COMMUNICATIONS IN NUMERICAL METHODS IN ENGINEERING, VOl.

9, 889-896 (1993)

COMPUTATION OF ISOTROPIC TENSOR FUNCTIONS

C. MIEHE
Institut fur Baumechanik und Numerische Mechanik, Universitat Hannover, D-3000 Honnover I , Appelstr. 9a,
Germany

SUMMARY
An explicit formulation and a procedure for the computation of isotropic tensor-valued tensor functions
is discussed. The formulation is based on a spectral decomposition in terms of second-order eigenvalue
bases, which avoids the costly computation of eigenvectors. As an important result a compact structure
of the fourth-order derivatives of general second-order isotropic tensor functions is presented.

1 . INTRODUCTION
The paper presents an explicit formulation and an algorithmic procedure for the computation
of isotropic tensor-valued tensor functions &B) of a symmetric positive-definite second-order
tensor B. Simple examples from applied continuum mechanics are the functions exp(B), ln(B)
or (B)* with cy E IR. More complicated functions appear in the context of large strain-elasticity
based on general isotropic strain-energy functions, see for example Ogden and references
therein.
Characteristic of the proposed procedure is the formulation of the spectral decomposition
in terms of second-order eigenvalue-bases associated with the eigenvalues of B, which can be
computed directly from B with low computational effort. This strategy, which avoids the costly
computation of eigenvectors, has been proposed by Betten, Ting3 and M ~ r m a n and , ~ has
been applied to computational finite-strain elasticity by Simo and Taylor.
As a central result we point out a compact structure of the fourth-order derivatives h G ( B )
of general second-order isotropic tensor functions. These fourth-order tensors appear as a
linear combination of fourth-order identity tensors and a simple sum of the dyadics of the
eigenvalue-bases, a structure which has not been pointed out in the literature to the authors
knowledge. This result can be exploited, for example, in the context of the linearization of
general isotropic tensor functions and particularly in the framework of elasticity for the
computation of elasticity moduli for complicated isotropic stress functions. It provides an
alternative to the formulations given in Reference 1, p. 336, in terms of eigenvectors (see also
Chadwick and Ogden6).
In Section 2 we point out fundamental results concerning the spectral decomposition of
second-order tensors and the derivatives of eigenvalues and eigenvalue-bases. Based on these
results we formulate in Section 3 second-order isotropic tensor functions and their fourth-
order derivatives in spectral representation. Finally the algorithmic procedure is summarized
for 3D and 2D tensor functions.

0748-8025/93/ 1 10889-08$09.00 Received 15 March 1992

0 1993 by John Wiley & Sons, Ltd. Revised 18 February 1993
890 C. MIEHE

2. SPECTRAL DECOMPOSITION
We start with a short review of some basic results concerning the spectral decomposition of
symmetric positive-definite second-order tensors from the standard literature of applied matrix
calculus (see, for example, Zurmiihl and Falk* and Strang'). We proceed with the formulation
of the spectral decomposition in terms of second-order eigenvalue-bases and point out their
direct computation without knowledge of the eigenvectors. The main objective of this section
is the derivation of explicit expressions for the derivatives of the eigenvalues and the
eigenvalue-bases.

2.1. Review of basic results

Let B be a second-order tensor which we view as a linear map B: R3 + IR3 defined on the
Euclidian space R 3 with dimension N = 3. We consider the eigenvalue problem
B - nA = XAnA (1)
{ X A ) , A = 1, ...,3, are the eigenvalues of B, i.e. roots of the characteristic polynomial
j ( X ) := det [B - Xl] = X 3 - Z,X2 + ZZX - 1 3 = 0
formulated in terms of the principal in variants (ZA], A = 1, ...,3, of B,
Zl:= tr B, ZZ := [Z: - tr B'], 1 3 := det B
The characteristic polynomial can be solved in closed form by Cardano's formula (see, for
example, Bronstein and Semendyayevlo), and has at least three real solutions if B is symmetric.
In what follows we restrict our attention to (i) symmetric positive-definite tensors, i.e.
- - -
BT= B and XA = n.4 B n A / n A nA E R+\o, with (ii) three distinct eigenvalues XA # X B for
A # B. See Section 3.3 for a discussion of the latter point. Thus the eigenvectors (na),
A = 1 , ..., 3, R3,are linearly independent and form a basis for the R3. Furthermore they are
-
orthogonal, i.e. nA n B = nA * (BT- B) * n B / ( X A - X B ) = 0 for A # B , since B is symmetric. By
scaling the eigenvectors to unit length we obtain
nA nB = 6AB (2)
where ISMI denotes the Kronecker symbol. Introduction of the eigenvector tensor S and the
diagonal tensor A yields, based on a reformulation of (l), the representation
B = S - A - S - ' , S:= [nl, ...,n31, A:=diag[Xl, ..., A31 (3)
From (2) we obtain the result ST.S = 1. Thus we have S-' = ST and therefore S E S0(3),
provided that ( n A ) , A = 1 , ...,3, form a positive-orientated orthonormal system. Here
S0(3):= (Q: IR3 --t R 3 I Q T * Q= 1, det Q = +1) is the proper orthogonal group. We refer to
(3)1 as the spectral decomposition of the symmetric tensor B into the eigenvector tensor S and
the diagonal tensor A.

2.2. Introduction of eigenvalue-bases

Insertion of the definitions for S and A yields with (2) the representation
3
B= 2
A=l
XAMA with @ nA
MA:=~A
 COMPUTATION OF ISOTROPIC TENSOR FUNCTIONS 891

1 . Eigenvalues: Set I I := tr B, Iz := ; ( I : - tr B2), 4 := det B, 0 := c0s-I [f(2I: - 91112+ 2 7 I 3 ) / ( I ?- 31z)3/2]and

compute for A = 1,2,3

AA = $ [II+ 2J(I: - 3h)COS f(0+ 2rA))

IF [I XA - XB I/max(l XA \ , I XB 1, I XC I )I < tola THEN

+ 6)

ENDIF
XA = XA(I
XB=XB(l
xc = k/(l
2. Eigenvalue-bases: Set D A : = ~ ~ = I \ A X( B~) A
-6)
+ 6)(1 - 6) I with perturbation 6 4 1

for- A = 1,2,3 and compute

3. Diagonal functions and derivatives: Compute for A = 1,2,3

4. Tensor-valued isotropic tensor function: Compute

5 . Derivative of tensor-valued isotropic tensor function: Set coefficients

and compute

~~~~~~~~~ ~ ~ ~ ~

Figure 1 . 3D isotropic tensor functions and their derivatives

alternatively to (3) 1. In what follows we refer to the second-order tensors (MA), A = 1, ..., 3,
as the eigenvalue-bases of B associated with the eigenvalues [ XA), A = 1, ..., 3. They have the
properties
3
C
A=l
MA=^, M A : M B = ~M
AAB-
,M B = ~ A B M B
The eigenvalue-bases [MA), A = 1, ..., 3, can be computed directly from B without know-
ledge of the eigenvectors (nAl, A = 1, ..., 3. This observation has been pointed out in
References 2-4. To show this, consider the expression
3
B- XAI= C (XB- XA)MB
B = 1\A
892 C. MIEHE

which follows by using (4)' and ( 5 ) ' . As a result of the condition (5)3 we derive
3
IT
B = l\A
(B- b l ) = [ fl
3

B = l\A
(XA- MA
which can be solved for the eigenvalue-base MA,
3
l 3
MA=-
DAE = I \ A
fl
(B- X B ~ ) with DA:=
E = 1\A
(XA - X E ) (6)

Observe carefully that (6) is valid only for the case of distinct eigenvalues (see Section 3.3 for
a discussion).

2.3. Derivatives of eigenvalues and eigenvalue-bases

The derivative of the eigenvalues XA of the tensor B are given by the fundamental formula
aBhA = MA (7)
which results from a representation of the rate 'XA := BAA :B = nA * B * nA by exploiting
-
the condition n~ nA = 0 and the symmetry of B. With this result in hand and the standard
derivative of the invariants
= 1, dB12 = Z1I - BT, a B 1 3 = Z3B-T

(see, for example, Reference 11, p. 26), we compute the derivative of the eigenvalue-base (6) 1
in a straightforward manner, starting from the representation

MA = &! [B + (AA - Z l ) l + Z3X,i1B-']

DA
We obtain the lengthy expression
~ B M A [ I=I - Z ~j X A 1 O ~ - l ] + - XA [Z3XA1(B-' @ B - ' ) - ( l 0 1 )
DA DA
- 13XA2(B-' @ MA +MA @ B-') + (1 @ MA +MA @ 1)
- (2 - 4 X A 3 ) ( M ~@ MA)]
A conceptually identical result has been given by Simo and Taylor' in the context of large-
strain elasticity. Here ( O ) I J K L = $ [ B I K ~ J L+ B I L ~ J K ] denotes the fourth-order identity tensor
-
and UB - I := B-' I B-' = - i3~B-l a further fourth-order tensor with components
a

(IB- 1)IjKL = [BFiBji! + BE'Bjj!]. The basic idea is now to express the dyadics in the
second bracket of the expression above by dyadics of the known eigenvalue-bases MA. By
inserting the spectral decomposition (4) we end up with the compact representation

This formula gives the fourth-order derivative of the eigenvalue-base MA as a linear

combination of the two fourth-order tensors I and UB- I and a simple sum of the dyadics
MB @ MB of the eigenvalue-bases. A matrix representation of the fourth-order tensors Iand
OB - 1 is given in the Appendix.
 COMPUTATION OF ISOTROPIC TENSOR FUNCTIONS 893

2.4. SimpliJcations for plane problems

We investigate now some particular representations of plane problems. In this case, the
tensor B has the structure
B = b + X3M3 with b = XIMI+ X2M2 (9)
b denotes the in-plane part of B, and X3 is the only out-plane component relative to the given
and fixed eigenvalue-base M3 := e3 @ e3, with unit normal e3 of the plane. The in-plane
eigenvalues ( X ) A = ~ ,can
~ be computed in closed form by solving a simple quadratic equation
(see Figure 2). With a known base M3 we compute the first in-plane baseM~by eliminating the
second in-plane base M2 via the representation of the unit tensor 1 = MI + M2 + M3. Thus (6)
is replaced by the simple expression
- X2)
Mi = [B - A21 - (A3 - X~)M~I/(XI (10)
The derivatives of the in-plane eigenvalue bases can be written after some straightforward
algebraic manipulations based on (7) in the form
3
&Mi = - 8 ~ M 2= ~ IX2 [I.-1 - C XAMA @MA] (1 1)
Xi - Xz A= 1

which simplifies the representation (8). Thus, for plane problems, the fourth-order derivative

1. Eigenvalues: The tensor B = b + X3M3 is given. Set in-plane invariants I1 := tr b/2. I2 := det b and compute
in-plane eigenvalues

X1.2= - 12)
I1 2 J(I:

IF [I XI - X2 I/max(l XI I, I X2 111 < tols THEN

ENDIF
XI = Xi(1 +6)
A2 = X 2 ( 1 - 6)
A3 = h / ( l + 6)(l -6)

2. Eigenvalue bases: Compute in-plane bases

I with perturbation 6 4 1

3. Diagonal functions and derivatives: Compute for A = 1,2,3

gA=&XA) and hA=g(ht)

4. Tensor-valued isotropic tensor function: Compute

5 . Derivative of isotropic tensor function: Set /3 := X 1 X 2 ( g 1 - g2)/(X1- X2), ?A := (hA - /3)X,i2 and compute

aeC(B) = boa- I + 2 YAMC3 MA

Figure 2. 2D isotropic tensor functions and their derivatives

894 C . MIEHE

of the eigenvalue-base MI appears as a linear combination of the fourth-order tensor Be- 1

and a simple sum of the dyadics M B Q MB of the eigenvalue-bases.

3. ISOTROPIC TENSOR FUNCTIONS

To formulate spectral representations in terms of second-order eigenvalue bases for isotropic
tensor-valued tensor functions and their fourth-order derivatives we exploit the results
obtained above in a straightforward manner.

3.1. Tensor-valued isotropic tensor functions

Let e: IR3 x IR3 -,R 3 x IR3 denote a tensor-valued tensor function, i.e. G = G(B). The
function is said to be isotropic if it satisfies the relation
Q C(B) - Q - ~ QQ - B - Q-
= l) VQ E so(3)
Setting Q = S-' we obtain with equation (3)1 the relation
-
S-' . e ( B ) S = C(f3-l - B -S) = e(A), which results in the standard representation
3
G(B) = S * G ( A ).S-' = C ~(XA)MA (12)
A= 1

where 2 :R+\o + R is the diagonal function associated with the tensor function c, i.e. 2 and
G are of the same type.

3.2. Derivative of tensor-valued tensor functions

Let us now derive an explicit expression for the fourth-order derivative of the tensor-valued
isotropic tensor function C(B) by the tensor B. Starting from the spectral representation (12)
we obtain in a straightforward manner by application of the product rule
3 3
d ~ e ( B=) c ~ ' ( X A ) M0A~ B X +A c H X A ) ~ B M A
A= 1 A=l

Insertion of the fundamental results (7) and (8) yields the representation

with coefficients a, /3 and * / A given in Figure 1. Thus the derivative of C(B) by B can be
represented by a linear combination of the two fourth-order tensors 0 and O B - 1 and a simple
sum over the dyadics MA @ MAof the eigenvalue bases. This compact structure has not been
pointed out in the literature to the author's knowledge. The simplified formulation for plane
problems based on the results (10) and (11) is summarized in Figure 2.

