Geometry of Logarithmic Strain Measures in Solid Mechanics

We consider the two logarithmic strain measures ωiso=||devnlogU||=||devnlogFTF||andωvol=|tr(logU)=|tr(logFTF)|=|log(detU)|,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{ll} {\omega_{\mathrm{iso}}} = ||{{\mathrm dev}_n {\mathrm log} U} || = ||{{\mathrm dev}_n {\mathrm log} \sqrt{F^TF}}|| \quad \text{ and } \quad \\ {\omega_{\mathrm{vol}}} = |{{\mathrm tr}({\mathrm log} U)} = |{{\mathrm tr}({\mathrm log}\sqrt{F^TF})}| = |{\mathrm log}({\mathrm det} U)|\,,\end{array}$$\end{document} which are isotropic invariants of the Hencky strain tensor logU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\log U}$$\end{document}, and show that they can be uniquely characterized by purely geometric methods based on the geodesic distance on the general linear group GL(n)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm GL}(n)}$$\end{document}. Here, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}$$\end{document} is the deformation gradient, U=FTF\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${U=\sqrt{F^TF}}$$\end{document} is the right Biot-stretch tensor, log\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\log}$$\end{document} denotes the principal matrix logarithm, ‖·‖\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\| \cdot \|}$$\end{document} is the Frobenius matrix norm, tr\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm tr}$$\end{document} is the trace operator and devnX=X-1ntr(X)·1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\text dev}_n X = X- \frac{1}{n} \,{\text tr}(X)\cdot {\mathbb{1}}}$$\end{document} is the n\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}$$\end{document}-dimensional deviator of X∈Rn×n\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X\in{\mathbb {R}}^{n \times n}}$$\end{document}. This characterization identifies the Hencky (or true) strain tensor as the natural nonlinear extension of the linear (infinitesimal) strain tensor ε=sym∇u\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon={\text sym}\nabla u}$$\end{document}, which is the symmetric part of the displacement gradient ∇u\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\nabla u}$$\end{document}, and reveals a close geometric relation between the classical quadratic isotropic energy potential μ‖devnsym∇u‖2+κ2[tr(sym∇u)]2=μ‖devnε‖2+κ2[tr(ε)]2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu {\| {\text dev}_n {\text sym} \nabla u \|}^2 + \frac{\kappa}{2}{[{\text tr}({\text sym} \nabla u)]}^2 = \mu {\| {\text dev}_n \varepsilon \|}^2 + \frac{\kappa}{2} {[{\text tr} (\varepsilon)]}^2$$\end{document}in linear elasticity and the geometrically nonlinear quadratic isotropic Hencky energy μ‖devnlogU‖2+κ2[tr(logU)]2=μωiso2+κ2ωvol2,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu {\| {\text dev}_n log U \|}^2 + \frac{\kappa}{2}{[{\text tr}(log U)]}^2 = \mu {\omega_{{\text iso}}^2} + \frac{\kappa}{2}{\omega_{{\text vol}}^2},$$\end{document}where μ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mu}$$\end{document} is the shear modulus and κ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\kappa}$$\end{document} denotes the bulk modulus. Our deduction involves a new fundamental logarithmic minimization property of the orthogonal polar factor R\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}$$\end{document}, where F=RU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F=RU}$$\end{document} is the polar decomposition of F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}$$\end{document}. We also contrast our approach with prior attempts to establish the logarithmic Hencky strain tensor directly as the preferred strain tensor in nonlinear isotropic elasticity.

which are isotropic invariants of the Hencky strain tensor log U , and show that they can be uniquely characterized by purely geometric methods based on the geodesic distance on the general linear group GL(n). Here, F is the deformation gradient, U = √ F T F is the right Biot-stretch tensor, log denotes the principal matrix logarithm, . is the Frobenius matrix norm, tr is the trace operator and dev n X = X − 1 n tr(X ) · 1 is the n-dimensional deviator of X ∈ R n×n . This characterization identifies the Hencky (or true) strain tensor as the natural nonlinear extension of the linear (infinitesimal) strain tensor ε = sym ∇u, which is the symmetric part of the displacement gradient ∇u, and reveals a close geometric relation between the classical quadratic isotropic energy potential μ dev n sym ∇u 2 + κ 2 [tr(sym ∇u)] 2 = μ dev n ε 2 + κ 2 [tr(ε)] 2 in linear elasticity and the geometrically nonlinear quadratic isotropic Hencky energy where μ is the shear modulus and κ denotes the bulk modulus. Our deduction involves a new fundamental logarithmic minimization property of the orthogonal In memory of Giuseppe Grioli (*10.4.1912 - †4. 3.2015), a true paragon of rational mechanics.

What's in a strain?
The concept of strain is of fundamental importance in elasticity theory. In linearized elasticity, one assumes that the Cauchy stress tensor σ is a linear function of the symmetric infinitesimal strain tensor ε = sym ∇u = sym(∇ϕ − 1) = sym(F − 1), where ϕ : → R n is the deformation of an elastic body with a given reference configuration ⊂ R n , ϕ(x) = x + u(x) with the displacement u, F = ∇ϕ is the deformation gradient, sym ∇u = 1 2 (∇u + (∇u) T ) is the symmetric part of the displacement gradient ∇u and 1 ∈ GL + (n) is the identity tensor in the group of invertible tensors with positive determinant. 1 In geometrically nonlinear elasticity models, it is no longer necessary to postulate a linear connection between some stress and some strain. However, nonlinear strain tensors are often used in order to simplify the stress response function, and many constitutive laws are expressed in terms of linear relations between certain strains and stresses [15,16,24] (cf. Appendix A.2 for examples). 2 There are many different definitions of exactly what the term "strain" encompasses: while Truesdell and Toupin [205, p. 268] consider "any uniquely invertible isotropic second order tensor function of [the right Cauchy-Green deformation tensor C = F T F]" to be a strain tensor, it is commonly assumed [106, p. 230] (cf. [23,107,108,160]) that a (material or Lagrangian 3 ) strain takes the form of a primary matrix function of the right Biot-stretch tensor U = √ F T F of the deformation gradient F ∈ GL + (n), that is an isotropic tensor function E : Sym + (n) → Sym(n) from the set of positive definite tensors to the set of symmetric tensors of the form with a scale function e : (0, ∞) → R, where ⊗ denotes the tensor product, λ i are the eigenvalues and e i are the corresponding eigenvectors of U . However, there is no consensus on the exact conditions for the scale function e; Hill (cf. [107, p. 459] and [108, p. 14]) requires e to be "suitably smooth" and monotone with e(1) = 0 and e (1) = 1, whereas Ogden [163, p. 118] also requires e to be infinitely differentiable and e > 0 to hold on all of (0, ∞). The general idea underlying these definitions is clear: strain is a measure of deformation (that is the change in form and size) of a body with respect to a chosen (arbitrary) reference configuration. Furthermore, the strain of the deformation gradient F ∈ GL + (n) should correspond only to the non-rotational part of F. In particular, the strain must vanish if and only if F is a pure rotation, that is if and only if F ∈ SO(n), where SO(n) = {Q ∈ GL(n) | Q T Q = 1, det Q = 1} denotes the special orthogonal group. This ensures that the only strain-free deformations are rigid body movements: 1 Although F is widely known as the deformation "gradient", F = ∇ϕ = Dϕ actually denotes the first derivative (or the Jacobian matrix) of the deformation ϕ. 2 In a short note [32], Brannon observes that "usually, a researcher will select the strain measure for which the stress-strain curve is most linear". In the same spirit, Bruhns [33, p. 147] [74]).
where the last implication is due to the rigidity [175] inequality Curl R 2 c + ∇ R 2 for R ∈ SO(n) (with a constant c + > 0), cf. [152]. A similar connection between vanishing strain and rigid body movements holds for linear elasticity: if ε ≡ 0 for the linearized strain ε = sym ∇u, then u is an infinitesimal rigid displacement of the form where so(n) = {A ∈ R n×n : A T = −A} denotes the space of skew symmetric matrices. This is due to the inequality Curl A 2 c + ∇ A 2 for A ∈ so(n), cf. [152].
In the following, we will use the term strain tensor (or, more precisely, material strain tensor) to refer to an injective isotropic tensor function U → E(U ) of the right Biot-stretch tensor U mapping Sym + (n) to Sym(n) with where O(n) = {Q ∈ GL(n) | Q T Q = 1} is the orthogonal group and 1 denotes the identity tensor. In particular, these conditions ensure that 0 = E(U ) = E( √ F T F) if and only if F ∈ SO(n). Note that we do not require the mapping to be of the form (1).
All strain tensors, by the definition employed here, can be seen as equivalent: since the mapping U → E(U ) is injective, for every pair E, E of strain tensors there exists a mapping ψ : Sym(n) → Sym(n) such that E (U ) = ψ(E(U )) for all U ∈ Sym + (n). Therefore, every constitutive law of elasticity can -in principle -be expressed in terms of any strain tensor and no strain tensor can be inherently superior to any other strain tensor. 7 Note that this invertibility property also holds if the definition by Hill or Ogden is used: if the strain is given via a scale function e, the strict monotonicity of e implies that the mapping U → E(U ) is strictly monotone [130], that is 6 Bruhns [37, p. 41-42] emphasizes the advantages of the Hencky strain tensor over the other Seth-Hill strain tensors in the one-dimensional case: "The significant advantage of this logarithmic (Hencky) measure lies in the fact that it tends to infinity as F tends to zero, thus in a very natural way bounding the regime of applicability to the case F > 0. This behavior can also be observed for strain [tensors] with negative exponent n. Compared with the latter, however, the logarithmic measure also goes to infinity as F does, whereas it is evident that for negative values of n the strain [ 1 n (F n − 1)] is bound to the limit − 1 n . All measures with positive values of n including the Green  for all U 1 , U 2 ∈ Sym + (n) with U 1 = U 2 , where X, Y = tr(X T Y ) denotes the Frobenius inner product on Sym(n) and tr(X ) = n i=1 X i,i is the trace of X ∈ R n×n . This monotonicity in turn ensures that the mapping U → E(U ) is injective.
In contrast to strain or strain tensor, we use the term strain measure to refer to a nonnegative real-valued function ω : GL + (n) → [0, ∞) depending on the deformation gradient which vanishes if and only if F is a pure rotation, that is ω(F) = 0 if and only if F ∈ SO(n).
Note that the terms "strain tensor" and "strain measure" are sometimes used interchangeably in the literature (for example [108,160]). A simple example of a strain measure in the above sense is the mapping F → E( √ F T F) of F to an orthogonally invariant norm of any strain tensor E.
There is a close connection between strain measures and energy functions in isotropic hyperelasticity: an isotropic energy potential [84] is a function W depending on the deformation gradient F such that for all F ∈ GL + (n), Q ∈ SO(n) and While every such energy function can be taken as a strain measure, many additional conditions for "proper" energy functions are discussed in the literature, such as constitutive inequalities [11,44,106,107,127,203], generalized convexity conditions [10,13] or monotonicity conditions to ensure that "stress increases with strain" [155,Section 2.2]. Apart from that, the main difference between strain measures and energy functions is that the former are purely mathematical expressions used to quantitatively assess the extent of strain in a deformation, whereas the latter postulate some physical behavior of materials in a condensed form: an elastic energy potential, interpreted as the elastic energy per unit volume in the undeformed configuration, induces a specific stress response function, and therefore completely determines the physical behavior of the modelled hyperelastic material. 8 The connection between "natural" strain measures and energy functions will be further discussed later on.
In particular, we will be interested in energy potentials which can be expressed in terms of certain strain measures. Note carefully that, in contrast to strain tensors, strain measures cannot simply be used interchangeably: for two different strain measures (as defined above) ω 1 , ω 2 , there is generally no function f : R + → R + such that ω 2 (F) = f (ω 1 (F)) for all F ∈ GL + (n). Compared to "full" strain tensors, this can be interpreted as an unavoidable loss of information for strain measures (which are only scalar quantities).
Sometimes a strain measure is employed only for a particular kind of deformation. For example, on the group of simple shear deformations (in a fixed plane) consisting of all F γ ∈ GL + (3) of the form we could consider the mappings We will come back to these partial strain measures in Section 3.2.
In the following we consider the question of what strain measures are appropriate for the theory of nonlinear isotropic elasticity. Since, by our definition, a strain measure attains zero if and only if F ∈ SO(n), a simple geometric approach is to consider a distance function on the group GL + (n) of admissible deformation gradients, that is a function dist : which satisfies the triangle inequality and vanishes if and only if its arguments are identical. 9 Such a distance function induces a "natural" strain measure on GL + (n) by means of the distance to the special orthogonal group SO(n): In this way, the search for an appropriate strain measure reduces to the task of finding a natural, intrinsic distance function on GL + (n).

The search for appropriate strain measures
The remainder of this article is dedicated to this task: after some simple (Euclidean) examples in Section 2, we consider the geodesic distance on GL + (n) in Section 3. Our main result is stated in Theorem 3.3: if the distance on GL + (n) is induced by a left-GL(n)-invariant, right-O(n)-invariant Riemannian metric on GL(n), then the distance of F ∈ GL + (n) to SO(n) is given by where F = RU with U = √ F T F ∈ Sym + (n) and R ∈ SO(n) is the polar decomposition of F. Section 3 also contains some additional remarks and corollaries which further expand upon this Riemannian strain measure.
In Section 4, we discuss a number of different approaches towards motivating the use of logarithmic strain measures and strain tensors, whereas applications of our results and further research topics are indicated in Section 5.
Our main result (Theorem 3.3) has previously been announced in a Comptes Rendus Mécanique article [148] as well as in Proceedings in Applied Mathematics and Mechanics [149].
The idea for this paper was conceived in late 2006. However, a number of technical difficulties had to be overcome (cf. [29,118,129,146,157]) in order to prove our results. The completion of this article might have taken more time than was originally foreseen, but we adhere to the old German saying: Gut Ding will Weile haben.

The Euclidean strain measure in linear isotropic elasticity
An approach similar to the definition of strain measures via distance functions on GL + (n), as stated in equation (5), can be employed in linearized elasticity theory: let ϕ(x) = x + u(x) with the displacement u. Then the infinitesimal strain measure may be obtained by taking the distance of the displacement gradient ∇u ∈ R n×n to the set of linearized rotations so(n) = {A ∈ R n×n : A T = −A}, which is the vector space of skew symmetric matrices. 10 An obvious choice for a distance measure on the linear space R n×n ∼ = R n 2 of n × n-matrices is the Euclidean distance induced by the canonical Frobenius norm We use the more general weighted norm defined by which separately weights the deviatoric (or trace free) symmetric part dev n sym X = sym X − 1 n tr(sym X )·1, the spherical part 1 n tr(X )·1, and the skew symmetric part skew X = 1 2 (X − X T ) of X ; note that X μ,μ c ,κ = X for μ = μ c = 1, κ = 2 n , and that . μ,μ c ,κ is induced by the inner product Scale functions e r ,ẽ r associated with the strain tensors E r and E r = 1 2 (E r − E −r ) via eigenvalue λ on R n×n , where X, Y = tr(X T Y ) denotes the canonical inner product. 11 In fact, every isotropic inner product on R n×n , that is every inner product ·, · iso with for all X, Y ∈ R n×n and all Q ∈ O(n), is of the form (7), cf. [50]. The suggestive choice of variables μ and κ, which represent the shear modulus and the bulk modulus, respectively, will prove to be justified later on. The remaining parameter μ c will be called the spin modulus.
Of course, the element of best approximation in so(n) to ∇u with respect to the weighted Euclidean distance dist Euclid,μ,μ c ,κ (X, Y ) = X − Y μ,μ c ,κ is given by the associated orthogonal projection of ∇u to so(n), cf. Fig. 2. Since so(n) and the space Sym(n) of symmetric matrices are orthogonal with respect to ·, · μ,μ c ,κ , this projection is given by the continuum rotation, that is the skew symmetric part skew ∇u = 1 2 (∇u − (∇u) T ) of ∇u, the axial vector of which is curl u. Thus the 11 The family (7) of inner products on R n×n is based on the Cartan-orthogonal decomposition of the Lie algebra gl(n) = R n×n . Here, sl(n) = {X ∈ gl(n) | tr X = 0} denotes the Lie algebra corresponding to the special linear group SL(n) = {A ∈ GL(n) | det A = 1}. Euclid,μ,μ c ,κ (∇u, so(n)) = μ dev n ε 2 + κ 2 [tr(ε)] 2 of ∇u to so(n) in R n×n in the infinitesimal strain setting. The strain tensor ε = sym ∇u is orthogonal to the infinitesimal continuum rotation skew ∇u distance is 12 dist Euclid,μ,μ c ,κ (∇u, so(n)) := inf We therefore find for the linear strain tensor ε = sym ∇u, which is the quadratic isotropic elastic energy, that is the canonical model of isotropic linear elasticity with σ = D ∇u W lin (∇u) = 2μ dev n ε + κ tr(ε) · 1.
This shows the aforementioned close connection of the energy potential to geometrically motivated measures of strain. Note also that the so computed distance to so(n) is independent of the parameter μ c , the spin modulus, weighting the skewsymmetric part in the quadratic form (6). We will encounter the (lack of) influence of the parameter μ c subsequently again. 12 The distance can also be computed directly: since for all A ∈ so(n), the infimum inf Fig. 3. The "flat" interpretation of GL + (n) ⊂ R n×n endowed with the Euclidean distance. Note that F − R = R (U − 1) = U − 1 by orthogonal invariance of the Frobenius norm, where F = RU is the polar decomposition of F Furthermore, this approach motivates the symmetric part ε = sym ∇u of the displacement gradient as the strain tensor in the linear case: instead of postulating that our strain measure should depend only on ε, the above computations deductively characterize ε as the infinitesimal strain tensor from simple geometric assumptions alone.

The Euclidean strain measure in nonlinear isotropic elasticity
In order to obtain a strain measure in the geometrically nonlinear case, we must compute the distance dist(∇ϕ, SO(n)) = dist(F, SO(n)) = inf Q∈SO(n) dist(F, Q) of the deformation gradient F = ∇ϕ ∈ GL + (n) to the actual set of pure rotations SO(n) ⊂ GL + (n). It is therefore necessary to choose a distance function on GL + (n); an obvious choice is the restriction of the Euclidean distance on R n×n to GL + (n). For the canonical Frobenius norm . , the Euclidean distance between Thus the computation of the strain measure induced by the Euclidean distance on GL + (n) reduces to the matrix nearness problem [104] dist Euclid (F, SO(n)) = inf By a well-known optimality result discovered by Giuseppe Grioli [82] (cf. [31,83,131,151]), also called "Grioli's Theorem" by Truesdell and Toupin [205, p. 290], this minimum is attained for the orthogonal polar factor R. Theorem 2.1. (Grioli's Theorem [82,151,205]) Let F ∈ GL + (n). Then where F = RU is the polar decomposition of F with R = polar(F) ∈ SO(n) and U = √ F T F ∈ Sym + (n). The minimum is uniquely attained at the orthogonal polar factor R.
Remark 2.2. The minimization property stated in Theorem 2.1 is equivalent to [132] max Q∈SO(n) Thus for nonlinear elasticity, the restriction of the Euclidean distance to GL + (n) yields the strain measure dist Euclid (F, SO(n)) = U − 1 .
In analogy to the linear case, we obtain where E 1/2 = U − 1 is the Biot strain tensor. Note the similarity between this expression and the Saint-Venant-Kirchhoff energy [117] where E 1 = 1 2 (C − 1) = 1 2 (U 2 − 1) is the Green-Lagrangian strain. The squared Euclidean distance of F to SO(n) is often used as a lower bound for more general elastic energy potentials. Friesecke, James and Müller [78], for example, show that if there exists a constant C > 0 such that for all F ∈ GL + (3) in a large neighborhood of 1, then the elastic energy W shows some desirable properties which do not otherwise depend on the specific form of W . As a starting point for nonlinear theories of bending plates, Friesecke et al. also use the weighted squared norm where λ is the first Lamé parameter, as an energy function satisfying (13). The same energy, also called the Biot energy [150], has been recently motivated by applications in digital geometry processing [43]. However, the resulting strain measure ω(U ) = dist Euclid (F, SO(n)) = U −1 does not truly seem appropriate for finite elasticity theory: for U → 0 we find U − 1 → 1 = √ n < ∞, thus singular deformations do not necessarily correspond to an infinite measure ω. Furthermore, the above computations are not compatible with the weighted norm introduced in Section 2.1: in general [70,71,150], thus the Euclidean distance of F to SO(n) with respect to . μ,μ c ,κ does not equal √ F T F − 1 μ,μ c ,κ in general. In these cases, the element of best approximation is not the orthogonal polar factor R = polar(F).
In fact, the expression on the left-hand side of (14) is not even well defined in terms of linear mappings F and Q [150]: the deformation gradient F = ∇ϕ at a point x ∈ is a two-point tensor and hence, in particular, a linear mapping between the tangent spaces T x and T ϕ(x) ϕ( ). Since taking the norm of X requires the decomposition of X into its symmetric and its skew symmetric part, it is only well defined if X is an endomorphism on a single linear space. 13 Therefore F − Q μ,μ c ,κ , while being a valid expression for arbitrary matrices F, Q ∈ R n×n , is not an admissible term in the setting of finite elasticity. We also observe that the Euclidean distance is not an intrinsic distance measure on GL + (n): in general, A − B / ∈ GL + (n) for A, B ∈ GL + (n), hence the term A − B depends on the underlying linear structure of R n×n . Since it is not a closed subset of R n×n , GL + (n) is also not complete with respect to dist Euclid ; for example, the sequence 1 n · 1 n∈N is a Cauchy sequence which does not converge. Most importantly, because GL + (n) is not convex, the straight line {A + t (B − A) | t ∈ [0, 1]} connecting A and B is not necessarily contained 14 in GL + (n), which shows that the characterization of the Euclidean distance as the length of a shortest connecting curve is also not possible in a way intrinsic to GL + (n), as the intuitive sketches in Figs. 4 and 5 indicate. 15 13 If X : V 1 → V 2 is a mapping between two different linear spaces V 1 , V 2 , then X T is a mapping from V 2 to V 1 , hence sym X = 1 2 (X + X T ) is not well-defined. 14 The straight line connecting F ∈ GL + (n) to its orthogonal polar factor R (that is the shortest connecting line from F to SO(n)), however, lies in GL + (n), which easily follows from the convexity of Sym + (n): for all t ∈ [0, 1], t U + (1 − t) 1 ∈ Sym + (n) and thus 15 Note that the representation of GL + (n) as a sphere only serves to visualize the curved nature of the manifold and that further geometric properties of GL + (n) should not be inferred These issues amply demonstrate that the Euclidean distance can only be regarded as an extrinsic distance measure on the general linear group. We therefore need to expand our view to allow for a more appropriate, truly intrinsic distance measure on GL + (n).

