16.

Solution of elliptic partial differential equation

12/22/2015
Recall in the first lecture of this course ….
Assume…
• you know how to use a computer to compute;
• but have not done any serious numerical computations,
• and only know some very basic numerics.

This course will teach you…

• FORTRAN programing language to compute,
• good habits to write structural Fortran programs,
• to write test driver to test the Fortran program written by you or others,
• to use well developed and optimized libraries, such as LAPACK, FFTW,
• to do parallel computing in multi-process computer using OpenMP
• to use the these skills to solve partial differential equations governing simple
physical systems, such as Poisson equation, wave equation, heat equation.
… and using these basic blocks plus some extra hard work, eventually you are able to
develop your own numerical program to simulate the more realistic physical system,
such as …
Numerical simulation of flow motion
• A popular tool due to the burst in power of computing.
• Substitute for experiment when measurement is inaccessible.
 Governing equations for motion of an incompressible fluid
• Property variables to describe the flow:
velocity: = , , ,
= , , , ̂+ , , , ̂+ , , ,
pressure: = ( , , , )

• Equations governing the flow:

incompressibility of fluid (solenoidal condition):

+ + =0 ⋅ =0

momentum conservation:
1 1
=− − − − + + + =− ⋅ − +

1 =− − ( )
=− − − − + + +

1
=− − − − + + +

• 4 unknowns with 4 equations

Numerical solution based on explicit scheme
Harlow, F.H. & Welch, J.E., 1965. Numerical calculation of time-dependent viscous
incompressible flow of fluid with free surface. Phys. Fluids, 8(12), 2182−2189.

• Define velocity and pressure at discrete time instance = Δ : and

⋅ =0 ⋅ =0

−
=− − ( ) ≈ = ≈− − ( )
Δ
i.e., after obtaining and at , marching towards to get

• But, 1. How is the solenoidal condition ⋅ = 0 satisfied?

2. What is the equation for the pressure ?

• Take divergence of the discretized momentum equation:

−
⋅ ≈− − ( )
Δ
• Take divergence of the time-discretized momentum equation:
− ⋅ − ⋅
⋅ ≈− − ⟹ ≈− ⋅ − ⋅ ( )
Δ Δ

• Ideally, ⋅ = 0 and ⋅ = 0, since the solenoidal condition ⋅ = 0 must be

satisfied at every time instance.

• But due to numerical approximation, the solenoidal condition may not be satisfied exactly.
Let say, there is a small value of ⋅ at = .

• We then project that ⋅ = 0 at = .

− ⋅ 1
∴ ≈− ⋅ − ⋅ ⟹ ≈ ⋅ − ⋅ ( )
Δ Δ
• This is an elliptic-type partial differential equation called Poisson equation:

+ + = , ,

• So the solenoidal condition at = is satisfied by solving the pressure Poisson equation.

This means that pressure is a tuning property to make sure the flow remain solenoidal.
Discrete Fourier Transform in Higher Dimension

• One-dimensional: =

=0 1 2 −1
= =
0 =

1
=

• Two-dimensional:

−1

, = , = ,
=

1 1 2
, = , 1
=0
=0 1 2 −1
=
 Numerical Solution of Poisson Equation

• Elliptic-type partial differential equation for

= , in a rectangular domain:

+ = ,

• Discrete representation of the continuous function:

= , ⟹ , = ( , ) −1

= , ⟹ , = ( , )
=

• Numerical solution of the equation means 2

1
finding discrete , satisfies the discretized equation:
=0
=0 1 2 −1
+ ≈ , =
, ,
Rectangular domain with periodic conditions in both directions

, at each grid point can be represented by a discrete Fourier series with

the coefficients to be determined:

, = ̂ , + = ,

= 0, 1, 2, … , −1 = 0, 1, 2, … , −1 −1

=
Similarly,

2
, = , 1
=0
=0 1 2 −1
=

Since , are given ⟹ , are known.

 + = ,

, = ̂ , , = ,

• Substitute , and , into Poisson equation:

2 2
̂ , + ̂ , = ,

2 2
− ̂ , − ̂ , = , = , =

,
⟹ ̂ , =−
+

• This is valid when and are not both equal to zero.

