BASIC NUMERICAL METHODS FOR IMAGE

PROCESSING

CHANG-OCK LEE

The purpose of this lecture note is to show how finite-difference schemes

can be used in image analysis. We first introduce in Section 1 the main
notation and consider the 2-D heat equation. The remainder of this note is
concerned with the discretization of certain PDEs used in image analysis:
• Restoration by energy minimization (Section 2): We detail the dis-
cretization of the divergence term, which can also be found for the
Perona and Malik equation.
• Enhancement by Osher and Rudin’s shock filters (Section 3): The
main interest is to use a flux limiter called minmod.
• Curve evolution with level sets and especially segmentation with
geodesic active contours (Section 4). For the sake of simplicity. we
examine separately each term of the model (mean curvature motion,
constant speed motion, advection equation). We essentially write
their discretization and give some experimental results.
For more details of finite difference methods, see [2, 6, 12]; for hyperbolic
equations, see [3, 5]
Note: This lecture note is for the three hours lecture in the NIMS Work-
shop and Minicourse “Mathematical Analysis, Numerics and Applications
in Fluid and Gas Dynamics”. This is a part of the book [1].

1. Getting Started
In this lecture note we will show how certain PDEs used in image analysis
can be discretized. The generic form of these PDEs is


 Lv = F (t, x, y) ∈ R+ × Ω,


∂v
(t, x, y) = 0 on R+ × ∂Ω,

 ∂N

v(0, x, y) = f (x, y),

where Ω is the image domain and N is the normal to the boundary of Ω,

denoted by ∂Ω. L is generally a second-order differential operator such as

∂v
(t, x, y) + H(x, y, v(t, x, y), ∇v(t, x, y), ∇2 v(t, x, y)) = 0.
∂t
1
2 CHANG-OCK LEE

Figure 1: Grid on the space domain. The circles indicate the vertices
that belong to the 3 × 3 neighborhood of the vertex (i, j).

Example: One of the simplest PDEs presented in image analysis is the

heat equation:
 µ 2 ¶
 ∂v ∂ v ∂2v

 = ν∆v = ν + (t, x, y) ∈ R+ × Ω,

 ∂t ∂x2 ∂y 2
(1.1) ∂v

 (t, x, y) = 0 on R+ × ∂Ω,

 ∂N

v(0, x, y) = f (x, y),

where ν is a positive constant.

In image processing, finite differences are widely used, which is due to the
digital structure of an image as a set of pixels uniformly distributed. It is
then very easy and natural to associate with an image a uniform grid, as
presented in Figure 1. Since there is no reason to choose it differently, the
grid spacing in the x and y directions is usually equal:

∆x = ∆y = h.

Notice that in many articles from the computer vision literature, it is even
choose as h = 1, which means that the pixel size is chosen as the unit
of reference. We will call the positions (ih, jh) vertices, nodes, or pixels
equivalently. We will denote by vi,jn (respectively un ) the values of the
i,j
exact solution (respectively the discrete solution) at location (ih, jh) and
tine n∆t.
Example: The PDE (1.1) is an initial-boundary value problem. To dis-
cretize it, we need to consider the following:
• The equation. To find the difference scheme associated with the
heat equation (1.1), we use Taylor expansions expanded about the
index (n∆t, ih, jh). Naturally, the simplest method is to consider
separately the discretization of each second-order derivative in x and
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 3

Figure 2: Example of results with the scheme (1.2) at different times

(iterations), as applied to a simple and small image (32 × 32).

Figure 3: Representation of the vertices involved in the finite difference

schemes. • approximation (1.3), ° approximation (1.4)

y. By doing so, we obtain

¯n n+1 n
∂v ¯ vi,j − vi,j
− ν∆v ¯¯ =
∂t i,j ∆t
n
vi+1,j n
+ vi−1,j n
+ vi,j+1 n
+ vi,j−1 n
− 4vi,j
−ν
h2
2
+ O(∆t) + O(h ) .
Then the difference scheme that we can propose is
ν∆t ¡ n ¢
(1.2) un+1 n
i,j = ui,j + 2
ui+1,j + uni−1,j + uni,j+1 + uni,j−1 − 4uni,j .
h
It is of order (2, 1).
• The boundary condition. The Neumann boundary condition can be
taken into account by a symmetry procedure. If the value of a pixel
(vertex) that is outside the domain is needed, we use the value of
the pixel that is symmetric with respect to the boundaries.
• The initial condition is simply: u0i,j = gi,j , where g is the discretiza-
tion of f .
To illustrate this algorithm, we show in Figure 2 some iteration as applied to
a very simple image. Notice that this example shows clearly the propagation
of the information as the number of iteration increases.
4 CHANG-OCK LEE

