ZAMM · Z. Angew. Math. Mech. 86, No. 4, 291 – 311 (2006) / DOI 10.1002/zamm.

200410242

Fourth-order tensors – tensor differentiation

with applications to continuum mechanics.
Part I: Classical tensor analysis
O. Kintzel∗ and Y. Başar∗∗
1
Institute for Structural Mechanics, Ruhr University Bochum, Bochum, Germany

Received 29 May 2004, revised 24 April 2005, accepted 27 June 2005

Published online 19 October 2005

Key words fourth-order tensors, tensor differentiation, linearization, isotropic tensor functions, exponential function
MSC (2000) 03E25
The present contribution provides a tensor formalism for fourth-order tensors in the so-called absolute notation and focusses
in particular on the use of this notation in the process of tensor differentiation with respect to a second-order tensor. Three
tensor products, two new double contraction rules and a set of well-defined notations are introduced which in combination
with the tensor differentiation rules simplify analytical derivation procedures considerably and provide significant advan-
tages for various tasks in continuum mechanics. The suitability of the proposed rules and definitions is demonstrated in a
number of relevant problems of continuum mechanics such as linearization of the generalized midpoint-rule and the ex-
ponential function. Special attention is given to the differentiation with respect to symmetric, skew-symmetric and inverse
second-order tensors.


c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

1 Introduction
Fourth-order tensors are widely used in continuum mechanics in particular for the formulation of constitutive models. A
particular problem is to represent a fourth-order tensor and its properties by a set of well-defined notations in absolute form.
Tensor algebra in the so-called absolute notation is increasingly used in continuum mechanics [23, 3, 21, 9] in contrast
to classical formulations in component form [8, 2, 4, 22] since it allows a compact representation of certain physical and
geometric phenomena by means of invariant quantities. In this context invariant is to be understood as independent from
any special coordinate system. In spite of considerable advances in the last two decades [10, 9, 18, 17, 7, 16] there are many
aspects still requiring further improvements or suitable extensions of the existing theoretical fundamentals in the field of
absolute notation of tensors.
The differentiation of a tensor function with respect to its argument tensors plays an important role in continuum
mechanics especially for the linearization of constitutive models. Following an approach of Itskov [10] the process
of tensor differentiation can be carried out advantageously in absolute tensor notation and thereby simplifies and accelerates
analytical derivation procedures considerably in various applications. This formalism is especially suited to the differentiation
of isotropic tensor functions e.g. in the form of power series. However, in our view, the tensor formalism used in [10], which
had been proposed in [9], was not general enough to exploit all aspects of the differentiation rules. By introducing three
kinds of tensor products and three types of double contraction laws, we have been able to remove some of these restrictions.
Furthermore, the differentiation with respect to symmetric and skew-symmetric tensors (the latter has not been considered
in [10]) is discussed in detail. To avoid ambiguity with the well-established double contraction law (:) the two further
contraction rules, which are in particular advantageous for tensor differentiation, are introduced with the notations ( q a) and
( a q).
This contribution is divided into two parts. In this part a mathematical framework is used referred to as classical tensor
analysis. However, dealing with large strain problems many authors employ a set of new operations like push-forward-,
pull-back-operations or Lie-derivatives which are typically borrowed from tensor analysis on manifolds. These operations
are often considered within the scope of classical tensor analysis solely [3] which, in certain points, may cause confusion.
E.g. in classical tensor analysis the component decomposition of an invariant tensor is treated as irrelevant. Such a distinction
of component variance, however, is vitally important to allow a unique application of push- or pull-operations. Also, as has
∗ Corresponding author, E-mail: [email protected]; Address: IA 6, Universitätsstraße 150, 44780 Bochum, Germany, Web-page:

http://www.sd.rub.de
∗∗ In dear memory of Prof. Dr.-Ing. Yavuz Başar who died on 30th of August, 2002

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
292 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

been stated in [10], the well-known Doyle-Ericksen-formula [19, 24] loses its meaning. To avoid such inconsistencies
the mathematical framework will be broadened in part II by introducing certain concepts from tensor analysis on manifolds
[14].
Part I is divided into eight sections in the following manner: In the second section some frequently used notations are
introduced. In the third section the definitions of tensor products and contraction rules are given. The fourth section is con-
cerned with transposition operations for fourth-order tensors which will be employed in the fifth section, in which symmetry
properties of fourth-order tensors are discussed. The following section is devoted to the process of basis rearrangement as
prerequisite to the seventh section, in which rules concerning the tensor differentiation are introduced and are used e.g. to
construct partial derivatives of tensor powers of arbitrary order. Particular attention is given to the differentiation with respect
to symmetric, skew-symmetric and inverse second-order tensors. Finally, in the eighth section some examples are presented
demonstrating the suitability of the new rules and definitions.

2 Notations, definitions
Notations. In this section some notations are summarized which will be used in the sequel. Use is made of curvilinear
coordinate systems with co-variant and contra-variant base vectors gi and gj such that the relations
gi · gj = δji , gi = g ij gj , gi = gij gj , g ik gkj = δji (1)
are satisfied with g ij = gi · gj and gij = gi · gj as so-called metric tensor components. The basis of each tensor is formed by
a dyadic product of independent base vectors denoted by (⊗). The Einstein summation convention over repeated indices
is adopted. Vectors are represented by bold lower-case letters
a = ai gi , b = bi gi , ... . (2)
Second-order tensors are denoted by bold upper-case letters and for a clear distinction fourth-order tensors are introduced
by blackboard upper-case letters, thus
A = Aij gi ⊗ gj , B = Bi j. gi ⊗ gj , ... (3)
D = Dijmn g ⊗ g ⊗ g ⊗ g , E = E
i j m n ijmn
gi ⊗ gj ⊗ gm ⊗ gn , ... . (4)
By definition, the kind of component variance of an invariant tensor is arbitrary. For instance, a second-order tensor A
may be given in the following equivalent representations:
A = Aij gi ⊗ gj = Ai j. gi ⊗ gj = Ai. j gi ⊗ gj = Aij gi ⊗ gj , (5)
where the components are obtained as follows:
Aij = gi · A · gj , Ai j. = gi · A · gj , Ai. j = gi · A · gj , Aij = gi · A · gj . (6)
For tensors defined in one specific vector space this is unproblematic. Situations, in which a tensor may belong to different
vector spaces or may be related to different configurations of a body, are considered in part II.
The simple contraction of tensors and the double contraction (notation::) which are well established in literature [14, 3]
are defined as usual:
simple contraction:
A B C = Ai .j B j .m C mn gi ⊗ gn , (7)
E A = E ijmn An .r gi ⊗ gj ⊗ gm ⊗ gr , (8)
double contraction:
E : A = E ijmn Amn gi ⊗ gj , (9)
A : E = Aij E ijmn gm ⊗ gn . (10)
The transpose of a second-order tensor A (notation: (. . . )T ) is defined by:
a · (A b) = b · (AT a) , (11)
where a and b are arbitrary vectors. The inverse of a second-order tensor (notation: (. . . )−1 ) is, finally, defined by
A−1 A = A A−1 = I , (12)
where I = gi ⊗ gi = gi ⊗ gi = gij gi ⊗ gj = g ij gi ⊗ gj is the second-order identity tensor.

