Reproducing Kernel Hilbert Space of A Set Indexed Brownian Motion

International
OPEN ACCESS Journal
Of Modern Engineering Research (IJMER)
| IJMER | ISSN: 2249–6645 www.ijmer.com | Vol. 7 | Iss. 5 | May. 2017 | 22 |
Reproducing Kernel Hilbert Space of A Set Indexed
Brownian Motion
Arthur Yosef
Tel Aviv-Yaffo Academic College, 2Rabenu Yeruham St., Tel Aviv-Yaffo, Israel
I. INTRODUCTION
In this article, we present the representation of set indexed Brownian motion { : }AX X A A via
orthonormal basis, based on reproducing kernel Hilbert space (RKHS). Set indexed Brownian motion is a
natural generalization of planar Brownian motion where A is a collection of compact subsets of a fixed
topological space ( , )T  . The frame of a set-indexed Brownian motion is not only a new stepto generalize
theclassical Brownian motion, but it was proven asa new look upon a Brownian motion (see [Yo09], [Yo15],
[MeYo], [He], [IvMe], [Kh], [MeNu]).
RKHS is a robust tooland can be used in a wide variety of areas such as curve fitting,signal analysis
and processing, function estimation and model description, differential equations, probability, statistics,
nonlinear Burgers equations, empirical risk minimization, fractals, machine learning and etc. (see [Par67],
[Par],[Ha],[Sc], [Va], [ScSm], [Dan], [Be], [Cu], [Ge], [Ad]).
Let’s assume we havea set indexed Brownian motion on topological and separable spaceT , with a
continuous covariance kernel :R   A A . We can associate a Hilbert space, which is the reproducing
kernel Hilbert space of real-valued functions on T that is naturally isometric to
2
( )L A .The isometry between
these Hilbert spaces leads to useful spectral representations of the set indexed Brownian motion, notably the
Karhunen-Loève (KL) theorem.(The KL theorem is a representation of a stochastic process as an infinite linear
combination of orthogonal functions, analogous to a Fourier series representation of a function on a bounded
interval).In this work, the KL representation of a set indexed Brownian motion is:
[ ]A i A iX e E X e 
Where 1{ }n ne 
 an orthonormal sequences of centered Gaussian variables. In addition, in this study we present
two special casesof a KL representationof a set indexed Brownian motion, when:
(a) [0,1]d
T  and ([0,1] ) {[0, ]: [0,1] }d d
x x  A A
(b) [0,1]d
T  and ( )LsA = A
In the first case, the KL representation of a set indexed Brownian motion is:
ABSTRACT: This study researches a representation of set indexed Brownian motion
{ : }AX X A A via orthonormal basis, based on reproducing kernel Hilbert space (RKHS). The
RKHS associated with the set indexed Brownian motion X is a Hilbert space of real-valued functions
on T that is naturally isometric to
2
( )L A . The isometry between these Hilbert spaces leads to useful
spectral representations of the set indexed Brownian motion, notably the Karhunen-Loève (KL)
representation: [ ]A n A nX e E X e  where { }ne is an orthonormal sequence of centered Gaussian
variables.
In addition, we present two special cases of a representation of a set indexed Brownian motion, when
([0,1] )d
A A and ( )LsA = A .
Keywords: Brownian motion, orthonormal basis, Hilbert Space, Karhunen-Loève.

Reproducing Kernel Hilbert Spaceof A Set Indexed Brownian Motion
1 2
2 2 1 1
[0, ] [0, ] ... [0, ] 2(2 1)
1 1
( ) sin( (2 1) )d
d
d
d
A x x x n in
n i
X X e n x 

   
 
   
In the second case:
1 2
2 2 2
2 2 1 1
( ) 2(2 1) 2([0, ] [0, ] ... [0, ])
1 1
lim( ) sin( (2 1) )i
kk k d mdm m m m m m m
d
kd
A g A nng m
n i
X X X e n 

     
 
     .
II. PRELIMINARIES
As in earlier set-indexed works (see [IvMe]), processes and filtrations will be indexed by a class A
whose elements are compact subsets of a fixed  -compact metric space T . In addition, we assume A satisfies
several natural conditions. We use the definition and notation from [IvMe] and all this section comefrom there.
Definition 1.Let ( , )T  be a non-void sigma-compact connected topologicalspace. A nonempty class A of
compact, connected subsets of T is called an indexed collection if it satisfies the following:
1.  A . In addition, there is an increasing sequence ( )nB of sets in A s.t. 1n nT B
 
