J.M. Landsberg and L. Manivel
J.M. Landsberg and L. Manivel
J.M. Landsberg and L. Manivel
Contents
1. Overview 1
2. Construction of complex simple Lie algebras via geometry 2
2.1. Differential geometry, algebraic geometry and representation theory 2
2.2. First algorithm 6
2.3. Second algorithm 9
2.4. Outline of the proofs 10
2.5. Applications, generalizations and related work 11
2.6. Why secant and tangent lines? 12
2.7. Tits correspondences and applications 13
3. Triality and exceptional Lie algebras 15
3.1. Division algebras 15
3.2. Derivations 15
3.3. Triality 16
3.4. Triality and so4,4 : 4-ality 17
3.5. The magic square 19
3.6. Inclusions 20
3.7. Automorphisms, symmetric and trisymmetric spaces 22
3.8. Dual pairs 24
3.9. The quaternionic form 25
4. Series of Lie algebras via knot theory and geometry 27
4.1. From knot theory to the universal Lie algebra 27
4.2. Vogel’s decompositions and Tits correspondences 29
4.3. The exceptional series 29
4.4. Freudenthal geometries 32
4.5. A geometric magic square 33
4.6. The subexceptional series 33
References 35
1. Overview
This article has two purposes. The first is to provide an elementary introduction to papers
[59, 60, 61, 62, 63], related works and their historical context. Sections 2.1–2.3, 2.5–2.6, 3.1–3.5,
4.1–4.3 should be accessible to a general audience of mathematicians. The second is to provide
generalizations, new perspectives, and complements to results in these papers, i.e., things we
thought of after the papers were published. In particular, we mention 2.5, 2.7, 3.4 and 3.6–3.9.
Each section begins with a description of its contents.
Simply put, our goals are to use geometry to solve questions in representation theory and to
use representation theory to solve questions in geometry.
On the geometric side, the objects of interest are homogeneous varieties X = G/P ⊂ PV .
Here G is a complex semi-simple Lie group, V is an irreducible G-module, X is the unique closed
orbit (projectivization of the orbit of a highest weight vector) and P is the stabilizer of a point.
Date: July 2002.
1
2 J.M. LANDSBERG AND L. MANIVEL
For example, let W = Cm and let X = G(k, W ) ⊂ P(Λk W ), the Grassmannian of k-planes
through the origin in W . Here G/P = SL(W )/P , V = Λk W . We are more generally interested
in the geometry of orbit closures in PV .
Basic questions one can ask about a variety are its dimension, its degree, and more generally
its Hilbert function. We may ask the same questions for varieties associated to X. For example
the degrees of the dual varieties of Grassmannians are still unknown in general (see [67, 26, 78]).
Other types of problems include recognition questions. For example, given a variety with
certain geometric properties, are those properties enough to characterize the variety? For an
example of this, see Zak’s theorem on Severi varieties below in §2.6. For another example, Hwang
and Mok characterize rational homogeneous varieties by the variety of tangent directions to
minimal degree rational curves passing through a general point, see [44] for an overview. Some
results along those lines are described in §2.5 below. We also mention the LeBrun-Salamon
conjecture [68, 69] which states that any Fano variety equipped with a holomorphic contact
structure must be the closed orbit in the adjoint representation Xad ⊂ Pg for a complex simple
Lie algebra. In this context also see [8, 52].
On the representation theory side, the basic objects are g, a complex semi-simple Lie algebra
and V , an irreducible g-module (e.g., g = sl(W ), V = Λk W ). Problems include the classification
of orbit closures in PV , to construct explicit models for the group action, to geometrically inter-
pret the decomposition of V ⊗ k into irreducible g-modules. We discuss these classical questions
below, primarily for algebras occuring in “series”.
Vassiliev theory points to the need for defining objects more general than Lie algebras. We
have nothing to add about this subject at the moment, but the results of [62, 63] were partly
inspired by work of Deligne [24] and Vogel [82, 83] in this direction.
For the mystically inclined, there are many strange formulas related to the exceptional groups.
We present some such formulas in §4.3, 4.6 below. Proctor and Gelfand-Zelevinski filled in
“holes” in the classical formulas for the ospn series using the non-reductive odd symplectic
groups. Our formulas led us to exceptional analogues of the odd symplectic groups. These
analogues are currently under investigation (see [65]).
When not otherwise specified, we use the ordering of roots as in [9].
We now turn to details. We begin with some observations that lead to interesting rational
maps of projective spaces.
X
X
Tx X Tx X
IP T x X IP T x X
Y Y
X,x X,x
In an attempt to get global information from the asymptotic directions, restrict to the case
where x ∈ X is a general point
Question 2.2. How much of the geometry of X can be recovered from YX,x ⊂ PTx X?
Answer: Usually not much. For example, if X is a smooth hypersurface, then Y is always a
smooth quadric hypersurface, i.e., all smooth hypersurfaces look the same to second order.
The set of asymptotic directions is the zero set of a system of quadratic equations generically
of dimension equal to the codimension of X (unless the codimension of X is large, in which case
it is generically the complete system of quadrics). If Y is sufficiently pathological one might
hope to recover important information about X.
Consider the Segre variety, Seg(Pk × Pl ) ⊂ P(Ck+1 ⊗ Cl+1 ) of rank one matrices in the space
of all (k + 1) × (l + 1) matrices. A short calculation shows YSeg,x = Pk−1 ⊔ Pl−1 , the disjoint
union of a Pk−1 with a Pl−1 . (Note that the codimension is sufficiently large here that one would
expect Y to be empty based on dimension considerations.)
Griffiths and Harris [37] conjectured that if Z ⊂ P8 is a variety of dimension 4 and z ∈ Z a
general point, if YZ,z ⊂ PTz Z is the disjoint union of two lines, then Z = Seg(P2 × P2 ) ⊂ P8 .
4 J.M. LANDSBERG AND L. MANIVEL
Theorem 2.3. [58] Let k, l > 1, let Z k+l ⊂ PN be a variety, and let z ∈ Z be a general point.
If YZ,z = Pk−1 ⊔ Pl−1 , then Z = Seg(Pk × Pl ).
Moreover, the analogous rigidity is true for varieties having the same asymptotic directions
as X = G(2, m) ⊂ P(Λ2 Cm ), the Grassmannian of two-planes in Cm in its Plucker embedding,
with m ≥ 6, and the complex Cayley plane E6 /P6 = OP2C ⊂ P26 [58]. We recently prove similar
rigidity results for homogeneous Legendrian varieties [64].
The rigidity results hold in the C ∞ category if one replaces the word “general” by the word
“every”.
If one adds the global assumption that Z is smooth, the results also hold for X = G(2, 5) and
X = S5 ⊂ P15 , the spinor variety using different methods.
We are unaware of any second order rigidity results for nonhomogeneous varieties. Thus,
purely from the most naive differential geometry, one already encounters homogeneous varieties
as examples of the most rigid projective varieties. In fact, so far we have just encountered the
most homogeneous ones, the ones admitting Hermitian symmetric metrics. For more on this
and the rigidity of other homogenous varieties see §2.5.
We define a homogeneous variety X = G/P ⊂ PV to be minuscule if G is simple, X admits
a Hermitian symmetric metric and X is in its minimal homogeneous embedding. X is said to
be generalized minuscule if it is homogeneously embedded and admits a Hermitian symmetric
metric.
Remark 2.4. If X ⊂ PV is a variety cut out by quadratic polynomials, then the asymptotic
directions Y ⊂ PTx X are actually the tangent directions to lines (linearly embedded P1 ’s) on
X. Rational homogeneous varieties are cut out by quadratic equations, in fact if X ⊂ PVλ , then
the ideal is generated by V2λ ⊥ ⊂ S 2 Vλ∗ .
