Left-invariant G 2 ∗ {\rm G}_{2}^{*} -structures of type III Viviana del Barco, Ana Cristina Ferreira, Ines Kath Abstract We investigate left-invariant G 2 ∗ {\rm G}_{2}^{*} -structures on 7-dimensional Lie groups, focusing on those whose holonomy algebras are indecomposable and of type III, the latter meaning that the socle of the holonomy representation is maximal. Building on the classification of indecomposable holonomy algebras contained in 𝔤 2 ∗ {{\mathfrak{g}}^{*}_{2}} , we determine which ones arise as infinitesimal holonomy algebras of type III for left-invariant G 2 ∗ {\rm G}_{2}^{*} -structures. Our main result shows that only abelian subalgebras occur, and these are necessarily of dimension two or three. Moreover, we provide explicit Lie groups with left-invariant G 2 ∗ {\rm G}_{2}^{*} -structures realizing these abelian holonomies. Mathematics Subject Classification (2020): Primary 53C29; Secondary 53C50, 53C10, 53C30. 1 Introduction This article will deal with holonomy groups contained in the non-compact group of type {\rm G}_{2} , which is denoted by {\rm G}_{2}^{*} . Such holonomy groups appear on pseudo-Riemannian manifolds of signature (4,3) . In particular, the full group {\rm G}_{2}^{*} is a holonomy group, as it can indeed be realized as a holonomy of compact nilmanifolds [ 3 ] . Besides {\rm G}_{2}^{*} itself, also proper subgroups occur as holonomy groups. We will be mainly interested in these subgroups, where we want to consider only connected ones which, moreover, are ‘genuine’ in the sense that the holonomy representation is indecomposable. Instead of studying the holonomy group itself, we consider its Lie algebra, which we call holonomy algebra. For a Lie subalgebra of \mathop{{\mathfrak{s}\mathfrak{o}}}(p,q) , a necessary condition for it to be a holonomy is that it is a Berger algebra, i.e. that it satisfies Berger’s first criterion. In [ 1 ] , the indecomposable Berger algebras properly contained in the Lie algebra {{\mathfrak{g}}^{*}_{2}} of {\rm G}_{2}^{*} were classified. They were distinguished by the dimension of the socle of their natural representation on the tangent space, this dimension is 1, 2 or 3 (see Section 2 for more information). Accordingly, we have three families of indecomposable Berger algebras, namely those of type I, II or III. For each type, [ 1 ] provides a complete list of all Berger algebras of that type. All these Berger algebras can indeed be realized as holonomy algebras of local metrics [ 2 , 8 , 9 ] . For some of these holonomy algebras a global metric has also been found. Here we mean by ‘global’ that there is a ‘nice’ manifold with such a metric and not just a coordinate patch. In [ 6 ] , indecomposable indefinite symmetric spaces with {\rm G}_{2}^{*} -structure are classified. Their holonomy algebras are abelian and of dimension two or three. In [ 3 ] , Fino and Lujan studied torsion-free {\rm G}_{2}^{*} -structures on nilpotent Lie groups. In addition to the already mentioned example whose holonomy is equal to {{\mathfrak{g}}^{*}_{2}} , they found an example of a torsion-free {\rm G}_{2}^{*} -structure with 6 -dimensional indecomposable holonomy. In [ 1 ] , new left-invariant metrics with indecomposable holonomy in {{\mathfrak{g}}^{*}_{2}} were found, where, for type III, only the 3-dimensional abelian Berger algebra was realized. In addition to these examples, it is known that torsion-free left-invariant {\rm G}_{2}^{*} -structures whose associated metric is bi-invariant on Lie groups with a cocompact lattice have a trivial holonomy [ 4 ] . However, the question of which holonomy algebras contained in {{\mathfrak{g}}^{*}_{2}} are realizable by a left-invariant metric has not been studied systematically. Here we will attack this question for holonomy algebras of type III, i.e. for those whose socle is maximal. We solve it completely under the additional assumption that not only the metric but also the parallel {\rm G}_{2}^{*} -structure is left-invariant. More precisely, we determine which of the holonomy algebras of type III listed in [ 1 ] are the Lie algebra of the holonomy group of a Lie group endowed with a left-invariant torsion-free {\rm G}_{2}^{*} -structure and its associated metric. We can state our main result as follows. Theorem 1.1 . Let {\mathfrak{h}} be a subalgebra of {{\mathfrak{g}}^{*}_{2}} whose natural representation on {\mathbb{R}}^{4,3} is indecomposable and of type III . Then {\mathfrak{h}} is the infinitesimal holonomy of a left-invariant {\rm G}_{2}^{*} -structure if and only if {\mathfrak{h}} is abelian. If {\mathfrak{h}} is abelian, then it has dimension two or three. Our proof of the “if” direction is constructive. It consists of giving a one-parameter family of Lie groups such that all of them carry a parallel left-invariant {\rm G}_{2}^{*} -structure whose holonomy is 3-dimensional abelian for non-zero values of the parameter, and 2-dimensional abelian otherwise. It is remarkable that their left-invariant metrics can also be induced by non left-invariant parallel {\rm G}_{2}^{*} -structures, as we show in Section 7 . To make a better description of the main result of the paper, let us recall some properties of {\rm G}_{2}^{*} and introduce some notation. Recall that the complex Lie group {\rm G}_{2}^{\mathbb{C}} contains two conjugacy classes of 9-dimensional maximal parabolic subgroups. These correspond to two conjugacy classes P_{1} , P_{2} of real subgroups in {\rm G}_{2}^{*} . These conjugacy classes can be characterized in the following way. The group {\rm G}_{2}^{*} acts transitively on the set of isotropic lines in {\mathbb{R}}^{4,3} . Groups in P_{1} appear as the stabilizer of an isotropic line under this action. Similarly, {\rm G}_{2}^{*} acts transitively on the set of isotropic 2-planes with vanishing cross product and groups in P_{2} can be understood as the stabilizer of such a 2-plane (see [ 1 ] for more explanations). Let us denote by {\mathfrak{p}}_{1} and {\mathfrak{p}}_{2} the conjugacy classes of Lie algebras corresponding to P_{1} and P_{2} . Lie algebras in {\mathfrak{p}}_{1} are isomorphic to \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes{\mathfrak{m}} , where {\mathfrak{m}} is a 3-step nilpotent Lie algebra, while Lie algebras belonging to {\mathfrak{p}}_{2} are isomorphic to \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes{{\mathfrak{n}}} for a 5-dimensional Heisenberg algebra {\mathfrak{n}} . Holonomy algebras of type III are subalgebras of Lie algebras in {\mathfrak{p}}_{1} . More exactly, they are contained in a subalgebra isomorphic to \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes[{\mathfrak{m}},{\mathfrak{m}}] . In Section 2 we will give an explicit embedding of {{\mathfrak{g}}^{*}_{2}} into \mathfrak{so}(4,3) realizing {\rm G}_{2}^{*} inside \mathrm{SO}(4,3) . Then a representative of P_{1} is given by the stabilizer of the real line spanned by the first basis vector. Its Lie algebra is equal to the Lie algebra {\mathfrak{h}}^{\rm I}\cong\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes{\mathfrak{m}} defined in ( 4 ) and the subalgebra \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes[{\mathfrak{m}},{\mathfrak{m}}] is equal to {\mathfrak{h}}^{\rm III} defined in ( 8 ). In ( 10 ), we define two abelian subalgebras {\mathfrak{m}}(1,0,1)\subset{\mathfrak{m}}(1,0,2)=[{\mathfrak{m}},{\mathfrak{m}}] of {\mathfrak{h}}^{\rm III} . The abelian subalgebras appearing in Theorem 1.1 are conjugate to these two subalgebras. The paper is organized as follows. In Section 2 we fix the standard embedding of {\rm G}_{2}^{*} into \mathop{{\rm SO}}(4,3) and of its Lie algebra that will be used all along the paper. We introduce notation and recall the list of possible indecomposable holonomy algebras of type III obtained in [ 1 ] ; all of them are of the form {\mathfrak{h}}={\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k) with {\mathfrak{a}}\subset\mathop{{\mathfrak{g}\mathfrak{l}}}(2,\mathbb{R}) and {\mathfrak{m}}(1,0,k) abelian (see Theorem 2.4 ). Section 3 specializes in left-invariant {\rm G}_{2}^{*} -structures on Lie groups. We introduce geometric tensors for later use and show a restriction for the Levi-Civita connection on left-invariant vector fields, when the holonomy of the pseudo-Riemannian manifold is assumed to be contained {\mathfrak{h}}^{\rm III} . Section 4 summarizes technical results that will be used in the sequel; these are general results for pseudo-Riemannian Lie groups endowed with a left-invariant metric of signature (4,3) , whose Levi-Civita connection restricts to a particular subalgebra of {\mathfrak{g}}_{2}^{*} . Section 5 shows that if {\mathfrak{h}} is the holonomy algebra of a Lie group endowed with a left-invariant {\rm G}_{2}^{*} -structure, then its {\mathfrak{a}} part cannot be a non-trivial subalgebra of the Lie algebra of diagonal matrices. This is a first step towards the proof of Theorem 1.1 . Indeed, Section 6 contains the second and final step of the “only if” part, that shows that actually {\mathfrak{a}} has to be zero dimensional and thus {\mathfrak{h}} is abelian. Finally, Section 7 is devoted to constructing explicit examples of Lie groups with left-invariant {\rm G}_{2}^{*} -structures whose holonomies are the remaining two abelian Lie algebras that are mentioned in our main theorem above. The detailed proof of Theorem 1.1 is also part of this section. Acknowledgments. We would like to thank the Banff International Research Station (Banff, Canada) for its support during the Women in Geometry 3 Workshop (23w5062), where the work on this paper was first developed. The research of ACF was supported by Portuguese Funds through FCT (Fundação para a Ciência e a Tecnologia, I.P.) within the Projects UIDB/00013/2020, UIDP/00013/ 2020 and UID/00013: Centro de Matemática da Universidade do Minho (CMAT/UM). VdB is partially supported by FAPESP grant 2023/15089-9. VdB and ACF would like to thank the Department of Mathematics of Philipps Universität Marburg and VdB the Centro de Matemática da Universidade do Minho for their hospitality during the visits on which this work was advanced. IK’s stay at the Banff International Research Station was supported by the Programme: Kongressreisen 23, funded by Deutscher Akademischer Austauschdienst (DAAD). 