Numerical Methods for the Solution of Partial Differential

Equations: A Review∗

Pedro González–Casanova
May 2005

The main purpose of this note is to provide a wide view of the different
numerical methods for the solution of partial differential equations. We
aim that this text can help the reader to be aware of some of actual main
trends in this area of knowledge. With in the text, we have included
several references to different detailed reviews related to each research
sub area of this field.

1. Introduction
Numerical methods for the solution of partial differential equations can
be broadly separated into two major groups regarding the Lagrangian
and Eulerian descriptions of continuous motion. In Lagrangian algo-
∗
Please cite as: “Pedro González–Casanova, Numerical Methods for the Solution of
Partial Differential Equations: A Review. Report, No 1-2005. Project: Pderbf. Insti-
tuto de Matemáticas, UNAM. http://paginas.matem.unam.mx/pderbf/index.php/”

1
Section 1: Introduction 2

rithms the nodes move with continuum, in Eulerian algorithms the

nodes stay in place while the the continuum moves through the station-
ary mesh or through Eulerian coordinate system. Moreover, these nu-
merical methods can be further classified in mesh-structurated and
mesh-unstrucrurated methods. Classical discretization techniques, like
finite differences, finite elements and finite volumes, which have been de-
veloped within both Eulerian and Lagrangian approaches, emerged as
mesh unstructured methods.
In recent years mesh-free and mesh-adaptive methods gained much
attention, not only in the engineering but also in the mathematics com-
munity. One of the reasons for this development is the fact that mesh-free
and mesh-adaptive discretizations are often better suited to cope with ge-
ometric changes of the domain of interest, e.g. free surfaces and large de-
formations, than the classical structured-mesh discretization techniques.
Both mesh structurated and unstructurated techniques have been devel-
oped within the frame of the mesh-adaptive approach, thus extending the
classical finite element, finite difference and finite volume techniques.
Numerical methods that are used to model partial differential equations
with steep solution regions often involves high computational cost if a
uniform mesh is used. The family of mesh-adaptive methods -also known
as moving mesh methods- adapts the mesh to the features of the computed
Section 2: Mesh Adaptive Methods 3

solution. The nodal density is high in regions of large variation and low
in regions where the solution variation is small.
On the other hand, mesh-free algorithms are truly knot base methods.
Mesh generation is still the most time consuming part of any mesh based
numerical simulation. Typically, more than 70 percent of the overall com-
puting time is spent by mesh generators. Since mesh-free discretization
techniques are based only on a set of independent points these costs of
mesh generation are eliminated.

2. Mesh Adaptive Methods

The diversity of ways that grids are generated has prevented effectively
the development of universal adaptive techniques. Since there are many
grid generation techniques, there are also many adaptive grid techniques.
A reasonable set of goals for the application of adaptation can be proposed
based on those stated in [24], by Y. Kallinderis in the preface to the recent
special issue [15], and in remarks by the editors of [13]:
1. The fundamental goal of mesh adaptation must be to reduce spatial
discretization error and solution grid dependence.
2. Adaptation should result in quantifiable solution accuracy improve-
ment.
3. Temporal accuracy and conservation should be preserved if required.
Section 2: Mesh Adaptive Methods 4

4. Additional error introduced by the adaptive algorithm should not

reduce benefits significantly.
5. Adaptation should be both automatic and efficient.
These goals/criteria provide initial measures against which adaptation
algorithms may be compared.

2.1. Structured and Unstructured Grid Methods

Structured grid methods take their name from the fact that the grid is laid
out in a regular repeating pattern called a block. Mesh generated by a
structured grid generator is typically all quad or hexahedral. Although
the element topology is fixed, the grid can be shaped to be body fitted
through stretching and twisting of the block. Algorithms employed gener-
ally involve complex iterative smoothing techniques that attempt to align
elements with boundaries or physical domains. Really good structured
grid generators utilize sophisticated elliptic equations to automatically
optimize the shape of the mesh for orthogonality and uniformity. Where
non-trivial boundaries are required, “block-structured” techniques can be
employed which allow the user to break the domain up into topological
blocks. Strictly speaking, a structured mesh can be recognized by all
interior nodes of the mesh having an equal number of adjacent elements.
Unstructured grid methods utilize an arbitrary collection of elements to
fill the domain. Because the arrangement of elements have no discernible
Section 2: Mesh Adaptive Methods 5

pattern, the mesh is called unstructured. These types of grids typically

utilize triangles in 2D and tetrahedral in 3D. As with structured grids,
the elements can be stretched and twisted to fit the domain. These meth-
ods have the ability to be automated to a large degree. The automatic
meshing algorithm typically involves meshing the boundary and then ei-
ther adding elements touching the boundary (advancing front) or adding
points in the interior and reconnecting the elements (Delaunay). Unstruc-
tured mesh generation relaxes the node valence requirement, allowing any
number of elements to meet at a single node.
While there is certainly some overlap between structured and unstruc-
tured mesh generation technologies, the main feature which distinguish
the two fields are the unique iterative smoothing algorithms employed by
structured grid generators.

