PROGRAMMING OF FINITE DIFFERENCE METHODS IN MATLAB

LONG CHEN

We discuss efficient ways of implementing finite difference methods for solving the
Poisson equation on rectangular domains in two and three dimensions. The key is the ma-
trix indexing instead of the traditional linear indexing. With such an indexing system, we
will introduce a matrix-free and a tensor product matrix implementation of finite difference
methods.

1. M ATRIX FREE IMPLEMENTATION

Here the ‘matrix free’ means that the matrix-vector product Au can be implemented
without forming the matrix A explicitly. Such matrix free implementation will be useful if
we use iterative methods to compute A−1 f , e.g., the Conjugate Gradient methods which
only requires the computation of Au. Ironically the matrix-free implementation is possible
because a matrix instead of a vector is used to store the function.
Let us use a matrix u(1:m,1:n) to store the function. The following double loops will
compute Au for all interior nodes. The h2 scaling will be moved to the right hand side.
For Neumann boundary conditions, additional loops for boundary nodes are needed since
the boundary stencils are different; see Introduction to Finite Difference Methods.
1 for i = 2:m-1
2 for j = 2:n-1
3 Au(i,j) = 4*u(i,j) - u(i-1,j) - u(i+1,j) - u(i,j-1) - u(i,j+1);
4 end
5 end

Since MATLAB is an interpret language, every line will be complied when it is exe-
cuted. A general guideline for efficient programming in MATLAB is:
avoid large for loops.
A simple modification of the above double loops is to use the vector indexing.
1 i = 2:m-1;
2 j = 2:n-1;
3 Au(i,j) = 4*u(i,j) - u(i-1,j) - u(i+1,j) - u(i,j-1) - u(i,j+1);

To evaluate the right hand side, we can use coordinates (x,y) in the matrix form.
For example, for f (x, y) = 8π 2 sin(2πx) cos(2πy), the h2 scaled right hand side can be
computed as
1 [x,y] = ndgrid(0:h:1,0:h:1);
2 fh2 = hˆ2*8*piˆ2*sin(2*pi*x).*cos(2*pi*y);

Note that .* is used to compute the component-wise product for two matrices. For non-
homogenous boundary conditions, one needs to evaluate boundary values and add to the
right hand side. The evaluation of a function on the whole grid is of complexity O(m × n).
For boundary condition, we can reduce to O(m + n) by restricting to bdidx only.
1
 2 LONG CHEN

1 u(bdidx) = sin(2*pi*x(bdidx)).*cos(2*pi*y(bdidx));

The array bdidx for collecting all boundary nodes can be generated as follows
1 isbd = true(size(u));
2 isbd(2:end-1,2:end-1) = false;
3 bdidx = find(isbd(:));

One Jacobi iteration for solving the matrix equation Au = f can be implemented as
1 j = 2:n-1;
2 i = 2:m-1;
3 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;

The weighted Jacobi iteration can be obtained as a combination of current approximation

of Jacobi iteration. Let uJ be the updated using Jacobi iteration and ω ∈ (0, 1) be a weight.
Then the weighted Jacobi iteration is
1 u = omega*u + (1-omega)*uJ;

A more efficient iterative methods, Gauss-Seidel (G-S) iteration updates the coordinates
sequentially one at a time. Here is the implementation using for loops.
1 for j = 2:n-1
2 for i = 2:m-1
3 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;
4 end
5 end

The ordering does matter in the Gauss-Seidel iteration. The backwards G-S can be imple-
mented by inverse the ordering of i,j indexing.
1 for j = n-1:-1:2
2 for i = m-1:-1:2
3 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;
4 end
5 end

Note that for the matrix-free implementation, there is no need to modify the right hand
side for the Dirichlet boundary condition. The boundary values of u is assigned before the
iteration and remains the same since only the interior nodal values are updated during the
iteration. For Neumann boundary conditions, an additional update on boundary nodes is
needed.
The symmetric version Gauss-Seidel will be the combination of one forwards and one
backwards GS iteration and is an SPD operator which can be used in pcg to accelerate the
computation of an approximated solution to the linear system Au = f .
Vectorization of Gauss-Seidel iteration is subtle. If we simply remove the for loops,
it is the Jacobi iteration since the values of u on the right hand side is the old one. To
vectorize G-S, let us first classify the nodes into two category: red nodes and black nodes;
see Fig 1. Black nodes can be identified as mod(i+j,2) == 0. A crucial observation is
that to update red nodes only values of black nodes are needed and vice verse. Then Gauss-
Seidel iteration applied to this red-black ordering can be implemented as Jacobi iterations.