2.1.2. Lie groups and homogeneous varieties. Let G be a complex (semi)-simple Lie group and
let V be an irreducible G-module. Then there exists a unique closed orbit X = G/P ⊂ PV .
Examples:
1. G = SL(n, C), the group preserving det ∈ Λn Cn∗ , V = Λk Cn , X = G(k, n), the Grass-
mannian of k-planes through the origin in Cn .
2. G = SO(n, Q), the group preserving a nondegenerate Q ∈ S 2 Cn∗ , V = Λk Cn , X =
GQ (k, n), the Grassmannian of Q-isotropic k-planes through the origin in Cn .
3. G = Sp(n, ω), the group preserving a nondegenerate ω ∈ Λ2 Cn∗ , V = Λk Cn /(Λk−2 Cn ∧ ω),
X = Gω (k, n), the Grassmannian of ω-isotropic k-planes through the origin in Cn . Here n is
usually required to be even (but see [73, 36], and §4.3).
Since linear algebra is easier than global geometry, we work with g = TId G, the associated
Lie algebra.
About a century ago, Killing and Cartan classified complex simple Lie algebras. Thanks to
Coxeter and Dynkin, the classification can be expressed pictorially. (See [40] for a wonderful
account of their work and the history surrounding it.)
◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
@◦
An Dn
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ E6 ◦ E7 ◦ E8
If a diagram has symmetry, we are allowed to fold it along the symmetry and place an arrow
pointing away from the hinge to get a new one:
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 5
◦ ◦ ◦ ◦ ◦> ◦ ◦ ◦ ◦ ◦ ◦<◦
Bn Cn
◦ ◦> ◦ ◦ ◦>◦
F4 G2
For example, Cn is the fold of A2n−1 . Given a semi-simple Lie algebra g, we let D(g) denote its
Dynkin diagram.
◦
◦ ◦ ◦• ◦ ◦ ◦ ◦ ◦ •◦ ◦ ◦
@◦
G(3, 7) GQ (3, 14)
◦• ◦ ◦ ◦ ◦ ◦
Pn
Example 2.6.
◦ ◦
◦ ◦ ◦ •◦ ◦ −→ ◦ ◦ • •
@◦ @◦
X = GQ (4, 12) Y = Seg(P3 × G(2, 4))
6 J.M. LANDSBERG AND L. MANIVEL
The H-modules hY i are studied in [49] where they are called type-I θ-representations.
The embedding of Y is minimal iff the diagram is simply laced. In the case of a double edge,
one takes the quadratic Veronese embedding, for a triple edge, one takes the cubic Veronese (see
the algorithms below).
Remark 2.7. In [59] we explicitly determine Y in the case of short roots and in fact arbitrary
parabolics. We also give geometric models. In the case of X = G/Pα with α short, Y is the
union of exactly two G-orbits, and the closed orbit has codimension one in Y .
If Y is still complicated, one can continue, studying the asymptotic directions of Y at a point.
Eventually one gets (Segre products of re-embeddings of) CP1 , the one homogeneous space we
all pretend to understand.
We describe our discovery that one can reverse this infinitesimalization procedure below. First
we need to review some elementary algebraic geometry.
2.1.3. Embeddings of projective space. Recall the Veronese embeddings of projective space:
vd : PV → P(S d V )
[w] 7→ [wd ].
Dually, let P0 , ..., PN be a basis of S d V ∗ . The map is
[w] 7→ [P0 (w), ..., PN (w)].
A remarkable fact is that all maps between projective spaces PV → PW are obtained by
projecting a Veronese re-embedding.
If we project to P(S d V /L) where L⊥ ⊂ S d V ∗ has basis Q0 , ..., Qm , the map is
f : PV → P(S d V /L).
[w] 7→ [Q0 (w), ..., Qm (w)]
The image of f is smooth and isomorphic to PV iff no secant line of vd (PV ) intersects L.
L
vd (PV )
P(S d V /L)
In the following algorithm we will actually be interested in images that get squashed in the
projection so that they are no longer isomorphic to Pn (but are still smooth).
2.2. First algorithm. We are about to describe an algorithm, which you might like to think
of as a game. Starting with P1 as initial input we build some new algebraic varieties subject to
certain rules. The game has rounds, and in each new round, we are allowed to use the outputs
from previous rounds as new inputs.
We fix some notation.
Let X ⊂ PV , Y ⊂ PW , be varieties.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 7
Seg(X × Y ) ⊂ P(V ⊗ W ) is the Segre product, ([x], [y]) 7→ [x ⊗ y]. Sometimes we just write
X × Y with the Segre product being understood.
σk (X) ⊂ PV is the union of all secant Pk−1 ’s to X. We let σ(X) = σ2 (X).
T (X) ⊂ G(2, V ) ⊂ P(Λ2 PV ) denotes the union of all tangent lines to V . (Recall G(k, V ) is
also the space of Pk−1 ’s in PV .)
X̂ ⊂ V is the cone over X and hXi ⊆ V denotes the linear span of X̂.
Given Y ⊂ PW , we let Ik (Y ) ⊂ S k W ∗ denote the component of the ideal in degree k. In the
flowchart below we slightly abuse notation by letting Ik (Y ) denote a set of generators of Ik (Y ).
Similarly, we let T ∗ be shorthand for a basis of T ∗ .
First Algorithm
Initial input
P1 ⊂ P1 - RESERVOIR
Is hT (Y )i = Λ2 T ?
@
@
NO @ YES
@
@
R
@
B Let d be the smallest integer such that
B TRASH
B σd (Y ) = PT = Pn−1
B Linearly embed PT ⊂ Pn as the hyperplane {l = 0}.
B
Consider the rational map
ϕ : Pn −→ PN ⊂ P(S d Cn+1 )
||
d d−1 ∗
Phl , l T , ld−2 I2 (Y ), . . . , Id (σd−1 (Y ))i.
Let X = ϕ(Pn ) ⊂ PN . Call X an OUTPUT.
For the first run through the algorithm, the admissible varieties and their outputs are as
follows
Y ⊆ Pn−1 Xn ⊆ PN
P1 ⊆ P1 P2 ⊆ P2
P1 × P1 ⊂ P3 Q4 ⊂ P5
v2 (P1 ) ⊂ P2 Q3 ⊂ P4 .
Here G(k, l) denotes the Grassmannian of k-planes in Cl and Gω (k, 2k) denotes the Grassman-
nian of Lagrangian k-planes for a given symplectic form.
Question 2.8. What comes out? Since the algorithm goes on forever, is it even possible to
answer this question?
Proposition 2.9. [60] The algorithm is effective. A priori, r, dj ≤ 2 and the algorithm stabilizes
after six rounds.
So at least our question is reasonable. Now for the answer:
Theorem 2.10. [60] OUTPUTS = MINUSCULE VARIETIES.
Corollary 2.1. [60] A new proof of the classification of the compact Hermitian symmetric spaces
without any reference to Lie groups.
We have the following stable round:
Y ⊆ Pn−1 X n ⊆ PN
Pn−1 ⊆ P n−1 Pn ⊆ Pn
m+1
v2 (Pm−1 ) ⊂ P( 2 )−1 Gω (m, 2m) ⊂ PCm+1 −1
k+l
Pk−1 × Pl−1 ⊂ Pkl−1 G(k, k + l) ⊂ P( k )−1
Q2m−2 ⊂ P2m−1 Q2m ⊂ P2m+1
Q 2m−1 ⊂ P 2m Q 2m+1 ⊂ P2m+2
m
G(2, m) ⊂ P( 2 )−1
m−1
Sm ⊂ P2 −1 .
1 2m+2
Here Cm+1 = m+2 m+1 is the (m + 1)-st Catalan number. The spinor variety Sm of Dm
consists of one family of maximal isotropic subspaces of C2m endowed with a nondegenerate
quadratic form and embedded in the projectivization of one of the two half-spin representations.