2.2. Mesh Adaptive Strategies

Recent discussions of adaptation have used five categories [13]. Specif-
ically, the different techniques to adapt the Finite Element space (for
stationary problems), are
• h-refinement: we enrich the finite element space by (locally) refin-
ing the underlying spatial partition The insertion/deletion of mesh
nodes resulting in an overall increase/decrease of the number of cells,
Section 2: Mesh Adaptive Methods 6

• p-refinement is performed by increasing for fixed mesh the polyno-

mial degree of the ansatz space,
• h-p-refinement is a combination of the two last items. Adaptive
finite element methods that are capable of exploiting both local
polynomial -degree-variation (p-refinement) and local mesh subdi-
vision (h-refinement) offer greater flexibility and improved efficiency
than mesh refinement methods which only incorporate h-refinement
or p-refinement in isolation,
• r-refinement: one relocates the mesh points in order to get a better
resolution of the solution with fixed amount of unknowns. The
number of nodes remain constant but are relocated physically in
the domain while maintaining identity and data structure,
• m-refinement: one switches to a different equation (= physical model)
depending on the local behavior of the approximated solution. As
an example one may use linearized equations only if the nonlinear
terms of the physical model are negligible.

2.3. Eulerian and Lagrangian Adaptive Methods

In the Eulerian approach, an adaptive mesh technique (h-version, hp-
version) must follow the time-dependent features of the data or solution
by local refinement and coarsening of the mesh ( [2], [4], [14], [20],
Section 3: Mesh-free Methods 7

[32]). But time-dependent adaptive mesh refinement and coarsening is

not simple, especially for three-dimensional (3D) problems. It is quite
involved, programming is complicated, data structures are not easy to
handle, and the storage overhead is significant. Besides, good local and
global error estimators are necessary. Therefore, there exist only a few
unstructured adaptive programs which are able to handle 3D application-
oriented problems with time-dependent change of the geometry, the data,
or the solution.
The Lagrangian viewpoint allows the mesh itself to be moved (r-method)
( [1], [25], [26]). But an implementation is still cumbersome, since
the mesh may become tangled and twisted, elements may collapse, or
angles of some elements may degenerate over time due to the movement
of the nodes. The proper treatment of these problems is not an easy task,
especially in 3D applications.

3. Mesh-free Methods
Generally speaking, there are two different types of mesh-free approaches
the classical particle methods ( [29], [30], [28], [31]), and griddles
discretizations based on data fitting techniques ( [3], [7] ).
Section 3: Mesh-free Methods 8

3.1. Particle Methods

Traditional particle methods stem from physics applications like Boltz-
mann equations [12]. They are truly Lagrangian methods, i.e., they
are based upon a time-dependent formulation or conservation law. In a
particle method we use a discrete set of points to discretize the domain
of interest and the solution at a certain time. The PDE is transformed
into equations of motion for the discrete set of particles such that the
particles can be moved via these equations. After time discretization of
the equations of motion we obtain a certain particle distribution for ev-
ery time step. Therefore, we get an approximate solution to the PDE
via the definition of a density function for these particle distributions.
These methods are easy to implement. However, they exhibit in general
relatively poor convergence properties in weak norms.

3.2. Gridless Methods

The so-called gridless methods follow a different approach. Here, patches
or volumina are attached to each point whose union forms an open cov-
ering of the domain. Then, local shape functions are constructed with
the help of methods from data fitting. These shape functions are used in
a Galerkin or collocation discretization process to set up a linear system
of equations. Finally this system must be solved efficiently. In contrast
to particle methods, such gridless discretizations may also be applied to
Section 3: Mesh-free Methods 9

