Computational Methods in Applied Mathematics

Vol. 5, No. 10(2011) 1 — 15

c Institute of Mathematics

ENERGY PRINCIPLES AND FINITE ELEMENT METHODS

FOR PURE TRACTION LINEAR ELASTICITY
1

PAVEL BOCHEV
Numerical Analysis and Applications, Sandia National Laboratories
P.O. Box 5800, MS 1320, Albuquerque, NM 87185, USA
E-mail: [email protected]

RICHARD LEHOUCQ
Multiphysics Simulation Technologies, Sandia National Laboratories
P.O. Box 5800, MS 1320, Albuquerque, NM 87185, USA
E-mail: [email protected]

Abstract. A conforming finite element discretization of the pure traction elasticity boundary value
problem results in a singular linear system of equations. The singularity of the linear system is removed
through various approaches. In this report, we consider an alternative approach in which discrete finite
element formulations are derived directly from the principle of minimum potential energy. This point
of view turns out to be particularly well suited to handle the rigid body modes, which are the source of
the singularity in the finite element linear system. Our approach allows us to formulate a regularized
potential energy principle and show that the associated weak problem is coercive in H 1 (Ω). This
guarantees nonsingular algebraic problems, enables simplified solution algorithms and leads to more
efficient and robust preconditioners for the iterative solution linear equations.
2000 Mathematics Subject Classification: 65N30; 65N12.
Key words: finite elements, elasticity, pure traction, least-squares regularization, quadratic pro-
gramming..

1. Introduction
We consider an isotropic elastic body occupying a bounded open region Ω in R I n for n = 2, 3. The surface Γ
of the body is assumed to be Lipschitz continuous and consisting of two disjoint pieces denoted by ΓD and
ΓT , respectively. For small deformations the static behavior of this elastic body is governed by the linear
equilibrium equation [7, pp.200-206]

−∇ · (σ(u)) = f in Ω (1.1)
u = g on ΓD (1.2)
σ(u) · n = t on ΓT . (1.3)

In (1.1)–(1.3) u is the displacement,

1
ε(u) = (∇u + (∇u)T ), (1.4)
2
is the infinitesimal strain, and
σ(u) = 2µε(u) + λtr(ε(u))I (1.5)
is the stress. The vector functions g and t prescribe the surface displacement and the surface traction,
respectively, and the scalar functions λ and µ are the Lame moduli. If the material is homogeneous λ and µ
are constants. When ΓD = ∅, or ΓT = ∅, the problem is called a pure displacement or pure traction problem,
respectively.
1 Sandia National Laboratories is a multi-program laboratory operated by Sandia Corporation, a wholly owned subsidiary of

Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under contract
DE-AC04-94AL85000.
2 P.B. Bochev & R.B. Lehoucq

When ΓD is not empty the Partial Differential Equations (PDEs) (1.1)–(1.3) have a unique solution. In
contrast, the pure traction problem has a nontrivial nullspace consisting of all infinitesimal rigid displace-
ments. The dimension of this nullspace is 3 in R I 2 and 6 in R
I 3 . As a result, the pure traction problem is
solvable if and only if the body force f and the surface traction t satisfy the compatibility relations
Z Z
f+ t=0 (1.6)
Ω Γ
Z Z
r × f + r × t = 0, (1.7)
Ω Γ

I 3.
where r = [ρ1 , ρ2 , ρ3 ]T is position vector in R
Equations (1.6)–(1.7) express the equilibrium of all external forces applied to an elastic body. For the
remainder of our report, we consider pure traction problems so that ΓD = ∅ and t = 0.
A standard Galerkin procedure applied directly to the pure traction boundary value problem leads
to a linear system with a non-trivial nullspace. To ensure consistency of this singular linear system a
discrete version of (1.6)-(1.7) must hold. However, in finite precision arithmetic discrete consistency may
become problematic. For instance, if the discrete load vector is computed using numerical quadrature, then
quadrature error introduces components that do not discretely satisfy (1.6)-(1.7).
In this report we focus on the finite element solution of the pure traction boundary value problem.
The motivation for this work is two-fold. The first is to understand carefully the relationship between the
variational form and the ensuing algebraic equations. The second is that this information is important for
the construction of preconditioners and/or when the linear system is solved with a preconditioned iterative
method. For example, the dual Schur domain-decomposition and algebraic multigrid preconditioners [5, 10, 9]
involve floating structures containing rigid body motions. Moreover, when an iterative method such as the
conjugate gradient algorithm is used, knowledge of the matrix is only assumed through a matrix vector
product.
We are not aware of any other study that carefully links variational methods and the resulting algebraic
equations for linear elasticity for use with preconditioned iterative methods. For example, the recent papers
[6, 8] instead consider and propose techniques for computing a basis for the rigid body modes via algebraic
and geometric techniques. The connection with an energy minimization principle is never discussed.
Our approach utilizes the existing connection between the partial differential equation (1.1)–(1.3) and
unconstrained minimization of a quadratic functional expressing the potential energy of an elastic body. This
minimization problem, known as the minimum potential energy principle, is the starting point of our finite
element development. We use it to derive three other minimization problems that posses unique minimizers.
The first two utilize non-standard spaces of kinematically admissible states but retain the original potential
energy functional. The third energy principle uses the original admissible space but augments the potential
energy functional by terms characterizing the rigid body deformations.
When restricted to a finite element subspace each one of these problems gives rise to a quadratic program
whose solution represents the finite element approximant. Thus, the principal difference between a conven-
tional Galerkin scheme and our approach is that the former treats finite element approximants as solutions
of linear systems, while here they are considered as minimizers of a potential energy functional.
The reasons to place the minimization problem, rather than the PDE, at the core of our approach are not
merely pedagogical. The minimum potential energy principle is the primal mathematical model describing
admissible states of an elastic body. The PDE (1.1)–(1.3) gives the strong form of the necessary optimality
condition for this principle. Consequently, our approach emphasizes finite element methods for elasticity as
an instance of a Rayleigh-Ritz principle. Characterization of finite element approximants as solutions of a
minimization problem offers the additional flexibility of utilizing ideas and solution methods from numerical
optimization. Of course, our approach does not rule out conventional solution techniques applied to linear
systems representing the necessary condition of the quadratic functional. Third, for conforming finite element
subspaces, the coefficient matrix is guaranteed to be symmetric and positive definite because of the coercivity
of the corresponding bilinear form. This precludes the need to maintain a discrete form of the compatibility
relations and allows for effortless convergence of iterative methods such as conjugate gradients.
In this report we use the standard notation H s (Ω) for a Sobolev space of order s with norm and inner
product given by k · Rks and (·, ·)s , respectively. Semi-norms on H s (Ω) will be denoted by | · |k , 0 ≤ k ≤ s.
For example, |u|1 = Ω |∇u|2 dx. For s = 0 we write L2 (Ω) instead of H 0 (Ω) and denote the resulting inner
product by (·, ·).
Because our study also makes use of matrix theory, we introduce some useful notation. Let {ei }N i=1 denote
the canonical basis of R I N ; IN be the identity matrix of order N ; I`N be IN with the `-column removed. The
 CMAM Energy principles and finite element methods for pure traction linear elasticity 3