2 Holonomy algebras of type III The action of \mathop{{\rm GL}}(7,{\mathbb{R}}) on \bigwedge^{3}{\mathbb{R}}^{7} has two open orbits. Elements of these orbits are called generic. Their stabilizers are isomorphic to {\rm G}_{2} for one orbit and to {\rm G}_{2}^{*} for the other one. A generic three-form \omega induces an orientation and a non-degenerate symmetric inner product {\langle\cdot\,,\cdot\rangle} on {\mathbb{R}}^{7} . If the stabilizer of \omega equals {\rm G}_{2}^{*} , then this inner product has signature (4,3) . In particular, {\rm G}_{2}^{*} is contained in \mathop{{\rm SO}}({\mathbb{R}}^{7},{\langle\cdot\,,\cdot\rangle})\cong\mathop{{\rm SO}}(4,3) . We want to fix a standard embedding of {\rm G}_{2}^{*} into \mathop{{\rm SO}}(4,3) and describe its Lie algebra. Let e_{1},\dots,e_{7} be the standard basis of {\mathbb{R}}^{7} . The 3-form \omega_{0}=\sqrt{2}(e^{167}+e^{235})-e^{4}\wedge(e^{15}-e^{26}-e^{37})\, (1) is generic and will serve as the standard 3-form here. It induces the metric \displaystyle{\langle\cdot\,,\cdot\rangle} \displaystyle= \displaystyle 2(e^{1}\cdot e^{5}+e^{2}\cdot e^{6}+e^{3}\cdot e^{7})-(e^{4})^{2}. (2) The stabilizer of \omega_{0} in \mathop{{\rm GL}}(7,{\mathbb{R}}) is the standard embedding of {\rm G}_{2}^{*} that we will use in this article. Its Lie algebra {{\mathfrak{g}}^{*}_{2}}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}(4,3) consists of all elements of the form \begin{pmatrix}s_{1}+s_{4}&-s_{10}&s_{9}&\sqrt{2}s_{6}&0&-s_{11}&-s_{12}\\ -s_{8}&s_{1}&s_{2}&\sqrt{2}s_{9}&s_{11}&0&s_{6}\\ s_{7}&s_{3}&s_{4}&\sqrt{2}s_{10}&s_{12}&-s_{6}&0\\ \sqrt{2}s_{5}&\sqrt{2}s_{7}&\sqrt{2}s_{8}&0&\sqrt{2}s_{6}&\sqrt{2}s_{9}&\sqrt{2}s_{10}\\ 0&s_{13}&s_{14}&\sqrt{2}s_{5}&-s_{1}-s_{4}&s_{8}&-s_{7}\\ -s_{13}&0&-s_{5}&\sqrt{2}s_{7}&s_{10}&-s_{1}&-s_{3}\\ -s_{14}&s_{5}&0&\sqrt{2}s_{8}&-s_{9}&-s_{2}&-s_{4}\end{pmatrix} (3) for s_{1},\dots,s_{14}\in{\mathbb{R}} . Let us introduce some general notions before we study subalgebras of {{\mathfrak{g}}^{*}_{2}} . Let {\mathfrak{h}} be a subalgebra of \mathop{{\mathfrak{s}\mathfrak{o}}}(p,q) on the pseudo-Euclidean space {\mathbb{R}}^{p,q} . Its natural representation on {\mathbb{R}}^{p,q} is called indecomposable if it does not leave invariant any proper non-zero subspace which is non-degenerate with respect to the pseudo-Euclidean inner product. The socle of a representation is its maximal semisimple subrepresentation. Suppose that the natural representation of {\mathfrak{h}}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}(p,q) is indecomposable but not irreducible. Then its socle is totally isotropic. In this case the dimension of the socle will be called the type of {\mathfrak{h}} . Now let us assume that {\mathfrak{h}} is a subalgebra of {{\mathfrak{g}}^{*}_{2}} whose natural representation of {\mathbb{R}}^{4,3} is indecomposable but not irreducible. Since the socle is totally isotropic, its dimension is 1, 2, or 3. Accordingly, the type of {\mathfrak{h}} is denoted by I, II or III. The following subalgebra {\mathfrak{h}}^{\rm I}\subset{\mathfrak{g}}_{2}^{*} is an example of a subalgebra of type I: {\mathfrak{h}}^{\rm I}:=\{h(A,v,u,y)\mid A\in\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}),\ v\in{\mathbb{R}},\ u,y\in{\mathbb{R}}^{2}\},\ (4) where h(A,v,u,y):=\left(\begin{array}[]{ccccccc}\mathop{{\rm tr}}(A)&-u_{2}&u_{1}&\sqrt{2}v&0&-y_{1}&-y_{2}\\ 0&a_{1}&a_{2}&\sqrt{2}u_{1}&y_{1}&0&v\\ 0&a_{3}&a_{4}&\sqrt{2}u_{2}&y_{2}&-v&0\\ 0&0&0&0&\sqrt{2}v&\sqrt{2}u_{1}&\sqrt{2}u_{2}\\ 0&0&0&0&-\mathop{{\rm tr}}(A)&0&0\\ 0&0&0&0&u_{2}&-a_{1}&-a_{3}\\ 0&0&0&0&-u_{1}&-a_{2}&-a_{4}\end{array}\right), (5) for A=\left(\begin{array}[]{cc}a_{1}&a_{2}\\ a_{3}&a_{4}\end{array}\right) , y=(y_{1},y_{2})^{\top} , u=(u_{1},u_{2})^{\top} . In the following y and u will always be understood as vertical vectors, but we will drop the \top sign. The Lie algebra {\mathfrak{h}}^{\rm I} is isomorphic to \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes_{\rho}{\mathfrak{m}} , where {\mathfrak{m}} is equal to the 3-step nilpotent Lie algebra {\mathfrak{m}}=\{h(0,v,u,y)\mid v\in{\mathbb{R}},\ u,y\in{\mathbb{R}}^{2}\} and the representation of \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}) on {\mathfrak{m}} is given by \rho(A)(v,u,y)=(\mathop{{\rm tr}}(A)v,Au,(A+\mathop{{\rm tr}}(A)I_{2})y). (6) For a subalgebra {\mathfrak{h}}\subset{\mathfrak{h}}^{\rm I} , we define {\mathfrak{a}}:={\mathfrak{a}}({\mathfrak{h}}):=\{A\in\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\mid h(A,v,u,y)\in{\mathfrak{h}}\mbox{ for some }v\in{\mathbb{R}},\,u,y\in{\mathbb{R}}^{2}\} (7) and identify {\mathfrak{a}} with the subalgebra \{h(A,0,0,0)\mid A\in{\mathfrak{a}}\} of {\mathfrak{h}}^{\rm I} . We thus have defined the projections {\rm pr}_{A}:{\mathfrak{h}}\to{\mathfrak{a}} , {\rm pr}_{v}:{\mathfrak{h}}\to{\mathbb{R}} and {\rm pr}_{u},{\rm pr}_{y}:{\mathfrak{h}}\to{\mathbb{R}}^{2} , such that for h=h(A^{\prime},v^{\prime},u^{\prime},y^{\prime})\in{\mathfrak{h}}^{\rm I} , A^{\prime}={\rm pr}_{A}(h) , v^{\prime}={\rm pr}_{v}(h) , u^{\prime}={\rm pr}_{u}(h) and y^{\prime}={\rm pr}_{y}(h) . The Lie algebra {\mathfrak{h}}^{\rm I} is a representative of the conjugacy class {\mathfrak{p}}_{1} of maximal parabolic subalgebras mentioned in the introduction. Since {\rm G}_{2}^{*} acts transitively on the set of isotropic lines, any subalgebra {\mathfrak{h}}\subset{{\mathfrak{g}}^{*}_{2}} of type I is conjugate to a subalgebra of {\mathfrak{h}}^{\rm I} . We further define the following subalgebra of type III: {\mathfrak{h}}^{\rm III}:=\{h(A,v,y):=h(A,v,0,y)\mid A\in\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}),\ v\in{\mathbb{R}},\ y\in{\mathbb{R}}^{2}\}\subset{\mathfrak{h}}^{\rm I}. (8) We have that \rho preserves the abelian subalgebra [{\mathfrak{m}},{\mathfrak{m}}]=\{h(0,v,y)\mid v\in{\mathbb{R}},\ y\in{\mathbb{R}}^{2}\} and {\mathfrak{h}}^{\rm III}=\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}})\ltimes_{\rho}[{\mathfrak{m}},{\mathfrak{m}}]. (9) If {\mathfrak{h}}\subset{{\mathfrak{g}}^{*}_{2}} is of type III, then it is conjugate by an element g\in{\rm G}_{2}^{*} to a subalgebra of {\mathfrak{h}}^{\rm III} . Indeed, {\mathfrak{h}} leaves invariant the isotropic three-dimensional socle. But then, by [ 1 , Lemma 2.2] , it leaves invariant also an isotropic line contained in the socle and therefore also a complementary two-dimensional subspace of the socle. For more detailed information see [ 1 , Lemma 2.10] . The converse is not always true. In fact, a subalgebra {\mathfrak{h}}\subset{\mathfrak{h}}^{\rm III} admitting an indecomposable representation on {\mathbb{R}}^{4,3} is of type III only if its subrepresentation on \mathop{{\rm span}}\{e_{2},e_{3}\} is semisimple. For later use, we adopt the following notation from [ 1 ] . For k=1,2 , let {\mathfrak{m}}(1,0,k)\subset{\mathfrak{h}}^{\rm III} be the abelian subalgebras {\mathfrak{m}}(1,0,1)=\{h(0,v,(y_{1},0))\mid v,y_{1}\in\mathbb{R}\},\ \ {\mathfrak{m}}(1,0,2)=\{h(0,v,y)\mid v\in\mathbb{R},\,y\in\mathbb{R}^{2}\}, (10) which are invariant under \rho . Lemma 2.1 . Let {\mathfrak{h}}\subset{{\mathfrak{g}}^{*}_{2}} be a subalgebra whose natural representation on {\mathbb{R}}^{4,3} is indecomposable and of type III . If {\mathfrak{h}} is abelian, then {\mathfrak{h}} is conjugate to {\mathfrak{m}}(1,0,1) or {\mathfrak{m}}(1,0,2) . Proof. Since {\mathfrak{h}} is of type III, we may assume that it is contained in {\mathfrak{h}}^{\rm III} . Assume, {\mathfrak{h}} would contain an element h(A_{0},v_{0},y_{0}) such that \mathop{{\rm tr}}(A_{0})\not=0 . If we conjugate {\mathfrak{h}} by \exp\big(h(0,v_{0}/\mathop{{\rm tr}}(A_{0}),0)\big) , we obtain a subalgebra {\mathfrak{h}}^{\prime} \subset{\mathfrak{h}}^{\rm III} , which contains h_{0}:=h(A_{0},0,y_{0}) . Since the natural representation of {\mathfrak{h}}^{\prime} is also indecomposable, {\mathfrak{h}}^{\prime} contains an element of the form h_{1}:=h(A,1,y) . But {\rm pr}_{v}([h_{0},h_{1}])=\mathop{{\rm tr}}(A_{0}) , which is impossible since {\mathfrak{h}} is abelian. Consequently, {\mathfrak{a}}={\mathfrak{a}}({\mathfrak{h}}^{\prime}) is an abelian subalgebra of \mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) . In particular, \dim{\mathfrak{a}}\leq 1 . Assume that \dim{\mathfrak{a}}=1 . Then there exists an element h(A_{0},v_{0},y_{0}) , A_{0}\not=0 . Since {\mathfrak{h}} is of type III, the adjoint action of {\mathfrak{a}} is semisimple. Since {\mathfrak{a}}=\mathop{{\rm span}}\{A_{0}\} , we see that A_{0} is a semisimple matrix. In particular, A_{0} is invertible because of \mathop{{\rm tr}}(A_{0})=0 . Conjugating {\mathfrak{h}} by \exp\big(h(0,0,A_{0}^{-1}y_{0})\big) we obtain a subalgebra {\mathfrak{h}}^{\prime}\subset{\mathfrak{h}}^{\rm III} , which contains h_{0}:=h(A_{0},v_{0},0) . For each element h_{1}=h(0,v,y)\in{\mathfrak{h}}^{\prime} , we have 0={\rm pr}_{y}([h_{0},h_{1}])=A_{0}(y) . Since A_{0} is invertible, this implies {\rm pr}_{y}({\mathfrak{h}}^{\prime})=0 . Hence the non-degenerate subspace \mathop{{\rm span}}\{e_{1},e_{4},e_{5}\} is invariant under {\mathfrak{h}}^{\prime} , which contradicts the indecomposability of the natural representation of {\mathfrak{h}}^{\prime} . We obtain {\mathfrak{a}}=0 . If {\mathfrak{h}}^{\prime}={\mathfrak{m}}(1,0,2) , we are done. If not, then {\mathfrak{h}}^{\prime}=\mathop{{\rm span}}\{h(0,1,y_{0}),h(0,0,y)\} , y\not=0 , by indecomposability. Conjugating first by \exp(h) for h=h(0,0,-y_{0}/3,0)\in{\mathfrak{h}}^{\rm I} and after that by a suitable element \exp\big(h(A,0,0)\big) , we obtain {\mathfrak{m}}(1,0,1) . ∎ Let M be a 7-dimensional manifold. A {\rm G}_{2}^{*} -structure on M is a reduction P_{{\rm G}_{2}^{*}}(M) of the frame bundle P(M) to {\rm G}_{2}^{*} . Equivalently, a {\rm G}_{2}^{*} -structure is a 3-form \omega on M which is generic at each point of M and whose stabilizer is isomorphic to {\rm G}_{2}^{*} . The correspondence is given by P_{{\rm G}_{2}^{*}}(M)=\{(e_{1},\dots,e_{7})\in P(M)\mid\omega=\omega_{0}\mbox{ w.r.t. }e_{1},\dots,e_{7}\}. The bundle P_{{\rm G}_{2}^{*}}(M) defines a pseudo-Riemannian metric g of signature (4,3) on M . This metric is given by ( 2 ) with respect to every frame (e_{1},\dots,e_{7})\in P_{{\rm G}_{2}^{*}}(M) . The {\rm G}_{2}^{*} -structure is called parallel if the Levi-Civita connection \nabla^{g} of g reduces to P_{{\rm G}_{2}^{*}}(M) or, equivalently, if \nabla^{g}\omega=0 . We now assume that we are given a parallel {\rm G}_{2}^{*} -structure P_{{\rm G}_{2}^{*}}(M) on M together with the reduced Levi-Civita connection. We fix a point x\in M . Let H_{x} denote the holonomy group of P_{{\rm G}_{2}^{*}}(M) at x . Let \mathfrak{e}=(e_{1},\dots,e_{7}) be a frame at x which belongs to P_{{\rm G}_{2}^{*}}(M) . Using this frame, we can identify H_{x} with a subgroup of {\rm G}_{2}^{*} . The conjugacy class of this subgroup in {\rm G}_{2}^{*} depends neither on x nor on the chosen basis \mathfrak{e} . So each element of this congugacy class is called a holonomy group of the given {\rm G}_{2}^{*} -structure and its Lie algebra is called a holonomy algebra. The natural representation of a holonomy algebra {\mathfrak{h}}\subset{{\mathfrak{g}}^{*}_{2}} on {\mathbb{R}}^{4,3} is called holonomy representation. A holonomy algebra is called indecomposable if its holonomy representation is indecomposable. Definition 2.2 (Type of the holonomy) . An indecomposable holonomy {\mathfrak{h}}\subset{\mathfrak{g}}_{2}^{*} is of type I , II , or III , if the socle of the holonomy representation has dimension 1 , 2 or 3 , respectively. Of course, the type of a holonomy algebra only depends on its conjugacy class. In this paper we are interested in holonomy algebras {\mathfrak{h}} of type III. Let {\mathfrak{h}} be such a holonomy algebra. As we have seen above, there exists an element g\in{\rm G}_{2}^{*} such that \mathop{{\rm Ad}}(g)({\mathfrak{h}})\subset{\mathfrak{h}}^{\rm III} . So we may assume that {\mathfrak{h}} is contained in {\mathfrak{h}}^{\rm III} . An important tool in the study of holonomy groups is Berger’s first criterion, which we want to recall here. For a subalgebra {\mathfrak{h}}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}(p,q) , {\mathcal{K}}({\mathfrak{h}})=\{R\in{\textstyle\bigwedge^{2}}({\mathbb{R}}^{p,q})^{*}\otimes{\mathfrak{h}}\mid\forall x,y,z\in{\mathbb{R}}^{p,q}:R(x,y)z+R(y,z)x+R(z,x)y=0\} is the space of formal curvature tensors with values in {\mathfrak{h}} . If {\mathfrak{h}} is a holonomy algebra, then {\mathfrak{h}}=\mathop{{\rm span}}\{R(x,y)\mid x,y\in{\mathbb{R}}^{p,q},\ R\in{\cal K}({\mathfrak{h}})\}. For a holonomy algebra {\mathfrak{h}}\subset{\mathfrak{h}}^{\rm III} , we obtain in particular {\mathfrak{h}}\subset\mathop{{\rm span}}\{R(x,y)\mid x,y\in{\mathbb{R}}^{4,3},\ R\in{\cal K}({\mathfrak{h}}^{\rm III})\}. So, let us describe {\cal K}({\mathfrak{h}}^{\rm III}) . For an algebraic curvature tensor R\in{\cal K}({\mathfrak{h}}^{\rm III}) we put R_{ij}:=R(e_{i},e_{j}) , where as above e_{1},\dots,e_{7} is the standard basis of {\mathbb{R}}^{4,3} , and denote A_{ij}:={\rm pr}_{A}(R_{ij}) . The following result is an immediate consequence of [ 1 , Table 1] , see also [ 9 , Table 1] . Proposition 2.3 . The space {\mathcal{K}}({\mathfrak{h}}^{\rm III}) of algebraic curvature tensors can be parametrized by a_{1},a_{2},a_{3}\in{\mathbb{R}} , b_{k},j_{k}\in{\mathbb{R}} ( k=1,\dots,4 ) and c_{3},c_{4},w_{1},w_{2},t\in{\mathbb{R}} , where R=h(A,v,y)\in{\mathcal{K}}({\mathfrak{h}}^{\rm III}) is given by the data in Table 1 in Appendix A.1 . Besides the well known Lie algebra \mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) , the Lie algebras \mathfrak{c}\mathfrak{o}(2):=\left\{\left.