19

FEM Convergence
Requirements

191

Chapter 19: FEM CONVERGENCE REQUIREMENTS

TABLE OF CONTENTS
Page

19.1.
19.2.
19.3.

19.4.

19.
19.
19.

Overview
The Variational Index
Consistency Requirements
19.3.1. Completeness
. . . . . . .
19.3.2. Compatibility . . . . . . . .
19.3.3. Matching and Non-Matching Meshes
Stability
19.4.1. Rank Sufficiency . . . . . . .
19.4.2. Jacobian Positiveness . . . . .
Notes and Bibliography
. . . . . . . . . . . . .
References . . . . . . . . . . . .
Exercises . . . . . . . . . . . . .

192

. . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
.
. .
.
. .
.

. .
.
. .
.
. .

.
. .
.
. .
.

. .
.
. .
.
. .

.
. .
.
. .
.

. .
. .
. .
. .
. .

193
193
194
194
194
196
197
198
199
1911
1911
1912

19.2 THE VARIATIONAL INDEX

19.1. Overview
Chapters 11 through 18 have discussed, in piecemeal fashion, requirements for shape functions
of isoparametric elements. These are motivated by convergence: as the mesh is refined, the FEM
solution should approach the analytical solution of the mathematical model.1 This attribute is
obviously necessary to instill confidence in FEM results from the standpoint of mathematics.
This Chapter provides unified information on convergence requirements. These requirements can
be grouped into three:
Completeness. The elements must have enough approximation power to capture the analytical
solution in the limit of a mesh refinement process. This intuitive statement is rendered more precise
below.
Compatibility. The shape functions should provide displacement continuity between elements.
Physically these insure that no material gaps appear as the elements deform. As the mesh is refined,
such gaps would multiply and may absorb or release spurious energy.
Stability. The system of finite element equations must satify certain well posedness conditions that
preclude nonphysical zero-energy modes in elements, as well as the absence of excessive element
distortion.
Completeness and compatibility are two aspects of the so-called consistency condition between
the discrete and mathematical models. A finite element model that passes both completeness and
continuity requirements is called consistent. This is the FEM analog of the famous Lax-Wendroff
theorem,2 which says that consistency and stability imply convergence.
Remark 19.1. A deeper mathematical analysis done in more advanced courses shows that completeness is

necessary for convergence whereas failure of the other requirements does not necessarily precludes it. There
are, for example, FEM models in common use that do not satisfy compatibility. Furthermore, numerically
unstable models may be used (with caution) in situations where that property is advantageous, as in the
modeling of local singularities. Nonetheless, the satisfaction of the three criteria guarantees convergence and
may therefore be regarded as a safe choice for the beginner user.

19.2. The Variational Index

For the mathematical statement of the completeness and continuity conditions, the variational index
alluded to in previous sections plays a fundamental role.
The FEM is based on the direct discretization of an energy functional [u], where u (displacements
for the elements considered in this book) is the primary variable, or (equivalently) the function to
be varied. Let m be the highest spatial derivative order of u that appears in . This m is called the
variational index.
1

Of course FEM convergence does not guarantee the correctness of the mathematical model in capturing the physics. As
discussed in Chapter 1, model verification against experiments is a different and far more difficult problem.

Proven originally for classical finite difference discretizations in fluid mechanics. More precisely, it states that a numerical
scheme for the scalar conservation law, du/dt + d f /d x = 0 converges to a unique (weak) solution, if it is consistent,
stable and conservative. There is no equivalent theorem for systems of conservation laws.

193

Chapter 19: FEM CONVERGENCE REQUIREMENTS

Example 19.1. In the bar problem discussed in Chapter 11,

[u] =

1
2

u  E Au  qu

d x.

(19.1)

The highest derivative of the displacement u(x) is u  = du/d x, which is first order in the space coordinate x.
Consequently m = 1. This is also the case on the plane stress problem studied in Chapter 14, because the
strains are expressed in terms of first order derivatives of the displacements.
Example 19.2. In the plane beam problem discussed in Chapter 12,

1

v  E I v  qv d x.
2

[v] =

(19.2)

The highest derivative of the transverse displacement is the curvature = v  = d 2 v/d x 2 , which is of second
order in the space coordinate x. Consequently m = 2.