The most interesting (but terminal) path is:
Seg(P1 × P2 ) = A1 /P1 × A2 /P2
↓
G(2, 5) = A4 /P2
↓
S5 = D5 /P5
↓
OP2C = E6 /P6
↓
Gw (O3 , O6 ) = E7 /P7 .
Here OP2C = E6(−14) is the complexification of OP2 = F4 /Spin9 , the celebrated Cayley plane
discovered by Chevalley. As a topological space, OP2 is built out of three cells of dimension 0, 8
and 16. It is F4(−20) in the notation of Tits (see e.g. [70]). The notation Gw (O3 , O6 ) is discussed
in §4.6.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 9
2.3. Second algorithm. Now that was fun, but it was a shame to throw away some of our
favorite varieties like v3 (P1 ) and Seg(P1 × P1 × P1 ). Let’s revise the algorithm slightly and have
just one consolation round. Given a smooth variety Y ⊂ PV , we let τ (Y ) ⊂ PV denote the
union of all points on all embedded tangent lines to Y .
MINUSCULE VARIETIES
Is codim hT (Y )i = 1 ?
@
@
NO @ YES
@
@R
@
B Linearly embed PT1 ⊂{l=0} Pn−1 ⊂{m=0} Pn .
B TRASH Consider the rational map
B
B ϕ : Pn −→ PN ⊂ P(S 4 Cn+1 )
B
||
Phl , l m, l T1 , l I2 (Y ), l mT1∗ − lI3 (τ (Y )sing ), l2 m2 − I4 (τ (Y ))i.
4 3 3 ∗ 2 2
?
OUTPUTS
Y ⊂ Pn−2 g
1
v3 (P ) ⊂ P3 g2
P1 × Qm−4 ⊂ P2m−5 som
Gω (3, 6) ⊂ P13 f4
G(3, 6) ⊂ P19 e6
S6 ⊂ P31 e7
Gw (O3 , O6 ) ⊂ P55 e8 .
The two exceptional (i.e., non-fundamental) cases are
Note that the algorithm works in these two cases, we just didn’t have the varieties in our
reservoir. (In the case of sp2m , it may have been there and just difficult to see.)
2.4. Outline of the proofs. The proofs have three ingredients: differential invariants, local
Lie algebras, and relating Casimirs to geometry.
2.4.1. Differential invariants. Given a variety X ⊂ PV and x ∈ X, we can recover X from its
Taylor series at x. The projective differential invariants comprise a series of tensors encoding
the geometric (i.e., invariant under GL(V )) information in the Taylor series. We prove a priori
facts about the differential invariants of any putative minuscule variety or fundamental adjoint
variety.
2.4.2. Local Lie algebras. A local Lie algebra is a graded vector space
g−1 ⊕ g0 ⊕ g1
equipped with a bracket for which g0 is a Lie algebra. The bracket must respect the grading
but the Jacobi identity need not hold. If the Jacobi identity fails, one can construct a (unique)
Z-graded Lie algebra from the local Lie algebra. The traditional way to do this (see, e.g., [48])
is to take the free algebra generated by the brackets and then mod out by the relations. Note
that if one does that, one has no idea how many factors (if any at all) one will be adding on to
obtain the final result.
The data Y = H/Q ⊆ PT1 furnishes (up to scale) a local Lie algebra with g0 = h ⊕ C, g1 = T1 ,
g−1 = T1∗ . Here the action of C is as a scalar times the identity and we do not initially specify
the scalar.
Since this does give rise to a unique Z-graded Lie algebra, (in particular, a Lie algebra equipped
with a representation V supported on one fundamental weight), we can study the resulting
homogeneous variety X ′ = G/P ⊂ PV , and calculate its differential invariants.
Note that if X ′ is minuscule, g = g−1 ⊕ g0 ⊕ g1 and if X ′ is adjoint, g = g−2 ⊕ g−1 ⊕ g0 ⊕ g1 ⊕ g2
with g±2 = C.
2.4.3. Compare X ′ with the constructed variety X. We compare the differential invariants of X
and X ′ and show that they are the same. The key point in the minuscule case is that a local
Lie algebra is already a Lie algebra iff the Jacobi identity holds. We show this is the case iff
Λ2 g1 is an eigenspace for the Casimir operator of h. We then connect this to the geometry by
showing that hT (Y )i is a Casimir eigenspace!
For the adjoint case the idea is similar: one shows that one can get away with a one-
dimensional correction.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 11
particular role, but it must be present at every point so in particular it determines a family of
distributions on X.
Hong’s theorem is a variant of an earlier rigidity result of Hwang and Mok [45], where a
geometric structure may be understood in the analogous way.
Theorem 2.14. A Fano manifold with a geometric structure modeled after a compact irreducible
Hermitian symmetric space S of rank ≥ 2 is biholomorphic to S and the geometric structure is
locally isomorphic to the standard geometric structure on S.
The work of Hwang and Mok (also see [44]) relies on studying an intrinsic analog of the set of
asymptotic directions. Namely a Fano manifold X is uniruled by rational curves, and, fixing a
reference line bundle, one has a subvariety Yx ⊂ PTx X of tangent directions to minimal degree
rational curves. They also study deformation rigidity of homogeneous varieties with Picard
number one. It is worth remarking that in the case of Hermitian symmetric spaces [46], a key
fact used in their proof is that T (Yx ) ⊂ PTx X is linearly nondegenerate, i.e., the same condition
we use in the minuscule algorithm.
2.5.4. Normal forms for singularities. In [4], Arnold classified the simple singularities. Like
many interesting things in life they are in correspondence with Dynkin diagrams, in fact just
the simply laced ones. He also gave normal forms in a minimal set of variables (two). It was
also known (but evidently unpublished) that the simple singularities can be realized as degree
three hypersurface singularities if one allows the number of variables to grow with the Milnor
number. Holweck [41] has found a nice realization of these hypersurface singularities using a
theorem of Knop [54] and the construction in §2.3.
2.6. Why secant and tangent lines? The idea that secant and tangent lines should so
strongly control the geometry of homogeneous varieties was inspired by Zak’s theorem on Severi
varieties.
Theorem 2.15 (Zak’s theorems on linear normality and Severi varieties). Let X n ⊂ Pn+a be a
smooth variety, not contained in a hyperplane and such that σ(X) 6= Pn+a Then
i. a ≥ n2 + 2.
ii. If a = n2 + 2 then X is one of v2 (P2 ) ⊂ P5 , Seg(P2 × P2 ) ⊂ P8 , G(2, 6) ⊂ P14 , OP2C ⊂ P26 .
The four critical varieties are called the Severi varieties after Severi who proved the n = 2
case of the theorem.
It is not known if there is a bound on the secant defect δσ (X) := 2n + 1 − dim σ(X) for
smooth subvarieties of projective space with degenerate secant varieties. On the other hand,
Zak established an upper bound on the codimension of a smooth variety of a given secant defect.
He then went on to classify the varieties achieving this bound, which he calls the Scorza varieties.
They are all closed orbits G/P ⊂ PV , namely v2 (Pn ), Seg(Pn × Pn ), G(2, n) and OP2C .
Using Zak’s result, Ein and Shepherd-Baron proved these four varieties also classify quadro-
quadro Cremona transformations [29]. There are numerous other characterization problems
where the answer is the Severi varieties.
Zak’s proofs of his theorems rely on looking at entry loci. Namely let X ⊂ PV be a variety
and let y ∈ PV \X. Define the entry locus of y to be
Σy := {x ∈ X | ∃z ∈ X, y ∈ P1xz }.
Here P1xz denotes the projective line spanned by x and z. Zak shows that for a Severi variety, the
entry locus of a general point of σ(X) is a quadric hypersurface in X. He then goes on to show
that a variety so uniruled by quadric hypersurfaces and satisfying the dimension requirements
must satisfy further dimension restrictions and eventually must be one of the four Severi varieties.