ordering of the eigenvalues of the standard stiffness matrix A for (1.1) is 0 = λ1 = · · · = λ6 ≤ λ7 ≤ · · · ≤ λN

in three-dimensions and 0 = λ1 = · · · = λ3 ≤ λ4 ≤ · · · ≤ λN in two-dimensions.
We call attention to our specific use of bold font for tensors, vectors and the associated Sobolev spaces
(denoted by H s (Ω)). Elements of matrices and vectors will be denoted by lower-case Greek letters. The
I 3 , and the unit matrix in R
symbols e1 , e2 , e3 , and I3 denote the oriented Cartesian basis in R I 3×3 , respectively.
Our report is organized as follows. In the next section we review the original principle of minimum
potential energy. Section 3 deals with several schemes of purging the rigid displacement modes from the
pure traction elasticity. These are later used in §4 to formulate three alternative potential energy principles.
The quadratic programs defining finite element methods for these principles are derived in §5. We conclude
with a discussion of properties of algebraic problems related to each quadratic program.

2. Minimum potential energy principle

The potential energy of an elastic isotropic body Ω under a given load f and a surface traction t is a quadratic
function of the displacement v:
Z Z Z
1
J (v; f , t) = σ(v) : ε(v) dΩ − f · v dΩ − t · v dΓ . (2.1)
2 Ω Ω ΓT

The load f and the traction t are assumed to satisfy the force equilibrium constraint (1.6)–(1.7). The first
term in J (·) represents the strain energy E(v) of the body. Because our main focus is on approaches for
handling the rigid body modes without loss of generality we set t = 0 for the rest of this paper and denote
the resulting energy functional J (v; f ).
The original principle of minimum potential energy [7, p.208] postulates that among all kinematically
admissible states v the state of the elastic body Ω is the one that minimizes the potential energy J (v).
Kinematically admissible states for (2.1) can be identified with the Sobolev space H1 (Ω). Thus, a formal
mathematical statement of the minimum energy principle is

inf J (v; f ). (2.2)

v∈H1 (Ω)

The minimum energy principle also states that if u1 and u2 are two minimizers, then u1 − u2 is a rigid
displacementt [7, p.55] of Ω, that is a vector field u with constant and skew gradient terms. We denote the
I3
set of all rigid displacements by N . In R

N = {u ∈ H1 (Ω) | u = a + b × r; I 3} .
a, b ∈ R (2.3)

With the tacit understanding that a ∈ R I 2 , b = b is a constant, and that the vector product b × r is the
defined by be3 × (ρ1 e1 + ρ2 e2 ), definition (2.3) remains valid in R I 2 . Note that dim N is 3 in R
I 2 and 6 in
3 3
R
I . This corresponds to the fact that in R I there are three rotational and translational rigid displacements,
I 2 there’s only one rotational and two translational modes. The compatibility conditions for the
while in R
pure traction problem in two and three dimensions can be expressed by the compact statement
Z
f · w = 0 for all w ∈ N , (2.4)
Ω

that is, f ∈ N ⊥ .
I 3 . Specialization of all results
For the remainder of this report we confine the discussion to the case of R
to two-dimensions is straightforward. The set

{wi }6i=1 = {ei }3i=1 ∪ {ei × r}3i=1 (2.5)

is a basis for N ; the first three vectors span the translational modes, whereas the last three span the rotational
modes. In mathematical terms, existence of the nontrivial set N means that the minimization problem (2.2)
has multiple minimizers.
Our plan is to formulate finite elements for the pure traction problem by directly using the minimum
energy principle, albeit in a modified form. To carry out this plan we will introduce three alternative versions
of (2.2). Two of these new energy principles will use the same energy functional but a different space of
minimizers, while the third one will use a different, regularized energy functional but the original minimizer
space. In contrast to (2.2), where minimizers are determined up to a rigid displacement, all three alternative
energy principles will posses unique minimizers. The key to our approach is the proper handling of the rigid
modes of the pure traction problem. The next section considers this issue.
4 P.B. Bochev & R.B. Lehoucq

Before leaving this section, we define several matrices and vectors that we will use in the remainder of
our report. Let  
W = w1 · · · w6 (2.6)
be the matrix of nullspace vectors (2.5) with three rows and six columns. Any w ∈ N has the form
6
X
w = Wx = ξi wi
i=1

where x = [ξ1 , . . . , ξ6 ] are the coordinates of w with respect to the basis (2.5). Define
 Z 
Θ1 u
B(u) ≡  Z Ω (2.7)
 

Θ2 ∇ × u
Ω

where the entries of the diagonal matrices

Θ1 = diag(θ1,1 , θ1,2 , θ1,3 ) and Θ2 = diag(θ2,1 , θ2,2 , θ2,3 ) (2.8)

are elements of H −1 (Ω) satisfying the hypothesis

Z Z
θ1,i wi 6= 0 and θ2,i wi+3 6= 0
Ω Ω

for i = 1, 2, 3. These dual functions serve the useful purpose of allowing us to enforce the mean and mean
of the curl of the displacement along a portion of Ω. p For the remainder of the report, we’ll assume that
Θ1 = I3 = Θ2 unless otherwise stated. Let |B(u)| ≡ B(u)T B(u). Finally, define the 6 × 6 matrix

S = B(W) . (2.9)

A simple calculation shows that the above matrix is

 
µ(Ω)I3 S12
S= (2.10)
0 2µ(Ω)I3

where µ(Ω) is the volume of the region Ω and

 Z Z 
0 ρ3 e3 − ρ2 e2
Ω Z Ω
 Z 
 
S12 ≡ − ρ3 e 3 0 ρ1 e1 .
 
