You are on page 1of 18

Kiran R RBF - 1D Date: Feb 10, 2020

Electric Field Equations using RBF

∆t
en+1
y = eny − P (x)hn+1/2
z (1)

P (x) is m x n matrix with m = number of test points and n = number of collocation points.
Here the input parameter r is assumed a Gaussian random variable (N [¯r , σ]). Polynomial Chaos
Expansion is applied on the random variables en+1 n
y , ey and  = 0 r . Thus each of them has the form

m
X
e= eiΨi(ξ) (2)
i=1

ei are constants and Ψi(ξ) are the basis functions which are chosen according to the Wiener-Askey
scheme. Substitution of the PCE expansion of the random variable into the electric field equations
gives (spacial indexing of the fields are omitted for convenience)

m m m n+1/2
X X ∆t Xh Ψi(ξ)
en+1
i Ψi(ξ) = n
ei Ψi(ξ) − P (x) i
(3)
i=1 i=1
0 i=1
r (ξ)

1
Kiran R RBF - 1D Date: Feb 10, 2020

Can it be written this way? OR is it

m m Pm n+1/2
X X ∆t h Ψi(ξ)
en+1
i Ψi(ξ) = eniΨi(ξ) − P (x) Pm i
i=1
(4)
i=1 i=1
0 i=1 r,i (ξ)

Applying Galerkin procedure on the above equation [3] for some Ψk (ξ) we can write

m  
∆tP X Ψ i (ξ)Ψ k (ξ)
en+1
k = enk − (5)
< Ψ2k > 0 i=1 r (ξ)
2 /2
The inner product integrals are evaluated seperately, with weight function w(x) = e−x
Z ∞
2 /2
< Ψk , Ψk >= Ψk (ξ)Ψk (ξ)e−ξ dξ f or k ∈ [1, m] (6)
−∞

∞ 2 /2
Ψi(ξ)Ψk (ξ)e−ξ
  Z
Ψi(ξ)Ψk (ξ)
= dξ f or i, k ∈ [1, m] (7)
r (ξ) −∞ ¯r + σξ
where ¯r and σ are the mean and variance of the input random variable.

2
Kiran R RBF - 1D Date: Feb 10, 2020

Magnetic Field Equations using RBF

∆t
hn+1/2
z = hn−1/2
z −P (x)eny (8)

Substitution of the PCE expansion of the random variable into the magetic field equations gives (spacial
indexing of the fields are omitted for convenience)

m m m
X n+1/2
X n−1/2 ∆t X
hi Ψi(ξ) = hi Ψi(ξ) − P (x) eniΨi(ξ) (9)
i=1 i=1
µ i=1

Applying Gallerkin procedure on this, for some k ∈ [1, m]


n+1/2 n−1/2 ∆t
hk = hk − P (x)enk (10)
µ
Difficulty in finding these integrals
· Equation [5] and [10] are the design equations which is to be run ’f or i = 1 : m’
· There is a singularity at ξ = −¯r /σ
· Infinite limit does not allow to evaluate them accurately using numerical methods

3
Kiran R RBF - 1D Date: Feb 10, 2020

Conclusion

· Adopting the Gaussian distribution (whose domain Ω = [−∞, ∞] is unrealistic and introduces a
sigularity in the integral equation of the Gallerkin form. This may be leading to errors in the
evaulation of integrals.
· One sided distributions like Gamma or bounded distributions like Uniform, Beta may be adopted
to represent the uncertainity.

4
Kiran R RBF - 1D Date: Feb 20, 2020

Work Plan

Figure 1 is the work plan showing the possible ways to proceed, and the green shades are attempted.

5
Kiran R RBF - 1D Date: Feb 20, 2020

Figure 1: Work plan.

6
Kiran R RBF - 1D Date: Feb 20, 2020

Cauchy’s Integral Theorem

Let f (z) be an analytic function in a simply connected domain D. If z = z0 be a singular point such
that
Z
f (z)
dz = 2πif (z0) (11)
z − z0
We can transform the integral of interest (which is a line integral from −∞ to +∞) into a complex
domain by substituting x = z where z = rejθ . This can be evaluated on a simply connected domain
which is an infinite half circle enclosing the upper half of the complex plane.
Figure 2 shows the exponential Gaussian function and the denominator contributing to the singularity.
Figure 3 shows the region of integration.

2 −b/a− 2 2 ∞ 2 2
e−x /2 e−x /2 e−x /2 e−x /2 e−x /2
I Z Z Z Z
dx = dx + dx + dx + dx (12)
C ax + b −∞ ax + b C1 ax + b −b/a+ ax + b C2 ax + b

Last factor goes to zero as r → ∞

7
Kiran R RBF - 1D Date: Feb 20, 2020

Figure 2: Singular point of the function.

8
Kiran R RBF - 1D Date: Feb 20, 2020

Figure 3: Region of integration in complex plane.

9
Kiran R RBF - 1D Date: Feb 20, 2020

2
e−x /2
I
2πi −(−b/a)2/2
∴ dx = e (13)
C ax + b a
Ambiguity in these results

· If we take the curve C1 such that the singularity is excluded, the result is zero. But both ultimately
indicating same integral when r → 0 on C1 !!
· This means that when the contour passes through the singular point, result may be different.
R
· Check in wolframalpha → (1/(x − 1)) . According to Cauchy what is this ?

10
Kiran R RBF - 1D Date: Feb 20, 2020

Cauchy’s Principal Value Theorem

This is another form to evaluate this integral approximately.

∞ 2 p− 2 ∞ 2
e−x /2 e−x /2 e−x /2
Z Z Z
dx = lim dx + dx (14)
−∞ ax + b p→−b/a −∞ ax + b p+ ax + b

In MATLAB this is estimated with values ofp − p− ranging from 0.01 to 0.000001 and the estimated
value is found to be 0.627049465994397. This agrees with wolframalpha results (They also estimate
using the same method).

MATLAB function integral() gives 0.627049465994421

MATLAB function evaluation estimation near the singular point.


feval(fun, -40.0000000000000035527136788) ans = NaN
feval(fun, -40.0000000000000035527136789) ans = 0

11
Kiran R RBF - 1D Date: Feb 20, 2020

Finally the limitting values obtained by the analytical integration by Cauchy’s Principal Value theorem
are used in the look up table for the PCE algorithm to run. The results obtained in comparison with
the Monte Carlo results are as in Figure 4 below.

12
Kiran R RBF - 1D Date: Feb 20, 2020

13

Figure 4: Comparison between Monte Carlo and S-RBF.


Kiran R RBF - 1D Date: Feb 28, 2020

Increasing the terms in PCE


Recursive relatio for the Hermite polynomials is
Pn+2 − xPn+1(x) + (n − 1)Pn(x) = 0, n = 1, 2, ... (15)
Using htis we obtain the polynomials as

P1(x) = 1
P2(x) = x
P3(x) = x2 − 1
P4(x) = x3 − 3x
P5(x) = x4 − 6x2 + 3
P6(x) = x5 − 10x3 + 15x
Increasing the number of coefficient terms beyond 4 does not make considerable changes in the distri-
bution.

Next, a change in the initial condition of the second moment calculation is intoduced and the simulation
results are comapred. But this could not be justified. Actually if the source is deterministic, it will
14
Kiran R RBF - 1D Date: Feb 28, 2020

have only the mean factor in its polynomial chaose expansion. Following is the algorithm for S-RBF.
(I have included the second moment factor in the initial condition for the sake of completeness of the
discussion.)

15
Kiran R RBF - 1D Date: Feb 28, 2020

Initialization;
[P ] ← Get the matrix;
[P C f actor] ← Get the marix from look-up table ;
while n <= n step do
while k <= pc count do
Hzk = Hzk − ∆t µ [P ]Eyk ;
end
while k <= pc count do
Ppc count
Ck = j=1 Hzj P C f actor(j, k);
∆t
Eyk = Eyk − 0P C f actor sq(k) [P ]Ck
end
t = n∆t;
source = e−(t−α)/β ;
Ey(source loc, 1) = source;
Ey(source loc, 2) = (σ 2/µ)2source;
Apply PEC boundary conditions on all moment equations;
end
Post Processing: Generate the distribution using the moment coefficients;
Algorithm 1: Algorithm for S-RBF 16
Kiran R RBF - 1D Date: Feb 28, 2020

Figure 5: Comparison between Monte Carlo and S-RBF (Change in σ = 0.1, 0.2, 0.3).

Figure 6: Comparison between Monte Carlo and S-RBF (Change in µ = 4, 8, 16).

17
Kiran R RBF - 1D Date: Feb 28, 2020

Figure 7: Comparison between Monte Carlo and S-RBF (Change in observation location x = 1.83m, 1.85m, 1.87m).

18

You might also like