More precisely, he shows that each Severi variety is the image of a rational map of a projective
space. Our algorithms generalize his construction of the Severi varieties.
Recently Chaput [15] has shown that such a uniruling by quadrics immediately implies ho-
mogeneity, which gives a quicker proof of Zak’s theorem.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 13
The rigidity results for the Severi varieties other than v2 (P2 ), are an outgrowth of a different
proof of Zak’s theorem, where one first shows that any putative Severi variety infinitesimally
looks like an actual Severi variety to second order at a general point, and then uses the rigidity
to finish the proof. (v2 (P2 ) requires special treatment as its tangential variety is nondegenerate.)
See [57] and [58].
Many other important problems are related to the geometry of secant varieties. Consider the
classical Waring problem: given a generic homogeneous polynomial of degree d in n+1 variables,
what is the minimal k such that the polynomial may be written as a sum of k d-th powers?
Phrased geometrically, the problem is to find the minimal k such that σk (vd (Pn )) = P(S d Cn+1 ).
This problem was solved by Alexander and Hirschowitz [1]. Generalizations and variants are
still open, see [13, 47, 75] for recent progress.
2.7. Tits correspondences and applications. In §2.7.1 we describe a construction, Tits cor-
respondences [81], to determine the homogeneous (in the sense described below) unirulings of a
rational homogeneous variety X ⊂ PV . We explain how to use Tits correspondences to explicitly
construct the homogeneous vector bundles over X in §2.7.2. In §2.7.3 we relate Tits correspon-
dences to other orbit closures in PV and systematize Kempf’s method for desingularizing orbit
closures. Tits correspondences led us to the decomposition formulas described in §4.2 and we
believe they will have more applications in the future.
2.7.1. Tits correspondences. Tits associated to any simple group a full set of geometries, encoded
by the parabolic subgroups and their relative positions. This culminated in the definition of
buildings, which has since known formidable developments. The Tits correspondences we now
describe come from this perspective.
Let G be a simple Lie group, let S, S ′ be two subsets of simple roots of G. Consider the
diagram
G/PS∪S ′
π ւ ց π′
X = G/PS X ′ = G/PS ′
Let x′ ∈ X ′ and consider Y := π(π ′−1 (x′ )) ⊂ X. We call Y = Yx′ the Tits transform of x′ . The
variety X is covered by such varieties Y . Tits shows that Y = H/Q where D(H) = D(G)\(S\S ′ ),
and Q ⊂ H is the parabolic subgroup corresponding to S ′ \S. We call such subvarieties Y of X,
G-homogeneous subvarieties.
2.7.2. Unirulings.
Unirulings by lines. Let X = G/Pα ⊂ PV be a homogeneous variety. In Theorem 2.5 above we
determined the parameter space of lines through a point of X. Here we determine the space
of all lines on X. Let F1 (X)G ⊂ G(2, V ) ⊂ PΛ2 V denote the variety parametrizing the G-
homogeneous P1 ’s on X. Then F1 (X)G = G/PS where S is the set of simple roots adjacent to α
in D(g). If P is not maximal, there is one family of G-homogeneous P1 ’s for each marked node.
If α is not a short root, all lines are G-homogeneous.
Unirulings by Pk ’s. The diagram for Pk is (ak , ω1 ). Here there may be several G-homogeneous
families of Pk ’s on X = G/Pα , each arising from a subdiagram of D(g, ω) isomorphic to a
D(ak , ω1 ). If α is not short, then all Pk ’s on X are G-homogeneous.
Example 2.16. The largest linear space on En /P1 is a Pn−1 , via the chain terminating with
αn , so En /P1 is maximally uniruled by Pn−1 ’s. There is a second chain terminating with α2 ,
so En /P1 is also maximally uniruled by P4 ’s. (The unirulings by the P4 ’s are maximal in the
sense that none of the P4 ’s of the uniruling are contained in any P5 on En /P1 .) The varieties
parametrizing these two unirulings of En /P1 are respectively En /P2 and En /P5 .
α1 α1
• • • • • • • • • ◦∗ ◦ ◦
∗
◦ •
14 J.M. LANDSBERG AND L. MANIVEL
Unirulings by quadrics. Here we look for subdiagrams isomorphic to the diagram of a quadric
hypersurface, i.e. D(son , ω1 ). Note that there are at most two possible such subdiagrams except
for the case of (g, V ) = (so8 , Vω2 ) where there are three isomorphic ones.
Example 2.17. (g, V ) = (so2m , Vωk ) has two such unirulings, by Q4 ’s parametrized by Dm /Pαk−2 ,αk+2
and by Q2(m−k−1) ’s parametrized by Dm /Pαk−1 .
◦ ◦
∗◦ ◦ •◦ ◦ ∗◦ ◦ ◦∗ ◦• ◦ ◦
αk @◦ αk @◦
Example 2.18 (The Severi varieties). The G-homogeneous quadrics uniruling the Severi vari-
eties are parametrized by the same Severi variety in the dual projective space.
◦ •◦ ◦ ∗◦ ◦ •◦ ◦ ◦ ◦ ◦∗
G(2, 6) ◦ OP2C
Unirulings by quadrics give rise to subvarieties of σ(X) as follows. The union of the spans
of the quadrics produce a subvariety of σ(X) whose entry loci contain the quadrics (and in all
examples we know, are equal to the quadrics). It also appears that, when X is homogeneous,
these give the subvarieties of σ(X) with maximal entry loci. As explained below, these orbit
closures admit uniform desingularizations by Kempf’s method.
2.7.3. Homogeneous vector bundles. The G-homogeneous vector bundles over rational homoge-
neous varieties Z = G/Q are defined by Q-modules. (We change notation to reserve G/P for a
different role below.) If W is a Q-module, one obtains the vector bundle EW := G×Q W → G/Q
where (gp, w) ≃ (g, pw) for all p ∈ Q. Note that D(f) is D(g) with the nodes corresponding to
Q deleted.
Some homogeneous vector bundles over rational homogeneous varieties Z = G/Q can be
understood in terms of Tits fibrations. Let X = G/P ⊂ PV . For each z ∈ Z we obtain a variety
Yz ⊂ X and thus a linear space hYz i ⊂ V . As we vary z, we obtain a vector bundle E → Z
whose fibers are the hYz i’s. In particular, E is a subbundle of the trivial bundle V ⊗ OZ → Z.
Recipe for Kempf ’s method. Let V be a G-module with closed orbit X ⊂ PV . Let Z = G/Q be
a rational homogeneous variety of G, defining a Tits correspondence with X.
Consider the G-variety, which often appears to be an orbit closure
O = ∪z∈Z hYz i ⊂ PV.
If O 6= PV , then O is a singular variety. It admits a desingularization by PE, where E → Z is
the vector bundle constructed above as the natural map PE → Z is generically one to one.
Note that the G-variety τ (X) is of course desingularized by PT X, which is a special case of
the above with Q = P .
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 15
We begin with a review of division algebras via the Cayley-Dickson process in §3.1 and
explicitly describe the derivations of O in §3.2 for later use. In §3.3 we review the triality
principle for the octonions and its extension to all structurable algebras. We pause for a detour
in §3.4, exploring the triality model for so4,4 in detail, which ends up being a 4-ality. We
review various constructions of the magic square in §3.5. We present a new result regarding the
compatiblity of the Cayley-Dickson process and inclusions of triality algebras in §3.6, which at
first glance appears to be trivial, but whose “obvious” generalization to structurable algebras
is false. In §3.7 we study automorphisms related to the algebras in the magic square and the
resulting magic squares of symmetric and trisymmetric spaces. In §3.8 we use our result in §3.6
to show how some towers of dual pairs naturally occur in series. In §3.9 we discuss quaternionic
symmetric spaces, their relation to adjoint varieties and the conjecture of LeBrun and Salamon.