19.3. Consistency Requirements

Using the foregoing definition of variational index, we can proceed to state the two key requirements
for finite element shape functions.
19.3.1. Completeness

The element shape functions must represent exactly all polynomial terms
of order m in the Cartesian coordinates. A set of shape functions that
satisfies this condition is called m-complete.
Note that this requirement applies at the element level and involves all shape functions of the
element.
Example 19.3. Suppose a displacement-based element is for a plane stress problem, in which m = 1. Then
1-completeness requires that the linear displacement field

u x = 0 + 1 x + 2 y,

u y = 0 + 1 x + 2 y

(19.3)

be exactly represented for any value of the coefficients. This is done by evaluating (19.3) at the nodes to
form a displacement vector ue and then checking that u = Ne ue recovers exactly (19.3). Section 16.6 presents
the details of this calculation for an arbitrary isoparametric plane stress element. That analysis shows that
completeness is satisfied if the sum of the shape functions is unity and the element is compatible.
Example 19.4. For the plane beam problem, in which m = 2, the quadratic transverse displacement

v = 0 + 1 x + 2 x 2

(19.4)

must be exactly represented over the element. This is easily verified in for the 2-node beam element developed
in Chapter 13, because the assumed transverse displacement is a complete cubic polynomial in x. A complete
cubic contains the quadratic (19.4) as special case.

194

19.3

(b)

(a)

CONSISTENCY REQUIREMENTS

(c)

2-node bars
Figure 19.1. An element patch is the set of all elements attached to a patch node, labeled i.
(a) illustrates a patch of triangles; (b) a mixture of triangles and quadrilaterals; (c) a mixture of
triangles, quadrilaterals, and bars.

19.3.2. Compatibility
To state this requirement succintly, it is convenient to introduce the concept of element patch, or
simply patch. This is the set of all elements attached to a given node, called the patch node. The
definition is illustrated in Figure 19.1, which shows three different kind of patches attached to patch
node i in a plane stress problem. The patch of Figure 19.1(a) contains only one type of element:
3-node linear triangles. The patch of Figure 19.1(b) mixes two plane stress element types: 3-node
linear triangles and 4-node bilinear quadrilaterals. The patch of Figure 19.1(c) combines three
element types: 3-node linear triangles, 4-node bilinear quadrilaterals, and 2-node bars.
We define a finite element patch trial function as the union of shape functions activated by setting
a degree of freedom at the patch node to unity, while all other freedoms are zero.
A patch trial function propagates only over the patch, and is zero beyond it. This property follows
from the local-support requirement stated in 18.1: a shape function for node i should vanish on
all sides or faces that do not include i.
With the help of these definitions we can enunciate the compatibility requirement as follows.
Patch trial functions must be C (m1) continuous between interconnected
elements, and C m piecewise differentiable inside each element.

If the variational index is m = 1, the patch trial functions must be C 0 continuous between elements,
and C 1 inside elements.
A set of shape functions that satisfies the first requirement is called conforming. A conforming
expansion that satisfies the second requirement is said to be of finite energy. Note that this condition
applies at two levels: individual element, and element patch. An element endowed with conforming
shape functions is said to be conforming. A conforming element that satisfies the finite energy
requirement is said to be compatible.3
3

The FEM literature is a bit fuzzy as regards these terms. It seems better to leave the qualifier conforming to denote
interelement compatibility; informally an element that gets along with its neighbors. The qualifier compatible is
used in the stricter sense of conforming while possessing sufficient internal smoothness.

195

Chapter 19: FEM CONVERGENCE REQUIREMENTS

(a)

(b)

(c)

3-node bars

(e)
(f)

(d)

Figure 19.2. Examples of 2D non-matching meshes. Interelement boundaries that fail matching
conditions are shown offset for visualization convenience. In (a,b,c) some nodes do not match. In
(d,e,f) nodes and DOFs match but some sides do not, leading to violations of C 0 continuity.