stationary and elliptic problems. According to the data fitting method

involved we can distinguish basically the following three approaches: Shep-
ard’s method [33], which has a consistency of first order only, the moving
least squares method (MLSM) ( [18], [19]), which generalizes Shepard’s
approach implicitly to the case of higher order shape functions, and the
partition of unity p-version method, which generalizes Shepard s ap-
proach explicitly to higher consistency orders. Meanwhile, different real-
izations of these approaches exist. First, there is the smoothed particle
hydrodynamics (SPH) technique of Lucy and Monaghan ( [10], [11],
[23], [27], [28] ), which resembles (up to area weighted scaling) Shepard
s method. Then, Duarte and Oden ( [7], [6]) used in their hp-cloud
approach the moving least squares (MLS) idea. Belytschko and cowork-
ers ( [26], [29]) apply similar techniques based on the MLS approach to
engineering problems. Furthermore, Dilts, [5] used the MLS technique
to extend the SPH method to the so-called MLS particle hydrodynam-
ics (MLSPH) method. Babuska and Melenk [28] proposed the so-called
partition of unity method (PUM), which mainly has been applied to uni-
form point distributions up to now. Liu, Jun, and Zhang [22] proposed
variants of the SPH method based on the idea of reproducing kernels of
higher order and wavelets. There also exist generalizations of the finite
difference approach to the gridless setting [21]. Furthermore, Kansa
( [16], [17]), Franke and Schaback ( [8], [9]), and Wendland [35] used the
radial basis approach from approximation theory to construct meshless
Section 3: Mesh-free Methods 10

methods for the discretization of PDEs. The mass-packet method of

Yserantant ( [36], [37]) is somewhat different from the classical particle
methods. Here, the particles are not considered in the sense of statistical
mechanics but they are understood as comparatively big mass-packets,
and the conservation of mass is automatically guaranteed by this ansatz.
For an overview on meshless methods see ( [34]) and the references therein.
All these data fitting approaches do not depend (at least to a great ex-
tent) upon a mesh or any fixed relation between gridpoints (particles).
However, the realization and implementation of such a method is not so
simple in general: there are often problems with stability and consistency.
Furthermore, in a Galerkin method, the discretization of the differential
operator, i.e., the integration of the stiffness matrix entries, is in general
quite involved in comparison with the conventional grid-based approach.
Another challenging task is the discrete formulation of Dirichlet boundary
conditions, since the constructed shape functions are in general noninter-
polatory. Nevertheless, the different variants of gridless methods are in-
teresting from both the practical and the theoretical point of view. These
methods, which are up to now merely in an experimental premature state,
possess some potential and might have an interesting future.

References
[1] M. J. Baines, Moving finite elements, Monographs on Numerical
Section 3: Mesh-free Methods 11

Analysis, The Clarendon Press Oxford University Press, New York,

1994. 7
[2] R. E. Bank, PLTMG: a software package for solving elliptic partial
differential equations, vol. 15 of Frontiers in Applied Mathematics,
Society for Industrial and Applied Mathematics (SIAM), Philadel-
phia, PA, 1994. 6
[3] T. Belytschko, Y. Krongauz, D. Organ, M. Fleming, and
P. Krysl, Meshless methods: An overview and recent developments,
Comput. Methods Appl. Mech. Engrg, 139 (1996), pp. 3–47. 7
[4] L. Demkowicz, J. T. Oden, W. Rachowicz, and O. Hardy,
Toward a universal h-p adaptive finite element strategy. I. Con-
strained approximation and data structure, Comput. Methods Appl.
Mech. Engrg., 77 (1989), pp. 79–112. 6
[5] G. A. Dilts, Moving-least-squares-particle hydrodynamics. I. Con-
sistency and stability, Internat. J. Numer. Methods Engrg., 44 (1999),
pp. 1115–1155. 9
[6] C. A. Duarte and J. T. Oden, H-p clouds—an h-p meshless
method, Numer. Methods Partial Differential Equations, 12 (1996),
pp. 673–705. 9
[7] C. A. M. Duarte, A review pf some meshless methods to solve
partial differential equations, Tech. Report 95-06, TICAM, University
of Texas, East Lansing, Michigan, 1995. 7, 9
Section 3: Mesh-free Methods 12

[8] C. Franke and R. Schaback, Convergence order estimates of

meshless collocation methods using radial basis functions, Adv. Com-
put. Math., 8 (1998), pp. 381–399. 9
[9] C. Franke and R. Schaback, Solving partial differential equa-
tions by collocation using radial basis functions, Appl. Math. Com-
put., 93 (1998), pp. 73–82. 9
[10] R. Gingold and J. Monaghan, Smoothed particle hydrody-
namics: theory and application to non-spherical stars, Mon. Not.
R. astr. Soc., 181 (1977), pp. 375–389. 9
[11] R. A. Gingold and J. J. Monaghan, Kernel estimates as a basis
for general particle methods in hydrodynamics, J. Comput. Phys., 46
(1982), pp. 429–453. 9
[12] R. T. Glassey, The Cauchy problem in kinetic theory, Society
for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,
1996. 8
[13] N. W. J.F. Thompson, B. Soni, Handbook of grid generation,
CRC Press, Boca Raton, FL, 1999. 3, 5
[14] C. Johnson, Numerical solution of partial differential equations by
the finite element method, Studentlitteratur, Lund, 1987. 6
[15] E. Kallinderis, Y., Adaptive methods for compressible cfd, Com-
puter Methods in Applied Science and Engineering, 189 (2000). 3
Section 3: Mesh-free Methods 13