3.1. Division algebras. A classical theorem of Hurwitz, published in 1898, asserts that there
are only four normed division algebras over the field of real numbers : R itself, C, the quaternions
H, and the octonions O. A division algebra is an algebra without zero divisor and a normed
division algebra is a division algebra endowed with a norm such that the norm of a product is
the product of the norms.
Quaternions were discovered by Hamilton in 1843, and octonions shortly afterwards by Graves
and Cayley. A nice way to define these algebras is to imitate the definition of complex numbers
by a pair of real numbers. Let A be a real algebra endowed with a conjugation x 7→ x, such that
x + x and xx = xx are always real numbers (more precisely scalar multiples of the unit element),
the later being positive when x is nonzero. If A is alternative, i.e., any subalgebra generated by
two elements is associative [5], then A is a normed algebra for the norm kxk2 = xx.
One can then define a new algebra with conjugation B = A ⊕ A by letting (x, y)(z, t) =
(xz − ty, xt + zy) with conjugation (x, y) = (x, −y). This is the Cayley-Dickson process. The
new algebra B will be alternative, hence a normed algebra, iff A is associative.
In particular, A = C gives the quaternion algebra B = H, which is associative, so the Cayley-
Dickson process can be applied once more and A = H gives the octonion algebra B = O. This
algebra is no longer associative (although it is alternative, hence normed), and the process then
fails to produce new normed algebras.
A useful variant of the Cayley-Dickson process is obtained by changing a sign in the formula
for the product in B, letting (x, y)(z, t) = (xz + ty, xt + zy). The resulting algebra B̃ is called
split. It is no longer normed, but endowed with a nondegenerate quadratic form of signature
(a, a) compatible with the product. (Here a denotes the dimension of A.) In particular, it is a
composition algebra. The algebra C̃ of split complex numbers is the algebra R ⊕ R with termwise
multiplication, while the split quaternion algebra H̃ is equivalent to M2 (R).
3.2. Derivations. Any automorphism of H is inner, so that any derivation of H is of the form
La − Ra for some imaginary quaternion a ∈ ImH, where La and Ra denote the operators of left
and right mulplication by a. In particular DerH = ImH = so3 .
The derivation algebra of O is the compact Lie algebra g2 . The algebra DerH = so3 does not
imbed in DerO in a canonical way, but there is one prefered embedding for each decomposition
O = H ⊕ eH, where the product in O is deduced from that in H through the Cayley-Dickson
process. Indeed, for φ ∈ DerH, the endomorphism φ̃ of O defined by φ̃(x + ey) = φ(x) + eφ(y) is
a derivation. Note that the subalgebra Der(O, H) of the derivations of O stabilizing H is strictly
greater than DerH. It contains, for each imaginary quaternion h, the map ψh (x + ey) = e(yh).
This gives another copy of so3 in Der(O, H) = so3 × so3 , and the restriction of these derivations
to eH gives the full so4 = so3 × so3 . Explicitly, choose a standard basis e0 = 1, e1 , e2 , e3 = e1 e2
of H, and let e4 = e, e5 = ee1 , e6 = ee2 , e7 = ee3 . Using this basis of O, we obtain a matrix
16 J.M. LANDSBERG AND L. MANIVEL
representation of g2 = Der(O):
0 0 0 0 0 0 0 0
0 0 −α2 −α3 −α4 −α5 −α6 −α7
0 α2
0 −β 3 −β 4 −β 5 −β 6 −β 7
0 α3 β3 0 α6 + β5 −α5 − β4 −α4 + β7 α7 − β6
g2 =
.
0 α4 β4 −α6 − β5 0 −γ5 −γ6 −γ7
0 α5 β5
α5 + β4 γ5 0 −α2 + γ7 −α 3 − γ6
0 α6 β6 α4 − β7 γ6 α2 − γ7 0 β3 + γ5
0 α7 β7 −α7 + β6 γ7 α3 + γ6 −β3 − γ5 0
3.3. Triality.
3.3.1. Cartan’s triality. The Dynkin diagram of type D4 is the only one with a threefold sym-
metry.
D4 ◦ ◦
J
J
J
◦
3.3.2. Triality algebras. Allison and Faulkner [3], following work of Allison [2] and Kantor [50]
define the following vast generalization of the triality principle. Let A be any algebra with unit
and involution a 7→ a over a ring of scalars. Let ImA = {a ∈ A, a = −a}. Define the triality
algebra of A
t(A) := {T = (T1 , T2 , T3 ) ∈ gl(A) ⊕ 3 | T1 (ab) = T2 (a)b + aT3 (b) ∀a, b ∈ A},
with T1 (a) = T1 (a). This is a Lie subalgebra of gl(A) ⊕ 3 . In [3] it is call the set of partially
related Lie triples of A. If 3 is invertible in the ring of scalars, they provide (Corollary 3.5) a
general description of t(A), which in the case where A is alternative gives
t(A) = {(D + Ls − Rt , D + Lt − Rr , D + Lr − Rs ) | D ∈ Der(A), r, s, t ∈ Im(A), r + s + t = 0}
= Der(A) ⊕ Im(A) ⊕ 2 .
3.4. Triality and so4,4 : 4-ality. An interesting version of triality holds for so4,4 , the split
version of so8 . We use a model is related to the fourfold symmetry of the extended Dynkin
diagram D̃4 .
◦ ◦
@
@
@
D̃4 ◦
@
@@◦
◦
18 J.M. LANDSBERG AND L. MANIVEL
3.5.1. Jordan algebras. The relation between exceptional Lie groups and normed division alge-
bras, especially the octonion algebra, was first noticed by E. Cartan, at least as early as 1908 (see
the encyclopedia article for the French version of the Mathematische Wissenschaften). Cartan
observed that the automorphism group of O is the compact group G2 . In 1950, Chevalley and
Schafer obtained the compact group F4 as the automorphism group of the Jordan algebra
(
r1 x3 x2
)
J3 (O) := x3 r2 x1 , ri ∈ R, xj ∈ O .
x2 x1 r3
This is an algebra with product A.B = 21 (AB + BA). It is a commutative but nonassociative
algebra, in which the characteristic identity A2 (AB) = A(A2 B) of Jordan algebras holds. By
remarks in [12], it appears Cartan was aware of this construction at least as early as 1939.
Jordan algebras were introduced in the 1930’s by Jordan, von Neumann and Wigner as a
natural mathematical framework for quantum theory. Any associative algebra (over a field of
characteristic not equal to 2) can be considered as a Jordan algebra for the symmetrized matrix
product a.b = 21 (ab + ba). A Jordan algebra is said to be exceptional if it cannot be embedded
as a Jordan subalgebra of an associative algebra. The algebra J3 (O) is an exceptional Jordan
algebra. Moreover, if A is an algebra such that Jn (A), with its symmetrized product, is a Jordan
algebra, then A must be alternative if n ≥ 3, and associative if n ≥ 4.
The closed orbits X ⊂ Jn (A), with A the complexification of a division algebra are exactly
the Scorza varieties while the Severi varieties are exactly the varieties X ⊂ J3 (A), see [85]. The
close relationship between Scorza varieties and Jordan algebras, which is not explained in [85],
is investigated in [16].