1 [m,n] = size(u);
2 % case 1 (red points): mod(i+j,2) == 0
 Red-Black Gauss-Seidel
PROGRAMMING OF FINITE DIFFERENCE METHODS IN MATLAB 3

Red depends
F IGURE 1. only on black,
Red-Black and
Ordering vice-versa.
of vertices
Generalization: multi-color orderings
3 i = 2:2:m-1; j = 2:2:n-1;
4 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;
5 i = 3:2:m-1; j = 3:2:n-1;
6 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;
7 % case 2 (black points): mod(i+j,2) == 1
8 i = 2:2:m-1; j = 3:2:n-1;
9 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;
10 i = 3:2:m-1; j = 2:2:n-1;
11 u(i,j) = (fh2(i,j) + u(i-1,j) + u(i+1,j) + u(i,j-1) + u(i,j+1))/4;

2. I NDEXING USING MATRICES

Geometrically a 2-D grid is naturally linked to a matrix. When forming the matrix
equation, we need to use a linear indexing to transfer this 2-D grid function to a 1-D vector
function. We can skip this artificial linear indexing and treat our function u(x, y) as a
matrix function u(i,j). The multiple subscript indexing to the linear indexing is build
into the matrix. The matrix is still stored as a 1-D array in memory. The default linear
indexing in MATLAB is column wise. For example, a matrix A = [2 9 4; 3 5 11] is
stored in memory as the array [2 3 9 5 4 11]’. One can use a single index to access
an element of the matrix, e.g., A(4) = 5.
In MATLAB, there are two matrix systems to represent a two dimensional grid: the
geometry consistent matrix and the coordinate consistent matrix. To fix ideas, we use the
following example. The domain Ω = (0, 1) × (0, 2) is decomposed into a uniform grid
with mesh size h = 0.5. The linear indexing of these two systems are illustrate in the
following figures.
The command [x,y] = meshgrid(0:0.5:1,2:-0.5:0) will produce 5 × 3 matri-
ces. Note that the flip of the ordering [2:-0.5:0] in the y-coordinate makes the matrix is
geometrically consistent with the domain in the sense that the shape of the matrix matches
the shape of the domain. This index system is illustrated in Fig. 2(a).
1 >> [x,y] = meshgrid(0:0.5:1,2:-0.5:0)
2 x =
3 0 0.5000 1.0000
4 0 0.5000 1.0000
5 0 0.5000 1.0000
6 0 0.5000 1.0000
7 0 0.5000 1.0000
 4 LONG CHEN

