Algebraic Stochastic Calculus

arXiv:1407.6784v1 [[Link]] 25 Jul 2014

∗
Renaud Gauthier

July 28, 2014

Abstract
We develop the foundations of Algebraic Stochastic Calculus, with
an aim to replacing what is typically referred to as Stochastic Calculus
by a purely categorical version thereof. We first give a sheaf theoretic
reinterpretation of Probability Theory. We regard probability spaces
(Ω, F, P ) as Grothendieck sites (F, τP ) on which Brownian motions
are defined via sheaves in symmetric monoidal ∞-categories. Due
to the complex nature of such a formalism we are naturally led to
considering a purely categorical, time independent formalism in which
stochastic differential equations are replaced by studying problems in
deformation theory.

∗
[Link]@[Link]

1
1 Introduction
Brownian motions display a fractal-like behavior (scale invariance ([L])) and
if one wants to use Brownian motions in modeling the dynamics of financial
variables, this seems to contradict the fact that as time increments become
smaller, the amount of information on the market one can collect should get
smaller, and such information should be made more precise, hence a first
motivation for departing from a purely Stochastic interpretation of Financial
Mathematics. If financial instruments display a fractal-like behavior, one
would want to resolve those phenomena. In order to do so one has to first
investigate the formal foundations underlying the standard usage of financial
variables. Our approach aims at explicating what it means to work with a
probability space (Ω, F , P ), what are Brownian motions W defined on them,
and how can one eliminate the sources of uncertainty, all of which contribute
to a certain volatility, and from there eliminate the need for Brownian mo-
tions (also known as Wiener processes). Thus in a first time we deconstruct
probability spaces (Ω, F , P ). We regard Ω as the collection of objects (events)
of a larger ∞-category Ω+ of events with F seen as decorating objects of
Ω. We argue that F can be regarded as a category on which we can put a
Grothendieck topology, either operadic in nature, or induced by the probabil-
ity P itself, giving rise to a site (F , τ ). Then we construct a sheaf-like object
in the category of symmetric monoidal ∞-categories [Lu1] on this site Brow-
nian motions will be sections of. Due to the complex nature of such a sheaf,
we consider instead a time independent formalism by working with Ω+ only
with a Grothendieck topology on it induced by monomorphisms of simplicial
sets. In this manner one can then construct a sheaf of symmetric monoidal
∞-categories whose sections do not have any volatility, thereby providing a
categorical answer to the problem of having to work with Brownian motions.
Alongside this development we also consider the tropical realization [M] of
the usual problem of having a geometric Brownian motion and study what
that entails from the perspective of our formalism.

In the first section we deconstruct probability spaces (Ω, F , P ) by regard-

ing them as a set of objects Ω of some ∞-category Ω+ of events, which we
decorate using a filtered σ-algebra F = {Ft }, and we regard the latter as
an element of Map(R, Cat). We can put on F two distinct Grothendieck
topologies, one induced by focusing on the operadic aspect of F , the other
by taking P into consideration. If we denote by τ the latter Grothendieck

2
topology, we then construct on (F , τ ) a sheaf-like object W in the ∞-category
of ∞-categories on wich we put a symmetric monoidal structure, and which
we regard as a sheaf of Brownian motions. This we do in section 3. In sec-
tion 4, we abandon using σ-algebras and probabilities in favor of a purely
categorical picture where τP is replaced by a Grothendieck topology induced
by monomorphisms on events on a category of events E, and time is taken
into consideration by considering roof diagrams under products ω × ⊳(ω),
⊳(ω) the forward cone at an event ω being the collection of events ω ′ that
can be reached from ω via a morphism ω → ω ′ . What we preserve from the
Brownian motion picture however is the fact that for a suitable Grothendieck
topology τ on E, we consider over (E, τ ) sheaves of symmetric monoidal ∞-
categories for which deformations of objects replace stochastic log differen-
tials in stochastic calculus ([KS1], [KS2], [O]).

2 Algebraic Geometry and Probability The-

ory
The starting point in this work is a random variable X on a probability
space (Ω, F , P ). The log differential of such a variable sometimes depends
on the differential of a Brownian motion. That latter is not well-defined
however. We will consider the definition of such a differential and develop
a formalism in which using such a differential takes its full meaning. In a
first time we will be developing a sheaf theoretic formalism underlying the
usual classical discussion of geometric Brownian motions, and that will serve
as a platform towards a purely categorical approach to solving stochastic
differential equations. This means we have to see probability spaces (Ω, F , P )
as sites and we also have to define the sheaf over those sites Brownian motions
are sections of.