 Z Ω Z Ω 
 
ρ2 e2 − ρ1 e1 0
Ω Ω

Therefore, S is nonsingular.

3. Elimination of rigid displacements

A rigid displacement cannot increase the strain energy of an elastic body. This can be verified by evaluating
the strain energy functional E(u) and showing that it vanishes for all u ∈ N . Mathematically, this means
that E(·) is only a seminorm on the space H1 (Ω). Our main goal will be to elevate the status of E(·) to that
of a true norm. Once this is accomplished, the finite element solution of the pure traction problem can be
treated within the classical Rayleigh-Ritz variational framework.
There are two distinct ways of turning E(·) into an equivalent norm. Both eliminate the rigid displace-
ments from the pure traction problem, however they differ in the manner in how this is accomplished.
The first one is to purge the rigid modes from the admissible space H1 (Ω) while retaining the original
definition of the strain energy. Two cases will be discussed, one uses a factor space and the other uses
a complement space. The second method retains the original admissible space, but regularizes the strain
energy in a manner that flushes out rigid displacements. We will pay special attention to the second approach
because it leads to a class of finite elements that has not yet been carefully examined in the literature.
 CMAM Energy principles and finite element methods for pure traction linear elasticity 5

3.1. Complement space

Define the space
H1C (Ω) = {v ∈ H1 (Ω) | B(v) = 0}. (3.1)
Evidently, H1C (Ω) 1
is a closed subspace of H (Ω). The next Lemma explains why H1C (Ω) is an algebraic
complement of N in H1 (Ω).

Lemma 3.0.1. For any u ∈ H1 (Ω) there exist unique uC ∈ H1C (Ω) and uN ∈ N , such that

u = uC + uN .

I 3 , the proof in the planar case is identical. Let u ∈ H1 (Ω) be arbitrary

Proof. We prove the theorem in R
but fixed. We seek uN = Wx as solution of the 6 × 6 system of equations
Z Z Z Z
Wx = u and ∇ × (Wx) = ∇ × u. (3.2)
Ω Ω Ω Ω

In terms of the matrix defined in (2.9) and the functional B system (3.2) has the following algebraic repre-
sentation:
Sx = B(u) . (3.3)
Because S is nonsingular (3.3), resp., (3.2) has a unique solution x, which defines the component uN =
Wx ∈ N . Set uC = u − uN . Then
Z Z
(u − uN ) = ∇ × (u − uN ) = 0 ,
Ω Ω

that is, uC ∈ H1C (Ω). The conclusion of the lemma is established.

We remark that N ⊥ 6= H1C (Ω) and so (uC , uN ) 6= 0 in general.

Corollary 3.1. The complement space H1C (Ω) has a trivial intersection with N : N ∩ H1C (Ω) = {0}.
Proof. Assume that u ∈ N ∩ H1C (Ω). Because u ∈ N it has the form u = Wx where x solves the linear
system (3.3). Because u ∈ H1C (Ω) there holds B(u) = 0, i.e., the linear system (3.3) has homogenous right
hand side. It follows that x = 0 and u = Wx = 0 as well.
This Corollary asserts that rigid modes have been purged from H1C (Ω), i.e., the complement space contains
only a trivial rigid displacement. For us, the most notable property of H1C (Ω) is that strain energy is an
equivalent norm on H1C (Ω). This follows from a result known as Korn’s second inequality2 that we quote for
future reference.

Theorem 3.0.1. [3, p.222] There exists a positive constant C such that

Ckvk1 ≤ kε(v)k0 for all v ∈ H1C (Ω) (3.4)

Using Korn’s second inequality it is a simple matter to show that E(·) is an equivalent norm on H1C (Ω).

3.2. Factor space

A standard way of removing an undesirable subspace is to factor it from the respective function space. In
the present context this leads to the factor space

H1N (Ω) = {b
u ⊂ H1 (Ω) | u1 , u2 ∈ u
b ⇔ u1 − u2 ∈ N }. (3.5)

When equipped with the factor norm

kb
ukH1N (Ω) ≡ inf kvk1 (3.6)
v∈b
u

H1N (Ω) is a Hilbert space. We recall that a factor space does not contain usual functions, instead its members
are equivalence classes of functions. For instance, the subspace of rigid displacements N is confined to the
zero element b0 of the factor space H1N (Ω). This is the trick that turns E(·) into an equivalent norm on
H1N (Ω).
2 Second Korn’s inequality can be established on other subspaces as well. For example, it is valid on H1 (Ω)
⊥ = {v ∈
H1 (Ω) | (v, w) = 0 ∀w ∈ N }.
6 P.B. Bochev & R.B. Lehoucq

Lemma 3.0.2. The mapping H1N (Ω) 3 u

b 7→ kε(u)k0 , where u ∈ u
b is arbitrary, defines an equivalent
norm on H1N (Ω).

Proof. Let ub be a given equivalence class and u ∈ ub an arbitrary but fixed function in that class. Using
this fixed function the right hand side in (3.6) can be expressed as

inf kvk1 = inf ku − wk1

v∈b
u w∈N

because for u, v ∈ u
b we have that u − v = w ∈ N . Therefore, we need to prove that there are positive
constants C1 and C2 , independent of u and such that

C1 kε(u)k0 ≤ inf ku − wk1 ≤ C2 kε(u)k0

w∈N

The lower bound is straightforward because ε(u) = ε(u − w). To prove the upper bound we use Korn’s
second inequality (3.4) and Lemma 3.0.1. According to the lemma we can write the function u ∈ u
b as
u = uC + uN where u ∈ H1C (Ω) and uN ∈ N . Because ε(u) = ε(u − uN ) and u − uN = uC ∈ H1C (Ω), the
Korn’s second inequality yields the statement of the lemma:

kε(u)k0 = kε(u − uN )k0 ≥ Cku − uN k1 ≥ inf Cku − wk1 .

w∈N

3.3. Regularized strain energy

If we wish to retain the original admissible space H1 (Ω), removal of the rigid displacements must necessarily
be accomplished by regularization of the strain energy. To motivate the regularization process, we prove a
modified version of Korn’s inequality.