Figure 4: Example of a binary image representing a vertical edge and the

same image after a rotation of π/4 radians. Rotationally invariant
operators estimated at the point in the middle should yield the same value
in both situations

Remark The scheme (1.2) was obtained by discretizing the Laplacian as a

sum of the second-order derivatives in the x and y directions (see Figure 3):
vi+1,j + vi−1,j + vi,j+1 + vi,j−1 − 4vi,j
(1.3) ∆v|i,j ≈ .
h2
Clearly, this discretization does not take into account the 2-D nature and
properties of this operator. To illustrate what we mean by “2-D nature
and properties,” we can remark that the Laplacian operator is rotationally
invariant. If we apply a rotation of center (x, y) to the image v (with any
angle θ ∈ [0, 2π)), then ∆v(x, y) remains constant for all θ, as it should be
for the discretization. Naturally, as we consider a discrete domain, we may
ask only that ∆v|i,j keep constant under rotations of π/4, as depicted in
Figure 4. That is not the case for discretization (1.3), since we obtain 1 or
2 (for h = 1) depending on the situation.
To overcome this difficulty, we need to use the complete 3 × 3 neighbor-
hood. We may propose the following approximation:
vi+1,j + vi−1,j + vi,j+1 + vi,j−1 − 4vi,j
∆v|i,j ≈ λ
(1.4) h2
vi+1,j+1 + vi−1,j+1 + vi+1,j−1 + vi−1,j−1 − 4vi,j
+ (1 − λ) ,
2h2
where λ ∈ [0, 1] is a constant to be chosen. We can verify that this approxi-
mation is consistent. Applying this operator (1.4) in the two situations from
Figure 4 and saying that both results should be equal yields λ = 31 , hence
the approximation.
Similarly, we can propose a discretization for the first-order derivatives
in x and y that is consistent with the fact that the norm of the gradient
is invariant under rotation. As we will see further, a second-order centered
approximation of the first derivative in x is
¯
∂v ¯¯ vi+1,j − vi−1,j
(1.5) ¯ ≈ δx vi,j = ,
∂x i,j 2h
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 5

Figure 5: Representation of the vertices involved in the finite difference

schemes. • approximation (1.5), ° approximation (1.6)

which can also be written in the y direction. The vertices involved in the
estimation (1.5) are represented in Figure 5. As for the case of the Laplacian,
these approximations are in fact “one-dimensional” and do not really take
advantage of the fact that the data is of dimension 2. This is visible if
we consider the value of the norm of the gradient of u in the two situations
described in Figure 4: We obtain either 12 or √12 . The solution is to use more
pixels in the estimation of the derivatives. In particular, we may suggest
the following approximation:
¯
∂v ¯¯ vi+1,j − vi−1,j
¯ ≈λ
∂x i,j 2h
(1.6) µ ¶
(1 − λ) vi+1,j+1 − vi−1,j+1 vi+1,j−1 − vi−1,j−1
+ + ,
2 2h 2h

where λ is a parameter to be chosen. Applying the operator (1.6) in the two

situations
√ of Figure 5 and by saying that both results should be equal yields
λ = 2 − 1, hence the approximation.
Finally we want to mention that not only is using more points in the
approximations good for the rotation-invariance properties, but, practically,
the result is also less sensitive to noise.The reason is that it is equivalent to
perform a smoothing of the data before the estimation.