 .
2. A is closed under arbitrary intersections and if ,A BA are nonempty, then A B is nonempty. If
( )iA is an increasing sequence in A and if there exists n such that i nA B for everyi , then i iA A .
3. ( ) A = B where B is the collection of Borel sets ofT .
4. Separability from above: There exist an increasing sequence of finite sub-classes 1{ ,..., }n
n n
n kA A A A
closed under intersection with ( )n n,B A u ( ( )nA u is the class of union of sets in nA ), and a sequence
of functions : ( )n ng T A A u such that:
(i) ng preserves arbitrary intersections and finite unions.
(ii) For each , ( )nA A g A A 
and ( )n nA g A  , ( ) ( )n mg A g A if n m .
(iii) ( ) 'ng A A A if , 'A A A and ( ) 'n ng A A A if A Aand ' nA A .
(iv) ( )ng    for all n .
Examples.
a. The classical example is
d
T   and ( ) {[0, ]: }d d
x x   A = A (or [0,1]d
T  and
([0,1] )d
A = A ). This example can be extended to
d
T   and ( ) {[0, ]: }d d
x x  A = A ,
which will give rise to a sort of 2d
-sides process.
b. The example (a) may be generalized as follows. Let
d
T   or
d
T   or [0,1]d
T  and take A to be
the class of compact lower sets, i.e. the class of compact subsets A of T satisfying t A implies
[0, ]t A. (We denote the class of compact lower sets by ( )LsA ).
c. Additional examples have been given when T is a “continuous” rooted tree (see [Sl]) and T a subspace of
the Skorokhod space, [0,1]D (see [IvMe]).
We will require other classes of sets generated by A . The first is ( )A u , which is the class of finite
unions of sets in A . We note that ( )A u is itself a lattice with the partial order induced by set inclusion. Let C
consists of all the subsets of T of the form
, , ( )C A B A B  A A u .
Any A -indexed function which has a (finitely) additive extension to C will be called additive (and is
easily seen to be additive on ( )C u as well). For stochastic processes, we do not necessarily require that each
sample path be additive, but additivity will be imposed in an almost suresense:

A set-indexed stochastic process { : }AX X A A is additive if ithas an (almost sure) additive extension to
C : 0X  and if 1 2, ,C C C C with 1 2C C C  and 1 2C C   then almost surely 1 2C C CX X X 
. In particular, if C C and 1 n
i iC A A  , 1, ,..., nA A A A then almost surely
11
... ( 1) n
i i j i i
n n
C A A A A A A A Ai i j
X X X X X
 
         
.
We shall always assume that our stochastic processes are additive. We note that a process with an (almost sure)
additive extension to C also has an (almost sure) additive extension to ( )C u .
Definition 2. A positive measure  on ( , )T B is called strictly monotone on A if: ' 0  and A B 
for all A B , ,A BA . The collection of these measures is denoted by ( )M A . ( ,' A A A   A , note
that '   )
The classical examples for definition is Lebesgue measure or Radon measure when
d
T   and
( ) {[0, ]: }d d
x x   A = A
Definition 3. Let ( )M  A . We say that the A -indexed process X is a Brownian motion with variance 
if X can be extended to a finitely additive process on ( )C u and if for disjoint sets 1,..., nC C C,
1
,..., nC CX X are independent mean-zero Gaussian random variables with variances 1
,..., nC C  , respectively.
(For any ( )M  A , there exists a set-indexed Brownian motion with variance [IvMe]).
III. RKHS OF A SET INDEXED BROWNIAN MOTION
2
( )L A is a separable space,and thusit must have a countable orthonormal basis. Then every set
indexed stochastic process { : }AX X A A has the representation
1
,A i i i
i
X a a


  where i are
random variables. Our purpose is to find the representation forthe set indexed Brownian motion via orthonormal
basis.
We restart with the RKHS of a set indexed Brownian motion. A RKHSis a Hilbert Space of functions. It can be
thought of as a space containing smoother function than the general Hilbert space.
Let A A and { : }AX X A A be a set indexed Brownian motion. We define the function
( , ) :R A   A by ( , ) ( , ) [ ]A AR A Cov X X E X X    .
Now, we define the set
1
{ : : ( ) ( , ), , , 1}
n
i i i i
i
f f a R A A a n