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 293

3 Fourth-order tensors
Tensor products. Besides the usual dyadic product (⊗) two further tensor products for second-order tensors, (×) and ( ×
2 ),
are introduced. The tensor product (×) has been used in the same form by Itskov [9]. The tensor product ( × 2 ) dates back
to a paper of del Piero [7] and has been also recently used by Rosati [18]:

D = A ⊗ B = Aij B mn gi ⊗ gj ⊗ gm ⊗ gn , (13)
E=A×B=A B ij mn
gi ⊗ g m ⊗ g n ⊗ g j , (14)
F=A2
× B = Aij B mn gi ⊗ gm ⊗ gj ⊗ gn . (15)

As can be seen from the definitions (13) to (15) the base vectors are arranged in three different forms. The introduced tensor
products (◦ = ⊗, ×, 2× ) satisfy the

the associative law: α(A ◦ B) = (αA) ◦ B = A ◦ (αB) , (16)

and

the distributive law: A ◦ (B + C) = A ◦ B + A ◦ C , (17)

where α is a scalar-valued variable. Note that in general the commutative rule does not hold: A ◦ B = B ◦ A.
Double contraction rules. Let us consider a fourth-order tensor

E = E ijmn gi ⊗ gj ⊗ gm ⊗ gn . (18)

For a clear explanation of the following rules we call the basis gj ⊗ gm placed at the center the inner basis while the base
vectors gi and gn placed outside are said to form the outer basis. Similarly, the bases gi ⊗ gj and gm ⊗ gn are called left-
and right-hand side basis, respectively.
For the double contraction of two fourth-order tensors we introduce three different operations:

  
E : F = E ijmn gi ⊗ gj ⊗ gm ⊗ gn : Fklrs gk ⊗ gl ⊗ gr ⊗ gs
= E ijmn Fmnrs gi ⊗ gj ⊗ gr ⊗ gs , (19)

  
E q aF = E ijmn gi ⊗ gj ⊗ gm ⊗ gn q a Fklrs gk ⊗ gl ⊗ gr ⊗ gs

= E ijmn Fjlrm gi ⊗ gl ⊗ gr ⊗ gn , (20)

  
E a qF = E ijmn gi ⊗ gj ⊗ gm ⊗ gn a q Fklrs gk ⊗ gl ⊗ gr ⊗ gs

= E ijmn Fkins gk ⊗ gj ⊗ gm ⊗ gs . (21)

All contraction rules are defined such that all base vectors which are not involved in the contraction process appear in the
resulting fourth-order tensor at the same position as before. This requirement is important in order to fullfill the associative
law (see (30) and (32)). The first rule (19) is the generalization of the double contraction rule (9), (10) to a combination of
fourth-order tensors. The contraction rules (20) and (21) are new and will play an important role in the context of tensor
differentiation (see Sect. 7). It becomes immediately clear from the notation how the double contraction rules are applied e.g.
for ( q a) the left filled circle ( q ) refers to the inner basis of the fourth-order tensor at the left-hand side and the right unfilled
circle ( a ) refers to the outer basis of the fourth-order tensor at the right-hand side which are contracted and analogously for
the contraction rule ( a q). For tensors of second order, the distinction between outer and inner bases is irrelevant. Accordingly,
the operations ( q a) and ( a q), if applied to a combination of tensors of fourth- and second-order, lead in view of the above
definitions to the following results:


E q aC = E ijmn gi ⊗ gj ⊗ gm ⊗ gn q a (Crs gr ⊗ gs )
= E ijmn Cjm gi ⊗ gn , (22)


C q aE = (Crs gr ⊗ gs ) q a E ijmn gi ⊗ gj ⊗ gm ⊗ gn
= Cin E ijmn gj ⊗ gm , (23)

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
294 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis


E a qC = E ijmn gi ⊗ gj ⊗ gm ⊗ gn a q (Crs gr ⊗ gs )
= E ijmn Cin gj ⊗ gm = C q aE , (24)


C a qE = (Crs gr ⊗ gs ) a q E ijmn gi ⊗ gj ⊗ gm ⊗ gn
= Cjm E ijmn gi ⊗ gn = E q aC . (25)
Finally, by considering a double contraction of second-order tensors the contraction rules (:), ( q a) and ( a q) produce the same
result:
A : B = A q aB = A a qB = Aij Bij . (26)
Identity tensors. We introduce the following identity tensors of fourth-order:
I = I ⊗ I = gi ⊗ gi ⊗ gm ⊗ gm , (27)
I =I×
R
2 I = gi ⊗ gm ⊗ gi ⊗ gm , (28)
I = I × I = gi ⊗ gm ⊗ gm ⊗ gi = gi ⊗ I ⊗ gi .
L
(29)
R L
Note that the notations (. . . ) and (. . . ) used above denote operations of basis rearrangement which will be defined in
Sect. 6.
Proposition 3.1. The set of fourth-order tensors is a ring (G, +, ∗) (see [1] for background information) i.e. it is a group
(G, +) with respect to the sum operation and for each operation (∗ = :, q a, a q) the associative and the distributive law are
fulfilled. Furthermore, for each operation (∗) there exists a unique identity element.
In case of (∗ = :) the following properties hold:
• The associative law holds
(A : B) : C = A : (B : C) . (30)

• The identity element is I = I 2

R
× I so that

IR : B = (I 2
× I) : B = B : I = B : (I 2
R
× I) = B . (31)
In case of (∗ = q a, a q) the following properties hold:
• The associative law holds
(A q aB) q aC = A q a(B q aC) ,
(32)
(A a qB) a qC = A a q(B a qC) .

• The identity element is I = I ⊗ I so that

I q aB = (I ⊗ I) q aB = B a q(I ⊗ I) = B a qI = B ,
(33)
I a qB = (I ⊗ I) a qB = B q a(I ⊗ I) = B q aI = B .
For all operations the distributive law is fulfilled:
(A + B) : C = A : C + B : C ,
(A + B) q aC = A q aC + B q aC ,
(A + B) a qC = A a qC + B a qC . (34)
Tables 1–3 summarize important identities obtained by means of the above defined rules. Table 1 is concerned with
simple contractions of fourth- and second-order tensors. Tables 2 and 3 contain results obtained by double contractions of
fourth-order tensors with second- and fourth-order tensors, respectively.
Using Table 2 we may state the following:
(I ⊗ I) : A = (tr A) I , (I ⊗ I) q aA = A , (I ⊗ I) a qA = A , (35)
× I) : A = A ,
(I 2 × I) q aA = A ,
(I 2 T × I) a qA = A ,
(I 2 T
(36)
(I × I) : A = AT , (I × I) q aA = (tr A) I , (I × I) a qA = (tr A) I . (37)

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 295

Table 1 Simple contractions of fourth- and second-order tensors.

(A ⊗ B) C = A ⊗ (B C) C (A ⊗ B) = (C A) ⊗ B

(A × B) C = (A C) × B C (A × B) = (C A) × B
× B) C = A 2
(A 2 × (B C) × B) = (C A) 2
C (A 2 ×B

Table 2 Double contractions of fourth- and second-order tensors.

(A ⊗ B) : C = A (B : C) C : (A ⊗ B) = (C : A) B

(A ⊗ B) q aC = A C B C q a(A ⊗ B) = AT C BT

(A ⊗ B) a qC = AT C BT C a q(A ⊗ B) = A C B

(A × B) : C = A CT BT C : (A × B) = BT CT A

(A × B) q aC = (B : C) A C q a(A × B) = (A : C) B

(A × B) a qC = (A : C) B C a q(A × B) = (C : B) A

(A ×
2 B) : C = A C BT C : (A ×
2 B) = AT C B

× B) q aC = A CT B
(A 2 C q a(A 2
× B) = B CT A

× B) a qC = B CT A
(A 2 C a q(A 2
× B) = A CT B

4 Transpositions of fourth-order tensors

Transposition operations. For a tensor of fourth-order E = E
ijmn
gi ⊗ gj ⊗ gm ⊗ gn two sets of transposition operations
are introduced. Transposition operations leading to a transposition of inner and outer bases are defined in the form:

T
ET = E ijmn gi ⊗ gj ⊗ gm ⊗ gn = E ijmn gj ⊗ gi ⊗ gn ⊗ gm
= E jinm gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn ti
E = E
ti
gi ⊗ g j ⊗ gm ⊗ g n = E ijmn gi ⊗ gm ⊗ gj ⊗ gn
= E imjn gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn to (38)
E = E
to
gi ⊗ g j ⊗ gm ⊗ g n = E ijmn gn ⊗ gj ⊗ gm ⊗ gi
= E njmi gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn t
E = E
t
gi ⊗ g j ⊗ gm ⊗ g n = E ijmn gn ⊗ gm ⊗ gj ⊗ gi
(E )to = (E )ti
ti to
= = E nmji gi ⊗ gj ⊗ gm ⊗ gn .
Similarly, transposition operations leading to a transposition of left- and right-hand side bases are introduced in the
following form:

D
ED = E ijmn gi ⊗ gj ⊗ gm ⊗ gn = E ijmn gm ⊗ gn ⊗ gi ⊗ gj
= E mnij gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn dl
E dl
= E gi ⊗ g j ⊗ g m ⊗ g n = E ijmn gj ⊗ gi ⊗ gm ⊗ gn
= E jimn gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn dr (39)
E dr
= E gi ⊗ g j ⊗ g m ⊗ g n = E ijmn gi ⊗ gj ⊗ gn ⊗ gm
= E ijnm gi ⊗ gj ⊗ gm ⊗ gn ,

ijmn d
E d
= E gi ⊗ g j ⊗ g m ⊗ g n = E ijmn gj ⊗ gi ⊗ gn ⊗ gm
(E )dr = (E )dl
dl dr
= = E jinm gi ⊗ gj ⊗ gm ⊗ gn .