\begin{pmatrix}a&-b\\ b&a\end{pmatrix}\ \right|\ a\in{\mathbb{R}}\right\},\quad{\mathfrak{d}}=\{\mathop{{\rm diag}}(d_{1},d_{2})\mid d_{1},d_{2}\in{\mathbb{R}}\} will appear in the theorem below. Furthermore, we will use the abelian Lie algebras {\mathfrak{m}}(1,0,k) , k=1,2 introduced in ( 10 ). Theorem 2.4 . [ 1 , 2 , 8 , 9 ] A subalgebra {\mathfrak{h}}\subset{\mathfrak{g}}_{2}^{*} is an indecomposable holonomy algebra of type III if and only if it is conjugate by an element of {\rm G}_{2}^{*} to one of the following Lie algebras: 1. {\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,2) with {\mathfrak{a}}\in\{\mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}),\,\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}),\,\mathfrak{c}\mathfrak{o}(2),{\mathfrak{d}}\} , 2. {\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k) for some k\in\{1,2\} with {\mathfrak{a}}\in\{0,\,{\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0)\} . 3 Left-invariant {\rm G}_{2}^{*} -structures and their holonomy Let G be a 7-dimensional Lie group endowed with a parallel left-invariant {\rm G}_{2}^{*} -structure \omega . Let {\langle\cdot\,,\cdot\rangle} be the induced left-invariant metric on G and let {\langle\cdot\,,\cdot\rangle} denote also the corresponding metric on the Lie algebra {\mathfrak{g}} of G . For X\in{\mathfrak{g}} , let \hat{X} be the right-invariant (i.e., the fundamental) vector field on G . We use Wang’s description of invariant connections by \Lambda(X)=(\nabla_{\hat{X}}-{\cal L}_{\hat{X}})_{e} . Then \Lambda(X)Y=(\nabla_{\hat{X}}Y)_{e} , where we identify elements of {\mathfrak{g}} with left-invariant vector fields. Since the commutator of left- and right-invariant vector fields vanishes, we obtain \Lambda(X)Y=\nabla_{X}Y,\quad\forall\,X,Y\in{\mathfrak{g}}. Thus \Lambda can be seen as linear map \Lambda:{\mathfrak{g}}\to\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) . Let R denote the curvature tensor of (G,{\langle\cdot\,,\cdot\rangle}) , which is also left-invariant and thus can be seen as a linear tensor on {\mathfrak{g}} . Explicitly, R(X,Y)=[\Lambda(X),\Lambda(Y)]-\Lambda([X,Y]) for X,Y\in{\mathfrak{g}} . Let {\mathfrak{h}}_{e}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) denote the holonomy algebra of the pseudo-Riemannian Lie group (G,{\langle\cdot\,,\cdot\rangle}) at e . We define {\mathfrak{r}}_{0}:=\mathop{{\rm span}}\{R(X,Y)\mid X,Y\in{\mathfrak{g}}\}=\mathop{{\rm span}}\{[\Lambda(X),\Lambda(Y)]-\Lambda([X,Y])\mid X,Y\in{\mathfrak{g}}\}. (11) According to [ 7 , Cor. 4.2.] , the holonomy algebra {\mathfrak{h}}_{e} is given by {\mathfrak{h}}_{e}={\mathfrak{r}}_{0}+[\Lambda({\mathfrak{g}}),{\mathfrak{r}}_{0}]+[\Lambda({\mathfrak{g}}),[\Lambda({\mathfrak{g}}),{\mathfrak{r}}_{0}]]+\dots. (12) Let \mathfrak{e}=(e_{1},\ldots,e_{7}) be a basis of its Lie algebra {\mathfrak{g}} such that \omega=\omega_{0} , where \omega_{0} is as in ( 1 ), then {\langle\cdot\,,\cdot\rangle} is as in ( 2 ). Using this basis we identify {\mathfrak{h}}_{e} with a subalgebra {\mathfrak{h}}\subset{{\mathfrak{g}}^{*}_{2}} , which is also called holonomy algebra. If it is necessary to be more precise, we will say that {\mathfrak{h}} is the holonomy with respect to \mathfrak{e} . We use the basis also to identify \Lambda(X)\in\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) , X\in{\mathfrak{g}} , with a map in \mathop{{\mathfrak{s}\mathfrak{o}}}(4,3) , which we also denote by \Lambda(X) . The resulting map \Lambda with values in \mathop{{\mathfrak{s}\mathfrak{o}}}(4,3) will also be called Levi-Civita connection with respect to the basis \mathfrak{e} . Since our {\rm G}_{2}^{*} -structure is parallel and left-invariant, we have 0=\nabla_{X}\omega=\Lambda(X)^{*}\omega if X\in{\mathfrak{g}} , which implies \Lambda({\mathfrak{g}})\subset{{\mathfrak{g}}^{*}_{2}} , i.e. \Lambda(X)\in{{\mathfrak{g}}^{*}_{2}} for all X\in{\mathfrak{g}} . The following result gives properties on the Levi-Civita connection of (G,{\langle\cdot\,,\cdot\rangle}) by assuming that the holonomy algebra is contained in the Lie algebra {\mathfrak{h}}^{\rm III} defined in ( 9 ). Proposition 3.1 . Suppose that {\mathfrak{h}} above is one of the Lie algebras listed in Theorem 2.4 , and thus {\mathfrak{h}}={\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k)\subset{\mathfrak{h}}^{\rm III} . Then: 1. If {\mathfrak{a}} is one of the Lie algebras \mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) , \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}) , {\mathfrak{c}\mathfrak{o}}(2) or {\mathfrak{d}} , then \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} . Moreover, if {\mathfrak{a}}\in\{\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}),\mathfrak{c}\mathfrak{o}(2),{\mathfrak{d}}\} , then {\rm pr}_{A}(\Lambda(X))\in{\mathfrak{a}} for all X\in{\mathfrak{g}} . 2. If {\mathfrak{a}}={\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) , then \Lambda({\mathfrak{g}})\subset\{h(A,v,u,y)\mid A\in{\mathfrak{d}},\ u=(0,u_{2})\}\subset{\mathfrak{h}}^{I} . Proof. Since G is equipped with a parallel left-invariant G_{2}^{*} -structure, we have that \Lambda({\mathfrak{g}})\subset{{\mathfrak{g}}^{*}_{2}} . Moreover, from ( 12 ), [\Lambda(X),{\mathfrak{h}}]\subset{\mathfrak{h}} . Therefore, for all X\in{\mathfrak{g}} , \Lambda(X) is in the normalizer of {\mathfrak{h}} inside {{\mathfrak{g}}^{*}_{2}} . Let N\in{{\mathfrak{g}}^{*}_{2}} be in the normalizer of {\mathfrak{h}} . Then conjugation by \exp(tN) maps {\mathfrak{h}} to {\mathfrak{h}} . In particular, \exp(tN) maps the socle to the socle. Thus the same is true for N . Since for all Lie algebras in Theorem 2.4 , the socle is spanned by e_{1} , e_{2} and e_{3} , we obtain N\in\big\{h\in{{\mathfrak{g}}^{*}_{2}}\mid h(\mathop{{\rm span}}\{e_{1},e_{2},e_{3}\})=\mathop{{\rm span}}\{e_{1},e_{2},e_{3}\}\big\}={\mathfrak{h}}^{\rm I} using ( 3 ). Hence N=h(A,v,u,y) , for some h(A,v,y,y)\in{\mathfrak{h}}^{\rm I} , thus [N,h(\bar{A},\bar{v},0,\bar{y})]=h([A,\bar{A}],*,\bar{A}u,*)\in{\mathfrak{h}}\subset{\mathfrak{h}}^{\rm III} for all h(\bar{A},\bar{v},0,\bar{y})\in{\mathfrak{h}} . This shows that A belongs to the normalizer of {\mathfrak{a}} in \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}) , which can be easily determined. Moreover we see that u must be annihilated by all elements of {\mathfrak{a}} . This proves the claim. ∎ We end this section by introducing some notation that will be used throughout the rest of the article. For each i,j=1,\ldots 7 , we denote \Lambda_{i}:=\Lambda(e_{i})\in{{\mathfrak{g}}^{*}_{2}} and R_{ij}:=R(e_{i},e_{j})=[\Lambda_{i},\Lambda_{j}]-\Lambda([e_{i},e_{j}])\in{{\mathfrak{g}}^{*}_{2}}. (13) Remark 3.2 . Let {\mathfrak{h}}\subset{\mathfrak{h}}^{\rm III} be a holonomy algebra of type III and let {\mathfrak{a}} be as in ( 7 ). Then {\mathfrak{a}} is contained in the Lie algebra generated by A_{i}:={\rm pr}_{A}(\Lambda(e_{i})) , i=1,\ldots,7 . In fact, by ( 12 ), {\mathfrak{a}} is linearly spanned by A_{ij}:={\rm pr}_{A}(R_{ij})={\rm pr}_{A}([\Lambda_{i},\Lambda_{j}]-\Lambda([e_{i},e_{j}]))=[A_{i},A_{j}]-\sum_{k=1}^{7}c_{ij}^{k}A_{k} (14) and their Lie brackets with A_{i} , where c_{ij}^{k} denote the structure coefficients. We will frequently use the second Bianchi identity B_{ijk}:=-\sum_{cycl}(\nabla_{e_{i}}R)(e_{j},e_{k})=\sum_{cycl}(R(\Lambda_{i}e_{j},e_{k})+R(e_{j},\Lambda_{i}e_{k})-[\Lambda_{i},R_{jk}])=0, (15) for all i,j,k=1,\dots,7 . We denote by B_{ijk}(m,n) the (m,n) -entry of B_{ijk}\in{{\mathfrak{g}}^{*}_{2}} . 4 Metric Lie algebras with Levi-Civita connection in {\mathfrak{h}}^{\rm III} Along this section {\mathfrak{g}} is a Lie algebra endowed with the metric ( 2 ) on a given basis \mathfrak{e}=(e_{1},\ldots,e_{7}) . Suppose that its Levi-Civita connection \Lambda with respect to \mathfrak{e} satisfies \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} , so that \Lambda_{i}=\Lambda(e_{i})=h(A_{i},v_{i},(y_{i1},y_{i2})),\ A_{i}=\begin{pmatrix}a_{i1}&a_{i2}\\ a_{i3}&a_{i4}\end{pmatrix},\ i=1,\ldots,7. (16) In particular, the left-invariant 3-form \omega_{0} defined by ( 1 ) is parallel. Using the torsion free property of the Levi-Civita connection and the form of {\langle\cdot\,,\cdot\rangle} in ( 2 ), we get: Proposition 4.1 . If \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} then the Lie bracket relations between the basis elements e_{1},\ldots,e_{7} are given in Table A.2 . Proof. Using that [x,y]=\Lambda(x)y-\Lambda(y)x , for all x,y\in{\mathfrak{g}} , and ( 8 ), one can compute the Lie brackets of basis vectors. ∎ Notice that the coefficients appearing in Table A.2 have to satisfy nontrivial relations due to the Jacobi identity. Let R denote the curvature tensor and let {\mathfrak{h}} denote the holonomy Lie algebra of ({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) (or of the associated simply connected Lie group with the induced left-invariant pseudo-Riemannian metric) with respect to the basis \mathfrak{e} . If \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} then, clearly, R\in\mathcal{K}({\mathfrak{h}}^{\rm III}) but also, for every fixed x\in{\mathfrak{g}} , \nabla_{x}R\in\mathcal{K}({\mathfrak{h}}^{\rm III}) since \nabla_{x}R has the same symmetries as R . Therefore, R and \nabla_{x}R have the form in Table 1 by Proposition 2.3 ; in addition, by ( 12 ), {\mathfrak{h}}\subset{\mathfrak{h}}^{\rm III} . We use these properties to give necessary conditions for {\mathfrak{h}} to be indecomposable. Lemma 4.2 . Suppose \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} . If R_{15}=R_{45}=0 or, equivalently, R_{15}=0 and w_{1}=w_{2}=0 , then {\mathfrak{h}} is decomposable. In particular, if \Lambda_{1}=0 and \Lambda_{4}=h(0,v_{4},0) , then {\mathfrak{h}} is decomposable. Proof. We use Table 1 : from R_{15}=0 , we get \mathop{{\rm tr}}(A_{56})=\mathop{{\rm tr}}(A_{57})=0 , and R_{45}=0 implies {\rm pr}_{v}(R_{56})={\rm pr}_{v}(R_{57})=0 . This implies that for all h\in{\mathfrak{r}}_{0} , where {\mathfrak{r}}_{0} is in ( 12 ), one has \mathop{{\rm tr}}({\rm pr}_{A}(h))=0={\rm pr}_{v}(h) . By ( 6 ) also the elements of [{\mathfrak{r}}_{0},\Lambda({\mathfrak{g}})] have this property. By induction we obtain from ( 12 ) that this is also true for all elements of {\mathfrak{h}} , thus e_{4} is