Theorem 3.0.2. There exists a positive C such that

kε(v)k0 + |B(v)| ≥ Ckvk1 (3.7)

for all v ∈ H1 (Ω).

Proof. We use standard compactness argument. Suppose that (3.7) does not hold for any positive C.
Then, there exists a sequence {vn } ∈ H1 (Ω) such that

1
kvn k1 = 1 and kε(vn )k0 + |B(vn )| < . (3.8)
n
From Lemma 3.0.1, given vn there are wn ∈ N and zn ∈ H1C (Ω) such that vn = wn + zn . The function
wn = Wx where the coefficient vector x solves the linear system (3.3), i.e.,

x = S−1 B(vn ) .

Because the right hand side B(vn ) involves at most integrals of the first derivatives of vn it easily follows
that there exists a positive constant C such that

kwn k1 ≤ Ckvn k1

for all integers n. Therefore,

kzn k1 + kwn k1 = kvn − wn k1 + kwn k1 ≤ kvn k1 + 2kwn k1 ≤ Ckvn k1 . (3.9)

The Korn’s second inequality (3.4) holds for all functions in H1C (Ω), in particular it is valid for zn . In
combination with the fact that the strain energy vanishes for wn and the hypothesis on vn we find that
1
Ckzn k1 ≤ kε(zn )k0 = kε(zn + wn )k0 = kε(vn )k0 < .
n
Therefore,
lim zn = 0 in H1 (Ω).
n→∞
 CMAM Energy principles and finite element methods for pure traction linear elasticity 7

This, and (3.9) imply that {wn } is bounded in H1 (Ω). Compactness of the imbedding H1 (Ω) ,→ L2 (Ω)
means that a subsequence of {wn }, denoted by the same symbol, converges in L2 (Ω). Since N is finite
dimensional, this subsequence must also converge in H1 (Ω). Let w ∈ N be its limit. Then

lim vn = lim zn + wn = w ∈ N .
n→∞ n→∞

Therefore,
kwk1 = lim kvn k1 = 1.
n→∞

On the other hand, taking the limit in (3.8) and noting that w has vanishing strain gives

B(w) = 0 ,

that is w ∈ N ∩ H1C (Ω). But Corollary 3.1 asserts that the intersection of these two spaces is trivial and so
we must have w = 0, a contradiction.

Remark 3.1. The modified Korn inequality in Theorem 3.0.2 plays the same role for the pure traction
elasticity problem as the modified Poincare-Friedrichs inequality [2, Lemma 2.2]
 Z 
kuk0 ≤ C |u|1 + | udx|
Ω

does for the pure Neumann problem.

The new Korn inequality prompts regularization of the strain energy E(·) by adding the mean and the
mean of the curl of the kinematic state. The new, regularized strain energy functional is
Z
1 1
ER (v) = σ(v) : ε(v) + B(v)T ΥB(v) (3.10)
2 Ω 2

where Υ is a diagonal matrix of order six. The regularization parameters are the diagonal elements of Υ.
The regularization terms serve to increase the frequency of the rigid body modes.
Together, the last term in (3.10) represent an “energy” measure of the rigid displacements. As a result,
the modified strain energy does not vanish on N .

4. Alternative energy principles

The factor space, the complement space and the regularized strain energy give rise to three functionals for
the minimum potential energy principle. The first one is to minimize the original potential energy J (·) over
the factor space H1N (Ω) leading to the minimization problem

inf J (b
v; f ). (4.1)
b∈H1N (Ω)
v

The second alternative is to minimize the same functional over the complement space H1C (Ω) and so the
minimization problem now is
inf J (v; f ). (4.2)
v∈H1C (Ω)

A third alternative is to minimize a regularized potential energy functional over the original admissible space
H1 (Ω). The regularized minimum potential energy principle is
Z
inf1 JR (v, f ) = inf1 ER (v) − f · v. (4.3)
v∈H (Ω) v∈H (Ω) Ω

The remainder of this section demonstrates that each one of the three alternative energy principles is a
well-posed minimization problem with a unique minimizer.
8 P.B. Bochev & R.B. Lehoucq

4.1. Factor Space minimization

Theorem 4.0.3. If the load f ∈ L2 (Ω) satisfies compatibility condition (2.4) then the minimization
b f from H1N (Ω).
problem (4.1) has a unique minimizer u
Proof. The energy principle (4.1) is a well-defined minimization problem for any load f that satisfies the
force equilibrium (2.4). In the context of a factor space well-defined means that the value of J (·) does not
depend on the particular class member u ∈ u b . To show this, let u1 and u2 be two functions from the same
b ∈ H1N (Ω). Then, u1 − u2 = w ∈ N ; E(u1 ) = E(u2 ) and so
class u
Z
J (u1 ) − J (u2 ) = f · (u1 − u2 ) = 0
Ω

because of (2.4). Let us now prove that (4.1) has a unique minimizer. The first-order optimality condition
b ∈ H1N (Ω) such that
(Euler-Lagrange equation) for (4.1) is represented by the variational problem: seek u

A(b b ) = F(b
u; v v) b ∈ H1N (Ω)
for all v (4.4)

where Z Z
A(b
u; v
b) = σ(b
u) : ε(b
v) and F(b
v) = f ·v
b. (4.5)
Ω Ω

The bilinear form A(·; ·) is well-defined in the same sense—its value does not depend on the members of the
classes used to evaluate it. Same holds true for F(·) provided (2.4) is valid.
Furthermore, Lemma 3.0.2 implies that A(·; ·) is coercive on H1N (Ω) × H1N (Ω). Continuity of the form
and the right hand side functional is obvious and from the Lax-Milgram theorem it follows that (2.1) has a
unique minimizer out of H1N (Ω).