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
296 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

Table 3 Double contractions of fourth-order tensors.

(A ⊗ B) : (C ⊗ D) = (B : C) A ⊗ D

(A ⊗ B) q a(C ⊗ D) = (A C) ⊗ (D B)

(A ⊗ B) a q(C ⊗ D) = (C A) ⊗ (B D)

(A ⊗ B) : (C × D) = A ⊗ (DT BT C) (A × B) : (C ⊗ D) = (A CT BT ) ⊗ D

(A ⊗ B) q a(C × D) = (A C B) × D (A × B) q a(C ⊗ D) = A × (CT B DT )

(A ⊗ B) a q(C × D) = C × (AT D BT ) (A × B) a q(C ⊗ D) = (C AD) × B

(A × B) : (C × D) = (A D) ×
2 (B C)

(A × B) q a(C × D) = (B : C) A × D

(A × B) a q(C × D) = (A : D) C × B

× D) = A ⊗ (CT B D)
(A ⊗ B) : (C 2 × B) : (C ⊗ D) = (A C BT ) ⊗ D
(A 2

(A ⊗ B) q a(C 2
× D) = (A C) 2
× (D B) × B) q a(C ⊗ D) = (A DT ) 2
(A 2 × (CT B)

(A ⊗ B) a q(C 2
× D) = (C BT ) 2
× (AT D) × B) a q(C ⊗ D) = (C A) 2
(A 2 × (B D)

× B) : (C 2
(A 2 × D) = (A C) 2
× (B D)

2 B) q a(C ×
(A × 2 D) = (A DT ) ⊗ (CT B)
× B) a q(C 2
(A 2 × D) = (C BT ) ⊗ (AT D)

× D) = (A D) × (B C)
(A × B) : (C 2 × B) : (C × D) = (A C) × (B D)
(A 2

(A × B) q a(C 2
× D) = A × (D BT C) × B) q a(C × D) = (A CT B) × D
(A 2

(A × B) a q(C ×
2 D) = (C AT D) × B 2 B) a q(C × D) = C × (B DT A)
(A ×

According to (38), the operation (. . . )T interchanges inner and outer bases. Through (. . . )ti and (. . . )to the position of
base vectors within the inner and outer bases, respectively, is changed. Finally, (. . . )t is based on a successive application
of (. . . )ti and (. . . )to , where the order of application is irrelevant. The application of the definitions (38) to a fourth-order
tensor A × B yields

(A × B)T = B × A, (A × B)ti = A × BT , (A × B)to = AT × B . (40)

Along the same lines, if we consider a fourth-order tensor constructed by means of the tensor product (⊗), the transposition
operations (39) yield:

(A ⊗ B)D = B ⊗ A, (A ⊗ B)dl = AT ⊗ B, (A ⊗ B)dr = A ⊗ BT . (41)

Both kinds of transposition operations are related by:

ED = ET t = Et T , Ed = ET , ED d = Ed D = Et . (42)

The results obtained for an application of the transposition operations (38) and (39) to a fourth-order tensor, which is
formed by means of the tensor products (13) to (15), are given in Table 4. The application of the transposition rules to a double

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 297

Table 4 Transposition operations for fourth-order tensors. Table 5 Transposition operations for double contractions of
fourth-order tensors.

(E : F)T = E : F (E : F)D = F :E
d d D D
(A ⊗ B)T = AT ⊗ BT (A ⊗ B)D = B ⊗ A

(E q aF)T = F q aE (E q aF)D = F q aEto T

T T ti T
(A × B)T = B × A (A × B)D = BT × AT
(E a qF)T = F a qE (E a qF)D = F a qEti T
T T to T
× B)T = B 2
(A 2 ×A × B)D = AT 2
(A 2 × BT

(E : F)dl = E :F
dl
×B
(A ⊗ B)ti = A 2 (A ⊗ B)dl = AT ⊗ B

(E q aF)ti = E q aF
ti
(A × B)ti = A × BT ×A
(A × B)dl = B 2
(E a qF)ti = E a qF
ti
(A ×
2 B)ti = A ⊗ B (A ×
2 B)dl = B × A

(E : F)dr = E : F
dr
× AT
(A ⊗ B)to = BT 2 (A ⊗ B)dr = A ⊗ BT

(E q aF)to = E q aF
to
(A × B)to = AT × B ×B
(A × B)dr = A 2
(E a qF)to = E a qF
to
× B)to = BT ⊗ AT
(A 2 × B)dr = A × B
(A 2

(E : F)t = F :E (E : F)d = E : F
Dd Dd d d
(A ⊗ B)t = BT ⊗ AT (A ⊗ B)d = AT ⊗ BT

(E q aF)t = E q aF (E q aF)d = F q aE
t t T T
(A × B)t = AT × BT (A × B)d = B × A
(E a qF)t = E a qF (E a qF)d = F a qE
t t T T
× B)t = BT 2
(A 2 × AT × B)d = B 2
(A 2 ×A

contraction of two fourth-order tensors is demonstrated in Table 5. If more than two fourth-order tensors are involved in
contraction the corresponding result can be obtained by a successive application of the rules given in Table 5. Finally, Table 6
points out some important connections between various transposition operations for a double application to a fourth-order
tensor. It will have become clear from the definitions (38) and (39) that a twofold application of the same transposition
operation to a fourth-order tensor is just the identity operation. Therefore these combinations are not listed in Table 6.

5 Symmetry properties
Symmetry properties. Symmetry properties of fourth-order tensors can be defined, similar to those ones of second-order
tensors, using the transposition operations introduced in (38) or, alternatively, in (39). Thus we define

• major symmetry of a fourth-order tensor E:

E = ET ,
(43)
A × B = (A × B)T = B × A ,

• minor symmetries of a fourth-order tensor E:

E = Eti , E = Eto ,
(44)
A × B = (A × B)ti = A × BT , A × B = (A × B)to = AT × B ,
or, alternatively,

• major symmetry of a fourth-order tensor E:

E = ED ,
(45)
A ⊗ B = (A ⊗ B)D = B ⊗ A ,

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
298 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

Table 6 Connections between transposition operations (E = E ).

T d

ET ti = Eto T ED dl = Edr D Eti D = ED to = Ed ti = Eto d

ET to = Eti T ED dr = Edl D ED ti = Eto D = Ed to = Eti d

Et T = ET t = ED ED d = Ed D = Et Et dl = Edr t Et dr = Edl t

Eti to = Eto ti = Et Edl dr = Edr dl = Ed Et D = ED t Et d = Ed t

ET dl = Edl T = Edr ET dr = Edr T = Edl ET D = ED T ET d = Ed T = E

• minor symmetries of a fourth-order tensor E:

E = Edl , E = Edr , (46)

A ⊗ B = (A ⊗ B)dl = AT ⊗ B , A ⊗ B = (A ⊗ B)dr = A ⊗ BT .
Note that each set of transposition operations leads to a specific symmetry definition ((43) and (44)) or ((45) and (46))
advantageously related to fourth-order tensors of the type A × B or A ⊗ B.
A fourth-order tensor E = E ijmn gi ⊗ gj ⊗ gm ⊗ gn is said to be supersymmetric, if

E = Eti , E = Eto , E = ET ,
E ijmn = E imjn , E ijmn = E njmi , E ijmn = E jinm , (47)

or, alternatively, if

E = Edl , E = Edr , E = ED ,
E ijmn = E jimn , E ijmn = E ijnm , E ijmn = E mnij . (48)

A supersymmetric tensor E = E ijkl gi ⊗ gj ⊗ gk ⊗ gl may have 21 independent components which can be presented
for tensors characterized by symmetry properties defined by means of (47) and (48) in the following forms [8]:

E 1111 E 1221 E 1331 E 1231 E 1131 E 1121

E 2222 E 2332 E 2232 E 2132 E 2122
E 3333 E 3233 E 3133 E 3123
(49)
E 2233 E 2133 E 2123
E 1133 E 1123
E 1122

E 1111 E 1122 E 1133 E 1123 E 1113 E 1112

E 2222 E 2233 E 2223 E 2213 E 2212
E 3333 E 3323 E 3313 E 3312
(50)
E 2323 E 2313 E 2312
E 1313 E 1312
E 1212

By considering the definitions (13), (15), and (38) the identity tensor I (27) can be split up into a symmetric S and a
skew-symmetric part A with respect to the inner basis in the form
1 1
I = I ⊗ I = (I + (I)ti ) + (I − (I)ti )
2 2