Figures 19.1(b,c) illustrates the fact that one needs to check the possible connection of matching
elements of different types and possibly different dimensionality.
19.3.3. Matching and Non-Matching Meshes
As stated, compatibility refers to the complete finite element mesh because mesh trial functions are
a combination of patch trial functions, which in turn are the union of element shape functions. This
generality poses some logistical difficulties because the condition is necessarily mesh dependent.
Compatibility can be checked at the element level by restricting attention to matching meshes. A
matching mesh is one in which adjacent elements share sides, nodes and degrees of freedom, as in
the patches shown in Figure 19.1.
For a matching mesh it is sufficient to restrict consideration first to a pair of adjacent elements, and
then to the side shared by these elements. Suppose that the variation of a shape function along
that side is controlled by k nodal values. Then a polynomial variation of order up to k 1 in the
natural coordinate(s) can be specified uniquely over the side. This is sufficient to verify interelement
compatibility for m = 1, implying C 0 continuity, if the shape functions are polynomials.
This simplified criterion is the one used in previous Chapters. Specific 2D examples were given in
Chapters 15 through 18.
Remark 19.2. If the variational index is m = 2 and the problem is multidimensional, as in the case of

plates and shells, the check is far more involved and trickier because continuity of normal derivatives along
a side is involved. This practically important scenario is examined in advanced FEM treatments. The case of
non-polynomial shape functions is, on the other hand, of little practical interest.

196

19.4
A mesh that does not satisfy the matching
criteria stated above is called a nonmatching
mesh. Several two-dimensional examples are
shown in Figure 19.2. As can be seen there
is a wide range of possibilities: nonmatching
nodes, matching nodes but different element
types, etc.
Figure 19.3 depicts a threedimensional example, in which case even more
variety can be expected.
Nonmatching meshes are the rule rather than
the exception in contact and impact problems
(which, being geometrically nonlinear, are
outside the scope of this book). See Figure
19.4 illustrates what happens in a problem of
slipping contact.

Initial shape

STABILITY

Mesh of
tetrahedra

Common surface

Mesh of
bricks

Figure 19.3. Example of a 3D non-matching

mesh. Top portion discretized with tetrahedra,
lower portion with bricks. Nodes and boundaryquad edges and DOFs match, but element types are
different, leading to violation of C 0 continuity.

Deformed shape

;;;;;;;;;
;;;;;;;;;
;;;;;;;;;
;;;;;;;;;
Figure 19.4. In contact and impact problems, matching meshes are the exception rather than the
rule. Even if the meshes match at initial contact, slipping may produce a nonmatching mesh in
the deformed configuration, as illustrated in the figure.

In multiphysics simulations nonmatching meshes are common, since they are often prepared separately for the different physical components, as illustrated in Figure 19.5.

Fluid

Structure

Figure 19.5. Nonmatching meshes are common in multiphysics problems, as in this example of
fluid-structure interaction (FSI). Two-dimensional model to simulate flow around a thin plate. If
the meshes are independenly prepared node locations will not generally match.

197

Chapter 19: FEM CONVERGENCE REQUIREMENTS

19.4. Stability
Stability may be informally characterized as ensuring that the finite element model enjoys the same
solution uniqueness properties of the analytical solution of the mathematical model. For example, if
the only motions that produce zero internal energy in the mathematical model are rigid body motions,
the finite element model must inherit that property. Since FEM can handle arbitrary assemblies of
elements, including individual elements, this property is required to hold at the element level.
In the present outline we are concerned with stability at the element level. Stability is not a property
of shape functions per se but of the implementation of the element as well as its geometrical
definition. It involves two subordinate requirements: rank sufficiency, and Jacobian positiveness.
Of these, rank sufficiency is the most important one.
19.4.1. Rank Sufficiency
The element stiffness matrix must not possess any zero-energy kinematic mode other than rigid
body modes.
This can be mathematically expressed as follows. Let n F be the number of element degrees of
freedom, and n R be the number of independent rigid body modes. Let r denote the rank of Ke . The
element is called rank sufficient if r = n F n R and rank deficient if r < n F n R . In the latter
case, the rank deficiency is defined by
d = (n F n R ) r

(19.5)

If an isoparametric element is numerically integrated, let n G be the number of Gauss points, while
n E denotes the order of the stress-strain matrix E. Two additional assumptions are made:
(i)

The element shape functions satisfy completeness in the sense that the rigid body modes are
exactly captured by them.

(ii) Matrix E is of full rank.

Then each Gauss point adds n E to the rank of Ke , up to a maximum of n F n R . Hence the rank
of Ke will be
(19.6)
r = min(n F n R , n E n G )
To attain rank sufficiency, n E n G must equal or exceed n F n R :
n E nG n F n R

(19.7)

from which the appropriate Gauss integration rule can be selected.