2 1 6 11 2 13 14 15

1.8 1.8

1.6 1.6
2 7 12 10 11 12
1.4 1.4

1.2 1.2

1 3 8 13 1 7 8 9

0.8 0.8

0.6 0.6
4 9 14 4 5 6
0.4 0.4

0.2 0.2

0 5 10 15 0 1 2 3
0 0.5 1 0 0.5 1

(a) meshgrid (b) ndgrid

F IGURE 2. Two indexing systems

8 y =
9 2.0000 2.0000 2.0000
10 1.5000 1.5000 1.5000
11 1.0000 1.0000 1.0000
12 0.5000 0.5000 0.5000
13 0 0 0
In this geometrically consistent system, however, there is an inconsistency of the con-
vention of notation of matrix and Cartesian coordinate. Let us figure out the mapping
between the algebraic index (i,j) and the geometric coordinate (xi , yj ) of a grid point.
In the command
[x,y] = meshgrid(xmin:hx:xmax,ymax:-hy:ymin),
the coordinate of the (i, j)-th grid point is
(xj , yi ) = (xmin + (j − 1)hx , ymin + (n − i + 1)hy ),
which violates the convention of associating index i to xi and j to yj . For a matrix en-
try A(i,j), the 1st index i is the row and the 2nd j is the column while in Cartesian
coordinate, i is usually associated to the x-coordinate and j to the y-coordinate.
The command ndgrid will produce a coordinate consistent matrix in the sense that the
mapping is (i,j) to (xi , yj ) and thus will be called the coordinate consistent indexing.
For example, [x,y] = ndgrid(0:0.5:1,0:0.5:2) will produce two 3 × 5 not 5 × 3
matrices; see Fig. 2(b).
1 >> [x,y] = ndgrid(0:0.5:1,0:0.5:2)
2 x =
3 0 0 0 0 0
4 0.5000 0.5000 0.5000 0.5000 0.5000
5 1.0000 1.0000 1.0000 1.0000 1.0000
6 y =
7 0 0.5000 1.0000 1.5000 2.0000
 PROGRAMMING OF FINITE DIFFERENCE METHODS IN MATLAB 5

8 0 0.5000 1.0000 1.5000 2.0000

9 0 0.5000 1.0000 1.5000 2.0000
In the output of
[x,y] = ndgrid(xmin:hx:xmax,ymin:hy:ymax),
the coordinate of the (i, j)-th grid point is
(xi , yj ) = (xmin + (i − 1)hx , ymin + (j − 1)hy ).
In this system, one can link the index change to the conventional change of the coordi-
nate. For example, the central difference u(xi + h, yj ) − u(xi − h, yj ) is transferred to
u(i+1,j) - u(i-1,j). When display a grid function u(i,j), however, one must be
aware of that the shape of the matrix is not geometrically consistent with the domain.
Remark 2.1. No matter which indexing system in use, when plotting a grid function using
mesh or surf, it results the same geometrically consistent figures.
As an example we discuss the access of boundary points. Using subscripts of meshgrid
system, the index of each part of the boundary of the domain is
meshgrid: left - (:,1) right - (:,end) top - (1,:) bottom - (end,:)
which is consistent with the boundary of the matrix. If using a ndgrid system, it becomes
ndgrid: left - (1,:) right - (end,:) top - (:,end) bottom - (:,1).
Remember the coordinate consistency: i to x and j to y. Thus the left boundary will be
i = 1 corresponding to x = x1 .
Which index system shall we choose? First of all, choose the one you feel more com-
fortable and thus has less chance to produce bugs. A more subtle issue is related to the
linear indexing of a matrix in MATLAB. Due to the column-wise linear indexing, it is
much faster to access one column instead of one row at a time. Depending on which co-
ordinate direction the subroutine will access more frequently, one chose the corresponding
coordinate-index system. For example, if one wants to use vertical line smoothers, then it
is better to use meshgrid system and ndgrid system for horizontal lines.
We now discuss the transfer between multiple subscripts and the linear indexing. The
commands sub2ind and ind2sub is designed for such purpose. We include two ex-
amples below and refer to the documentation of MATLAB for more comprehensive ex-
planation and examples. The command k=sub2ind([3 5],2,4) will give k=11 and
[i,j]=ind2sub([3 5],11) produces i=2, j=4. In the input sub2ind(size, i,j),
the i,j can be arrays of the same dimension. In the input ind2sub(size, k), the k can
be a vector and the output [i,j] will be two arrays of the same length of k. Namely these
two commands support vectors arguments.
For a matrix function u(i,j), u(:) will change it to a 1-D array using the column-wise
linear indexing and reshape(u,m,n) will change a 1-D array to a 2-D matrix function.
A more intuitive way is to explicitly store an index matrix. For meshgrid system, use
idxmat = reshape(uint32(1:m*n), m, n);

1 >> idxmat = reshape(uint32(1:15),5,3)