4.2. Complement space minimization

Theorem 4.0.4. The minimization problem (4.2) has a unique minimizer from H1C (Ω) for any load
f ∈ L2 (Ω).
Proof. Since the complement space contains only the trivial rigid displacement, J (·) is well-defined on
H1C (Ω) without any compatibility condition on the data. To prove that (4.2) has a unique minimizer consider
again the necessary condition: seek u ∈ H1C (Ω) such that

A(u; v) = F(v) for all v ∈ H1C (Ω). (4.6)

The bilinear form A(·; ·) and functional F(·) are the same as in (4.5), except that the arguments are drawn
from H1C (Ω) and hence are functions.
From Korn’s second inequality (3.4) it follows that this form is coercive on H1C (Ω). It is easy to see that
both the form and the functional are also continuous. Existence of a unique minimizer thus follows from
Lax-Milgram theorem.
Solutions of (4.2) are implicitly constrained by the definition of H1C (Ω) to have a zero mean and zero
mean curl. These constraints can also be imposed explicitly producing an equivalent, but constrained,
minimization problem:
inf1 J (v) subject to B(v) = 0. (4.7)
v∈H (Ω)

Using a Lagrange multiplier y ∈ R I 6 to enforce the constraints gives another problem that is equivalent to
(4.2) but is of the saddle-point type:

sup inf J (v; f ) + yT B(v). (4.8)

I6
y∈R v∈H1 (Ω)

I 6 such that
The first-order optimality system for (4.8) is to find (u, y) ∈ H1 (Ω) × R

A(u; v) + B(v)T y = F(v) ∀v ∈ H1 (Ω)

(4.9)
zT B(u) = 0 I6
∀z ∈ R

Theorem 4.0.5. The saddle point system (4.9) is well-posed; a unique solution (u, y) exists for any
f ∈ L2 (Ω).
 CMAM Energy principles and finite element methods for pure traction linear elasticity 9

Proof. Equation (4.9) is a saddle-point problem and is well-posed if and only if an inf-sup condition holds.
For (4.9), we need to demonstrate that there exists a positive constant γ such that for any y ∈ R I6

yT B(v)
sup ≥ γ|y| . (4.10)
v∈H1 (Ω) kvk1

Given y ∈ R I 6 the proof of Lemma 3.0.1 allows us to construct uN = Wx ∈ N such that B(uN ) = y by
solving the system Sx = y. Furthermore,
v
6 u 6
X uX 1
kuN k1 = kWxk1 ≤ |ξi |kwi k1 ≤ |x|t kwi k21 ≤ |y|
i=1 i=1
γ
qP
6
where 1/γ = kS−1 k 2
i=1 kwi k1 . Therefore,

yT B(uN ) yT y
= ≥ γ|y| ,
kuN k1 kuN k1

which proves the inf-sup condition.

4.3. Regularized minimization

Theorem 4.0.6. The regularized energy functional (4.3) has a unique minimizer from H1 (Ω) for any
f ∈ L2 (Ω).

Proof. The Euler-Lagrange equation for (2.1) reads: find u ∈ H1 (Ω) such that

AR (u; v) = F(v) for all v ∈ H1 (Ω) (4.11)

where Z
AR (u; v) = σ(u) : ε(v) + B(u)ΥB(v)T (4.12)
Ω

and the functional F(·) is the same as in (4.5). Korn’s modified inequality (3.7) implies that form AR (·; ·)
is coercive on H1 (Ω) × H1 (Ω). Therefore, from the Lax-Milgram lemma we can infer that the regularized
problem has a unique minimizer out of H1 (Ω).

4.4. Relationship between Complement and Regularized Minimization

We point out that the results we established for the unique minimizers of the complement and regularized
energy principles did not require the force equilibrium condition (2.4). For (4.2) this was possible because
the complement space does not contain any nontrivial rigid displacements. In contrast, for the regularized
energy principle (4.3), this was possible because the functional was augmented so that rigid displacements
have positive energy. Hence, (4.2) and (4.3) will always generate nonsingular linear systems. However, the
fact that (4.2) and (4.3) possess solutions even for incompatible loads raises the question about the proper
interpretation of their minimizers as solutions of the pure traction problem. To answer this question we need
an auxiliary lemma.

Lemma 4.0.3. Given f ∈ L2 (Ω) there exists a unique decomposition f = Pf + Qf where Pf and Qf are
the L2 projections onto N and N ⊥ , respectively, and Qf satisfies the force equilibrium condition (2.4).

Proof. The orthogonal projection Pf onto N is defined by the variational equation

(Pf , w) = (f , w) ∀w ∈ N . (4.13)

The elements in N have the form Wx where W is the rigid displacement basis defined in (2.6). Using the
ansatz Pf = Wx it is easy to see that (4.13) is a 6 × 6 linear system
Z 
WT W x = (f , W) (4.14)
Ω
10 P.B. Bochev & R.B. Lehoucq

for the unknown coefficient vector x. The matrix Ω WT W is nonsingular as the Gramm matrix of the rigid
R

displacement basis W. Therefore, (4.14), resp. (4.13), has a unique solution x which defines the unique
projection Pf = Wx. Let Qf = f − Pf . Then from (4.13)

(Qf , w) = (f − Pf , w) = (f , w) − (Pf , w) = 0 ∀w ∈ N ,

Therefore, Qf ∈ N ⊥ , i.e., it satisfies the equilibrium condition.

The next result characterizes the minimizer of the regularized energy principle.

Theorem 4.0.7. If uR denotes the unique minimizer of the regularized energy principle (4.3) for some
given f ∈ L2 (Ω), then
B(uR ) = Υ−1 S−T (Pf , W) (4.15)
where S is given by (3.3) and Pf is the projection defined in Lemma 4.0.3.
Proof. A minimizer of the regularized energy principle (4.3) satisfies the weak equation (4.11), which is
the necessary optimality condition for (4.3). Choose the test function in (4.11) to be a rigid displacement
w ∈ N . Using that ε(w) = 0 for w ∈ N and the definition of Pf , the weak equation reduces to

B(uR )ΥB(w)T = (Pf , w) ∀w ∈ N .

Because W spans N , and B(W) defines the matrix S in (2.9), this equation is equivalent to the linear
system
B(uR )ΥST = (Pf , W) .
Formula (4.15) gives the unique solution of this linear system.
The following corollary provides further information about the regularized minimizer.

Corollary 4.1. If f satisfies the force equilibrium condition (2.4), then uR ∈ H1C (Ω).
Proof. Assume that the load f satisfies (2.4), i.e., f ∈ N ⊥ . Then, Pf = 0 and from (4.15) in Theorem
4.0.7 it follows that B(uR ) = 0, i.e., uR ∈ H1C (Ω).
Theorem 4.0.7 and Corollary 4.1 indicate that strain energy principle (3.10) represents a regularization
by selection rather than by penalty. For compatible loads, uR does not depend on the regularization matrix
Υ. When the load is incompatible, the regularization terms help to select a unique solution from H1 (Ω).
Consider next the potential energy principle (4.2) posed on the complement space.