2. Image Restoration by Energy Minimization

We first consider the image restoration problem which becomes to mini-
mize with respect to v and b the functional
Z Z
1
J² (v, b) = |Rv − v0 | dx + λ (b|∇v|2 + ψ² (b)) dx .
2
2 Ω Ω

The so-called half-quadratic minimization algorithm consists in minimizing

successively J² with respect to each variable. The algorithm is as follows:
For (v 0 , b0 ) given
6 CHANG-OCK LEE

Figure 6: Vertices involved in the approximation of the divergence term

for the different schemes. × approximation (2.1), • approximation (2.2),
° approximation (2.3)

• v n+1 = argminv J² (v, bn ) i.e.,

 ∗ n+1
 R Rv − div(bn ∇v n+1 ) = 0 in Ω,
 bn ∂v n+1
=0 on ∂Ω.
∂N
φ0 (|∇v n+1 |)
• bn+1 = argminb J² (v n+1 , b) i.e., bn+1 = .
2|∇v n+1 |
• Go back to the first step until convergence.
The limit (v ∞ , b∞ ) is the solution.
As far as discretization is concerned, the only term that may be difficult to
approximate is the divergence operator. So for b ≥ 0 and v given at nodes
(i, j) the problem is to find an approximation at the node (i, j) for div(b∇v).
This kind of term is naturally present as soon as there is a regularization
with a φ function. The diffusion operator used in the Perona and Malik
model is also of the same kind.
Since this divergence operator may be rewritten as:
µ ¶ µ ¶
∂ ∂v ∂ ∂v
div(b∇v) = b + b ,
∂x ∂x ∂y ∂y
we can use the previous one-dimensional approximation and combine them.
For example, if we use the central finite-difference approximation (1.5), we
have
(2.1)
div(b∇v)|i,j ≈ δx (bi,j δx vi,j ) + δy (bi,j δy vi,j )
1
= 2 (bi+1,j vi+2,j + bi−1,j vi−2,j + bi,j+1 vi,j+2 + bi,j−1 vi,j−2
4h
− (bi+1,j + bi−1,j + bi,j+1 + bi,j−1 )vi,j ) .
The main drawback of this representation is that it involves only the points
((i ± 2)h, (j ± 2)h), and none of the 3 × 3 neighborhood (see also Figure 6).
This may be nonrobust for noisy data or when there is considerable variation
in this region. Another possibility is to combine forward and backward
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 7

differences:
div(b∇v)|i,j ≈ δx+ (bi,j δx− vi,j ) + δy+ (bi,j δy− vi,j )
1
= 2 (bi+1,j vi+1,j + bi,j vi−1,j + bi,j+1 vi,j+1 + bi,j vi,j−1
h
− (bi+1,j + bi,j+1 + 2bi,j )vi,j ) .
This approximation now involves the 3 × 3 neighborhood, but it introduces
an asymmetry: The values of b at ((i − 1)h, jh) and (ih, (j − 1)h) are not
used. A solution is to use the following approximation for the derivatives:
vi+ 1 ,j − vi− 1 ,j vi,j+ 1 − vi,j− 1
δx∗ vi,j = 2 2
and δy∗ vi,j = 2 2
,
h h
where vi± 1 ,j± 1 is the value of v at location ((i ± 21 )h, (j ± 21 )h), which can be
2 2
obtained by interpolation. As for (1.5). it is a second-order approximation.
Then we have
div(b∇v)|i,j ≈ δx∗ (bi,j δx∗ vi,j ) + δy∗ (bi,j δy∗ vi,j )
1
(2.2) = 2 (b+0 vi+1,j + b−0 vi−1,j + b0+ vi,j+1 + b0− vi,j−1
h
− (b+0 + b−0 + b0+ + b0− )vi,j ) .
where b±0 = bi± 1 ,j and b0± = bi,j∓ 1 . Notice that since we applied the
2 2
operators δx∗ and δy∗ twice, this approximation uses the values of v only at
((i ± 1)h, (j ± 1)h) (see Figure 6). However, interpolation is needed for b.
As mentioned previously for the estimation of the Laplacian, it would also
be interesting to take into account the diagonal values. Then, we can look
for an approximation such that
(2.3)
div(b∇v)|i,j ≈
λp
(b+0 vi+1,j + b−0 vi−1,j + b0+ vi,j+1 + b0− vi,j−1 − βp vi,j )
h2
λd
+ 2 (b++ vi+1,j+1 + b−− vi−1,j−1 + b−+ vi−1,j+1 + b+− vi+1,j−1 − βd vi,j )
h
with b±± = bi± 1 ,j± 1 ,
2 2
½
βp = b0+ + b0− + b+0 + b−0 ,
βd = b++ + b−− + b+− + b−+ ,
and where λp and λd are two weights to be chosen. The first condition is
that the scheme must be consistent, and it can be verified that this implies
λp + 2λd = 1.
Now there remains one degree of freedom. Two possibilities can be consid-
ered:
• The first is to choose (λp , λd ) constant, and for instance equal to
( 12 , 14 ), giving a privilege to the principal directions.
• The second is to choose (λp , λd ) by taking into account the orienta-
tion of the gradient of v, as described in Figure 7.
8 CHANG-OCK LEE

Figure 7: Adaptive choice of the coefficients (λp , λd ) as a function of θ,

the orientation of ∇v. The two right-hand figures show which points will be
used in the discretization of the divergence term in two specific situations.