• = 0 and = 0 corresponds to the constant mode.

• This is because the solution of the Poisson equation subject to periodic boundary conditions in
both direction is indeterminant,
i.e. any constant, , = , can be the solution.

∴ For the well posedness of the problem, ̂ , ≡ 0.

 Rectangular domain periodic condition in one direction
= given

+ = ,
periodic periodic

=0 1 2 −1
= given

• Since , is periodic in direction only, it can be represented as:

= − ̂

= , = ̂ ⟹
̂
=
Similarly,

= , =
• Substitute the expansions of , and , into the Poisson equation:

̂ 2
− ̂ + = =

̂
⇒ − ̂ =

• This is an ordinary differential equation for ̂ subject to the upper and lower boundary conditions
of Dirichlet type for different .
• If 2nd-order finite-difference scheme is used to approximate the differentiation of :

̂ = given = ̂ , given

̂ = ̂ ,

̂ +1
̂ , −2 ̂ , + ̂ ,
− ̂ = − ̂ , = ,
−1 △
2
∆= =

̂ = 0 given =0 ̂ , given
 Discretized equation for ̂ = ̂ : ̂ , −2 ̂ , + ̂ ,
, − ̂ , = ,
△
known
−1 ̂ , − 2+ ∆ ̂ , + ̂ , =△ ,
−2

+1
̂ , − 2+ ∆ ̂ , + ̂ , =△ ,
−1

2
1 ̂ − 2+ ∆ ̂ + ̂ =△
, , , ,
=0 known

 The discretized equation can be represented as a system of linear equation:

− 2+ ∆ 1 , △ , − ,
, ⋮
⋮ ⋮
− 2+ ∆ 1 △ ,
1 ,
1 − 2+ ∆ 1 , = △ ,
1 − 2+ ∆ 1 , △ ,
⋮ ⋮
, ⋮
1 − 2+ ∆ , △ , − ,

 The ( − 1)× − 1 matrix is a tridiagonal, and can be solved efficiently for various .
 Only need to consider = 0 to /2 for real DFT.
Homework 18
￭ Write a subroutine solving Poisson equation in a rectangular domain with periodic boundary
conditions in direction and Dirichlet conditions on the other two boundaries.
￭ Use routine in Lapack to solve the tridiagonal system of linear equation. Use FFFTW to do
discrete Fourier transform. Use random numbers between −1 and 1.
￭ In calling the subroutine, the following data are input :
• the length and width of the rectangular domain and
• the numbers of discrete grids and
• the source function , , = 0 ~ ( − 1), = 1 ~ ( − 1)
• the boundary value on the bottom and top , and , , = 0 ~ ( − 1)
￭ Write a test driver (use the second approach in the following) to assess the maximum error
calling the subroutine to confirm if the Poisson solver has been implemented correctly.
given

( − 1)
periodic in periodic in

+ = ,
2
1
=0
=0 1 2 ( −1)
given
 Two approaches to test the Poisson solver
 The first approach:
• Invent a known analytical function , and compute ANALYTICALLY the right-hand-side
source function of the Poisson equation ( , ) and the boundary conditions:
′ ′
, = +

,0 = ′ ,0

, = ′ ,

• Given the analytical source function , and the boundary conditions , and , , the Poisson
equation is then solved for , , and the numerical solution is compared with the known
function , .
−1

+ ≈ , =
, ,
2
1
=0
=0 1 2 −1
=
• Such an approach can be used to test the convergence of the solver by increase the grid points.
 The second approach:

• Generating the solutions , and the boundary conditions , and , using random numbers.

• The source function is then computed numerically using spectral scheme in the periodic direction
and finite-difference scheme in the other direction.
′ ′ −1
, = +
, ,
=
spectral finite
scheme difference 2
1
, = , , = , =0
=0 1 2 −1
=
• Given the source function , and the boundary conditions , , , , the Poisson equation is
then solved numerically.

+ = ,
, ,

• The numerical results , should be equal to the original given random numbers , since the
forward (computing the source function) and inverse (solving the equation) operators are identical.

• Such approach is for debugging the code.