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 299

1 1
= (I ⊗ I + I 2 × I) = S + A .
× I) + (I ⊗ I − I 2 (51)
2
   2  
S A
Similar holds also, in view of (39), for the tensor I introduced in (28), thus
R

1 1
IR = I 2
× I = (I + (I )dr ) + (I − (I )dr )
R R R R
2 2
1 1
× I − I × I) = S + A ,
R R
× I + I × I) + (I 2
= (I 2 (52)
2
    2  
SR AR
where S and A are the symmetric and skew-symmetric parts of I with respect to the right- or left-hand side bases. It
R R R

then follows for a second-order tensor C:

1 1 1
S q aC = C a qS = (C + CT ) , S a qC = C q aS = (C + CT ) , SR : C = C : SR = (C + CT ) , (53)
2 2 2
1 1 1
A q aC = C a qA = (C − CT ) , A a qC = C q aA = (C − CT ) , AR : C = C : AR = (C − CT ) .(54)
2 2 2
The above relations point out that, in case of double contractions of the type ( q a), ( a q) and (:), the symmetry or skew-symmetry
of a second-order tensor C can be enforced by means of the fourth-order tensors defined in (51) and (52).
The tensors S and A are especially useful in the context of tensor differentiation for the construction of derivatives with
respect to symmetric or skew-symmetric tensors, respectively, by means of rule (65) which will be demonstrated in Sect. 7.

6 Process of basis rearrangement

Process of basis rearrangement. In close relation to the two sets of double contraction rules ( q a), ( a q), and (:) the derivative
of a second-order tensor with respect to another second-order tensor can be arranged in absolute notation in two different
forms which will be proposed in Sect. 7. In order to transform one form into the other, and vice versa, use is made of
the so-called basis rearrangement operations (. . . )R and (. . . )L . The process of basis rearrangement is understood as an
operation where the position of a single base vector is rearranged within a fourth-order tensor. For instance, the operation
(. . . )R moves the last base vector to a position after the first base vector. Conversely, the second base vector is moved to a
position after the last one by means of the operation (. . . )L , thus

ER = (E ijmn gi ⊗ gj ⊗ gm ⊗ gn )R = E ijmn gi ⊗ gn ⊗ gj ⊗ gm
= E imnj gi ⊗ gj ⊗ gm ⊗ gn , (55)

E = (E ijmn gi ⊗ gj ⊗ gm ⊗ gn )L = E ijmn gi ⊗ gm ⊗ gn ⊗ gj
L

= E injm gi ⊗ gj ⊗ gm ⊗ gn . (56)

As is evident from the definitions (55) and (56) the successive application of (. . . )R and (. . . )L is just the identity operation:
ER L = EL R = E, while this is of course not fullfilled for a twofold application of the same operation
ER R = E , EL L = E , (57)

in contrast to any transposition rule. Applied to a fourth-order tensor constructed by means of the tensor products (13) to
(15) the following relations hold:

(A ⊗ B)L = A × B , × BT ,
(A × B)L = A 2 × B)L = A ⊗ BT ,
(A 2
× BT ,
(A ⊗ B)R = A 2 (A × B)R = A ⊗ B , × B)R = A × BT .
(A 2 (58)
Again, in view of (55) and (56) the double contraction rules (:), ( q a) and ( a q) can be transformed into each other as follows

E : A = EL q aA = A a qEL , E q aA = A a qE = ER : A , (59)
(A : B)L = A q aB = B a qA , (A q aB)R = (B a qA)R = A : B .
L L L L R R
(60)

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
300 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

7 Tensor differentiation
Introduction. The tensor differentiation which is understood as the differentiation of an arbitrary tensor function with respect
to another tensor plays a very important role in continuum mechanics, e.g. for the differentiation of isotropic tensor functions,
the linearization of nonlinear functionals, the construction of tangent operators. If a tensorial expression in absolute notation
is expressed in component form its differentiation with respect to a set of tensor components can be performed by means of
the well-known rules of differential calculus and causes in this sense no particular difficulties. The aim in this section is to
provide such rules and definitions to allow the differentiation exclusively in absolute tensor notation. Since the derivative of
a second-order tensor with respect to a tensor of the same order is naturally a fourth-order tensor the rules to be introduced
are closely related to the theory of fourth-order tensors discussed in detail in the previous sections. One obvious requirement
imposed on the differentiation rules to be introduced is that the results (in component form) coincide with the solution
directly obtainable from the corresponding component relation. Another requirement is to satisfy, in addition, the product
rule of differential calculus for operations occurring most frequently in derivation procedures. In the present formulation
this will be achieved for simple contractions of second-order tensors being particularly relevant for tensor functions in form
of power series. To the best knowledge of the authors the above mentioned problem has been solved for the first time by
Itskov [9, 10]. However, in some respects the original contribution [9] was not satisfying. For instance, Itskov redefined
the usual double contraction rule (:) which, of course, simplified notation but could lead to some confusion in applications.
To improve in particular this aspect we decided to define two new double contraction rules (20) and (21) from the outset.
Gâteaux derivative. The directional derivative of a scalar-valued function f (a) with respect to its argument vector a in
the direction of an arbitrary vector h is, under the assumption that f (a) is continuously differentiable around f (a), defined
by (see [5]):
f (a + s h) − f (a) d
f  (a) h = lim = f (a + s h) |s=0 . (61)
s→0 s ds
We call f  (a) h the Gâteux derivative of f in the direction of h.
Scalar-valued function of a second-order tensor. The first variation ∆f of a scalar-valued tensor function f , which depends
on a second-order-valued tensor function A and is linearized around f (A) = 0, can be computed by means of the above
introduced Gâteaux derivative (61) which reads as
d d ∂f
∆f = f (A + s ∆A) |s=0 = f (Ai j. + s ∆Ai j. ) |s=0 = ∆Ai j. . (62)
ds ds ∂Ai j.
It is obvious that by raising or lowering the indices of Ai j. and ∆Ai j. with respect to the given curvilinear coordinate system
the first variation can be also given in the following representations:
∂f ∂f ∂f ∂f
∆f = j
∆Ai j. = ∆Aij = ij
∆Aij = ∆Ai. j = f,A : ∆A , (63)
∂Ai . ∂Aij ∂A ∂Ai. j
which, finally, justifies the transformation of the component relation (62) into absolute tensor notation in the form f,A : ∆A.
Accordingly, the partial derivative f,A reads as:
∂f ∂f ∂f ∂f ∂f
f,A = = j
gi ⊗ g j = gi ⊗ g j = ij
gi ⊗ gj = g i ⊗ gj , (64)
∂A ∂Ai . ∂Aij ∂A ∂Ai. j
i.e. f is differentiated with respect to the components of its argument tensor A and a basis is finally attached to it.
Second-order-valued tensor function of a second-order tensor. The directional and partial derivative of a second-order
tensor function A(C) with respect to a second-order tensor C is defined as follows [10]:
d
A = A (C) : ∆A = A(C + s∆C) |s=0 = A,C q a∆C ,
ds

∂A ∂Aij i
A,C == g ⊗ g m ⊗ gn ⊗ g j . (65)
∂C ∂Cmn
A,C can be also given in related forms by raising or lowering indices. As can be observed from (65) the basis of C is
placed at the center of A,C . This definition has the advantage that the product rule of differential calculus is satisfied for
a tensor function involving simple contractions of second-order tensors (e.g. polynomials), since for a second-order tensor
F(D) = AB(C(D)) the derivative with respect to D reads as:
F,D = (AB),C q aC,D = (A,C B + AB,C ) q aC,D . (66)

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 301

Remark 7.1. In many works [23, 3] the partial derivative of a second-order tensor function A(C) with respect to a
second-order tensor C is defined by:

∂Aij i
A;C = g ⊗ g j ⊗ g m ⊗ gn , (67)
∂Cmn
and is, therefore, related to the present definition (65) by means of the basis rearrangement operations (55) and (56):
R L
A;C = (A,C ) , A,C = (A;C ) . (68)

However, the product rule of differential calculus similar to (66) is not satisfied for this arrangement (in combination
with the contraction rule (:)).
Differentiation rules. To summarize, the definitions stated above are reiterated for the sake of clarity. Partial derivatives
with respect to a second-order tensor C are defined as follows

• for a scalar-valued function

∂ψ ∂ψ
ψ = ψ(C) : ψ,C = = gi ⊗ g j ,
∂C ∂Cij
∂2ψ ∂2ψ
ψ,(C×C) = = gi ⊗ g k ⊗ g l ⊗ g j , (69)
∂C∂C ∂C ∂C
2
R ij kl2
∂ ψ ∂ ψ
ψ;(C⊗C) = = gi ⊗ g j ⊗ g k ⊗ g l ,
∂C∂C ∂Cij ∂Ckl
• for a second-order tensor function
∂A ∂Aij i
A = A(C) : A,C = = g ⊗ gm ⊗ gn ⊗ g j ,
∂C
∂C

R mn
(70)
∂A ∂Aij i
A;C = = g ⊗ g j ⊗ gm ⊗ g n .
∂C ∂Cmn
In accordance with the above definitions and the rules (14) and (19) to (21) the following differentiation rules hold:

• for a product of a second-order tensor A with a scalar-valued function ψ

D = ψ (C) A(C) :
∂(ψAij ) i
D,C = g ⊗ gm ⊗ gn ⊗ gj = A × ψ,C + ψ A,C (71)
∂Cmn
= (A ⊗ ψ,C )L + ψ A,C ,

• for a simple contraction of second-order tensors A and B

D = A(C) B(C) :
∂(Aik B k .j ) i (72)
D,C = g ⊗ gm ⊗ gn ⊗ gj = A,C B + AB,C ,
∂Cmn
• for a double contraction of second-order tensors A and B

ψ = A (C) : B(C) :
∂(Aij B ij )
ψ,C = gm ⊗ gn = A,C a qB + A q aB,C
∂Cmn (73)
= (A,C )T q aB + A a q(B,C )T
= B : A;C + A : B;C .
Fundamental results. The fundamental identities
∂Aij ∂Ai .j
= δim δjn , = δim δnj , ... (74)
∂Amn ∂Am .n
can be formulated in absolute notation as follows:

A,A = AT ,AT = I ⊗ I , ×I.

AT ,A = A,AT = I 2 (75)

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
302 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

With these results at hand, Table 1 and the identity A−1 A = I we obtain in a straightforward manner:
A−1 ,A A + A−1 A,A = A−1 ,A A + A−1 (I ⊗ I) = 0 ,

A−1 ,A = −A−1 ⊗ A−1 , (76)

A−1 ,A−1 A + A−1 A,A−1 = (I ⊗ I)A + A−1 A,A−1 = 0 ,

A,A−1 = −A ⊗ A . (77)
By using the chain rule, relations (75), (76) and Table 3 we find:
× I) q a(−A−1 ⊗ A−1 ) = −A−T 2
A−T ,A = A−T ,A−1 q aA−1 ,A = (I 2 × A−T , (78)

A−1 ,AT = A−1 ,A q aA,AT = (−A−1 ⊗ A−1 ) q a(I ×

2 I) = −A−1 ×
2 A−1 . (79)
The above given results (75) to (77) are fundamental to every differentiation process which will be observed soon.
Differentiation with respect to symmetric and skew-symmetric tensors. Now attention is focussed on the differentiation
with respect to a symmetric (skew-symmetric) tensor. A symmetric (skew-symmetric) tensor forms a six-dimensional (three-
dimensional) subgroup of the nine-dimensional group of linear transformations. An arbitrary second-order tensor A can be
decomposed into its symmetric and its skew-symmetric part as follows:
1  1 
A= A + AT + A − AT . (80)
2   2  
Asym Askw

The differentiation with respect to a symmetric (skew-symmetric) second-order tensor can be carried out by following the
procedure advocated in [25]. At first, the argument tensor A in ψ(A) or F(A) is expressed in symmetric (skew-symmetric)
form and ψ(A) or F(A) is afterwards differentiated with respect to A as if A would be a second-order tensor having nine
independent components:

• if A = AT is symmetric


∂ψ( 12 (A + AT )) ∂ψ 12 (A + AT ) ∂ 12 (A + AT )
ψ,A = =
 q a
∂A ∂ 12 (A + AT )
∂A
1 ∂ψ (A) ∂ψ (A)
= q aA,A + q aA,AT (81)
2 ∂A ∂A
∂ψ(A) 1 ∂ψ(A)
= q a (I ⊗ I + I 2 × I) = q aS ,
∂A 2 ∂A
• if A = −AT is skew-symmetric


∂ψ( 12 (A − AT )) ∂ψ 12 (A − AT ) ∂ 12 (A − AT )
ψ,A = =
1  qa
∂A ∂ 2 (A − AT )
∂A
1 ∂ψ (A) ∂ψ (A)
= q aA,A − q aA,AT (82)
2 ∂A ∂A
∂ψ(A) 1 ∂ψ(A)
= q a (I ⊗ I − I 2 × I) = q aA .
∂A 2 ∂A
In view of (53) and (54) the partial derivative of a scalar-valued function ψ(A) with respect to a symmetric (skew-symmetric)
argument tensor A = AT (A = −AT ) must be a symmetric (skew-symmetric) second-order tensor. The relations (81),
(82) demonstrate the significance of the tensors S and A already introduced in Sect. 5 through the additive decomposition
(51) for tensor differentiation.
In analogy to (81) and (82), the derivative of a scalar-valued ψ(A) or second-order tensor function F(A) with respect to
a symmetric (skew-symmetric) second-order tensor A can be obtained by employing the chain rule. Thus the derivative of
e.g. A−1 with respect to a symmetric (skew-symmetric) second-order tensor A = AT (A = −AT ) reads as:
1 1
A−1 ,A = A−1 ,A q a (A,A +A,AT ) = A−1 ,A q a (I ⊗ I + I 2
× I)
2 2
1
−1 
= −(A−1 ⊗ A−1 ) q a S = − A ⊗ A−1 + A−1 2× A−1 (83)
2

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 303

1 1
A−1 ,A = A−1 ,A q a (A,A −A,AT ) = A−1 ,A q a (I ⊗ I − I 2
× I)
2 2
1
−1 
= −(A−1 ⊗ A−1 ) q a A = − A ⊗ A−1 − A−1 2 × A−1 (84)
2
In particular, we note that for a symmetric (skew-symmetric) tensor A the fourth-order identity tensor is given by A,A =
AT ,A = A,AT = AT ,AT = S (A,A = AT ,AT = −AT ,A = −A,AT = A).
Applications. For a first application of the above proposed differentiation rules we consider tensor powers An (n = 1, 2, ...)
and compute their partial derivatives with respect to A:
A,A = I ⊗ I = I , (85)
1



A2 ,A = (I ⊗ I)A + A(I ⊗ I) = I ⊗ A + A ⊗ I = Ar ⊗ A1−r , (86)
r=0
2



A3 ,A = I ⊗ A2 + A ⊗ A + A2 ⊗ I = Ar ⊗ A2−r , (87)
r=0

by setting A0 = I. The above results can be generalized for the n-th tensor power An in the form


n−1
n
(A ) ,A = Ar ⊗ A(n−1)−r , (n = 1, 2, . . . ) , (88)
r=0

which, by a double contraction ( a q) and ( q a) with the identity tensor I and by considering the identities from Table 2, yields
n−1
 
n−1  T
A ,A a qI = (n−1)−r a qI = (Ar ) I A(n−1)−r
T
n
A ⊗Ar

r=0 r=0


n−1 T
=n A , (89)
n−1
 
n−1
A ,An q aI = A ⊗A
r (n−1)−r q aI = Ar I A(n−1)−r
r=0 r=0

= n An−1 . (90)
Accordingly, the partial derivative of the scalar-valued tensor function tr A with respect to A can be expressed by using
n

(73) with B = I and C = A as

(tr An ) ,A = (An : I) ,A = An ,A a qI = (An ,A ) q aI
T

T  n−1
= n An−1 = n AT . (91)

By considering the identity (tr A),A = I, we further note that

(A a qI) ,A
n n−1 n−1
(tr A) ,A = n (tr A) (tr A) ,A = n (tr A)
n−1
= n (tr A) I. (92)
Remark 7.2. Note that the contraction laws ( q a) or ( a q) have to be applied in accordance with the existing contraction
of tensors. Therefore in the above appearing trace operation (91) the double contraction law ( a q) has to be used or the rule
( q a) by transposing the fourth-order tensor An ,A .
For a further application, attention is now given to the Cayley-Hamilton theorem:

A3 − IA A2 + IIA A − IIIA I = 0 (93)

with the invariants IA , IIA and IIIA . Their derivatives with respect to A are, by using the above results, obtained in the
form [3]:
(IA ) ,A := (tr A) ,A = I , (94)