Figure 8: Numerical tests for the different discretization of the

divergence term (the choice of (λp , λd ) is indicated below the images).

We tested these different discretizations as applied to a simple image with

geometric structure (see Figure 8). From left to right, an improvement in
the results can be perceived (by observing the restoration of the horizontal
and vertical edges). It is the adaptive choice that gives the best result.

3. Image Enhancement by Osher and Rudin Shock Filters

This section concerns the shock-filter equation proposed by Osher and
Rudin [7]:
∂v
(3.1) = −|∇v|F (L(v)),
∂t
where:
• F is a Lipschitz function satisfying F (0) = 0, sign(s)F (s) > 0 (s 6=
0), for example F (s) = sign(s).
• L is a second-order edge detector, for example
L(v) = ∆v = vxx + vyy
or
1
L(v) = (v 2 vxx + 2vx vy vxy + vy2 vyy ),
|∇v|2 x
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 9

Figure 9: Approximation of the first derivative using the minmod

function.

Figure 10: Example of the shock filter on a blurred image of a face. It

shows clearly that this filter reconstructs a piecewise constant image that
is not satisfying from a perceptual point of view (the result does not look
like a real image).

which corresponds to the second derivative of v in the direction of

the normal to the isophotes.
Equation (3.1) involves two kinds of terms: a first-order term |∇v| and a
second-order term F (L(v)):
• L is discretized with central finite differences.
• |∇v| has to be approximated with more care. vx and vy are approx-
imated using the minmod operator m(α, β). For instance,
vx |i = m(δx− vi , δx+ vi ),
where
½
sign(α) min(|α| , |β|) if αβ > 0,
m(α, β) =
0 if αβ ≤ 0.
This function is usually called a flux limiter. As shown in Figure 9,
it permits us to choose the lowest slope, or zero in case of a local
extremum (this prevents instabilities due to noise).
10 CHANG-OCK LEE

To summarize, the approximation of (3.1) is then given by

∆t q
un+1
i,j = u n
i,j − (m(δx+ uni,j , δx− uni,j ))2 + (m(δy+ uni,j , δy− uni,j ))2 Fi,j (L(un )),
h
where Fi,j (L(u)) = F (Li,j (u)). We show an example in Figure 10.

4. Curve Evolution with the Level-Set Method

In this section we briefly discuss the discretization of the PDEs governing
curve evolution. We examine only the case where these curves are identified
as level sets of the same function u(t, x) (Eulerian formulation). Of course,
we have in mind the geodesic active contours model given by
µ ¶
∂v ∇v
(4.1) = g(|∇I|) |∇v| div + αg(|∇I|) |∇v| + ∇g · ∇v.
∂t |∇v|
As mentioned before, (4.1) involves two kinds of terms: a parabolic term (the
first one) and hyperbolic terms (the last two). One can easily imagine that
the discretization of each term needs an appropriate treatment, according
to its nature (parabolic or hyperbolic). The main idea is that parabolic
terms can be discretized by central finite differences, while hyperbolic terms
need to be approximated by nonoscillatory upwind schemes. For the sake of
clarity we start by examining the evolution driven by each of these terms.
For a detailed description of the above scheme and other numerical ques-
tions not developed here, we refer the reader to [8, 11]