invariant under all elements of {\mathfrak{h}} . Therefore {\mathbb{R}}e_{4} is a non-degenerate subspace which is invariant under {\mathfrak{h}} , thus {\mathfrak{h}} is decomposable. As for the second assertion, if \Lambda_{1}=0 , then ( 54 ) and R_{ij}=[\Lambda_{i},\Lambda_{j}]-\Lambda([e_{i},e_{j}]) imply R_{15}=0 . Furthermore, [\Lambda_{4},\Lambda_{5}]=[h(0,v_{4},0),\Lambda_{5}]=h(0,-\mathop{{\rm tr}}(A_{5})v_{4},0) by ( 6 ) and thus R_{45}=h(0,-\mathop{{\rm tr}}(A_{5})v_{4},0)-\sqrt{v}_{4}\Lambda_{4} by ( 66 ) and \Lambda_{1}=0 . That is, {\rm pr}_{y}(R_{45})=0 . By Table 1 this implies R_{45}=0 . ∎ Lemma 4.3 . If \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} and if 2(\mathop{{\rm tr}}(A_{r}))^{2}+\det(A_{r})\not=0 and 6(\mathop{{\rm tr}}(A_{r}))^{2}+\det(A_{r})\not=0 for r=1 or r=4 , then {\mathfrak{h}} is decomposable. Proof. For r\in\{1,4\} , using Table 1 for \nabla_{e_{r}}R we get \displaystyle\mathop{{\rm tr}}({\rm pr}_{A}(B_{r56})) \displaystyle=\ \mathop{{\rm tr}}({\rm pr}_{A}((-\nabla_{e_{r}}R)(e_{5},e_{6}))) \displaystyle=\ (-2a_{r1}-a_{r4})\mathop{{\rm tr}}(A_{56})-a_{r2}\mathop{{\rm tr}}(A_{57}), \displaystyle\mathop{{\rm tr}}({\rm pr}_{A}(B_{r57})) \displaystyle=\ \mathop{{\rm tr}}({\rm pr}_{A}((-\nabla_{e_{r}}R)(e_{5},e_{7}))) \displaystyle=\ (-a_{r1}-2a_{r4})\mathop{{\rm tr}}(A_{57})-a_{r3}\mathop{{\rm tr}}(A_{56}), hence (A_{r}+\mathop{{\rm tr}}(A_{r})I_{2})(\mathop{{\rm tr}}(A_{56}),\mathop{{\rm tr}}(A_{57}))=0. (17) Moreover, 0={\rm pr}_{y}(B_{r67})={\rm pr}_{y}(-(\nabla_{e_{r}}R)(e_{6},e_{7})) gives (A_{r}+2\mathop{{\rm tr}}(A_{r})I_{2})(w_{1},w_{2})=v_{r}(\mathop{{\rm tr}}(A_{56}),\mathop{{\rm tr}}(A_{57})). (18) Since \det(A_{r}+n\mathop{{\rm tr}}(A_{r})I_{2})=n(n+1)(\mathop{{\rm tr}}(A_{r}))^{2}+\det(A_{r}) for n\in{\mathbb{N}} , our assumption ensures that ( 17 ) and ( 18 ) imply w_{1}=w_{2}=\mathop{{\rm tr}}(A_{56})=\mathop{{\rm tr}}(A_{57})=0 . The assertion now follows from Lemma 4.2 . ∎ Making use of the fixed basis, we define {\mathfrak{g}}_{1}:=\mathop{{\rm span}}\{e_{1}\},\ {\mathfrak{g}}_{2}:=\mathop{{\rm span}}\{e_{2},e_{3}\},\ {\mathfrak{g}}_{j}:=\{e_{1},\dots,e_{j}\},\ j=3,\dots,7. (19) As a consequence of Proposition 4.1 and Table A.2 one can deduce: Proposition 4.4 . If \Lambda({\mathfrak{g}}) is contained in {\mathfrak{h}}^{\rm III} , then {\mathfrak{g}}_{j} are subalgebras of {\mathfrak{g}} for j=1,\dots,5 and {\mathfrak{g}}_{3} is an ideal in {\mathfrak{g}}_{4} . Moreover, [e_{1},e_{4}]\in{\mathbb{R}}e_{1} . Remark 4.5 . Assume that \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} . Table 1 shows that A_{ij}=0 for i,j=1,\dots,5 . Thus e_{j}\mapsto A_{j} , j=1,\dots,5 , defines a representation of {\mathfrak{g}}_{5} . Remark 4.6 . Assume that \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} . Then \Lambda_{i}=h(A_{i},v_{i},y_{i}) as in ( 8 ). If v_{i}=0 for all i=1,\ldots,7 , then {\mathfrak{h}} is decomposable. Indeed, if this holds then \Lambda_{i}(e_{4})=0 for all i=1,\ldots,7 . This shows that the left-invariant vector field defined by e_{4} is parallel. Hence e_{4} is an invariant of the holonomy representation. Thus {\mathbb{R}}e_{4} is a non-degenerate subrepresentation of {\mathfrak{h}} , so {\mathfrak{h}} is decomposable. In the following we will sometimes change the basis of {\mathfrak{g}} in order to simplify calculations. Then instead of \mathfrak{e}=(e_{1},\dots,e_{7}) we use a basis \mathfrak{e^{\prime}}=(e_{1},\dots,e_{7})\cdot\Phi , where \Phi is a an element of {\rm G}_{2}^{*} which preserves {\mathbb{R}}e_{1} and \mathop{{\rm span}}\{e_{2},e_{3}\} . These transformations constitute a subgroup of {\rm G}_{2}^{*} whose Lie algebra is exactly {\mathfrak{h}}^{\rm III} . In particular, \Phi{\mathfrak{h}}^{\rm III}\Phi^{-1}={\mathfrak{h}}^{\rm III} . Let \Lambda^{\prime} denote the Levi-Civita connection with respect to \mathfrak{e}^{\prime} and {\mathfrak{h}}^{\prime} the holonomy algebra with respect to \mathfrak{e}^{\prime} . Then \Lambda^{\prime}(X)=\Phi^{-1}\Lambda(X)\Phi\in{\mathfrak{h}}^{\rm III} for all X\in{\mathfrak{g}} and {\mathfrak{h}}^{\prime}=\Phi^{-1}{\mathfrak{h}}\Phi\subset{\mathfrak{h}}^{\rm III} . In particular, we will use the following transformations: \displaystyle\Phi_{T} \displaystyle:= \displaystyle\mathop{{\rm diag}}(\det T,T,1,(\det T)^{-1},(T^{\top})^{-1}),\quad T\in\mathop{{\rm GL}}(2), (20) \displaystyle\Phi_{v} \displaystyle:= \displaystyle\exp h(0,v,0),\quad v\in{\mathbb{R}}. (21) Transformations of the form \Phi_{v} allow to add an arbitrary multiple of e_{1} to e_{4} and to extend this to a map in {\rm G}_{2}^{*} which leaves e_{1},e_{2},e_{3} invariant. For the next result we shall introduce some notation. Let J denote the operator J(x,y,z):=[[x,y],z]+[[y,z],x]+[[z,x],y] for x,y,z\in{\mathfrak{g}} . Furthermore, we put J_{ijk}:=J(e_{i},e_{j},e_{k}) and define J_{ijk}^{l} by J_{ijk}=J_{ijk}^{l}e_{l} . Lemma 4.7 . If \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} and if \mathop{{\rm tr}}(A_{1})=0 or \mathop{{\rm tr}}(A_{4})=0 , then v_{1}=0 . Proof. The assertion follows from the Jacobi identities J_{145}^{4}=2v_{1}^{2}-\sqrt{2}v_{4}\mathop{{\rm tr}}(A_{1})=0 and J_{145}^{5}=(\mathop{{\rm tr}}(A_{4})-\sqrt{2}v_{1})\mathop{{\rm tr}}(A_{1})=0 . ∎ Proposition 4.8 . If \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} and {\mathfrak{g}}_{3} defined in ( 19 ) is not abelian, then {\mathfrak{h}} is decomposable or the matrices A_{i} , i=1,\ldots,7 are either all lower or all upper triangular. Proof. By Proposition 4.1 the Lie bracket relations are as in Table A.2 . Assume that {\mathfrak{g}}_{3} is not abelian. By ( 57 ), [e_{2},e_{3}]=\lambda e_{2}+\mu e_{3} , for some \lambda,\mu\in\mathbb{R} . If [e_{2},e_{3}]\neq 0 then we can apply a transformation \Phi_{T} as defined in ( 20 ), for T=\begin{pmatrix}\lambda&\mu^{\prime}\\ \mu&\lambda^{\prime}\end{pmatrix} , where \lambda^{\prime}=\lambda/(\lambda^{2}+\mu^{2}),\mu^{\prime}=-\mu/(\lambda^{2}+\mu^{2}) , in order to obtain [e_{2},e_{3}]=e_{2} . So we split the proof into two cases: [e_{2},e_{3}]=0 or [e_{2},e_{3}]=e_{2} . Case 1: [e_{2},e_{3}]=0 . Depending on whether {\mathfrak{g}}_{2}=\mathop{{\rm span}}\{e_{2},e_{3}\} is preserved by e_{1} or not, we split into cases 1.1 and 1.2: Case 1.1: [e_{1},{\mathfrak{g}}_{2}]\subset{\mathfrak{g}}_{2} . By ( 51 ), ( 52 ) and ( 57 ) we get the following identities \mathop{{\rm tr}}(A_{2})=\mathop{{\rm tr}}(A_{3})=0,\;a_{24}=a_{33},\;a_{22}=a_{31}. (22) Then {\mathfrak{g}}_{3}={\mathfrak{g}}_{1}\ltimes{\mathfrak{g}}_{2} , where {\mathfrak{g}}_{2} is abelian and e_{1} acts by an arbitrary linear map on the abelian ideal {\mathfrak{g}}_{2} . Since {\mathfrak{g}}_{3} is not abelian, the adjoint action of e_{1} on {\mathfrak{g}}_{2} is non-trivial. It is given by A_{1} with respect to the basis e_{2},e_{3} of {\mathfrak{g}}_{2} . If \det(A_{1})>0 or if \det(A_{1})=0 and \mathop{{\rm tr}}(A_{1})\not=0 , then {\mathfrak{h}} is decomposable by Lemma 4.3 . So let us assume that \det(A_{1})<0 or that \det(A_{1})=\mathop{{\rm tr}}(A_{1})=0 . Then A_{1} has two real eigenvalues that are either both equal to zero or both unequal to zero with opposite signs. We may change the basis of {\mathfrak{g}} by a transformation \Phi_{T} as in ( 20 ). Therefore we may assume that the action of e_{1} has a certain normal form. This leads to the following two cases, which correspond to the two possibilities for the eigenvalues of A_{1} . Case 1.1.1: {\mathfrak{g}}_{3}=\{[e_{1},e_{3}]=e_{2}\} (The brackets between \{\,\cdot\,\} are the only non-zero Lie bracket relations defining {\mathfrak{g}}_{3} .) In addition to ( 22 ), the form of {\mathfrak{g}}_{3} introduces the additional identities a_{13}=a_{11}=a_{14}=0,\;a_{12}=1. In particular, this gives v_{1}=0 by Lemma 4.7 . Proposition 4.4 implies that \mathop{{\rm ad}}(e_{4}) is a derivation of {\mathfrak{g}}_{3} and [e_{4},e_{1}]\in{\mathbb{R}}e_{1} , so we have \mathop{{\rm ad}}(e_{4})=\left(\begin{array}[]{ccc}-r&0&-t\\ 0&-r-s&u\\ 0&0&-s\end{array}\right). We may add a multiple of e_{1} to e_{4} (extendable to a map of the form \Phi_{v} in ( 21 )). Therefore we may assume u=0 . By ( 53 ) we get \mathop{{\rm tr}}(A_{4})=-r . On the other hand a_{41}=-r-s by ( 58 ), and a_{44}=-s by ( 62 ). We obtain s=0 . Consequently, {\mathfrak{g}}_{4}=\{[e_{1},e_{3}]=e_{2},[e_{1},e_{4}]=re_{1},[e_{2},e_{4}]=re_{2},[e_{3},e_{4}]=te_{1}\}, which implies, by ( 58 ) and ( 62 ), \displaystyle a_{42}=a_{43}=a_{44}=0,\ v_{2}=0. Furthermore, we already got v_{1}=0 . We will show that a_{j3}=0 for all j=1,\dots,7 . We already know that a_{13}=a_{43}=0 . Using Remark 4.5 we obtain [A_{1},A_{2}]=0 and [A_{1},A_{5}]=y_{12}A_{3}+y_{11}A_{2}-\mathop{{\rm tr}}(A_{5})A_{1} , which gives a_{24}=a_{33}=a_{23}=0 and a_{53}=y_{12}a_{31} . Note that J_{137}^{3}=-a_{73}=0 . Now J_{137}^{2}=2a_{74}+y_{12} together with J_{157}^{3}=-y_{11}a_{73}+y_{12}(y_{12}-a_{74})=y_{12}(y_{12}-a_{74}) implies y_{12}=a_{74}=0 . Hence a_{53}=0 and finally, J_{136}^{3}=a_{74}-a_{63}=-a_{63} gives a_{63}=0 . Therefore, A_{i} is upper triangular for all i=1,\ldots,7 . Case 1.1.2: {\mathfrak{g}}_{3}=\{[e_{1},e_{2}]=e_{2},[e_{1},e_{3}]=re_{3}\} , r<0 In this case, in addition to ( 22 ), we have A_{1}=\mathop{{\rm diag}}(1,r) . Using this and the identities [A_{1},A_{2}]=A_{2} , [A_{1},A_{3}]=rA_{3} (see Remark 4.5 ), we get A_{2}=A_{3}=0 . Moreover, [A_{1},A_{4}]=(\sqrt{2}v_{1}-\mathop{{\rm tr}}(A_{1}))A_{1} yields a_{42}=0 . From the Jacobi identity we get \displaystyle J_{136}^{2}=a_{62}(2-r)\quad J_{236}^{2}=-y_{31},\quad J_{135}^{2}=-2y_{31}+2a_{52}, which implies a_{52}=a_{62}=0 . Thus a_{j2}=0 for all j=1,\ldots,6 . Using ( 13 ) we compute the curvature term R_{173}^{2}=a_{72}-v_{1}a_{22}=a_{72} which must be zero by Table 1 . Therefore all A_{i} are lower triangular. Case 1.2: [e_{1},{\mathfrak{g}}_{2}]\not\subset{\mathfrak{g}}_{2} Let [e_{1},e_{2}]=re_{1}+se_{2}+te_{3} and [e_{1},e_{3}]=pe_{1}+qe_{2}+le_{3} . We may assume r=1 , p=0 . J_{123}=0 implies [e_{1},e_{3}]=0 , which leads to {\mathfrak{g}}_{3}=\{[e_{1},e_{2}]=e_{1}+se_{2}+te_{3}\}. Here we have a_{11}=s,a_{13}=t , a_{12}=a_{14}=0 . If s\not=0 , then {\mathfrak{h}} is decomposable by Lemma 4.3 . So let us assume that s=0 . Then the structure of {\mathfrak{g}}_{3} implies a_{11}=a_{12}=a_{14}=0,\ a_{13}=t,\ \mathop{{\rm tr}}(A_{2})=-1,\ \mathop{{\rm tr}}(A_{3})=0,\ a_{24}=a_{33},\ a_{22}=a_{31}. In particular, v_{1}=0 by Lemma 4.7 . We will first show that t=0 . Using Remark 4.5 , we have that [A_{1},A_{2}]=A_{1}+tA_{3} and [A_{2},A_{3}]=0 . This implies that ta_{31}=0 , t(3a_{24}+2)=0 and 2a_{23}a_{31}+2a_{24}^{2}+a_{24}=0. (23) Thus t=0 , otherwise we obtain a contradiction. Furthermore, since t=0 , we have J_{124}^{1}=a_{41}=0 , J_{134}^{1}=a_{42}=0 , and J_{234}^{1}=\sqrt{2}v_{3}=0 . Since we can add a multiple of e_{1} to e_{4} , we may assume that [e_{2},e_{4}] does not have an e_{1} -component, thus v_{2}=0 . Now we find J_{245}^{1}=-2\sqrt{2}v_{5},\ J_{245}^{4}=\sqrt{2}v_{4}, which gives v_{1}=\ldots=v_{5}=0 . If a_{44}\not=0 , then \det(A_{4})=0 and \mathop{{\rm tr}}(A_{4})\not=0 , hence {\mathfrak{h}} is decomposable by Lemma 4.3 . So let a_{44}=0 in the following. This implies a_{43}=0 . Indeed, from [A_{2},A_{4}]=-a_{43}A_{3} , we conclude that a_{31}a_{43}=0 and (3a_{24}+1)a_{43}=0 . If a_{43}\neq 0 , this would contradict ( 23 ). Now J_{245}=y_{41}e_{2}+y_{42}e_{3} yields y_{4j}=0 , j=1,2 . Now J_{256}^{4}+J_{347}^{1}=2\sqrt{2}v_{6} implies v_{6}=0 . Assume v_{7}\not=0 , then J_{246}=\sqrt{2}a_{31}v_{7}e_{1}=0,\ J_{247}=\sqrt{2}v_{7}(a_{24}-1)e_{1}=0 would give a_{31}=0 and a_{24}=1 , which again contradicts ( 23 ). Therefore, v_{7}=0 and from the above all v_{i} vanish, so {\mathfrak{h}} is decomposable by Remark 4.6 . Case 2: [e_{2},e_{3}]=e_{2} . We have J_{123}^{1}=\mathop{{\rm tr}}(A_{2})=0 . This implies a_{13}=-J_{123}^{3}-J_{125}^{5}=0 , i.e. [e_{1},e_{2}]=re_{2} , which gives [e_{1},e_{3}]\in\mathop{{\rm span}}\{e_{1},e_{2}\} by Jacobi identity. Thus a_{14}=0 . In particular \mathop{{\rm tr}}(A_{1})=a_{11} and \det(A_{1})=0 . If a_{11}\not=0 , Lemma 4.3 implies that {\mathfrak{h}} is decomposable. So let us assume that a_{11}=0 . From the bracket relations of {\mathfrak{g}}_{3} we conclude that \displaystyle A_{1}=\begin{pmatrix}0&q\\ 0&0\end{pmatrix},\ a_{21}=-a_{33},\ a_{22}=1+a_{31},\ a_{24}=a_{33}. The above identities imply J_{234}^{3}=-a_{43}=0 and v_{1}=0 by Lemma 4.7 . Case 2.1.: q\neq 0 Inspecting the Jacobi identities, and using q\neq 0 , we obtain a_{44}=0=v_{2} from J_{134}^{2} and from J_{135}^{4} , respectively. The term J_{126} yields a_{73}=a_{23}=a_{33}=0 . Combining J_{135}^{3} and J_{125}^{2} , we get y_{12}=0 . From J_{137}^{2} with J_{156}^{5} results a_{71}=a_{74}=0 . Finally, from J_{156}^{6} and J_{126}^{2} , we conclude a_{53}=a_{63}=0 . Therefore, all A_{i} are upper-triangular. Case 2.2.: q=0 In this case A_{1}=0 , so J_{234}^{2}=a_{44}=0 . This implies J_{125}^{2}=y_{12}=0 and J_{135}^{2}=-y_{11}=0 , thus \Lambda_{1}=0 . In particular, we obtain R_{15}=0 . If a_{41}\neq 0 then {\mathfrak{h}} is decomposable by Lemma 4.3 . So we suppose a_{41}=0 for the remaining of the case. Let us first assume that a_{42}=0 , i.e. A_{4}=0 . From J_{245}^{2} we obtain that y_{42}=0 . Thus, if also y_{41}=0 , we would have \Lambda_{1}=0 and \Lambda_{4}=h(0,v_{4},0) , implying {\mathfrak{h}} decomposable by Lemma 4.2 . So let us assume that y_{41}\neq 0 . From Remark 4.5 , we have that [A_{4},A_{5}]=y_{41}A_{2} . Then y_{41}\neq 0 implies a_{23}=a_{33}=0 and a_{31}=-1 . Since J_{345}^{2}=(2-a_{34})y_{41} then a_{34}=2 . From [A_{3},A_{5}]=-A_{5}+(y_{32}-a_{54})A_{3} , we deduce a_{53}=0 . Now, from J_{234}^{1} we get v_{2}=0 and J_{12k}^{1} yields a_{k3}=0 with k=6,7 . Thus all A_{i} , i=1,\ldots,7 , are upper triangular. Let us now suppose that a_{42}\neq 0 . Since [A_{3},A_{4}]=-a_{42}A_{2} , then a_{23}=a_{33}=0 , a_{34}=2a_{31}+1 and [A_{2},A_{3}]=A_{2} implies a_{31}(a_{31}+1)=0 . From [A_{3},A_{5}] we obtain a_{53}(2a_{31}+1)=0 which implies a_{53}=0 . Taking J_{234}^{1}=v_{2}(3a_{31}+2) and J_{246}^{3}=-a_{42}a_{73} , we conclude v_{2}=a_{73}=0 . If a_{31}\neq 0 , then J_{267}^{7}=a_{63}a_{31} implies a_{63}=0 . If a_{31}=0 , then we have J_{246}^{2}=-(a_{63}+a_{71})a_{42},\quad J_{346}^{3}=(a_{63}-a_{74})a_{42},\quad J_{237}^{2}=a_{74}-a_{71} from which we finally obtain a_{63}=0 . Thus all A_{i} , i=1\cdots 7 , are upper triangular. ∎ 5 Exclusion of the case that {\mathfrak{a}} is diagonal Let G be a 7-dimensional Lie group endowed with a parallel left-invariant {\rm G}_{2}^{*} -structure \omega . Assume that its holonomy algebra {\mathfrak{h}} is indecomposable of type III. By Theorem 2.4 , {\mathfrak{h}} is conjugate to {\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k) where {\mathfrak{a}} can be 0,\,\mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}),\,\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}),\,\mathfrak{c}\mathfrak{o}(2),{\mathfrak{d}} , or {\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) , and {\mathfrak{m}}(1,0,k) is as in ( 10 ) for some k\in\{1,2\} . This section is devoted to showing that the last two cases for {\mathfrak{a}} cannot occur in our left-invariant context. Let us be more precise. Let {\mathfrak{h}}_{e}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) be an indecomposable holonomy algebra of type III. Then there exists a basis \mathfrak{e}=(e_{1},\dots,e_{7}) such that the holonomy with respect to \mathfrak{e} , which we denote {\mathfrak{h}} , is equal to one of the Lie algebras {\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k)\subset{\mathfrak{h}}^{\rm III} listed in Theorem 2.4 and such that \omega=\omega_{0} , where \omega_{0} is as in ( 1 ). In the following, {\mathfrak{a}} will always denote this Lie algebra {\mathfrak{a}}\subset\mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}) uniquely determined by {\mathfrak{h}}_{e} . In the basis \mathfrak{e} above, the Levi-Civita connection verifies \Lambda(X)\in{\mathfrak{h}}^{\rm I} for all X\in{\mathfrak{g}} by Proposition 3.1 . Thus for each i=1,\ldots,7 , we set \Lambda_{i}=h(A_{i},v_{i},(u_{i1},u_{i2}),(y_{i1},y_{i2})) as in ( 16 ). In addition, the curvature R is an element of \mathcal{K}({\mathfrak{h}}^{\rm III}) . From Table 1 , we know that {\mathfrak{a}}\subset{\mathfrak{d}} implies 0=R_{1i}=R_{2j}=R_{3k}=R_{4l},\quad\forall i\neq 5,\;\forall j\neq 6,\;\forall k\neq 7,\;\forall l\neq 5. (24) In addition, using the notation in ( 5 ), \displaystyle R_{15} \displaystyle= \displaystyle h(0,0,(r,s)), (25) \displaystyle R_{26} \displaystyle= \displaystyle h(0,0,(r,0)), (26) \displaystyle R_{37} \displaystyle= \displaystyle h(0,0,(0,s)), (27) \displaystyle R_{45} \displaystyle= \displaystyle h(0,0,(\sqrt{2}w_{1},\sqrt{2}w_{2})), (28) \displaystyle R_{56} \displaystyle= \displaystyle h\left(\mathop{{\rm diag}}(r,0),w_{1},(j_{3},t)\right), (29) \displaystyle R_{57} \displaystyle= \displaystyle h\left(\mathop{{\rm diag}}(0,s),w_{2},(t,j_{4})\right), (30) \displaystyle R_{67} \displaystyle= \displaystyle h(0,0,(-w_{1},-w_{2})). (31) Proposition 5.1 . Let {\mathfrak{h}}_{e}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) be indecomposable of type III . Then {\mathfrak{a}}\neq{\mathfrak{d}} . Proof. Let \mathfrak{e} and {\mathfrak{h}} be as above and assume {\mathfrak{a}}={\mathfrak{d}} . Proposition 3.1 implies that A_{i}\in{\mathfrak{d}} for i=1,\dots,7 , i.e. A_{i}=\mathop{{\rm diag}}(a_{i1},a_{i4}) . In particular, by ( 29 ) and ( 30 ), {\rm pr}_{A}([\Lambda_{i},R_{56}])={\rm pr}_{A}([\Lambda_{i},R_{57}])=0 for all i=1,\ldots,7 . Thus, r\neq 0 and s\neq 0 in ( 25 )–( 31 ), since otherwise we would get {\mathfrak{a}}\subsetneq{\mathfrak{d}} . Computing the curvature tensor using ( 13 ) and considering R_{26} and R_{37} as above, respectively, we have that \displaystyle r=(a_{61}+a_{64}-y_{11})y_{21}-v_{2}y_{31}-(3a_{21}+2a_{24})y_{61} \displaystyle s=(a_{71}+a_{74}-y_{12})y_{32}+v_{3}y_{22}-(3a_{31}+2a_{34})y_{72} We will make use of the second Bianchi identity ( 15 ) to infer that r=s=0 , thus leading to a contradiction. From B_{125}=0 and from B_{135}=0 we get that A_{2}=A_{3}=0 . Now, considering 0={\rm pr}_{y}(B_{257}) we conclude y_{21}=y_{22}=0 . Also, from {\rm pr}_{y}(B_{356}) we conclude y_{31}=y_{32}=0 . Thus r=s=0 , and the contradiction follows. ∎ Proposition 5.2 . Let {\mathfrak{h}}_{e} be indecomposable and of type III . Then {\mathfrak{a}}\not={\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) . Proof. Let \mathfrak{e} and {\mathfrak{h}} be as above and assume {\mathfrak{a}}={\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) . Using that [x,y]=\Lambda(x)y-\Lambda(y)x one can show that the Lie bracket relations between the basis elements e_{1},\ldots,e_{7} in \mathfrak{e} are as in Table A.3 . Note that Equations ( 24 ) – ( 31 ) hold for s=0 . By Proposition 3.1 , for i=1,\dots,7 , the map \Lambda_{i} has the form \Lambda_{i}=h(\mathop{{\rm diag}}(a_{i1},a_{i4}),v_{i},(0,u_{i2}),(y_{i1},y_{i2}))\in{\mathfrak{h}}^{\rm I}. This, together with ( 12 ), implies that r appearing in ( 24 )–( 31 ) verifies r\neq 0 . Indeed, if r=0 , then {\mathfrak{r}}_{0}=\mathop{{\rm span}}\{R(x,y)\mid x,y\in{\mathfrak{g}}\}\subset{\mathfrak{m}}(1,0,2) and thus {\mathfrak{h}}\subset{\mathfrak{m}}(1,0,2) , which implies {\mathfrak{a}}=0 contradicting our hypothesis. Now, we proceed as follows: Step 1: a_{k4}=-2a_{k1} holds for k=1,2,3,4,7 : We project onto {\mathfrak{a}} the second Bianchi identity in ( 15 ) for (i,j,k)=(5,6,k) , k=1,\dots,4,7 . Recall that A_{ij}={\rm pr}_{A}(R_{ij}) , i,j=1,\dots,7 . Since A_{56} is the only A_{ij} that does not vanish, this easily gives 0=(-2a_{k1}-a_{k4})A_{56} , which implies the assertion. Step 2: We have y_{11}=(2a_{61}+a_{64})/4 and y_{21}=-(2a_{51}+a_{54})/4 , and y_{31}=0 : Using ( 25 ) – ( 31 ) we compute the first component of {\rm pr}_{y} applied to both sides of ( 15 ). This gives 0=4ry_{11}-(2a_{61}+a_{64})r for (i,j,k)=(1,5,6) , 0=4ry_{21}+(2a_{51}+a_{54})r for (i,j,k)=(2,5,6) , and 0=y_{31}r for (i,j,k)=(3,5,6) . Since r\neq 0 , we obtain our claim. Step 3: v_{i}=-a_{i1}w_{1}/r for i=1,\dots,4 : Equations ( 24 ), ( 27 ) and Table 1 for \nabla_{e_{6}}R immediately imply (\nabla_{e_{6}}R)_{7i}=0 for i=1,2,3,4 . Furthermore, (\nabla_{e_{7}}R)_{i6}=0 holds for i=1,3,4 . For i=2 , the first component of {\rm pr}_{y} of the latter term equals (2a_{71}+a_{74})r , which vanishes by Step 1. Finally, the first component of {\rm pr}_{y}(R(\Lambda_{i}e_{6},e_{7})+R(e_{6},\Lambda_{i}e_{7})-[\Lambda_{i},R_{67}]) is equal to -a_{i1}w_{1}-v_{i}r for i=1,\dots,4 . Now ( 15 ) gives the assertion. Step 4: A_{3}=0 , u_{32}=0 : We have 0=R_{13}=[\Lambda_{1},\Lambda_{3}]-\Lambda(a_{14}e_{3}-\mathop{{\rm tr}}(A_{3})e_{1})=[\Lambda_{1},\Lambda_{3}]+\Lambda(2a_{11}e_{3}-a_{31}e_{1}) by Step 1. This implies 0=R_{132}^{2}=3a_{11}a_{31},\quad 0=R_{132}^{1}=-a_{31}u_{12},\quad 0=R_{135}^{2}=-a_{31}y_{11}, where we used y_{31}=0 for the third identity. On the other hand, we have R_{15}=[\Lambda_{1},\Lambda_{5}]-\Lambda(u_{12}e_{6}-\mathop{{\rm tr}}(A_{1})e_{5}+\sqrt{2}v_{1}e_{4}+y_{12}e_{3}+y_{11}e_{2}-\mathop{{\rm tr}}(A_{5})e_{1}). Using Steps 1 and 3 and y_{31}=0 from Step 2 we obtain, taking suitable \alpha,\beta,\gamma , R_{155}^{2}=\alpha a_{11}+\beta y_{11}+\gamma u_{12}=r\not=0. (32) This implies a_{31}=0 , thus A_{3}=0 . In particular, 0=R_{23}=[\Lambda_{2},\Lambda_{3}]-\Lambda(a_{24}e_{3}+u_{32}e_{1}). Proceeding as above we obtain u_{32}a_{11}=u_{32}u_{12}=u_{32}y_{11}=0 , thus ( 32 ) gives u_{32}=0 . Step 5: A_{1}=0 : First we consider R_{16}=[\Lambda_{1},\Lambda_{6}]+\Lambda(a_{11}e_{6}+v_{1}e_{3}+(y_{11}+\mathop{{\rm tr}}(A_{6}))e_{1}) , which implies A_{16}=a_{11}A_{6}+(y_{11}+\mathop{{\rm tr}}(A_{6}))A_{1} since A_{3}=0 . This gives a_{11}(2a_{61}+a_{64}+y_{11})=5a_{11}y_{11}=0. Assume that a_{11}\not=0 . Then the latter equation implies y_{11}=0 . Using Steps 1 and 2, we now get R_{165}^{2}=a_{11}y_{61}-a_{61}y_{11}+y_{11}^{2}=a_{11}y_{61}=0, thus y_{61}=0 . This together with Steps 1 and 2 gives R_{265}^{2}=r=a_{61}y_{21}\neq 0,\quad R_{125}^{2}=0=-a_{11}y_{21}, which is a contradiction. Hence a_{11}=0 , thus A_{1}=0 . Step 6: A_{4}=0 : Since A_{1}=A_{3}=0 , we have A_{24}=a_{41}A_{2} , and A_{46}=a_{41}A_{6} . This gives 0=R_{242}^{2}=a_{41}a_{21},\quad 0=R_{462}^{2}=a_{41}a_{61},\quad 0=R_{463}^{3}=a_{41}a_{64}. (33) On the other hand, R_{26}=[\Lambda_{2},\Lambda_{6}]+\Lambda(a_{21}e_{6}+v_{6}e_{3}+a_{61}e_{2}+(y_{21}-u_{62})e_{1}) . Since y_{13}=0 and y_{11}=(2a_{61}+a_{64})/4 , we obtain, for suitable \alpha , \beta , \gamma , 0\not=r=R_{265}^{2}=\alpha a_{21}+\beta a_{61}+\gamma a_{64}. Together with ( 33 ) this gives a_{41}=0 , thus A_{4}=0 . Step 7: a_{64}=2a_{61}\not=0 : We already know that a_{11}=0 and therefore also