2 idxmat =
3 1 6 11
4 2 7 12
5 3 8 13
6 4 9 14
7 5 10 15
 6 LONG CHEN

Then one can easily get the linear indexing of the j-th column of a m × n matrix by
using idxmat(:,j) which is equivalent to sub2ind([m n], 1:m, j*ones(1,m)) but
much easier and intuitive. The price to pay is the extra memory for the full matrix idxmat
which can be minimized using uint32.
For the ndgrid system, to get a geometrically consistent index matrix, we can use the
following command.
idxmat = flipud(transpose(reshape(uint32(1:m*n), n, m))));

1 >> idxmat = flipud(reshape(uint32(1:15),3,5)’)

2 idxmat =
3 13 14 15
4 10 11 12
5 7 8 9
6 4 5 6
7 1 2 3

For such coordinate consistent system, however, it is recommended to use the subscript
indexing directly.
Similarly we can generate matrices to store the subscripts. For the meshgrid system
1 >> [jj,ii] = meshgrid(1:3,1:5)
2 jj =
3 1 2 3
4 1 2 3
5 1 2 3
6 1 2 3
7 1 2 3
8 ii =
9 1 1 1
10 2 2 2
11 3 3 3
12 4 4 4
13 5 5 5

For the ndgrid system

1 >> [ii,jj] = ndgrid(1:3,1:5)
2 ii =
3 1 1 1 1 1
4 2 2 2 2 2
5 3 3 3 3 3
6 jj =
7 1 2 3 4 5
8 1 2 3 4 5
9 1 2 3 4 5

Then ii(k),jj(k) will give the subscript of the k-th node.

The linear indexes of all boundary nodes can be found by the following codes
1 isbd = true(size(u));
2 isbd(2:end-1,2:end-1) = false;
3 bdidx = find(isbd(:));

In the first line, we use size(u) such that it works for both meshgrid and ndgrid system.
 PROGRAMMING OF FINITE DIFFERENCE METHODS IN MATLAB 7

3. T ENSOR PRODUCT MATRIX IMPLEMENTATION

When the grid is in the tensor product type, the matrix-vector multiplication Au can be
also implemented using the tensor product of 1-D matrix which will be called the tensor-
product matrix implementation.
For a uniform grid in one dimension, the matrix of the central difference discretization
of the Poisson equation is tri-diagonal and can be generated by
1 e = ones(n,1);
2 T = spdiags([-e 2*e -e], -1:1, n, n);

The boundary condition can be build into T by changing the entries near the boundary.
Here T corresponds to the homogenous Dirichlet boundary condition. And T is stored as a
sparse matrix to save memory and operations.
For a two dimensional n × n uniform grid, the five point stencil can be decomposed into
(2ui,j − ui−1,j − ui+1,j ) + (2ui,j − ui,j−1 − ui,j+1 )
which can be realized by the left product and right product with the 1-D matrix
Au = u*T + T*u;

For different mesh size or different stencil in x and y-direction, one should generate
specific Tx and Ty and use
1 Au = u*Tx + Ty*u; % meshgrid system
2 Au = Tx*u + u*Ty; % ndgrid system

We can write the matrix as the tensor product of 1-D matrices. Let Am×n and Bp×q be
two matrices. Then the Kronecker (tensor) product of A and B is
 
a11 B . . . a1n B
A ⊗ B =  ... ... ... 
am1 B . . . amn B
The matlab command is kron(A,B).
Let Am×m , Bn×n and Xm×n be there matrices. Then it is straightforward to verify the
identities
(1) (AX)(:) = (In ⊗ A) · X(:).
(2) (XB)(:) = (B | ⊗ In ) · X(:).
Here we borrow the notation (:) to change a matrix to a vector by stacking columns from
left to right. Therefore the matrix A for the five point stencil is
A = In ⊗ T + T ⊗ In ,
and the corresponding matlab code is
1 A = kron(speye(nx),Ty) + kron(Tx,speye(ny)); % meshgrid system
2 A = kron(speye(ny),Tx) + kron(Ty,speye(nx)); % ndgrid system

Exercise 3.1. Write out a similar formulae for Neumann boundary condition.
Hint: Change both T and I at boundary indices.
Again in the computation, it is not needed to form A. Instead use the left and right
matrix-product to compute Au if only the matrix-vector product is of interest.
 8 LONG CHEN