2.1 Probability space as a site

2.1.1 Filtered σ-algebras
The set of all events Ω we regard as the set of objects of an underlying ∞-
category of events. If the word “event” is well-understood in a pedestrian
sense, in our formalism we will model events as simplicial sets. The filtration

3
F we regard as organizing events according to a fashion dictated by the
filtration itself. The filtration is a R-filtration. This gives an obvious map:
F : R → σ-alg(Ω)
t 7→ Ft (1)
where σ-alg(Ω) denotes the set of σ-algebras on a set Ω. Thus F ∈ Map(R, σ-alg(Ω)).
This filtration is increasing: Fs ⊆ Ft for s ≤ t. The manner in which in-
formation is collected however corresponds to summing graded sets; given
Fs for s < t, the information obtained at t relative to lims→t Fs is really
s<t
lims→t Ft /Fs .
s<t

A filtration F = {Ft } on Ω can be represented by an operad ([MSS],

[KM]) on Ω, meaning the collection of information up to time t can be mapped
out using an operad. This we can do as we regard events to follow from an
appropriate collection of previous events in the sense that for a given event ω
there is at least one non-empty collection {ω1 , · · · , ωN } of events along with
an appropriate map ω1 ⊗ · · · ⊗ ωN → ω. We can assume N to be finite; all
the marginal contributions to ω from a possibly infinite collection of events
can be repackaged into a single event ωN . Since events are indexed by N
by having the index t in Ft being real we can by considering successive σ-
algebras Ft arrive at the above picture where N is finite. Let Op(Ω) denote
the set of operads that can be defined on Ω. Thus we have a map:
F : Map(R, Op(Ω)) → Map(R, σ-alg(Ω))
(t 7→ O(t)) 7→ (t 7→ Ft ) (2)
where the map O(t) 7→ Ft gives Ft the structure of a σ-algebra over an
operad O(t) through the collection of maps O(t)(j) ⊗ Ftj → Ft . Thus the
value F(O) of F at an element O of Map(R, Op(Ω)) gives the filtration F
on Ω the structure of a σ-algebra over an operad O. We regard F as a
Map(R, Op(Ω))-σ-algebra as there may be many operads O that yield an O-
algebra in Map(R, σ-alg(Ω)) that corresponds to F . Denote by Op(Ω)R -σalg
the set of such filtered σ-algebras F on Ω.

2.1.2 Nature of the filtration

Time comes into the picture by virtue of having a filtration F that is R-
filtered. This means in particular that for a random variable X on a proba-

4
bility space (Ω, F , P ) for which dX = αXdt + σXdW , dt in that equation is
in reference to the index t ∈ R in F = {Ft }. Further for a Brownian motion
W dependent on t and at least one event ω, dW can be expressed in terms
of dω, the deformation of ω. Suppose W depends on a unique event ω. One
can write dW = adt + bdω, where dt is the same differential as in dX. Note
that due to the Brownian motion nature of W , a can be complicated but the
nature of a does not concern us here. However in Quantum Physics we have
the well-known uncertainty relation ∆E · ∆t ≥ ~/2 ([MT]). If energy trans-
lates into producing events, then ∆ω ·∆t is bounded below, say by a constant
c > 0. Here we make the assumption that from a formal deformation dω one
can get a quantifiable increment ∆ω by using an appropriately defined norm
on Ω. If (dW )2 = dt, (dt)2 = 0 as well as dW dt = dtdW = 0, as is typical
to write in stochastic calculus ([S]), dt in dW and the previous differential
dX cannot be the differential of physical time, otherwise dW − adt = bdω,
dt = b2 dω · dω and we would have ∆t · ∆ω = b2 (∆ω)3 could be as small as
we want.

Thus the indexing set R for the filtration does not correspond to a physical
time, so the gathering of information dictates what is t in Ft as a function
of physical time hence dictates how time must be contracted to obtain the
indexing set for F . Events that continuously depend on physical time and
are measured as such correspond to deformations of one given event, and
events that do not continuously depend on physical time are measured on a
discrete basis, so in all cases the collection of events corresponds to giving a
discretization of the real line R corresponding to physical time, and if mod
out that space by N we get a quotient map R → R/N with toroidal fibers, the
fibers corresponding to intervals in time between collection times that come
into the definition of individual σ-algebras Ft , t ∈ R. Here we are considering
that collection times are discrete. This can be pictured geometrically by
considering a framing on R and accordingly we denote this quotient map by
Rf . Note that if it is convenient to regard collection times as equivalent to
having a quotient map R → R/N, it is far better geometrically to instead
regard the situation as having a (0, 1]-bundle over R, the base corresponding
to the indexing set for F . We will still denote this bundle by Rf . To resolve
the ambiguity in the manner in which events are collected together between
consecutive collection times, one can extend F to be defined on fibers of
Rf hence we define the extension F f of F from a Map(R, Op(Ω))-σ-algebra
f
to a Map(Rf , Op(Ω))-σ-algebra. Denote by Op(Ω)R -σ-alg the set of such

5
algebras. Thus q : Rf → R induces:

q ∗ : Map(R, σ-alg(Ω)) → Map(Rf , σ-alg(Ω))

F 7→ q ∗ (F ) (3)

and more generally:

F f : Map(Rf , Op(Ω)) → Map(R, σ-alg(Ω)) (4)

is the extension of F whose morphisms between events in the fibers of Rf → R

are dictated by filtered operads in Map(Rf , Op(Ω)).

2.1.3 Probabilities on filtered σ-algebras

Note that working with (Ω, F , P ) is really the same as defining Pt on F f (−, t) =
Ftf and:
P = lim Pt (5)
→

P as used in Finance is a sub-homomorphism:

f
P : (F∞ , π ◦ multO ) → ([0, 1], +) (6)
f
where multO is the multiplication on O, an operad that defines F∞ and π
f
is the forgetful map that projects the O-σ-algebra F∞ to its underlying set.
f
We have for distinct ωi ∈ F∞ , 1 ≤ i ≤ j:
a X
P (π ◦ multO (ω1 , · · · , ωj )) = P ( ωi ) = P (ωi ) (7)
i i

For generic ωi ’s however:

X
P (π ◦ multO (ω1 , · · · , ωj )) = P (∪i ωi ) ≤ P (ωi ) (8)
i

If we denote Sub-Hom(F f , [0, 1]) by (F f )∧ , then we can represent this for-

malism with a diagram:
colim
Ftf −−−→ F∞f
 
∧y
 ∧
y (9)

Pt ∈ (Ftf )∧ −−−→ (F∞

f ∧
) ∋P

6
 f f ∧
Then a real valued random variable X is an element of Mor((F∞ ×(F∞ ) ), R) =
f f ∧ ∧
(F∞ × (F∞ ) ) .

Now P is really defined by its value on connected components of the

f f
Op(Ω)R -σ-algebra F∞ . Indeed, two events that are unrelated in a loose sense
cannot be compared, hence for two such events ω and ω ′ saying P (ω) < P (ω ′)
is not helpful. However for two events ω and ω ′ that are elements of a same
connected component saying P (ω) < P (ω ′) does correspond to a useful state-
f
ment. This means in particular P is not defined on the set underlying F∞
Rf f
but on the Op(Ω) -σ-algebra F∞ itself on which one has to define a notion
of connected components. Keeping the operadic picture in mind we regard
Ω as the set of objects of an ∞-category Ω+ of events and higher morphisms
between them. Further if we denote by Ev the category of ∞-categories of
events, we take Ω+ to be a universal object in that category in the sense that
Ob(Ω+ ) is the set of all events and for any two events ω and ω ′ if there is
a map f : ω → ω ′ then f ∈ MorΩ+ (ω, ω ′). This leads to defining an exten-
sion of F f in the following sense. An element F of Map(R, σ-alg(Ω)) is now
seen as a decoration of objects of Ω+ via a map Map(R, σ-alg) → Ω+ . Such
maps form what we denote by σdec(Ω) and which we refer to as the set of
σ-decorations of Ω. This carries over to the set of F f ’s. Each element of that
set is seen as a decoration over an operad in Map(Rf , Op(Ω+ )) and the set
f
of such elements we denote by Map(Rf , Op(Ω+ ))-σdec or Op(Ω+ )R -σdec for
short.

We have:
f
P ∈ (Op(Ω+ )R -σdec)∧ (10)
f
Finally a Brownian motion W is an element of (Op(Ω+ )R -σdec×Op(Ω+ )Rf -σdec
f
Op(Ω+ )R -σdec∧ )∧ . If we tentatively let H be the space in which Brownian
motions are valued, then observe that W∗ P = P (W −1) defines a push-forward
measure on H.

2.1.4 Filtered σ-algebras as sites

f
Now fix some F f ∈ Op(Ω+ )R -σdec.
f
Proposition [Link]. F f ∈ Op(Ω+ )R -σdec is an element of Map(R, Cat).

7
Proof. For each t ≥ 0, Ftf provides a decoration of Ob(Ω+ ) thereby defining a
category Ftf whose objects are Ftf -decorated elements of Ob(Ω+ ), morphisms
are morphisms of Ω+ between Ftf -decorated elements if any, id = idΩ+ , com-
position is associative, the identity is a right and left unit, and this for all
t ≥ 0, making F f = {Ftf }t≥0 an element of Map(R≥0 , Cat).

Definition [Link]. We define connected components of F f to be collections

of levelwise F f -decorated connected components of Ω+ , t ≥ 0, of the form
{{C(t)} | t ≥ 0} where C(t) is a Ftf -decoration of a connected component
C of Ω+ .

Definition [Link]. For S a set, f ∈ Map(S, Cat), we call a collection

τ = {τ (u)}u∈S , τ (u) a Grothendieck topology on f (u) ∈ Cat for all u ∈ S, a
filtered Grothendieck topology τ on f .

We will now define a filtered Grothendieck topology on F f that ties in

the underlying ∞-category Ω+ with the operadic aspect of F f .

Levelwise products in a filtered operad defining F f define an obvious fil-

tered Grothendieck topology τ on F f as we will see. We define t-coverings of
an event ω to be given by a collection Cov(t) (ω) = {ω ′ → ω} where ω ′ is in the
same connected component C(t) of Ftf ω is in and there is a multiplication
in an operad generating Ftf one of whose summands is ω ′ and whose target
is ω. Such coverings will define a Grothendieck topology τ (t) the collection
of which assembles into a filtered Grothendieck topology τ = {τ (t)}t≥0 .