3.3. The case of equal or nearly equal eigenvalues

As already mentioned, the formulation above is restricted to the case of distinct eigenvalues
of the tensor B. From the computational point of view it makes sense to reduce cases with
equal or nearly equal eigenvalues to the general case of distinct eigenvalues by application of
a perturbation technique. A possibility based on a machine-dependent tolerance to16 and
 COMPUTATION OF ISOTROPIC TENSOR FUNCTIONS 895

perturbation 6, which has been used by Simo and Taylor, is outlined in Figures 1 and 2, where
the procedure for the computation of 2D and 3D isotropic tensor functions is summarized.

4. CONCLUSION
Algorithms for the computation of 3D and 2D isotropic tensor-valued tensor functions and
their derivatives for symmetric positive-definite tensor arguments have been discussed. The
formulation is based on a spectral decomposition in terms of second-order eigenvalue bases
which avoids the costly computation of eigenvectors. The important results are compact
representations of the fourth-order derivatives of general second-order isotropic tensor
functions.

APPENDIX. STORAGE OF FOURTH-ORDER TENSORS

To obtain an impression of the computational effort for the implementation of the fourth-
order tensors 0 and BB in (8) and (13) we consider a matrix formulation used in computational
elasticity where B is associated with a strain tensor and G(B) with the elasticity moduli.
Assume that the symmetric tensor B is stored in a vector b as follows:
B := [bi b2 b3 b4 b5 b d T = tBii &2 B33 Biz B23 BidT
-
Consequently one stores the fourth-order tensors I and O B := B 0 B in matrices 0 and -OB. For
example, if one demands the invariance conditions
E:I :E= eTae and E:88:E = eTltse
for a symmetric tensor E stored in the particular vector form
e:= [el ez e3 e4 e5 eelT= [El1 E22 E33 2E12 2E23 2Ei3IT
(which is common in computational elasticity where E is associated with the symmetric part
of the displacement gradient), one obtains the matrix representation
I=diag[l
- 1 1 ) f $1

REFERENCES
1 . R. W. Ogden, Non-Linear Elastic Deformations, Ellis Horwood, Chichester, 1984.
2. J. Betten, Interpolation methods for tensor functions, in X. J. R. Avula et al. (Ed.), Mathematical
Modelling in Science and Technology, Pergamon Press, New York, 1984, pp. 52-57.
3. T. C. T. Ting, Determination of C, C- and more general isotropic tensor functions of C, J.
Elasticity, 15, 319-323 (1985).
896 C . MIEHE

4. K. N. Morman, The generalized strain measure with application to nonhomogeneous deformations

in rubber-like solids, J. Appl. Mech., 53, 726-728 (1986).
5 . J. C. Simo and R. L. Taylor, Quasi-incompressible finite elasticity in principal stretches. Continuum
basis and numerical algorithms, Comput. Methods Appl. Mech. Eng., 85, 273-310 (1991).
6. P. Chadwick and R. W. Ogden, On the definition of elastic moduli, Arch. Rat. Mech. Anal., 44,
41-53 (1971).
7. P. Chadwick and R. W. Ogden, A theorem of tensor calculus and its application to isotropic
elasticity, Arch. Rat. Mech. Anal., 44, 54-68 (1971).
8. R. Zurmuhl and S. Falk, Mutrizen und ihre Anwendungen, Part 1, Springer Verlag, Berlin,
Heidelberg, New York, 1984.
9. G. Strang, Linear Algebra and its Applications, 3 edn, Harcourt Brace Jovanovich, Publishers, San
Diego, 1988.
10. I. N. Bronstein and S. Semendyayev, Handbook of Mathematics, Van Nostrand Co., 1985.
11. C. Truesdell, and W. Noll, The Nonlinear Field Theories of Mechanics, in S . Flugge (Ed.),
Handbuch der Physik Bd. M / 3 , Springer-Verlag, Berlin, 1965.