4. T HREE AND H IGHER D IMENSIONS

The meshgrid can be used to generate a 3-D tensor product grid and ndgrid can gen-
erate an n-D grid for any positive integer n. The two dimensional matrix can be generalized
to multi-dimensional arrays with more than two subscripts (also called tensor). Please read
the help doc on Multidimensional Arrays in MATLAB first. In the following we
discuss issues related to the implementation of the Poisson equation in 3-D.
Slices in each direction are in different type. Only A(:,:,i) is a matrix stored con-
secutively, which is called the i-th page of A. But A(i,:,:) will be formed by elements
across pages and thus not a matrix. One can use squeeze(A(i,:,:)) to squeeze it into
a matrix. Again it is stored column wise and which coordinate (x, y or z) corresponds to
the column will depend on the index system.
The tensor product representation of the matrix is still valid in high dimensions. For
example, in 3-D the 7-point stencil Laplace matrix is
A = In ⊗ In ⊗ T + In ⊗ T ⊗ In + T ⊗ In ⊗ In .
The matrix-free computation of Au is straightforward
1 [n1,n2,n3] = size(u);
2 i = 2:n1-1;
3 j = 2:n2-1;
4 k = 2:n3-1;
5 Au(i,j,k) = 6*u(i,j,k) - u(i-1,j,k) - u(i+1,j,k) - u(i,j-1,k) - u(i,j+1,k) ...
6 u(i,j,k-1) - u(i,j,k+1);

The tensor product matrix implementation is less obvious since the basic data structure
in MATLAB is matrix not tensor. Denote the stencil matrix in each direction by Ti , i =
1, 2, 3. The first two dimensions can be computed as
1 for k = 1:n3
2 Au(:,:,k) = u(:,:,k)*T2 + T1*u(:,:,k);
3 end

The third one is different

1 for j = 1:n2
2 Au(:,j,:) = squeeze(u(:,j,:))*T3;
3 end

To vectorize the above code, i.e., avoid for loop, one can use reshape which operates
in a column-wise manner. First think about the original data as a long vector by stacking
columns. Then reshape will create the reshaped matrix by transforming consecutive
elements of this long vector into different shape.
We explain the index change by the following example.
1 >> u = reshape(1:3*5*2,3,5,2)
2 u(:,:,1) =
3 1 4 7 10 13
4 2 5 8 11 14
5 3 6 9 12 15
6
7 u(:,:,2) =
8 16 19 22 25 28
9 17 20 23 26 29
 PROGRAMMING OF FINITE DIFFERENCE METHODS IN MATLAB 9

10 18 21 24 27 30

1 >> u1 = reshape(u, n1, n2*n3)

2 u1 =
3 1 4 7 10 13 16 19 22 25 28
4 2 5 8 11 14 17 20 23 26 29
5 3 6 9 12 15 18 21 24 27 30
Reshape in this way is like to put pages consecutively on the plane. This is efficient since
it is just a rearrangement of columns.
The stencil in the first direction can be realized by
1 Au1 = reshape(T1*reshape(u, n1, n2*n3), n1, n2, n3);
However, we cannot use similar trick to implement the stencil in the other two direc-
tions. For example,
1 >> reshape(u, n1*n3, n2)
2 ans =
3 1 7 13 19 25
4 2 8 14 20 26
5 3 9 15 21 27
6 4 10 16 22 28
7 5 11 17 23 29
8 6 12 18 24 30
Multiplication of T2 to the right will not get the desired result. A simple fix is to use
permute to permute the desired direction to the first one and using the operation ipermute
to switch back afterwards.
1 up = permute(u, [2 1 3]);
2 Au2 = reshape(T2*reshape(up, n2, n1*n3), n2, n1, n3);
3 Au2 = ipermute(Au2, [2 1 3]);
Repeat this procedure for each direction and add them together to get Au. It seems
cumbersome to using the tensor product matrix implementation comparing with the matrix-
free one. The advantage is: one can easily build the boundary condition, the non-uniform
grid size, and non-standard stencil into the one dimensional matrix. No need to loop over
boundary indices to modify the stencil.