Theorem 4.0.8. Let uC denote the unique minimizer of (4.2) for some given f ∈ L2 (Ω). If f = Qf + Pf
where the decomposition is given by Lemma 4.0.3, then uC = uC,Q + uC,P where
1. uC,Q ∈ H1N (Ω) with respect to the compatible load Qf ;
2. uC,P solves the weak equation: seek uC,P ∈ H1C (Ω) such that

A(uC,P ; v) = (Pf , v) ∀v ∈ H1C (Ω) (4.16)

Proof. After the ansatz f = Qf + Pf , the weak equation (4.6) reads:

A(uC ; v) = (Pf , v) + (Qf , v) ∀v ∈ H1C (Ω).

Define uC,Q ∈ H1C (Ω) to be the solution of the weak equation

A(uC,Q ; v) = (Qf , v) ∀v ∈ H1C (Ω).

Because Qf satisfies the force equilibrium condition (2.4), it follows that uC,Q satisfies the weak equation
(4.4) with F(v) = (Qf , v). Thus, uC,Q belongs to a minimizing class of a potential energy principle set with
respect to Qf . The difference uC − uC,Q satisfies the weak equation

A(uC − uC,Q ; v) = (Pf , v) ∀v ∈ H1C (Ω),

and so it gives the second solution component uC,P in the statement of the theorem.
 CMAM Energy principles and finite element methods for pure traction linear elasticity 11

Corollary 4.2. If f satisfies the force equilibrium condition (2.4), then

1. uC belongs to the minimizing class of (4.1);
2. Minimizers of the regularized problem (4.3) and the complement space problem (4.2) coin-
cide: uC = uR .
The significance of this corollary is that the relationship between minimizers of the two alternative poten-
tial energy principles and the original pure traction problem is revealed. When f satisfies (2.4), minimizers
of (4.2) and (4.3) coincide with the unique member of u b associated with f . Thus, both problems lead to
a solution of the pure traction problem. If the data is not compatible, then (4.2) and (4.3) have distinct
minimizers. These minimizers are solutions of a perturbed pure traction problem.

5. Finite element methods

Throughout this section we use standard finite element notations. Th denotes a uniformly regular partition of
Ω into finite elements. For simplicity we consider only simplicial triangulations and standard C 0 Lagrangian
finite element spaces P k , where k ≥ 1 stands for the polynomial degree employed. The standard nodal basis
for P k will be denoted by {φhi }N
i=1 . Vector fields are approximated by the product space

d
Y
h
X = P k,
i=1

where d = 2 or d = 3 is the space dimension. We present details for d = 3, the planar case is very similar.
The basis of Xh can be expressed in terms of the nodal basis in P k as {φhi e1 , φhj e2 , φhk e3 }N
i,j,k=1 . We will
also use the compact notation {φhi } with the tacit understanding that the nodes and unknowns have been
ordered in some manner. We denote the finite element approximant
N
X
uh ≡ υi,1 φhi e1 + υj,2 φhj e2 + υk,3 φhk e3 .
i,j,k=1

With some abuse of notation we will use u to denote the coefficient vector of the finite element vector
function uh . The coefficient sets corresponding to finite element expansions of the components of uh will be
denoted by ui . With Ih we denote the standard nodal interpolation operator H 1 (Ω) ∩ C 0 (Ω) 7→ P k , while
I N that returns the nodal coefficients of a finite element function.
ih is the operator P k 7→ R
We now discuss the formulation of finite element methods for the pure traction linear elasticity. Let
us first point out that the linear equilibrium equation (1.1) and the minimum potential energy principle
(2.2) are equivalent in the sense that the differential equation problem is a strong form of the necessary
condition for the minimization problem. Traditionally, to derive a discrete problem finite element methods
have relied exclusively upon the differential equation description of the elastic body state. The common
Galerkin approach is to write (1.1) in weak form and then restrict this variational problem to finite element
subspaces. The result is a linear algebraic system whose solution defines the finite element approximation.
Thus, Galerkin finite elements use the extant optimization setting implicitly.
In contrast, here we adopt an alternative approach in which discrete problems are obtained directly from
a minimum potential energy principle by restricting its admissible states to a finite element subspace. This
results in a quadratic programming problem and so solution techniques from numerical optimization may
be used.
Another reason to favor the quadratic programming as a starting point in the determination of the finite
element solution has to do with the spaces used in (4.1) and (4.2). When rigid displacements are purged from
the admissible space or confined to a zero class, the result is a nonstandard space of kinematic states. By
nonstandard here we mean spaces whose members are either classes of functions, as in (3.5), or are subject
to constraints as in (3.1). The use of nonstandard spaces does not render these settings unusable for finite
element methods. However, it does require more careful handling of the ensuing algebraic problems. In
what follows we will see that (4.2) is related to an equality constrained quadratic program, (4.1) will give
a quadratic program with semi-definite matrix, while (4.3) leads to a problem with symmetric and positive
definite matrix.
To state the quadratic programs and discuss their properties we shall need some additional vectors
and matrices. Since we consider at least piecewise linear elements, the functions 1, x1 , x2 and x3 are
all represented exactly by P k . Let c, x1 , x2 and x3 denote the nodal coefficients of their finite element
expansions on the mesh Th , that is c = ih (1), x1 = ih (x1 ), x2 = ih (x2 ) and x3 = ih (x3 ). For example, it is
12 P.B. Bochev & R.B. Lehoucq

I N . The coefficients of the other three functions

easy to see that c = (1, . . . , 1)T is the constant vector in R
can also be easily determined once Th is given. Next, we introduce the vectors

c1 = ih (e1 ); c2 = ih (e2 ); c3 = ih (e3 );

and
r1 = ih (e1 × r); r2 = ih (e2 × r); r3 = ih (e3 × r).
The first three vectors contain the nodal coefficients of the three translational rigid displacements. We have
c1 = e1 ⊗ c, c2 = e2 ⊗ c and c3 = e3 ⊗ c. The ri ’s represent nodal coefficients of the three rotational modes.
For example,
r1 = (0, −x3 , x2 )T .
We arrange these vectors in a matrix as follows:
 
N = c1 c2 c3 r1 r2 r3 . (5.1)

The finite element functions corresponding to these coefficient vectors are wih , i = 1, . . . , 6.
We shall also need the symmetric and semi-definite stiffness matrix A with element i, j