3.5.2. The Freudenthal-Tits construction. The discovery that F4 = AutJ3 (O) led several au-
thors (including Freudenthal, Tits, Vinberg, Jacobson, Springer) to study the relations between
exceptional groups and Jordan algebras. Freudenthal and Tits obtained a unified construction
of all compact exceptional Lie algebras in the following way: take two normed division algebras
A and B, and let
g(A, B) := DerA ⊕ (ImA ⊗ J3 (B)0 ) ⊕ DerJ3 (B),
where J3 (B)0 ⊂ J3 (B) is the hyperplane of traceless matrices. One can then define a Lie algebra
structure on g(A, B), with DerA ⊕ DerJ3 (B) as a Lie subalgebra acting on ImA ⊗ J3 (B)0 in a
natural way [79]. The result of this construction is the Freudenthal-Tits magic square of Lie
algebras:
R C H O
R su2 su3 sp6 f4
C su3 su3×su3 su6 e6
H sp6 su6 so12 e7
O f4 e6 e7 e8
Let B = 0 denote the zero algebra, so Jn (0) ≃ Cn is the diagonal matrices. Let ∆ be
something slightly smaller, so that Jn (∆) ⊂ Jn (0) consists of the scalar matrices. With these
conventions, we obtain a magic rectangle:
∆ 0 R C H O
R 0 0 su2 su3 sp6 f4
C 0 0 su3 su3×su3 su6 e6
H su2 (su2 ) ⊕ 3 sp6 su6 so12 e7
O g2 so8 f4 e6 e7 e8
20 J.M. LANDSBERG AND L. MANIVEL
3.5.3. The Allison construction. While it is impressive that the vector space g(A, B) is a Lie
algebra, the symmetry of the square appears to be miraculous. Vinberg, in 1966, obtained
a symmetric construction (see [72] for an exposition). We present a variant of the Vinberg
construction first discovered by Allison [2], then rediscovered independently in more geometric
form by Dadok and Harvey [23] (who, after discovering Allison’s work decided not to publish
their manuscript), Barton and Sudbery [7], and ourselves [62].
Allison shows that given an arbitrary structurable algebra A, one can put a Lie algebra
structure on
t(A) ⊕ A ⊕ 3
Allison observes moreover that the algebra A = A ⊗ B with A, B division algebras is structurable
and his construction applied to them yields the triality model, see [2, 3].
Using the Allison construction, the magic square may be described as follows:
g(A, B) = t(A) × t(B) ⊕ (A1 ⊗ B1 ) ⊕ (A2 ⊗ B2 ) ⊕ (A3 ⊗ B3 ).
This bracket is defined so that t(A) × t(B) is a Lie subalgebra, acting on each Ai ⊗ Bi in the
natural way. The bracket of an element a1 ⊗ b1 ∈ A1 ⊗ B1 with a2 ⊗ b2 ∈ A2 ⊗ B2 is simply
a1 a2 ⊗ b1 b2 , considered as an element of A3 ⊗ B3 . This is the general rule for taking the bracket
of an element of Ai ⊗ Bi with one of Aj ⊗ Bj , although there are some slight twists whose details
can be found in [3] or [62]. Finally, the bracket of two elements in Ai ⊗ Bi is defined by the
quadratic forms on Ai , Bi and the natural maps Ψi : ∧2 Ai → t(A) obtained by dualizing the
action of t(A) on Ai , which can be seen as a map t(A) → ∧2 Ai . (Note that t(A) is always
reductive, hence isomorphic to its dual as a t(A)-module.) The key formulas that ensure that
the Jacobi identities hold in g(A, B) are then the following, which we state only for Ψ1 :
Ψ1 (u ∧ v)1 x = Q(u, x)v − Q(v, x)u,
Ψ1 (u ∧ v)2 x = 12 (v(ux) − u(vx)),
Ψ1 (u ∧ v)3 x = 21 ((xu)v − (xv)u).
These formulas are classical in the study of triality, that is for A = O. It is easy to check that
in the other cases, which are much easier, we can arrange so that they also hold.
Proposition 3.2. The real Lie algebra g(A, B) is compact, while g(Ã, B̃) is split.
Proof. Let KA be the non-degenerate quadratic form on t(A) which we used in the construction
of g(A, B). For A = O, this form is the Killing form on so8 and is therefore negative definite;
this is also true for the other division algebras. Now it is easy to check that the quadratic form
3
X
Q := KA + KB − QAi ⊗ QBi
i=1
is invariant and negative definite. Since g(A, B) is simple (except for A = B = C), this invariant
form must be proportional to the Killing form, which is therefore negative definite as well (it
cannot be positive definite!). Thus g(A, B) is compact.
Finally g(Ã, B̃) is clearly split since both t(Ã) and t(B̃) are split, and their product is a
maximal rank subalgebra of g(A, B).
3.6. Inclusions. It is not at all obvious that an inclusion B ⊂ B′ of composition algebras induces
an inclusion g(A, B) ⊂ g(A, B′ ) of Lie algebras. In fact the Allison-Faulkner process applied to
structurable algebras A ⊂ B does not imply inclusions g(A) ⊂ g(B). To discuss this problem
in our situation we may assume that B′ = B ⊕ eB is deduced from B via the Cayley-Dickson
process.
Theorem 3.1. There is a unique embedding of Lie algebras t(B) ֒→ t(B′ ) that makes g(A, B) a
Lie subalgebra of g(A, B′ ).
If B′ is deduced from B by the Cayley-Dickson process, the relevant embedding of t(B) inside
t(B′ ) is
t(B) ≃ {(U1 , U2 , U3 ) ∈ t(B′ ) | Ui (B) ⊂ B} ⊂ t(B′ ).
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 21
Proof. We want the inclusion g(A, B) ֒→ g(A, B′ ) to be a morphism of Lie algebras. It is easy
to see that the only non-obvious verification to be made is that the bracket of two elements
in Ai ⊗ Bi is the same in g(A, B) as in g(A, B′ ), which amounts to proving that we have a
commutative diagram
∧2 (Ai ⊗ Bi ) −→ t(A) × t(B)
↓ ↓
∧2 (Ai ⊗ B′i ) −→ t(A) × t(B′ ).
Since the norm on B is the restriction of the norm on B′ , it is actually enough to prove the
commutativity of the following diagram:
Ψ
∧2 Bi i
−→ t(B)
↓ ↓
Ψ′
∧2 B′i i
−→ t(B′ ).
We must check this for all i, so we consider the sum ΨB : ∧2 B1 ⊕ ∧2 B2 ⊕ ∧2 B3 −→ t(B). Note
that ΨB is dual to the action, hence surjective since the action map is injective by definition. So
the commutativity of the diagram above for i = 1, 2, 3 completely determines the inclusion of
t(B) inside t(B′ ). For this inclusion map to be well-defined, we need a compatibility condition,
namely that KerΨB ⊂ KerΨB′ . Everything being trivial for B = R, we will check that KerΨC ⊂
KerΨH ⊂ KerΨO and deduce the relevant embeddings t(C) ⊂ t(H) ⊂ t(O).
We begin with ΨC , which is dual to the inclusion map
t(C) = {(a, b, c) ∈ R3 , a = b + c} ⊂ ∧2 C1 ⊕ ∧2 C2 ⊕ ∧2 C3 ,
0 a
where for example a ∈ R maps to the skew-symmetric endomorphism −a 0
of C1 . On t(C)
we use the restriction of the canonical norm of R3 , and we find that the dual map is
! 2a + b + c a + 2b − c a − b + 2c
0 a 0 b 0 c
ΨC : , , 7→ , , ,
−a 0 −b 0 −c 0 3 3 3
!
0 1 0 0
whose kernel is generated by −1 0
,
1
−1
0
,
1
−1
0
.
multiplication by z ∈ H, and by zero on H⊥ . Note that with these conventions, it follows from
the multiplication rules in the Cayley-Dickson process that for z ∈ ImH, we have Lz = (Lz , −Lz )
and Rz = (Rz , Lz ).
It is easy to check that every skew-symmetric endomorphism of H is of the form La + Rb . In
matrix terms, if a = α1 i + α2 j + α3 k and b = β1 i + β2 j + β3 k, we have
0 −α1 −α2 −α3 0 −β1 −β2 −β3
α1 0 −α3 α2 , Rb = β1
0 β3 β2
La = .
α2 α3 0 −α1 β2 −β3 0 β1
α3 −α2 α1 0 β3 β2 −β1 0
Under the natural isomorphism so(H) ≃ ∧2 H induced by the norm on H, this means that
La = α1 (1∧i + j ∧k) + α2 (1∧j + k∧i) + α3 (1∧k + i∧j),
Rb = β1 (1∧i − j ∧k) + β2 (1∧j − k∧i) + β3 (1∧k − i∧j).
Now we can ask what is the element of t(O) defined by u1 = 1∧i ∈ so(H) ⊂ so(O)? To answer
this question, we just need to note that 1∧i = 12 (Li + Ri ), so u2 = 21 Li and u3 = 12 Ri . Also
if we take u1 = j ∧k, a straightforward computation gives u2 = j ∧k + 21 Ri = 21 (Li , Li ) and
u3 = j ∧k − 12 Li = 21 (−Ri , Li ). Letting automorphisms of H act, we deduce that the following
triplets (u1 , u2 , u3 ) belong to t(O), where z, u, v denote imaginary quaternions:
u1 1∧z u∧v (Lz , 0) (Rz , 0) (0, Lz )
1 1
u2 L
2 z u∧v + R
4 [u,v] (L z , 0) (0, −L z ) (R z , 0)
1 1
u3 2 Rz u∧v − 4 L[u,v] (0, Lz ) (Rz , 0) (−Lz , 0)
Thus we have
ΨO ((Lx , 0), 0, 0) = ((Lx , 0), (0, −Lx ), (Lx , 0)) = ΨO (0, 0, (Lx , 0)),
ΨO ((Ry , 0), 0, 0) = ((Ry , 0), (Ry , 0), (0, Ly )) = ΨO (0, (Ry , 0), 0),
ΨO (0, (Lz , 0), 0) = ((0, −Lz ), (Lz , 0), (−Rz , 0)) = ΨO (0, 0, (−Rz , 0)).
This immediately implies that ((Lx + Ry , 0), (Lz − Ry , 0), (−Lx + Rz , 0)) belongs to KerΨO ,
and we deduce from the preceding formulas that the induced imbedding of t(H) ⊂ t(O) is given
by
(La + Rb , La − Rc , Lc + Rb ) ∈ t(H)
7→ ((La + Rb , Lc ), (La − Rc , −Lb ), (Lc + Rb , La )) ∈ t(O).
In particular, one can check that as a subalgebra of t(O), t(C) is generated by (Li + Ri , Li , Ri )
and (Ri , −Ri , Li + Ri ).
A more careful inspection of these formulas yields the explicit conclusion stated in the theorem.
3.7. Automorphisms, symmetric and trisymmetric spaces. The symmetry in the triality
model in the roles of A and B allows one to exhibit automorphisms of exceptional Lie algebras
with interesting properties. For example:
Proposition 3.3. [62] The endomorphism of g(A, B) defined as the identity on h(A, B) = t(A)×
t(B) ⊕ A1 ⊗ B1 , and minus the identity on A2 ⊗ B2 ⊕ A3 ⊗ B3 , is a Lie algebra involution.
Exhibiting a Lie algebra involution or a symmetric space is more or less equivalent, and we
conclude that there exists a “magic square” of symmetric spaces of dimensions 2ab where a and
b denote the dimensions of A and B. In particular we always get powers of 2! These are not
all uniquely defined by our Lie algebra involution, but a version of this magic square is the
following:
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 23
R C H O
R RP2 CP2 HP2 OP2
C CP2 CP2 × CP2 GC (2, 6) OP2C
H HP 2 GC (2, 6) GR (4, 12) E7(−5)
O OP2 OP2C E7(−5) E8(8)
The symmetric space E8(8) = E8 /SO16 , of dimension 128, is particularly intriguing and it would
be very nice to have a direct construction of it. It is claimed in [76] that it can be interpreted as
a projective plane over O ⊗ O, and that in fact the whole square above of symmetric spaces can
be obtained by taking projective planes, in a suitable sense, over the tensor products A ⊗ B.
It is also easy to describe Lie algebra automorphisms of order three, reflecting the trialitarian
origin of the magic square. Recall that t(A) has a natural automorphism τA or order three.
τA,B (X + Y + U1 + U2 + U3 ) = τA (X) + τB (Y ) + U2 + U3 + U1
for X ∈ t(A), Y ∈ t(B), Ui ∈ Ai ⊗ Bi , is a Lie algebra automorphism of order three, whose fixed
points set is the subalgebra
When B = O, the conclusion also holds when we replace the usual triality automorphism τO
by its twisted version τO′ . In this case, the fixed point set is the subalgebra k′ (A, B) = DerA ×
su3 ⊕ A ⊗ B.
As in the case of involutions, we deduce from this statement a “magic square” of homogeneous
spaces which are no longer symmetric, but are sometimes called trisymmetric: they are quotients
of a semisimple Lie group G by a subgroup K which is, up to a finite group, the fixed point set
of an automorphism of G of order three. Their dimensions are equal to 2ab + 2a + 2b − 4.
R C H O
R U2 /U1 × U1 U3 /(U1 )3 Sp6 /U2 × Sp2 F4 /(Spin7 × S 1 /Z2 )
C U3 /(U1 )3 U3 × U3 /(U1 )6 U6 /(U2 )3 E6 /SO8 × (SO2 )2
H Sp6 /U2 × Sp2 U6 /(U2 )3 SO12 /U4 × SO4 E7 /S(U7 × U1 )/Z2
O F4 /(Spin7 × S 1 /Z2 ) E6 /SO8 × (SO2 )2 E7 /S(U7 × U1 )/Z2 E8 /SO14 × SO2
Here we denoted by F4 , E6 , E7 , E8 the compact centerless groups of these types. Note that
for example, U3 /U1 × U1 × U1 is the variety of complete flags in C3 . Similarly, U6 /U2 × U2 × U2
is the flag variety F(2, 4, 6). Also SO12 /U4 × SO4 is the space of 8-dimensional subspaces in
R12 endowed with an orthogonal complex structure. This is because SO2n acts transitively on
the space of orthogonal complex structures. Moreover, the stabilizer of a point is the group
of orthogonal transformations commuting with the corresponding complex structure, and this
subgroup of SO2n is a copy of Un .
In the twisted case, for B = O, we replace the last line by
F4 /(SU3 × SU3 /Z3 ) E6 /(SU3 × SU3 × SU3 /Z3 ) E7 /(SU3 × SU6 /Z3 ) E8 /SU3 × E6 ,
a series of trisymmetric spaces of dimension 18a + 18. Note that at the infinitesimal level, we
obtain the quotients g(A, O)/su3 × g(A, C).
Automorphisms of order three were studied and classified by Wolf and Gray [34]. In the
exceptional cases the construction above is not far from giving the whole classification.
24 J.M. LANDSBERG AND L. MANIVEL
3.8. Dual pairs. The triality model allows one to identify series of dual pairs in the exceptional
Lie algebras. Recall that a pair (h, h′ ) of Lie subalgebras of a Lie algebra g is a dual pair if h is
the centralizer of h′ and vice versa. Dual pairs of Lie algebras, or of Lie groups (with the same
definition), have been extensively studied after the discovery by R. Howe that certain dual pairs
of groups inside the real symplectic group Sp(2n, R) have very special properties with respect
to the (infinite dimensional) metaplectic representation [43].
First recall that, as we saw in 4.7, the triality model is compatible with inclusions, in the
sense that g(A, B) is naturally embedded in g(A′ , B′ ) when A ⊂ A′ and B ⊂ B′ .
Proposition 3.5. The centralizers in g(A, O) of the subalgebras g(A, H), g(A, C), g(A, R) and
t(A), are isomorphic to su2 , su3 , g2 and z(t(A)) × so8 , respectively. Moreover, the centralizers
of these centralizers are the subalgebras themselves.