[16] E. J. Kansa, Multiquadrics—a scattered data approximation scheme

with applications to computational fluid-dynamics. I. Surface approx-
imations and partial derivative estimates, Comput. Math. Appl., 19
(1990), pp. 127–145. 9
[17] , Multiquadrics—a scattered data approximation scheme with ap-
plications to computational fluid-dynamics. II. Solutions to parabolic,
hyperbolic and elliptic partial differential equations, Comput. Math.
Appl., 19 (1990), pp. 147–161. 9
[18] P. Lancaster, Moving weighted least-squares methods, in Polyno-
mial and spline approximation (Proc. NATO Adv. Study Inst., Univ.
Calgary, Calgary, Alta., 1978), vol. 49 of NATO Adv. Study Inst.
Ser., Ser. C: Math. Phys. Sci., Reidel, Dordrecht, 1979, pp. 103–120.
9
[19] P. Lancaster and K. Salkauskas, Surfaces generated by moving
least squares methods, Math. Comp., 37 (1981), pp. 141–158. 9
[20] J. Lang, Adaptive FEM for reaction-diffusion equations, Appl. Nu-
mer. Math., 26 (1998), pp. 105–116. 6
[21] T. Liszka and J. Orkisz, The finite difference method at arbitrary
irregular grids and its application in applied mechanics, Comput. &
Structures, 11 (1980), pp. 83–95. 9
[22] W. Liu, S. Jun, and Y. Zhang, Reproducing kernel particle meth-
ods, Int. J. Numer. Methods Eng., 20 (1995). 9
Section 3: Mesh-free Methods 14

[23] L. B. Lucy, A Numerical Approach to Testing the Fission Hypoth-

esis, Astron. J, 82 (1977), pp. 1013–1924. 9
[24] D. McRae, r-refinement grid adptation and issues, Computer Meth-
ods in Applied Science and Engineering, 189 (2000), pp. 1261–1282.
3
[25] K. Miller, Moving finite elements. II, SIAM J. Numer. Anal., 18
(1981), pp. 1033–1057. 7
[26] K. Miller and R. N. Miller, Moving finite elements. I, SIAM
J. Numer. Anal., 18 (1981), pp. 1019–1032. 7, 9
[27] J. Monaghan, An Introduction to SPH, Comp. Phys. Comm., 48
(1988), pp. 89–96. 9
[28] J. J. Monaghan, Why particle methods work, SIAM J. Sci. Statist.
Comput., 3 (1982), pp. 422–433. 7, 9
[29] K. Nanbu, Direct simulation scheme derived from the boltzmann
equation, J. Phys. Soc. Japan, 49 (1980). 7, 9
[30] , Theoretical basis on the direct simulation monte carlo method,
Rare Gas Dynamics, 1 (1986). 7
[31] H. Neunzert and J. Struckmeier, Boltzmann simulation by par-
ticle methods, 1997. 7
[32] J. T. Oden, T. Strouboulis, and P. Devloo, Adaptive finite
element methods for the analysis of inviscid compressible flow. I. Fast
Section 3: Mesh-free Methods 15

refinement/unrefinement and moving mesh methods for unstructured

meshes, Comput. Methods Appl. Mech. Engrg., 59 (1986), pp. 327–
362. 7
[33] D. S. Shepard, A two-dimensional interpolation function for irreg-
ularly spaced data, in Proceedings of the 1968 ACM National Con-
ference, New York, 1968, pp. 517–524. 9
[34] B. T., K. Y., O. D., F. M., and K. P, Meshless methods : An
overview and recent developments, Comp. Meth. in App. Mech. and
Eng., special issue on Meshless Methods, 39 (1996), pp. 3–47. 10
[35] H. Wendland, Meshless Galerkin methods using radial basis func-
tions, Math. Comp., 68 (1999), pp. 1521–1531. 9
[36] H. Yserentant, A new class of particle methods, Numer. Math.,
76 (1997), pp. 87–109. 10
[37] , A particle model of compressible fluids, Numer. Math., 76
(1997), pp. 111–142. 10