v_{1}=0 by Step 3. This implies R_{16}=[\Lambda_{1},\Lambda_{6}]+(y_{11}+\mathop{{\rm tr}}(A_{6}))\Lambda_{1} , hence 0=R_{165}^{2}=-(a_{61}+\mathop{{\rm tr}}(A_{6}))y_{11}+(y_{11}+\mathop{{\rm tr}}(A_{6}))y_{11}=\frac{1}{4}(-2a_{61}+a_{64})y_{11}. Thus in order to prove a_{64}=2a_{61} it suffices to prove y_{11}\not=0 . Assume y_{11}=0. Then R_{15}=[\Lambda_{1},\Lambda_{5}]-\Lambda(u_{12}e_{6}+y_{12}e_{3}-\mathop{{\rm tr}}(A_{5})e_{1}). In particular, A_{15}=-u_{12}A_{6} because of A_{1}=A_{3}=0 . This gives u_{12}a_{61}=u_{12}a_{64}=0 . Moreover, 0\not=r=R_{155}^{2}=-u_{12}y_{61} , thus u_{12}\not=0 . Consequently, a_{61}=a_{64}=0 . Thus we have A_{6}=0 besides A_{1}=A_{3}=A_{4}=0 , which gives \mathop{{\rm diag}}(r,0)=A_{56}=y_{61}A_{2} . However, this contradicts a_{24}=-2a_{21} since r\not=0 . Furthermore, the latter consideration also shows that A_{6}\not=0 , which is equivalent to a_{61}\neq 0 . Step 8: A_{2}=0 , u_{12}=0 : Since A_{1}=A_{3}=A_{4}=0 , we have 0=A_{26}=a_{21}A_{6}+a_{61}A_{2} , thus a_{21}a_{61}=0 . Since a_{61}\not=0 this implies a_{21}=0 and therefore also A_{2}=0 . Furthermore, we have 0=-R_{162}^{1}=-a_{64}u_{12}+(y_{11}+\mathop{{\rm tr}}(A_{6}))u_{12}=2a_{61}u_{12} , which implies u_{12}=0 . Step 9: final contradiction: Since A_{1}=A_{2}=A_{3}=A_{4}=0 we obtain \mathop{{\rm diag}}(r,0)=A_{56}=(a_{51}+u_{62})A_{6}-\mathop{{\rm tr}}(A_{6})A_{5} . Together with a_{64}=2a_{61} this gives \displaystyle r \displaystyle= \displaystyle(-2a_{51}+u_{62})a_{61}, \displaystyle 0 \displaystyle= \displaystyle(2a_{51}-3a_{54}+2u_{62})a_{61}, hence 2r=3(-2a_{51}+a_{54})a_{61}. (34) On the other hand, R_{15}=[\Lambda_{1},\Lambda_{5}]-\Lambda(y_{12}e_{3}+y_{11}e_{2}-\mathop{{\rm tr}}(A_{5})e_{1}) since u_{12}=0 and A_{1}=0 , which also implies v_{1}=0 by Step 3. Since A_{1}=0 by Step 5 and y_{31}=0 by Step 2, we get r=R_{155}^{2}=-(\mathop{{\rm tr}}(A_{5})+a_{51})y_{11}-y_{11}y_{21}+\mathop{{\rm tr}}(A_{5})y_{11}=-y_{11}(y_{21}+a_{51}). Applying Step 2 and Step 7, we obtain 4r={a_{61}}(-2a_{51}+a_{54}), which contradicts ( 34 ) since r\not=0 . ∎ Corollary 5.3 . Let {\mathfrak{h}}_{e} be indecomposable and of type III . If all A_{i} are upper triangular or all A_{i} are lower triangular, then {\mathfrak{a}}=0 . Proof. Let \mathfrak{e} , {\mathfrak{h}} and {\mathfrak{a}} be as above. By ( 11 ) and ( 12 ), {\mathfrak{a}} is contained in the subalgebra of \mathop{{\mathfrak{g}\mathfrak{l}}}(2,{\mathbb{R}}) spanned by A_{i} , i=1,\dots,7 . If all A_{i} are upper triangular or all are lower triangular, then we have {\mathfrak{h}}={\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k) with {\mathfrak{a}} being a subalgebra of the triangular 2\times 2 -matrices. However, by Theorem 2.4 , the only such subalgebras are 0,\,{\mathfrak{d}}, or {\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) . Since the last two cases are ruled out in Propositions 5.1 and 5.2 , we get {\mathfrak{a}}=0 . ∎ 6 Proving that {\mathfrak{a}} is trivial Let again {\mathfrak{h}}_{e}\subset\mathop{{\mathfrak{s}\mathfrak{o}}}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle}) be an indecomposable holonomy algebra of a left-invariant metric and assume that {\mathfrak{h}}_{e} is of type III. As in the previous section, let \mathfrak{e}=(e_{1},\dots,e_{7}) be a basis of {\mathfrak{g}} such that \omega=\omega_{0} , where \omega_{0} is as in ( 1 ) and the holonomy with respect to this basis is one of the Lie algebras {\mathfrak{h}}={\mathfrak{a}}\ltimes{\mathfrak{m}}(1,0,k)\subset{\mathfrak{h}}^{\rm III} listed in Theorem 2.4 . Lemma 6.1 . Let {\mathfrak{h}}_{e} be indecomposable and of type III . If {\mathfrak{a}}\not=0 , then (i) \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} , and therefore \Lambda_{i}=\Lambda(e_{i})=h(A_{i},v_{i},(y_{i1},y_{i2})) for some A_{i}=\begin{pmatrix}a_{i1}&a_{i2}\\ a_{i3}&a_{i4}\end{pmatrix},\quad v_{i}\in{\mathbb{R}},\quad(y_{i1},y_{i2})\in{\mathbb{R}}^{2},\quad i=1,\ldots,7. (ii) the Lie subalgebra {\mathfrak{g}}_{3} defined in ( 19 ) is abelian. Proof. If {\mathfrak{a}} is non-trivial, then by Proposition 5.2 , {\mathfrak{a}} is one of the Lie algebras in item 1 of Theorem 2.4 . In this case \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} by Proposition 3.1 . This gives (i). In particular, we can apply Proposition 4.8 , which leads to the following conclusion. Assume that {\mathfrak{g}}_{3} were not abelian. Then all A_{i} are upper triangular or all A_{i} are lower triangular. Now Corollary 5.3 implies {\mathfrak{a}}=0 , which contradicts our assumption. ∎ Lemma 6.2 . Let {\mathfrak{h}}_{e} be indecomposable and of type III . If {\mathfrak{a}}\not=0 , then A_{4}=0 . Proof. Assume that {\mathfrak{a}}\neq 0 . Then Lemma 6.1 implies that \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} and that {\mathfrak{g}}_{3} is abelian. Together with Eqns. ( 51 ), ( 52 ) and ( 57 ) this gives A_{1}=0,\ a_{21}=-a_{33},\ a_{22}=a_{31},\ a_{24}=a_{33},\ \mathop{{\rm tr}}(A_{3})=0. Since \mathop{{\rm tr}}(A_{1})=0 , from Lemma 4.7 , we obtain v_{1}=0 . For a contradiction, suppose that A_{4}\neq 0 . We may apply Lemma 4.3 for r=4 : If \det(A_{4})>0 or if \det(A_{4})=0 and \mathop{{\rm tr}}(A_{4})\not=0 , then {\mathfrak{h}} would be decomposable. Thus \det(A_{4})<0 or \det(A_{4})=\mathop{{\rm tr}}(A_{4})=0 . Then A_{4} has two real eigenvalues that are either both equal to zero or both unequal to zero with opposite signs. By choosing appropriate basis vectors e_{2} and e_{3} , we may assume that A_{4} is given in normal form. Case 1: A_{4}=\begin{pmatrix}r&0\\ 0&s\end{pmatrix} , rs<0 Analyzing the Jacobi equations, we get that J_{235}=0 implies v_{2}=v_{3}=0 . From J_{246}^{6}=-ra_{33} and J_{246}^{3}=-v_{4}a_{33}-a_{63} we obtain a_{33}=a_{63}=0 . Similarly, J_{247}^{6}=(2r-s)a_{23} and J_{247}^{3}=v_{4}a_{23}+(r-2s)a_{73} imply a_{23}=a_{73}=0 . Furthermore, J_{245}^{3}=2s(y_{22}-a_{53}) yields a_{53}=y_{22} . Using a_{23}=a_{33}=0 we obtain J_{247}^{1}=-2sy_{22} , which finally gives a_{53}=0 . Thus all A_{j} are upper triangular, which contradicts Corollary 5.3 . Case 2: A_{4}=\begin{pmatrix}0&1\\ 0&0\end{pmatrix} Here, we obtain from J_{235} that v_{2}=0 . That a_{23}=a_{33}=a_{73}=0 results from J_{246}^{3}=J_{246}^{6}=J_{246}^{7}=0 . From J_{236}^{7} we obtain a_{31}=0 and from J_{457}^{7} we conclude a_{53}=0 . Finally, -2a_{63}=2J_{467}^{7}+J_{456}^{5}-J_{347}^{2}=0 . This shows that all A_{i} are upper triangular matrices, which contradicts Corollary 5.3 . ∎ Proposition 6.3 . If {\mathfrak{h}}_{e} is indecomposable and of type III , then {\mathfrak{a}}=0 . Proof. Assume for a contradiction that {\mathfrak{a}}\neq 0 . Then \Lambda({\mathfrak{g}})\subset{\mathfrak{h}}^{\rm III} , {\mathfrak{g}}_{3} is abelian and A_{4}=0 by Lemmas 6.1 and 6.2 . Since {\mathfrak{g}}_{3} is abelian, ( 51 ), ( 52 ) and ( 57 ) imply A_{1}=0 , \mathop{{\rm tr}}(A_{2})=\mathop{{\rm tr}}(A_{3})=0 and a_{21}=-a_{33} , a_{22}=-a_{34} . In addition, v_{1}=0 by Lemma 4.7 . By Remark 4.5 and the fact that {\mathfrak{g}}_{3} is abelian, we get [A_{2},A_{3}]=0 . Therefore, \mathop{{\rm span}}\{A_{2},A_{3}\} is an abelian subalgebra of \mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) , thus of dimension \leq 1 . By Remark 3.2 , the subalgebra {\mathfrak{a}} of the holonomy algebra {\mathfrak{h}} is contained in the Lie algebra spanned by A_{i} , i=2,3,5,6,7 (since A_{1}=A_{4}=0 ). We shall prove that these matrices span a Lie algebra of dimension \leq 1 , from which we obtain {\mathfrak{a}}=0 by Theorem 2.4 and Proposition 5.2 . We shall split the proof in cases. Case 1: A_{2}\not=0 or A_{3}\not=0 . Then there exists a non-trivial linear combination \lambda A_{2}+\mu A_{3}=0 . After a change of basis in {\mathfrak{g}} by a transformation \Phi_{T} for T:=\begin{pmatrix}\mu&\lambda\\ -\lambda&\mu\end{pmatrix} as defined in ( 20 ) we have A_{2}\neq 0 and A_{3}=0 . Since a_{21}=-a_{33} , a_{22}=-a_{34} and \mathop{{\rm tr}}(A_{2})=0 , as shown above, we obtain A_{2}=\begin{pmatrix}0&0\\ a_{23}&0\end{pmatrix} with a_{23}\neq 0 . Recall the notation A_{ij}={\rm pr}_{A}(R_{ij}) , for i,j=1,\ldots,7 . From Table 1 , ( 14 ) and Eqns. ( 54 ), ( 66 ) we get A_{15}=-y_{11}A_{2}=0,\quad A_{45}=-y_{41}A_{2}=0, thus y_{11}=y_{41}=0 . By Table 1 and ( 68 ), we obtain A_{47}=-v_{4}A_{2}=0, so v_{4}=0 . Moreover, by Table 1 and ( 61 ), [A_{2},A_{7}]+a_{23}A_{6}-(v_{2}-a_{71})A_{2}=A_{27}\in\mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}), which implies \mathop{{\rm tr}}(A_{6})=0 . By A_{25}=[A_{2},A_{5}]+(a_{51}-y_{21})A_{2}=0 we get that [A_{5},A_{2}] is a multiple of A_{2} , so strictly lower triangular. This implies a_{52}=0 , since a_{23}\neq 0 . Using the table again one gets A_{37}=(a_{72}-v_{3})A_{2}=-A_{26}=-[A_{2},A_{6}]-a_{61}A_{2}. (35) Hence [A_{2},A_{6}] is a multiple of A_{2} and therefore strictly lower diagonal. This implies a_{23}a_{62}=0 , thus a_{62}=0 and [A_{2},A_{6}]=2a_{61}A_{2} . Equation ( 35 ) now gives (a_{72}-v_{3})A_{2}=-3a_{61}A_{2} , so v_{3}-a_{72}=3a_{61} . However, J_{237}^{3}=0 gives v_{3}=a_{61} so we conclude a_{72}=-2a_{61} . Using this and Table 1 , we get A_{67}=[A_{6},A_{7}]+a_{61}A_{7}+(a_{63}-a_{71})A_{6}-v_{6}A_{2}=0, As A_{6} , A_{2} are lower triangular, the upper right corner of the matrix in the last equation gives a_{61}=0 . Therefore, A_{2},A_{5},A_{6},A_{7} are all lower triangular. By Corollary 5.3 , the holonomy {\mathfrak{h}} would be decomposable, which is a contradiction. Case 2: A_{2}=A_{3}=0 . We may assume that y_{11}\in\{0,1\} and y_{12}=0 . Indeed, if one of the numbers is different from zero, we may assume that y_{11} is this number by possibly interchanging e_{2} , e_{3} and e_{6} , e_{7} using \Phi_{T} for T=\begin{pmatrix}0&1\\ 1&0\end{pmatrix} (see ( 20 )). After that, a base change by \Phi_{T} for T=\begin{pmatrix}y_{11}&0\\ y_{12}&\frac{1}{y_{11}}\end{pmatrix} yields y_{11}=1 . We now consider the cases y_{11}=1 and y_{11}=0 separately. Recall that {\mathfrak{a}} is generated by A_{5},A_{6},A_{7} , since A_{i}=0 for i=1,\ldots,4 . 1. y_{11}=1 . By J_{245}^{2}=\sqrt{2}v_{2} and J_{345}^{2}=\sqrt{2}v_{3} , we get v_{2}=v_{3}=0 . From J_{15i}^{j}=0 for i=6,7 , j=2,3 we get a_{61}=1 , a_{63}=a_{71}=a_{73}=0. In particular, A_{6},A_{7} are both upper triangular. If a_{53}=0 , then A_{i} is upper triangular for all i=1,\dots,7 , thus {\mathfrak{h}} is decomposable or {\mathfrak{a}}=0 by Corollary 5.3 . So we assume a_{53}\neq 0 . If a_{74}\not=0 , then J_{167}^{1}=(a_{72}-a_{64})a_{74} , J_{257}^{3}=2(a_{53}-y_{22})a_{74} would imply a_{64}=a_{72} and y_{22}=a_{53} , which would lead to the contradiction 0=J_{157}^{1}-J_{567}^{6}=-2a_{53}\not=0 . Hence a_{74}=0 . Now J_{567}^{6}=(a_{72}-2a_{64}-1)a_{53},\quad J_{367}^{2}=a_{72}^{2}-a_{72}=0 implies that either a_{72}=0 , a_{64}=-1/2 or a_{72}=1 , a_{64}=0 . Both cases lead to a contradiction. Indeed, in the first case we obtain 0=J_{357}^{3}=a_{53}/2\not=0 and in the second one 0=J_{157}^{1}+J_{357}^{3}=-3a_{53}\not=0 . 