GL + (n) as a Riemannian manifold
In order to find an intrinsic distance function on GL + (n) that alleviates the drawbacks of the Euclidean distance, we endow GL(n) with a Riemannian metric. 16 Such a metric g is defined by an inner product Then the length of a sufficiently smooth curve γ : [0, 1] → GL(n) is given by Footnote 15 continued from the figures. In particular, GL + (n) is not compact and the geodesics are generally not closed. 16 For technical reasons, we define g on all of GL(n) instead of its connected component GL + (n); for more details, we refer to [129], where a more thorough introduction to geodesics on GL(n) can be found. Of course, our strain measure depends only on the restriction of g to GL + (n).
, and the geodesic distance (cf. Fig. 5) between A, B ∈ GL + (n) is defined as the infimum over the lengths of all (twice continuously differentiable) curves connecting A to B: Our search for an appropriate strain measure is thereby reduced to the task of finding an appropriate Riemannian metric on GL(n). Although it might appear as an obvious choice, the metricǧ witȟ (15) provides no improvement over the already discussed Euclidean distance on GL + (n): since the length of a curve γ with respect toǧ is the classical (Euclidean) length the shortest connecting curves with respect toǧ are straight lines of the form Locally, the geodesic distance induced byǧ is therefore equal to the Euclidean distance. However, as discussed in the previous section, not all straight lines connecting arbitrary A, B ∈ GL + (n) are contained within GL + (n), thus length minimizing curves with respect toǧ do not necessarily exist (cf. Fig. 6). Many of the shortcomings of the Euclidean distance therefore apply to the geodesic distance induced byǧ as well.
In order to find a more viable Riemannian metric g on GL(n), we consider the mechanical interpretation of the induced geodesic distance dist geod : while our focus lies on the strain measure induced by g, that is the geodesic distance of the deformation gradient F to the special orthogonal group SO(n), the distance dist geod (F 1 , F 2 ) between two deformation gradients F 1 , F 2 can also be motivated directly as a measure of difference between two linear (or homogeneous) deformations F 1 , F 2 of the ) dx measures how much two deformations ϕ 1 , ϕ 2 of a body differ from each other via integration over the pointwise geodesic distances between ∇ϕ 1 (x) and ∇ϕ 2 (x) same body . More generally, we can define a difference measure between two inhomogeneous deformations ϕ 1 , under suitable regularity conditions for ϕ 1 , ϕ 2 (example if ϕ 1 , ϕ 2 are sufficiently smooth with det ∇ϕ i > 0 up to the boundary). This extension of the distance to inhomogeneous deformations is visualized in Fig. 7.
In order to find an appropriate Riemannian metric g on GL(n), we must discuss the required properties of this "difference measure". First, the requirements of objectivity (left-invariance) and isotropy (right-invariance) suggest that the metric g should be bi-O(n)-invariant, that is satisfy However, these requirements do not sufficiently determine a specific Riemannian metric. For example, (17) is satisfied by the metricǧ defined in (15) as well as by the metricǧ withǧ A (X, Y ) = A T X, A T Y . In order to rule out unsuitable metrics, we need to impose further restrictions on g. If we consider the distance measure dist(ϕ 1 , ϕ 2 ) between two deformations ϕ 1 , ϕ 2 introduced in (16), a number of further invariances can be motivated: if we require that the distance is not changed by the superposition of a homogeneous deformation, that is that for all A, B ∈ GL(n) and X, Y ∈ T A GL(n). The physical interpretation of this invariance requirement is readily visualized in Fig. 8. It can easily be shown [129] that a Riemannian metric g is left-GL(n)-invariant 17 as well as right-O(n)-invariant if and only if g is of the form where ·, · μ,μ c ,κ is the fixed inner product on the tangent space gl(n) = T 1 GL(n) = R n×n at the identity with for constant positive parameters μ, μ c , κ > 0, and where X, Y = tr(X T Y ) denotes the canonical inner product on gl(n) = R n×n . 18 A Riemannian metric g defined in this way behaves in the same way on all tangent spaces: for every 17 Of course, the left-GL(n)-invariance of a metric also implies the left-O(n)-invariance. 18 If μ = μ c = 1 and κ = 2 n , then the inner product ·, · μ,μ c ,κ is the canonical inner product, and the corresponding metric g is the canonical left-invariant metric on GL(n) with

Fig. 9. A left-GL(n)-invariant Riemannian metric on GL(n) transforms the tangent space at
A ∈ GL + (n) to the tangent space T 1 GL + (n) = gl(n) at the identity and applies a fixed inner product on gl(n) to the transformed tangents. Thus no tangent space is treated preferentially at the identity via the left-hand multiplication with A −1 and applies the fixed inner product ·, · μ,μ c ,κ on gl(n) to the transformed tangents, cf. Fig. 9.
In the following, we will always assume that GL(n) is endowed with a Riemannian metric of the form (19) unless indicated otherwise.
In order to find the geodesic distance of F ∈ GL + (n) to SO(n), we need to consider the geodesic curves on GL + (n). It has been shown [5,87,129,134] that every geodesic on GL + (n) with respect to the left-GL(n)-invariant Riemannian metric induced by the inner product (20) is of the form γ ξ with F ∈ GL + (n) and some ξ ∈ gl(n), where exp denotes the matrix exponential. 19 These curves are characterized by the geodesic equation 19 The mapping ξ → exp geod (ξ ) is also known as the geodesic exponential function at F. Note that in general , thus the geodesic curves are generally not one-parameter groups of the form t → F exp(t ξ), in contrast to bi-invariant metrics on Lie groups (for example SO(n) with the canonical bi-invariant metric [136]).
Since the geodesic curves are defined globally, GL + (n) is geodesically complete with respect to the metric g. We can therefore apply the Hopf-Rinow theorem [111,129] to find that for all F, P ∈ GL + (n) there exists a length minimizing geodesic γ ξ F connecting F and P. Without loss of generality, we can assume that γ and the length of the geodesic γ ξ F starting in F with initial tangent F ξ ∈ T F GL + (n) (cf. (21) and Fig. 11) is given by [129] The geodesic distance between F and P can therefore be characterized as that is the minimum of ξ μ,μ c ,κ over all ξ ∈ gl(n) which connect F and P, that is satisfy Although some numerical computations have been employed [216] to approximate the geodesic distance in the special case of the canonical left-GL(n)-invariant metric, that is for μ = μ c = 1, κ = 2 n , there is no known closed form solution to the highly nonlinear system (23) in terms of ξ for given F, P ∈ GL + (n) and thus no known method of directly computing dist geod (F, P) in the general case exists. However, this parametrization of the geodesic curves will still allow us to obtain a lower bound on the distance of F to SO(n).  11. The geodesic (intrinsic) distance to SO(n); neither the element Q of best approximation nor the initial tangent F ξ ∈ T F GL + (n) of the connecting geodesic is known beforehand

The geodesic distance to SO(n)
Having defined the geodesic distance on GL + (n), we can now consider the geodesic strain measure, which is the geodesic distance of the deformation gradient F to SO(n): Without explicit computation of this distance, the left-GL(n)-invariance and the right-O(n)-invariance of the metric g immediately allow us to show the inverse deformation symmetry of the geodesic strain measure: This symmetry property demonstrates at once that the Eulerian (spatial) and the Lagrangian (referential) points of view are equivalent with respect to the geodesic strain measure: in the Eulerian setting, the inverse F −1 of the deformation gradient appears more naturally, whereas F is used in the Lagrangian frame (cf. Fig.  10). 20 Equality (25) shows that both points of view can equivalently be taken if the geodesic strain measure is used. As we will see later on (Remark 3.5), the equality dist geod (B, SO(n)) = dist geod (C, SO(n)) also holds for the right Cauchy-Green deformation tensor C = F T F = U 2 and the Finger tensor B = F F T = V 2 , 20 Note that Cauchy originally introduced the tensors C −1 and B −1 in his investigations of the nonlinear strain [41,42,77,183], where C = F T F = U 2 is the right Cauchy-Green deformation tensor [77,81] and B = F F T = V 2 is the Finger tensor. Piola also formulated an early nonlinear elastic law in terms of C −1 , cf. [204, p. 347].
further indicating the independence of the geodesic strain measure from the chosen frame of reference. This property is, however, not unique to geodesic (or logarithmic) strain measures; for example, the Frobenius norm (4) and [17], which can be considered a "quasilogarithmic" strain measure, fulfils the inverse deformation symmetry as well. 21 However, it is not satisfied for the Euclidean distance to SO(n): in general, Now, let F = RU denote the polar decomposition of F with U ∈ Sym + (n) and R ∈ SO(n). In order to establish a simple upper bound on the geodesic distance dist geod (F, SO(n)), we construct a particular curve γ R connecting F to its orthogonal factor R ∈ SO(n) and compute its length L(γ R ). For It is easy to confirm that γ R is in fact a geodesic as given in (21) the length of γ R is given by We can thereby establish the upper bound 21 The quantity 1 for the geodesic distance of F to SO(n).
Our task in the remainder of this section is to show that the right hand side of inequality (29) is also a lower bound for the (squared) geodesic strain measure, that is that, altogether, However, while the orthogonal polar factor R is the element of best approximation in the Euclidean case (for μ = μ c = 1, κ = 2 n ) due to Grioli's Theorem, it is not clear whether R is indeed the element in SO(n) with the shortest geodesic distance to F (and thus whether equality holds in (28)). Furthermore, it is not even immediately obvious that the geodesic distance between F and R is actually given by the right hand side of (29), since a shorter connecting geodesic might exist [and hence inequality might hold in (29)].
Nonetheless, the following fundamental logarithmic minimization property of the orthogonal polar factor, combined with the computations in Section 3.1, allows us to show that (29) is indeed also a lower bound for dist geod (F, SO(n)). 22 is defined as the infimum of sym . over "all real matrix logarithms" of Q T F. Proposition 3.1, which can be seen as the natural logarithmic analogue of Grioli's Theorem (cf. Section 2.2), was first shown for dimensions n = 2, 3 by Neff et al. [157] using the so-called sum-of-squared-logarithms inequality [29,30,48,171]. A generalization to all unitarily invariant norms and complex logarithms for arbitrary dimension was given by Lankeit, Neff and Nakatsukasa [118]. We also require the following corollary involving the weighted Frobenius norm, which is not orthogonally invariant. 23 22 Of course, the application of such minimization properties to elasticity theory has a long tradition: Leonhard Euler, in the appendix "De curvis elasticis" to his 1744 book "Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes sive solutio problematis isoperimetrici latissimo sensu accepti" [62,165], already proclaimed that "[…] since the fabric of the universe is most perfect, and is the work of a most wise creator, nothing whatsoever takes place in the universe in which some rule of maximum and minimum does not appear." 23 While Q T X Q μ,μ c ,κ = X μ,μ c ,κ for all X ∈ R n×n and Q ∈ O(n), the orthogonal invariance requires the equalities Q X μ,μ c ,κ = X Q μ,μ c ,κ = X μ,μ c ,κ , which do not hold in general.

Corollary 3.2. Let
Proof. We first note that the equality det exp(X ) = e tr(X ) holds for all X ∈ R n×n . Since det Q = 1 for all Q ∈ SO(n), this implies that for all X ∈ R n×n with exp(X ) = Q T F, Note that Corollary 3.2 also implies the slightly weaker statement We are now ready to prove our main result. 24 Observe that μ dev n Y 2

Theorem 3.3. Let g be the left-GL(n)-invariant, right-O(n)-invariant Riemannian metric on GL(n) defined by
Then for all F ∈ GL + (n), the geodesic distance of F to the special orthogonal group SO(n) induced by g is given by where log is the principal matrix logarithm, tr(X ) = n i=1 X i,i denotes the trace and dev n X = X − 1 n tr(X ) · 1 is the n-dimensional deviatoric part of X ∈ R n×n . The orthogonal factor R ∈ SO(n) of the polar decomposition F = RU is the unique element of best approximation in SO(n), that is In particular, the geodesic distance does not depend on the spin modulus μ c .

Remark 3.5.
Since the weighted Frobenius norm on the right hand side of equation (32) only depends on the eigenvalues of U = √ F T F, the result can also be expressed in terms of the left Biot-stretch tensor V = √ F F T , which has the same eigenvalues as U : Applying the above formula to the case F = P with P ∈ Sym + (n), we find √ P T P = √ P P T = P and therefore since 1 is the orthogonal polar factor of P. For the tensors U and V , the right Cauchy-Green deformation tensor C = F T F = U 2 and the Finger tensor B = F F T = V 2 , we thereby obtain the equalities = dist geod (U, 1) = dist geod (U −1 , 1) = dist geod (U, SO(n)).
Note carefully that, although (34) for P ∈ Sym + (n) immediately follows from Theorem 3.3, it is not trivial to compute the distance dist geod (P, 1) directly: while the curve given by exp(t log P) for t ∈ [0, 1] is in fact a geodesic [87] connecting 1 to P with squared length μ dev n log P 2 + κ 2 [tr(log P)] 2 , it is not obvious whether or not a shorter connecting geodesic might exist. Our result ensures that this is in fact not the case.
Proof. (Proof of Theorem 3.3) Let F ∈ GL + (n) and Q ∈ SO(n). Then according to our previous considerations (cf. Section 3.1) there exists ξ ∈ gl(n) with and In order to find a lower estimate on ξ μ,μ c ,κ (and thus on dist geod (F, Q)), we compute Since exp(W ) ∈ SO(n) for all skew symmetric W ∈ so(n), we find with Q ξ = Q exp(−(1 + μ c μ ) skew ξ ) ∈ SO(n); note that sym Y = − sym ξ . According to (39), Y = − sym ξ + μ c μ skew ξ is "a logarithm" of Q T ξ F. 25 The weighted Frobenius norm of the symmetric part of Y = − sym ξ + μ c μ skew ξ is therefore bounded below by the infimum of sym X μ,μ c ,κ over "all logarithms" X of Q T ξ F: We can now apply Corollary 3.2 to find inf for U = √ F T F. Since this inequality is independent of Q and holds for all Q ∈ SO(n), we obtain the desired lower bound on the geodesic distance of F to SO(n). Together with the upper bound already established in (29), we finally find By equation (42), apart from computing the geodesic distance of F to SO(n), we have shown that the orthogonal polar factor R = polar(F) is an element of best approximation to F in SO(n). However, it is not yet clear whether there exists another element of best approximation, that is whether there is a Q ∈ SO(n) with Q = R and dist geod (F, Q) = dist geod (F, R) = dist geod (F, SO(n)). For this purpose, we need to compare geodesic distances corresponding to different parameters μ, μ c , κ. We therefore introduce the following notation: for fixed μ, μ c , κ > 0, let dist geod,μ,μ c ,κ denote the geodesic distance on GL + (n) induced by the left-GL(n)invariant, right-O(n)-invariant Riemannian metric g [as introduced in (19)] with parameters μ, μ c , κ. Furthermore, the length of a curve γ with respect to this metric will be denoted by L μ,μ c ,κ (γ ).
Assume that Q ∈ SO(n) is an element of best approximation to F with respect to g for some fixed parameters μ, μ c , κ > 0. Then there exists a length minimizing geodesic γ : [0, 1] → GL + (n) connecting Q to F of the form with ξ ∈ R n×n , and the length of γ is given by We first assume that skew ξ = 0. We chooseμ c > 0 withμ c < μ c and find dist 2 geod,μ,μ c ,κ (F, SO(n)) = inf since γ is a curve connecting F to Q ∈ SO(n); note that although γ is a shortest connecting geodesic with respect to parameters μ, μ c , κ by assumption, it need not necessarily be a length minimizing curve with respect to parameters μ,μ c , κ.
The fact that the orthogonal polar factor R = polar(F) is the unique element of best approximation to F in SO(n) with respect to the geodesic distance corresponds directly to the linear case [cf. equality (8) in Section 2.1], where the skew symmetric part skew ∇u of the displacement gradient ∇u is the element of best approximation with respect to the Euclidean distance: for F = 1 + ∇u we have hence the linear approximation of the orthogonal and the positive definite factor in the polar decomposition is given by skew ∇u and sym ∇u, respectively. The geometric connection between the geodesic distance on GL + (n) and the Euclidean distance on the tangent space R n×n = gl(n) at 1 is illustrated in Fig. 12.
The right-GL(n)-invariant Riemannian metric can be motivated in a way similar to the left-GL(n)-invariant case: it corresponds to the requirement that the distance between two deformations F 1 and F 2 should not depend on the initial shape of , meaning it should not be changed if is homogeneously deformed beforehand (cf. Fig. 13). A similar independence from prior deformations (and so-called "pre-stresses"), called "elastic determinacy" by Prandtl [172], was postulated by Hencky in the deduction of his elasticity model; cf. [ According to Theorem 3.3, the squared geodesic distance between F and SO(n) with respect to any left-GL(n)-invariant, right-O(n)-invariant Riemannian metric on GL(n) is the isotropic quadratic Hencky energy where the parameters μ, κ > 0 represent the shear modulus and the bulk modulus, respectively. The Hencky energy function was introduced in 1929 by Hencky [101], who derived it from geometrical considerations as well: his deduction was based on a set of axioms including a law of superposition (cf. Section 4.2) for Fig. 13. The right-GL(n)-invariance of a distance measure on GL(n): the distance between two homogeneous deformations F 1 , F 2 is not changed by a prior homogeneous deformation the stress response function [147], an approach previously employed by Becker [18,153] in 1893 and later followed in a more general context by Richter [177], cf. [176,178,179]. 26 A different constitutive model for uniaxial deformations based on logarithmic strain had previously been proposed by Imbert [114] and Hartig [89]. While Ludwik is often credited with the introduction of the uniaxial logarithmic strain, his ubiquitously cited article [124] (which is even referenced by Hencky himself [102, p. 175]) does not provide a systematic introduction of such a strain measure.
While the energy function W H (F) = dist 2 geod (F, SO(n)) already defines a measure of strain as described in Section 1.1, we are also interested in characterizing the two terms dev n log U and |tr(log U )| as separate partial strain measures.  26 Hencky's approach is often misrepresented as empirically motivated. Truesdell claims that "Hencky himself does not give a systematic treatement" in introducing the logarithmic strain tensor [200, p. 144] and attributes the axiomatic approach to Richter [177] instead [205, p. 270]. Richter's resulting deviatoric strain tensors dev 3 log U and dev 3 log V are disqualified as "complicated algebraic functions" by Truesdell and Toupin [205, p. 270]. Then and where the geodesic distances dist geod, SL(n) and dist geod, R + ·1 on the Lie groups SL(n) = {A ∈ GL(n) | det A = 1} and R + · 1 are induced by the canonical left-invariant metric Remark 3.8. Theorem 3.7 states that ω iso and ω vol appear as natural measures of the isochoric and volumetric strain, respectively: if F = F iso F vol is decomposed multiplicatively [73] into an isochoric part F iso = (det F) −1/n · F and a volumetric part F vol = (det F) 1/n · 1, then ω iso (F) measures the SL(n)-geodesic distance of F iso to SO(n), whereas 1 √ n ω vol (F) gives the geodesic distance of F vol to the identity 1 in the group R + · 1 of purely volumetric deformations.
Proof. First, observe that the canonical left-invariant metrics on SL(n) and R + · 1 are obtained by choosing μ = μ c = 1 and κ = 2 n and restricting the corresponding metric g on GL + (n) to the submanifolds SL(n), R + · 1 and their respective tangent spaces. Then for this choice of parameters, every curve in SL(n) or R + · 1 is a curve of equal length in GL + (n) with respect to g. Since the geodesic distance is defined as the infimal length of connecting curves, this immediately implies dist geod, SL(n) (F iso , SO(n)) dist geod, GL + (n) (F iso , SO(n)) as well as dist geod, R + ·1 (F vol , 1) dist geod, GL + (n) (F vol , 1) dist geod, GL + (n) (F vol , SO(n)) for F iso := (det F) −1/n · F and F vol := (det F) 1/n ·1. We can therefore use Theorem 3.3 to obtain the lower bounds and 27 To obtain an upper bound on the geodesic distances, we define the two curves where F = RU with R ∈ SO(n) and U ∈ Sym + (n) is the polar decomposition of F. Then γ iso connects (det F) −1/n · F to SO(n): while γ vol connects (det F) 1/n · 1 and 1: The lengths of the curves compute to as well as showing that which completes the proof.

Remark 3.9.
In addition to the isochoric (distortional) part F iso = (det F) −1/n · F and the volumetric part F vol = (det F) 1/n · 1, we may also consider the cofactor Cof F = (det F)· F −T of F ∈ GL + (n). Theorem 3.3 allows us to directly compute (cf. Appendix A.4) the distance

Riemannian geometry applied to Sym + (n)
Extensive work on the use of Lie group theory and differential geometry in continuum mechanics has already been done by Rougée [181][182][183][184], Moakher [137,139], Bhatia [26] and, more recently, by Fiala [64][65][66][67][68] (cf. [119,120,164,167,168]). They all endowed the convex cone Sym + (3) of positive definite symmetric (3 × 3)-tensors with the Riemannian metric where C ∈ Sym + (3) and X, Y ∈ Sym(3) = T C Sym + (3). 28 Fiala and Rougée deduced a motivation of the logarithmic strain tensor log U via geodesic curves connecting elements of Sym + (n). However, their approach differs markedly from our method employed in the previous sections: the manifold Sym + (n) already corresponds to metric states C = F T F, whereas we consider the full set GL + (n) of deformation gradients F (cf. Appendix A.3 and Table 1 in Section 6). This restriction can be viewed as the nonlinear analogue of the a priori restriction to ε = sym ∇u in the linear case, which means that the nature of the strain measure is not deduced but postulated. Note also that the metric g cannot be obtained by restricting our left-GL(3)-invariant, right-O(3)-invariant metric g to Sym + (3). 29 Furthermore, while Fiala and Rougée aim to motivate the Hencky strain tensor log U directly, our focus lies on the strain measures ω iso , ω vol and the isotropic Hencky strain energy W H . The geodesic curves on Sym + (n) with respect to g are of the simple form with C 1 ∈ Sym + (n) and M ∈ Sym(n) = T C 1 Sym + (n). 30 These geodesics are defined globally, that is Sym + (n) is geodesically complete. Furthermore, for given C 1 , C 2 ∈ Sym + (n), there exists a unique geodesic curve connecting them; this easily follows from the representation formula (51) or from the fact that the curvature of Sym + (n) with g is constant and negative [25,65,116]. Note that this implies that, in contrast to GL + (n) with our metric g, there are no closed geodesics on Sym + (n). An explicit formula for the corresponding geodesic distance was given by Moakher: 31 dist geod, Sym + (n) (C 1 , C 2 ) = log(C −1/2 2 28 Note the subtle difference with our metric g C (X, Y ) = C −1 X, C −1 Y . Pennec [167, p. 368] generalizes (50) by using the weighted inner product X, Y * = X, Y + β tr(X ) tr(Y ) with β > − 1 n . 29 Since Sym + (n) is not a Lie group with respect to matrix multiplication, the metric g itself cannot be left-or right-invariant in any suitable sense. 30 While Moakher gives the parametrization stated here, Rougée writes the geodesics in the form γ (t) = exp(t · Log(C 2 C −1 1 )) C 1 with C 1 , C 2 ∈ Sym + (n), which can also be written as γ (t) = (C 2 C −1 1 ) t C 1 ; a similar formulation is given by Tarantola 31 Moakher [137, eq. (2.9)] writes this result as Log(C −1 2 C 1 ) = n i=1 ln 2 λ i , where λ i are the eigenvalues of C −1 2 C 1 . The right hand side of this equation is identical to the result stated in (52). However, since C −1 2 C 1 is not necessarily normal, there is in general no logarithm Log(C −1 2 C 1 ) whose Frobenius norm satisfies this equality. Note that the eigenvalues of the matrix C −1 2 C 1 are real and positive due to its similarity to C 1/2 In the special case C 2 = 1, this distance measure is equal to our geodesic distance on GL + (n) induced by the canonical inner product: Theorem 3.3, applied with parameters μ = μ c = 1 and κ = 2 n to R = 1 and U = C 1 , shows that dist geod, GL + (n) (C 1 , 1) = log C 1 = dist geod, Sym + (n) (C 1 , 1).
More generally, assume that the two metric states C 1 , C 2 ∈ Sym + (n) commute. Then C −1 2 C 1 ∈ Sym + (n), and the left-GL(n)-invariance of the geodesic distance implies However, since C −1 2 C 1 / ∈ Sym + (n) in general, this equality does not hold on all of Sym + (n).
A different approach towards distance functions on the set Sym + (n) was suggested by Arsigny et al. [7][8][9] who, motivated by applications of geodesic and logarithmic distances in diffusion tensor imaging, directly define their Log-Euclidean metric on Sym + (n) by where . is the Frobenius matrix norm. If C 1 and C 2 commute, this distance equals the geodesic distance on GL + (n) as well: where equality in (55) holds due to the fact that C 1 and C −1 2 commute. Again, this equality does not hold for arbitrary C 1 and C 2 .
Using a similar Riemannian metric, geodesic distance measures can also be applied to the set of positive definite symmetric fourth-order elasticity tensors, which can be identified with Sym + (6). Norris and Moakher applied such a distance function in order to find an isotropic elasticity tensor C : Sym(3) → Sym(3) which best approximates a given anisotropic tensor [138,158].
The connection between geodesic distances on the metric states in Sym + (n) and logarithmic distance measures was also investigated extensively by the late Albert Tarantola [198], a lifelong advocate of logarithmic measures in physics. In his view [198, 4.3.1], "…the configuration space is the Lie group GL + (3), and the only possible measure of strain (as the geodesics of the space) is logarithmic."

Further mechanical motivations for the quadratic isotropic Hencky model based on logarithmic strain tensors "At the foundation of all elastic theories lies the definition of strain, and before introducing a new law of elasticity we must explain how finite strain is to be measured."
Heinrich Hencky: The elastic behavior of vulcanized rubber [103].
Apart from the geometric considerations laid out in the previous sections, the Hencky strain tensor E 0 = log U can be characterized via a number of unique properties.
For example, the Hencky strain is the only strain tensor (for a suitably narrow definition, cf. [153]) that satisfies the law of superposition for coaxial deformations: for all coaxial stretches U 1 and U 2 , that is U 1 , U 2 ∈ Sym + (n) such that U 1 · U 2 = U 2 ·U 1 . This characterization was used by Heinrich Hencky [97,102,103,197] in his original introduction of the logarithmic strain tensor [99][100][101]147] and, indeed much earlier, by the geologist George Ferdinand Becker [133], who postulated a similar law of superposition in order to deduce a logarithmic constitutive law of nonlinear elasticity [18,153] (cf. Appendix A.2). In the case n = 1, this superposition principle simply amounts to the fact that the logarithm function f = log satisfies Cauchy's [40] well-known functional equation or, in other words, that the logarithm is an isomorphism between the multiplicative group (R + , ·) and the additive group (R, +). This means that for a sequence of incremental one-dimensional deformations, the logarithmic strains e i log can be added in order to obtain the total logarithmic strain e tot log of the composed deformation [72]: log + e 2 log + · · · + e n log = log where L i denotes the length of the (one-dimensional) body after the i-th elongation. This property uniquely characterizes the logarithmic strain e log among all differentiable one-dimensional strain mappings e : R + → R with e (1) = 1.
Since purely volumetric deformations of the form λ·1 with λ > 0 are coaxial to every stretch U ∈ Sym + (n), the decomposition property (56) allows for a simple additive volumetric-isochoric split of the Hencky strain tensor [177]: In particular, the incompressibility condition det F = 1 can be easily expressed as tr(log U ) = 0 in terms of the logarithmic strain tensor.

From Truesdell's hypoelasticity to Hencky's hyperelastic model
As indicated in Section 1.1, the quadratic Hencky energy is also of great importance to the concept of hypoelasticity [83,Chapter IX]. It was found that the Truesdell equation 32 [76,[200][201][202] d dt with constant Lamé coefficients μ, λ > 0, under the assumption that the stress rate d dt is objective 33 and corotational, is satisfied if and only if d dt is the so-called logarithmic corotational rate d dt log and τ = 2 μ log V + λ tr(log V ) · 1 [159,173,174,210,[212][213][214][215], that is if and only if the hypoelastic model is exactly Hencky's hyperelastic constitutive model. Here, τ = det F · σ (V ) denotes the Kirchhoff stress tensor and D is the unique rate of stretching tensor (that is the symmetric part of the velocity gradient in the spatial setting). A rate d dt is called corotational if it is of the special form which means that the rate is computed with respect to a frame that is rotated. 34 This extra rate of rotation is defined only by the underlying spins of the problem. Upon specialization, for μ = 1, λ = 0 we obtain [ 32 It is telling to see that equation (58) had already been proposed by Hencky himself in [100] for the Zaremba-Jaumann stress rate [cf. (62)]. Hencky's work, however, contains a typographical error [100, eq. (10) and eq. (11e)] changing the order of indices in his equations (cf. [33]). The strong point of writing (58) is that no discussion of any suitable strain tensor is necessary. 33 is the spatial Almansi strain tensor and d dt is the upper Oldroyd rate [as defined in (63)]. The quadratic Hencky model was generalized in Hill's generalized linear elasticity laws [108, eq. (2.69)] with work-conjugate pairs (T r , E r ) based on the Lagrangian strain measures given in (3); cf. Appendix A.
2 for examples. 36 The concept of work-conjugacy was introduced by Hill [106] via an invariance requirement; the spatial stress power must be equal to its Lagrangian counterpart: by means of which a material stress tensor is uniquely linked to its (material rate) conjugate strain tensor. Hence it generalizes the virtual work principle and is the foundation of derived methods like the finite element method. For the case of isotropic materials, Hill [106, p. 242] (cf. [109]) shows by spectral decomposition techniques that the work-conjugate stress to log U is the backrotated Cauchy stress σ multiplied by det F, hence σ, D = R T σ R, d dt log U , which is a generalization of Hill's earlier work [106,108]. Sansour [186] additionally found that the Eshelby-like stress tensor = C S 2 is equally conjugate to log U ; here, S 2 denotes the second Piola-Kirchhoff stress tensor. For anisotropy, however, the conjugate stress exists but follows a much more complex format than for isotropy [109]. The logarithm of the left stretch log V in contrast exhibits a work conjugate stress tensor only for isotropic materials, namely the Kirchhoff stress tensor τ = det F · σ [109,163].
While hyperelasticity in its potential format avoids rate equations, the use of stress rates (that is stress increments in time) may be useful for the description of inelastic material behavior at finite strains. Since the material time derivative of an Eulerian stress tensor is not objective, rates for a tensor X were developed, like the (objective and corotational) Zaremba-Jaumann rate d dt or the (objective but not corotational) lower and upper Oldroyd rates 36 Hooke's law [110] (cf. [141]) famously states that the strain in a deformation depends linearly on the occurring stress ("ut tensio, sic vis"). However, for finite deformations, different constitutive laws of elasticity can be obtained from this assumption, depending on the choice of a stress/strain pair. An idealized version of such a linear relation is given by (60), that is by choosing the spatial Hencky strain tensor log V and the Kirchhoff stress tensor τ . Since, however, Hooke speaks of extension versus force, the correct interpretation of Hooke's law is T Biot = 2 μ (U − 1) + λ tr(U − 1) · 1, that is the case r = 1 2 in (61).
to name but a few (cf. [90,Section 1.7] and [187]). Which one of these or the great number of other objective rates should be used seems to be rather a matter of taste, hence of arbitrariness 37 or heuristics, 38 but not a matter of theory. The concept of dual variables 39 as introduced by Tsakmakis and Haupt in [91] into continuum mechanics overcame the arbitrariness of the chosen rate in that it uniquely connects a particular (objective) strain rate to a stress tensor and, analogously, a stress rate to a strain tensor. The rational rule is that, when stress and strain tensors operate on configurations other than the reference configurations, the physically significant scalar products S 2 ,Ė 1 , Ṡ 2 , E 1 , S 2 , E 1 and Ṡ 2 ,Ė 1 (with the second Piola-Kirchhoff stress tensor S 2 and its work-conjugate Green strain tensor E 1 ) must remain invariant, see [90,91].

Advantageous properties of the quadratic Hencky energy
For modeling elastic material behavior there is no theoretical reason to prefer one strain tensor over another one, and the same is true for stress tensors. As discussed in Section 1.1, stress and strain are immaterial. 40 Primary experimental data (forces, displacements) in material testing are sufficient to calculate any strain tensor and any stress tensor and to display any combination thereof in stress-strain curves, while only workconjugate pairs are physically meaningful.
However, for modeling finite-strain elasticity, the quadratic Hencky model exhibits a number of unique, favorable properties, including its functional simplicity and its dependency on only two material parameters μ and κ that are determined in the infinitesimal strain regime and remain constant over the entire strain range. In view of the linear dependency of stress from logarithmic strain in (64), it is obvious that any nonlinearity in the stress-strain curves can only be captured in Hencky's model by virtue of the nonlinearity in the strain tensor itself. There is a surprisingly large number of different materials, where Hencky's elasticity relation provides a very good fit to experimental stress-strain data, which is true for different length 37 Truesdell and Noll [204, p. 404] declared that "various such stress rates have been used in the literature. Despite claims and whole papers to the contrary, any advantage claimed for one such rate over another is pure illusion", and that "the properties of a material are independent of the choice of flux [that is of the chosen rate], which, like the choice of a [strain tensor], is absolutely immaterial" [204, p. 97]. 38 For a shear test in Eulerian elasto-plasticity using the Zaremba-Jaumann rate (62), an unphysical artefact of oscillatory shear stress was observed, first in [122]. A similar oscillatory behavior was observed for hypoelasticity in [52]. 39 Hill [108] used the terms conjugate and dual as synonyms. 40 Cf. Truesdell [200, p. 145]: "It is important to realize that since each of the several material tensors […] is an isotropic function of any one of the others, an exact description of strain in terms of any one is equivalent to a description in terms of any other" or Antman [6, p. 423]: "In place of C, any invertible tensor-valued function of C can be used as a measure of strain." Rivlin [180] states that strain need never be defined at all, cf. [204, p. 122]. scales and strain regimes. In the following we substantiate this claim with some examples.
Nonlinear elasticity on macroscopic scales for a variety of materials. Anand [3,4] has shown that the Hencky model is in good agreement with experiments on a wide class of materials, as for example vulcanized natural rubber, for principal stretches between 0.7 and 1.3. More precisely, this refers to the characteristic that in tensile deformation the stiffness becomes increasingly smaller compared with the stiffness at zero strain, while for compressive deformation the stiffness becomes increasingly larger.
Nonlinear elasticity in the very small strain regime. We mention in passing that a qualitatively similar dependency of material stiffness on the sign of the strain has been made much earlier in the regime of extremely small strains (10 −6 -10 −3 ). In Hartig's law [89] from 1893 this dependency was expressed as dσ dε = E 0 + b σ , where E 0 is the elasticity modulus at zero stress and b < 0 is a dimensionless constant, cf. the book of Bell [19] and [126] in the context of linear elasticity with initial stress. 41 Hartig also observed that the stress-stretch relation should have negative curvature in the vicinity of the identity, as shown in Fig. 14. 42 41 The negative curvature (b < 0) was already suggested by Jacob Bernoulli [21] (cf. [20, p. 276]): "Homogeneous fibers of the same length and thickness, but loaded with different weights, neither lengthen nor shorten proportional to these weights; but the lengthening or the shortening caused by the small weight is less than the ratio that the first weight has to the second." 42  Crystalline elasticity on the nanoscale. Quite in contrast to the strictly stressbased continuum constitutive modeling, atomistic theories are based on a concept of interatomic forces. These forces are derived from potentials V according to the potential relation f a = −∂ x a V, which endows the model with a variational structure. 43 A further discussion of hybrid, atomistic-continuum coupling can be found in [60]. Thereby the discreteness of matter at the nanoscale and the nonlocality of atomic interactions are inherently captured. Here, atomistic stress is neither a constitutive agency nor does it enter a balance equation. Instead, it optionally can be calculated following the virial stress theorem [196,Chapter 8] to illustrate the state of the system.
With their analyses in [53] and [54], DŁuzewski and coworkers aim to link the atomistic world to the macroscopic world of continuum mechanics. They search for the "best" strain measure with a view towards crystalline elasticity on the nanoscale. The authors consider the deformation of a crystal structure and compare the atomistic and continuum approaches. Atomistic calculations are made using the Stillinger-Weber potential. The stress-strain behavior of the best-known anisotropic hyperelastic models are compared with the behavior of the atomistic one in the uniaxial deformation test. The result is that the anisotropic energy based on the Hencky strain energy 1 2 C. log U, log U , where C is the anisotropic elasticity tensor from linear elasticity, gives the best fit to atomistic simulations. More in detail, this best fit manifests itself in the observation that for considerable compression (up to ≈20 %) the material stiffness is larger than the reference stiffness at zero strain, and for considerable tension (up to ≈20 %) it is smaller than the zero-strain stiffness, again in good agreement with the atomistic result. This is also corroborated by comparing tabulated experimentally determined third order elastic constants [53]. 44 Elastic energy potentials based on logarithmic strain have also recently been motivated via molecular dynamics simulations [93] by Henann and Anand [94].

The exponentiated Hencky energy
As indicated in Section 1.1 and shown in Sections 2.1 and 3, strain measures are closely connected to isotropic energy functions in nonlinear hyperelasticity: similarly to how the linear elastic energy may be obtained as the square of the 43 For molecular dynamics (MD) simulations, a well-established level of sophistication is the modeling by potentials with environmental dependence (pair functionals like in the embedded atom method (EAM) account for the energy cost to embed atomic nuclei into the electron gas of variable density) and angular dependence (like for Stillinger-Weber or Tersoff functionals). 44 Third order elastic constants are corrections to the elasticity tensor in order to improve the response curves beyond the infinitesimal neighborhood of the identity. They exist as tabulated values for many materials. Their numerical values depend on the choice of strain measure used which needs to be corrected. DŁuzewski [53] shows that again the Henckystrain energy 1 2 C. log U, log U provides the best overall approximation.
Euclidean distance of ∇u to so(n), the nonlinear quadratic Hencky strain energy is the squared Riemannian distance of ∇ϕ to SO(n). For the partial strain measures ω iso (F) = dev n log √ F T F and ω vol (F) = |tr(log √ F T F)| defined in Theorem 3.7, the Hencky strain energy W H can be expressed as However, it is not at all obvious why this weighted squared sum should be viewed as the "canonical" energy associated with the geodesic strain measures: while it is reasonable to view the elastic energy as a quantity depending on some strain measure alone, the specific form of this dependence must not be determined by purely geometric deductions, but must take into account physical constraints as well as empirical observations. 45 For a large number of materials, the Hencky energy does indeed provide a very accurate model up to moderately large elastic deformations [3,4], that is up to stretches of about 40 %, with only two constant material parameters which can be easily determined in the small strain range. For very large strains, however, the subquadratic growth of the Hencky energy in tension is no longer in agreement with empirical measurements. 46 In a series of articles [80,[154][155][156], Neff et al. have therefore introduced the exponentiated Hencky energy with additional dimensionless material parameters k 1 4 and k 1 8 , which for all values of k, k approximates W H for deformation gradients F sufficiently close to the identity 1, but shows a vastly different behavior for F → ∞, cf. Fig. 15.
The exponentiated Hencky energy has many advantageous properties over the classical quadratic Hencky energy; for example, W eH is coercive on all Sobolev spaces W 1, p for 1 p < ∞, thus cavitation is excluded [12,143]. In the planar case n = 2, W eH is also polyconvex [80,156] and thus Legendre-Hadamard-elliptic [10], whereas the classical Hencky energy is not even LH-elliptic (rank-one convex) outside a moderately large neighborhood of 1 [36,145] (see also [113], where the 45 Leibniz, in a letter to Jacob Bernoulli [123, p. 572], stated as early as 1690 that "the [constitutive] relation between extension and stretching force should be determined by experiment", cf. [19, p. 10]. 46 The elastic range of numerous materials, including vulcanized rubber or skin and other soft tissues, lies well above stretches of 40 %. While the behavior of elasticity models for extremely large strains might not seem important due to physical restraints and intermingling plasticity effects outside a narrow range of perfect elasticity, it is nevertheless important to formulate an idealized law of elasticity over the whole range of deformations; cf. Hencky [99, p. 215] (as translated in [147, p.2]): "It is not important that such an idealized elastic [behavior] does not actually exist and our ideally elastic material must therefore remain an ideal. Like so many mathematical and geometric concepts, it is a useful ideal, because once its deducible properties are known it can be used as a comparative rule for assessing the actual elastic behavior of physical bodies." loss of ellipticity for energies of the form dev 3 log U β with hardening index 0 < β < 1 are investigated). Therefore, many results guaranteeing the existence of energy-minimizing deformations for a variety of boundary value problems can be applied directly to W eH for n = 2.
Furthermore, W eH satisfies a number of constitutive inequalities [155] such as the Baker-Ericksen inequality [127], the pressure-compression inequality and the tension-extension inequality as well as Hill's inequality [107,161,162], which is equivalent to the convexity of the elastic energy with respect to the logarithmic strain tensor [193]. 47 47 Hill's inequality [162] can be stated more generally as d dt  [194, p. 309]." Moreover, for W eH , the Cauchy-stress-stretch relation V → σ eH (V ) is invertible (a property hitherto unknown for other hyperelastic formulations) and pure Cauchy shear stress corresponds to pure shear strain, as is the case in linear elasticity [155]. The physical meaning of Poisson's ratio [79,170] is also similar to the linear case; for example, ν = 1 2 directly corresponds to incompressibility of the material and ν = 0 implies that no lateral extension or contraction occurs in uniaxial tensions tests.

Related geodesic distances
The logarithmic distance measures obtained in Theorems 3.3 and 3.7 show a strong similarity to other geodesic distance measures on Lie groups. For example, consider the special orthogonal group SO(n) endowed with the canonical biinvariant Riemannian metriĉ for Q ∈ SO(n) and X, Y ∈ T Q SO(n) = Q·so(n). 48 Then the geodesic exponential at 1 ∈ SO(n) is given by the matrix exponential on the Lie algebra so(n), that is all geodesic curves are one-parameter groups of the form with Q ∈ SO(n) and A ∈ so(n) (cf. [136]). It is easy to show that the geodesic distance between Q, R ∈ SO(n) with respect to this metric is given by where . is the Frobenius matrix norm and log : SO(n) → so(n) denotes the principal matrix logarithm on SO(n), which is uniquely defined by the equality exp(log Q) = Q and the requirement λ i (log Q) ∈ (−π, π] for all Q ∈ SO(n) and all eigenvalues λ i (log Q).
This result can be extended to the geodesic distance on the conformal special orthogonal group CSO(n) consisting of all angle-preserving linear mappings: where the bi-invariant metric g CSO(n) is given by the canonical inner product: Then where log again denotes the principal matrix logarithm on SO(n). Note that the punctured complex plane C\{0} can be identified with CSO(2) via the mapping

Outlook
While first applications of the exponentiated Hencky energy, which is based on the partial strain measures ω iso , ω vol introduced here, show promising results, including an accurate modeling of so-called tire-derived material [140,144], a more thorough fitting of the new parameter set to experimental data is necessary in order to assess the range of applicability of W eH towards elastic materials like vulcanized rubber. A different formulation in terms of the partial strain measures ω iso and ω vol , that is an energy function of the form with : [0, ∞) 2 → [0, ∞), might even prove to be polyconvex in the threedimensional case. The main open problem of finding a polyconvex (or rank-one convex) isochoric energy function F → ( dev 3 log U ) has also been considered by Sendova and Walton [190]. 49 Note that while every isotropic elastic energy W can be expressed as W (F) = h(K 1 , K 2 , K 3 ) with Criscione's invariants 50 [45,46,51,209] not every elastic energy has a representation of the form (68); for example, (68) implies the tension-compression symmetry 51 W (F) = W (F −1 ), which is not 49 Ideally, the function should also satisfy additional requirements, such as monotonicity, convexity and exponential growth. 50 The invariants K 1 and K 2 2 = tr (dev 3 log U ) 2 as well as K 3 = tr (dev 3 log U ) 3 had already been discussed exhaustively by Richter in a 1949 ZAMM article [177, §4], while K 1 and K 2 have also been considered by Lurie [125, p. 189]. Criscione has shown that the invariants given in (69) enjoy a favorable orthogonality condition which is useful when determining material parameters. 51 The tension-compression symmetry is often expressed as τ (V −1 ) = −τ (V ), where τ (V ) is the Kirchhoff stress tensor corresponding to the left Biot stretch V . This condition, which is the natural nonlinear counterpart of the equality σ (−ε) = −σ (ε) in linear elasticity, is equivalent to the condition W (F −1 ) = W (F) for hyperelastic constitutive models. Fig. 17. The tension-compression symmetry for incompressible materials: if det ∇ϕ ≡ 1 and necessarily satisfied by energy functions in general. 52 In terms of the Shield transformation 53 [39,192] the tension-compression symmetry amounts to the requirement 1 det F W * (F) = W (F) or, for incompressible materials, W * (F) = W (F). Moreover, under the assumption of incompressibility, the symmetry can be immediately extended to arbitrary deformations ϕ : → ϕ( ) and ϕ −1 : ϕ( ) → : if det ∇ϕ ≡ 1, we can apply the substitution rule to find , thus the total energies of the deformations ϕ, ϕ −1 are equal, cf. Fig. 17.
Since the function 52 Truesdell and Noll [204, p. 174] argue that "…there is no foundation for the widespread belief that according to the theory of elasticity, pressure and tension have equal but opposite effects". Examples for isotropic energy functions which do not satisfy this symmetry condition in general but only in the incompressible case can be found in [92]. For an idealized isotropic elastic material, however, the tension-compression symmetry is a plausible requirement (with an obvious additive counterpart in linear elasticity), especially for incompressible bodies. 53 Further properties of the Shield transformation can be found in [194, p.288]; for example, it preserves the polyconvexity, quasiconvexity and rank-one convexity of the original energy.
in planar elasticity is polyconvex [80,156], it stands to reason that a similar formulation in the three-dimensional case might prove to be polyconvex as well. A first step towards finding such an energy is to identify where the function W with which is not rank-one convex [155], loses its ellipticity properties. For that purpose, it may be useful to consider the quasiconvex hull of W . There already are a number of promising results for similar energy functions; for example, the quasiconvex hull of the mapping can be explicitly computed [56,57,195], and the quasiconvex hull of the similar Saint-Venant-Kirchhoff energy W SVK (F) = μ 4 C − 1 2 + λ 8 [tr(C − 1)] 2 has been given by Le Dret and Raoult [121]. For the mappings with n 2, however, no explicit representation of the quasiconvex hull is yet known, although it has been shown that both expressions are not rank-one convex [24].
It might also be of interest to calculate the geodesic distance dist geod (A, B) for a larger class of matrices A, B ∈ GL + (n): 54 although Theorem 3.3 allows us to explicitly compute the distance dist geod (1, P) for P ∈ Sym + (n) and local results are available for certain special cases [129], it is an open question whether there is a general formula for the distance dist geod, GL + (n) (Q, R) between arbitrary rotations R, Q ∈ SO(n) for all parameters μ, μ c , κ > 0. Since restricting our left-GL(n)invariant, right-O(n)-invariant metric on GL(n) to SO(n) yields a multiple of the canonical bi-SO(n)-invariant metric on SO(n), we can compute if for all Q, R ∈ SO(n) a shortest geodesic in GL + (n) connecting Q and R is already contained within SO(n), cf. Fig. 18. However, whether this is the case depends on the chosen parameters μ, μ c ; a general closed-form solution for dist geod, GL + (n) on SO(n) is therefore not yet known [128]. Moreover, it is not known whether our result can be generalized to anisotropic Riemannian metrics, that is if the geodesic distance to SO(n) can be explicitly computed for a larger class of left-GL(n)-invariant Riemannian metrics which are not necessarily right-O(n)-invariant. A result in this direction would have immediate impact on the modeling of finite strain anisotropic elasticity [14,188,189]. The difficulties with such an extension are twofold: one needs a representation formula for Riemannian metrics which are right-invariant under a given symmetry subgroup of O(n), as well as an understanding of the corresponding geodesic curves. 54 An improved understanding of the geometric structure of mechanical problems could, for example, help to develop new discretization methods [85,185].  18. If SO(n) contains a length minimizing geodesic connecting Q, R ∈ SO(n) with respect to our left-GL(n)-invariant, right-O(n)-invariant metric g on GL(n), then the GL + (n)-geodesic distance between Q and R is equal to the well-known SO(n)-geodesic distance μ c log(Q T R) 2

Conclusion
We have shown that the squared geodesic distance of the (finite) deformation gradient F ∈ GL + (n) to the special orthogonal group SO(n) is the quadratic isotropic Hencky strain energy: if the general linear group is endowed with the left- with X, Y = tr(X T Y ). Furthermore, the (partial) logarithmic strain measures ω iso = dev n log U = dev n log F T F and ω vol = |tr(log U )| = |tr(log F T F)| have been characterized as the geodesic distance of F to the special orthogonal group SO(n) and the identity tensor 1, respectively: where the geodesic distances on SL(n) and R + · 1 are induced by the canonical left We thereby show that the two quantities ω iso = dev n log U and ω vol = |tr(log U )| are purely geometric properties of the deformation gradient F, similar to the invariants dev n ε and |tr(ε)| of the infinitesimal strain tensor ε in the linearized setting.
While there have been prior attempts to deductively motivate the use of logarithmic strain in nonlinear elasticity theory, these attempts have usually focussed on the logarithmic Hencky strain tensor E 0 = log U (or E 0 = log V ) and its status as the "natural" material (or spatial) strain tensor in isotropic elasticity. We discussed, for example, a well-known characterization of log V in the hypoelastic context: if the strain rate d dt is objective as well as corotational, and if d dt for some strain tensor E, then d dt = d dt log must be the logarithmic rate and E = E 0 = log V must be the spatial Hencky strain tensor. However, as discussed in Section 1.1, all strain tensors are interchangeable: the choice of a specific strain tensor in which a constitutive law is to be expressed is not a restriction on the available constitutive relations. Such an approach can therefore not be applied to deduce necessary conditions or a priori properties of constitutive laws.
Our deductive approach, on the other hand, directly motivates the use of the strain measures ω iso and ω vol from purely differential geometric observations. As we have indicated, the requirement that a constitutive law depends only on ω iso and ω vol has direct implications; for example, the tension-compression symmetry W (F) = W (F −1 ) is satisfied by every hyperelastic potential W which can be expressed in terms of ω iso and ω vol alone.
Moreover, as demonstrated in Section 4, similar approaches oftentimes presuppose the role of the positive definite factor U = √ F T F as the sole measure of the deformation, whereas this independence from the orthogonal polar factor is obtained deductively in our approach (cf. Table 1).
Note also that the specific distance measure dist geod on GL + (n) used here is not chosen arbitrarily: the requirements of left-GL(n)-invariance and right-O(n)invariance, which have been motivated by mechanical considerations, uniquely determine g up to the three parameters μ, μ c , κ > 0. This uniqueness property further emphasizes the generality of our results, which yet again strongly suggest that Hencky's constitutive law should be considered the idealized nonlinear model of elasticity for very small strains outside the infinitesimal range.
Acknowledgmets. The second author acknowledges support by the Deutsche Forschungsgemeinschaft (DFG) through a Heisenberg fellowship under grant EI 453/2-1. We are grateful to Prof. Alexander Mielke (Weierstraß-Institut, Berlin) for pertinent discussions on geodesics in GL(n); the first parametrization of geodesic curves on SL(n) known to us is due to him [134]. We also thank Prof. Robert Bryant (Duke University) for his helpful remarks regarding geodesics on Lie groups and invariances of inner products on gl(n), as well as a number of friends who helped us with the draft. We also thank Dr. Andreas Fischle (Technische Universität Dresden) who, during long discussions on continuum mechanics and differential geometry, inspired many of the ideas laid out in this paper. The first author had the great honor of presenting the main ideas of this paper to Richard Toupin on the occasion of the Canadian Conference on Nonlinear Solid Mechanics 2013 in the mini-symposium organized by Francesco dell'Isola and David J. Steigmann, which was dedicated to Toupin.

A.1: Notation
• R is the set of real numbers, • R + = (0, ∞) is the set of positive real numbers, • R n is the set of real column vectors of length n, • R n×m is the set of real n × m-matrices, • 1 is the identity tensor; • SO(n) is the special orthogonal group of all Q ∈ O(n) with det Q = 1, • Sym(n) is the set of symmetric, real n × n-matrices, that is S T = S for all S ∈ Sym(n), • Sym + (n) is the set of positive definite, symmetric, real n × n-matrices, that is x T Px > 0 for all P ∈ Sym + (n), 0 = x ∈ R n , • gl(n) = R n×n is the Lie algebra of all real n × n-matrices, • so(n) = {W ∈ R n×n | W T = −W } is the Lie algebra of skew symmetric, real n × n-matrices, • sl(n) = {X ∈ R n×n | tr(X ) = 0} is the Lie algebra of trace free, real n × nmatrices, that is tr(X ) = 0 for all X ∈ sl(n), • ⊂ R n is the reference configuration of an elastic body, • ∇ϕ = Dϕ is the first derivative of a differentiable function ϕ : ⊂ R n → R n , often called the deformation gradient, • curl v denotes the curl of a vector valued function v : R 3 → R 3 , • Curl p denotes the curl of a matrix valued function p : R 3 → R 3×3 , taken row-wise, • ϕ : → R n is a continuously differentiable deformation with ∇ϕ(x) ∈ GL + (n) for all x ∈ , • F = ∇ϕ ∈ GL + (n) is the deformation gradient, • U = √ F T F ∈ Sym + (n) is the right Biot-stretch tensor, • V = √ F F T ∈ Sym + (n) is the left Biot-stretch tensor, • B = F F T = V 2 is the Finger tensor, • C = F T F = U 2 is the right Cauchy-Green deformation tensor, • F = RU = V R is the polar decomposition of F with R = polar(F) ∈ SO(n), • E 0 = log U is the material Hencky strain tensor, • E 0 = log V is the spatial Hencky strain tensor, • S 1 = D F W (F) is the first Piola-Kirchhoff stress corresponding to an elastic energy W = W (F), • S 2 = F −1 S 1 = 2 D C W (C) is the second Piola-Kirchhoff stress corresponding to an elastic energy W = W (C) (Doyle-Ericksen formula), [125, p. 116] is the Kirchhoff stress tensor, • σ = 1 det F τ is the Cauchy stress tensor, • T Biot = U S 2 = D U W (U ) is the Biot stress tensor corresponding to an elastic energy W = W (U ), • L =Ḟ F −1 is the spatial velocity gradient, • D = sym L is the rate of stretching or spatial strain rate tensor, • W = skew L is the spatial continuum spin.

A.2: Linear stress-strain relations in nonlinear elasticity
Many constitutive laws commonly used in applications are expressed in terms of linear relations between certain strains and stresses, including Hill's family of generalized linear elasticity laws (cf. Section 4.2.1) of the form T r = 2 μ E r + λ tr(E r ) · 1 (71) with work-conjugate pairs (T r , E r ) based on the Lagrangian strain measures given in (3). A widely known example of such a constitutive law is the hyperelastic Saint-Venant-Kirchhoff model S 2 = 2 μ E 1 + λ tr(E 1 ) 1 = μ (C − 1) + λ 2 tr(C − 1) · 1 for r = 1 and T 1 = S 2 , where S 2 denotes the second Piola-Kirchhoff stress tensor. Similarly, a number of elasticity laws can be written in the form with a spatial strain tensor E r and a corresponding stress tensor T r . Examples include the Neo-Hooke type model for r = 0 and T 0 = τ . A thorough comparison of these four constitutive laws can be found in [16]. Another example of a postulated linear stress-strain relation is the model where T Biot denotes the Biot stress tensor, which measures the "stress per unit initial area before deformation" [28]. This constitutive relation was first given in an 1893 article by the geologist G.F. Becker [18,153], who deduced it from a law of superposition in an approach similar to that of Hencky. The same constitutive law was considered by Carroll [38] as an example to emphasize the necessity of a hyperelastic formulation in order to ensure physical plausibility in the description of elastic behavior. Note that of the constitutive relations listed in this section, only the Hencky model and the Saint-Venant-Kirchhoff model are indeed hyperelastic (cf. [23, Chapter 7.4]).

A.3: Tensors and tangent spaces
In the more general setting of differential geometry, the linear mappings F, U, C, V, B and R as well as various stresses at a single point x in an elastic body are defined as mappings between different tangent spaces: for a point x ∈ and a deformation ϕ, we must then distinguish between the two tangent spaces T x and T ϕ(x) ϕ( ). The domains and codomains of various linear mappings are listed below and indicated in Fig. 19. Note that we do not distinguish between tangent and cotangent vector spaces (cf. [63]).

F, R
: The right Cauchy-Green tensor C = F T F, in particular, is often interpreted as a Riemannian metric on ; Epstein [61, p. 113] explains that "the right Cauchy-Green tensor is precisely the pull-back of the spatial metric to the body manifold", cf. [127]. If and ϕ( ) are embedded in the Euclidean space R n , this connection can immediately be seen: while the length of a curve x : [0, 1] → is given by 1 0 √ ẋ,ẋ dt, where ·, · is the canonical inner product on R n , the length of the deformed curve ϕ • x is given by (cf. Fig. 19) The quadratic form g x (v, v) = C(x) v, v at x ∈ therefore measures the length of the deformed line element Fv at ϕ(x) ∈ ϕ( ). Thus locally, dist Euclid,ϕ( ) (ϕ(x), ϕ(y)) = dist geod, (x, y), where dist Euclid,ϕ( ) (ϕ(x), ϕ(y)) = ϕ(x) − ϕ(y) is the Euclidean distance between ϕ(x), ϕ(y) ∈ ϕ( ) and dist geod, (x, y) denotes the geodesic distance between x, y ∈ with respect to the Riemannian metric g x (v, w) = C(x) v, w . Moreover, this interpretation characterizes the Green-Lagrangian strain tensor E 1 = 1 2 (C − 1) as a measure of change in length: the difference between the squared length of a line element v ∈ T x in the reference configuration and the squared length of the deformed line element F(x) v ∈ T ϕ(x) ϕ( ) is given by where . denotes the Euclidean norm on R n . Note that for F(x) = 1 + ∇u(x) with the displacement gradient ∇u(x), the expression F(x) v 2 can be linearized to where h.o.t. denotes higher order terms with respect to ∇u(x). Thus where ε = sym ∇u is the linear strain tensor.

A.4: Additional computations
Let Cof F = (det F) · F −T denote the cofactor of F ∈ GL + (n). Then the geodesic distance of Cof F to SO(n) with respect to the Riemannian metric g introduced in (19) can be computed directly by applying Theorem 3.3: A. 5 The principal matrix logarithm on Sym + (n) and the matrix exponential The following lemma states some basic computational rules for the matrix exponential exp : R n×n → GL + (n) and the principal matrix logarithm log : Sym + (n) → Sym(n) involving the trace operator tr and the deviatoric part dev n X = X − tr(X ) n ·1 of a matrix X ∈ R n×n . Lemma A.1. Let X ∈ R n×n , P ∈ Sym + (n) and c > 0. Then (i) det(exp(X )) = e tr(X ) , (ii) exp(dev n X ) = e − tr(X ) n · exp(X ), (iii) log(c · 1) = ln(c) · 1, (iv) log((det P) −1/n · P) = log P − ln(det P) n · 1 = dev n log P.
according to (ii). Then the injectivity of the matrix exponential on Sym(n) shows (iv).
Hencky received his diploma in civil engineering from TH München in 1908 and his Ph.D from TH Darmstadt in 1913. The title of his thesis was "Über den Spannungszustand in rechteckigen, ebenen Platten bei gleichmäßig verteilter und bei konzentrierter Belastung" ("On the stress state in rectangular flat plates under uniformly distributed and concentrated loading"). In 1915, the main results of his thesis were also published in the Zeitschrift für angewandte Mathematik und Physik [96]. After working on plasticity theory and small-deformation elasticity, he began his work on finite elastic deformations in 1928. In 1929 he introduced the logarithmic strain e log = log final length original length in a tensorial setting [99] and applied it to the description of the elastic behavior of vulcanized rubber [103]. Today, Hencky is mostly known for his contributions to plasticity theory: the article "Über einige statisch bestimmte Fälle des Gleichgewichts in plastischen Körpern" [98] ("On statically determined cases of equilibrium in plastic bodies"), published in 1923, is considered his most famous work [197].