Definition [Link]. For S a set, f ∈ Map(S, Cat), τ a filtered Grothendieck

topology on f , we call (f, τ ) a filtered site.

Proposition [Link]. τ defines a filtered Grothendieck topology on F f ∈

f
Op(Ω+ )R -σdec making (F f , τ ) into a filtered site.

Proof. If ω ′ → ω is an isomorphism then {ω ′ → ω} is a t-covering for all t.

If {ωi → ω} is a t-covering, γ → ω is any arrow in Ftf , by the universality
of Ω+ since ωi ×ω γ is an event there is a morphism ωi ×ω γ → γ in Ω+ .
It suffices now to show ωi ×ω γ ∈ Ftf and there is a product in an operad
defining Ftf sending ωi ×ω γ to γ. Since Ftf is a σ-algebra if ωi and γ
are in Ftf so is ωi ∪ γ. If ωi and γ are distinct then ωi ∪ γ ≃ ωi × γ ∈ Ftf ,
assembled at that point. In particular there is at some time s ≤ t a morphism

8
ωi ×ω γ → ωi ×γ, so ωi ×ω γ ∈ Fsf , in particular it is in Ftf . If ωi = γ however,
ωi ×ω ωi = ωi × ωi , which as an object is equivalent to considering ωi . In all
cases then ωi ×ω γ ∈ Ftf . Since we have maps ωi → ω and γ → ω, we have
an operad giving a map ωi ⊗ γ → ω, and a universal such map we can find
in an operad O(s) for s ≤ t with a commutative diagram:

ωi ×ω γ −−−→ γ
 
(11)
 
y y
ωi −−−→ ω

The top map ωi ×ω γ → γ in in O(s), so it yields a morphism in Fsf ,

in particular it is in Ftf , and consequently we have an operad O(t) giv-
ing rise to a map ωi ×ω γ → γ. Thus {ωi ×ω γ → γ} is an element of
Cov(t) (γ). Finally if {ωi → ω} is a covering Cov(t) (ω), {ωij → ωi } is a cover-
ing Cov(t) (ωi ) then {ωij → ωi → ω} is a covering Cov(t) (ω). Thus τ (t) defines
a Grothendieck topology on Ftf , and this for all t, hence τ = {τ (t)}t≥0 is a
filtered Grothendieck topology on F f = {Ftf }t≥0 .

2.2 Probability space as a site - version 2

What we have achieved in the previous subsection is really get an enhanced
site, a category F f with a Grothendieck topology and a probability P defined
on it as well. In the present subsection we use P to define a Grothendieck
topology τP on F f instead of trying to tie in the operadic picture. This
has the advantage of working with (F f , τP ) instead of (F f × (F f )∧ ) as it is
f
implied in Finance. The Grothendieck topology τP on F f ∈ Op(Ω+ )R -σdec
induced by P is defined as follows. t-coverings of ω ∈ F f are defined lev-
elwise by Cov(t) (ω) = {ω ′ → ω | ω ′ ∈ C(t), P (ω) ≥ P (ω ′)} with ω ∈ C(t).
We will first show those define a Grothendieck topology on Ftf , and then
we will define τP = {τP (t)}t≥0 as giving a filtered Grothendieck topology on
F f = {Ftf }t≥0 ∈ Map(R, Cat).

Proposition 2.2.1. τP defines a filtered Grothendieck topology on a given

f
F f ∈ Op(Ω+ )R -σdec making (F f , τP ) into a filtered site.

Proof. If ω ′ → ω is an isomorphism, then P (ω ′) = P (ω) so {ω ′ → ω} is a

covering for all t. If {ωi → ω} is a t-covering, γ → ω any arrow in Ftf , since

9
P (ωi ×ω γ) ≤ P (ωi × γ) ≤ P (γ), ωi ×ω γ → γ is a morphism in Ω+ , ωi ×ω γ
is an element of Ftf , then ωi ×ω γ → γ is a morphism in Ftf as observed
before, so that {ωi ×ω γ → γ} is a t-covering Cov(t) (γ). Finally if {ωi → ω}
is a t-covering and so is {ωij → ωi } then the morphism ωij → ωi → ω is a
morphism in Ftf and P (ωij ) ≤ P (ωi ) ≤ P (ω) so that P (ωij ) ≤ P (ω) and thus
{ωij → ωi → ω} is a t-covering, hence we have a Grothendieck topology τP (t)
on Ftf , and this for all t, so we can assemble the τP (t) into τP = {τP (t)}t≥0 ,
a filtered Grothendieck topology on F f .
We will denote (F f , τP ) by (C, τ ) for simplicity and the probability will be
implied. This will be preferred over working with (F f , τ ) from the previous
subsection which is actually the object being implied in Quantitative Finance.

2.3 Time independent formalism

As we will argue in the next subsection, much of the uncertainty arising from
f
having a time dependency in Op(Ω+ )R -σdec leads to having a volatility that
can be eliminated by just considering a time independent base site. This
also eliminates the problem of having to work with Brownian motions. Thus
f
instead of considering an element F f of Op(Ω+ )R -σdec we consider instead
Ω+ made into a site by taking a Grothendieck topology very much like the
one induced by the probability P in the previous section. Then random
variables will be upgraded to being sections of some sheaf over Ω+ . Events
being simplicial sets we call a morphism ω ′ → ω a structural morphism if it
is a monomorphism of simplicial sets. We define coverings to be collections
of such monomorphisms. Let τ be the collection of such coverings.

Proposition 2.3.1. τ defines a Grothendieck topology on Ω+ making it into

a site.

Proof. If ω ′ → ω is an isomorphism then {ω ′ → ω} is trivially a covering. If

{ωi → ω} is a covering, γ → ω any morphism, then as before ωi ×ω γ is a
simplicial set, so we have a morphism ωi ×ω γ → γ in Ω+ and such a morphism
is a structural morphism as well, and this for all i ∈ I so {ωi ×ω γ → γ} is
a covering of γ. Finally if {ωi → ω} is a covering, {ωij → ωi } is a covering,
then so is {ωij → ωi → ω}.

10
3 Construction of the sheaf of Brownian mo-
tions
The problem we are now facing is that of determining in what space, loosely
speaking, are Brownian motions X on (Ω, F , P ) defined. As pointed out
f
above it is necessary to consider (C, τ ) instead with C = F f ∈ Op(Ω+ )R -σdec
and τ being given by the topology induced by P .

3.1 Differentials
In a first time observe that it is typical in Quantitative Finance to consider
the differential of a stock price S being given by dS = αSdt + σSdW ([S]),
where W is a Brownian motion. This means S itself depends on W in addition
to depending on time. Note that this notation is actually misleading. The
differential of a Brownian motion is ill-defined and is instead defined by
considering for a given interval a subdivision Π = {ti }1≤i≤n of that interval.
We then define as in [S]:
X
dW = lim ∆i W (12)
kΠk→0
i

where ∆i W = W (ti+1 ) − W (ti ). Writing ∆ without index and W (ti+1 ) as

Wi+1 observe that this implies:
∆X · ∆Y = (Xi+1 − Xi )(Yi+1 − Yi )
X∆Y = Xi (Yi+1 − Yi )
∆X · Y = (Xi+1 − Xi )Yi (13)
summing those lines we end up with Xi+1 Yi+1 − Xi Yi = ∆(XY ). Summing
over the index i we get d(XY ) = XdY + dX · Y + dX · dY as it is usual in
stochastic calculus ([S]). This means in particular this differential on Brow-
nian motions is not a derivation in the usual algebraic sense, so henceforth
we will denote it by δ instead of using d. Accordingly we will write:
DS = αSdt + σSδW (14)
An immediate question is whether we have δ 2 = 0. We will answer this ques-
tion shortly.

11
 Note also that when one develops Ito’s formula ([I]) for differentiating the
function of a Brownian motion, one writes:
1
df (t, W ) = ∂1 f dt + ∂2 f dW + ∂22 f (dW )2 (15)
2
so one assumes we can have a Taylor expansion of f for which loosely speak-
ing dW dt = 0 and (dW )2 = dt. This subsumes that for the purposes of doing
a Taylor expansion δW is considered to be a finite difference, which it is by
definition.

At this point we can perform a computation which will be needed later.

For X a random variable whose differential is given by DX = αXdt+σXδW
then using the above differential rule we have:
1 1 1
D log X = dX − (dX)2
X 2 X2
1 2 2
= αdt + σδW − σ X dt
2X 2
1
= (α − σ 2 )dt + σδW (16)
2

3.2 Construction of a sheaf of Brownian motions

Recall that σ stands for the volatility of S. If W itself is a function of events,
it has its own volatility. In probability events are well-defined. Not so in
Finance. Either an event ω is trivial, in which case its complement ω c is
infinitely complex, or ω itself is infinitely complex, hence a function of events
is bound to be highly volatile, a reflection that in practice events are not
fully defined and there is consequently a large room for error which we can
accomodate by seeing W itself as having a volatility, which is passed on to
the price of a stock, which itself adds some volatility of its own. Indeed as
dt → 0 one should be able to resolve events, hence resolve the uncertainty on
W , but saying W is a Brownian motion does not allow that. As a matter of
fact the curve for W cannot be resolved ([L]). Thus in developing a theory of
Brownian motions, this is one salient feature we have to incorporate in our
formalism.

As just mentioned W depends on events so W is really a section of some

space over (C, τ ) which we will argue is a sheaf in a sense to be precised, and

12
we regard δω to be a deformation of ω along a path in the base. We would
like to understand what δW means. For X and Y two Brownian motions we
have the well-known result ([S]):

δ(XY ) = XδY + δX.Y + δXδY (17)

A first observation is that we have elements of the form XδY with X and
δY objects of a different nature so we should really write that as X ⊗ δY .
That rule just given is typically taken as a formula when actually it is just
a consequence of the definition of δX for X a Brownian motion. Recall that
δX is defined as the limit of ∆X when the subdivision size goes to zero, and
∆X we regard as the boundary ∂X of X where ∂ is the boundary operator
on simplices. Thus δX carries some homological information relative to the
space of random variables. Further if we denote by H the space in which
Brownian motions are defined then we should write XY = X ⊗ Y with ⊗ a
tensor product on H. Without loss of generality we can put a structure of a
symmetrical monoidal category on H.

Collecting things together we first define a constant sheaf Wt = R valued

in R over Ftf for all t, the collection of which

W = {Wt } ∈ Map(R, Sh(Op(Ω+ )-σdec))

is what we call a filtered sheaf. Now W is not a sheaf as for ω fixed, t < t′ ,
there is no unique object Wt′ (ω) with a morphism Wt′ (ω) → Wt (ω). What
we have instead is a cone at each point Wt (ω) ∈ Wt (ω), a cone with apex
at Wt (ω), based at t′ in Wt′ (ω) in which possible sections of Wt′ (ω) origi-
nating from Wt (ω) are valued, the base of that cone being dependent on the
increment t′ − t as well as the volatility of W . What that means is possible
movements in Wt′ given Wt (ω) are found in the base of that cone. Thus in W
instead of having morphisms in t we have transversal cones. This prompts
us to generalize this formalism as follows.

Definition 3.2.1. For C ∈ Map(R, Cat), τ a filtered Grothendieck topology

on C, D a fixed category, Wt a sheaf in D over C(t) for all t, assembling into
a filtered sheaf W = {Wt } over (C, τ ), if we have transversal cones in time
defined for all W ∈ Γ(C, W) then such a filtered sheaf W we call a sheaf of
Brownian motions.

13
 On the sections of W we define a symmetric monoidal category structure.
Sections of W are regarded as geometric objects that carry some homological
information. In particular we have a boundary map δ on Γ(C, W). The set of
points Wt (ω) ∈ Wt (ω) inherit the monoidal structure from that of Γ(C, W),
but are only used in the definition of the homology of elements of such a
space. Transversally, for ω fixed, t fixed, t′ chosen, δW at (t, ω) with respect
to t′ is Wt′ (ω) − Wt (ω). We can also defined a boundary map in the shaves
Wt as follows: for ω fixed, ω → ω ′ a morphism, δω along that morphism is
defined to be ω ′ − ω, which induces δWt = Wt (ω ′ ) − Wt (ω). Note that this
provides a generalization of δX that is typically not considered in Finance.
We can make the notion of transversal homology more precise as follows:
recall that for a given category C we have the Yoneda embedding j : C →
op
SetC , C 7→ HomC (·, C). For us C is played by the C-diagram in a category D
formed by a functor F : C → D, which we regard as a simplicial set, in which
case denoting this simplicial set by F (C) we still have an ∞-generalization of
op
the Yoneda embedding: F (C) → S F (C) , S the ∞-category of spaces ([Lu2]).
Then for c, c′ ∈ C, F (c′ ) − F (c) 7→ HomF (C) (·, F (c′)) − HomF (C) (·, F (c)) under
the Yoneda embedding. Those presheaves are objects in Fun(F (C)op , S), an
∞-category since S is an ∞-category, hence we can regard F (c′ ) − F (c) as
the boundary of a morphism F (c) → F (c′ ) in Fun(F (C)op , S), which we take
to be F , which leads us to the following definition:
Definition 3.2.2. For C and D two categories, F : C → D a functor, we
define the q-boundary δF of F at c ∈ Ob(C) along a morphism c → c′ in C
to be defined by δF (c) = F (c′ ) − F (c).
Note that if this definition gives a meaning to having a difference F (c′ ) −
F (c), it does not explain how such a difference should be used in practice.
For a difference such as Wt′ (ω) − Wt (ω) Brownian motions being real valued
the difference is a real number. For more general categories D it is really
contextually that one can have a working definition of F (c′ ) − F (c).

For a fixed event ω, Wt (ω) as a function of the index t will display the
characteristics of a Brownian motion. Thus letting Wt (ω) = W (t, ω) for a
Brownian motion W , we picture a cone at W (t, ω) in the (t, W )-plane open-
ing to the right whose opening is dependent upon W ’s volatility. Possible
elements Wt′ (ω) for t′ > t are in that cone. Observe that for t fixed there
are morphisms between events in Ftf , which necessarily means some time
has elapsed when those sets are assembled, while t itself is fixed. This is no

14
contradiction. Recall that t is an index, that we have a bundle Rf → R,
and it is in the fiber of this map that we have morphisms between events.
Thus at t we have an (0, 1]-fiber in time in which morphisms in Ftf occur.
This also means that for two indexes t and t′ , t < t′ , ω fixed, the transi-
tion W (t, ω) → W (t′ , ω) is uncertain as we are missing all the morphisms
between events in the fibers over points s, t < s < t′ , whence the uncertainty
in defining W and Wt , which gives the Brownian motion picture.

3.3 Tropical Geometry realization

As a digression we consider the tropical realization ([M]) of stochastic dif-
ferential equations. If we have DX = αXdt + σXδW , that corresponds to
writing D log X = (α − 21 σ 2 )dt + σδW . Having a log differential invites us to
consider the tropicalization of this equation:
1 1
D log X = “(α − σ 2 )dt + σδW ” = max{α − σ 2 + dt, σ + δW } (18)
2 2
αi ν i ∈ A, αi ∈ C, we define an
P
Definition 3.3.1. For A a C-dga, α =
augmentation homomorphism:

ε:A→C
X
α 99K αi (19)

Then if β = βi ν i ∈ A we define:
P

max{α, β} = max{ε(α), ε(β)} (20)

With this definition we have:

1 1
D log X = max{α − σ 2 + 1, σ + 1} = max{α − σ 2 , σ} (21)
2 2
f
If we work on F f ∈ Op(Ω+ )R -σdec, then this is the problem we are fac-
ing. As an interesting note observe that α and σ are put on the same footing
in the tropical picture, which is not anything new as we have the market
price of risk variable θ = (α − r)/σ ([S]). If we work with Ω+ however, there
is no time dependency, thus no volatility, no Brownian motion and we just
have a linearization D log X = ∆ log X for deformations of events, something

15
we will develop in the last section.

In a categorical setting we need to define the logarithm of objects and

f
morphisms. For X a random variable on (F f , τP ) ∈ Op(Ω+ )R -σdec we
formally define:
X 1
exp(X) = ⊗n X (22)
n≥0
n!
where we would have a symmetrical monoidal structure on a filtered sheaf of
random variables G on (F f , τP ), so we would have a functor

exp : G → G
X 7→ exp(X) (23)

where G is the graded completion of the symmetric monoidal category as-

sociated with G. Then the formal inverse to such a functor would be a log
morphism:

G ← G : log
X L99 exp(X) (24)

Formally one has log X = n>0 (−1)n /n ⊗n X. On morphisms P if X and Y

P
are objects of G and f is a morphism from X to Y then exp(f ) = 1/n!⊗n f
and likewise for log(f ).

4 Of the use of Ω+ to resolve Wiener pro-

cesses
Note that if we have an event ω in a σ-algebra over Ω then ω c = Ω − ω is
also in that σ-algebra. If ω is simple enough, to have a function defined at
ω c one would have to resort to an approximation. Further having filtered ob-
jects as we have seen introduces a further uncertainty in defining transition
maps, so instead of working with filtered σ-algebras and probabilities, one
should work with categories and structural morphisms as done with Ω+ in 2.3.

One can retain a notion of time in this time independent formalism by

introducing roof diagrams such as the one below where ⊳(A), which we call
the forward cone at A, is defined to be the collection of events B that can be

16
reached from A via a morphism A → B in Ω+ . Thus we consider a category
E with Ob(E) = Ob(Ω+ ) and morphisms in E are roof diagrams such as the
one below with base an element of Mor(Ω+ ).

A × ⊳(A)

p1 ✁ ✁ ❆
❆
πB
☛✁✁ ❆❆❯
✲
A B
Identity:

A × ⊳(A)

p1 ✁ ✁ ❆
❆
πA
☛✁✁ ❆❆❯
✲
A A
Composition:

A × ⊳(A)

id × ⊳ ◦ ∆ ◦ p1 ✂ ❆ id × ⊳ ◦ ∆ ◦ πB
✂ ❆
✂ ❆
❆❯
✌✂✂
A × ⊳(A) B × ⊳(B)

✁ ❆ πB ✁✁ ❆ π
✁ ❆ ❆ C
✁
☛
✁ ❆❆❯ ☛✁✁ ❆❆❯
A ✲ B ✲ C

where ∆ is the diagonal map. With this composition being defined we

can verify that the identity defined above functions as an identity:

17
 A × ⊳(A)

id ✂✂ ❆
❆
id
✂ ❆
❆❯
✂✌✂
A × ⊳(A) A × ⊳(A)

p1 ✁ ✁ ❆
❆
✁p
✁ 1
❆
❆
✁✁
☛ id ❆❆❯ ☛✁✁ ❆❆❯
A ✲ A ✲ B

which simplifies as:

A × ⊳(A) A × ⊳(A)

id ✂✂ ❆
❆
id ❆
❆
id
✂ ❆ ❆
❆❯ ❆❯
✌✂✂
A × ⊳(A) A × ⊳(A) = A × ⊳(A)

❆ ✁p ❆ p1 ✁ ✁ ❆
❆ ✁ 1 ❆ ❆
❆❆ ✁
❯ ☛
✁ ❆❆❯ ☛✁✁ ❆❆❯
A ✲ B A ✲ B
which further simplifies as:

A × ⊳(A)

✁p ❆
✁ 1 ❆
☛✁✁ ❆❆❯
A ✲ B
thereby showing that the identity as we defined it acts as a left identity
and likewise one would show that it also acts as a right identity.

Associativity:

18
 A × ⊳(A)
✟✟
✟
✟
✙
✟ ❈
❈
A × ⊳(A) ❈
❈
✂ ❆ ❈
✂ ❆ ❈
✂ ❆ ❈
❆❯ ❈
✂✌✂ ❈❲
A × ⊳(A) B × ⊳(B) C × ⊳(C)

✁ ❆ ✁ ❆ ✁ ❆
✁ ❆ ✁ ❆ ✁ ❆
☛✁
✁ ❆❆❯ ✁☛✁ ❆❯❆ ✁☛✁ ❆❆❯
A ✲ B ✲ C ✲ D

and for the associated composition:

A × ⊳(A)
◗
◗
✄ ◗◗
s
✄
✄ B × ⊳(B)
✄
✄ ✂ ❇
✄ ✂ ❇
✄ ✂ ❇
✄
✄✎ ✌✂✂ ❇❇◆
A × ⊳(A) B × ⊳(B) C × ⊳(C)

✁ ❆ ✁ ❆ ✁ ❆
✁ ❆ ✁ ❆ ✁ ❆
✁✁
☛ ❆❆❯ ✁☛✁ ❆❯❆ ✁☛✁ ❆❆❯
A ✲ B ✲ C ✲ D

We have associativity by virtue of the equality πB→C ◦ πA→B = πA→C .

We put on E a Grothendieck topology generated by roof diagrams with base
structural morphisms - which we call structural roof diagrams - as introduced
in section 2.3. We verify that this does indeed give such a Grothendieck topol-
ogy. For coverings we take collections of structural roof diagrams {ω ′ → ω}.
If we have an isomorphism ω ′ → ω then {ω ′ → ω} is a covering. If {ωi → ω}

19
is a covering and γ → ω is any morphism for all i the projections ωi ×ω γ → γ
are structural morphisms, so all such morphisms form the base of a covering
of γ. Finally if {ωi → ω} is a covering, and so are the {ωij → ωi } then so is
the composition ωij → ωi → ω, so the collection of all such compositions is a
covering of ω. Denoting by τ the Grothendieck topology thus defined, (E, τ )
is a site.

Finally we consider the category of sheaves W : (E, τ )op → Cat⊗ ∞ of

symmetric monoidal ∞-categories. Since there is no time dependence, no
probability, no σ-algebra, all sources of uncertainties are eliminated.

Finally for X a section of such a sheaf W, ω ∈ E, we let ω o be the set of

morphisms ω → ω ′ in a same connected component C ω is in for which there
is no ω ′′ ∈ C such that ω → ω ′′ → ω. Then for ψ : ω → ω ′ ∈ ω o , we define:

Dψ X(ω) = X(ω ′) − X(ω) (25)

If one can define quotients in W then one would define this difference as being
X(ω ′)/X(ω). Note that if that solves the problem of having to solve a log dif-
ferential equation in Stochastic calculus, this is not very illuminating however
as it is quite a simple equation. What is most important perhaps is how X
behaves in between ω and ω ′ relative to other sections of W between the same
objects, and this from a homological perspective, so one could impose that
our sheaves be sheaves of stable symmetric monoidal ∞-categories ([Lu3])
and study their sheaf cohomology. To be complete one should also study
such a sheaf as an object of the topos of symmetric monoidal ∞-categories
([Lu2]) on E.

References
[I] K. Ito, Stochastic Integrals, Proc. Imperial Acad. Tokyo 20, 519-524
(1944).

[KS1] I. Karatzas, S.E. Shreve, Brownian Motion and Stochastic Calculus,

Springer-Verlag, New York (1991).

[KS2] I. Karatzas, S.E. Shreve, Methods of Mathematical Finance, Springer-

Verlag, New York (1998).

20
[KM] I. Kriz, P. May, Operads, Algebras, Modules and Motives, Soc. Math.
de France (1995).

[Lu1] J. Lurie, Higher Algebra, [Link]/ lurie/papers/[Link].

[Lu2] Jacob Lurie, Higher Topos Theory, Ann. of Math. Studies, 170 (2009).

[Lu3] J. Lurie, Stable ∞-Categories, arXiv:[Link]/0608228.

[L] G.F. Lawler, Geometric and Fractal Properties of Brownian Motions

and Random Walk Paths in two and three Dimensions, Random Walks
219-258, Bolyaj Society Math. Stud., Vol. 9 Budapest (1999).

[MT] L.I. Mandelshtam, I.E. Tamm, The Uncertainty Relation between En-
ergy and Time in Non-relativistic Quantum Mechanics, J. Phys (USSR),
9, 249-254 (1945).

[MSS] M. Martin, S. Shnider, J. Stasheff, Operads in Algebra, Topology and

Physics, Am. Math. Soc. (2002).

[M] G. Mikhalkin, Tropical Geometry and its Applications,

arXiv:[Link]/0601041.

[O] B. Oksendal, Stochastic Differential Equations, 4th Ed., Springer-Verlag,

New York (1995).

[S] S. Shreve, Stochastic Calculus for Finance II: Continuous-time Models,

Springer Finance (2013).

21