2. y_{11}=y_{12}=0 so that \Lambda_{1}=0 and thus R_{15}=0 . By Table 1 , \mathop{{\rm tr}}(A_{56})=\mathop{{\rm tr}}(A_{57})=0 and thus {\mathfrak{a}}\subset\mathfrak{sl}(2,{\mathbb{R}}) . By Theorem 2.4 , {\mathfrak{a}} is either \mathfrak{sl}(2,{\mathbb{R}}) or {\mathfrak{a}}=0 . (a) v_{4}=0 . If y_{41}=y_{42}=0 , then \Lambda_{4}=0 and Lemma 4.2 , implies that {\mathfrak{h}} is decomposable. So either y_{41} or y_{42} is non-zero. Similarly to the discussion for (y_{11},y_{12}) we see that we can assume y_{41}=1 and y_{42}=0 from now on. One has {\rm pr}_{v}(R_{45})=-v_{2} . Thus v_{2}=0 by Table 1 . From J_{456}^{3}=J_{457}^{3}=0 , we obtain a_{63}=a_{73}=0 , thus A_{6} , A_{7} are upper triangular. If a_{53} vanishes, then also A_{5} is upper triangular and we are done by Corollary 5.3 . So we assume a_{53}\neq 0 . The identies J_{456}^{2}=J_{457}^{2}=0 yield a_{64}=-2a_{61} and a_{74}=-2a_{71} . If a_{71}=0 , then also a_{74}=0 . Hence J_{157}^{1}=-a_{53}(a_{61}+a_{64}) , which implies a_{61}+a_{64}=0 . Thus a_{61}=a_{64}=0 . Now J_{567}^{6}=0 yields a_{72}=0 . Moreover, J_{346}^{1}=0 gives v_{3}=0 , which implies J_{367}^{1}+J_{256}^{2}=a_{53}a_{62} , hence a_{62}=0 . We obtain A_{6}=A_{7}=0 . Consequently, {\mathfrak{a}}\subset{\mathbb{R}}\cdot A_{5} , which gives {\mathfrak{a}}=0 . If a_{71}\not=0 , then J_{347}^{1}=-\sqrt{2}v_{3}a_{71} implies v_{3}=0 . Moreover, J_{367}^{3}=-2a_{71}(a_{72}+3a_{61}) gives a_{72}=-3a_{61} . Using the condition J_{567}^{6}=0 we now obtain \mathop{{\rm tr}}(A_{5})=0 . From J_{156}^{1}=J_{157}^{1}=0 we get a_{61}a_{51}+a_{52}a_{71}=0,\quad a_{53}a_{61}-a_{71}a_{51}=0. (36) After an explicit computation, this implies [A_{5},A_{7}]=3a_{71}A_{5} . Multiplying J_{367}^{2}=4(3a_{61}^{2}+a_{62}a_{71}) by a_{53}/a_{71} and using ( 36 ) we obtain a_{53}a_{62}+3a_{51}a_{61}=0. (37) Eq. ( 36 ) together with ( 37 ) imply that the non-trivial linear combination a_{53}A_{6}-a_{51}A_{7} vanishes. So in this case A_{5},A_{6},A_{7} span a 2-dimensional Lie subalgebra of \mathfrak{gl}(2,{\mathbb{R}}) containing {\mathfrak{a}} . As mentioned above, the only possibilities for {\mathfrak{a}} are \mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) or {\mathfrak{a}}=0 . Therefore {\mathfrak{a}}=0 . (b) v_{4}\neq 0 . Here J_{467}^{2}=J_{467}^{3}=0 gives v_{2}=-\mathop{{\rm tr}}(A_{7}) and v_{3}=\mathop{{\rm tr}}(A_{6}) . Then J_{156}^{1}=J_{157}^{1}=J_{256}^{4}=J_{257}^{4}=J_{346}^{1}=J_{347}^{1}=0 shows that (-\mathop{{\rm tr}}(A_{7}),\mathop{{\rm tr}}(A_{6})) is in the common kernel of A_{5}^{\top} , A_{6}^{\top} and A_{7}^{\top} . If this vector is non-zero, then there exists a basis such that A_{5},A_{6},A_{7} are all lower triangular and the proposition follows from Corollary 5.3 . Let us assume now that \mathop{{\rm tr}}(A_{7})=\mathop{{\rm tr}}(A_{6})=0 . Then v_{2}=v_{3}=0 . Using this we conclude from {\rm pr}_{v}(R_{45})=0 that \mathop{{\rm tr}}(A_{5})=-\sqrt{2}v_{4} . From J_{567}^{4}=J_{567}^{7}=J_{567}^{6}=0 we have \displaystyle-(a_{63}-a_{71})v_{6}+(a_{61}+a_{72})v_{7} \displaystyle= \displaystyle 0, (38) \displaystyle(a_{63}-a_{71})a_{52}+(a_{61}+a_{72})a_{51} \displaystyle= \displaystyle 0, (39) \displaystyle(a_{63}-a_{71})a_{54}+(a_{61}+a_{72})a_{53} \displaystyle= \displaystyle 0. (40) Recall that R_{15}=0 . Since {\mathfrak{h}} is indecomposable, we get w_{1}\not=0 or w_{2}\not=0 by Lemma 4.2 . On the other hand, w_{1}={\rm pr}_{v}(R_{56})=a_{51}v_{6}+a_{52}v_{7} and w_{2}={\rm pr}_{v}(R_{57})=a_{53}v_{6}+a_{54}v_{7} . Thus the system of linear equations ( 38 ) – ( 40 ) for a_{63}-a_{71} and a_{61}+a_{72} has only the trivial solution. Hence a_{63}=a_{71},\quad a_{61}=-a_{72}. (41) This implies 0=A_{67}=[A_{6},A_{7}]-{\rm pr}_{A}(\Lambda([e_{6},e_{7}]))=[A_{6},A_{7}] . Since A_{6},A_{7}\in\mathop{{\mathfrak{sl}}}(2,{\mathbb{R}}) , we see that A_{6} and A_{7} are linearly dependent. Applying a suitable transformation \Phi_{T} , we may assume A_{7}=0 . Equation ( 41 ) implies A_{6}=\begin{pmatrix}0&a_{62}\\ 0&0\end{pmatrix}. Moreover, A_{7}=0 implies {\rm pr}_{y}R_{37}=0 and therefore b_{4}=0 . We compute A_{57}=a_{53}A_{6} , which gives 0=b_{4}=a_{62}a_{53} . If a_{62}=0 , then the Lie algebra generated by A_{5},A_{6},A_{7} is at most one-dimensional, thus {\mathfrak{a}}=0 . If a_{53}=0 , then all A_{j} , j=1,\dots,7 , are upper triangular, hence {\mathfrak{a}}=0 also in this case. ∎ 7 Examples and proof of Theorem 1.1 In this section we first present examples of Lie groups endowed with parallel {\rm G}_{2}^{*} -structures whose holonomies are the abelian Lie algebras appearing in Theorem 1.1 . Later we conclude with the detailed proof of this main theorem, using the results in the previous sections. Let us consider the one-parameter family of Lie algebras {\mathfrak{g}}_{\varepsilon} defined by \displaystyle\textstyle[e_{1},e_{5}]=\frac{1}{2}e_{1},\ [e_{2},e_{5}]=-\frac{1}{2}e_{2},\ [e_{2},e_{6}]=-\frac{1}{2}e_{1},\ [e_{3},e_{5}]=2e_{3}, \displaystyle\textstyle[e_{3},e_{7}]=-\frac{1}{2}e_{1},\ [e_{4},e_{5}]=-e_{2},\ [e_{4},e_{6}]=-e_{1}, (42) \displaystyle\textstyle[e_{5},e_{6}]=-\varepsilon e_{2}-2e_{4}-e_{6},\ [e_{5},e_{7}]=-\varepsilon e_{3}+\frac{3}{2}e_{7},\ [e_{6},e_{7}]=-\sqrt{2}e_{2}, for \varepsilon\in\mathbb{R} . Obviously, {\mathfrak{p}}:=\mathop{{\rm span}}\{e_{1},\dots,e_{4}\} is an abelian ideal of {\mathfrak{g}}_{\varepsilon} . Thus {\mathfrak{g}}_{\varepsilon} is an extension of {\mathfrak{s}}:={\mathfrak{g}}_{\varepsilon}/{\mathfrak{p}} by {\mathfrak{p}} . The Lie algebra {\mathfrak{s}} is isomorphic to \{[X,Y]=-Y,[X,Z]=\frac{3}{2}Z\} . The Lie algebra {\mathfrak{g}}_{\varepsilon} can also be understood as a semi-direct product of the ideal {\mathfrak{n}} spanned by e_{j} , j\not=5 , by {\mathbb{R}} . Let G_{\varepsilon} be the simply connected Lie group with Lie algebra {\mathfrak{g}}_{\varepsilon} and N the nilpotent subgroup with Lie algebra {\mathfrak{n}} . Then G_{\varepsilon}={{\mathbb{R}}}\ltimes N . We write elements of G_{\varepsilon} as (t,n) according to this decomposition. Doing so, the left-invariant vector field e_{5} is equal to \partial_{t} . Let {\mathfrak{g}}_{\varepsilon} denote the Lie algebra defined by ( 42 ) endowed with the inner product ( 2 ) induced by the {\rm G}_{2}^{*} -structure \omega_{0} in ( 1 ). Given two parameters \varepsilon and \varepsilon^{\prime} , the Lie algebras {\mathfrak{g}}_{\varepsilon} and {\mathfrak{g}}_{\varepsilon^{\prime}} are isomorphic. For instance, an isomorphism is given by \varphi:{\mathfrak{g}}_{\varepsilon}\longrightarrow{\mathfrak{g}}_{\varepsilon^{\prime}} such that \varphi(e_{i})=e_{i} , for i=1,\ldots,5 , and \varphi(e_{6})=e_{6}+\tfrac{2}{3}(\varepsilon^{\prime}-\varepsilon)e_{2} , \varphi(e_{7})=e_{7}+\tfrac{2}{7}(\varepsilon-\varepsilon^{\prime})e_{3} . We will prove that {\mathfrak{g}}_{\varepsilon} is isometrically isomorphic to {\mathfrak{g}}_{\varepsilon^{\prime}} if and only if \varepsilon=\varepsilon^{\prime} by showing that the parameter \varepsilon is determined by the Lie algebra structure and the inner product. In the following we will choose vectors f_{4},f_{5},f_{6},f_{7} only using the Lie algebra structure and the inner product. After choosing, each of these vectors will be described with respect to the basis used in ( 42 ). The center {\mathfrak{z}} of {\mathfrak{g}}_{\varepsilon} is one-dimensional and non-degenerate. We choose f_{4}\in{\mathfrak{z}} such that \langle f_{4},f_{4}\rangle=-1 . Then f_{4}=\pm(e_{4}-2e_{2}) . There exists exactly one further one-dimensional subspace \mathfrak{u} which is invariant under the adjoint representation, namely \mathfrak{u}={\mathbb{R}}\cdot e_{1} . We choose an element f_{5} which acts by multiplication with -1/2 on \mathfrak{u} . Then f_{5}=e_{5}+f_{5}^{\prime} , where f_{5}^{\prime}\in\mathop{{\rm span}}\{e_{1},\dots,e_{4},e_{6},e_{7}\} . The action of f_{5} on \mathfrak{q}:=[{\mathfrak{g}}_{\varepsilon},{\mathfrak{g}}_{\varepsilon}]/{\mathfrak{g}}_{\varepsilon}^{(2)}=\mathop{{\rm span}}\{e_{1},e_{2},e_{3},2e_{4}+e_{6},e_{7}\}/\mathop{{\rm span}}\{e_{1},e_{2}\} is independent of the choice of f_{5} . It is diagonalizable with eigenvalues -2 , -1 and 3/2 . We denote the corresponding eigenspaces by E_{-2} , E_{-1} and E_{3/2} . Choose a representative of an eigenvector in E_{-1} such that |\langle f_{6},f_{4}\rangle|=4 . Then f_{6}=\pm(e_{6}+2e_{4})+a_{6}e_{1}+b_{6}e_{2} for some a_{6},b_{6}\in{\mathbb{R}} . Note that the inner product on {\mathfrak{g}}_{\varepsilon} induces a non-degenerate inner product on \mathfrak{q} , and let f_{7} be a representative of a non-zero vector in an isotropic complement of E_{-2} in E_{-1}^{\perp}\subset\mathfrak{q} . We ‘normalize’ f_{7} by requiring that \langle[f_{6},f_{7}],f_{6}\rangle=-\sqrt{2} , which yields f_{7}=e_{7}+a_{7}e_{1}+b_{7}e_{2} for some a_{7},b_{7}\in{\mathbb{R}} . Then \langle[f_{5},f_{7}],f_{7}\rangle depends only on {[\cdot\,,\cdot]} and {\langle\cdot\,,\cdot\rangle} , but not on the choices we made. Our assertion now follows from \langle[f_{5},f_{7}],f_{7}\rangle=-\varepsilon. Consider the left-invariant metric on G_{\varepsilon} defined by ( 2 ). The linear maps \Lambda_{i} belong to {\mathfrak{h}}^{\rm III} . We have \Lambda_{1}=0 and \displaystyle\Lambda_{2}=h(0,0,y_{2}),\quad\ y_{2}=(1/2,0), (43) \displaystyle\Lambda_{3}=h(0,0,y_{3}),\quad\ y_{3}=(0,1/2), (44) \displaystyle\Lambda_{4}=h(0,0,y_{4}),\quad\ y_{4}=(-1,0), (45) \displaystyle\Lambda_{5}=h(A_{5},0,0),\quad A_{5}=\mathop{{\rm diag}}(1,-3/2), (46) \displaystyle\Lambda_{6}=h(0,v_{6},y_{6}),\quad v_{6}=\sqrt{2},\ y_{6}=(\varepsilon,0), (47) \displaystyle\Lambda_{7}=h(0,0,y_{7}),\quad\ y_{7}=(0,\varepsilon). (48) The holonomy of our metric is equal to {\mathfrak{m}}(1,0,2) for \varepsilon\not=0 and to {\mathfrak{m}}(1,0,1) for \varepsilon=0 . The vectors e_{1},e_{2},e_{3}\in T_{e}G_{\varepsilon} are invariants of the holonomy representation. The parallel vector fields corresponding to these vectors are not left-invariant. In terms of the left-invariant vector fields e_{1},e_{2},e_{3} , they are equal to e^{\frac{1}{2}t}e_{1} , e^{-t}e_{2} , and e^{\frac{3}{2}t}e_{3} , respectively. Indeed, e.g. \textstyle\nabla_{e_{5}}(e^{\frac{1}{2}t}e_{1})=\partial_{t}(e^{\frac{1}{2}t})e_{1}+e^{\frac{1}{2}t}\Lambda_{5}(e_{1})=\frac{1}{2}e^{\frac{1}{2}t}e_{1}-e^{\frac{1}{2}t}\cdot\frac{1}{2}e_{1}=0. We want to deform \omega_{0} by a pointwise orthonormal transformation A depending on t . In the end we will obtain a family of parallel {\rm G}_{2}^{*} -structures that are not left-invariant but induce the same left-invariant metric like \omega_{0} . Consider M:=M(a,b,c)\in\mathop{{\mathfrak{s}\mathfrak{o}}}(4,3) given by e_{1},e_{2},e_{3}\in\ker M and \displaystyle\textstyle M(e_{4})=a\exp(\frac{1}{2}t)e_{1}+b\exp(-t)e_{2}+c\exp(\frac{3}{2}t)e_{3}, \displaystyle\textstyle M(e_{5})=a\exp(\frac{1}{2}t)e_{4},\quad M(e_{6})=b\exp(-t)e_{4},\quad M(e_{7})=c\exp(\frac{3}{2}t)e_{4} and put A:=A(a,b,c):=\exp M(a,b,c) . Now we obtain a 3-parameter family of three-forms \omega(a,b,c)=A(a,b,c)^{*}\omega_{0}. (49) Since A is orthogonal, all these 3-forms induce the same metric as \omega_{0} . Moreover, they are parallel. This can be seen as follows. Instead of showing that A^{*}\omega_{0} is parallel we prove that (A^{-1})^{*}\nabla(A^{*}\omega_{0}) vanishes. Since \omega_{0} is parallel, this is equivalent to \omega_{0}(D_{l}e_{i},e_{j},e_{k})+\omega_{0}(e_{i},D_{l}e_{j},e_{k})+\omega_{0}(e_{i},e_{j},D_{l}e_{k})=0 (50) for all i,j,k,l\in\{1,\dots,7\} , where D_{l}=A\circ\nabla_{e_{l}}A^{-1} . Equations ( 43 ) – ( 48 ) give D_{l}=0 for l\not=6 , D_{6}e_{i}=0 for i=1,2,3,4, and \displaystyle D_{6}(e_{5}) \displaystyle= \displaystyle\textstyle 2b\exp(-t)e_{2}+2c\exp(\frac{3}{2}t)e_{3}, \displaystyle\quad D_{6}(e_{6}) \displaystyle= \displaystyle-2b\exp(-t)e_{1}, \displaystyle