Here we denoted by z(t(A)) the center of t(A), which is t(A) itself when this algebra is
commutative, that is for A = R or C, and zero otherwise.
Proof. It is easy to see that the centralizer of g(A, B) inside g(A, O) is a subalgebra of t(O),
equal to the subalgebra of the centralizer of t(B) inside t(O) acting trivially on each Bi ⊂ Oi .
If B = R we get the subalgebra of elements (u1 , u2 , u3 ) ∈ t(O) killing the unit element in each
Oi . Making x = 1 and y = 1 in the triality relation u1 (xy) = u2 (x)y + xu3 (y), we see that
u1 = u2 = u3 is a derivation of O, so that the center of g(A, R) inside g(A, O) is DerO = g2 .
When B = C, the condition that (u1 , u2 , u2 ) kills Ci ⊂ Oi implies that u = u1 = u2 = u3 is a
derivation of O such that u(i) = 0 (i = e1 in the notations of 5.1). But then u(ix) = iu(x), which
means that u is a complex endomorphism of C⊥ ≃ C3 , endowed with the complex structure
defined by Li (and also that defined by Ri , which is the conjugate complex structure, since
u(xi) = u(x)i). Note that the corresponding element of t(O) automatically centralizes t(C),
since u commutes with Li , Ri and we saw that t(C) is generated as a subalgebra of t(O) by
(Li + Ri , Li , Ri ) and (Ri , −Ri , Li + Ri ). Finally, take the general matrix of g2 as written in §3.2.
The condition that u(i) = 0 means that α2 = α3 = α4 = α5 = α6 = α7 = 0. A complex basis
of the space C⊥ is (e2 , e4 , e6 ), in terms of which u is the complex endomorphism given by the
matrix
iβ3 −β4 + iβ5 −β6 + iβ7
β4 + iβ5 iγ5 −γ6 + iγ7 ,
β6 + iβ7 γ6 + iγ7 −i(β3 + γ5 )
In particular, we have pairs of reductive subalgebras in g(A, O), each of which is the centralizer
of the other: this is a reductive dual pair. The classification of reductive dual pairs inside
reductive complex Lie algebras was obtained in [77]. We deduce from the proposition above the
existence of uniform series of what Rubenthaler calls towers of reductive dual pairs.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 25
g(A, O) = g(A, O)
∪ ∪
g(A, H) so8 × z(t(A))
∪ ∪
g(A, C) g2
∪ ∪
g(A, R) su3
∪ ∪
t(A) su2
We conclude that so8 × t(A), g2 × g(A, R), su3 × g(A, C) and so3 × g(A, H) are maximal rank
reductive subalgebras of g(A, O). A natural question to ask is to describe the module structure
of g(A, O) over each of these subalgebras. For the first one, the answer is given by the triality
construction. For the second one, we note that g(A, R) = Der(J3 (A)) and g2 = DerO, so that
we are back to the original Tits construction of the exceptional series of Lie algebras. We obtain:
g(A, O) = g2 × g(A, R) ⊕ R7 ⊗ J3 (A)0 ,
= su3 × g(A, C) ⊕ R3 ⊗ J3 (A) ⊕ (R3 ⊗ J3 (A))∗ ,
= su2 × g(A, H) ⊕ R2 ⊗ Z3 (A).
Here Z3 (A) denotes the space of Zorn matrices, which is a g(A, H)-module, see §4.6. Note that
su2 acts not on R2 but on C2 . Therefore there must exist a complex structure on Z3 (A) =
Z30 (A) ⊕ iZ30 (A), which is not g(A, H)-invariant, so that R2 ⊗ Z3 (A) = C2 ⊗ Z30 (A), the su2 -
action coming from C2 .
These towers of dual pairs in series were discovered in joint work with B. Westbury.
3.9. The quaternionic form. The algebras g(A, B̃) are also interesting, especially when B =
O. Following Wolf [84], each simple complex Lie algebra has a unique real form such that the
associated compact or non-compact symmetric spaces have an invariant quaternionic structure.
We will say that this real form is quaternionic.
Recall that the rank of a symmetric space G/K, corresponding to some Cartan decomposition
g = k ⊕ p, is the maximal dimension of a subalgebra of g contained in p. Such a subalgebra is
automatically abelian and is called a Cartan subspace.
Theorem 3.6. The real Lie algebra g(A, Õ) is quaternionic. A Cartan subspace of t(Õ) = so4,4
embeds as a Cartan subspace of g(A, Õ), which in particular has real rank four (independent of
A).
Proof. Recall that each simple complex Lie algebra g, once a Cartan subalgebra and a set of
positive roots has been chosen, has a canonical 5-graduation defined by the highest root α̃, more
precisely by the eigenspaces of ad(Hα̃ ), where Hα̃ denotes the coroot of α.
If g is the complexification of the compact real Lie algebra g(A, O), we can arrange so that
the highest root comes from t(O)C = so8 (C). Note that the space of highest root vectors in
so8 (C) is the cone over the Grassmannian of isotropic planes in C8 . Using this, we then check
that we can arrange so that
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
iHα̃ =
.
0 0 0 0 0 −1 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 −1
0 0 0 0 0 0 1 0
26 J.M. LANDSBERG AND L. MANIVEL
The eigenvalues of ad(Hα̃ ) acting on g are 0, ±1, ±2. The corresponding eigenspaces are
g0 = t(AC ) × t(OC )0 ⊕ AC,1 ⊗ O0C,1 ⊕ AC,2 ⊗ O0C,2 ⊕ AC,3 ⊗ O0C,3 ,
g±1 = t(OC )±1 ⊕ AC,1 ⊗ O±1 ±1 ±1
C,1 ⊕ AC,2 ⊗ OC,2 ⊕ AC,3 ⊗ OC,3 ,
g±2 = t(OC )±2 ,
where we denoted by Oti the t-eigenspace for the action of Hα̃ on Oi . Note that iHα̃ belongs
to DerO = g2 (see our description of g2 above), so we can let i = 1 and consider the standard
action of Hα̃ . Then it is clear that the kernel of Hα̃ is the complexification of the standard
quaternion subalgebra H, and that the sum of the two other eigenspaces is the complexification
of eH.
Now, let θ denote the involution of g associated to its 5-graduation. By definition, θ acts by
(−1)k on gk . Then θ stabilizes g(A, O), and the corresponding Cartan decomposition g(A, O) =
k ⊕ p into eigenspaces of θ is given, in terms of the Cartan decomposition t(O) = kO ⊕ pO , by
k = t(A) × kO ⊕ A1 ⊗ H1 ⊕ A2 ⊗ H2 ⊕ A3 ⊗ H3 ,
p = pO ⊕ A1 ⊗ eH1 ⊕ A2 ⊗ eH2 ⊕ A3 ⊗ eH3 .
Finally (see [84, 17]), we obtain the quaternionic form g(A, O)H of g by twisting this decom-
position, that is, letting g(A, O)H = k ⊕ ip, which amounts to multiplying the brackets in g(A, O)
of two elements of p by −1. From the description of k and p we have just given, we see that
this twist amounts to doing two things. First, twist the Cartan decomposition of t(O), which
means that we replace t(O) = so8 by its quaternionic form so4,4 = t(Õ). Second, twist the
Cayley-Dickson process of construction of O from H by multiplying by −1 the product of two
elements in eH, which amounts to replace O by the split Cayley algebra Õ. This proves the first
part of the theorem. The rest of the proof is straightforward.
Adjoint varieties and quaternionic symmetric spaces. Symmetric spaces with quaternionic struc-
tures are closely related to adjoint varieties. Recall that on a Riemannian manifold M , a quater-
nionic structure is defined as a parallel field of quaternion algebras Hx ⊂ End(Tx M ), x ