jλadi
and, by definition of a local jet, d' and
di have the same restriction to an open
neighborhood V of a contained in U'.
The image of V by d is then the same as the image of V by jλd', so that it is included in jλd'(U'); thus d is
a continuous section of p on U, hence
it is a distribution on U.
2. Conversely, let d' be a distribution on U
and a an
element of U. By definition of the topology of JD', d'(a) has an open neighborhood of the form jλda(U')
for some finite order distribution da
on U'. Since d' is a continuous section of p, there exists an open neighborhood Ua of a sent by
d'
into jλd'(U'),
and (Ua, da) is a representative of d'(a).
These da for the different a in U are
compatible; indeed, let (db,
Ub) be a representative of
d'(b) and c an element of the intersection Uab of Ua
and Ub. Both da and db admit d'(c) as their local jet at c,
so that there exists a neighborhood Vc of c contained
in Uab on which da
and db have the same
restriction. It follows that da
and db have the same
restriction to the union Uab
of the Vc
so associated to the different elements c
of Ua
The family (da)a generates a complete
compatible family Δ on U, obtained by adding the finite order
distributions on subsets of U compatible with all the da, in particular all the restrictions of an element of Δ
and a distribution on U' whose restrictions to an open cover of U' belong to Δ.
This maximality of Δ ensures that there is exactly one such family
associated to d'. It follows that the map sending d' to Δ defines a
1-1 correspondence between the set D*U of distributions on U
and the set of complete compatible families of D'U.
3. Now we
suppose that E is finite dimensional. Let d be a distribution on U and U' a bounded open subset of U
whose closure is contained in U. We associate to d the complete compatible
family Δ = (di)i, di on Ui. The closure Ū' of U' admits an open covering by the traces of the Ui and, since it is compact
(U' being bounded in the finite dimensional space E) we can extract from it a
finite cover of Ū'; let U'1, ..., U'm
the corresponding finite cover of U' and d'j
the restriction of the distribution dj
to U'j for j = 1,...,m. By the process already used, we can select
representatives of those m distributions dj with the same first term,
say dj = [α, hj]. Since the dj
must have the same restriction to the intersection of two Uj it follows that the hj are restrictions of a same continuous map h on U', and the finite order distribution
[α, h] on U' which glues together the dj must belong to the maximal
family Δ. Thus the
restriction of d to U' identifies to
the finite order distribution [α, h].
Given the
distribution d we'll write d ≈ (di)i,
where (di)i is the corresponding
complete compatible family.
The
general construction of the associated Lcs-sheaf shows that the topology of the
locally convex space D*U of distributions on U is obtained as
follows: For each open subset Ui
of U, let Ai be the set of
distributions d ≈ Δ such that Δ contains an element di on Ui. We equip it with the topology inverse image of the topology
of D'Ui (space of
finite order distributions on Ui)
by the map sending d to di.
The locally convex topology of D*U is the finest topology making
continuous the inclusion of Ai
in D*U for each i. Given
an open cover (Uj)j of U and a convex open
neighborhood Wj of 0 in D'Uj for each Uj, we denote by Ω(Wj)j
the set of distributions d ≈ Δ such that Δ contains a di belonging to Wi for each j. These
sets, where the cover (Uj)j and
Wj vary, form a basis of
neighborhoods of 0 in D*U. We can also describe Ω(Wj)j in terms of local jets: it
is the set of distributions d such that, for each a in Uj the local jet d(a)
belongs to the set jλWj of jets of the elements of
Wj.
d) Comparison with Schwartz distributions
Theorem. If E is
finite dimensional and E' complete, there
is an isomorphism S* from the Lcs-sheaf
D* to the sheaf SD' of
Schwartz E'-valued distributions,
which has for restriction an isomorphism S' from D' to the
presheaf SDf of Schwartz finite order distributions.
Proof. 1. Let DU be the space of the
infinitely differentiable functions φ from E to R with compact
support contained in U, with the compact-open topology for the functions and
their derivatives. If d = [α, g] be an element of D'U of order n. We define
d(φ) = [α, g](φ) = (1/n)∫g(a)(φ.α)(a)da,
where the second member is a E'-valued integral
which exists since E' is complete and φ has a compact support in U. By definition of the equivalence R'
defining the finite order distributions, we deduce (using an integration by
parts) that this is independent of the chosen representative of d. The map Sd: φ |→ d(φ) is linear; it is continuous from DU to E' because the α-derivation is continuous from DU to DU and the integral of g depends continuously (uniformly on each compact) of the
continuous function φ.α (Bourbaki [6]). Thus Sd is a Schwartz E'-valued distribution [20].
2. This extends to any distribution d' on U. Indeed, since φ has a compact support K included in
U, we can find a bounded open neighborhood U' of K included in U. From
the first part of the proposition, the restriction d'/U' of d' to U' is a distribution
of finite order, to which corresponds (d'/U')(φ) defined as above; this
element of E' does not depend on the choice of U', so that we can denote it by d'(φ).
The map Sd': φ |→ d'(φ) is the Schwartz E'-valued distribution associated to d'.
3. We are going to
prove that, for each open set U of E' (not necessarily bounded) we have so
defined an isomorphism S*U from D*U on the locally
convex space SD'U of Schwartz E'-valued distributions
sending d on Sd, whose restriction S'U to
D'U maps this space on the space SDfU of finite order Schwartz distributions
on U.
S'U
is 1-1. Indeed, if d and d' are two finite order distributions
which are different, we can find representatives [α, g] and [α, g'] of d and d' such that g be different from g'. Then
it exists a function φ whose derivative φ.α has its support in
an open subset of U on which g and g' differ, and d(φ) is different from d'(φ), whence Sd is dfferent from Sd'..
The
topology of D'U being the final locally convex topology for
the maps [α, -] from C(U) to D'U, to prove that S'U
is continuous from D'U to SDfU,
it is sufficient to prove that its composite with each of these maps is
continuous; and that comes from the fact that S'U°[α, -] maps g on the above integral, and, for each
φ, this integral depends continuously on g, It ensures
that, if gn converge, the
S([α, gn])
converge in SDfU.
Schwartz
has proved that an E'-valued distribution, where E' is complete, is locally of
finite order ([20], p. 90) and ([20], Proposition 24, p. 86) that a finite
order distribution T on U is the finite sum of (distribution) derivatives βmgm of continuous
functions gm. Since such a
βmgm is
the image by S'U of [βm,
gm], it follows that their
sum T is the image by S'U of the finite order distribution d sum of the [βm, gm].
Thus S'U defines an isomorphism from D'U on SDfU. Whence an
isomorphism S' from D' to the presheaf SDf of Schwartz finite order E'-valued distributions.
Moreover
Schwartz's result implies (though he did not use this terminology) that SD'U is the fiber on U of a
sheaf SD'(E') on E which is the sheaf
associated to SDf(E').
Therefore the isomorphism S' from the Lcs-presheaf D' to SDf extends into an isomorphism
S* between their associated sheaves D* and SD'. In particular it gives an isomorphism S*E from D*E
to the space of Scwhartz distributions SD'
on E.
Remark. Distructures can be applied in many other cases,
for instance a construction similar enough to that we have just done for
distributions allows to define the analog of de Rham's currents on infinite
dimensional manifolds. Other kinds of 'generalized functions' can also be
defined through the construction of appropriate distructures (cf. examples in
my thesis [1]).
References
1. Bastiani(-Ehresmann), A., Différentiabilité dans les espaces localement convexes - Distructures, Thèse d'Etat, Univ, Paris (1962).
2. Bastiani(-Ehresmann), A., Sur les distributions vectorielles, Pub. Labo..
Automatique Théor. Univ. Caen (1964), 32 p.
3. Bastiani(-Ehresmann), A., Systèmes guidables et problèmes d'optimisation, Pub. Labo. Automatique Théor. Univ. Caen: I (1963), 31 p.; II (1964), 47 p.; III (1965), 30 p.; IV (1965), 62 p. Summarized in: Sur le problème général d'optimisation, in Actes Congrès d'Automatique Théorique, Dunod, Paris (1965), 125-136.
4. Bastiani(-Ehresmann), A. Applications différentiables et variétés différentiables de dimension infinie, J. Analyse Math. Jérusalem XIII (1964), 1-114.
5. Bastiani(-Ehresmann),
A. & Ehresmann, C., Multiple functors, Cahiers
Top. et Geom. Diff. XV-3 (1974), 215-292.
Reprinted in [7], Part IV.
6. Bourbaki, N., Intégration, Tome VI, Hermann 1952.
7. Ehresmann A.C. (Edition and Comments), Charles Ehresmann: Oeuvres complètes et commentées, Amiens 1980-83.
8. Ehresmann, A.C. &
Vanbremeersch, J.-P., Memory Evolutive
Systems: Hierarchy, Emergence, Cognition, Elsevier 2007.
9. Ehresmann, A.C. & Vanbremeersch,
J.-P., Memory Evolutive Systems,
http://pagesperso-orange.fr/vbm-ehr/
10. Ehresmann, C., Gattunken von Lokalen Strukturen, Jahresberichte d. Deutschen Math. 60-2 (1957), 49-77. Reprinted in [7], Part I.
11. Ehresmann C;
Catégories topologiques et catégories différentiables, Coll. Géom. Diff.Globales, Bruxelles, C.B.R.M. (1959), 137-150. Reprinted in [7],
Part I.
12. Ehresmann, C., Catégories ordonnées, Holonomie, Cohomologie, Ann. Inst. Fourier X (1964), 205-268. Repinted in [7], Part II.
13. Ehresmann, C., Catégories et Structures, Dunod 1965.
14. Ehresmann, C.,
Expansion des systèmes de structures dominés, CRAS Paris 262 (1966), 8-11. Reprinted in [7], Part III.
15. Eilenberg, S. & Mac Lane, S., General
theory of natural equivalences, Trans.
AMS 58 (1945), 231-294.
16. Farina, M.F. & Meloni, G.C., The
associated sheaf functor, Rend. Ist. Lomb. Milano (1982)
17. Gray, J. W., Formal category theory I, Lecture Notes in Mathematics 391,
Springer (1974)
18. Schröder, L. & Herrlich, H.,
Abstract initiality, Comment .Math.Univ.Carolinae 41,3 .(2000),.575.583.
19. Schwartz, L., Théorie des distributions, Tome I Hermann 1950.
20. Schwartz, L., Théorie des distributions à valeurs vectorielles I, Ann. Inst. Fourier VII (1957), 1-141.
21. Street,
R. & Walters, R., The comprehensive factorization of a functor, Bull. AMS 79 (1973), 936-941.
22. Vaugelade, E., Application des bicatégories à l'étude des catégories internes, Esquisses Math. 21 (1975), Paris.