D_{6}(e_{7}) \displaystyle= \displaystyle\textstyle-2c\exp(\frac{3}{2}t)e_{1}, which immediately implies ( 50 ). Although we have neither explained nor worked with the relation between spinors and {\rm G}_{2}^{*} -structures in this paper let us now roughly explain the idea behind the above construction using it. For details on this relation see [ 5 ] . Every (not necessarily parallel) {\rm G}_{2}^{*} -structure on a Lie group G that induces the same metric on G as the {\rm G}_{2}^{*} -structure \omega is given by a 3-form A^{*}\omega , where A is function from G to {\rm O}({\mathfrak{g}},{\langle\cdot\,,\cdot\rangle})\cong{\rm O}(4,3) . If A maps to \mathop{{\rm SO}}(4,3) , then A^{*}\omega induces the same orientation as \omega , otherwise we get the opposite orientation. Let us first consider the situation at the identity e\in G . The \mathop{{\rm SO}}(4,3) -orbit of \omega in \bigwedge^{3}({\mathfrak{g}}^{*}) corresponds bijectively to the set of non-isotropic elements in the projective real spinor representation (\Delta_{4,3}\setminus\{0\})/{\mathbb{R}}^{*}=({\mathbb{R}}^{4,4}\setminus\{0\})/{\mathbb{R}}^{*} of {\rm Spin}(4,3) . Thus it can be identified with (S^{4,3}/{\mathbb{Z}}_{2})\cup(H^{3,4}/{\mathbb{Z}}_{2}) , where S^{4,3} and H^{3,4} are the sets of space- and time-like unit vectors in {\mathbb{R}}^{4,4} , respectively. Furthermore, it is well known that given a non-isotropic spinor \psi_{0}\in\Delta_{4,3} the Clifford multiplication X\mapsto X\cdot\psi_{0} defines an isometry from {\mathfrak{g}} to \psi_{0}^{\perp}\subset\Delta_{4,3} . Now we again turn to 3-forms and spinor fields on G . Then of course a {\rm G}_{2}^{*} -structure A^{*}\omega is parallel if and only if the associated unit spinor field (unique up to sign) is parallel. Consider \omega as a parallel 3-form on G and let \psi_{0} be the unit spinor field that corresponds to \omega . Assume that there exists a parallel isotropic vector field X on G . Then \psi^{\prime}:=\psi_{0}+X\cdot\psi_{0} is a parallel unit spinor field. Hence it corresponds to a parallel {\rm G}_{2}^{*} -structure \omega^{\prime}=A^{*}\omega for some function A:{\mathfrak{g}}\to\mathop{{\rm SO}}(4,3) . In the present example we applied this construction to the 3-dimensional space of isotropic parallel vector fields, which gave us a 3-parameter family of parallel {\rm G}_{2}^{*} -structures. Conversely, let \omega^{\prime}=A^{*}\omega be parallel and let \psi^{\prime} be the corresponding parallel unit spinor field (unique up to sign). We decompose \psi^{\prime}=f\cdot\psi_{0}+\psi^{\perp} , where f is a function on G and \psi^{\perp}\perp\psi_{0} . Since \psi^{\prime} and \psi_{0} are parallel, f is constant and \psi^{\perp} is also parallel. Thus we can define a parallel vector field X by \psi^{\perp}=X\cdot\psi_{0} . If the holonomy {\mathfrak{h}} is indecomposable, then X is isotropic. In particular, the three-parameter family in ( 49 ) contains all parallel {\rm G}_{2}^{*} -structures that induce the same metric and the same orientation on G_{\varepsilon} as \omega_{0} . The parallel {\rm G}_{2}^{*} -structures that induce the same metric and but the reverse orientation on G are then exactly the 3-forms -\omega(a,b,c) for a,b,c\in{\mathbb{R}} . Proof of Theorem 1.1 . Let {\mathfrak{h}} be a subalgebra of {{\mathfrak{g}}^{*}_{2}} whose natural representation on {\mathbb{R}}^{4,3} is indecomposable and of type III. By Lemma 2.1 , {\mathfrak{h}} is abelian if and only if it is conjugate to the two-dimensional Lie algebra {\mathfrak{m}}(1,0,1) or the three-dimensional Lie algebra {\mathfrak{m}}(1,0,2) . In particular, this proves the second assertion of the Theorem. Furthermore, the holonomy algebras of the examples described above are {\mathfrak{m}}(1,0,1) and {\mathfrak{m}}(1,0,2) . This shows one direction of the first assertion. Assume now that {\mathfrak{h}} is an indecomposable type III infinitesimal holonomy algebra of a left-invariant {\rm G}_{2}^{*} -structure. Then, on the one hand, {\mathfrak{h}} is on the list in Theorem 2.4 . On the other hand, {\mathfrak{a}}=0 holds by Proposition 6.3 . Thus {\mathfrak{h}}={\mathfrak{m}}(1,0,1) or {\mathfrak{h}}={\mathfrak{m}}(1,0,2) . ∎ Appendix A Appendix A.1 Algebraic curvature tensors with values in {{\mathfrak{h}}^{\rm III}} R(e_{i},e_{j}) A v y R_{15} 0 0 (b_{1}+b_{4},b_{3}+c_{4}) R_{26} \left(\begin{array}[]{cc}-a_{1}&-a_{2}\\ -a_{3}&a_{1}\end{array}\right) 0 (b_{1},b_{3}) R_{27} \left(\begin{array}[]{cc}-a_{3}&a_{1}\\ j_{1}&a_{3}\end{array}\right) 0 (b_{3},c_{3}) R_{36} \left(\begin{array}[]{cc}-a_{2}&j_{2}\\ a_{1}&a_{2}\end{array}\right) 0 (b_{2},b_{4}) -R_{67}=\frac{1}{\sqrt{2}}R_{45} 0 0 (w_{1},w_{2}) R_{56} \left(\begin{array}[]{cc}b_{1}&b_{2}\\ b_{3}&b_{4}\end{array}\right) w_{1} (j_{3},t) R_{57} \left(\begin{array}[]{cc}b_{3}&b_{4}\\ c_{3}&c_{4}\end{array}\right) w_{2} (t,j_{4}) R_{12}=R_{13}=R_{14}=R_{16}=R_{17}=R_{23}=R_{24}=0 R_{25}=R_{34}=R_{35}=R_{46}=R_{47}=0 R_{37}=R_{15}-R_{26} Table 1: Curvature A.2 Commutators for \Lambda_{i}\in{\mathfrak{h}}^{\rm III} , i=1,\ldots,7 . \displaystyle\,[e_{1},e_{2}] \displaystyle= \displaystyle a_{13}e_{3}+a_{11}e_{2}-\mathop{{\rm tr}}(A_{2})e_{1} (51) \displaystyle\,[e_{1},e_{3}] \displaystyle= \displaystyle a_{14}e_{3}+a_{12}e_{2}-\mathop{{\rm tr}}(A_{3})e_{1} (52) \displaystyle\,[e_{1},e_{4}] \displaystyle= \displaystyle(\sqrt{2}v_{1}-\mathop{{\rm tr}}(A_{4}))e_{1} (53) \displaystyle\,[e_{1},e_{5}] \displaystyle= \displaystyle-\mathop{{\rm tr}}(A_{1})e_{5}+\sqrt{2}v_{1}e_{4}+y_{12}e_{3}+y_{11}e_{2}-\mathop{{\rm tr}}(A_{5})e_{1}, (54) \displaystyle\,[e_{1},e_{6}] \displaystyle= \displaystyle-a_{12}e_{7}-a_{11}e_{6}-v_{1}e_{3}+(-y_{11}-\mathop{{\rm tr}}(A_{6}))e_{1}, (55) \displaystyle\,[e_{1},e_{7}] \displaystyle= \displaystyle-a_{14}e_{7}-a_{13}e_{6}+v_{1}e_{2}+(-y_{12}-\mathop{{\rm tr}}(A_{7}))e_{1}, (56) \displaystyle\,[e_{2},e_{3}] \displaystyle= \displaystyle(a_{24}-a_{33})e_{3}+(a_{22}-a_{31})e_{2}, (57) \displaystyle\,[e_{2},e_{4}] \displaystyle= \displaystyle-a_{43}e_{3}-a_{41}e_{2}+\sqrt{2}v_{2}e_{1}, (58) \displaystyle\,[e_{2},e_{5}] \displaystyle= \displaystyle-\mathop{{\rm tr}}(A_{2})e_{5}+\sqrt{2}v_{2}e_{4}+(y_{22}-a_{53})e_{3}+(y_{21}-a_{51})e_{2}, (59) \displaystyle\,[e_{2},e_{6}] \displaystyle= \displaystyle-a_{22}e_{7}-a_{21}e_{6}+(-v_{2}-a_{63})e_{3}-y_{21}e_{1}-a_{61}e_{2}, (60) \displaystyle\,[e_{2},e_{7}] \displaystyle= \displaystyle-a_{24}e_{7}-a_{23}e_{6}-a_{73}e_{3}+(v_{2}-a_{71})e_{2}-y_{22}e_{1}, (61) \displaystyle\,[e_{3},e_{4}] \displaystyle= \displaystyle-a_{44}e_{3}-a_{42}e_{2}+\sqrt{2}v_{3}e_{1}, (62) \displaystyle\,[e_{3},e_{5}] \displaystyle= \displaystyle-\mathop{{\rm tr}}(A_{3})e_{5}+\sqrt{2}v_{3}e_{4}+(y_{32}-a_{54})e_{3}+(y_{31}-a_{52})e_{2}, (63) \displaystyle\,[e_{3},e_{6}] \displaystyle= \displaystyle-a_{32}e_{7}-a_{31}e_{6}+(-v_{3}-a_{64})e_{3}-y_{31}e_{1}-a_{62}e_{2}, (64) \displaystyle\,[e_{3},e_{7}] \displaystyle= \displaystyle-a_{34}e_{7}-a_{33}e_{6}-a_{74}e_{3}+(v_{3}-a_{72})e_{2}-y_{32}e_{1}, (65) \displaystyle\,[e_{4},e_{5}] \displaystyle= \displaystyle-\mathop{{\rm tr}}(A_{4})e_{5}+\sqrt{2}v_{4}e_{4}+y_{42}e_{3}+y_{41}e_{2}-\sqrt{2}v_{5}e_{1}, (66) \displaystyle\,[e_{4},e_{6}] \displaystyle= \displaystyle-a_{42}e_{7}-a_{41}e_{6}-v_{4}e_{3}+(-y_{41}-\sqrt{2}v_{6})e_{1}, (67) \displaystyle\,[e_{4},e_{7}] \displaystyle= \displaystyle-a_{44}e_{7}-a_{43}e_{6}+v_{4}e_{2}+(-y_{42}-\sqrt{2}v_{7})e_{1}, (68) \displaystyle\,[e_{5},e_{6}] \displaystyle= \displaystyle-a_{52}e_{7}-a_{51}e_{6}+\mathop{{\rm tr}}(A_{6})e_{5}-\sqrt{2}v_{6}e_{4}-(v_{5}+y_{62})e_{3}-y_{51}e_{1}-y_{61}e_{2},\qquad\ \ (69) \displaystyle\,[e_{5},e_{7}] \displaystyle= \displaystyle-a_{54}e_{7}-a_{53}e_{6}+\mathop{{\rm tr}}(A_{7})e_{5}-\sqrt{2}v_{7}e_{4}-y_{72}e_{3}+(v_{5}-y_{71})e_{2}-y_{52}e_{1}, (70) \displaystyle\,[e_{6},e_{7}] \displaystyle= \displaystyle(-a_{64}+a_{72})e_{7}+(-a_{63}+a_{71})e_{6}+v_{7}e_{3}+v_{6}e_{2}+(-y_{62}+y_{71})e_{1} (71) A.3 Commutators for {\mathfrak{a}}={\mathbb{R}}\cdot\mathop{{\rm diag}}(1,0) \displaystyle\,[e_{1},e_{2}] \displaystyle= \displaystyle a_{11}e_{2}-(u_{12}+\mathop{{\rm tr}}(A_{2}))e_{1}, \displaystyle\,[e_{1},e_{3}] \displaystyle= \displaystyle a_{14}e_{3}-\mathop{{\rm tr}}(A_{3})e_{1}, \displaystyle\,[e_{1},e_{4}] \displaystyle= \displaystyle\sqrt{2}u_{12}e_{3}+(\sqrt{2}v_{1}-\mathop{{\rm tr}}(A_{4}))e_{1}, \displaystyle\,[e_{1},e_{5}] \displaystyle= \displaystyle u_{12}e_{6}-\mathop{{\rm tr}}(A_{1})e_{5}+\sqrt{2}v_{1}e_{4}+y_{12}e_{3}+y_{11}e_{2}-\mathop{{\rm tr}}(A_{5})e_{1}, \displaystyle\,[e_{1},e_{6}] \displaystyle= \displaystyle-a_{11}e_{6}-v_{1}e_{3}+(-y_{11}-\mathop{{\rm tr}}(A_{6}))e_{1}, \displaystyle\,[e_{1},e_{7}] \displaystyle= \displaystyle-a_{14}e_{7}+\sqrt{2}u_{12}e_{4}+v_{1}e_{2}+(-y_{12}-\mathop{{\rm tr}}(A_{7}))e_{1}, \displaystyle\,[e_{2},e_{3}] \displaystyle= \displaystyle a_{24}e_{3}-a_{31}e_{2}+u_{32}e_{1}, \displaystyle\,[e_{2},e_{4}] \displaystyle= \displaystyle\sqrt{2}u_{22}e_{3}-a_{41}e_{2}+(\sqrt{2}v_{2}+u_{42})e_{1}, \displaystyle\,[e_{2},e_{5}] \displaystyle= \displaystyle u_{22}e_{6}-\mathop{{\rm tr}}(A_{2})e_{5}+\sqrt{2}v_{2}e_{4}+y_{22}e_{3}+(y_{21}-a_{51})e_{2}+u_{52}e_{1}, \displaystyle\,[e_{2},e_{6}] \displaystyle= \displaystyle-a_{21}e_{6}-v_{2}e_{3}-a_{61}e_{2}+(-y_{21}+u_{62})e_{1}, \displaystyle\,[e_{2},e_{7}] \displaystyle= \displaystyle-a_{24}e_{7}+\sqrt{2}u_{22}e_{4}+(v_{2}-a_{71})e_{2}+(-y_{22}+u_{72})e_{1}, \displaystyle\,[e_{3},e_{4}] \displaystyle= \displaystyle(\sqrt{2}u_{32}-a_{44})e_{3}+\sqrt{2}v_{3}e_{1}, \displaystyle\,[e_{3},e_{5}] \displaystyle= \displaystyle u_{32}e_{6}-\mathop{{\rm tr}}(A_{3})e_{5}+\sqrt{2}v_{3}e_{4}+(y_{32}-a_{54})e_{3}+y_{31}e_{2}, \displaystyle\,[e_{3},e_{6}] \displaystyle= \displaystyle-a_{31}e_{6}-(v_{3}+a_{64})e_{3}-y_{31}e_{1}, \displaystyle\,[e_{3},e_{7}] \displaystyle= \displaystyle-a_{34}e_{7}+\sqrt{2}u_{32}e_{4}-a_{74}e_{3}+v_{3}e_{2}-y_{32}e_{1}, \displaystyle\,[e_{4},e_{5}] \displaystyle= \displaystyle u_{42}e_{6}-\mathop{{\rm tr}}(A_{4})e_{5}+\sqrt{2}v_{4}e_{4}+(y_{42}-\sqrt{2}u_{52})e_{3}+y_{41}e_{2}-\sqrt{2}v_{5}e_{1}, \displaystyle\,[e_{4},e_{6}] \displaystyle= \displaystyle-a_{41}e_{6}-(v_{4}+\sqrt{2}u_{62})e_{3}-(y_{41}+\sqrt{2}v_{6})e_{1}, \displaystyle\,[e_{4},e_{7}] \displaystyle= \displaystyle-a_{44}e_{7}+\sqrt{2}u_{42}e_{4}-\sqrt{2}u_{72}e_{3}+v_{4}e_{2}-(y_{42}+\sqrt{2}v_{7})e_{1}, \displaystyle\,[e_{5},e_{6}] \displaystyle= \displaystyle-(a_{51}+u_{62})e_{6}+\mathop{{\rm tr}}(A_{6})e_{5}-\sqrt{2}v_{6}e_{4}-(v_{5}+y_{62})e_{3}-y_{51}e_{1}-y_{61}e_{2}, \displaystyle\,[e_{5},e_{7}] \displaystyle= \displaystyle-a_{54}e_{7}-u_{72}e_{6}+\mathop{{\rm tr}}(A_{7})e_{5}+\sqrt{2}(u_{52}-v_{7})e_{4}-y_{72}e_{3}+(v_{5}-y_{71})e_{2}-y_{52}e_{1}, \displaystyle\,[e_{6},e_{7}] \displaystyle= \displaystyle-a_{64}e_{7}+a_{71}e_{6}+\sqrt{2}u_{62}e_{4}+v_{7}e_{3}+v_{6}e_{2}+(-y_{62}+y_{71})e_{1} Viviana del Barco, Instituto de Matemática, Estatística e Computação Científica, Universidade Estadual de Campinas, UNICAMP, Rua Sergio Buarque de Holanda, 651, Cidade Universitaria Zeferino Vaz, 13083-859, Campinas, São Paulo, Brazil. Email address: delbarc@ime.unicamp.br Ana Cristina Ferreira, Centro de Matemática, Universidade do Minho, Campus de Gualtar, 4710-057 Braga, Portugal. Email address: anaferreira@math.uminho.pt Ines Kath, Institut für Mathematik und Informatik, Universität Greifswald, Walther-Rathenau-Str. 47, D-17487 Greifswald, Germany Email address: ines.kath@uni-greifswald.de