4.1. Mean Curvature Motion. Let us consider

 µ ¶
 ∂v = |∇v| div ∇v ,
(4.2) ∂t |∇v|

v(0, x, y) = v0 (x, y).
Equation (4.2) is a parabolic equation and has diffusive effects (like the heat
equation). So, the use of upwind schemes is inappropriate, and classical
central differences are used,
q
un+1
i,j = u n
i,j + ∆t n
(δx uni,j )2 + (δy uni,j )2 Ki,j ,
n is the central finite-difference approximation of the curvature:
where Ki,j
µ ¶
∇v vxx vy2 + vyy vx2 − 2vx vy vxy
K = div = .
|∇v| (vx2 + vy2 )3/2
n . Unfortunately, the discretiza-
We let the reader write the expression of Ki,j
tion of (4.2) is not as easy as it may appear. At some points (t, x), ∇v can
be undefined, or |∇v| = 0, or |∇v| = +∞. This situation can occur even in
very simple cases [9]. For example. let us consider the shrinking of a unit
circle in two dimensions which corresponds to
p
v(0, x, y) = x2 + y 2 − 1,
i.e., v0 is the signed distance to the unit circle. Equation (4.2) is rotationally
invariant, and if we search for the solution of the form
p
v(t, x, y) = φ(t, x2 + y 2 ),
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 11

Figure 11: Example of mean curvature motion.

p
we easily get v(t, x, y) = x2 + y 2 + 2t − 1, from which we deduce
µ ¶ p
1 x x2 + y 2
∇v = p , |∇v| = p ,
x2 + y 2 + 2t y x2 + y 2 + 2t
µ ¶ µ ¶
∇v 1 x ∇v 1
=p , div =p ,
|∇v| 2
x +y 2 y |∇v| x + y2
2

so the two last quantities are not defined at the origin, and effectively a
spike occurs at the origin (see [9]). Moreover, the interface
Γ(t) = {(x, y) : v(t, x, y) = 0}
√
is the circle x2 + y 2 = 1 − 2t, and on Γ(t) we have |∇v| (t) = 1 − 2t.
Therefore v(t, x, y) becomes more and more flat as the interface evolves and
disappears at t = 21 . To circumvent this type of problem, we have to find a
numerical trick that prevents the gradient norm form vanishing (or blowing
up). This can be realized by reinitializing the function v from time to time
to a signed distance function.
More precisely, we run (4.2) until some step n; then we solve the auxiliary
PDE

 ∂φ + sign(φ)(|∇φ| − 1) = 0,
(4.3) ∂t
 φ(0, x, y) = v(n∆t, x, y).

The resulting solution (as t tends to infinity), denoted by φ∞ , is a signed

distance function whose zero-level set is the same as that of the function
v(n∆t, x, y). Then we can run again (4.2) with the initial data v(0, x, y) =
φ∞ (x, y). Practically, this reinitialization has to be done every n = 20
iterations of the curve evolution equation, and it is usually performed about
every 5 to 10 iterations of (4.3).1
Remark There exists another way to avoid doing the reinitialization step.
It consists in considering a modified equation (4.2) that has the property
of maintaining the norm of the gradient of the solution equal to one. For
further details see [4, 9].

1These numbers are just an indication and naturally depend on the kind of equation
to be solved and on the time steps
12 CHANG-OCK LEE

Figure 12: Example of constant speed motion (c = 1, grass fire).

Figure 13: Example of constant speed motion (c = −1).

An example of mean curvature motion is shown in Figure 11. Notice that

if we let the evolution run until convergence, any curve transforms into a
circle and then collapses.
4.2. Constant Speed Evolution. The second example is given by

 ∂v = c |∇v| ,
(4.4) ∂t
 v(0, x, y) = v (x, y),
0

where c is a constant. This equation describes a motion in the direction

normal to the front (the corresponding Lagrangian formulation of (4.4) is
∂Γ
∂t (t, p) = cN (t, p), where N is the normal to Γ(t)). For c = 1, it is also
referred to as grass fire, since it simulates a grass fire wave-front propagation.
Equation (4.4) is approximated by a nonoscillatory upwind scheme:
un+1 n + n
i,j = ui,j + ∆t∇ ui,j ,
where
£
∇+ uni,j = max(δx− uni,j , 0)2 + min(δx+ uni,j , 0)2
¤1/2
+ max(δy− uni,j , 0)2 + min(δy+ uni,j , 0)2 .
We show in Figures 12 and 13 two examples of constant speed motions.
Remark Motions like equation (4.4) and more generally with a monotone
speed have the following property: Every point is crossed once and only
once by the curve during its evolution. Notice that this is not the case for
mean curvature motion. This property can be used to derive an efficient
numerical approach called the fast marching algorithm [10, 13]. It is beyond
the scope of this lecture note to explain this method, and we refer to the
original articles and to [11] for more detail.
 BASIC NUMERICAL METHODS FOR IMAGE PROCESSING 13

Figure 14: Example of segmentation. Different iterations are displayed.