In the plane stress problem, n E = 3 because E is a 3 3 matrix of elastic moduli; see equation
(14.5)2 . Also n R = 3. Consequently r = min(n F 3, 3n G ) and 3n G n F 3.
Remark 19.3. The fact that each Gauss point adds n E n G to the rank can be proven considering the following
property. Let B be a n E n F rectangular real matrix with rank r B n E , and E an n E n E positive-definite (p.d.)
symmetric matrix. Then the rank of BT E B is r B . Proof: let u = 0 be a non-null n F -vector. If BT E B u = 0
then 0 = uT BT E B u = ||E1/2 B u||. Therefore B u = 0. Identify now B and E with the strain-displacement
and stress-strain (constitutive) matrix, respectively. In the plane stress case n E = 3, n F = 2n > 3 is the
number of element freedoms. Thus B has rank 3 and a fortiori BT E B must also have rank 3 since E is p.d.
At each Gauss point i a contribution of wi BT E B, which has rank 3 if wi > 0, is added to Ke . By a theorem
of linear algebra, the rank of Ke increases by 3 until it reaches n F n R .

198

19.4

STABILITY

Table 19.1 Rank-sufficient Gauss Rules for Some Plane Stress Elements
Element

nF

nF 3

Min n G

Recommended rule

3-node triangle
6-node triangle
10-node triangle
4-node quadrilateral
8-node quadrilateral
9-node quadrilateral
16-node quadrilateral

3
6
10
4
8
9
16

6
12
20
8
16
18
32

3
9
17
5
13
15
29

1
3
6
2
5
5
10

centroid
3-point rules
6-point rule
2x2
3x3
3x3
4x4

These triangle integration rules are introduced in 24.2.

Figure 19.6. Effect of displacing node 4 of the four-node bilinear quadrilateral shown on the
leftmost picture, to the right.

Example 19.5. Consider a plane stress 6-node quadratic triangle. Then n F = 2 6 = 12. To attain the proper
rank of 12 n R = 12 3 = 9, n G 3. A 3-point Gauss rule, such as the midpoint rule defined in 24.2,
makes the element rank sufficient.
Example 19.6. Consider a plane stress 9-node biquadratic quadrilateral. Then n F = 2 9 = 18. To attain

the proper rank of 18 n R = 18 3 = 15, n G 5. The 2 2 product Gauss rule is insufficient because
n G = 4. Hence a 3 3 rule, which yields n G = 9, is required to attain rank sufficiency.

Table 19.1 collects rank-sufficient Gauss integration rules for some widely used plane stress elements
with n nodes and n F = 2n freedoms.
19.4.2. Jacobian Positiveness
The geometry of the element should be such that the determinant J = det J of the Jacobian matrix
defined4 in 17.2, is positive everywhere. As illustrated in Equation (17.20), J characterizes the
local metric of the element natural coordinates.
4

This definition applies to quadrilateral elements. The Jacobian determinant of an arbitrary triangular element is defined
in 24.2.

199

Chapter 19: FEM CONVERGENCE REQUIREMENTS

3
7

7
4

8
2

6
9

8
5

5 2

5=2

Figure 19.7. Effect of moving midpoint 5 of a 9-node biquadratic quadrilateral

tangentially toward corner 2.

For a three-node triangle J is constant and in fact equal to 2A. The requirement J > 0 is equivalent
to saying that corner nodes must be positioned and numbered so that a positive area A > 0 results.
This is called a convexity condition. It is easily checked by a finite element program.
But for 2D elements with more than 3 nodes distortions may render portions of the element metric
negative. This is illustrated in Figure 19.6 for a 4-node quadrilateral in which node 4 is gradually
moved to the right. The quadrilateral gradually morphs from a convex figure into a nonconvex
one. The center figure is a triangle; note that the metric near node 4 is badly distorted (in fact
J = 0 there) rendering the element unacceptable. This clearly contradicts the erroneous advice of
some FE books, which state that quadrilaterals can be reduced to triangles as special cases, thereby
rendering triangular elements unnecessary.
For higher order elements proper location of corner nodes is not enough. The non-corner nodes
(midside, interior, etc.) must be placed sufficiently close to their natural locations (midpoints,
centroids, etc.) to avoid violent local distortions. The effect of midpoint motions in quadratic
elements is illustrated in Figures 19.7 and 19.8.
Figure 19.7 depicts the effect of moving midside node 5 tangentially in a 9-node quadrilateral
element while keeping all other 8 nodes fixed. When the location of 5 reaches the quarter-point of
side 1-2, the metric at corner 2 becomes singular in the sense that J = 0 there. Although this is
disastrous in ordinary FE work, it has applications in the construction of special crack elements
for linear fracture mechanics.
1910

19.