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
304 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

1 2

(IIA ) ,A := (tr A) − tr A2 ,A = (tr A) I − AT , (95)
2
 
1 1 1 3
(IIIA ) ,A := det(A),A = tr A3 − tr A2 tr A + (tr A) ,A
3 2 6

2 T
= A − IA AT + IIA I = IIIA A−T , (96)

where, in the derivation of the last expression, (93) has been used.
For a last application, attention is given to the derivative of a scalar-valued function ψ(A) with respect to the inverse
A−1 which in view of (77), Table 2 and by means of the chain rule yields:

ψ = ψ (A) :
ψ,A−1 = ψ,A q aA,A−1 = ψ,A q a(−A ⊗ A) = −AT ψ,A AT . (97)

8 Applications to continuum mechanics

In this section the above given rules are applied to some important problems of continuum mechanics. At first the linearization
of the generalized midpoint-rule is presented. A special section is devoted to the computation and differentiation of second-
order isotropic tensor functions in the form of power series. Finally, the derivative of the exponential tensor function is
studied in detail.
Generalized midpoint-rule. We are first concerned with the differentiation of a special second-order tensor function.
According to [13] a differential equation in the form ḞF−1 = A can be integrated by means of the generalized midpoint-
rule in the form:

Fn+1 − Fn = ∆tAn+α Fn+α = ∆tAn+α (α Fn+1 + (1 − α) Fn ) , (98)

1
which by assuming α = 2 and ∆A = ∆tAn+ 12 delivers:

−1

1 1
Fn+1 = I − ∆A I + ∆A Fn . (99)
2 2

With An+ 12 = 12 (An + An+1 ) (99) can be solved e.g. by means of a Newton-Raphson procedure. However, our
concern is to find the derivative of an equation of the form
−1
χ(A) = (I − A) (I + A) (100)

with respect to the argument tensor A. Obviously (100) is defined over the set of second-order tensors A which donnot have
1 as an eigenvalue. As presented in [13] (100) can be expressed by means of the Cayley-Hamilton-theorem in the form
1

χ(A) = (1 − IA + IIA + IIIA ) I + 2 (1 − IA ) A + 2 A2 (101)
det(I − A)

with
−1 1

(I − A) = (1 − IA + IIA ) I + (1 − IA ) A + A2 (102)
det(I − A)

and

det(I − A) = 1 − IA + IIA − IIIA , (103)

by using the invariants IA , IIA , and IIIA defined in (94) to (96). The derivative of the right-hand side term in (99) may be
expressed in symbolic form as χ(A/2),A . This derivative can be obtained by means of the chain rule, (75) and (76) in a
remarkably simple manner starting from (100) in the form:


−1
χ (A/2)
∂χ ∂ (I − A/2) −1 ∂ (I + A/2)
= (I + A/2) + (I − A/2)
∂A ∂A ∂A

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 305

  ∂ (I − A/2)

−1 −1
= − (I − A/2) ⊗ (I − A/2) (I + A/2) q a
∂A

  ∂ (I + A/2)

−1
+ (I − A/2) (I ⊗ I) q a
∂A

1 −1 1 −1
= (I − A/2) ⊗ χ(A/2) + (I − A/2) ⊗ I
2 2

1 −1
= (I − A/2) ⊗ (χ
χ(A/2) + I) , (104)
2
which is obviously at variance with the approximation for χ(A/2),A presented in [13] which in the current notation by
means of (58) reads as:
 L
−1 −T −1
χ(A/2),A = (I − A/2) × (I − A/2) = (I − A/2) × (I − A/2)−1 .
2 (105)

Considering (94) to (96), the following identities hold:

 1
 1
 1
 1
 1
IA/2 ,A = IA/2 ,A/2 = I , IIA/2 ,A = IIA/2 ,A/2 = IA/2 I − AT ,
2 2 2 2 4 (106)



 1
IIIA/2 ,A = IIIA/2 A−T , III(I−A/2) ,A = − III(I−A/2) (I − A/2)−T .
2
Thus, starting directly from relation (101), we obtain by means of (71) and (86):

χ(A/2)
∂χ 1 1 1
= I × ( (IA/2 − 1) I − AT + IIIA/2 A−T )
∂A det(I − A/2) 2 4

1 1
+(1 − IA/2 ) I ⊗ I − A × I + (I ⊗ A + A ⊗ I)
2 2

1
+ det(I − A/2) χ(A/2) × (I − A/2)−T . (107)
2
The identity of both solutions (104) and (107) can be proved e.g. by a numerical analysis. It is worthwile to specify the
above given second-order tensor function χ(A) for a skew-symmetric argument tensor A = −AT . Such a tensor function
is important for the time-integration of a differential equation of the form ṘRT = A by means of the generalized mid-point
rule (99) holding for arbitrary orthogonal tensors R with the property R−1 = RT and ṘRT being skew-symmetric. In this
particular case we find [23]:

2 (A + A2 )
χ(A) = I + (108)
1 − 12 tr(A2 )
with
−1 A + A2
(I − A) =I+ (109)
1 − 12 tr(A2 )
and
1
det(I − A) = 1 + IIA = 1 − tr(A2 ) . (110)
2
The corresponding derivative χ(A/2),A is obtained by means of the chain rule in the form
χ(A/2)
∂χ χ(A/2)
∂χ
= q aA (111)
∂A ∂A

χ(A/2) 1
∂χ
= q a (I ⊗ I − I 2
× I)
∂A 2

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
306 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

 
1 −1 1
= χ(A/2) + I) q a (I ⊗ I − I ×
(I − A/2) ⊗ (χ 2 I)
2 2

1  −1

= χ(A/2) + I) − (I − A/2)−1 2
(I − A/2) ⊗ (χ × (χ
χ(A/2) + I) .
4
Finally, using expression (108) we arrive at the following result for the derivative χ(A/2),A :
χ(A/2)
∂χ χ(A/2)
∂χ
= q aA (112)
∂A ∂A

χ(A/2) 1
∂χ
= q a (I ⊗ I − I 2
× I)
∂A 2
1 1


2
× I) +
(I ⊗ I − I 2 4 I⊗A+I2 × AT + A ⊗ I + AT 2
×I
= 1 2
1 − 8 tr(A )

1


4 A + A2 /2 × AT
+
2 ,
1 − 18 tr(A2 )
where use has been made of:
χ(A/2)
∂χ (I ⊗ I) + 12 (I ⊗ A + A ⊗ I) (A + A2 /2) × (− 14 AT )
= −
2 . (113)
∂A 1 − 18 tr(A2 ) 1 − 18 tr(A2 )

Isotropic tensor functions. Now attention is focussed on the computation of isotropic tensor functions and their derivatives.
A second-order isotropic tensor function G(A) satisfies the following identity:

G(QAQT ) = QG(A)QT , (114)

holding for arbitrary orthogonal tensors Q. An important class of isotropic tensor functions are tensor functions in form of
power series. It is a well-known fact that by means of a successive application of the Cayley-Hamilton-theorem (93)
tensor powers An with n ≥ 3 can be expressed in terms of the first three tensor powers I, A, A2 by taking coefficients cnr
into consideration:
2

An = cnr (IA , IIA , IIIA ) Ar , (115)
r=0

such that for an isotropic tensor function with the argument tensor A the well-known relation

G(A) = φ0 (IA , IIA , IIIA ) I + φ1 (IA , IIA , IIIA ) A + φ2 (IA , IIA , IIIA ) A2 (116)

holds. The corresponding coefficients cnr belonging to a tensor power An can be obtained by means of a recursive procedure
[15, 20, 10, 11] defined by:

c00 = 1 , c01 = 0 , c02 = 0 ,

(117)
cn0 = cn−1
2 IIIA , cn1 = cn−1
0 − cn−1
2 IIA , cn2 = cn−1
1 + c2n−1 IA .

For tensor powers with negative exponents the corresponding coefficients cnr can be obtained analogously. Under the
assumption that the inverse exists, the three invariants of the inverse A−1 of a tensor A are given by:
IIA IA 1
IA−1 = , IIA−1 = , IIIA−1 = , (118)
IIIA IIIA IIIA
The coefficients belonging to tensor powers with negative exponents can be found using the following computations:

c00 = 1 , c01 = 0 , c02 = 0 ,

(119)
cn−1
0 = cn1 + cn0 IA−1 , cn−1
1 = cn2 − cn0 IIA−1 , cn−1
2 = cn0 IIIA−1 .

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 307

With these relations at hand a given power series can be evaluated with any desired accuracy. For the differentiation of a
tensor power with a positive exponent a series expansion with terms of the form (88) is obtained which, in turn, can be
expressed by means of (115) in the form:

n−1 2 2

An ,A = ckr Ar ⊗ c(n−1)−k
p Ap , (120)
k=0 r=0 p=0


n−1 2
2 
= ckr c(n−1)−k
p Ar ⊗ Ap , (121)
k=0 r=0 p=0

2 
 2
= χnrp (IA , IIA , IIIA ) Ar ⊗ Ap , (n > 0) . (122)
r=0 p=0

The coefficients χnrp can be computed again by means of a recursive procedure by taking the symmetry of χnrp = χnpr into
account. The results are given as follows:

χn00 = χn−1 n−1

20 IIIA + c0 10 − χ12 IIA ,
, χn11 = χn−1 n−1
χ100 = 1 , χ111 = 0 ,
χn22 = χn−1 n−1
12 + χ22 IA , 00 − χ20 IIA ,
χn10 = χn−1 n−1
χ122 = 0 , χ110 = 0 , (123)
χn12 = χn−1
20 − χn−1
22 IIA , χn20 = χn−1
10 + χn−1
20 IA , χ112 =0, χ120 =0.

In extension to the above well-known results, we will consider also the derivative of a tensor power with negative exponent
(n < 0):
−n−1

An ,A = − A−k−1 ⊗ Ak+n , (124)
k=0

−n−1
  2 2

=− cr−k−1 Ar ⊗ ck+n
p Ap , (125)
k=0 r=0 p=0

−n−1
  2
2 
=− cr−k−1 ck+n
p Ar ⊗ Ap , (126)
k=0 r=0 p=0

2 
 2
=− χnrp (IA−1 , IIA−1 , IIIA−1 ) Ar ⊗ Ap , (n < 0) , (127)
r=0 p=0

with the following coefficients χnrp (n < 0):

2 2 2
χ−1
00 = (IA−1 ) , χ−1
11 = (IIA−1 ) , χ−1
22 = (IIIA−1 ) ,

χ−1 −1 −1
10 = −IIA−1 IA−1 , χ12 = −IIIA−1 IIA−1 , χ20 = IIIA−1 IA−1 ,

n 
n 
χn−1 11 = χ12 − χ10 + c1
n−1
n
00 = χ10 + χ00 + c0 IA−1 , χn−1 n n−1
IIA−1 , (128)



χn−1 n
22 = χ20 + c2
n−1
IIIA−1 , χn−1 n n n−1
10 = χ11 + χ10 + c1 IA−1 ,



χ12 = χ22 − χ20 + c2
n−1 n n n−1 n−1 n n n−1
IIA−1 , χ20 = χ12 + χ20 + c2 IA−1 .

For the sake of completeness we will give the calculation scheme for the differentation of an arbitrary tensor power (n) with
respect to the inverse A−1 of a tensor A which is, firstly for (n > 0) , given by:

n−1
An ,A−1 = − Ak+1 ⊗ An−k , (129)
k=0

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
308 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis


n−1 2 2

=− ck+1
r A ⊗
r
cn−k
p Ap , (130)
k=0 r=0 p=0


n−1 2
2 
=− ck+1
r cpn−k Ar ⊗ Ap , (131)
k=0 r=0 p=0

2 
 2
=− χ̄nrp (IA , IIA , IIIA ) Ar ⊗ Ap , (n > 0) , (132)
r=0 p=0

with the coefficients χ̄nrp (for positive n):

1 1
χ̄n00 = χ̄n−1 10 − χ̄12 IIA + c1 , χ̄00 = 0 , χ̄11 = 1 ,
χ̄n11 = χ̄n−1 n−1 n
20 IIIA ,

χ̄n22 = χ̄n−1 n−1 n n−1

12 + χ̄22 IA , χ̄10 = χ̄12 IIIA , χ̄122 = 0 , χ̄110 = 0 , (133)
χ̄n12 = χ̄n−1
11 + χ̄n−1
12 IA , χ̄n20 = χ̄n−1
22 IIIA , χ̄112 =0, χ̄120 =0,

and is, secondly for (n < 0), obtained as:

−n−1

An ,A−1 = A−k ⊗ A(n+1)+k , (134)
k=0

−n−1
  2 2

= c−k
r A ⊗
r
cp(n+1)+k Ap , (135)
k=0 r=0 p=0

−n−1
  2 
2
= c−k (n+1)+k
r cp Ar ⊗ Ap , (136)
k=0 r=0 p=0

2 
 2
= χ̄nrp (IA−1 , IIA−1 , IIIA−1 ) Ar ⊗ Ap , (n < 0) , (137)
r=0 p=0

with the coefficients χ̄nrp (for negative n) defined by:

χ̄−1 −1 −1
00 = 1 , χ̄11 = 0 , χ̄22 = 0 ,
χ̄−1 −1 −1
10 = 0 , χ̄12 = 0 , χ̄20 = 0 ,

00 = χ̄10 + c0 + χ̄00 IA−1 , χ̄11 = χ̄12 − χ̄10 IIA−1 ,

χ̄n−1 n n n n−1 n n
(138)
χ̄n−1
22 = χ̄n20 IIIA−1 , n−1
χ̄10 = χ̄n20 − χ̄n00 IIA−1 ,
χ̄n−1
12 = χ̄n10 IIIA−1 , n−1
χ̄20 = χ̄n00 IIIA−1 .

The exponential mapping. The above results for a tensor power and its derivative are useful for the computation of power
series in terms of a second-order tensor A. Here we want to focus our attention to a particular isotropic tensor function in
the form of a power series, the exponential function, which can be employed for a time integration of a differential equation
of the form ḞF−1 = A [23]:

Fn+1 = exp(∆t An+α ) Fn . (139)

By using the following relation for the time derivative of the exponential function:

˙
exp(∆t An+α ) = exp(∆t An+α ),(∆t An+α ) q aAn+α = An+α exp(∆t An+α ) = exp(∆t An+α )An+α ,
(140)

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 309

the identity (139) can be proved:

 

ḞF−1 = exp(∆t An+α ),(∆t An+α ) Fn F−1
n exp(−∆t An+α )
q aAn+α = An+α . (141)
n+1

For α = 0 relation (139) leads to an explicit (forward) Euler rule. If relation (139) is fulfilled at tn+1 (i.e. α = 1) an
implicit (backward) Euler rule is obtained. The midpoint rule, which is is second-order accurate (see [23]), is obtained for
α = 12 .

The exponential function can be expanded in a power series in the form :

∞
1 2 1 3 1 4 1
exp(A) = I + A + A + A + A + ··· = An . (142)
2! 3! 4! n=0
n!

By utilizing the coefficients cnr (117) and χnrp (123) the exponential function and its derivative, which is important for its
linearization, can be computed by [6]:
∞
 ∞ 2

1 n  1 n
exp(A) = A = cr Ar , (143)
n=0
n! n=0 r=0
n!

∞ ∞  2  2

1 1 n
exp(A),A = An ,A = χrp Ar ⊗ Ap . (144)
n=0
n! n=1 r=0 p=0
n!