4.3. The Pure Advection Equation. We consider here the equation


 ∂v = A(x, y) · ∇v,
(4.5) ∂t
 v(0, x, y) = v (x, y),
0

where A(x, y) = (A1 (x, y), A2 (x, y)). For (4.5) we use a simple upwind
scheme, i.e., we check the sign of each component of A and construct a
one-side upwind difference in the appropriate direction:
£
un+1 n n − n n + n
i,j = ui,j + ∆t max((A1 )i,j , 0)δx ui,j + min((A1 )i,j , 0)δx ui,j
¤
+ max((A2 )ni,j , 0)δy− uni,j + min((A2 )ni,j , 0)δy+ uni,j .
4.4. Image Segmentation by the Geodesic Active Contour Model.
Now we can consider the geodesic active contour model (4.1), which can be
seen as the sum of the previous discretization. So the discrete scheme is
h £ ¤1
un+1
i,j = un
i,j + ∆t n
gi,j n
Ki,j (δx uni,j )2 + (δy uni,j )2 2
£ n
¤
+ α max(gi,j , 0)∇+ + min(gi,j n
, 0)∇− uni,j
+ max((gx )ni,j , 0)δx− uni,j + min((gx )ni,j , 0)δx+ uni,j
¤
+ max((gy )ni,j , 0)δy− uni,j + min((gy )ni,j , 0)δy+ uni,j ,
where ∇− uni,j is obtained from ∇+ uni,j by inverting the signs plus and minus.
Figure 14 shows a typical example.
Remark From a numerical point of view, all the equations presented in this
section involve local operations. Since we are interested only in the curve,
it is enough to update the values in a band around the current position of
the curve, also called a narrow band. Naturally, this region (band) has to
be updated as the curve evolves. See, for instance, [11] for more details.

References
[1] G. Aubert and P. Kornprobst, Mathematical Problems in Image Processing, Applied
Mathematical Sciences 147, Springer, 2002.
[2] P.G. Ciarlet and J.L. Lions, Handbook of Numerical Analyss. Vol. 1. Finite Difference
Methods. Solution of Equations in Rn , North Holland, 1990.
[3] E. Godlewski and P.A. Raviart, Hyperbolic Systems of Conservation Laws, Vol. 3/4
of Mathématiques et Applications, Ellipses, 1991.
[4] J. Gomes and O. Faugeras, Reconciling distance functions and level sets, J. Vis.
Comm. Image Rep. 11 (1990), 919–940.
14 CHANG-OCK LEE

[5] R.J. LeVeque, Numerical Methods for Conservation Laws, Birkhäuser, Basel, 1992.
[6] K.W. Morton and D.F. Mayers, Numerical Solution of Partial Differential Equations:
An Introduction, Second Edition, Cambridge University Press, 2005.
[7] S. Osher and L.I. Rudin, Feature-oriented image enhancement using shock filters,
SIAM J. Numer. Anal. 27 (2000), 209–223.
[8] S. Osher and J. Sethian, Fronts propagating with curvature dependent speed: al-
gprithms based on the Hamilton-Jacobi formulation, J. Comput. Physics. 79 (1988),
12–49.
[9] D. Peng, B. Merriman, S. Osher, H. Zhao, and M. Kang, A PDE-based fast local level
set method, J. Comput. Physics. 155 (1999), 420–438.
[10] J.A. Sethian, A fast marching level set method for monotonically advancing fronts,
In Proceedings of the National Academy of Sciences. 93 (1996), 1592–1694.
[11] J.A. Sethian, Level Set Methods and Fast Marching Methods: Evolving Interfaces in
Computational Geometry, Fluid Mechanics, Computer Vision, and Material Sciences,
Cambridge Monogarphs on Applied and Computational Mathematics. Cambridge
Univeristy Press, 1999.
[12] J.W. Thomas, Numerical Partial Differential Equations: Finite Difference Methods,
Springer-Verlag, 1995.
[13] J.N. Tsitsiklis, Efficient algorithms for globally optimal trajectories, IEEE Tran. Auto.
Control. 40 (1995), 1528–1538.

Department Mathematical Sciences, KAIST, Daejeon, 305-701 Korea

E-mail address: [email protected]