Ai,j = A(φhj ; φhi ), i, j = 1, . . . , 3N. (5.2)

and the discrete source term

bi = F(φhi ), i = 1, . . . , 3N. (5.3)
Our ordering convention induces a natural splitting of A and b into blocks of size N × N and N × 1,
respectively. For example the ijth element in the klth block of A is given by A(el φhj ; ek φhi ). For a nodal
basis
ker(A) = span{ci , ri }3i=1
and so AN = 0.
We will have occasion to use two discrete inner products. The first is the standard Euclidean inner
product of x and y denoted by xT y. The Euclidean norm of a vector will be denoted by k · k. The second
choice of inner product is the discrete L2 (Ω) inner product of x and y denoted by xT My where M is the
mass matrix containing the L2 (Ω) inner products of finite Pelement basis functions. Recall that if x and y
ξi φhi and h = ηi φhi , then (g, h) = xT My is
P
are the nodal coefficient associated with the functions g =
the discrete L2 (Ω) inner product of g and h.
To complete our formulation of finite element methods for pure traction problems in linear elasticity, we
need to discuss the computation of the finite element minimizers. This is the topic of the next few sections.

5.1. Factor space FEM

Consider first the minimization problem (4.1) and its restriction to a finite element subspace XhN = Xh /N
I 3N / ker(A) and so
of H1N (Ω). This space is isomorphic to R

1 T
uh ; f ) ≡
min J (b min u
b Ab b T b.
u−u (5.4)
b h ∈Xh
u N
I 3N / ker(A)
b ∈R
u 2

Therefore, the quadratic program on the right hand side of (5.4) represents the minimum potential energy
principle (4.1) on the finite dimensional subspace XhN .
The first-order optimality condition for (5.4) is

Ab
u = b. (5.5)

A solution of (5.5) should be interpreted in the sense of computing any member of the minimizing class u b.
Since A is singular, a direct method cannot be used for this purpose (at least not without modification so
that a zero pivot can be detected; see [4] for a direct solver that handles the hydrostatic pressure mode in
the incompressible Stokes equations).
However, an iterative method can be used to solve the minimization problem (5.4) instead of computing
a solution to the Euler-Lagrange equation (5.5). Indeed, as long as the b is in the range of A the quadratic
functional in (5.4) has a finite lower bound. As a result, an iterative method such as the conjugate gradient
method will generate a minimizing sequence that will converge modulo ker(A); see Theorem 13.11, [1,
 CMAM Energy principles and finite element methods for pure traction linear elasticity 13

p. 583]. The rate of convergence of an iterative method depends on the ratio3 κ (A) = λ3N (A)/λ7 (A) or
the effective condition number. The paper [11] provides further information about the iterative solution of
consistent singular systems arising in incompressible flows.
An important practical consideration for (5.5) is that the discrete source b must be discretely orthogonal
to the vectors ci and ri or equivalently NT b = 0; this is a consequence of (2.4) because
Z Z
T T
ci b = f · ei and ri b = f · (ei × r). (5.6)
Ω Ω

In other words, solvability of the pure traction problem implies solvability of (5.5) provided integration is
carried out exactly.
In practice, the source b and the stiffness matrix A are computed in floating point arithmetic via
quadrature and may not satisfy (5.6). To restore consistency of the system we introduce the discrete
analogue of the projection P onto the rigid displacements space N of Lemma 4.0.3 as

(P h f , wih ) = (f , wih ) i = 1, . . . , 6 .

The functions wih are the finite element functions whose coefficients are the six vectors in (5.1). The matrix
P = N(NT MN)−1 NT M is the equivalent algebraic form of P h acting directly on the coefficients of finite
element functions. A direct calculation reveals that the matrix form of the companion projection Qh onto
the complement of N in the finite element space Xh is

Q = I − N(NT MN)−1 NT M (5.7)

Application of the projector to the linear system (5.5) results in

(QAQT )u = Qb.

The computed solution QT u(= u b ) satisfies wT QT u = 0 and so the FEM solution is in N ⊥ .

We remark that that the algebraic projection Qa = I − N(NT N)−1 NT may also be used. When a
structured triangulation is employed, then Qa is essentially an orthogonal discrete L2 (Ω) projector. If the
triangulation is unstructured, then Qa is orthogonal with respect to the Euclidean inner product but not
the discrete L2 (Ω) inner product. The use of a projector to maintain consistency of the linear system is not
unfounded, especially for unstructured triangulations.

5.2. Constrained FEM

Consider the energy principle (4.2) posed on the complement space H1C (Ω). We restrict (4.2) to a finite
I 3N −6 , the discrete problem is
element subspace XhC of H1C (Ω). Since XhC is isomorphic to R
1 T
min J (uh ; f , t) ≡ min u AC u − uT bC , (5.8)
uh ∈Xh
C
I 3N −6
u∈R 2

where AC , bC , and u denote a stiffness matrix, right hand side and a coefficient vector relative to the nodal
basis of XhC . Again, the right hand side in (5.8) is a quadratic program that represents (4.2) on the finite
dimensional subspace XhC .
The first-order optimality condition is
AC uC = bC .
The caveat here is that in order to form the matrix AC and the discrete load bC , a conforming basis for XhC
is required. If the triangulation Th is non-uniform, imposing even the zero mean condition alone upon the
basis may lead to significant increase in condition numbers; see [2].
An alternative quadratic program can be obtained from (4.8). The first-order necessary conditions for
(4.9) are
A BT
    
uC b
= (5.9)
B 0 y 0
The matrix in (5.9) is called the Karush-Kuhn-Tucker (KKT) matrix. The matrix B is the FEM implmen-
tation of the operator B(·) defined in (2.7).
There are several ways to solve the KKT linear system. The first approach would solve the indefinite
linear system by either a sparse direct method or an iterative method. The second and third methods exploit
3 We remind that the eigenvalues of the stiffness matrix are orderered as 0 = λ1 = . . . = λ6 ≤ λ7 , . . . , ≤ λ3N .
14 P.B. Bochev & R.B. Lehoucq

the structure in the KKT matrix. The range-space method for the KKT system requires that A is nonsingular
and so it is not applicable to (5.9). The null-space method removes the requirement of nonsingularity of A
and can instead be used. However knowledge of the (3N − 6)-dimensional null-space

I 3N | Bu = 0}
Z = {u ∈ R

of B is needed. Let C ∈ R I 3N ×3N −6 denote a matrix whose columns form a basis for Z. Then u = Cv for
3N −6
v∈R I and the constrained problem in terms of u is equivalent to
1 T T
min v C ACv − vT CT b. (5.10)
I 3N −6
v∈R 2
The minimizer of (5.10) is computed by solving the symmetric positive-definite linear system

CT ACv = CT b. (5.11)

The null-space method for (5.9) amounts to constructing the matrix C and solving the linear system of
equations (5.11).

5.3. Regularized FEM

The last quadratic program is derived from the regularized energy principle (4.3). After minimization in
(4.3) is restricted to Xh we obtain the following discrete problem:
1 T
min JR (uh ; f , t) ≡ min u AR u − uT b. (5.12)
uh ∈Xh I 3N
u∈R 2
The elements of AR are given by AR (φi ; φj ) of (4.12).
In this class of methods, the finite element solution of the pure traction problem is obtained by computing
a minimizer of (5.12). The first-order necessary condition for (5.12) is

AR u = b. (5.13)

The elements of the matrix AR are determined from (4.12) and result in

AR ≡ A + BΥBT

where A is the singular stiffness matrix of order 3N , B is a rank-six matrix and Υ is a diagonal matrix of
order six containing the regularization parameters. Depending on the choice of Θ1 and Θ2 of (2.7), AR may
be dense even if A is sparse because B is a dense rank six matrix. This precludes the use of direct solver
unless Θ1 and Θ2 are selected properly.
However, the possible disadvantage is only formal as we can solve the regularized system iteratively with
very little additional cost added. For instance, to solve (5.13) by the preconditioned conjugate gradient
algorithm, only the matrix vector product AR u is needed. The following two steps lead to an efficient
implementation:
1. compute w = Au;
2. compute the vector y = Υ(BT u);
3. update w + By.
Step (1) is a standard part of any finite element solver, so the only additional work involved is in the rank-six
update. Step (2) performs the six inner products needed followed by the rank size update in Step (3). The
rank six matrix B can be precomputed and stored rendering the computation of y efficient.
The following theorem proves fundamental for understanding the structure of AR and how the rank-six
update modifies the null-space of A.

Theorem 5.0.9. Suppose that A = ZΛZT is an eigen-decomposition, where

 
Z6 ≡ ZE6 ≡ Z e1 · · · e6

and AZ6 = 0. If B = PB + QB where Q is defined in (5.7) and P = I − Q are discrete L2 projectors, then

kAR − Z(Λ + E6 ΥET6 )ZT k ≤ kΥk kQBk(2 + kQBk) (5.14)

 CMAM Energy principles and finite element methods for pure traction linear elasticity 15

Proof. From the identity

AR = A + BΥBT = Z Λ + (ZT B)Υ(ZT B)T ZT

(5.15)

the conclusion of the theorem will follow by examining ZT B. By the hypothesis on Z6 , we have ZT B =
ZT PB + ZT QB = E6 B + ZT QB and so

ZT BΥ(ZT B)T = E6 ΥET6 + E6 Υ(BT QZ)T + BT QZΥET6 +

(5.16)
BT QZΥ(BT QZ)T .

Substituting (5.16) into (5.15) and using the identity that kZk = kZT k = 1 gives the desired result.
The theorem demonstrates that the role of Υ is to shift the zero eigenvalues of A while only perturbing
the eigenvectors. Values that are at least as large as the smallest positive eigenvalue of A but no larger than
the largest eigenvalue will suffice.

6. Conclusions
We have developed a framework for the finite element solution of the pure traction linear elasticity based on
direct discretization of a minimum potential energy principle. Our approach reveals important connections
between finite element approximants and solutions of quadratic programs and prompts a novel regularized
formulation of the pure traction problem. Among the principal advantages of this formulation are the
possibility to obtain symmetric and positive definite algebraic systems using standard nodal finite element
spaces and the elimination of the force equilibrium constraint. These systems can be solved efficiently by
preconditioned conjugate gradients with the added cost of just few rank-six updates per iteration. The
regularization approach is not a penalty approach and does not require large values of the weights in Υ.
Theorem 5.0.9 reveals that weights just slightly larger than the smallest positive eigenvalue of A are enough
to shift its zero eigenvalues. In practice this means that we can always choose Υ so that conditioning of the
regularized system does not exceed the effective condition number of the consistent singular system.

References
[1] O. Axelsson. Iterative Solution methods. Cambridge University Press, Cambridge, 1994.
[2] Pavel Bochev and R. B. Lehoucq. On the finite element solution of the pure neumann problem. SIAM Review, 47(1):50–66,
2005.
[3] S. Brenner and R. Scott. The mathematical theory of finite element methods. Springer-Verlag, 1994.
[4] M. S. Engelman, R. L. Sani, and P. M. Gresho. The implementation of normal and/or tangential boundary conditions in
finite element codes for incompressible fluid flow. International Journal for Numerical Methods in Fluids, 2(3):225–238,
1982.
[5] C. Farhat and F.-X. Roux. An unconventional domain-decomposition method for an efficient parallel solution of large
scale finite element systems. SIAM J. Sci.Stat. Comp., 13(1):379–396, 1992.
[6] Charbel Farhat and Michel Géradin. On the general solution by a direct method of a large-scale singular system of linear
equations: application to the analysis of floating structures. International Journal on Numerical Methods in Engineering,
41:675–696, 1998.
[7] Morton Gurtin. An introduction to Continuum Mechanics. Academic Press, Boston, MA, 1981.
[8] M. Papadrakakis and Y. Fragakis. An integrated geometric-algebraic method for solving semi-definite problems in structural
mechanics. Computer Methods in Applied Mechanics and Engineering, 190:6513–6532, 2001.
[9] P. Vanek, M. Brezina, and J. Mandel. Convergence of algebraic multigrid based on smoothed aggregation. Numerische
Mathematik, 88:559–579, 2001.
[10] P. Vanek, J. Mandel, and M. Brezina. Smoothed aggregation for second and fourth order elliptic problems. Computing,
65:179–196, 1996.
[11] Andrew Yeckel and Jeffrey J. Derby. On setting a pressure datum when computing incompressible flows. International
Journal for Numerical Methods in Fluids, 29(1):19–34, 1999.