Remark 8.1. For a symmetric (skew-symmetric) argument tensor A the tensor product Ar ⊗ Ap in (144) should be
replaced with 12 (Ar ⊗ Ap + Ar 2 × Ap ) (or likewise 12 (Ar ⊗ Ap − Ar 2 × Ap )).
Remark 8.2. Note that the recursive computation of cnr and χnrp may involve a large number of terms when the invariants
of the argument tensor are relatively large. However, for moderate values the recursive computation scheme works well
and converges after a few steps. An alternative procedure has been proposed in [12] relying on a spectral decomposition
of the argument tensor even in the case of non-symmetry of A. A special consideration is needed for the case of multiple
eigenvalue s. The latter problem has been solved in [11].
For a skew-symmetric argument tensor A = −AT the exponential function can be expressed in the following form
referred to as Rodrigues formula
sin( a ) 1 − cos( a ) 2
exp(A) = I + A+ A (145)
a  a 2
with the axial vector a of A (aa × b = Ab b) and the norm  a 2 = − 12 tr(A2 ). Using  a ,A = − 21a AT = 1
2a A the
derivative of (145) with respect to the skew-symmetric tensor A is given by:

cos( a ) sin( a )
exp(A),A = 2
A×A− (A − A2 ) × A
2 a 2  a 3

(1 − cos( a )) 2 sin( a )

− A ×A+ (I ⊗ I)
 a 4 a

1 − cos( a )
+ (A ⊗ I + I ⊗ A) q aA
 a 2

1 cos( a ) 1 sin( a )
= A×A− (A − A2 ) × A
2  a 2 2  a 3

(1 − cos( a )) 2 sin( a ) 1

− A ×A+ × I)
(I ⊗ I − I 2
 a 4 a 2

1 − cos( a ) 1
+ ×I+I⊗A+I2
(A ⊗ I + AT 2 × AT ) . (146)
 a 2 2

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
310 O. Kintzel and Y. Başar: Fourth-order tensors – tensor differentiation. Part I: Classical tensor analysis

Linearization of a second-order-valued tensor function with rational exponent. As last example a special second-order-
m
valued tensor function H = F n is considered with an exponent given as a rational fraction of two integer values. The tensor
function F may depend, in turn, in nonlinear form on a second-order tensor D. The first variation ∆H of the second-order-
valued tensor function H, linearized around H(F(D)) = 0 with respect to D, can be obtained according to a procedure
explained in the following. First H is multiplied n times such that
m
Hn = (F(D)) (147)

holds with integer values n and m as exponents. The procedure is illustrated best by using the specific values n = 2 and
m = 3. Thus the following holds:

∂F
(H ⊗ I + I ⊗ H) q a∆H = (F2 ⊗ I + F ⊗ F + I ⊗ F2 ) q a q a∆D , (148)
∂D
which can be expressed as:

∂F
∆H = (H ⊗ I + I ⊗ H)−1 q a(F2 ⊗ I + F ⊗ F + I ⊗ F2 ) q a q a∆D (149)
∂D
which is the required result. In case of symmetric tensors F, D and H the result can be further transformed into:

−1 ∂F
∆H = H q aF q a q a∆D (150)
∂D

with H = 12 (H ⊗ I + H 2 × H) consistent with H =

× I+I⊗H+I 2 Ht = Hti = Hto owing to the symmetry of H
(H = H ) and I, and
T

1
2 
F= F ⊗ I + F ⊗ F + I ⊗ F2 + F2 2
×I+F2 × F2 .
×F+I2 (151)
2
H itself can be computed by solving an eigenvalue problem.

9 Conclusion
To conclude, in this contribution new rules and definitions have been introduced for fourth-order tensors as well as for partial
derivatives of scalar-valued or second-order-valued tensor functions to enable the tensor differentiation, which in component
form requires lengthy derivation procedures, in a decisively simplier form using absolute tensor notation. Concerning the
effectivity of the proposed rules we may state the following:

• The double contraction rules ( :, a q, q a) hold without ambiguity for second- and fourth-order tensors as well as for their
arbitrary combinations and they are, thus, very advantageous in solutions of problems in continuum mechanics. The
rules (:) and (( a q), ( q a)) are closely related to partial derivatives of the forms A;C (70.2) and A,C (70.1), respectively.
The derivatives of both kinds A;C and A,C are, in turn, transformable easily into each other by means of new operations
introduced in Sect. 6, the so-called process of basis rearrangement, (see (58)).
• The set of fourth-order tensors forms with each of the proposed double contraction operations (:) and (( a q),( q a)) a ring
in the sense that it satisfies the associative and, moreover, the distributive law and is, in addition, associated with a
well-defined identity element.
• The consideration of the identities presented in Tables 1–3 for simple and double contractions of fourth-order and
second-order tensors may simplify derivation procedures in terms of these tensors considerably.
• The differentiation rule adopted from Itskov [10] (65) here denoted by A,C obeys the well-known product rule of
differential calculus in dealing with simple contractions of second-order tensors and is, therefore, very effective for
expressions governed by this operation, e.g. tensor powers An . In applications we have found it to be superior to the
more usual definition [3] denoted in this contribution by A;C .
The introduced rules have been applied to relevant problems of continuum mechanics such as linearization of the generalized
midpoint-rule and the exponential function. Special attention has been given to the differentiation of the above mentioned
functions with respect to a skew-symmetric argument tensor. Additionally, a recursive calculation scheme has been presented
to express arbitrary tensor powers and their derivatives in terms of the first three tensor powers I, A, A2 by means of the
Cayley-Hamilton-theorem.

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.zamm-journal.org
ZAMM · Z. Angew. Math. Mech. 86, No. 4 (2006) 311

Acknowledgements The financial support of the National Science Foundation (DFG) within the collaborative research center (SFB)
398 is gratefully acknowledged. The authors thank Prof. Dr.-Ing. Mikhail Itskov for helpful discussions.

References
[1] M. Artin, Algebra (Prentice Hall, Englewood Cliffs, N.J., 1991).
[2] Y. Başar and W. B. Krätzig, Mechanik der Flächentragwerke (Vieweg & Sohn, Braunschweig, 1985).
[3] Y. Başar and D. Weichert, Nonlinear Continuum Mechanics of Solids (Springer, Berlin, 2000).
[4] J. Betten, Kontinuumsmechanik (Springer, Berlin 2001).
[5] P. G. Ciarlet, Mathematical Elasticity, Vol. 1: Three-dimensional Elasticity (Elsevier, New York, 1988).
[6] E.A. de Souza Neto, The exact derivative of the exponential of an unsymmetric tensor, Comput. Methods Appl. Mech. Eng. 82,
535–544 (2001).
[7] G. del Piero, Some properties of the set of fourth-order tensors with application to elasticity, J. Elast. 3, 245–261 (1979).
[8] A. E. Green and W. Zerna, Theoretical Elasticity (Clarendon Press, Oxford, 1968).
[9] M. Itskov, On the theory of fourth-order tensors and their applications in computational mechanics, Comput. Methods Appl. Mech.
Eng. 189, 419–438 (2000).
[10] M. Itskov, The derivative with respect to a tensor: Some theoretical aspects and applications, Z. Angew. Math. Mech. 82, 535–544
(2002).
[11] M. Itskov, Computation of the exponential and other isotropic tensor functions and their derivative, Comput. Methods .Appl Mech.
Eng. 192, 3985–3999 (2003).
[12] M. Itskov and N. Aksel, A closed-formula representation for the derivative of non-symmetric tensor power series, Int. J. Solids
Struct. 39, 5963–5978 (2002).
[13] B. Loret, F. Hammoum and Y. F. Dafalias, Computational aspects of large elastoplastic deformations in the presence of anisotropy
and plastic spin, Comput. Methods Appl. Mech. Eng. 105, 151–180 (1993).
[14] J. E. Marsden and T. J. R. Hughes, Mathematical Foundations of Elasticity (Prentice Hall, Englewood Cliffs, N.J., 1983).
[15] C. Miehe, Exponential map algorithm for stress updates in anisotropic multiplicative elastoplasticity for single crystals, Int. J. Numer.
Methods Eng. 39, 3367–3390 (1996).
[16] P. Podio-Guidugli, A Primer in Elasticity, J. Elast. 58, 1–104 (2000).
[17] E. Rizzi and I. Carol, A formulation of anisotropic elastic damage using compact tensor formalism, J. Elast. 64, 85–109 (2001).
[18] L. Rosati, A novel approach to the solution of the tensor equation AX+XA=H, Int. J. Solids Struct. 37, 3457–3477 (2000).
[19] C. Sansour, On the spatial description in elasticity and the Doyle-Ericksen formula, Comp. Methods Appl. Mech. Eng. 107, 239–249
(1993).
[20] C. Sansour and F. G. Kollmann, Large viscoplastic deformations of shells, theory and finite element formulation, Comput. Mech.
21, 512–525 (1998).
[21] M. Scheidler, The tensor equation AX+XA=φ(A,H), with Applications to Kinematics of Continua, J. Elast. 36, 117–153 (1994).
[22] J.A. Schouten, Ricci Calculus (Springer, Berlin, 1954).
[23] J. C. Simo and T. J. R. Hughes, Computational Inelasticity (Springer, New York, 1998).
[24] J. C. Simo and J. E. Marsden, On the rotated stress tensor and the material version of the Doyle-Ericksen formula, Arch. Rational
Mech. Anal. 86, 213–231 (1984).
[25] C. Truesdell and W. Noll, The nonlinear field theories of mechanics, Handbuch der Physik Vol. III/3, edited by S. Flügge (Springer,
Berlin, 1965).

www.zamm-journal.org

c 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim