You are on page 1of 10

Numerical Results

four
four
In this section we present 4 examples which act as numerical verification for the e↵ective-
ness of the stabilisation method using biorthogonal projection. Examples 1,2 and 3 are
taken from [1] and example 4 is taken from [2]. For each example we present the numeri-
cal convergence rates achieved when using the biorthogonal stabilisation on the standard
linear triangular element, as well as the standard bi-linear and bi-quadratic quadrilateral
elements shown in Figure 1.
From this point we will refer to the standard linear triangular element as the triangular
order 1 element, the standard bi-linear quadrilateral element as the quadrilateral order 1
element and the standard bi-quadratic quadrilateral element as the quadrilateral order 2
element.
The numerical results demonstrate the e↵ectiveness of the biorthogonal stabilisation
method when using weak (Nitsche) boundary conditions and also when using strong
(Dirichlet) boundary conditions.

(a) Triangular Order 1 (b) Quadrilateral Order 1 (c) Quadrilateral Order 2

Figure 1: The standard elements, where the red dots represent the nodal points

Example 1
fameters
We consider the PDE (1.1) with coefficients " = 10 8 , b = (2, 3), c = 1, and where
⌦ = (0, 1)2 , D = @⌦, N = ; for which the exact solution is

u (x, y) = x3 (1 x) y 1 y4 .

This problem has a smooth solution with no discontinuities on the domain, demon-
strating how well the stabilisation using the biorthogonal method works on simple prob-
lems.

(a) Initial Triangular mesh (b) Initial Quadrilateral mesh

Figure 2: Initial mesh for example 1

For the numerical testing, the initial mesh used for the triangular and quadrilateral
elements are shown in Figure 2. Tables 1 and 2 display the L2 , H 1 and LP norm errors
over each iteration for the stabilised triangular and quadrilateral first order elements with

1
strong boundary condition and weak boundary condition respectively. Both tables show
that the optimal rate of convergence for all three norm measures is achieved, and in the
cases of the L2 and H 1 norms it is exceeded.
Table 1: Errors over whole domain ⌦ for solutions generated using Triangular and Quadri-
lateral Order 1 elements for Example 1 using biorthogonal stabilisation with strong
boundary conditions

Triangular Element Quadrilateral Element


h L2 error 1
H error LP error h L2 error H 1 error LP error
0.500000 0.022646 0.135897 0.193057 0.707107 0.023343 0.139208 0.104650
0.250000 0.016596 0.117779 0.109686 0.353553 0.020405 0.130760 0.131703
0.125000 0.006566 0.065190 0.037727 0.176777 0.010546 0.087555 0.061550
0.062500 0.001491 0.030854 0.011083 0.088388 0.003065 0.041988 0.019217
0.031250 0.000310 0.012755 0.002899 0.044194 0.000668 0.017917 0.005398
0.015625 0.000058 0.004958 0.000705 0.022097 0.000131 0.007351 0.001429
0.007812 0.000010 0.001986 0.000170 0.011049 0.000025 0.003049 0.000366
Order of cng 2.48 1.32 2.06 - 2.41 1.27 1.97

Table 2: Errors over whole domain ⌦ for solutions generated using Triangular and Quadri-
lateral Order 1 elements for Example 1 using biorthogonal stabilisation with weak bound-
ary conditions

Triangular Element Quadrilateral Element


h L2 error 1
H error LP error h L2 error H 1 error LP error
0.500000 0.021017 0.124419 0.187807 0.707107 0.032455 0.149298 0.094746
0.250000 0.014983 0.104198 0.104016 0.353553 0.021507 0.123931 0.122183
0.125000 0.005676 0.058529 0.037068 0.176777 0.009862 0.083924 0.059800
0.062500 0.001344 0.027227 0.010914 0.088388 0.002954 0.040594 0.019007
0.031250 0.000289 0.011652 0.002887 0.044194 0.000650 0.017316 0.005293
0.015625 0.000054 0.004657 0.000702 0.022097 0.000127 0.007110 0.001388
0.007812 0.000010 0.001918 0.000168 0.011049 0.000024 0.002971 0.000354
Order of cng 2.49 1.28 2.06 - 2.43 1.26 1.97

The norm error of the solutions generated by solving the weak boundary formulation
achieve the expected rates of convergence. These three graphs in Figure 3 represent the
solutions obtained using the triangular and quadrilateral first order elements, as well as
the second order quadrilateral element.

Example 2
The second example considers the PDE (1.1) with coefficients " = 10 8 , b = (1, 1), c = 1,
and where ⌦ = (0, 1)2 , D = @⌦, N = ; for which the exact solution is
! !
exp x " 1 1 exp y " 1 1
u (x, y) = + (x 1) + (y 1) .
exp "1 1 exp "1 1

The solution over the entire domain for this problem shows the error term calculated
either converges suboptimally, as shown in Figure 5 a) b) and c) for the L2 and H 1 norms,

2
Single row

L 2 error L 2 error
10-1
H1 error 10-1 H1 error
LP error LP error
2
O(h ) O(h 2)
O(h 1) O(h 1)
-2
10
10-2
Error

Error
10-3
10-3

10-4
10-4

10-5
102 103 104 101 102 103 104
Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral

L 2 error
H1 error
-2
10 LP error
O(h 3)
O(h 2)

approximation 10-4
Error

is as 10-6

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 3: The L2 , H 1 and LP error convergence using the Biorthogonal method on the
Nitsche formulation plotted against the number of degrees of freedom

I
or diverges completely, in the way of the LP norm. This is due to the exponential layer
near the x = 1 and y = 1 boundaries, where the space over which it acts is much less than
the mesh size h. But when the solution is analysed in a subdomain that does not contain
Éhwhorthmen.si
the exponential layer, such as the subdomain [0, 3/4]2 , then the solution fits and acts as h
expected when using the biorthogonal stabilisation with weak boundary condition. This
is shown for all three element types in Figure 6.
As displayed in Figure 4b, if stabilisation is not employed in the discrete solver, then

Lenin Y
the solution reached does not resemblance the exact solution as shown in Figure 4a. When
stabilisation is employed, the solutions improve vastly as displayed in Figures 4c and 4d,
which employ stabilisation for the strong and weak boundary conditions respectively.
This result is true not only for the triangular case presented in the figure but also for the
order 1 and 2 quadrilateral elements.

Example 3 the approximate solution has spurious


oscillations
8
Here, we consider the PDE (1.1) with coefficients " = 10 , b = (2, 3), c = 2, and where
⌦ = (0, 1)2 , D = @⌦, N = ; for which the exact solution is
0 ⇣ ⇣ ⌘⌘ 1
1 1 1 2 1 2
1 tan 200 42
x 2
y 2
u (x, y) = 16x (1 x) y (1 y) @ + A.
2 ⇡

3
(a) (b)

(c) using(d) solution wins


Isomtion E
Figure 4: Triangular Order 1 discrete solutions generated for example
y
2 with mesh size
h = 0.0055. (a) Exact solution. (b) Non-stabilised formulation. (c) Stabilised formu-
lation with strong boundary condition. (d) Stabilised formulation with weak boundary
condition.

For this example there is a transition layer imposed by the solution, much like in
Example 2. However unlike Example 2, the transition layer is not located close to the
boundary. In this problem the transition layer lies on a circle in the domain with radius
0.25 and centre (0.5, 0.5). A di↵erent issue emerges than the one caused in Example 2
since here the transition layer encompasses a large proportion of the domain.
The results indicate that the transition layer causes the convergence rate to be below
that expected by the theory, as shown in Figure 7. The discrepancy will eventually be
corrected when the meshsize reaches the size of ".
The transition layer is the main source of error for the solutions generated, as shown
in Figure 8. These graphs numerically demonstrate that most of the error occurs in the
immediate nodal neighbourhood of the transition layer.

Example 4
In the final example we consider the PDE (1.1) with coefficients " = 10 8 , b = (2, 3),
c = 1, and where ⌦ = (0, 1)2 , D = @⌦, N = ; for which the exact solution is
✓ ◆ ✓ ◆ ✓ ◆
2 (x 1) 3 (y 1) 2 (x 1) + 3 (y 1)
u (x, y) = xy 2 y 2 exp x exp + exp .
" " "

4
100

10-1

10-1

10-2
Error

Error
10-2

10-3 L 2 error L 2 error


H1 error H1 error
LP error LP error
O(h ) 2 10-3 O(h 2)

10-4 O(h 1) O(h 1)

102 103 104 101 102 103


Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral


100

10-1

10-2
Error

10-3

10-4 L 2 error
H1 error
LP error
O(h 3)
10-5
O(h 2)

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 5: The L2 , H 1 and LP error convergence using the biorthogonal stabilisation


method with weak boundary condition over the whole domain plotted against the number
of degrees of freedom for E
example 2

In this problem transition layers are again lying close to the x = 1 and y = 1 boundary.
The error rates for this problem also perform poorly when taken to be over the entire
domain for each respective element type, as shown in Figure 9. However, when evaluated
over a subdomain that excludes the transition layers, the error rates perform as expected,
E
which agrees with the observation made for the same phenomenon in example 2. This
is shown in Figure 10, where the expected rates of convergence are achieved or exceeded
within the first iterations of the method for the order 1 elements and requires additional
iterations for the order 2 element.

Bibliography
References
[1] Asha K. Dond and Thirupathi Gudi. Patch-wise local projection stabilized finite
element methods for convection–di↵usion–reaction problems. Numerical Methods for
Partial Di↵erential Equations, 35(2):638–663, 2019.
[2] S. Ganesan and L. Tobiska. Stabilization by local projection for convection–di↵usion
and incompressible flow problems. Journal of Scientific Computing, 43(3):326–342,
2010.

5
100
L 2 error L 2 error
1
H error H1 error
10-1 LP error LP error
2 -1
O(h ) 10 O(h 2)
O(h 1) O(h 1)
10-2

10-2
Error

Error
10-3

10-3
10-4

10-5 10-4

102 103 104 101 102 103


Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral


100

10-1

10-2
Error

10-3

10-4 L 2 error
H1 error
LP error
O(h 3)
10-5
O(h 2)

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 6: The L2 , H 1 and LP error convergence using the Biorthogonal method on the
Nitsche formulation over the subdomain [0, 3/4]2 plotted against the number of degrees
of freedom for example 2
E

Table 3: Errors over subdomain (0, 3/4)2 for solutions generated using Triangular and
E
Quadrilateral Order 1 elements for example 4 using strong boundary condition

Triangular Element Quadrilateral Element


h L2 error H error 1
LP error h L2 error H 1 error LP error
0.707107 0.009594 0.085339 0.162990 0.707107 0.011666 0.081466 0.161722
0.353553 0.045212 0.359969 0.543790 0.353553 0.042260 0.344633 0.513518
0.176777 0.142821 0.687921 0.407319 0.176777 0.131462 0.552725 0.344662
0.088388 0.051366 0.416846 0.183173 0.088388 0.036759 0.351024 0.174760
0.044194 0.007062 0.106084 0.042170 0.044194 0.004750 0.068089 0.024117
0.022097 0.000459 0.016221 0.003115 0.022097 0.000239 0.009019 0.001991
Order of cng 3.94 2.71 3.76 - 4.31 2.92 3.60

6
101

100

100

10-1
Error

Error
10-1

10-2
L 2 error 10-2 L 2 error
H1 error H1 error
LP error LP error
2
O(h ) O(h 2)
10-3 O(h 1) 10-3 O(h 1)

102 103 104 101 102 103


Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral

101

100

10-1
Error

10-2

L 2 error
10-3 H1 error
LP error
O(h 3)
10-4 O(h 2)

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 7: The L2 , H 1 and LP error convergence using the Biorthogonal method on the
Nitsche formulation plotted against the number of degrees of freedom for example 3

Table 4: Errors over subdomain (0, 3/4)2 for solutions generated using Triangular and
Quadrilateral Order 1 elements for example 4 using weak boundary condition
E
Triangular Element Quadrilateral Element
h L2 error H error1
LP error h L2 error H 1 error LP error
0.707107 0.015826 0.131967 0.146696 0.707107 0.013348 0.068400 0.063526
0.353553 0.050160 0.367266 0.345191 0.353553 0.035989 0.257036 0.247741
0.176777 0.033979 0.191772 0.090052 0.176777 0.023733 0.138203 0.057120
0.088388 0.005574 0.065618 0.024101 0.088388 0.003236 0.038609 0.016040
0.044194 0.000724 0.018912 0.002590 0.044194 0.000354 0.009038 0.001645
0.022097 0.000094 0.008293 0.000530 0.022097 0.000053 0.003666 0.000345
Order of cng 2.94 1.19 2.29 - 2.74 1.30 2.25

7
(a) (b)

(c)

Figure 8: Triangular Order 1 discrete solutions generated for example 3 with mesh size
h = 0.0055. (a) Exact solution. (b) Stabilised formulation with weak boundary condition.
(c) Magnitude of the error between exact solution and the weak boundary solution.

and numerical solution with the weak

boundary condition
refer to numerical
It will be good if you
number
schemes wilt equation
8
100 100

10-1 10-1
Error

Error
10-2 10-2

10-3 L 2 error 10-3 L 2 error


H1 error H1 error
LP error LP error
2
O(h ) O(h 2)

10-4 O(h 1) 10-4 O(h 1)

1 2 3 4 1 2
10 10 10 10 10 10 103 104
Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral


100

10-2
Error

10-4

L 2 error
H1 error
LP error
O(h 3)
10-6 O(h 2)

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 9: The L2 , H 1 and LP error convergence over the entire domain ⌦ using the
Biorthogonal method on the Nitsche formulation plotted against the number of degrees
of freedom for example 4
E

9
100 100
L 2 error L 2 error
H1 error H1 error
LP error LP error
O(h 2) O(h 2)
10-1 10-1
O(h 1) O(h 1)
Error

Error
10-2
10-2

10-3
10-3

10-4
10-4
1 2 3
10 10 10 101 102 103
Degrees of Freedom Degrees of Freedom

(a) Bilinear Triangle (b) Bilinear Quadrilateral


100

10-1

10-2
Error

10-3

10-4 L 2 error
H1 error
LP error
O(h 3)
10-5
O(h 2)

102 103 104


Degrees of Freedom

(c) Biquadratic Quadrilateral

Figure 10: The L2 , H 1 and LP error convergence using the Biorthogonal method on the
Nitsche formulation over the subdomain [0, 3/4]2 plotted against the number of degrees
of freedom for example 4

10

You might also like