Vous êtes sur la page 1sur 10

Correlator Solution

Arslan Sikandar
December 26, 2015

Collocation Method

Consider a linear integral equation bounded as [-1,1], we wish to calculate the solution u(x)
that satisfies:
Z 1
u(x) = x +
xtu(t)dt
(1)
1

here is a degenerate kernel K(x, t) = xt. This equation has exact solution u(x) = 3x. We
will solve it numerically by collocation method. Consider an ansatz,
u(x) =

3
X

cn n (x) = c1 + c2 x + c3 x2

(2)

n=1

where (1, x, x2 ) are linearly independent functions. We expanded just upto 3 terms. Substitute this into (1).
Z 1
2
c1 + c2 x + c3 x = x + x
(c1 t + c2 t2 + c3 t3 )dt + 
1

c1 t2 c2 t3 c3 t4
=x+x
+
+
2
3
4
3
= x(1 + c2 )
2


1
1

Comparing coefficients on both sides, we have c1 = 0, c2 = 3 and c3 = 0. Putting these


coefficients to our ansatz eq(2), we get u(x) = 3x which coincides with the exact solution.
This was a pretty easy integral equation. Sometimes, the integrand is so nasty that we cant
do an analytical integration and have to rely upon numerical methods like trapezoidal rule.
In integral equations, Nystroms method is usually used that i will introduce later.

1.1

Sinc Interpolation

Sinc methods are based on the Cardinal function;


S(k, h)(x) =

sin[(x jh)/h]
,
(x jh)/h
1

j = 0, 1, 2

(3)

here h > 0 is the mesh size (a little bit on it later). j is a positive integer. So for example
j = 1, 2, 3 we get three linearly independent functions which are translated by mesh size
h. Later if we have interpolating points xk = kh in the integration space[a,b](limits of an
integral), we get.
(0)
S(j, h)(kh) = jk
(4)
Given a function u(x) bounded for all (, ), the series;
C(u, h)(x) =

u(jh)Sinc(

j=

x jh
)
h

(5)

is called the Whittaker cardinal expansion of u if this series converges. For finite N, the
series may be defined as;
C(u, h)(x) =

N
X

u(jh)Sinc(

j=N

1.2

x jh
)
h

(6)

Double Exponential Transformation-Linear

In previous sub-section, we have defined an orthogonal basis in the form of translated Sinc
functions just like we did in solving the first example. But we sometimes face highly nonlinear equations with singularities. Such simple interpolation series doesnt guarantee us the
desired solution. Also, sometimes we are concerned with the rate of convergence of our approximate solution. There exist some conformal mappings that in general treat singularities
and in particular end-point singularities. These are exponential transformations single(SE)
or double(DE). The difference is there convergence rate to exact solution. For example, SE
converges to exact solution as  e N while DE converges as  eN . Consider again a
general non-homogeneous linear integral equation;
Z
u(x) = f (x) + k(x, t)u(t)dt, x, t = [a, b]
(7)

We have the basis in the form of Sinc functions. We define a conformal mapping such that
(a) = and (b) = . The j th function will then become;


(x) jh
(8)
Sj (x) = S(j, h) (x) = Sinc
h
Now the point is to choose the conformal mappings whether single or double and use the
above basis in eq(7). We will first do the general treatment with (x) and later choose the
mappings when doing the examples. Defining a new ansatz with conformal functions;


N
X
(x) jh
u(x) =
uj Sinc
(9)
h
j=N
Substitute in (7);
N
X
j=N


uj Sinc

(x) jh
h


= f (x) + h

N
X
j=N

K(x, tj )0 (tj )uj

(10)

where we used Nystroms method to transform the integral into a sum. Here tj = (jh).
Now comes the interesting part. As we have transform the integral into sum, we need to
define collocation points xk [a, b] that the coefficients uj satisfies. We can define any
arbitrary points but if we define it as xk = (kh) and using eq(4), we get things a little bit
simplified as;
N n
o
X
(0)
kj hK(xk , tj )0 (tj ) uj = f (xk )
(11)
j=N

We can easily transform this into a matrix equation. We have to find the coefficients uj
P
(0)
which is an N 1 column vector. Considering, N
where I is an N N identity
j=N = I
PN
matrix. The kernel j=N K(xk , tj ) = K is an N N matrix and f (xk ) = F as a column
vector, we have the following matrix form;
I(0) hK = F

(12)

By solving this system we get the required coefficients and using (9) we get the required
solution.
So we kind of defined the algorithm to solve the linear non-homogeneous integral equations.
But before applying it, we have to discuss the conformal transformations.
Sugihara and Mori have worked alot on different sort of DE transformation that are efficient in evaluating integral equation that have different bounds [a, b]. The relevant DE
transformation is;

 b+a
ba
tanh
sinh(t) +
(13)
(t) =
2
2
2
Its derivative is;
ba
cosh(t)

(t) =
(14)
4 cosh2 2 sinh(t)
There is one confusing point that needs to be clear. In DE, we defined for the independent
variable x the collocation points which are exactly those of the dependent variable t integrated out. Like i said, defining the collocation points xk is arbitrary, we could have defined
any other points as long as they are in the bound.

1.3

Linear Example

Consider a linear non-homogeneous integral equation:


Z
2

3
3
(15)
u(x)
(xt) 2 u(t)dt = x x3/2
24
0

It has exact solution; u(x) = x. In the next page the exact and approximate solutions are
both plotted.

1.2

1.0

0.8

0.6

0.4

0.2

0.5

1.0

1.5

Figure 1: Exact solution of Example 1.


While using the DE transformation in mathematica for N=4, we get the plot as;

0.8

0.6

0.4

0.2

0.2

0.4

0.6

Figure 2: Approxiamte solution of Example 1.

0.8

1.0

As can be seen there is some error which can be reduced if we take N higher. But the
prominent thing is how the approximate solution is converging into exact solution even when
N is lower.

1.4

Double Exponential Transformation- Non-Linear

We can extend our analysis of using Sinc-collocation method to solve non-linear integral
equations. In linear example, we transformed our problem into a matrix problem. As it was
linear, so we just had to diagonalize the matrix. We can also transform non-linear equations
into matrix form but then we have to solve the system of non-linear simultaneous algebraic
equations. Fortunately we can, but not with ease. Consider a general non-linear equation:
Z
(16)
u(x) = K(x, t)P (u(t), t)dt, x, t [a, b]

here the function P (u(t), t) introduces non-linearity. For example, P (u(t), t) = u(t), then we
get our linear case but if it is any other function of u(t), lets say eu(t) then the corresponding
integral equation is non-linear. Using the same ansatz as in (9) and substituting in eq(16),
we get;
N
X
K(xk , tj )Pj 0 (tj ) = f (xk )
(17)
uj h
j=N

here tj = (jh). Also the collocation points xk = (kh). There are (2N + 1) unknowns uj ,
j = N, N + 1, ..., N . In order to determine these (2N + 1) unknowns, we rewrite this
system which is the non-linear system of algebraic equations in matrix form as:
UK P=F

(18)

here,
U = [uN , uN +1 , ........., uN 1 , uN ]T
K = [K(xk , tj )0 (jh)], j = N, ..., +N , which implies it is an N N matrix.
P = [PN , PN +1 , ........, PN 1 , PN ]T
F = [fN , fN +1 , ........, fN 1 , fN ]T
The point noticeable is the form of vector P. For example, if we have a non-linearity of the
form P (u(t), t) = eu(t) , then the vector P will be
P = [euN , euN +1 , ..........., euN 1 , euN ]

2.5

2.0

1.5

0.2

0.4

0.6

0.8

1.0

Figure 3: Exact solution of Example 2.

1.5

Example 2: Non-linear

Consider a non-linear integral equation defined in the interval[0,1],


Z 1
u(x) +
e(x2t) u3 (t)dt = e(1+x)

(19)

It has exact solution u(x) = ex . We will again use the conformal map defined in (13) and
the sinc-basis to end up with a matrix equation (17). This will allow us to write (2N+1)nonlinear algebraic equations. I used FindRoot in mathematica, which pretty much working for
me. Solving them simultaneously give us the required coefficients. One thing that i missed
1
earlier, its the mesh size h. We use for Sinc functions the mesh size as h = N1 log( N
) 2 After
2
finding the coefficients, we write our approximate solution as;
uN (x) = f (x) + h

N
X

K(x, tj )0 (jh)P (u(tj ), tj )

(20)

j=N

In figure 3, the exact solution of is first plotted as a reference. In figure 4, we have the
approximate solution for N = 2.

3.5

3.0

2.5

2.0

1.5

1.0

0.5

0.2

0.4

0.6

0.8

1.0

Figure 4: Approximate of Example 2 for N=2.

1.6

2-point 1PI function

We found the two point function of non-commutative O(N) vector model at large N in 2
dimensions as;
Z
cos(pk)
kdk
(21)
y(p) = f (p) +
y(k)
0
2

here f (p) = p2 + m2 , = g2N (the N here defined is the one came from the theory not the
approximation parameter). is the non-commutativity parameter. We convert this equation
in x and t variables so as to be consistent with the overall gist of this presentation. Also, t
shouldnt be confused with time. We can write it as;
Z
t
cos (t)
dt
(22)
u(x) = f (x) +
u(t)
0
here we just say the product x . The reason will be clear soon. This equation has a
reciprocal non-linearity which means there could be singularities that we should aware of.
Secondly, it is defined on the semi-infinite plane [0, ) which means we cannot take the
conformal transformation that we used earlier to solve the example 1 and 2. The integral is
slowly decaying which means we need to solve this equation by adopting a new conformal
transformation discussed in []. We have the DE transformation for cosine integral form as;
t M (t

)/
2M

(23)

here M is just a constant pretty much like the mesh size earlier. Here (t) is defined as
(t) =

t
1 exp (2t (1 et ) (et 1))
7

(24)

1 106

800 000

600 000

400 000

200 000

200

400

600

800

1000

Figure 5: Approxiamte 2-point function for N=10 and w=1.

where =

1
4

and =

.
1+M log(1+M )/4

we again follow the procedure as done before, by using the sinc basis function as in (9) and
using the transformation (23), we have
N
X
N

N
X
(x) h
) = f (x) + M h
cos [M (jh)] P [u((jh)), (jh)] 0 (jh)/ (25)
uj Sinc(
h
j=N

where P (u(t), t) = t/u(t). Defining the collocation points exactly as same as of t as xk =


(kh). This gives the following equation for the coefficients;
N
M h X
(jh)
1
cos (M (jh))

0 (jh) = f ((kh))
uj
j=N
uj
(kh)

(26)

where (kh) in the denominator came from the = x present there. We can again turn this
into a matrix equation. Defining M h = will make things a bit easier. I have approximated
the solution for N=10.
From here, we can study the solution for different values of non-commutativity parameter.
For 0 we need to solve the equation with different conformal transformation. On
the next page i draw the two figures for the different non-commutative parameters, i.e,
= 1 and = 100 and for clarity, i just drew the non-perturbative correction part; it
shows that at high momentum the propagator gets less correction to mass term for high
noncommutativity(which makes sense but we need to discuss it).

-0.2

-0.3

-0.4

-0.5

-0.6

Figure 6: Approxiamte 2-point function for N=10 and w=1.

10

-0.05

-0.10

-0.15

-0.20

Figure 7: Approxiamte 2-point function for N=10 and w=100.

10

10

Vous aimerez peut-être aussi