6
5

References

4
4

1
2 1

2 1

4
4

4
Figure 19.8. Effect of displacing midpoints 4, 5 and 6 of an equilateral 6-node triangle along
the midpoint normals. Motion is inwards in first two top frames, outwards in the last four. In
the lower leftmost picture nodes 1 through 6 lie on a circle.

Displacing midside nodes normally to the sides is comparatively more forgiving, as illustrated in
Figure 19.8. This depicts a 6-node equilateral triangle in which midside nodes 4, 5 and 6 are moved
inwards and outwards along the normals to the midpoint location. As shown in the lower left
picture, the element may be even morphed into a parabolic circle (meaning that nodes 1 through
6 lie on a circle) without the metric breaking down.
Notes and Bibliography
The literature on the mathematics of finite element methods has grown exponentially since the monograph of
Strang and Fix [698]. This is very readable but out of print. A more up-to-date exposition is the textbook by
Szabo and Babuska [714]. The subjects collected in this Chapter tend to be dispersed in recent monographs
and obscured by overuse of functional analysis.
References
Referenced items have been moved to Appendix R.

1911

Chapter 19: FEM CONVERGENCE REQUIREMENTS

Homework Exercises for Chapter 19

FEM Convergence Requirements
EXERCISE 19.1 [D:20] Explain why the two-dimensional meshes pictured in Figure 19.2(d,e,f) fail interelement compatibility although nodes and DOFs match.
EXERCISE 19.2 [A:20] The isoparametric definition of the straight 3-node bar element in its local system x

is

 
1
x
v

1
x1
u 1

1
x2
u 2

1
x3
u 3



N1e ( )
N2e ( )
N3e ( )

(E19.1)

Here is the isoparametric coordinate that takes the values 1, 1 and 0 at nodes 1, 2 and 3, respectively, while
N1e , N2e and N3e are the shape functions found in Exercise 16.3 and listed in (E16.2).
For simplicity, take x1 = 0, x2 = L, x3 = 12 L + L. Here L is the bar length and a parameter that
characterizes how far node 3 is away from the midpoint location x = 12 L. Show that the minimum s
(minimal in absolute value sense) for which J = d x/d

vanishes at a point in the element are 1/4 (the

quarter-points). Interpret this result as a singularity by showing that the axial strain becomes infinite at a an
end point. (This result has application in fracture mechanics modeling.)
EXERCISE 19.3 [A:15] Consider one dimensional bar-like elements with n nodes and 1 degree of freedom
per node so n F = n. The correct number of rigid body modes is 1. Each Gauss integration point adds 1 to
the rank; that is N E = 1. By applying (19.7), find the minimal rank-preserving Gauss integration rules with
p points in the longitudinal direction if the number of node points is n = 2, 3 or 4.
EXERCISE 19.4 [A:20] Consider three dimensional solid brick elements with n nodes and 3 degrees of
freedom per node so n F = 3n. The correct number of rigid body modes is 6. Each Gauss integration point
adds 6 to the rank; that is, N E = 6. By applying (19.7), find the minimal rank-preserving Gauss integration
rules with p points in each direction (that is, 111, 222, etc) if the number of node points is n = 8, 20,
27, or 64. Partial answer: for n = 27 the minimal rank preserving rule is 3 3 3.
EXERCISE 19.5 [A/C:35] (Requires use of a CAS help to be tractable). Repeat Exercise 19.2 for a 9-node

plane stress element. The element is initially a perfect square, nodes 5,6,7,8 are at the midpoint of the sides
12, 23, 34 and 41, respectively, and 9 at the center of the square. Displace 5 tangentially towards 2 until
the Jacobian determinant at 2 vanishes. This result is important in the construction of singular elements for
fracture mechanics.
EXERCISE 19.6 [A/C:35] Repeat Exercise 19.5 but moving node 5 along the normal to the side. Discuss
the range of motion for which det J > 0 within the element.
EXERCISE 19.7 [A:20] Discuss whether the deVeubeke triangle presented in Chapter 15 satisfies completeness and interelement-compatbility requirements.

1912