         A A . Define an inner
product on  by:
1 1
, ( , )
n m
i j i j
i j
f g a b R A B
 
    when
1 1
( ) ( , ), ( ) ( , )
n m
i i j i
i j
f a R A g b R B
 
       (1)
The fact that R is non-negative definite implies 2
1
, ( , ) 0
n
i i i
i
f f a R A A

    for all f .From the inner
product, we get the following property:
1 1
( ) ( , ) ( , ), ( , ) , ( , )
n n
i i i i
i i
f B a R A B a R A R B f R B
 
           (2)
Then for all f , B  A,
2 2
| ( ) | | , ( , ) | , ( , ), ( , )f B f R B f f R B R B          . The inequality
being merely the Schwartz inequality for semi-inner products, which holds as long as , 0f f   . Thus, if
, 0f f   then (2) implies that 0f  for all B  A. Consequently, (1) defines a proper inner product on
 , and so we thus obtain a norm:

,f f f  
For 1{ }n nf 
  we have:
2 22 2
| ( ) ( ) | | , ( , ) | ( ) ( ) ( , )n m n m n mf B f B f f R B f B f B R B         
2
( ) ( ) ( , )n mf B f B R B B 
Thusitfollows that if 1{ }n nf 
 is Cauchy in  then it is a pointwise Cauchy. The closure of  denoted
by H . SinceT is separable and R continuous then H is also separable.Since H is a separable Hilbert space, it
must have a countable orthonormal basis.
Define ( )H X , the so-called “linear part” of the 2
( )L A space of the set indexed stochastic process X , as the
closure in 2
( )L A of
1
, , , 1i
n
i A i i
i
a X A a n

 
   
 
 A ,
thinned out by identifying all elements indistinguishable in 2
( )L A (In other word, elements ,x y for which
2
[( ) ] 0E x y  ). This contains all distinguishable random variables, with finite variance, obtainable as linear
combination of values of the process. There is a linear, one-one mapping
2
: ( )L T A defined by:
1 1
( ) ( , ) i
n n
i i i A
i i
f a R A a X
 
 
   
 
 Τ T
Note that T is clearly norm preserving and so extends to all H with range equal to all of ( )H X . (In
other words, there is exists Tr such that | Tr T , ( )Dom HTr and Im( ) ( )H XTr ).
Since H is separable, we now know that ( )H X isalso separable. We can use this to build an orthonormal
basis for ( )H X . If 1{ }n n 
 is an orthonormal basis for H then 1{ }n ne 
 an orthonormal basis for ( )H X when
( )n ne  T . X is a set indexed Brownian motion therefore [ ] 0nE e  for all n and
1
[ ]A i A i
i
X e E X e


  almost surely,
where the series converges in 2
( )L A . Easy to see that 1{ }n ne 
 an orthonormal sequences of centered Gaussian
variables. Since T was an isometry, it follows that
[ ] ( , ), ( )A i i iE X e R A A      for all i
Then
1
( )A i i
i
X e A


  almost surely.
RKHS and the KL representation of a set indexed Brownian motion on ([0,1] )d
A and on ( )LsA = A
Now, we willpresent two special cases of a RKHS and the KL representation of a set indexed Brownian motion,
when:
i. [0,1]d
T  and ([0,1] ) {[0, ]: [0,1] }d d
x x A (see Examples after Definition 1).
ii. [0,1]d
T  and ( )LsA = A (see Examples afterDefinition 1).
First case: [0,1]d
T  and ([0,1] ) {[0, ]: [0,1] }d d
x x A = A
Let { : }AX X A A be a set indexed Brownian motion when [0,1]d
T  and
([0,1] ) {[0, ]: [0,1] }d d
x x A = A .

