Beruflich Dokumente
Kultur Dokumente
Driver
Contents
5
6
p
2.1
2.2
2.3
2.4
2.5
9
10
13
14
15
18
21
21
23
25
27
28
34
43
48
51
51
53
54
Contents
59
59
60
61
64
Homogeneous Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1 Lie group p variation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Homogeneous Metrics on G (V ) and Ggeo (V ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Carnot Caratheodory Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
69
70
73
77
77
79
83
88
Rough ODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
9.1 Local Existence and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
9.2 A priori-Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Page:
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Part I
Here are a few suggested references for this course, [12,15,1]. The latter two
references are downloadable if you are logging into MathSci net through your
UCSD account. For a proof that all p variation paths have some extension
to a rough path see, [14] and also see [6, Theorem 9.12 and Remark 9.13]. For
other perspectives on the the theory, see [3] and also see Gubinelli [7, 8] Also
see, [9, 4, 7] look interesting. A recent paper which deals with global existence
issues for non-bounded vector fields is Lejay [11].
1
From Feynman Heuristics to Brownian Motion
1
Zt
In the physics literature one often finds the following informal expression,
dT () =
RT 0
2
1
1
e 2 0 | ( )| d DT for WT ,
Z (T )
(1.1)
Z
=
= C (s, t)
s
( (t) (s)) and ( ) := ( ) ( ) ,
ts
(t)(s)
,
ts
( ) ( ) d =
s
1 Es ()
Z Z
1 |y(s)|2
C (s, t)
e 2
F |[0,s] , y e 2 ts dy
D[0,s]
Zt
Z (s)
Rd
Z Z
2
1 |y(s)|
C (s, t)
=
F (, y) e 2 ts dy ds () .
Zt
Ws
Rd
0 ( ) 0 ( ) d
(t) (s)
( (t) (s)) = 0.
ts
| (t) (s)|
+ E[s,t] () .
ts
Thus if f () = F |[0,s] , (t) , we will have,
=
e 12 Es () 1 |(t)(s)|2
ts
F |[0,s] , (t)
e 2
.
Z (s)
D(s,t)
Wt
|(t)(s)|2
+E[s,t] ()
Es ()+
ts
Multiplying this equation by Z1t D[0,s] d (t) and integrating the result then
implies,
Z
F |[0,s] , (t) dt ()
If we decompose ( ) as ( ) + ( ) where
then we have, 0 ( ) =
21
F |[0,s] , (t) e
Z
( ) := (s) +
1
F |[0,s] , (t) e 2 Et () Dt
Wt
Z
1
1
F |[0,s] , (t) e 2 [Es ()+E[s,t] ()] Dt
=
Z t Wt
and now fixing |[0,s] and (t) and then doing the integral over |(s,t) implies,
Z
1
F |[0,s] , (t) e 2 [Es ()+E[s,t] ()] D(s,t)
0<tT
(1.2)
Thus the heuristic expression in Eq. (1.1) leads to the following Markov property for t , namely.
Proposition 1.1 (Heuristic). Suppose that F : Ws Rd R is a reasonable
function, then for any t s we have
F |[0,s] , (t) dt ()
Wt
Z Z
F (, y) pts ( (s) , y) dy ds () ,
=
Ws
Rd
and
p
ps (x, y) :=
1
2 (t s)
d/2
e
|yx|2
12 ts
Z
WT
(Rd )n
f (y1 , . . . , yn )
n
Y
i=1
(1.4)
where by convention, y0 = 0.
Theorem 1.3 (Wiener 1923). For all t > 0 there exists a unique probability measure, t , on Wt , such that Eq. (1.4) holds for all n and all bounded
n
measurable f : Rd R.
Definition 1.4. Let Bt () := (t) . Then {Bt }0tT as a process on (WT , T )
R
is called Brownian motion. We further write Ef for WT f () dT () .
The following lemma is useful for computational purposes involving Brownian motion and follows readily form from Eq. (1.4).
Lemma 1.5. Suppose that 0 = s0 < s1 < s2 < < sn = t and fi : Rd R
are reasonable functions, then
" n
#
n
Y
Y
E
fi Bsi Bsi1 =
E fi Bsi Bsi1 ,
(1.5)
i=1
i=1
E [f (Bt Bs )] = E [f (Bts )] ,
and
E [f (Bt )] = Ef
tB1 .
p/2
(1.8)
Z
f ( (s1 ) , . . . , (sn )) dT () =
E [|B1 | ] = Cp |t s|
(1.3)
Corollary
1.2 (Heuristic). If 0 = s0 < s1 < s2 < < sn = T and f :
n
Rd R is a reasonable function, then
Page:
p/2
E [|Bt Bs | ] = |t s|
where
job:
rpaths
(1.6)
(1.7)
In this section we sketch one method of constructing Wiener measure ore equivalently Q
Brownian motion. We begin with the existence of a measure T on the
T :=
W
0sT R which satisfies Eq. (1.4) where R is a compactification of R
i=1
R
along with the fact that Rd pt (x, y) dy = 1 for all t > 0, one shows that
I (F ) is well defined independently of how we represent F as a finitely based
continuous function.
T is a compact Hausdorff space. By the Stone
By Tychonoffs Theorem W
Weierstrass
Theorem,
the
finitely
based continuous functions are dense inside
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
for all s,
t [0, T] . Then for any (0, /) there is a modification, X, of X
t = 1 for all t) which is H
(i.e. P Xt = X
older continuous. Moreover, there
is a random variable K such that,
(Xt , Xs ) K |t s|
and EKp < for all p <
(1.10)
1 .
t : W
T R be the projection map, B
t () = (t) . Then
Corollary 1.8. Let B
n o
t for which t Bt is H
there is a modifications, {Bt } of B
older continuous T almost surely for any (0, 1/2) .
Proof. Applying Theorem 1.7 with := p n
and o := p/2 1 for any p
t which is almost surely
(2, ) shows there is a modification {Bt }t0 of B
Holder continuous for any
p/2 1
= (0, 1/2 1/p) .
(0, /) = 0,
p
Letting p shows that {Bt }t0 is almost surely Holder continuous for
all < 1/2.
We will see shortly that these Brownian paths are very rough. Before we
do this we will pause to develop a quantitative measurement of roughness of a
continuous path.
2
p Variations and Controls
Let (E, d) be a metric space which will usually be assumed to be complete.
Definition 2.1. Let 0 a < b < . Given a partition :=
{a = t0 < t1 < < tn = b} of [a, b] and a function Z C ([a, b] , E) , let
(ti ) := ti1 , (ti )+ := ti+1 , with the convention that t1 := t0 = a and
tn+1 := tn = T. Furthermore for 1 p < let
1/p
!1/p
n
X
X
p
p
Vp (Z : ) :=
d Ztj , Ztj1
=
d Zt , Zt
.
(2.1)
j=1
||0 t
1/p
n
X
dp Ztj , Ztj1 .
Vp (Z) := sup Vp (Z : ) = sup
(2.2)
P(a,b)
P(a,b)
j=1
sup
n
X
dp Ztj , Ztj1 .
(2.3)
Remark 2.3. We can define Vp (Z) for p (0, 1) as well but this is not so interesting. Indeed if 0 s T and P (0, T ) is a partition such that s ,
then
X
X
d (Z (s) , Z (0))
d (Z (t) , Z (t )) =
d1p (Z (t) , Z (t )) dp (Z (t) , Z (t ))
t
1p
max d
(Z (t) , Z (t ))
Vpp
is decreasing in p and
(p) := ln
n
X
!
api
is convex in p.
i=1
i=1
d p
f = f p ln f, it follows that and
Using dp
(f p ln f )
and
(f p )
2
f p ln2 f
(f p ln f )
00 (p) =
.
(f p )
(f p )
0 (p) =
P(a,b) j=1
!1/p
api
i=1
Furthermore, let P (a, b) denote the collection of partitions of [a, b] . Also let
mesh () := maxt |t t | be the mesh of the partition, .
(Z : )
Using the uniform continuity of Z (or d (Z (s) , Z (t)) if you wish) we know that
lim||0 maxt d1p (Z (t) , Z (t )) = 0 and hence that,
10
i
d h
d 1
1
1
ln kf kp =
(p) = 0 (p) 2 (p)
dp
dp p
p
p
1
= 2
[p (f p ln f ) (f p ) ln (f p )]
p (f p )
1
= 2
[ (f p ln f p ) (f p ) ln (f p )]
p (f p )
1
fp
p
= 2
f ln
.
p (f p )
(f p )
Proof. Given Lemma 2.4, it suffices to prove Eq. (2.4) and the continuity assertion on p Vp (Z) . Since p Vp (Z) is a decreasing function,
we know that limpp0 Vp (Z) and limpp0 Vp (Z) always exists and also that
limpp0 Vp (Z) = supp>p0 sup Vp (Z : ) . Therefore,
lim Vp (Z) = sup sup Vp (Z : ) = sup sup Vp (Z : ) = sup Vp0 (Z : ) = Vp0 (Z)
pp0
n
X
ap+r
j
1
p
ln Vp (Z)
follows
directly from the fact that ln Vp (Z) is convex in p and that convex functions
are continuous (where finite).
p
Here is a proof for this case. Let (p) := ln Vp (Z) , 1 p0 < p1 such that
Vp0 (Z) < , and ps := (1 s) p0 + sp1 , then
(ps ) (1 s) (p0 ) + s (p1 ) .
Letting s 1 then implies ps p1 and (p1 ) (p1 ) , i.e. Vp1 Vp1
Vp1 . Therefore Vp1 = Vp1 and along with Eq. (2.4) proves the continuity of
p Vp (Z) .
r X
n
r
p
q
max aj
apj kakp kakp = kakp ,
j
j=1
p>p0
Up to now our computation has been fairly general. The point where being
counting measure comes in is that in this
casei (f p ) f p everywhere and
h
fp
d
therefore ln (f p ) 0 and therefore, dp ln kf kp 0 as desired.
Alternative proof that kf kp is decreasing in p. If we let q = p + r, then
q
kakq
p>p0
j=1
max apj
j
1/p
n
X
1/p
apj
= kakp .
j=1
Remark 2.5. It is not too hard to see that the convexity of is equivalent to
the interpolation inequality,
1s
where 0 s 1, 1 p0 , p1 , and
(a + b)p ap + bp
1
1
1
:= (1 s)
+s .
ps
p0
p1
pp0
Page:
10
job:
rpaths
(2.5)
(2.4)
a+b
a+b
a+b a+b
Moreover the latter inequality is strict if if a, b > 0 and p > 1.
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Vp (x : D0) Vp (x : D) = (x(ti+1 ) x(ti1 ))p (x(ti+1 ) x(ti ))p (x(ti ) x(ti1 ))p
p
p
p
= ti x + ti+1 x (ti x) ti+1 x 0
and the inequality is strict if ti x > 0, ti+1 x > 0 and p > 1.
In other words, on any monotone increasing segment, we should not include
any intermediate points, because they can only hurt us.
Example 2.9. Consider a path like the following: If we partition [0, T ] at the
11
11
job:
rpaths
svmonob.cls
date/time:
11-Mar-2009/12:03
12
Remark 2.13. Here is another way to state the jog-free condition. Let x be
piecewise monotone with extrema s0 , s1 , . . . . Let i = |x(si+1 ) x(si )| . Then
x is jog-free iff 1 2 . . . . The idea is that the oscillations are shrinking.
(Notice that we dont need 0 1 ; this is because s0 = 0 is not required to be
a forward extremum.)
Remark 2.14. It is also okay if s1 , s2 , . . . are backwards extrema; this corresponds to the oscillations getting larger. Just reverse time, replacing x(t) by
x(T t), which again doesnt change the p-variation. Note that if i are as
above, this corresponds to having 0 1 2 . . . (note that 0 is included
now, but m1 would not be). This case seems less useful, however.
Lemma 2.15. Let x be jog-free with extrema s0 , . . . , sm . Let D = {t0 , . . . , tn }
be any partition not containing all the sj . Then there is some sj
/ D such that
if D0 = D {sj }, Vp (x : D0) Vp (x : D).
Proof. Let sj be the first extremum not contained in D (note s0 = 0 D
already, so j is at least 1 and sj is also a forward extremum). Let ti be the last
element of D less than sj . Note that sj1 ti < sj < ti+1 .
Now x is monotone on [sj1 , sj ]; say WLOG its monotone increasing, so
that sj is a local maximum and also a forward maximum. Since ti [sj1 , sj ],
where x is monotone increasing, x(sj ) x(ti ). And since sj is a forward maximum, x(sj ) x(ti+1 ).
Therefore we have
x(sj ) x(ti ) x(ti+1 ) x(ti )
x(sj ) x(ti+1 ) x(ti ) x(ti+1 ).
12
since one of the terms on the left is already the term on the right. This shows
p
p
that Vp (x : D0) Vp (x : D) .
In other words, we should definitely include the extreme points, because
they can only help.
Putting these together yields the desired result.
PropositionP
2.16. If x is jog-free with extrema S = {s0 , . . . , sm }, then Vp (x) =
Vp (x : S) = ( ip )1/p .
Proof. Fix > 0, and let D be a partition such that Vp (x : D) Vp (x) .
By repeatedly applying Lemma 2.15, we can add the points of S to D one
by one (in some order), and only increase the p-variation. So Vp (x : D S)
Vp (x : D). Now, if t D\S, it is inside some interval [sj , sj+1 ] on which x is
monotone, and so by Lemma 2.8 t can be removed from D S to increase the
p-variation. Removing all such points one by one (in any order), we find that
Vp (x : S) Vp (x : D S). Thus we have Vp (x : S) Vp (x : D) Vp (x) ;
since was arbitrary we are done.
Notice that we only considered the case of jog-free paths with only finitely
many extrema. Of course, in order to get infinite p-variation for any p we would
need infinitely many extrema. Lets just check that the analogous result holds
there.
Proposition 2.17. Suppose we have a sequence s0 , s1 , . . . increasing to T ,
where x is alternately monotone increasing and decreasing on the intervals
[sj , sj+1 ]. Suppose also that the sj are forward extrema for x. Letting j =
|x(sj+1 ) x(sj )| as before, we have
Vp (x) =
One of the quantities on the right is equal to |x(ti+1 ) x(ti )|, and so it follows
that
Page:
Here is an example.
job:
rpaths
1/p
jp
j=0
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Actually, the extreme points sj can converge to some earlier time than T ,
but x will have to be constant after
Pm that time.
p
Proof. For any m, we have j=0 jp = Vp (x : D) for D = {s0 , . . . , sm+1 },
P
P
m
tm
Let us define
Cov (A, B) := E [AB] EA EB and
h
i
2
2
Var (A) := Cov (A, A) = EA2 (EA) = E (A EA) .
and observe that
Var
1/p
X
p
Vp (x : D) Vp (x) .
j
13
n
X
!
Ai
n
X
i=1
Var (Ai ) +
i=1
Cov (Ai , Aj ) .
i6=j
j=0
t
2
= Var(N )
(t) Var(N 2 ) |m |
t
2
Corollary 2.18. For all p > 2 and T < , Vp B|[0,T ] < a.s. (We will see
later that Vp B|[0,T ] = a.s. for all p < 2.)
= T Var(N ) |m | 0 as m .
(By explicit Gaussian integral computations,
|Bt Bs | Kp |t s|
for all 0 s, t T.
(2.6)
Thus we have
p X
X
X
p
1/p
|i B|
Kp |ti ti1 |
Var(N 2 ) = EN 4 EN 2
Kpp T
If
< a.s.
13
job:
rpaths
= 3 1 = 2 < .)
2
|m | < , then
"
#
X
X
X
2
2
E
(Qm T ) =
E (Qm T ) =
Var(Qm )
m=1
m=1
m=1
m=1
Var(N 2 ) T
mesh(m ) <
m=1
svmonob.cls
m=1
date/time:
11-Mar-2009/12:03
14
pq
max d
(Z (t) , Z (t ))
dq (Z (t) , Z (t ))
t
Vqq (Z
C [|m |]
T 0 as m
Proposition 2.24. Suppose that Z (t) is a real continuous function such that
Z0 = 0 for simplicity. Define
Z T
Z T
f ( ) dZ ( ) :=
f ( ) Z (t) d + f (t) Z (t) |T0
0
Corollary 2.21. If p < 2, then Vp B|[0,T ] = a.s.
Proof. Choose partitions, {m } , of [0, T ] such that limm Qm = T a.s.
where Qm := V22 (B : m ) and let 0 := {limm Qm = T } so that P (0 ) =
1. If Vp B|[0,T ] () < for then by Proposition 2.20,
lim Qm () = lim V22 (B () : m ) = 0
tm
: )
||0
tm
21
and
hence
,
i.e.
V
B|
()
<
0c . Therefore 0
0
p
[0,T
]
Vp B|[0,T ] () = and hence
P Vp B|[0,T ] () = P () = 1.
(2.8)
then V1 (Z) < (See Definition 2.2 above) and the best possible choice for C
in Eq. (2.8) is V1 (Z) .
Proof. Given a partition, := {0 = t0 < tP
1 < < tn = T } be a partition
n
n
of [0, T ] , {k }k=1 R, and f (t) := 1 1{0} + k=1 k 1(tk1 ,tk ] . Choose fm (t)
1
in C ([0, T ] , R) well approximating f (t) as in Figure 2.3. It then is fairly
1 if p > 2
.
0 if p 2
Page:
14
job:
rpaths
easy to show,
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
2.4 Controls
fm ( ) Z (t) d
n1
X
(k+1 k ) Z (tk )
k=1
and therefore,
Z
m
fm (t) dZ (t) =
lim
n1
X
k=1
n
X
Therefore we have,
Z
n
T
X
fm ( ) dZ ( )
k (Z (tk ) Z (tk1 )) = lim
m 0
k=1
f ( ) Z (t) d =
which still does not make sense as a Youngs integral when B is a Brownian
motion because for any p > 2, p1 + p1 =: < 1. For more on this point view see
the very interesting work of Terry Lyons on rough path analysis, [13].
2.4 Controls
(t) |T0
for all f C 1 . Thus if Vp (Z) < the integral extends to those f C ([0, T ])
such that Vq (f ) < .
Taking
k
=
sgn(Z (tk ) Z (tk1 )) for each k, then shows
Pn
|Z
(t
)
Z
(t
k
k1 )| C. Since this holds for any partition , it
k=1
follows that V1 (Z) C.
If V1 (Z) < , then
T
and
0 T
Therefore C can be taken to be V1 (Z) in Eq. (2.8) and hence V1 (Z) is the best
possible constant to use in this equation.
Combining Fact 2.22
R t with Proposition 2.24 explains why we are going to
have trouble defining 0 fs dBs when B is a Brownian motion. However, one
might hope to use Youngs integral in this setting.
Page:
15
+ 1q =: >
k (Z (tk ) Z (tk1 )) .
k=1
1
p
15
job:
rpaths
(2.9)
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
16
(2.10)
a+b
a+b
a+b
V1 ( : ) =
(2.11)
(t , t) .
(2.13)
We now assume:
1. is continuous,
2. (t, t) = 0 for all t [0, T ] (This condition is redundant since next condition
would fail if it were violated.), and
3. V1 () := (0, T ) := supP(0,T ) V1 ( : ) < .
Under these assumptions, : [0, ) is a control.
(2.12)
1/p
V1 ( : ) ,
d (Zs , Zt ) (s, t)
sup
P(s,t)
We will give the proof of Theorem 2.32 after a Lemma 2.37 below.
Corollary 2.33 (The variation control). Let p [1, ) and suppose that
Z Cp ([0, T ] , E). Then Z,p : [0, ) defined in Eq. (2.3) is a control
1/p
satisfying, d (Z (s) , Z (t)) Z,p (s, t)
for all (s, t) .
Proof. Apply Theorem 2.32 with (s, t) := dp (Z (s) , Z (t)) and observe
that with this definition, Z,p = .
Lemma 2.34. Let : [0, ) satisfy the hypothesis in Theorem 2.32, then
= (defined in Eq. (2.12)) is superadditive.
Proof. If 0 u s v T and 1 P (u, s) , 2 P (s, v) , then
1 2 P (u, v) . Thus we have,
V1 ( : 1 ) + V1 ( : 2 ) = V1 ( : 1 2 ) (u, v) .
Therefore,
p,Z (s, t) := Vpp Z|[s,t] =
sup
P([s,t])
i.e. is superadditive.
Page:
16
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
2.4 Controls
17
Lemma 2.35. Let Z Cp ([0, T ] , E) for some p [1, ) and let := Z,p :
[0, ) defined in Eq. (2.3). Then is superadditive. Furthermore if p = 1,
is additive, i.e. Equality holds in Eq. (2.9).
Proof. The superadditivity of Z,p follows from Lemma 2.34 and since
Z,p (s, t) = (s, t) where (s, t) := dp (Z (s) , Z (t)) . In the case p = 1, it is
easily seen using the triangle inequality that if 1 , 2 P (s, t) and 1 2 ,
then V1 (X : 1 ) V1 (X, 2 ) . Thus in computing the supremum of V1 (X : )
over all partition in P (s, t) it never hurts to add more points to a partition.
Using this remark it is easy to show,
(u, s) + (s, v) =
sup
[V1 (X : 1 ) + V1 (X : 2 )]
1 P(u,s),2 P(s,v)
V1 (X : 1 2 )
sup
1 P(u,s),2 P(s,v)
sup
V1 (X : ) = (u, v)
P(u,v)
The second and third case are similar. For example if u < t0 , we will have,
as desired.
(2.14)
t 0
In particular, if (, ) 2 then
(, ) V1 ( : [, ]) + .
With these lemmas as preparation we are now ready to complete the proof
of Theorem 2.32.
Proof. Proof of Theorem 2.32. Let (s, t) := (s, t) be as in Theorem
2.32. It is clear by the definition of , the (t, t) = 0 for all t and we have
already seen in Lemma 2.34 that is superadditive. So to finish the proof we
must show is continuous.
Using Remark 2.27, we know that (s, t) is increasing in t and decreasing in s and therefore (u+, v) = limsu,tv (s, t) and (u, v+) =
lims u,,tv (s, t) exists and satisfies,
(2.15)
(u+, v) (u, v) (u, v+) .
(2.16)
The main crux of the continuity proof is to show that the inequalities in Eq.
(2.16) are all equalities.
1. Suppose that > 0 is given and > 0 is chosen as in Lemma 2.37 and
suppose that u < s < t < v with |s u| < and |v t| < . Further let
P (u, v) be a partition of [u, v] , then according to Lemma 2.37,
V1 ( : ) V1 ( : {s, t}) + 2
= (u, s) + (t, v) + V1 ( : [s, t] {s, t}) + 2
(u, s) + (t, v) + (s, t) + 2.
t 0
(a, b) V1 ( : ) = .
17
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
18
V1 ( : ) (u+, v) + 2
and then taking the supremum over P (u, v) and then letting 0 shows
(u, v) (u+, v) . Combined this with the first inequality in Eq. (2.16)
shows, (u+, v) = (u, v) .
2. We will now show (u, v) = (u, v+) by showing (u, v+) (u, v) .
Let > 0 and > 0 be as in Lemma 2.37 and suppose that s < u and t > v
with |u s| < and |t v| < . Let us now choose a partition P (s, t) such
that
(s, t) V1 ( : ) + .
Proof. Let (s, t) := p,x (s, t) = Vpp x|[s,t] be the control associated to x
and define h (t) := (0, t) . Observe that h is increasing and for 0 s t T
that h (s) + (s, t) h (t) , i.e.
(s, t) h (t) h (s) for all 0 s t T.
Let g : [0, h (T )] E be defined by g (h (t)) := x (t) . This is well defined
since if s t and h (s) = h (t) , then (s, t) = 0 and hence x|[s,t] is constant
and in particular x (s) = x (t) . Moreover it now follows for s < t such that
u := h (s) < h (t) =: v, that
(s, t) V1 ( : 1 ) + 3
where 1 = {u, v} . As above, let u and v+ be the elements in 1 just
before u and just after v respectively. An application of Lemma 2.36 then shows,
(u, v+) (u , v+ ) V1 ( : 1 [u , v+ ]) + 3
= V1 ( : 1 [u, v]) + (u , u) + (v, v+ ) + 3
(u, v) + 5.
As > 0 was arbitrary we may conclude (u, v+) (u, v) which completes
the proof that (u, v+) = (u, v) .
I now claim all the other limiting directions follow easily from what we have
proved. For example,
!1/p
s u, t v
s u, t v
s u, t v
Proposition 2.38 (See [6, Proposition 5.15 from p. 83.]). Let (E, d) be
a metric space, and let x : [0, T ] E be a continuous path. Then x is of
finite p-variation if and only if there exists a continuous
increasing (i.e. non
decreasing) function h : [0, T ] 0, Vpp (Z) and a 1/p H
older path g :
p
0, Vp (Z) E such that such that x = g h. More explicitly we have,
1/p
d (g (v) , g (u)) |v u|
for all u, v 0, Vpp (Z) .
(2.17)
18
d (t x, t y)
Page:
where t x := x1
t xt for all t . The claim is that this should now be a
complete metric space.
(u, v) = (u+, v) lim inf (s, t) lim inf (s, t) (u, v+) = (u, v)
s u, t v
P(0,T )
dpvar (x, y) :=
sup
job:
rpaths
i
vD
(X)
r
p/i
X
i
Xt`1 t`
!i/p
`=1
is a semi-norm on C0,p (, T (n) (V )) and in particular satisfies the triangle inequality. Moreover,
i
i
i
i
i
vD
(X + Y ) sup[vD
(X) + vD
0 (Y )] sup vD 0 (X) + sup vD 0 (Y )
D0
macro:
svmonob.cls
D0
date/time:
D0
11-Mar-2009/12:03
and therefore
i
i
i
sup vD
(X + Y ) sup vD
(X) + sup vD
(Y )
D
which shows
i
sup vD
(X)
D
3
The Bounded Variation Theory
3.1 Integration Theory
Let T (0, ) be fixed,
S := {(a, b] : 0 a b T } {[0, b] R : 0 b T } .
(3.1)
[0,T ]
End(V,W )
The basic linearity properties of this integral are explained in Proposition ??.
For later purposes, it will be useful to have the following substitution formula
at our disposal.
TheoremR 3.2 (Substitution formula). Suppose that f and Z are as above
and Yt = [0,t] f dZ W. Further suppose that g : R+ End (W, U ) is another
A simple function with finite support. Then
Z
Z
gdY =
gf dZ .
[0,T ]
[0,T ]
[0,T ]
[0,T ]
Therefore,
Z
gdY =
[0,T ]
X
Z
() dZ ,
End(W,U )
1{g=} f dZ
[0,T ]
Z
1{g=} f dZ =
gf dZ
[0,T ]
End(W,U )
as desired.
Let us observe that
Z
f (t) dZt
[0,T ]
Y (g = ) =
X
[0,T ]
[0,T ]
[0,T ]
End(W,U )
kk kZ (f = )k .
End(V,W )
n
X
Zt Zt
: a = t0 < t1 < < tn = b and n N
= sup
j
j1
j=1
22
Z
f (t) dZt
[0,T ]
Proof. The additivity on S was already verified in Lemma 2.34. Here is the
proof again for sake of convenience.
Suppose that = {a = t0 < t1 < < tn = b} , s (tl1 , tl ) for some l,
and 0 := {s} . Then
kZ k ((a, b]) :=
=
n
X
X
End(V,W )
Ztj Ztj1
j=1
n
X
kk kZ k (f = ) =
j=1:j6=l
n
X
Ztj Ztj1
+ kZt Zs k +
Zs Zt
l
l1
[0,T ]
((a, s]) + kZ k
Proof. These results are elementary soft analysis except possibly for the
last assertion for the statement in Eq. (3.4). To prove this, to any partition,
P (0, T ) , let
X
f :=
f (t ) 1(t ,t] + f (0) 1{0}
in which case,
and therefore,
Z
f (t ) (Z (t) Z (t )) =
job:
rpaths
f (t) dZ (t) .
[0,T ]
Page:
kf ()k kZ k (d)
[0,T ]
kf k kZ k ([0, T ]) .
Zt Zt
+
Zt Zs + Zs Zt
j
j1
l
l1
= kZ k
X
End(V,W )
j=1:j6=l
kZ k
kk kZ (f = )k
[0,T ]
In short, dY = f dZ.
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Z
kYt Ys k
kf k kdZk =: (s, t)
s
where the right side is a continuous control. This follows from the fact that
kdZk is a continuous measure. Therefore
Z
Let
A = {t [, b] : kf (t) f ()k (t )}
kf k kdZk < .
0
Z
gdY = (Yb Ya ) =
[0,T ]
Second Proof (with out Hahn Banach). Let > 0 and (a, b)
be given. (We will later let 0.) By the definition of the derivative, for all
(a, b) there exists > 0 such that
kf (t) f ( )k =
f (t) f ( ) f( )(t )
|t | if |t | < . (3.6)
V1 (Y )
Z
f dZ =
Z
f dZ =
gf dZ.
0
For the sake of contradiction, suppose that t0 < b. By Eqs. (3.6) and (3.8),
Remark 3.8. If we keep the same hypothesis as in Theorem 3.7 but now take
RT
Yt := t f dZ instead. In this case we have,
Z
Z
gdY =
gf dZ.
[0,T ]
Rt
0
f dZ. It is now
f` (t) = 0 for all t (a, b) . Therefore by the mean value theory, it follows that
f` (t) is constant, i.e. ` (f (t) f (a)) = 0 for all t [a, b] . Since ` V is
arbitrary, it follows from the Hahn Banach theorem that f (t) f (a) = 0, i.e.
f (t) = f (a) independent of t.
Page:
23
job:
rpaths
(3.7)
and t0 be the least upper bound for A. We will now use a standard argument
which is sometimes referred to as continuous induction to show t0 = b. Eq.
(3.6) with = shows t0 > and a simple continuity argument shows t0 A,
i.e.
kf (t0 ) f ()k (t0 ).
(3.8)
Thus Eq. (3.5) holds for all A simple functions and hence also for all uniform
limits of simple functions. In particular this includes all continuous g : [0, T ]
End (W, U ) .
[0,T ]
23
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
24
where (h) := max [t,t+h] k(f ( ) f (t))k. Combining this with a similar computation when h < 0 shows, for all h R sufficiently small, that
Z
Z t
t+h
f ( )d
f ( )d f (t)h
|h|(h),
a
a
where now (h) := max [t|h|,t+|h|] k(f ( ) f (t))k. By continuity of f at t,
Rt
d
(h) 0 and hence dt
f ( ) d exists and is equal to f (t).
a
Rt
For the second item, set G(t) := a F ( ) d F (t). Then G is continuous
and G(t)
= 0 for all t (a, b) by item 1. An application of Proposition 3.9 shows
Rb
G is a constant and in particular G(b) = G(a), i.e. a F ( ) d F (b) = F (a).
Alternative proof of Eq. (3.9). It is easy to show
! Z
Z b
Z b
b
d
f (t) dt.
T (c)
d
F (T (s)) = F 0 (T (s)) T 0 (s) = f (T (s)) T 0 (s) .
ds
Integrating this equation on s [c, d] and using the chain rule again gives
d
T (d)
f (t) dt.
T (c)
Exercise 3.1 (Fundamental Theorem of Calculus II). Prove the fundamental theorem of calculus in this context. That is; if f : V W be a C 1
function and {Zt }t0 is a V valued function of locally bounded variation,
then for all 0 a < b T,
Equation (3.9) now follows from these identities after an application of the Hahn
Banach theorem.
Corollary 3.11 (Mean Value Inequality). Suppose that f : [a, b] V is
a continuous function such that f(t) exists for t (a, b) and f extends to a
continuous function on [a, b]. Then
Z b
kf (b) f (a)k
kf(t)kdt (b a)
f
.
(3.10)
T (d)
`
F (t) dt F (b) + F (a) = 0 for all ` V .
Rt
Proof. For t (a, b) define F (t) := T (c) f ( ) d. Then F C 1 ((a, b) , V )
and by the fundamental theorem of calculus and the chain rule,
f dt = (b a) f .
f (T (s)) T 0 (s) ds =
Rb
a
f(t)dt
f (Zb ) f (Za ) =
f 0 (Z ) dZ ,
f (Z ) dZ :=
a
[a,b]
d
dt |0 f
(z + tv) . In particular
where
t f (Z ) = f (Zt ) f Zt = f Zt + t Z f Zt
Z 1
=
f 0 Zt + st Z t Z ds = f 0 Zt t Z +
t t Z
Page:
24
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
and
:=
Zt + st Z f
Zt
ds.
Thus we have,
f (Zb ) f (Za ) =
Zt t Z + =
k k
Zt dZ (t) + (3.11)
[a,b]
where :=
t t Z. Since,
1 n
Now let Hn (t) := n!
h (t) , by an application of Exercise 3.1 with f (x) =
n+1
x
/ (n + 1)! implies,
Z t
Z t
0
Hn+1 (t) = Hn+1 (t) Hn+1 (a) =
f (h ( )) dh ( ) =
Hn ( ) dh ( )
a
X
X
X
t
kt Zk max
kt Zk
t
t t Z
t
25
V1 (Z) ,
max
t
t
and
t
:=
n ({(s1 , . . . , sn ) [a, b] : si = sj })
Z
n2
= ([a, b])
({sj }) d (sj ) = 0.
0
f Zt + st Z f 0 Zt
ds.
0
0
[0,t]
[a,b]
X Z
(0,t)
X Z
Sn
25
job:
rpaths
(3.12)
Sn
d (s1 ) . . . d (sn )
as1 s2 sn b
Z
= n!
d (s1 ) . . . d (sn ) .
as1 s2 sn b
Exercise 3.3 (Gronwalls Lemma). If (t) and f (t) are continuous nonnegative functions such that
Z t
f (t) (t) +
f ( ) d ( ) ,
(3.13)
0
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
26
then
If we further assume that is increasing, then from Eq. (3.17) and Exercise 3.2
we have
Z
X
f (t) (t) + (t)
d (s1 ) . . . d (sk )
Z
f (t) (t) +
([,t])
( ) d ( ) .
(3.14)
(3.15)
Solution to Exercise (3.3). Feeding Eq. (3.13) back into itself implies
Z t
Z
f (t) (t) +
( ) +
f (s) d (s) d ( )
0
(t) +
Z
(s2 ) +
Z
(s1 ) d (s1 ) +
0s2 s1 t
= (t) +
Z
+
0s2 s1 t
Z
(s1 ) d (s1 ) +
s2
(s1 ) d (s1 ) +
([,t])
d eZt Z
dZ =
t
eZt Z 0
Z
=e
Zt
1.
f (t) (t) + (t) eZt 1 = (t) eZt .
Exercise 3.4. Suppose that {n (t)}n=0 is a sequence of non-negative continuous functions such that
Z t
n+1 (t)
n ( ) d ( ) for all n 0
(3.18)
0s3 s2 s1 t
(3.16)
0sk s2 s1 t
Zt Z
Therefore,
k=1
d ( ) =
0s2 s1 t
f (s3 ) d (s3 ) d (s1 ) d (s2 )
([0, t])
= (t) e([0,t]) .
k!
k=1
0
t
= (t) +
Z
= (t) + (t)
0sk s2 s1 t
k=1
n (t) (t)
([0, t])
for all n 0.
n!
0sk+1 s2 s1 t
N +1
([0, t])
max f (t)
0st
(N + 1)!
0 as N .
Z
1 (t)
So passing to the limit in Eq. (3.16) and again making use of Exercise 3.2 shows,
Z
X
f (t) (t) +
(sk ) d (s1 ) . . . d (sk )
(3.17)
= (t) +
( )
= (t) +
0
Z
= (t) +
k1
X
([, t])
k=1
(k 1)!
26
0 ( ) d ( ) (t)
d (s1 ) ,
0s1 t
Z t Z
1 (s2 ) d (s2 ) (t)
d (s1 ) d (s2 )
0
0
0s1 t
Z
= (t)
d (s1 ) d (s2 ) ,
Z
2 (t)
0s2 s1 t
..
.
d ( )
Z
n (t) (t)
d (s1 ) . . . d (sn ) .
0sn s1 t
e([,t]) ( ) d ( ) .
Page:
k=1 0sk s2 s1 t
Z t
k1
X
([sk , t])
(sk )
d (sk )
(k 1)!
k=1 0
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Z
F (, y ( )) dx ( )
(t) = y0 w0 +
G (, w ( )) dz ( )
Z t
Z t
G (, w ( )) dz ( )
F (, w ( ) + ( )) dx ( )
= y0 w0 +
0
0
Z t
Z t
= y0 w0 +
[F (, w ( )) G (, w ( ))] dx ( ) +
G (, w ( )) d (x z) ( )
0
0
Z t
+
[F (, w ( ) + ( )) F (, w ( ))] dx ( ) .
0
(3.19)
27
where (t) is given in Eq. (3.24). The estimate in Eq. (3.23) is now a consequence
of this inequality and Exercise 3.3 with d ( ) := K ( ) kdx ( )k .
Corollary 3.14 (Uniquness of solutions). If F satisfies the Lipschitz hypothesis in Eq. (3.22), then there is at most one solution to the ODE in Eq.
(3.20).
Proof. Simply apply Proposition 3.13 with F = G, y0 = w0 , and x = z. In
this case 0 and the result follows.
(3.22)
Then
Then
Z
K ( ) kdx ( )k .
(3.23)
ky (t)k
Z
ky0 k +
Z t
a ( ) d ( ) exp
b ( ) d ( ) , where
where
(3.25)
(3.26)
kF (, w ( )) G (, w ( ))k kdx ( )k
Z t
+
kG (, w ( ))k kd (x z) ( )k
ky (t)k ky0 k +
(3.24)
kF (, y ( ))k d ( )
0
ky0 k +
(a ( ) + b ( ) ky ( )k) d ( )
Z
0
t
ky ( )k d ( )
= (t) +
0
Page:
27
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
28
where
Z
(t) := ky0 k +
a ( ) d ( ) and d ( ) := b ( ) d ( ) .
0
Hence we may apply Exercise 3.3 to learn ky (t)k (t) e([0,t]) which is the
same as Eq. (3.25).
Theorem 3.16 (Global Existence). Let us now suppose U = X and F satisfies the Lipschitz hypothesis in Eq. (3.22). Then there is a unique solution,
y (t) to the ODE in Eq. (3.20).
F
(,
y
(
))
dx
(
)
n
0
kF (, yn ( )) F (, y ( ))k kdx ( )k
0
Proof. We will use the standard method of Picard iterates. Namely let
y0 (t) W be any continuous function and then define yn (t) inductively by,
Z t
yn+1 (t) := y0 +
F (, yn ( )) dx ( ) .
(3.27)
K ( ) kyn ( ) y ( )k kdx ( )k
Z t
sup kyn ( ) y ( )k
K ( ) kdx ( )k 0 as n .
0
0 t
Then from our assumptions and the definition of yn (t) , we find for n 1 that
Z t
Z t
F (, yn ( )) dx ( )
F (, yn1 ( )) dx ( )
kyn+1 (t) yn (t)k =
0
t
Remark 3.17 (Independence of initial guess). In the above proof, we were allowed to choose y0 (t) as we pleased. In all cases we ended up with a solution to
the ODE which we already knew to be unique if it existed. Therefore all initial
guesses give rise to the same solution. This can also be see directly. Indeed, if
z0 (t) is another continuous path in W and zn (t) is defined inductively by,
Z t
zn+1 (t) := y0 +
F (, zn ( )) dx ( ) for n 0.
0
Since,
Then
Z
Z t
ky1 (t) y0 (t)k =
y
+
F
(,
y
(
))
dx
(
)
y
(t)
0
0
0
Z
max ky0 ( ) y0 k +
and therefore,
kF (, y0 )k kdx ( )k =: (t) ,
0 t
Z
Z
n
K ( ) kdx ( )k /n!.
(3.28)
28
K ( ) kzn ( ) yn ( )k kdx ( )k .
RT
0
K( )kdx( )k
n=0 0tT
Page:
kF (, zn ( )) F (, yn ( ))k kdx ( )k
0
Z
that
[F (, zn ( )) F (, yn ( ))] dx ( )
0
job:
rpaths
< .
svmonob.cls
date/time:
11-Mar-2009/12:03
dx ( ) y ( ) + f (t)
y (t) =
(3.29)
29
(I A) y = f.
where x (t) End (W ) and f (t) W are finite variation paths. To put this
in the form considered above, let V := End (W ) and define, for y W, F (y) :
V W W by,
y = (I A)
f=
An f.
(3.32)
n=0
But
(An f ) (t) =
dx (n ) An1 f (n ) =
Z
dx (n )
dx (n1 ) An2 f (n1 )
..
.
Notice that
k[F (y) F (y 0 )] (x, f )k = kx (y y 0 )k kxk ky y 0 k
Z
=
Z
dx (n )
Zs
and therefore,
kF (y) F (y 0 )k ky y 0 k
Z
dx (n1 )
n1
dx (n2 ) . . .
s
dx (1 ) f (1 )
s
where we use any reasonable norm on V W, for example k(x, w)k := kxk +
kwk or k(x, w)k := max (kxk , kwk) . Thus the theory we have developed above
guarantees that Eq. (3.29) has a unique solution which we can construct via
the method of Picard iterates.
and therefore, Eq. (3.31) now follows from Eq. (3.32) and (3.33).
For those not happy with this argument one may use Picard iterates instead.
So we begin by setting y0 (t) = f (t) and then define yn (t) inductively by,
Z
F (yn ( )) d (x, f ) ( )
s
Z
X
n=1
Z t
= f (s) +
[dx ( ) yn ( ) + df ( )]
s
Z t
=
dx ( ) yn ( ) + f (t) .
dx (n ) dx (n1 ) dx (n2 ) . . . dx (1 ) f (1 ) .
01 n t
(3.30)
Therefore,
is given by
y (t) = f (t) +
y1 (t) =
n=1
dx ( ) f ( ) + f (t)
s
Z
X
dx (n ) dx (n1 ) dx (n2 ) . . . dx (1 ) f (1 ) .
y2 (t) =
dx (2 ) y1 (2 ) + f (t)
Z 2
Z t
dx (1 ) f (1 ) + f (2 ) + f (t)
=
dx (2 )
s1 n t
(3.31)
Proof. Let us first find the formula for y (t) . To this end, let
Z
Z
(Ay) (t) :=
dx ( ) y ( ) .
Z
dx (2 ) dx (1 ) f (1 ) +
s1 2 t
dx (2 ) f (2 ) + f (t)
s
Page:
29
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
30
and therefore we may conclude in this case that T x (t, s) = e(ts)A where
and likewise,
Z
y3 (t) =
dx (3 ) dx (2 ) dx (1 ) f (1 )
etA :=
s1 2 3 t
dx (2 ) dx (1 ) f (1 ) +
s1 2 t
dx (2 ) f (2 ) + f (t) .
n
X
t n
A .
n!
n=0
dx (k ) . . . dx (1 ) f (1 ) + f (t) .
s1 k t
k=1
y (t) = f (t)
T (t, d ) f ( ) .
(3.39)
dx (k ) . . . dx (1 ) f (1 ) .
s1 k t
(3.35)
s1 n t
T x (t, s) =
X
n=0
Tn (t, s) = I+
Z
X
n=1
dx (n ) dx (n1 ) dx (n2 ) . . . dx (1 ) .
s1 n t
(3.36)
Example 3.20. Suppose that x (t) = tA where A End (V ) , then
Z
dx (n ) dx (n1 ) dx (n2 ) . . . dx (1 )
s1 n t
Z
=An
dn dn1 dn2 . . . d1
=
Page:
30
s1 n t
n
(t s) n
n!
(3.38)
(3.37)
(3.40)
Proof. 1. One may directly conclude that T (t, s) solves Eq. (3.37) by applying Theorem 3.18 with y (t) and f (t) now taking values in End (V ) with
f (t) I. Then Theorem 3.18 asserts the solution to dy (t) = dx (t) y (t) with
y (0) = I is given by T x (t, s) with T x (t, s) as in Eq. (3.36). Alternatively it is
possible to use the definition of T x (t, s) in Eq. (3.36) to give a direct proof the
Eq. (3.37) holds. We will carry out this style of proof for Eq. (3.38) and leave
the similar proof of Eq. (3.37) to the reader if they so desire to do it.
2. Proof of the semi-group property. Simply observe that both t
T (t, s) T (s, u) and t T (t, u) solve the same differential equation, namely,
dy (t) = dx (t) y (t) with y (s) = T (s, u) End (V ) ,
hence by our uniqueness results we know that T (t, s) T (s, u) = T (t, u) .
3. Proof of Eq. (3.38). Let Tn (t, s) := Tnx (t, s) and observe that
1
This is a key algebraic idenitity that we must demand in the rough path theory to
come later.
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Tn (t, s) =
(3.41)
Thus if let
T (N ) (t, s) :=
N
X
N
X
Tn (t, s) = I +
n=0
N Z
X
n=1
Z
=I+
n=1
Tn1 (t, ) dx ()
1
tN
X
Z
Tn (t, ) dx () = I +
s n=0
Corollary 3.22 (Duhamels principle II). Equation (3.39) may also be expressed as,
Z t
y (t) = T x (t, s) f (s) +
T x (t, ) df ( )
(3.43)
s
Tn (t, s) ,
it follows that
T (N ) (t, s) = I +
31
T (N 1) (t, ) dx () . (3.42)
We already now that T (N ) (t, s) T (t, s) uniformly in (t, s) which also follows
from the estimate in Eq. (3.34) as well. Passing to the limit in Eq. (3.42) as
N then implies,
Z t
T (t, s) = I +
T (t, ) dx ()
(3.44)
which is the integrated form of Eq. (3.38) owing to the fundamental theorem
of calculus which asserts that
Z t
ds
T (t, ) dx () = T (t, s) dx (s) .
s
4. Proof of Eq. (3.39). From Eq. (3.31) and Eq. (3.41) which reads in
differential form as, Tn (t, d) = Tn1 (t, ) dx () , we have,
y (t) = f (t)
= f (t)
= f (t)
Z
X
n=1 s
Z t
X
n=1 s
Z tX
Tn (t, d) f ()
Tn1 (t, ) dx () f ()
Tn1 (t, ) dx () f ()
s n=1
Z t
= f (t)
[(x ( ) + x) (y ( ) + y) x ( ) y ( )]
T (t, d) f () .
[x ( ) y + ( x) y ( ) + ( x) y] . (3.46)
Page:
31
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
32
X
X
X
k xk k yk max k xk
k yk
( x) y
Exercise 3.8 (Abelian Case). Suppose that [x (s) , x (t)] = 0 for all 0 s, t
T, show
T x (t, s) = e(x(t)x(s)) .
(3.49)
max k xk V1 (y) 0 as || 0.
T x (dt, s)
= T x (dt, s)
= I.
dx ( )
s
x ( )
(n 1)!
n1
1 n
n! x
(3.47)
T (t, s) =
X
1
n
(x (t) x (s)) = e(x(t)x(s))
n!
n=0
as desired.
Exercise 3.9 (Abelian Factorization Property). Suppose that V is a Banach space and x : [0, T ] End (V ) and y : [0, T ] End (V ) are continuous
finite 1 variation paths such that [x (s) , y (t)] = 0 for all 0 s, t T, then
dt [g (t)] = dt [T x (t, s) S (t, s)] = T x (t, s) S (t, s) dx (t) + dx (t) T x (t, s) S (t, s)
= dx (t) g (t) g (t) dx (t) with g (s) = I.
Observe that g (t) I solves this ODE and therefore by uniqueness of solutions to such linear ODE we may conclude that g (t) must be equal to I, i.e.
T x (t, s) S (t, s) = I.
As usual we say that A, B End (V ) commute if
0 = [A, B] := AB BA.
(3.50)
Hint: show both sides satisfy the same ordinary differential equations see the
next problem.
Exercise 3.10 (General Factorization Property). Suppose that V is a Banach space and x : [0, T ] End (V ) and y : [0, T ] End (V ) are continuous
finite 1 variation paths. Show
T x+y (s, t) = T x (s, t) T z (s, t) ,
where
Z
(3.48)
z (t) :=
T x (s, )
dy ( ) T x (s, )
32
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
dg (t) = T x (s, t)
33
x+y
= T (s, t) dy (t) T
(s, t)
1
1
= T x (s, t) dy (t) T x (s, t) T x (s, t) T x+y (s, t)
is invertible
Solution to Exercise (3.12). This can be deduced from what we have already done. In order to do this, let y (t) := Rx(t) End (B) so that for a B,
d
1
1
1
g (t) = g (t) g (t) g (t) .
dt
T y (t, s) a = a +
=a+
dy (n ) . . . dy (1 ) a
n=1 s1 n t
Z
X
n=1
adx (1 ) . . . dx (n )
s1 n t
= aX (s, t) .
Therefore, taking a = 1, we find,
Z
X
n=0
Xnx (s, t) ,
n=1
where
Xnx (s, t) :=
(3.52)
Z
dx (1 ) . . . dx (n )
X
k,l=0
k l Xk (s, t) Xl (t, u)
n=0
Xk (s, t) Xl (t, u)
k+l=n
s1 n t
Page:
33
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
34
n
X
(3.53)
k=0
Terry. Lyons refers the identities in Eq. (3.53) as Chens identities. These identities may be also be deduced directly by looking at the multiple integral expressions defining Xk (s, t) .
Proof. but we can check this easily directly as well. From Eq. (3.41) we conclude,
Z t
kTn (t, s)k
kTn1 (t, )k kdx ()k
s
and hence one learn inductively that Tn (t, s) is a finite variation process in t,
Z
kTn (t, s)k
kdx (n )k kdx (n1 )k kdx (n2 )k . . . kdx (1 )k
s1 n t
Z t
n
1
=
n!
kdxk
and
V1 Tn (, s) |[s,T ]
1
(n 1)!
Z
s
n1
1
kdxk
kdx ( )k =
n!
!n
kdxk
V1 Tn (, s) |[s,T ] exp
!
kdxk
1.
n=1
P
Hence we learn that n=0 Tn (t, s) converges uniformly to T (t, s) so that T (t, s)
is continuous. Moreover, if P (s, T ) , then
X
kT (t, s) T (t , s)k
X
X
n=1 t
V1 Tn (, s) |[s,T ] < .
n=1
Page:
34
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
kdxk
V1 T (, s) |[s,T ] exp
1 < .
Similarly we have,
N
X
X
X
Tn (t , s)
V1 Tn (, s) |[s,T ]
T (t, s)
n=0
n=N +1
and therefore,
"
V1
T (, s)
N
X
#
Tn (, s) |[s,T ]
!
0 as N .
n=0
N Z
X
f ( ) T (d, s) = lim
n=0
f ( ) Tn (d, s) =
Z
X
n=0
f ( ) Tn (d, s) .
In particular,
T (t, s) T (s, s) =
[Tn (t, s)
n=0
Z t
X
Tn (s, s)]
#
"Z t
X
dx ( ) Tn1 (t, s) =
dx ( )
Tn1 (, s)
n=1
Z t
dx ( ) T (, s) .
s
n=1
4
Banach Space p variation results
In this chapter, suppose that V is a Banach space and assume x
C ([0, T ] V ) with x (0) = 0 for simplicity. We continue the notation used
in Chapter 2. In particular we have,
!1/p
Vp (x : ) :=
X
xt xt
p
Vp (x) :=
sup
!1/p
=
kt xk
Vpp (x, ) = Vpp (x, [s, u]) + Vpp (x, [u, t])
and
(s, u) + (u, t) .
Vp (x : ) .
(4.3)
P(0,T )
Vpp (x, ) 2p1 Vpp (x, {u}) 2p1 [ (s, u) + (u, t)] .
(4.1)
Proof. Since
p
Taking the supremum of this inequality over P (s, t) then gives the desired
result.
These results may be significantly improve upon. For example, we have the
following proposition whose proof we leave to the interested reader who may
wish to consult Lemma (4.13) below.
we have,
then
Vpp (x, )
p1
Corollary 4.2. As above, let (s, t) := x,p (s, t) := Vpp x|[s,t] be the control
associated to x. Then for all 0 s < u < t T, we have
(s, u) + (u, t) (s, t) 2p1 [ (s, u) + (u, t)] .
(4.2)
The second inequality shows that if x|[s,u] and x|[s,t] both have finite p variation, then x|[s,t] has finite p variation.
p1
# ( 0 ( , ])
Vpp ( 0 ( , ])
Vpp (x, 0 ) ,
(4.4)
(4.5)
where
k := max {#(i1 , i ] 0 : i = 1, 2, . . . , n} .
(4.6)
Proof. This follows by the same methods used in the proof of Lemma 4.1
or Lemma (4.13) below. The point is that,
38
p
X
X
p
x
Vpp (x, ) =
k xk =
s
s 0 ( , ]
X
X
p1
p
# ( 0 ( , ])
ks xk
!1/p
s 0 ( , ]
Vp (x + y : ) =
kt x + t yk
kt xk
(kt xk + kt yk)
(4.7)
kt yk
= Vp (x : ) + Vp (y : )
and therefore,
Vpp (x, 0 )
(4.8)
Hence it follows from this triangle inequality and items 1. and 2. that
C0,p ([0, T ] , V ) is a linear space and that Vp () is a norm on C0,p ([0, T ] , V ) .
4. To finish the proof we must now show (C0,p ([0, T ] , V ) , Vp ()) is a
complete space. So suppose that {xn }n=1 C0,p ([0, T ] , V ) is a Cauchy sequence. Then by Eq. (4.10) we know that xn converges uniformly to some
x C ([0, T ] V ) . Moreover, for any partition, P (0, T ) we have
Vp (x xn : ) Vp (x xm : ) + Vp (xm xn : )
Vp (x xm : ) + Vp (xm xn ) .
while
Vpp (x, 0 [ , ])
Vp (x + y) Vp (x) + Vp (y) .
p1
!1/p
X
Vp (x) + Vp (y) ,
!1/p
Vpp (x, 0 ) =
!1/p
X
Vpp (x : [ , ]) .
(4.9)
So combining these two inequalities and then taking the supremum over
P (s, t) gives Eq. (4.7).
p 1/p
Vp (x) .
(4.10)
Proof. Let P (0, T ) , then
38
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Vqq (x : ) =
kt xk =
max kt xk
t
qp
kt xk
kt xk
39
t
qp
qp
kt xk 2 kxku
Vpp (x) .
Taking the supremum over P (0, T ) and then taking the q th roots of both
sides gives the result.
Notation 4.8 For x C ([0, T ] V ) and P (0, T ) , let x (t) be the
piecwise linear path defined by
x (t) := x (t ) +
(t t )
t x for all t [0, T ] ,
t+ t
We will give the proof of this theorem at the end of this section.
Corollary 4.11. Suppose
x C0,p ([0, T ] V ) and q (p, ) . Then
lim||0 Vq x x = 0, i.e. x x as || 0 in C0,q for any q > p.
Proof. According to Proposition 4.7 and Theorem 4.10,
Fig. 4.1. Here = {0 = s0 < s1 < < s7 = T } and should be x. The red lines
indicate the image of x .
qp p
V p x x
Vqq x x 2
x x
u
qp
p
2
x x
u
Vp (x) + Vp x
qp p1 p
2
x x
u
2
Vp (x) + Vpp x
qp p1
2
x x
u
2
1 + 3p1 Vpp (x) .
The latter expression goes to zero because of Proposition 4.9.
We refer the reader to in [5] for more results in this vain. In particular, as a
corollary of Theorem 23 and 24 one sees that the finite variation paths are not
dense in C0,p ([0, T ] V ) .
Notation 4.12 Suppose that , 0 P (0, T ) the we let
Page:
39
job:
rpaths
= (, 0 ) = {t (0, T ] : (t , t) 6= }
S := S (, 0 ) = t {t , t} ,
see Figure 4.3 below for an example.
(4.12)
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
40
api
n
X
i=1
Fig. 4.3. In this figure the red xs correspond to and the vertical hash marks
correspond to . The green circles indicate the points which make up S.
kx (u3 ) x (u2 )k (kx (u3 ) x (t2 )k + kx (t6 ) x (t2 )k + kx (u3 ) x (t6 )k)
p
(4.13)
ai
i=1
!p
Since x is linear on each of the intervals [ti , ti+1 ] , we may apply Lemma 4.14
to find,
Vpp x : [t1 , t2 ] +Vpp x : [t6 , t7 ] + Vpp x : [t9 , t10 ]
Vpp x : {t1 , t2 } + Vpp x : {t6 , t7 } + Vpp x : {t9 , t10 }
and for the remaining terms we have,
Vpp x : [0, t1 ] + Vpp x : [t2 , t6 ]
+Vpp x : [t7 , t9 ] + Vpp x : [t10 , T ]
Vpp x : {0, t1 } + Vpp x : {t2 , t6 }
=
+Vpp x : {t7 , t9 } + Vpp x : {t10 , T }
which gives the inequality in Eq. (4.15).
with the inequality being strict if p > 1 and is a strict refinement of {u, v} .
Proof. Here we have,
X
X
p
p
Vpp (x : ) =
kx (t) x (t )k =
k(t t ) bk
t
p
= kbk
job:
rpaths
Page:
40
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Taking the supremum over all 0 P (0, T ) then gives the estimate in Eq.
(4.12).
For an alternative proof of Theorem 4.10 the reader is referred to [5] and [6,
Chapter 5] where controls are used to prove these results. Most of these results
go over to the case where V is replaced by a complete metric space (E, d) which
is also geodesic. That is if a, b E there should be a path : [0, 1] E such
that (0) = a, (t) = b and d ( (t) , (s)) = |t s| d (a, b) for all s, t [0, 1] .
More invariantly put, there should be a path : [0, 1] E such that if ` (t) :=
d (a, (t)) , then d ( (t) , (s)) = |` (t) ` (s)| d (a, b) for all s, t [0, 1] .
5
Youngs Integration Theory
Theorem 2.24 above shows that if we insist upon integrating all continuous
functions, f : [0, T ] R, then then the integrator, x, must be of finite variation.
This suggests that if we want to allow for rougher integrators, x, then we must
in turn require the integrand to be smoother. Youngs integral, see [18] is a
result along these lines. Our fist goal is to prove some integral bounds.
In this section we will assume that V and W are Banach spaces and g :
[0, T ] V and f : [0, T ] End (V, W ) are continuous functions. Later we will
assume that Vp (f ) + Vq (g) < where p, q 1 with := 1/p + 1/q > 1.
Notation 5.1 Given a partition P (s, t) , let
X
S (f, g) :=
f ( ) g.
r
X
i=2
r
X
Si (f, g) Si1 (f, g)
.
i=2
So in more detail,
= {s = 0 < 1 < < r = t}
then
S (f, g) :=
r
X
f (l1 ) l g = f (0 ) 1 g + + f (r1 ) r g.
(5.1)
l=1
(5.4)
We also have
min xi =
1in
min xri
i
1/r
1X r
x
n i=1 i
!1/r
=
1/r
1
kxkr .
n
44
min ai bi
1in
1/r
1/r
1
1
ka b kr
kakp kbkq .
n
n
where
r () :=
n
. . . a
n
n
X
Consequently,
kS (f, g)k kf (s)k kg (t) g (s)k + r () Vp f |[s,t] Vq g|[s,t]
kf (s)k + r () Vp f |[s,t] Vq g|[s,t] .
(5.6)
Proof. Without loss of generality we may assume that ai > 0 for all i. By
Jensens inequality
!
n
n
n
X
X
X
1
n
a1 . . . an = exp
i ln ai
i exp (ln ai ) =
i ai .
(5.7)
i=1
i=1
Replacing ai by api in this inequality and then taking the pth root gives the
inequality in Eq. (5.5). (Remark: if p 1, Eq. (5.5) follows from Eq. (5.7) by
an application of Holders inequality.) Making use of Eq. (5.6), we find again
that
1/n
n
1X p
a
n i=1 i
!1/p
1/n
= [a1 . . . an ]
n
1X q
b
n i=1 i
!1/q
1
=
kakp kbkq .
n
44
job:
rpaths
1
r1
1/p
p
k f k
1/q
q
+ g
(s,t)
(s,t)
Vp f |[s,t] Vq g|[s,t] .
1/n
[b1 . . . bn ]
1
r1
Vp f |[s,t] Vq g|[s,t] .
r1
r
X
Si (f, g) Si1 (f, g)
i=2
r
X
1
Vp f |[s,t] Vq g|[s,t]
i1
i=2
r1
X
1
Vp f |[s,t] Vq g|[s,t]
=
i
i=1
= r () Vp f |[s,t] Vq g|[s,t] .
(5.12)
(5.5)
min ai bi [a1 b1 . . . an bn ]
(5.11)
(s,t)
!1/p
i api
a
.
n i=1 i
1in
(5.10)
l=1
l=1
i=1
i=1
r1
X
X
1
1
and
()
=
()
:=
.
l
l
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Z
X
X ,
p
Vp (X : ) .
(5.16)
(5.17)
More generally,
Z
Vq
f dg |[s,t]
f |[s,t]
u + () Vp f |[s,t] Vq g|[s,t]
0
kf (s)k + [1 + ()] Vp f |[s,t] Vq g|[s,t] .
(5.13)
wherein we have used (s, t) := Vqq g|[s,t] is a control for the last inequality.
Taking the supremum over P (0, T ) then implies,
Vq (X) () Vp (f ) Vq (g) .
Vqq g|[ , ] q () Vpq (f ) Vqq (g) ,
(5.14)
(5.19)
q () Vpq (f )
Vq Y |[u,v]
f |[u,v]
u Vq g|[u,v] .
(5.18)
X
X
q
X ,
=
Vqq (X : ) =
f dg f ( ) (g ( ) g ( ))
X
then
(5.20)
kf ( )k k(g ( ) g ( ))k
f dg = Xst + Yst ,
q X
q
q
f |[u,v]
u
k(g ( ) g ( ))k =
f |[u,v]
u Vqq (g : ) .
[kf ku + () Vp (f )] Vq (g) ,
which is Eq. (5.16). Equation (5.17) is an easy consequence of Eq. (5.16) and
the simple estimate,
kf (t)k kf (t) f (0)k + kf (0)k Vp (f ) + kf (0)k .
(5.21)
The estimates in Eqs. (5.18) and (5.19) follow by the same techniques or a
simple reparameterization argument.
and
Page:
[kf ku + () Vp (f )] Vq (g)
P(s,t)
Proof. Inequality (5.15) follows from Eq. (5.9) upon letting || 0. For
the remaining inequalities let Yst be as in Eq. (5.13) and define,
Z t
Z t
Xs,t :=
f dg f (s) (g (t) g (s)) =
f dg Ys,t .
Vp X|[s,t] :=
45
f dg
Vq
Young gives examples showing that Eq. (5.9) fails if one only assume that
p1 +q 1 = 1. See Theorem 4.26 on p. 33 of Dudley 98 [2] and Young (1936) [18]
Young constant is not as good as the one in [2].
45
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
46
Theorem 5.9. If Vp (g) < and Vq (f ) < with := 1/p + 1/q > 1, then
Z
Z
n
f dgn exists
f dg := lim
2
kf
(s)k
+
1
+
f dg S (f, g)
lim sup
(5.22)
||0
(5.23)
f dg is a bilinear form in f and g, the estimates in Eqs. (5.18) and (5.19)
s
continue to hold, and
Z t
f dg :=
lim
S (f, g) .
(5.24)
Rt
(s, t) := Vpp g|[s,t] + Vqq f |[s,t] ,
f
(s)
(g
(t)
g
(s))
46
job:
rpaths
1/p
(t , cn (t))
1/p
(t , t)
1/q
(t , t)
1/q
(t , t)
(t , t)
tn
1
sup (t , t)
tn
(t , t)
tn
1
sup (t , t)
(0, T ) .
(5.25)
tn
Page:
tn
kf (cn (t)) f (t )k kt gk
tn
We may now let q q to get the estimate in Eq. (5.23). This estimate gives those
in Eqs. (5.18) and (5.19). The independence of the limit on the approximating
sequence and the resulting bilinearity statement is left to the reader.
If P (s, t) it follows from the estimates in Eq. (5.23) and Eq. (5.12) that
Z t
Z t
Z t
Z t
f dg S (f, g)
f dg
f dgn
+
f dgn S (f, gn )
+ kS (f, gn ) S (f, g)k
h
h
i
i
2 kf (s)k + 1 + Vp f |[s,t] Vq (g gn ) |[s,t]
Z t
+
f dgn S (f, gn )
.
X
tn
tn
Proof. Let
tn
n :=
macro:
svmonob.cls
max
|ts||n |
kf (t) f (s)k
q 0 p
p
0 as n ,
date/time:
11-Mar-2009/12:03
we find,
X
kf (cn (t)) f (t )k kt gk
tn
q0
!1/q0
!1/q
kf (cn (t)) f (t )k
tn
!1/q0
X
Vq (g : )
kf (cn (t)) f (t )k
Exercise 5.1 (Product Rule). Suppose that V is a Banach space and p, q > 0
such that := p1 + 1q > 1, x C ([0, T ] End (V )) with Vq (x) < and
y C ([0, T ] End (V )) with Vp (y) < . Show for all 0 s < t T that,
Z t
Z t
x (t) y (t) x (s) y (s) =
dx ( ) y ( ) +
x ( ) dy ( ) ,
(5.26)
kt gk
tn
tn
47
n Vqp/q (f ) Vq (g) 0 as n .
[(x ( ) + x) (y ( ) + y) x ( ) y ( )]
f dg =
and therefore,
Z t
XZ
f dg =
XZ
d
f () g
f dg
S (f, g) =
Taking the limit at || 0, the terms corresponding to the first two summands
Rt
Rt
converge P
to s x ( ) dy ( ) and s dx ( ) y ( ) respectively. So it suffices to show,
lim||0 ( x) y = 0. However for every > 0 we have,
X
X
1
k xk k yk max k xk
k xk
k yk
d
.
[f () f ( )] g
sup ( , )
(s, t) 0 as || 0.
Z
f dg =
Taking norms of this equation and using the obvious inequalities implies,
Z t
XZ
d
f
dg
S
(f,
g)
kf () f ( )k k gk
s
XZ
1/p
1/q d
( , ) ( , )
=
( , )
[x ( ) y + ( x) y ( ) + ( x) y] .
XZ
Remark 5.11. The same methods used to prove the estimate in Eq. (5.25) allows
to give another proof of Eq. (5.24). Observe that
lim
P(s,t), ||0
max k xk
job:
k yk
p
p1
where p0 =
or equivalently, p10 + p1 = 1. As p1 + 1q = > 1 it follows that
0
q < p and therefore we may choose > 0 such that p0 (1 ) = q. For this
we then have,
X
Alternatively: let (s, t) := Vqq x|[s,t] + Vpp y|[s,t] which is a control on
[0, T ] . We then have,
X
X
X
1/q
1/p
k xk k yk
( , ) ( , )
=
( , )
max ( , )
f dg .
rpaths
!1/p
X
(1)
macro:
svmonob.cls
( , )
47
!1/p0
max ( , )
Page:
k xk
p0 (1)
(s, t) 0 as || 0.
date/time:
(5.27)
11-Mar-2009/12:03
48
=K
Vpp (Z
: ) K
Vpp
0
f Zt + st Z f 0 Zt ds.
Thus we have,
f (Zb ) f (Za ) =
:=
Letting (s, t) :=
f 0 (Z ) dZ :=
a
0
f 0 (Z ) dZ ,
d
dt |0 f
t t Z.
Z|[s,t] , we have,
0
f Zt + st Z f 0 Zt
ds
K
0
1
1/p
s kt Zk ds K (t , t)
2
and hence
X
X
1/p
1/p
kt Zk K
(t , t) (t , t)
t
2
t
t
K X
(t , t) 0 as || 0
2
k k
(5.29)
[a,b]
0
(Z) .
Exercise 5.3 (Fundamental Theorem of Calculus II). Prove the fundamental theorem of calculus in this context. That is; if f : V W be a C 1
function such that f 0 is Lipschitz and {Zt }t0 is a continuous V valued
function such that Vp (Z) < for some p (1, 2). Then Vp (f 0 (Z )) < and
for all 0 a < b T,
b
(5.30)
Vpp
f 0 Zt t Z +
Therefore taking the supremum over all P (0, T ) gives Eq. (5.28).
f (Zb ) f (Za ) =
where
t :=
(z + tv) . In particular
as we saw in Eq. (5.27). Thus letting || 0 in Eq. (5.30) completes the proof.
Exercise 5.4. See what you can say about a substitution formula in this case.
Namely, suppose that
Z t
y (t) =
f (s) dx (s)
0
where
t f (Z ) = f (Zt ) f Zt = f Zt + t Z f Zt
Z 1
=
f 0 Zt + st Z t Z ds = f 0 Zt t Z +
t t Z
0
and
Page:
48
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
49
Then
f (s) (g (t) g (s)) f (s) (g (u) g (s))
kXst Xsu Xut k =
f (u) (g (t) g (u))
of [s, t] , let
X() :=
X , =
r
X
Xti1 ,ti .
i=1
[s,t]{s,t}
The following lemma explains the reason for introducing this notation.
[s,t]
!1/p
Vp (X) :=
sup
X
Xt ,t
p
V
< .
P(0,T ) t
(5.31)
X (n {u})s,t X (n )s,t
=
Xu ,u + Xu,u+ Xu ,u
Xu ,u
+
Xu,u+
+
Xu ,u
3 n
where
n := max {kXs,t k : (s, t) 3 |t s| |n |} .
(s, t) := Vpp f |[s,t] + Vqq g|[s,t] .
2
This is redundant since Vp (X) < can only happen if Xt,t = 0 for all t.
Page:
49
(5.32)
job:
rpaths
svmonob.cls
date/time:
11-Mar-2009/12:03
50
kX () Xs,t k () (s, t) .
kX () X ( ( ))k =
X ,+ X , X,+
( , + ) .
(5.34)
U ,
2C
kUst k =
U ,
( , ) 0 as || 0.
min
(s,t)
kX () X ( ( ))k
min
(s,t)
( , + ) min
2
, 1 (s, t).
|| 2
||2
X 1
(s, t) () (s, t) .
kX () Xs,t k
k
k=1
Theorem 5.24. Let X : V be a continuous (, ) almost additive functional and denote a partition of [0, T ] . Then
Yst := lim X ()st exists uniformly in (s, t) .
||0
(5.35)
(5.36)
k=0
for some C < . In fact according to Proposition 5.23 we know that C may be
chosen to be () .
and
k=0
() := max 1 (, ) .
1 X
2
(tj1 , tj+1 )
(s, t)
r 1 j=1
r1
| |
Page:
50
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
X
0
0
X ( ) , X ,
kX ( ) X ()k =
X
X ( 0 ) , X ,
()
( , )
()
max
| |||
= () (||)
(, )
max
| |||
51
( , )
(, )
(5.38)
(s, t) .
Now let 1 , 2 P (0, T ) be arbitrary and apply the previous inequality with
0 = [1 2 ][s,t] and being either [1 ][s,t] or [2 ][s,t] to find,
k[X (1 ) X (2 )]st k k[X (1 ) X (1 2 )]st k
+ k[X (1 2 ) X (2 )]st k
() (s, t) [ (|1 |) + (|2 |)]
() (0, T ) [ (|1 |) + (|2 |)] .
(5.40)
where k is the Lipschitz constant for f and hence the integral in Eq. (5.39) is
well defined. In order to consider existence, uniqueness, and continuity in the
driving path x of Eq. (5.46) we will need a few more facts about p variations.
Therefore,
(s,t)
Vp (Y ) Vp (X) + C (0, T ) .
(5.37)
Gronwalls
then
un n1 . . . 1 0 u0 +
Vp (Y ) Vp (Y X) + Vp (X)
C p (0, T )
p1
un n u0 +
(0, T ) = C p (0, T )
job:
n1k
Y
j k .
(5.41)
j=1
n1
X
n1i i
(5.42)
i=0
n u0 + n1
51
n1
X
k=0
Suppose that
ui+1 i ui + i ,
Inequalities).
n1
X
i if 1.
(5.43)
i=0
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
52
Proof. Let (s, t) := Vpp (x : [s, t]) be the control associated to x and define
.
(5.45)
= (p) := 1 + (2/p) .
If y solves Eq. (5.39), then
Z t
y (t) = y (s) +
f (y ( )) dx ( ) for all 0 s t T.
s
or equivalently, with c := k,
1/p
1/p
1 c (s, t)
Vp (y : [s, t]) kf (y (s))k (s, t) .
Therefore it follows that
which is to say,
vi+1 vi
(i+1)
i .
1/p
1/p
kf (y (s))k if (s, t)
1/2c.
(5.47)
un u0 = vn v0 =
n1
X
(vi+1 vi )
n1
X
Since
(i+1)
and
i=0
i=0
which upon solving for un gives Eq. (5.42). When i = is constant, we use
n1
X
i=0
n1i
n1
X
n 1
=
=
1
i=0
i
1/p
n 1
1 n
= n u0 +
1
1
ky (t)k
52
job:
rpaths
(5.49)
and
(5.48)
1
1+
1/p
ky (s)k + 2 (s, t)
1/p
kf (0)k if (s, t)
1/2c.
(5.50)
svmonob.cls
date/time:
11-Mar-2009/12:03
2c
p
1
p
p
p
ky (ti )k + 22p1 (ti , ti+1 ) kf (0)k
ky (ti+1 )k 2p1 1 +
ip
ky (ti )k 4 ky0 k + 4
(i1)p
kf (0)k
i1
X
(tl , tl+1 )
kt (f x)k = k(f + t f ) (x + t x) f x k
= kf t x + t f x + t f t xk
1
kf ku kt xk + kxku kt f k + (kt f t xk + kt f t xk)
2
1
kf ku kt xk + kxku kt f k + (2 kf ku kt xk + 2 kxku kt f k)
2
2 (kf ku kt xk + kxku kt f k) .
and therefore,
p
1
p
p
(y : [ti , ti+1 ]) 2
ky (ti )k + 22p1 (ti , ti+1 ) kf (0)k
p h
i
1
p
p
p1
2
4ip ky0 k + 4(i1)p kf (0)k (0, ti )
p1
(5.51)
n
X
(0,T )
((2c) (0, T ) + 1)
l=0
p
p1
(0,T )+1)p
Vpp (y : [0, T ]) (n + 1)
C (p) eC(p)k
53
Vpp
1 p
Vp (f x : ) 2 kf ku Vp (x : ) + 2 kxku Vp (f : )
from which the result follows.
Theorem 5.30. Suppose that W and Z are Banach spaces and
f
C 2 (W Z) with f 0 C 1 (W End (W, Z)) and f 00
C (W End (W, End (W, Z))) both being bounded functions. If y0 , y1
Cp ([0, T ] W ) , then
Vp (f (y1 ) f (y0 )) 2 kf 0 ku Vp (y1 y0 )+kf 00 ku [Vp (y0 ) + Vp (y1 )] ky1 y0 ku .
(5.52)
i=0
4(n+1)p 1
41
p
p
p1
(n+1)p
C (p) (ky0 k + kf (0)k (0, T )) 4
(n + 1)
.
(n + 1)
Page:
53
p1
Proof. Let
job:
rpaths
k := kf 0 ku ,
M := kf 00 ku ,
svmonob.cls
date/time:
11-Mar-2009/12:03
p1
54
ys := y0 + s (y1 y0 ) = (1 s) y0 + sy1 ,
we have, by the fundamental theorem of calculus,
Z 1
Z 1
d
f (y1 )f (y0 ) =
f (ys ) ds =
f 0 (ys ) (y1 y0 ) ds. =: F (y0 , y1 ) (y1 y0 )
0 ds
0
Thus if we define,
1
f 0 (ys ) ds,
F (y0 , y1 ) :=
1
M [Vp (y0 ) + Vp (y1 )] .
2
Putting this all together gives the result in Eq. (5.52).
Vp (F (y0 , y1 ))
(5.53)
then
f (y1 ) f (y0 ) = F (y0 , y1 ) (y1 y0 ) .
(5.54)
Let us observe that if y0 and y1 are two more such points in W, then
Z 1
Z 1
0
0
kF (y0 , y1 ) F (
y0 , y1 )k =
f (ys ) ds
f (
ys ) ds
0
0
Z 1
[f 0 (ys ) f 0 (
ys )] ds
=
Z
0
1
kf 0 (ys ) f 0 (
ys )k ds M
We are now prepared to prove our basic existence, uniqueness, and continuous
dependence on data theorem for Eq. (5.39).
Theorem 5.31 (Local Existence of Solutions). Let p (1, 2) and :=
1 + (2/p) < . Suppose that f : W End (V, W ) is a C 2 function such that f 0 and f 00 are both bounded functions. Then there exists
0 = 0 (kf 0 ku , kf 00 ku , p, kf (y0 )k) such that for all x Cp ([0, T ] V ) with
Vp (x) 0 , there exists a solution to Eq. (5.39).
kys ys k ds
Proof. Let
WT := {y Cp ([0, T ] , W ) : y (0) = y0 } .
M
0
1
= M [ky0 y0 k + ky1 y1 k] .
2
So in summary we have,
f (y1 ) f (y0 ) = F (y0 , y1 ) (y1 y0 )
where F : W W Z is bounded Lipschitz function satisfying,
kF ku k and
1
kF (y0 , y1 ) F (
y0 , y1 )k M [ky0 y0 k + ky1 y1 k] .
2
Therefore
Vp (f (y1 ) f (y0 )) = Vp (F (y0 , y1 ) (y1 y0 ))
2 [kVp (y1 y0 ) + ky1 y0 ku Vp (F (y0 , y1 ))] .
Our goal is to show that S is a contraction and then apply the contraction
mapping principle to deduce the result. For this to work we are going to have
to restrict our attention to some ball, C , about the constant path y0 and at the
same time shrink T in such a way that S (C ) C and S|C is a contraction.
We now carry out the details.
First off if y C WT , then
Vp (S (y)) [kf (y0 )k + Vp (f (y))] Vp (x) [kf (y0 )k + kf 0 ku Vp (y)] Vp (x)
kf (y0 )k Vp (x) + kf 0 ku Vp (x) .
(5.55)
Secondly, if y, z C WT , then
Since,
Z
[S (y) S (z)] (t) :=
(f (y) f (z)) dx
0
Page:
54
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Therefore, making use of the fact that f (y) = f (z) at t = 0, we find that
Vp (S (y) S (z)) Vp (f (y) f (z)) Vp (x)
{2 kf 0 ku Vp (y z) + kf 00 ku [Vp (z) + Vp (y)] ky zku } Vp (x)
Vp (x) {2 kf 0 ku + kf 00 ku [Vp (z) + Vp (y)]} Vp (y z)
2Vp (x) {kf 0 ku + kf 00 ku } Vp (y z) .
(5.56)
So in order for this scheme to work we must require, for some (0, 1) , that
0
Lemma 5.32. Let a, A, and B be positive constants and (0, 1) . Then there
exists 0 (, a, A, B) > 0 such that for 0 , there exists (0, ) such that
(a + As) and 2A + 2Bs for 0 s .
(5.59)
Proof. Our goal is to satisfy Eq. (5.59) while allowing for to be essentially
as large as possible. The worst case scenarios in these inequalities is when s =
and in this case the inequalities state,
a
2A
1 A
2B
provided A < 1.
Letting M := 1/ we may rewrite the condition on as M > A and
M 2A
a
M A
2B
0 := min A1 ,
(5.57)
(5.58)
2 +
2
s
A+
1
2
2
aB A
2+
A +2
.
which gives,
2aB (M A) (M 2A) ,
on k implies,
i.e.
2
y (Ti1 ) y0 =
M (2 + ) AM + 2 A aB 0
2+
A2 aB
AM + 2
0M
2
2
2
2+
A aB
2+
= M
A +2
A .
2
2
2
55
X Z
1k<i
or equivalently,
Page:
55
2
Thus we must choose (if 2 A aB
2+
0),
2 A
s
2
aB A2
1
2+
2+
=M
A+
A +2
.
2
2
job:
rpaths
Tk1
Ti1
f (y) dx =
Tk1
f (y) dx
0
(5.61)
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
56
and therefore,
and therefore,
Z
f (y ( )) d (x u) ( ) +
w (t) = w (s) +
s
[f (y ( )) f (v ( ))] du ( )
s
and hence
Vp (w : [s, t]) A (s, t) + B (s, t)
where
Z
f (y ( )) d (x u) ( ) : [s, t]
A (s, t) = Vp
Zs
and
[f (y ( )) f (v ( ))] du ( ) : [s, t] .
and
Z
B (s, t) = Vp
Y : W Cp ([0, T ] , V ) Cp ([0, T ] , W )
is uniformly continuous on bounded subsets of W Cp ([0, T ] , V ) .
Page:
56
job:
rpaths
[f (y ( )) f (v ( ))] du ( ) : [s, t]
wherein we have used the a priori bounds that we know of y and v. Thus we
have,
Vp (w : [s, t]) C1 Vp (x u : [s, t]) + C2 Vp (u : [s, t]) (kw (s)k + Vp (w : [s, t])) .
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
The result follows from this estimate via the usual iteration procedure based
on keeping C2 Vp (u : [s, t]) 1/2 so that
Then
57
d
z(xt x0 ) = a(z(xt x0 ))x t with z(xt x0 )|t=0 = .
dt
d
Theorem P
5.38. If {Ai }m
i=1 are commuting Lipschitz vector fields on R ,
m
m
d
1
m
f (y) x :=
and y R , and x C ([0, T ] , R )
i=1 xi Ai (y) for x R
then Eq. (5.62) becomes,
y t =
m
X
Ai (yt )x it with y0 = .
i=1
yt = e
P i
(xt xi0 )Ai
xm
t )Am
().
Remark 5.36. This result could be used to give another proof of existence of
solutions, namely, we just extend
The next example shows however that life is more complicated when the
vector fields do not commute.
Y : W C 1 ([0, T ] , V ) Cp ([0, T ] , W )
by continuity to W Cp ([0, T ] , V ) .
and A2 :=
.
r
r
[0, T ] , R2 so that the corresponding ODE be-
A1 := r
y t = A1 (yt ) x 1t + A2 (yt ) x 2t
(5.62)
= yt x 1t + x 2t with y0 = .
The solution to this equation is given by Duhamels principle as
Z t
1
1
x1t x10
yt = e
+
ext xs dx2s .
0
yt = e(xt x0 )a x ().
In particular, yt (x, ) depends continuously on x in the sup-norm topology and
hence easily extends to rough paths.
57
job:
rpaths
This last expression is not continuous in x in the Vp norm for any p > 2
which follows from Lemma 5.40 with x = (u (n) , v (n)) .
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
58
Z
0
while
1
eus (n) dvs (n) = o (1) t 9 0 as n
2
Vp (u (n) , v (n)) = O n2/p1
1
n
cos n2 t
(5.63)
By Greens theorem, At , is the signed area swept out by x|[0,t] then followed by
the straight line path from x (t) back to the origin. Taking
1
1
2
2
xn (t) := (ut (n) , vt (n)) =
cos n t 1 , sin n t
n
n
we find,
(5.64)
Z
1 2 t
n
cos n2 s 1 cos n2 s + sin n2 s sin n2 s ds
2
2n
0
Z
1
1
1
1 t
1 cos n2 s ds = t 2 sin n2 t t as n .
=
2 0
2
2n
2
Whereas we have seen from above, that Vp (xn )
= O n2/p1 . This shows that
the area process is not continuous in the Vp - norm when p > 2 since Vp (xn ) 0
but At (n) 21 t 6= 0.
At (n) =
= o(1) O
1 cos 2n2 s ds
n
2 0
1
= o (1) t,
2
which proves Eq. (5.63).
By Theorem 2.10,
Vp (v (n)) Vp (u (n))
p
1/p
2
1 2/p
n2
=O
n
= O n2/p1 .
n
n
Equation (5.64) now follows from this estimate and the fact that
Vp (u, 0) Vp (u, v) Vp (u, 0) + Vp (0, v) .
Example 5.41. Let x (t) := (u (t) , v (t)) be a smooth path such that x (0) = 0
and consider the area process,
Z
1 t
At :=
(udv vdu) .
2 0
Page:
58
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
6
Rough Paths with 2 p < 3
We now are going to consider paths with finite p variation for some p
[2, 3). As we have seen above we have to add some additional information to get
a reasonable theory going. We first need to introduce the appropriate algebra.
As usual V with be a Banach space which we will sometimes assume is finite
dimensional in order to avoid technical details.
X
X
X
Q (vi , wi )
|Q (vi , wi )| kQk
kvi k kwi k ,
Q () =
E
i
where kQk is the best constant for which the last inequality holds. Taking the
infimum over all such decomposition of shows
Q () kQk kk
E
and therefore
Q
op
from which it follows that
Q
. Since (0, kQk) was arbitrary, it
op
follows that
Q
kQk
.
op
Before checking kk is a norm (see Corollary 6.3), let us observe the following
property.
Lemma 6.2. Suppose that E is another Banach space and Q : V W E is
: V W E be corresponding linear map on V W
a bilinear form
and let Q
to E. Then
Q
= kQk . Because of this property, in the future we will no
op
vi wi , we have
It is easy to check that this satisfies the properties of norm modulo showing
kk = 0 implies = 0.
:=
n
X
vi wi
i=1
n
with {vi }i=1 being a linearly independent set and each wi 6= 0. Choose V
such that (vi ) = i1 and W such that (w1 ) = 1. Then Q (v, w) :=
(v) (w) is a continuous bilinear form on V W with kQk = kk kk > 0.
Thus we have
1 = |Q ()| kQk kk
from which it follows that kk > 0.
60
for all 0 s u t T.
(6.1)
S 2 (V ) := { V V : = } .
1
2
Xst = 1 + Xst
+ Xst
.
[a, b] := ab ba (V ) and
Let us observe by the multiplicative property that for all t [0, T ] , Xt,t =
Xt,t Xt,t , i.e. Xt,t = 1 for all t [0, T ] .
a b := ab + ba S 2 (V ) .
Observe that if V V, then
1
1
= ( + ) + ( ) S 2 (V ) 2 (V )
2
2
and in particular,
1
a b = (a b + [a, b]) ,
2
(a + v + ) (b + w + ) = ab + (aw + vb) + (b + a + v w) .
We are now going to drop the tensor symbol from the notation.
Lemma 6.6. The subset,
G (V ) := {g A (V ) : g = 1 + v + with v V and V V } ,
1
1
Xst
Xtu
w = v and = vw = + v .
rpaths
2
Xtu
1
2
1
2
1
2
1 + Xsu
+ Xsu
= 1 + Xst
+ Xst
1 + Xtu
+ Xtu
1
2
1
2
1
1
= 1 + Xst
+ Xtu
+ Xst + Xtu
+ Xst
Xtu
job:
(6.3)
2
Xst
V V = S 2 (V ) 2 (V ) .
60
(6.2)
2
Xsu
for all 0 s t u T.
so that
Page:
1
1
1
Xsu
= Xst
+ Xtu
and
:= 1 + x (t) x (s) +
(6.4)
where the latter integral is the Young integral. An alternative way to look at
this function is to observe that
Z t
Xst = 1 +
Xs dx ( )
(6.5)
s
and this differential equation uniquely determines Xst . Let us also check directly
that Eqs. (6.3) holds. Equation (6.2) holds trivially.
Rt
1
2
We have in this case Xst
= x (t) x (s) and Xst
= s (x (v) x (s)) dx (v) ,
therefore,
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Proof. We have,
Z
(x (v) x (s)) dx (v)
2
Xst
=
s
2
Xst
61
1
1
(x (t) x (s)) dx (v) = Xst
Xtu
1
dv [x (v) x (s)] = [x (t) x (s)] = Xst
2
as desired.
and
Example 6.10. Suppose again that p < 2 and W := x+A with x Cp ([0, T ] , V )
and A Cp ([0, T ] , V V ) . Then
Z
Xst = 1 + W (t) W (s) +
dW (u) dW (v)
2
2
Xst
Xst
=
Z
(x (v) x (s)) dx (v)
suvt
Z
:= 1 + x (t) x (s) + A (t) A (s) +
(6.6)
Z
Xst = 1 +
Xs dW ( )
(6.7)
Ggeo (V ) := g = 1 + a + A G (V ) : A + A = a2 ,
1
Xst
2
2
where Xst
is chosen so that Xst := 1+x (t)x (s)+Xst
satisfies Chens identity.
Xst
1 2
Xst
Z
+
[x (v) x (s) , dx (v)] .
61
1
1
g 1 = 1 a a2 A + a2 = 1 a + a2 A
2
2
1
2
= 1 a + (a) A Ggeo (V )
2
(6.8)
2
This shows that it is only the anti-symmetric part of Xs,t
which does not depend
continuously on x in the sup-norm topology.
Page:
(6.9)
so that
1
1
= 1 + Xst
+
2
1
g = 1 + a + a2 + A with A 2 (V ) .
2
1 2
2a .
job:
rpaths
and
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
62
1 2
1 2
gh = 1 + a + a + A
1+b+ b +B
2
2
1
1
= 1 + a + b + a2 + b2 + ab + A + B
2
2
1
1
2
= 1 + a + b + (a + b) (ab + ba) + ab + A + B
2
2
1
1
2
= 1 + a + b + (a + b) + (ab ba) + A + B
2
2
1
1
2
= 1 + a + b + (a + b) + [a, b] + A + B Ggeo (V ) .
2
2
h i
]
which we now
The Lie bracket is then determined by the formula, [,
] = ,
work out. Let f : G R be a smooth function, then
f (g) = (g f 0 (g) (g)) = f 00 (g) (g, g) + f 0 (g) .
Since f 00 (g) is symmetric (mixed partial derivative commute) we find,
h i
^
f (g) = f 0 (g) ( ) = (
,
)f (g) .
We summarize these results in the following proposition.
As with any Lie group, H, we may associate a Lie algebra, Lie (H) = Te H.
Lemma 6.14. For G = G (V ) and Ggeo = Ggeo (V ) we have
Lie G = V V V and Lie Ggeo = V 2 (V ) .
(6.10)
Proof. Let g (t) = 1+x (t)+A (t) be a smooth path in G such that g (0) = 1,
then
g (0) = x (0) + A (0) V V V.
Conversely if a + A V V V then g (t) = 1 + t (a + A) G is a smooth
path such that g (0) = 0 and g (0) = a + A. Therefore Lie (G) = V V V.
A smooth path g (t) Ggeo may be written as
1
g (t) = 1 + x (t) + x2 (t) + A (t) with A (t) 2 (V ) .
2
where h (t) is any smooth curve in G such that h (0) = 0 and h (0) = for
example take h (t) = 1 + t. Writing g = 1 + , we find,
d
(g) = |0 (1 + ) (1 + t) = g .
dt
Page:
62
job:
rpaths
et (1) = g (t) = 1 + at +
t2 2
a + At.
2
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
1
1
e = 1 + a + a2 + A = 1 + + 2 = exp () .
2
2
Proof. We have
1 2
1 2
e e = 1++
1++
2
2
1
1
= 1 + + + 2 + 2 +
2
2
1
1
2
= 1 + + + ( + ) ( + ) +
2
2
1
1
2
= 1 + + + ( + ) + [, ]
2
2
++ 21 [,]
=e
.
Alternatively; compute
1 2 1 2
log (exp () exp ()) = log 1 + + + + +
2
2
1
1
1
2
= + + 2 + 2 + ( + )
2
2
2
1
= + + [, ] .
2
1
log (1 + ) = 2 .
2
Moreover, g G is in Ggeo iff log (g) L (V ) = Lie (Ggeo ) .
Proof. To prove the last assertion, observe that if
1
1 + = e = 1 + + 2
2
Corollary 6.18. Suppose that U (t) Lie (G) is a finite variation curve with
U (0) = 0 and g (t) solves,
1
1
= 2 =
2
2
1
2
2
1
= 2 .
2
then
Z
1 t
g (t) = exp U (t) +
[U ( ) , dU ( )] .
2 0
If U (t) = u (t) + A (t) with u (t) V and A (t) V V, we may write g (t) as,
Z
1 t
[u ( ) , du ( )] + A (t) .
g (t) = exp u (t) +
2 0
Proof. We know that the solution to the ODE is given by
Z t
Z
g (t) = 1 +
dU ( ) +
dU () dU ( )
0 t
e e =e
++ 12 [,]
U ( ) dU ( )
Z
1 2
1 t
= 1 + U (t) + U (t) +
[U ( ) , dU ( )]
2
2 0
Z t
1
= exp U (t) +
[U ( ) , dU ( )] .
2 0
0
1
[, ] for all , Lie (G) ,
2
then Lie (G) becomes a group such that exp : Lie (G) G and exp : L (V )
Ggeo are Lie group isomorphisms.
Page:
63
= 1 + U (t) +
In particular if we define
=++
63
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
64
=e
u v u v
e e .
[ a, a] d = 0.
[xa ( ) , dxa ( )] =
Now let
u v u v
e e =e
=e
xi (t) =
u v u+v+ 21 [u,v]
Then
Z 1
Z
[xi (t) , dxi (t)] =
= eu eu+[u,v] = e
u+[u,v]u+ 21 [u,u+[u,v]]
1
(ui (cos 2t 1) + vi sin 2t) for 0 t 1.
2
= e[u,v] .
Z
= [ui , vi ]
e =e
Pm
m
Y
[ui ,vi ]
i=1
m
Y
(6.11)
Pm
i=1
eui evi eui evi .
i=1
(cos 2t 1) cos 2t + sin2 2t dt
= [ui , vi ]
Let x := x1 xm xa by which we mean follows x1 , then x2 , . . . , then xm ,
and then xa . Then
Z
1 1
x (1) +
[x ( ) , dx ( )]
2 0
Z
m Z
1X 1
1 1
[xa ( ) , dxa ( )] +
[xi ( ) , dxi ( )]
=a+
2 0
2 i=1 0
m
If in addition, a V, then
=a+
ea+A = ea eA = ea
m
Y
eui evi eui evi .
1X
[ui , vi ]
2 i=1
i=1
It is now easy to see how to construct the desired path x (t) . We determine
x (t) so that it is continuous and satisfies, x (t) = a for 0 t 1, x (t) = ui
for t 4 (i 1) + (1, 2) , x (t) = vi for t 4 (i 1) + (2, 3) , x (t) = ui for
t 4 (i 1) + (3, 4) , and x (t) = vi for t 4 (i 1) + (4, 5) for i = 1, 2, . . . , m.
With this definition it follows that
g (4 (m 1) + 5) = ea+A
as desired.
Second Proof. We know the solution to Eq. (6.11) is given by
Z
1 t
[x ( ) , dx ( )] .
gx (t) = exp x (t) +
2 0
Z t
Xst = 1 + (x (t) x (s)) +
(x (v) x (s)) dx (v)
s
Z
1 t
[x (v) x (s) , dx (v)] .
= exp (x (t) x (s)) +
2 s
64
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Lemma 6.23 (Characterization of MFs). The space of (algebraic geometric) multiplicative functionals are in one to one correspondence with functions
(Y : [0, T ] Ggeo (V )) Y : [0, T ] G (V ) such that Y (0) = 1. (The latter
conditions is an arbitrary normalization.) The correspondence is given by
Y Xst := Ys1 Yt and Xst Yt := X0t .
Proof. Given a multiplicative function, X, let Yt := X0,t . Then for (s, t)
we have
Yt = X0t = X0s Xst = Ys Xst = Xst = Ys1 Yt .
Conversely if Y : [0, T ] G (V ) and Xst := Ys1 Yt , then
Xst Xtu = Ys1 Yt Yt1 Yu = Ys1 Yu = Xsu
st
such that X = X , then st := Xst
X
is an additive functional.
may be written as in Eq. (6.12) with
Proof. By Proposition 6.24, X and X
Therefore,
A replaced by A for X.
2
2
st
st = Xst
X
= At As At As
= At At As As
which is an additive functional.
Alternatively, we make use of Chens identity to find,
2
2
2
2
1
1
2
2
1 1
2
su
st
tu
st
X2 X
= Xsu
X
= Xst
+ Xtu
+ Xst
Xtu
X
+X
+X
Xtu
su
2
2
2
2
st
tu
= Xst
X
+ Xtu
X
2
2
= X2 X
+ X2 X
.
as desired.
Proposition 6.24. If Yt = 1 + xt + At G (V ) then
Xst = 1 + (xt xs ) + At As + x2s xs xt
1
1
1
= exp (xt xs ) + At x2t As x2s [xs , xt ]
2
2
2
st
(6.13)
(6.14)
(6.15)
Proof.
1
Xst = Ys1 Yt = (1 + xs + As ) (1 + xt + At )
= 1 xs As + x2s (1 + xt + At )
= 1 + (xt xs ) + At As + x2s xs xt
1
2
= 1 + (xt xs ) + (xt xs )
2
1 2
xt + x2s xs xt xt xs + At As + x2s xs xt
2
1
1
1
1
2
= 1 + (xt xs ) + (xt xs ) + At x2t As x2s [xs , xt ]
2
2
2
2
Page:
65
job:
rpaths
tu
(6.12)
65
macro:
svmonob.cls
k
k
X
X
ai bki
(ea )i (eb )ki =
i! (k i)!
i=0
i=0
(a + b)k
= (ea+b )k ,
k!
date/time:
11-Mar-2009/12:03
66
st
k
(a + b)
= (a + b) (a + b)
k
= (a + b) 1
k
X
ai bki k
=
1
i! (k i)!
i=0
k
X
ai b(ki)
.
i! (k i)!
i=0
Definition 6.30. Let p [1, ). A p (geometric) rough path is a multiplicative functional, (X : Ggeo ) X : G such that;
Therefore
Xst = Ys1 Yt = exs ext = e(xt xs )
is a multiplicative functional. Indeed if s < u < t then
k
X
ki
i
Xsu
Xut
=
i=0
xs xt
(xt xs )2
:=
xs (xs xt ) = (xs xt ) xs +
2
2
xs + xt
1 2
= (xs xt )
= (xs x2t )
2
2
1
= (x2t x2s ).
2
k
X
(xs xu )j (xt xu )ki
i!
(k i)!
i=0
1
k
= (xt xu + xs xu )k = Xs,t
k!
as desired.
Example 6.29. Suppose dim(V ) = 1, i.e. V = R, and x : [0, T ] R is a contin
P
uous path and now let Yt = 1 xt . Then Yt1 = 1 +
xkt so that
1. X is continuous,
2. Vp X 1 + Vp/2 X 2 < .
Definition 6.31. If x Cp ([0, T ] V ) is given. We say that X is a (geomet1
= x (t) x (s)
ric) p - lift if X is a p (geometric) rough path such that Xst
for all (s, t) .
Theorem 6.32. Let x Cp ([0, T ] , V ) with p < 2. Then x has precisely one p
lift which is given by
Z t
2
Xst
=
(x ( ) x (s)) dx ( )
(6.16)
s
Z
1
1 t
2
[x ( ) x (s) , dx ( )] ,
(6.17)
= (x (t) x (s)) +
2
2 s
k=1
2i
i
0
2
1
1
2
0
Xsu
Xut
= Xsu
Xut
+ Xsu
Xut
+ Xsu
Xut
i=0
as desired.
Let us now consider (at level 2) the difference, , between the two multiplicative functionals in Examples 6.28 and 6.29,
Page:
66
job:
rpaths
2
Xst
(2/p) Vp (x () x (s) : [s, t]) Vp (x : [s, t]) = (2/p) (s, t)2/p .
2/p
Hence it follows that Vp/2 X 2 (2/p) (0, T )
< and the existence
assertion is proved.
For uniqueness, suppose that Y is another lift. Then we know st := Yst2
2
Xst V V is an additive functional with Vp/2 () < . As p/2 < 1, this
implies that st = 0 for all (s, t) which gives the uniqueness assertion of
the theorem.
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
(6.20)
and hence v (X) (0, T ) < for all i. This implies that X p (V ) .
For the converse, an application of Theorem 2.32 shows that both
p/2
Vpp X 1 : [s, t] and Vp/2 X 2 : [s, t] are controls. It is now easy to verify that
Eq. (6.19) holds for this control.
The following theorem due to Lyons and Victoir [14] shows that there are
plenty of p rough paths.
Theorem 6.34 (Extension Theorem). Let 1 p < and x
Cp ([0, T ] , V ) , then there always exists a geometric p lift, X : Ggeo ,
of x.
Proposition 6.35 (Non-Uniqueness of Lifts). If 2 p < and x
Cp ([0, T ] , V ) , there exists an infinite number of (geometric) p lifts of x.
2
7
Homogeneous Metrics
In this chapter we are going to introduce some metrics on G and Ggeo which
will allow us to relate p rough paths to the more familiar paths of finite p
variation relative to one of these metrics. We begin with some generalities
about left invariant metrics and the associated p variation spaces.
!1/p
(x, y : ) =
d (t x, t y)
!1/p
[d (t x, e) + d (e, t y)]
Vp (x : ) + Vp (y : ) Vp (x) + Vp (y)
and hence,
(x, y) Vp (x) + Vp (y) < .
dp xt , xt
!1/p
Vp (x) :=
sup
kt xk
(dp (t x, t y) + dp (T x, T y))
As t [0, T ] was arbitrary it follows that
(7.2)
0tT
and
!1/p
Vp (x : ) .
P(0,T )
(x, z : ) =
d (t x, t z)
We also define
!1/p
!1/p
(x, y : ) :=
(x, y) .
!1/p
X
Definition 7.1. Let C0,p ([0, T ] , G) := {x C ([0, T ] G) : x (0) = e and Vp (x) < }
dp (t x, t y)
p
X
1
(t x) t y
!1/p
[d (t x, t y) + d (t y, t z)]
!1/p
and set,
(x, y) :=
sup
P(0,T )
(x, y : ) .
(7.1)
d (t x, t y)
!1/p
+
= (x, y : ) + (y, z : )
(x, y) + (y, z) .
X
t
d (t y, t z)
70
7 Homogeneous Metrics
!1/p
which shows that satisfies the triangle inequality. It is clear from the definition
of that (x, y) = (y, x) and from Eq. (7.2) that (x, y) = 0 implies x = y.
Moreover we now see that if x, y C0,p ([0, T ] V ) , then (x, y) (x, e) +
(e, z) < so that is finite on C0,p ([0, T ] V ) .
3. To finish the proof we must now show is complete. So suppose that
{xn }n=1 C0,p ([0, T ] , G) is a Cauchy sequence. Then by Eq. (7.2) we know
that xn converges uniformly to some x C ([0, T ] G) . Moreover, for any
partition, P (0, T ) we have
(x, xn : ) (x, xm : ) + (xm , xn : )
(x, xm : ) + (xm , xn ) .
and therefore,
d (e, t (xy)) = d e, t y yt1 t xyt = d t y 1 , yt1 t xyt
d t y 1 , e + d e, yt1 t xyt
= d (e, t y) + d e, yt1 t xyt
d (e, t y) + C (yt ) d (e, t x) .
70
job:
rpaths
= yt1
yt yt1 t xyt = t y yt1 t xyt
Page:
d (e, t (xy))
We now go back to the specific case at hand. In our case the groups G and Ggeo
are equipped with a dilation structure.
(e, xy : ) =
2. k
1(g)k
= || kgk (homogeneous) for all R ,
3.
g
= kgk (symmetric), and
4. kghk kgk + khk (subadditive).
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
We will give one example of such a norm in Corollary 7.8 below. We will
give another example on Ggeo in Section 7.3 below.
71
kgk=1
By compactness, 0 < c < C < . For general g G \ {1} , choose > 0 such
that k (g)kG = 1, i.e. take := 1/ kgkG . Then we know that
and
( (g)) = || (g) for all g G (V ) and R .
Proof. Only the subadditivity requires any proof here. Let := (g) and
:= (h) where h = 1+h1 +h2 . Observe that kg1 k , kh1 k , 2 kg2 k 2 ,
and 2 kh2 k . With this notation we have,
c | (g)| =
|g|
C
kgkG
and therefore,
c kgkG |g| C kgkG for all g G.
gh = 1 + g1 + h1 + (g2 + h2 + g1 h1 ) ,
and
p
(gh) = max kg1 + h1 k , 2 kg2 + h2 + g1 h1 k
p
max kg1 k + kh1 k , 2 kg2 k + 2 kh2 k + 2 kg1 k kh1 k
p
max + , 2 + 2 + 2 = max ( + , + )
defines a left invariant homogeneous (i.e. d ( (g) , (h)) = || d (g, h)) metric
on G.
Proof.
The proof of this proposition is easy. For example, d (g, h) = 0 iff
1
g h
= 0 iff g 1 h = 1 iff g = h;
G
= + = (g) + (h) .
d (h, g) =
h1 g
G =
g 1 h
G = d (g, h) , and
d (g, k) =
g 1 k
G =
g 1 h h1 k
G
g 1 h
+
h1 k
= d (g, h) + d (h, k) .
Page:
71
(7.4)
job:
rpaths
For the rest of this section we will assume that () and kkG are as defined
in Lemma 7.7 and 7.8 above.
Theorem 7.11. Let X : G (or Ggeo ) be a continuous multiplicative functional and Y C ([0, T ] G) be associated to X via, Y (t) := X0,t or equiva1
lently by, Xst = Y (s) Y (t) for all (s, t) . Then X is a p rough path iff
Y has finite p variation.
Proof. The point is that
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
72
7 Homogeneous Metrics
Vpp (Y : ) =
dp (Y (t ) , Y (t)) =
Xst =
B dB +
1 Bs
X : +
p/2
Vp/2
X : .
0
t
B dB Bs (Bt Bs )
s
Z
p/2
B dB + Bs2 Bs Bt
B dB
= 1 + Bt Bs +
= 1 + Bt Bs +
1
= 1 + Bt Bs +
Z t
B dB
1 + Bt +
0
Z
Z
Bs2
! !
r
p
2
Xt ,t
1
p
Xt ,t
+
X
X
Xt ,t
p
G
Vpp
p
X
1
Y (t ) Y (t)
(B Bs ) dB a.s.
s
2
Therefore,
E [dp (Ys , Yt )] = E dp 1, Ys1 Yt = E [dp (1, Xst )]
"
Z t
1/2 !p #
CE
|Bt Bs | +
(B Bs ) dB
.
Using this theorem we can see fairly easily that finite dimensional Brownian
motions have geometric p - lifts for all p > 2. This is the content of the next
theorem.
Theorem 7.12 (Enhance Brownian Motion). Let {Bt }t0 be a Rd valued
Brownian motion.
Then for all (0, 1/2) , there exists a Xst = 1 + Bt Bs +
2
Ggeo Rd such that
Xst
Z
1/2 p
T
E [dp (Ys , Yt )] CE |bT | +
b db
0
"
Z 1
1/2 !p #
b db
= CE
T |b1 | + T
0
2 1/2
|Bt Bs | + Xst
C |t s| a.s.,
where C is a random finite (a.s.) constant.
Proof. Let
Z
Z
B dB = Yt := 1 + Bt +
Yt := 1 + Bt +
0
B dB + tC
0
= C (p, n) T
p/2
= C (p, n) |t s|
d
For the second line we have used the Brownian scaling, b = T bT 1 () , to
conclude that
d
b + b+ b+ b = T bT 1 + bT 1 + bT 1 + bT 1
and therefore,
Z
Pd
p/2
b db = T
0
b db .
0
d (Ys , Yt ) C |t s|
we have,
Page:
72
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
g (t1 ) g (t0 ) =
Proof.
Notice that
S is a continuously differentiable function such that
S (t) =
g 1 (t) g (t)
. Moreover if S (t0 ) = S (t1 ) for some 0 t0 < t1 T,
then g 1 ( ) g ( ) = 0 for [t0 , t1 ] and therefore g ( ) is constants on [t0 , t1 ] .
Thus it makes sense to define, : [0, ` (g)] Ggeo by the equation,
(S (t)) := g (t) for all 0 t T.
Now suppose that 0 s0 < s1 ` (g) . Using the intermediate value theorem, there exists 0 t0 < t1 T such that S (t0 ) = t0 and S (t1 ) = s1 .
Therefore,
Z t1
1
g ( ) g ( )
d = S (t1 ) S (t0 ) .
d ( (S (t1 )) , (S (t0 ))) ` g|[t0 ,t1 ] =
t0
g ( ) d.
t0
Lemma 7.14. Suppose that g : [0, T ] Ggeo is a smooth horizontal path, then
there exists a unique path, : [0, ` (g)] Ggeo , such that g (t) = (S (t)) for
Rt
g|[0,t] . Moreover,
all 0 t T where S (t) := 0
g 1 ( ) g (
)
d arc-length
is absolutely continuous, horizontal, and
1 (s) 0 (s)
= 1 for a.e. s.
73
t1
t0
Z
1
(S ( )) 0 (S ( ))
S ( ) d =
S(t1 )
S(t0 )
1
(s) 0 (s)
ds
1
This then implies that
(s) 0 (s)
= 1 for a.e. s showing that is
parametrized by arc-length as desired.
Theorem 7.15. Assuming dim (V ) < , we have:
1. d is a metric on Ggeo compatible with the natural induced topology.
2. d is left invariant, i.e. d (uw, u y) = d (w, y) for all u, w, y Ggeo .
3. d is homogeneous, i.e. for all R and w, y Ggeo ,
d ( (w) , (y)) = || d (w, y) .
(7.5)
4. Let (w) := d (e, w) . Then there are constants, 0 < c < C < such that,
p
p
c kak + kAk ea+A C kak + kAk .
(7.6)
S(t0 )
t0
Page:
73
g (t)
job:
rpaths
macro:
g (t) =
svmonob.cls
d
d
1
1
|0 g (t) g (t + s) =
|0 g (t) g (t + s) = w (t)
ds
ds
date/time:
11-Mar-2009/12:03
74
7 Homogeneous Metrics
which shows that g is a horizontal path joining (w) to (y) and moreover
` (g ) = || ` (g) . Equation (7.5) follows easily from this observation.
We now check that d is a metric. Since g (T t) is a path taking y to w with
` (g (T )) = ` (g) , it follows that d (w, y) = d (y, w) . If g is a horizontal path
from w to y and k is a horizontal path from y to z, then g k is a horizontal
path from w to z such that
d (w, z) ` (g k) = ` (g) + ` (k) .
Taking the infimum over g and k joining w to y and y to z respectively shows
that d satisfies the triangle inequality.
We now must still show that d (w, y) = 0 implies w = y. To prove this let
us consider another metric,
d0 (w, y) := inf {`0 (g) : g (0) = w, and g (T ) = y}
where now; g is not assumed to be horizontal and
Z
`0 (g) :=
0
and therefore,
1
g g
dt
` (g) =
0
Page:
74
job:
rpaths
t1
2
kxk
+
A
dt
exp
1
A [u, v]
2
` (gx ) =
kx (t)k dt
0
1p
|A|.
i<j
ea+A = ea eA (ea ) + eA
Y
kak + eAij [ei ,ej ]
2
1
2
2
g g
= kxk
+
A x x
2
2
2
x x
= kxk
+
A
+ kx xk
2 A,
2
2 1
2
2
2
kxk
+
A
+ kx xk
A
kx xk
2
2
2
= kxk
1 1 kx xk
+ (1 )
A
2
n
o
2
2
= kxk
1 1 1 kxk + (1 ) A .
2
1
2
2
g g
kxk
+ (1 )
A
1
g g
dt C (, )
Therefore if A =
t1
1
g (t) g (t)
dt.
V V V
i<j
X
kak +
eAij [ei ,ej ]
i<j
q
1X
2 |Aij | C ea+A
i<j
p
C kak + kAk = C ea+A
kak +
where
p
ea+A = kak + kAk.
This gives the upper bound in Eq. (7.6). This bound also shows is continuous.
Indeed,
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
= lim
c ( (w)) = (w) =
(w)
(w)
which gives the lower bound in Eq. (7.6). The bounds in Eq. (7.6) shows that
the metric topology associated to d is the same as the vector space topology.
In particular, from Eq. (7.6), we know that {wn }n=1 is a bounded sequence
and therefore has a convergent subsequence in the usual topology and therefore
in the d topology. It is now easy to conclude that {wn } is d convergent.
Therefore (Ggeo , d) is a complete metric space.
We now prove the last assertion about geodesics we will follow Montgomery [16]. Using Lemma 7.14, we may choose n : [0, `n ] Ggeo which are
absolutely continuous horizontal paths with unit speed a.e. such that (0) = x
and (`n ) = y and `n d (x, y) as n . By letting gn (t) := n (t`n /d (x, y)) ,
1
we then have gn (0) = x and gn (`) = y with gn (t) g n (t) = `n /d (x, y) for a.e.
t. It now follows by the AscolliArzela and Banach Anolouge theorem, that after
passing to a subsequence if necessary, we may assume that gn g uniformly
on [0, `n ] and g n (t) u (t) weakly in L2 ([0, `]) . For any A , we have for
any bounded measurable , that
Page:
75
job:
rpaths
1
(t) g (t) u (t) dt = 0
1
allowing us to conclude that g (t) u (t) = 0 a.e. t and therefore
g (t)
Z `
1
(t) g (t) u (t) dt
(t) kk dt
1
(t) gn (t) g n (t) dt
We are now ready to use the dilation invariance to prove the lower bound.
By compactness, we have,
where in the we have used the fact that gn g uniformly and supn kg n k < .
Now if (V ) = 0, we find that
Z
| (w) (z)| C
and hence,
s
!
1
kb ak +
B A 2 [a, b a]
0 as z w.
75
Z `
1
1
(t) g (t) g n (t) dt
(t) g (t) u (t) dt = lim
1
g (t) u (t) kk a.e. t.
1
Since V was arbitrary we may conclude that
g (t) u (t)
1 for a.e.
V
t. Moreover we have,
g (t1 ) g (t0 ) = lim [gn (t1 ) gn (t0 )]
n
Z `
Z `
= lim
1(t0 ,t1 ] ( ) g n ( ) d =
1(t0 ,t1 ] ( ) u ( ) d
n
from which it follows that g is absolutely continuous and g
(t) = u (t) a.e.
t.
1
Thus g is a horizontal absolutely continuous path such that
g (t) g (t)
1
for a.e. t. Therefore we may conclude that
d (x, y) ` (g) =
Z `
1
g (t) g (t)
dt ` = d (x, y) .
0
1
Thus we must in fact d (x, y) = ` (g) and
g (t) g (t)
= 1 for a.e. t as desired.
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
8
Rough Path Integrals
Throughout this chapter we will be assuming that 2 p < 3. Our first goal
is to show how to make p rough paths out of almost p rough paths.
Proof. (Uniqueness.) Suppose Zst is another such rough path so that Eq.
replaced by Z. . Then by the triangle inequality we have
(8.2) holds with X
h
ii
st
2Ci (s, t) .
Zst X
(8.3)
1
1
st
is an additive functional, it follows from Lemma 5.21 and Eq. (8.3)
X
As Zst
1
1 . Now that X
1 = Z 1 we further know that X
2 Z 2 is also an
that Z = X
additive functional. So another application of Lemma 5.21 along with Eq. (8.3)
2 = Z 2 . Thus we have shown X
= Z as desired.
implies that X
(Existence.) 1. Notice that the condition in Eq. (8.1) for i = 1 is the
1
is an almost (, p) additive functional. Therefore we may
statement that Xst
1
st
, such
apply Theorem 5.24 to find a finite p variation additive functional, X
that Eq. (8.2) holds for i = 1, with K () = () in this case.
1
2
st
2. Let Zst := 1 + X
+ Xst
. I now claim that Z is still an (, p) almost
rough path. Indeed,
2
2
2
2
1 X
1
Xst
Xtu
X
[Zsu Zst Ztu ]
=
Xsu
st tu
2
1 1
2
2
1
1
1 X
1
Xsu
Xst
Xtu
Xst
Xtu
+
Xst
Xtu X
st tu
1 1
1 1
1 1
1 1
st
C (s, u) +
Xst
Xtu Xst
Xtu
+
Xst
Xtu X
Xtu
1
1
1
1
1
1
tu
st
C (s, u) +
Xst
X
X
Xtu
+
Xst
Xtu
1
1
C (s, u) +
Xst
+
X
tu
C () (s, u)
"
!
#
1/p
1/p
C (s, t) + C (t, u)
C 1+
() (s, u)
+C () (t, u)
K (, (0, T )) (s, t) .
3. Now suppose that P (s, t) and . Then
Z ( \ { }) Z () = Z [s, ] Z ,+ Z , Z,+ Z [+ ,t]
2
= Z [s, ] Z ,+ Z , Z,+ Z [+ ,t]
2
= Z ,+ Z , Z,+ ,
78
Lemma 8.5. Let a, b be elements of a Banach algebra, i.e., |ab| |a||b|, then
2
a b2 |a + b| |a b|.
i
= 0 for i = 0, 1.
2
kZ ( \ { }) Z ()k = Z ,+ Z , Z,+ C ( , + ) .
a2 b2 =
Comparing this identity with that of Eq. (5.34), we now see that we may follow
the proof of Proposition 5.23 and Theorem 5.24 verbatim with X replaced by
2
2
st
exists and satisfies,
Z 2 = X 2 in order to learn, lim||0 X ()st := X
2
2
CK (, (0, T )) (s, t) for all (s, t) .
Xst Xst
|a2 b2 | =
r
X
n
n
X
X
X
=1+
ai +
Ai +
ai aj .
(8.4)
i=1
i=1
r
X
X () =
a1 . . . ar+1 b1 . . . br+1
= (a1 . . . ar b1 . . . br ) ar+1 + b1 . . . br (ar+1 br+1 )
r
X
=
b1 . . . bi1 (ai bi ) ai+1 . . . ar ar+1 + b1 . . . br (ar+1 br+1 )
X2 , +
X1 , X1 , .
1
st
X
= lim
||0
X1 , and
(8.8)
2
st
X
= lim
||0
job:
rpaths
(8.7)
, : <
i=1
78
(8.6)
i=1
Page:
Proof. This is easily proved by induction. Indeed, for i = 1, the right side
of Eq. (8.4) is a1 b1 and by induction,
i<j
|a1 . . . ar b1 . . . br | r r .
i=1
i=1
r+1
X
1
[(a + b)(a b) + (a b)(a + b)]
2
macro:
svmonob.cls
X2 , +
X1 , X1 , .
(8.9)
, : <
date/time:
11-Mar-2009/12:03
||0
lim [X () Z ()]st = 0.
||0
1
Xst
r
X
Zt2i1 ti ,
Z t
1
Zst
:= z (t) z (s) =
f (x ( )) dx ( ) , and
s
Z t
Z t
2
Zst
:=
(z ( ) z (s)) dz ( ) =
(z ( ) z (s)) f (x ( )) dx ( ) .
Z ( [s, ]) X , Z , X ( [, t])
Z ( [s, ]) X1 , Z1 , X ( [, t]) .
Z t
1
Zst
:= z (t) z (s) =
f (x ( )) dx ( )
s
Z t
o
Xn
1
1
1
1
1
1
Zs,
X
Z
+
X
Z
X
(
[,
t])
.
,
,
,
,
s
1
2
= f (x (s)) Xst
+ f 0 (x (s)) Xst
2
Zst
Z
=
K (, , X)
( , ) 0 as || 0.
f (x ()) dx () f (x ( )) dx ( )
s t
Z
f (x ()) f (x ( )) dx () dx ( )
=
s t
Z
dz () dz ( ) =
s t
X
1
+ () (, t) K ( , )
Zs,
+
X,t
(8.10)
and
(x ( ) x (s)) dx ( )
s
then gives,
Zt0 t1 . . . Zti2 ti1 Xti1 ti Zti1 ti Xti ti+1 . . . Xtr1 tr
i=1
2
Xst
:= x (t) x (s) ,
:=
Z t
z (t) :=
f (x ( )) dx ( ) ,
Xt2i1 ti
79
2
Xst
. (8.11)
s t
2
Proposition 8.8. Let x Cp ([0, T ] , V ) and Xst = 1 + (x (t) x (s)) + Xst
is
a p lift of x to a p rough path. Further suppose that is a control such that
i
Xst
(s, t)i/p for all (s, t) .
(8.12)
Page:
79
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
80
p/2
1
ft fs s Xst
C (s, t)
kk1 = |0 | + N1 () ,
kyk1 = |y0 | + N1 (y) , and
k(y, )k2 = |0 | + |y0 | + N1 () + N2 (y, )
= kk1 + |y0 | + N2 (y, ) .
and
1/p
|t s | C (s, t)
Similarly,
if X is a p rough path controlled by , we define N1 X 1 and
N2 X 2 to be the best constants such that
2
1
N2 X 2 (s, t)2/p .
Xst N1 X 1 (s, t)1/p and Xst
We also let
2/p
C (f 00 ) (s, t)
kXk := N1 X 1 + N2 (X) .
and
1/p
|s | |0 | + N () (0, s)
2/p
|st | N1 () (s, t)
(8.13)
(8.14)
such that
1
1/p
2/p
1
2
Zst
=
fs Xst
+ s Xst
|fs | N X 1 (s, t) + |s | N X 2 (s, t)
and
2
2
2/p
2
Zst
=
[fs fs ] Xst
|fs | N X 2 (s, t) .
Proof. Let 0 s u t T, then
and
2/p
1
1
1
2
1
2
Zsu
+ Zut
= fs Xsu
+ s Xsu
+ fu Xut
+ u Xut
1
1
2
1
2
+ su Xut
+ [s + su ] Xut
= fs Xsu
+ s Xsu
+ fs + s Xsu
1
1
2
1
1
2
1
2
= fs Xsu + Xut
+ s Xsu
+ Xsu
Xut
+ Xut
+ su Xut
+ su Xut
1
1
2
= Zst
+ su Xut
+ su Xut
Page:
80
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
and therefore,
2
1
1
1
1
+ |su | Xut
Zsu + Zut
|su | Xut
Zst
N2 (f, ) (s, u)
Z
(f, ) dX =
f dX 1 + dX 2
t
Z
1/p
f dX + dX
=1+
N1 (X) (u, t)
2/p
+ N1 () (s, u) N2 X 2 (u, t)
3/p
N2 (f, ) N1 (X) + N2 X 2 N1 () (s, t) .
2/p
81
1
Z
t
1
f dX + dX
2
.
1/p
(See Remark 8.6 for the formulas stated in Eqs. (8.15) and (8.16).)
Proposition 8.13. Continuing the notation used above we have
Similarly,
1/p
|t | |0 | + N1 () (0, t)
[Zsu Zut ] =
2
Zsu
2
Zut
1/p
|0 | + N1 () (0, T )
(8.17)
1
1
Zsu
Zut
2
2
1
2
= [fs fs ] Xsu
+ [fu fu ] Xut
+ fs Xsu
+ s Xsu
1
2
fu Xut
+ u Xut
h
i
1/p
1/p
2/p
|fst | |0 | + N1 () (0, T )
N1 (X) (s, t) + N (f, ) (s, t)
2
2
1
1
= [fs fs ] Xsu
+ [(fs + fsu ) (fs + fsu )] Xut
+ fs (fs + fsu ) Xsu
Xut
2
2
1
1
2
= [fs fs ] Xsu
+ Xut
+ Xsu
Xut
+ [fs fsu + fsu fs + fsu fsu ] Xut
(8.18)
h
1/p
max 1, (0, T )
1/p
1
1
+ fs fsu Xsu
Xut
2
Zsu
(8.19)
2
fsu ] Xut
+ fs
1
fsu Xsu
1
Xut
.
N1 (f ) max 1, (0, T )
C (, f, X) (s, t)
1/p
1/p
f dX 1 + dX
1
2
f dX 1 + dX
(
:= lim
||0
:= lim
||0
Z
Xh
f X1 , + X2
and
, : <
)
+ X2 ,
(8.16)
so that
Z
Zst := 1 +
f dX 1 + dX 2
job:
(8.22)
2/p
Proof. The proofs of these estimate are all fairly straight forward. The
first estimate is a trivial consequence of the definition of N1 and the fact that
t = 0 + 0,t . For the estimate in Eq. (8.18) we have
1
+ |st |
|fst | |s | Xs,t
h
i
1/p 1
|0 | + N1 () (0, T )
Xs,t + |st |
h
i
1/p
1/p
2/p
|0 | + N1 () (0, T )
N1 (X) (s, t) + N (f, ) (s, t) .
Page:
1/p
2
f f X
,
f X1 , + X2 , f X1 ,
(8.21)
h
i
1/p
2/p
+ N1 () N1 (X) (0, T ) + N (f, ) (s, t) ,
(8.23)
(8.15)
and
2
2
(8.20)
Z
1/p
(s, t)
rpaths
Eq. (8.21) is a simple consequence of Eq. (8.18) and Eq. (8.23) is a consequence
of Eqs. (8.17) and (8.21) and Eq. (8.24) follows directly from Eq. (8.23). For
Eq. (8.22) we have
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
82
h
i
1/p
1/p
kf k1 = |f0 | + N1 (f ) |f0 | + (0, T ) N (f, ) + N1 (X) |0 | + (0, T ) N1 ()
1/p
max (1, N1 (X)) max 1, (0, T )
k(f, )k2 .
it follows that
Z
N
F dX, f
1/p
K3/p N (F ) N1 X 1 + N1 () N2 (X) (0, T ) + kk N2 (X)
1/p
K3/p N (F ) N1 X 1 + N1 () N2 (X) (0, T )
h
i
1/p
+ |0 | + N1 () (0, T )
N2 (X)
1/p
= K3/p N (F ) N1 X 1 + N1 () N2 (X) + N1 () N2 (X) (0, T )
+ |0 | N2 (X)
h
i
1/p
1/p
K3/p kXk kF k2 (0, T ) + |0 | + N1 () (0, T )
N2 (X)
1/p
1 + K3/p kXk kF k2 (0, T ) + |0 | N2 (X)
(8.25)
1/p
kF k2
(8.26)
where
Cp (kXk) := 1 + 2 + K3/p kXk .
(8.27)
Z
1
t
1/p
F dX K (X, f ) (s, t)
s
(8.28)
Furthermore,
and thus that
Z
N
F dX, f |0 | N2 (X)
(8.30)
where
1/p
K (X, f ) := kf k N1 X 1 + kk N2 (X) (0, T )
2/p
+ K3/p N (F ) N1 X 1 + N1 () N2 (X) (0, T ) .
(8.29)
Therefore,
Z
1
t
1
F dX fs Xst
s
Z
1
t
1
2
F dX Wst + s Xst
s
2
3/p
K3/p N (F ) N1 X 1 + N1 () N2 (X) (s, t) + |s | Xst
3/p
2/p
K3/p N (F ) N1 X 1 + N1 () N2 (X) (s, t) + |s | N2 (X) (s, t)
3/p
1/p
+ K3/p N (F ) N1 (X) + K3/p + 1 N1 () N2 (X) (0, T )
1/p
(8.31)
1 + K3/p kXk kF k2 (0, T ) + |0 | N2 (X)
Moreover from Eq. (8.17) we find
1/p
1/p
kF k2 .
(8.32)
(8.33)
(8.34)
(8.35)
2/p
+ |s | N2 (X) (s, t)
h
i
3/p
1/p
2/p
K3/p kXk kF k2 (s, t) + |0 | + N1 () (0, t)
N2 (X) (s, t)
Page:
82
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
Z
F dX, f
83
1/p
Z
F dX, f
F dX + N
0
Z
|f0 | + N1 (f ) + N
F dX, f
Z
= |f0 | + N1 (f ) +
1/p
|f0 | + kXk |0 | + 1 + K3/p kXk + 1 N1 (X) (0, T ) kF k2
1/p
|f0 | + kXk |0 | + 1 + 2 + K3/p kXk (0, T ) kF k2
1/p
Hence it follows that (n) in (1 (U, ) , kk1 ) and the proof is complete.
because
1
1
st = yst s Xs,t
= ysu + yut s Xs,t
1
1
1
= s Xsu
+ su + u Xut
+ ut s Xs,t
1
1
1
= s Xsu
+ su + [s + su ] Xut
+ ut s Xs,t
1
1
1
1
= su + ut + s Xsu + Xut
Xs,t
+ su Xut
for all (y, ) D (U, ) . Hence if {(y (n) , (n))}n=1 is a Cauchy sequence
y
(m)
(m)
X
=
st
st
st
s
st
s st
2/p
1
= su + ut + su Xut
.
kk1 := |0 | + N1 () .
2/p
and hence if { (n)}n=1 is a Cauchy sequence in 1 (U, ) then it is also uniformly Cauchy. Therefore there exist a continuous function : U such
that (n) uniformly in t. Since
Page:
83
job:
rpaths
Since the right side of this equation goes to zero as n , it follows that
limn N (y (n) y, (n) ) = 0 and we have shown (y (n) , (n)) (y, )
as n in (D (U, ) , k(, )k2 ) .
We now wish to consider the mapping properties of the spaces 1 and D.
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
84
(8.38)
(8.39)
(8.40)
(8.41)
(8.42)
(8.44)
and working as above one easily proves Eq. (8.44) as well. Alternatively,
kab uvk1 = k(a u) b + u (b v)k1 k(a u) bk1 + ku (b v)k1
1/p
1/p
max 2 (0, T ) , (0, T ) + 1 [k(a u)k1 kbk1 + kuk1 kb vk1 ] .
Theorem 8.19. Suppose that f : U S is a smooth map of Banach spaces
and for a 1 (U, ) , let f (a) := f a. Then f (a) := f (a) 1 (S, ) and
the map f : 1 (U, ) 1 (S, ) satisfies the following estimates,
(8.45)
(8.46)
where
1/p
and
(8.47)
Now suppose that a and b are two paths of finite (, p) variation and that f
is a smooth function, then
N1 (f (a) f (b)) kf 0 k,b N1 (a b)
h
i
1/p
N1 (a)
+ kf 00 k,a,b |a0 b0 | + N1 (a b) (0, T )
(8.48)
and
h
i
1/p
1/p
kak1 + |a0 | + N1 (a) (0, T )
N1 (b)
kabk1 |b0 | + N1 (b) (0, T )
h
i
1/p
1/p
|b0 | + N1 (b) (0, T )
kak1 + 1 (0, T )
kak1 N1 (b)
h
n
oi
1/p
1/p
= kak1 |b0 | + N1 (b) (0, T ) + 1 (0, T )
1/p
1/p
max 2 (0, T ) , (0, T ) + 1 kak1 kbk1 .
Page:
84
job:
rpaths
h
i
1/p
kf (a) f (b)k1 kf 0 k,a,b + kf 00 k,a,b N1 (a) max 1, (0, T )
ka bk1
(8.49)
where
00
|f (bs + bst + r [as + ast (bs + bst )])| :
00
kf k,a,b = sup
s, t [0, T ] & r, [0, 1]
sup {|f 00 ()| : || kak kbk } .
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
(8.50)
Z
f (as , at )
1
0
f (as + ast ) d
(8.51)
and
The second inequality in Eq. (8.47) follows from the fact that
1/p
+ kf 0 k,b N1 (a b) (s, t)
1/p
Moreover,
(8.52)
Now
f (as , at ) f (bs , bt )
Z 1
d
dr f
+r
[a
+
(b
+
b
)]
s
st
s
st
0
0
kf 00 k,a,b |as bs + [ast bst ]|
(8.53)
where
kf 00 k,a,b = sup
.
As above, we have
where
1/p
1/p
1/p
1/p
C (0, T )
= max 1, (0, T )
max 2 (0, T ) , (0, T ) + 1
85
85
job:
rpaths
(8.55)
h
1 + 2 (0, T )
macro:
svmonob.cls
1/p
i2
(8.56)
date/time:
11-Mar-2009/12:03
86
and
kf (y, y ) f (z, z )k2 |f (y0 ) f (z0 )| + C k(y, y ) (z, z )k2
kf 0 k,y N (y, )
h
i2
1/p
+ kf 00 k,y |0 | N1 (X) + [N1 () N1 (X) + N (y, )] (0, T )
(8.57)
where
(8.64)
1/p
C = C (0, T ) , N1 (X) , kf 0 k,y,z , kf 00 k,y,z , kf 000 k,y,z , k(y, y )k2 , k(z, z )k2
(8.58)
y
z
0
00
depends on (k(y, )k2 , k(z, )k2 ) quadratically and on kf k,y,z , kf k,y,z , kf 000 k,y,z
linearly.
Proof. Let f : W S be a smooth map of Banach space, st = yst :=
1
, and
yst s Xst
f (y)
st
1
1
:= f (yt ) f (ys ) f 0 (ys ) s Xst
= f (y)st f 0 (ys ) s Xst
,
(8.59)
2
f (yt ) f (ys ) = f 0 (ys ) yst + R (ys , yt ) yst
2
1
= f 0 (ys ) s Xst
+ st + R (ys , yt ) yst
(8.60)
then
where
Z
R (ys , yt ) =
(8.61)
f 00 (ys + yst ) (1 ) d.
Thus we have
df (y) = f 0 (y) dX 1 + df (y)
f (y)
kf k,y N (y, )
2
2/p
2
+ kf 00 k,y max 1, N1 (X) max 1, (0, T )
k(y, )k2
(8.65)
wherein the last two inequalities we have made use of Eqs. (8.21) and (8.22).
Moreover by Eqs. (8.41) and (8.46)
1/p
1/p
kf 0 (y) k1 max 2 (0, T ) , (0, T ) + 1 kf 0 (y)k1 kk1
1/p
1/p
max 2 (0, T ) , (0, T ) + 1 kf 00 k,y kyk1 kk1
1/p
1/p
max 2 (0, T ) , (0, T ) + 1
(8.66)
1/p
2
max 1, (0, T )
max 1, N1 (X) kf 00 k,y k(y, )k2 kk1
1/p
2
= C (0, T )
max 1, N1 (X) kf 00 k,y k(y, )k2 kk1
1/p
where C (0, T )
is as in Eq. (8.55). Combining Eqs. (8.65) and (8.66) then
shows
kf (y, )k
where
st
2
= f 0 (ys ) yst + R (ys , yt ) yst
(8.62)
y
y 2
0
y 1
= f (ys ) st + R (ys , yt ) s Xst + st
h
i
2
2
1 2
1
= f 0 (ys ) yst + R (ys , yt ) (sy ) Xst
+ sy Xst
yst + (yst ) .
2/p
2
|f (y0 )| + kf 0 k,y N (y, ) + kf 00 k,y max 1, (0, T )
max 1, N1 (X) k(y, )k
1/p
2
+ C (0, T )
max 1, N1 (X) kf 00 k,y k(y, )k2 kk1
1/p
2
2
|f (y0 )| + kf 0 k,y N (y, ) + 2C (0, T )
max 1, N1 (X) kf 00 k,y k(y, )k2
1/p
2
2
|f (y0 )| + kf 0 k,y k(y, )k2 + 2C (0, T )
max 1, N1 (X) kf 00 k,y k(y, )k2
This proves Eq. (8.54).
Now suppose that Y and z are two X differentiable paths so that
dY = Y dX 1 + dY and dz = z dX 1 + dz
1 00
2
2/p
kf k,y N1 (y) (s, t) ,
2
(8.63)
where
Y
d N z, Y (s, s + ds)2/p and
and so
Page:
2/p
86
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
1
f (Y )st f (z)st = f 0 (Ys ) sY Xst
+ st
f (z)
1
f 0 (zs ) sz Xst
+ st
f (Y )
f (z)
1
= f 0 (Ys ) sY f 0 (zs ) sz Xst
+ st st
1
= f 0 (Ys ) sY Xst
+ f 0 (Ys ) Yst + R (Ys , Yt ) Yst2
0
2
1
f (zs ) sz Xst
+ f 0 (zs ) zst + R (zs , zt ) zst
st
f (Y )
= st
f (z)
st
2
= f 0 (Ys ) Yst f 0 (zs ) zst + R (Ys , Yt ) Yst2 R (zs , zt ) zst
= [f 0 (Ys ) f 0 (zs )] Yst + f 0 (zs ) Yst zst
2
+ [R (Ys , Yt ) R (zs , zt )] Yst2 + R (zs , zt ) Yst2 zst
.
Let us observe
1
f 00 (x + (y x)) (1 ) d
R (x, y) =
2 000
kf k,x,y .
3
so that
Z
1
00
(v,w) [f (x + (y x))] (1 ) d
(v,w) R (x, y) =
0
87
1/p
1/p
d
|0 [f 00 (x + tv + (y x + t (w v)))] (1 ) d
dt
0
Z 1
=
v+ (wv) f 00 (x + (y x)) (1 ) d
=
(8.69)
Z
=
[f 000 (x + (y x)) (v + (w v) , , )] (1 ) d.
2/p
+ kf 0 k,z N (y z, y z ) (s, t)
2
2
+ kf 000 k,y,z |(ys zs , yt zt )| yst
3
2
1
2
+ kf 00 k,z yst
zst
.
2
0
000
kf k,x,y
[|v| + |w v|] (1 ) d
0
1
1 1
= kf 000 k,x,y
|v| +
|w v|
2
2 3
1
1
1
1
1
= kf 000 k,x,y
|v| + |w v| kf 000 k,x,y
|v| + |v| + |w|
2
6
2
6
6
2
1
2
kf 000 k,x,y
|v| + |w| kf 000 k,x,y (|v| + |w|)
3
6
3
Page:
87
job:
rpaths
Moreover
2
2
yst zst
|yst + zst | |yst zst | [|yst | + |zst |] |yst zst |
1/p
1/p
N1 (y z) (s, t)
2/p
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
88
and
Combining the last three equations and using Proposition 8.13 shows
1/p
+ kf 0 k,z N (y z, y z ) (s, t)
h
i
4
1/p
2
2/p
+ kf 000 k,y,z |y0 z0 | + N1 (y z) (0, T )
N1 (y) (s, t)
3
1
2/p
+ kf 00 k,z [N1 (y) + N1 (z)] N1 (y z) (s, t)
2
from which it follows that
N (f (y) f (z) , f 0 (y) y f 0 (z) z )
h
i
1/p
kf 00 k,y,z N (y, y ) |y0 z0 | + N1 (y z) (0, T )
+ kf 0 k,z N (y z, y z )
h
i
4
1/p
2
+ kf 000 k,y,z |y0 z0 | + N1 (y z) (0, T )
N1 (y)
3
+ kf 00 k,z [N1 (y) + N1 (z)] N1 (y z)
kf 0 (y) y f 0 (z) z k1
h
i
1/p
C (p, ) kf 00 k,y,z + kf 000 k,y,z N1 (y) max 1, (0, T )
ky zk1 ky k1
+ C (p, ) kf 00 k,z kzk1 ky z k1
kf 00 k,y,z
C (p, , N1 (X))
k(y z, y z )k2 k(y, y )k2
+ kf 000 k,y,z k(y, y )k2
+ C (p, ) kf 00 k,z k(z, z )k2 k(y z, y z )k2
"
#
kf 00 k,y,z (k(y, y )k2 + k(z, z )k2 )
C (p, , N1 (X))
k(y z, y z )k2 .
2
+ kf 000 k,y,z k(y, y )k2
Assembling all of these estimates shows Eq. (8.57) with a constant as described in Eq. (8.58).
k=0
where
1
Rn+1 (v, h) =
(n + 1)!
n+1 times
n+1
z }| {
f (v + rh)( h, . . . , h )dn (r)
and
dn (r) = (n + 1) (1 r)n dr.
Notice that n is a probability measure on [0, 1] for each n = 0, 1, 2, . . . In
particular we have
Z 1
f (v + h) = f (v) +
f 0 (v + rh) dr
0
00
88
job:
rpaths
f 00 (v + rh) (h, h) (1 r) dr
and
1
1
f (v + h) = f (v)+f (v) h+ f 00 (v) (h, h)+
2
2!
0
f (v + h) = f (v) + f 0 (v) h +
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
1
X0,s
89
3. For , B,
|R1 (, )|
M
2
| |
2
and
|R2 (, )| = M | | .
and in particular we have
for
|R1 (Xs , Xt )| =
M 1 2
M 2
2/p
Xst
C (s, t)
2
2
and
1/p
R1 (, ) =
(1 r) D2 ( + r( ))( , )dr
where
1
1
1
1
(1 r) D2 (Xs + rXst
)(Xst
, Xst
)dr.
R1 (Xs , Xt ) =
0
R2 (, ) =
D2 ( + r( ))( , , )dr
and in particular
0 (Xt ) = 0 (Xs ) + R2 (Xs , Xt )
where
Z
R2 (Xs , Xt ) =
1
1
1
D2 (Xs + rXst
)(Xst
, , )dr.
Page:
89
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
9
Rough ODE
Let us not go on to understanding the meaning of the differential equation,
Z t
y (t) = y0 +
f (y ( )) dx ( ) ,
(9.1)
0
Proof. We have,
Z
st = f (y (t)) f (y (s)) =
where
Z
st :=
and in particular,
2/p
1
yst f (y (s)) Xst
C (s, t) .
(9.3)
The following lemmas shows the right side of Eq. (9.2) makes sense provided y
satisfies the constraint in Eq. (9.3).
Lemma 9.1. Suppose that p [1, 3), y Cp ([0, T ] W ) , and X : G is
a p rough path. Further suppose that
2/p
1
y (t) y (s) f (y (s)) Xst
= O (s, t)
.
0
(9.5)
92
9 Rough ODE
f (y, y ) dX
yt = y0 +
0
Z
= y0 +
(9.6)
y = f (y)
(9.7)
where
C = C((0, T )1/p , kf kC 3 (B(0,kY k kYe k )) , kY k2 ,
Ye
, kXk).
2
or equivalently that
Z t
Z t
Zt =
f (y0 + X + Z)dX Xt =
[f (y0 + X + Z) ]dX
0
0
Z t
=
g(X + Z)dX
0
Z t
d(X + Z)
dX 2
=
g(X + Z)dX 1 + g 0 (X + Z)
dX
0
Z t
=
g(X + Z)dX 1 + g 0 (X + Z)( + Z )dX 2 .
0
C
(kXk)(0,
T
)
(Y
)
f
(
Y
Y Y
=
f
p
2
2
2
1/p
C Cp (kXk)(0, T )
Y Ye
.
2
Hence on the interval [0, t] such that C Cp (kXk)(0, t)1/p < 1, it follows that
Y = Ye .
n
o
Let := inf t > 0 : Yt 6= Yet . By continuity of Y and Ye it follows that
Y = Ye . If < T , it follows from the above argument with the time interval
shifted to [, +t], then in fact Ys = Yes for all s +t. But this contradicts
the definition of unless = T .
When f = L(V, U ) is constant, it follows that the solution to Eq. (9.4)
or more precisely Eq. (9.6) is
1
yt = y0 + Xt := y0 + X0,t
.
We will solve the general case as a perturbation of the solution of the constant
case with := f (y0 ). So let
yt = y0 + Xt + zt
Page:
92
job:
rpaths
or equivalently that
Z t
Zt =
[f (y)dX 1 + f 0 [(y)y dX 2 ] Xt
0
Z t
=
([f (y0 + X + Z) f (y0 )]dX 1 + f 0 (y0 + X + Z)( + Z )dX 2 )
0
Z t
=
(g(X + Z)dX 1 + g 0 (X + Z)( + Z )dX 2 )
0
Z t
=
g (X + Z, + Z ) dX.
0
svmonob.cls
date/time:
11-Mar-2009/12:03
Letting N := 2kXk|f 0 (y0 )f (y0 )| and making the assumption that kZk2 N,
we find
1
M [|f (y0 )| + N
]+
1/p
k (Z)k2 N +Cp (kXk)(0, T )
+2C (0, T )1/p 1 + kXk2 M 2 [|f (y0 )| + N ]2
2
(9.11)
from which it follows that k (Z)k2 N provided we choose T sufficiently small
so that
1
M [|f (y0 )| + N
]+
Cp kXk(0, T )1/p
1/p
2
2
2 N.
+2C (0, T )
1 + kXk M [|f (y0 )| + N ]
2
g (X + z, + Z ) dX, g(X + z)
where as usual,
g (X + z, + Z ) = g(X + z), g 0 (X + z) + Z
which at t = 0 is given by
(9.8)
g(X
+
Z)
dX
(Z)
(
Z)
=
(Z
Z)
=
2
1 + kXk2 M 2 [|f (y0 )|+kZk2 ]2 .
(9.10)
Page:
93
job:
rpaths
93
(9.12)
svmonob.cls
date/time:
11-Mar-2009/12:03
94
9 Rough ODE
Z
principle implies there exists a unique Z F such that (Z) = Z. The desired
solution to Eq. (9.4) is then
f (yu ) f (ys ) =
Z
yt := y0 + f (y0 )Xt + Zt .
f 0 (ys + Ysu ) d
1
s Xsu
+ su
For the last assertion, notice that when f and its derivatives are assumed
to be bounded, then for Eqs. (9.11) and (9.12) to be valid, we need only
choose T such that (0, T ) with > 0 being a constant independent
of y0 . Thus if we want to solve the equation, we may choose a partition
= {0 = t0 < t1 < < tr = T } such that (tl1 , tl ) for all l. Thus we
may inductively construct the solution on [0, tl ] for l = 1, 2, . . . , r.
we see that
1 1
A = f 0 (ys ) s Xsu
Xut
Z
f 0 (ys + Ysu ) d
1
1
s Xsu
+ su Xut
Z
1
0
1 1
s Xsu
Xut
Z
1
f (ys + Ysu ) d su Xut
9.2 A priori-Bounds
1
1 1
Xut + M1 |su | Xut
|A| M2 |Ysu | |s | Xsu
3/p
Furthermore,
2
B = (f 0 (ys ) s f 0 (yu ) u ) Xut
2
= ([f 0 (ys ) f 0 (yu )] s + f 0 (yu ) [s u ]) Xut
,
Proof. Let
1
2
Wst := f (ys ) Xst
+ f 0 (ys ) s Xst
and thus
where
Yst =
1
s Xst
2
(M2 |s | N (Y ) + M1 N ()) (s, t)3/p .
|B| (M2 |Ysu | |s | + M1 |su |) Xut
+ st .
3/p
3/p
3/p
2
+ [f 0 (ys ) s f 0 (yu ) u ] Xut
1
1
1
= [f (ys ) f (yu )] Xut
+ f 0 (ys ) s Xsu
Xut
=: A + B.
2
+ [f 0 (ys ) s f 0 (yu ) u ] Since
Xut
s = f (ys ) , we have kk M0 ,
1/p
Since
and hence
N () M1 N (Y ) .
Page:
94
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
9.2 A priori-Bounds
Therefore,
95
N (Y ) + N ()
M0
(1 + M1 + M1 ) .
1
0
1
2
|Yst | |Wst | + |Yst Wst | f (ys ) Xst
+ f (ys ) s Xst
+ |Yst Wst |
2/p
M0 (s, t) + M1 M0 (s, t)
3/p
=
1/p
N (Y ) M0 + M1 M0 (0, T )
2/p
+ K (3/p) 2M2 M0 + M12 N (Y ) + M1 N () (0, T ) .
Moreover we also have
1
1
|Yst Wst | + Wst f (ys ) Xst
|st | = Yst f (ys ) Xst
2
|Yst Wst | + f 0 (ys ) s Xst
|Yst Wst | + M1 M0 (s, t)
In summary, if we define
:= 2M0 1 +
and assume
M12
, M1
1
(0, t) =
4K (3/p) (2M2 M0 + M12 )
p
1
=: <
2
2
1/p
(9.13)
p
,
|y0 | + .
(N (Y ) + N ()) and
2
(N (Y ) + N ()) .
2
1/p
1/p
|y0 | + (0, t)
Choosing 0 = t0 < t1 < t2 < < tn < tn+1 = T such that (tl1 , tl ) = p for
l = 1, 2, . . . , n and (tn , T ) p . It then follows that
N () M0 M1 +
M1
+ M1
4K (3/p) (2M2 M0 + M12 )
N (Y ) M0 + M1 M0 +
1
4K (3/p) (2M2 M0 + M12 )
N (Y ) + N () 2M0 (1 + M1 + M1 )
M1
2M0 1 +
+
M
.
1
4K (3/p) (2M2 M0 + M12 )
2/p
1/p
2M2 M0 + M12 N (Y ) + M1 N () (0, T ) .
so that
N () M0 M1 + K (3/p)
1
2
n p + (tn , T ) =
N (Y ) + N () M0 + M1 M0 + M0 M1 + (N (Y ) + N ())
n+1
X
for t [tn , T ] .
(tl1 , tl ) (0, T )
l=1
Page:
95
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
96
9 Rough ODE
(9.14)
where
p1
C := 1p = 4K (3/p) 2M2 M0 + M12
p1
M1
= 2M0 1 +
+ M1 4K (3/p) 2M2 M0 + M12
2
4K (3/p) (2M2 M0 + M1 )
(9.15)
p1
M1
(0, T )
|yt | |y0 | + 2 [M1 (|y0 | + )] 1 +
+ M1 4K (3/p) M12
4K (3/p) M12
M1
1
+ 2 [M1 (|y0 | + )] 1 +
+ M1
(9.17)
4K (3/p) M12
4K (3/p) M12
C1 (p, M1 ) [|y0 | + ] + C2 (p, M1 , )
(9.18)
Since
1/p
(tn , T )
h
i
1/p1
(tn , T ) p = (tn , T ) (tn , T )
p
i
h
1p
= (tn , T ) (tn , T ) p p
(tn , T ) 1p 1p 0,
(9.16)
M0p (0, T ) .
Remark 9.6. From Eqs. (9.16) and (9.15), in the case that f is linear, so that
f 00 0, we then have M2 = 0 and in this case we learn that
p1
M1
(M1 , p) M0 .
C = 2M0 1 +
+
M
4K (3/p) M12
=K
1
2
4K (3/p) M1
To make use of this sort of nonsensical statement (nonsensical, since M0 = if
f is linear) we must restrict our attention to solutions in a big open ball. STOP
In more detail, let T be the first exit time of y (t) from B (y0 , ) for some
cutoff . In this case we will have for y B (y0 , ) that
|f (y)| M1 (|y0 | + )
so that
M0 M1 (|y0 | + ) .
Using this estimate back in Eq. (9.16) implies that
Page:
96
job:
rpaths
macro:
svmonob.cls
date/time:
11-Mar-2009/12:03
10
Some Open Problems
Problem 10.1 (Measure theoretic approach). Is there are more measure
theoretic approach to the rough path theory? I would guess this would require
are reasonable notion of simple functions. Also, can one get rid of the continuity assumptions? (I have not done a literature search on this point, so there
probably is some work in this direction already.)
Problem 10.2 (Non explosion criteria). It is shown in Fritz and Victoir [6,
Exercise 10.61 on p. 259] that if dy = f (y) dx, x is a geomentric p variation
rough path on Rd and f has bounded derivatives to sufficiently high order,
then the equation has solutions for all time. It is reasonable to ask the following
questions;
1. What happens for non-geometric rough paths?
2. What happens in infinite dimensions, i.e. when d ?
3. What are other sufficient conditions for non-explosion?
4. Can one find necessary condititions as well?
Theorem 10.3. Let p [1, ) and n Z+ such that n 1 p < n. Suppose
that X : G(n) (V ) is a p rough path. Then for any m n there is a
: G(m) (V ) . In particular,
unique extension of X to a p rough path, X
we may extend X to a p rough path,
=
X
X
k=0
References