Let 1 2 1 2[0, ] [0, ] ... [0, ] , [0, ] [0, ] ... [0, ]d dA a a a B b b b         A A then
1 1 2 2( , ) [ ] min{ , }min{ , } min{ , }A B d dR A B E X X a b a b a b  
Values in set  are:
1 1 2 2
1 1
( ) ( , ) min{ ( ), }min{ ( ), } min{ ( ), }
n n
i i i d d
i i
f A a R A A a i a a i a a i a
 
   
1 1 2 2
1 1
( ) ( , ) min{ ( ), }min{ ( ), } min{ ( ), }
n n
j j j d d
j i
g A R B A b j a b j a b j a 
 
   
When
1 2[0, ( )] [0, ( )] ... [0, ( )]i dA a i a i a i    A and 1 2[0, ( )] [0, ( )] ... [0, ( )]j dB b j b j b j    A
The inner product on  is:
1 1
, ( , )
n m
i j i j
i j
f g R A B 
 
   
1 1 2 2
1 1
min{ ( ), ( )}min{ ( ), ( )} min{ ( ), ( )}
n m
i j d d
i j
a i b j a i b j a i b j 
 
 
but
1
[0, ( )] [0, ( )]
0
min{ ( ), ( )} ( ) ( )k kk k a i b ja i b j I x I x dx  for 1,2,...,k d
When DI is the indicator function of DA (
1 ,
( )
0 ,
D
x D
I x
x D

 

).
Therefore, we can rewrite the above as follows:
1 1
1 1
[0, ( )] 1 [0, ( )] 1 1 [0, ( )] [0, ( )]
1 1 0 0
, ( ) ( ) ( ) ( )d d
n m
i j a i b j a i d b j d d
i j
f g I x I x dx I x I x dx 
 
      
1 1[0, ( )] 1 [0, ( )] 1 [0, ( )] [0, ( )] 1
1 1 [0,1]
( ) ( ) ( ) ( )d d
d
n m
i j a i b j a i d b j d d
i j
I x I x I x I x dx dx 
 
    
1 1[0, ( )] 1 [0, ( )] [0, ( )] 1 [0, ( )] 1
1 1[0,1]
[ ( ) ( )] [ ( ) ( )]d d
d
n m
i a i a i d j b j b j d d
i j
I x I x I x I x dx dx 
 
     
1 1 1
[0,1] [0,1]
( ,..., ) ( ,..., )
d d
I I I I
n n df x x g x x dx dx f g dS   
When
11 2 [0, ( )] 1 [0, ( )]
1
( , ,..., ) [ ( ) ( )d
n
I
d i a i a i d
i
f x x x I x I x

 
11 2 [0, ( )] 1 [0, ( )]
1
( , ,..., ) [ ( ) ( )d
m
I
d j b j b j d
j
g x x x I x I x

 
Finally, ,f g  
[0,1]d
I I
f g dS .

We define 2
[0,1]
{ : : ( ) , ( ) }
d
I I I
f f f dS f dS

       A then ( , ) I
R A   and
11 [0, ] 1 [0, ] 1
[0,1]
( ) , ( , ) ( ,..., ) ( ) ( )d
d
I
d a a d df A f R A f x x I x I x dx dx       then
I
  . The mapping
2
: ( )L T A is
1
( ) i
n
I I
i A
i
f f g dS X

 
  
 
Τ T and from that we get in the same way
1
[ ]A i A i
i
X e E X e


  almost surely.
Karhunen-Loève expansion of a set indexed Brownian motion:
Let 1 2, ,..., d   and 1 2, ,..., d   be a eigenvalues and normalized eigenfunctions of operator
2 2
: ( ) ( )L L A A
1 2( ( )) ( ( , ,..., ))dA x x x    
[0,1]
( , ) ( )
d
R A S S dS 
( , )
1 1 2 2 1 2 1 2
[0,1]
min{ , }min{ , } min{ , } ( , ,..., )
d
R A S
d d d ds x s x s x s s s ds ds ds  

When 1 2[0, ] [0, ] ... [0, ]dA x x x    A , 1 2[0, ] [0, ] ... [0, ] .dS s s s    A
That is i and i solve the integral equation
1 2 1 1 2 2 1 2 1 2
[0,1]
( , ,..., ) min{ , }min{ , } min{ , } ( , ,..., )
d
d d d d dx x x s x s x s x s s s ds ds ds   
And
1 2 1 2 1 2
[0,1]
1 ,
( , ,..., ) ( , ,..., )
0 ,d
i d j d d
i j
s s s s s s ds ds ds
i j
 

  


Suppose that, there exist a 1{ }d
i i  such that 1 2 1 1 2 2( , ,..., ) ( ) ( ) ( )d d dx x x x x x     for all
1 2( , ,..., ) [0,1]d
dx x x  then
1 1
1 1 1 1 1 1 1
0 0
( ) ( ) ( )min{ , } ( )min{ , }d d d d d d dx x s s x ds s s x ds       (3)
Denote
d
  . If for all i ,
1
0
( ) ( )min{ , }i i i i i i ix s s x ds   then we get the required on (3).It is clearthat,
1 1
0 0
( ) ( )min{ , } ( )
i
i
x
i i i i i i i i i i i i i
x
x s s x ds s ds x s ds       for all i . Differentiating both sides with
respect to ix generates:
1
'( ) ( )
i
i i i i i
x
x s ds   , ''( ) ( )i i i ix x   for all i , togetherwith boundary
condition (0) 0i  . The solutions of this pair of differential equations are given by:
1
, 2( ) 2sin( (2 1) )i n i ix n x   ,  
2
2
(2 1)n n  
Then
1
1 2 1, 1 2, 2 , 2
1
( , ,..., ) ( ) ( ) ( ) 2 sin( (2 1) )
d
n d n n d n d i
i
x x x x x x n x    

    ,  
2
2
(2 1)
d
n n  
The Karhunel-Loève expansion of X is obtained by setting i i i  , we get:

1
( )A n n n
n
X e A 


 
In other words, 1 2
2 2 1 1
[0, ] [0, ] ... [0, ] 2(2 1)
1 1
( ) sin( (2 1) )d
d
d
d
x x x n in
n i
X e n x 

   
 
   (4)
Second case: [0,1]d
T  and ( )LsA = A
Let 1 2( , ,..., ) [0,1]d
dx x x x  , wedenote 1 2[0, ] [0, ] [0, ] ... [0, ]dx x x x    . There existsan increasing
sequence of finite sub-classes 1([0,1] ) { ,..., }m
m d m m
kA A A Aclosed under intersection:
2
([0,1] ) {[0, ]: , 0 2 , 1,2,..., }i
m
km d m
i ix x k i d     A .From the Definition 1 we derive that:
1 1 2
2 2 2 2 2
([0, ]) [0,( ,..., )] [0, ] [0, ] ... [0, ]d d
m m m m m
k kk k k
mg x      , ( )mg    and for all ( )A LsA ,
( )m mA g A  .
Let ( )A LsA then from (4),
1 21 2
2 2 2
2 2 1 1
([0, ] [0, ] ... [0, ]) 2(2 1) 2[0, ] [0, ] ... [0, ]
1 1
( ) sin( (2 1) )i
kk k d mdm d
m m m
d
kd
g x x x nn
n i
X X e n 

     
 
   
The process X is continuous and then from Definition 1 we derive:
1 2
2 2 1 1
([0, ] [0, ] ... [0, ]) 2(2 1) 2
1 1
lim lim( ) sin( (2 1) )i
d m
m d
d
kd
g x x x nnm m
n i
X e n 

    
 
  
1 2
2 2 2
2 2 1 1
( ) 2(2 1) 2([0, ] [0, ] ... [0, ])
1 1
lim( ) sin( (2 1) )i
kk k d mdm m m m m m m
d
kd
A g A nng m
n i
X X X e n 

     
 
    
REFERENCES
[1]. [Ad] Adler R. J., "An introduction to continuity, extreme a and relatedtopics for general Gaussian processes", IMS
Lect. Notes–Monograph series, Institute of Mathematics Statistics, Hayward, California, vol. 12, 1990.
[2]. [Be] Berlinet A. and Thomas-Agnan C., Reproducing Kernel Hilbert Spaces in Probability and Statistics, Kluwer
Academic Publishers, Boston, MA, 2004.
[3]. [BoSa] Borodin, A.B., Salminen, P., Handbook of Brownian motion – Facts and Formulae. Probability and Its
Applications. Birkhäuser Verlag (1996).
[4]. [CaWa] Cairoli, R., Walsh, J.B., Stochastic integrals in the plane. Acta Math. 134, 111–183 (1975).
[5]. [Cu] Cui M. G. and Geng F. Z., Solving singular two-point boundary value problem in reproducing kernel space, J.
Comput. Appl. Math. 205 (2007), no. 1, 6–15.
[6]. [Dal] Dalang R. C., Level Sets and Excursions of Brownian Sheet, in Capasso V., Ivano B.G., Dalang R.C.,
Merzbach E., Dozzi M., Mountford T.S., Topics in Spatial Stochastic Processes, Lecture Notes in Mathematics,
1802, Springer, 167-208, 2001.
[7]. [Dan] Daniel A., Reproducing Kernel Spaces and Applications, Operator Theory, Advances and Application,
Birkhauser Verlag, Basel, 2003.
[8]. [Du] Durrett, R., Brownian motion and Martingales in Analysis. The Wadsworth Mathematics Series. Wadsworth,
Belmont, California (1971).
[9]. [Fr] Freedman, D., Brownian motion and Diffusion. Springer, New York, Heidelberg, Berlin (1971).
[10]. [Ge] Geng F. Z. and Cui M. G., Solving singular nonlinear second-order periodic boundary value problems in the
reproducing kernel space, Appl. Math. Comput. 192 (2007), no. 2, 389–398. [5] F. Z. Geng and M. Cui, Solvi
[11]. [Ha] Haykin S., Adaptive Filter Processing. Prentice-Hall, 4nd edition, 2002.
[12]. [He] Herbin, E., Merzbach, E., A characterization of the set-indexed Brownian motion by increasing paths. C. R.
Acad. Sci. Paris, Sec. 1 343, 767–772 (2006).
[13]. [IvMe] Ivanoff, G., Merzbach, E., Set-Indexed Martingales. Monographs on Statisticsand Applied Probability,
Chapman and Hall/CRC (1999).
[14]. [Kh] Khoshnevisan, D., Multiparameter Processes: An Introduction to Random Fields.Springer (2002).
[15]. [MeYo] Merzbach E. and Yosef A., Set-indexed Brownian motion on increasing paths, Journal of Theoretical
Probability, (2008), vol. 22, pages 883-890.
[16]. [MeNu] Merzbach, E., Nualart, D., Different kinds of two parameter martingales. Isr. J. Math. 52(3), 193–207
(1985).

[17]. [Par] Parzen, E., Probability density functionals and reproducing kernel Hilbert spaces. Time Series Analysis. 155–
169 M. Rosenblatt, Ed. New York: Wiley 1963.
[18]. [Par67] Parzen. E, Time Series Analysis Papers. Holden-Day, San Francisco, CA, 1967.
[19]. [ReYo] Revuz, D., Yor, M., Continuous Martingales and Brownian Motion. Springer, New York,Heidelberg, Berlin
(1991).
[20]. [Sc] Scholkopf B., C. J. C. Burges, and Smola A., editors. Advances in Kernel Methods: Support Vector Learning.
MIT Press, 1999.
[21]. [ScSm] Scholkopf B. and Smola A., Learning with Kernels. MIT Press, Cambridge, MA, 2002.
[22]. [Sl] Slonowsky, D., 2001. Strong martingales: their decompositions and quadratic variation, J. Theor. Probab. 14,
609-638.
[23]. [St] Steinwart I., Hush D. and Scovel C., An explicit description of the reproducing kernel Hilbert spaces of
Gaussian RBF kernels IEEE Trans. Inform. Theory, 52 (10) (2006), pp. 4635–4643
[24]. [Yo09] Yosef A., Set-indexed strong martingale and path independent variation, Journal of Statistics and Probability
Letters, (2009), vol. 79, issue 8, pages 1083-1088.
[25]. [Yo15] Yosef A., Some classical-new set-indexed Brownian motion, Advances and Applications in Statistics (Pushpa
Publishing House) (2015), vol. 44, number 1, pages 57-76.
[26]. [Va] Vapnik V. N., Statistical Learning Theory. Wiley, New York, 1998.
[27]. [Za] Zakai, M., Some classes of two-parameter martingales. Ann. Probab. 9, 255–265 (1981).

Reproducing Kernel Hilbert Space of A Set Indexed Brownian Motion

More Related Content

What's hot (20)

Similar to Reproducing Kernel Hilbert Space of A Set Indexed Brownian Motion (20)

More from IJMERJOURNAL (20)

Recently uploaded (20)

Reproducing Kernel Hilbert Space of A Set Indexed Brownian Motion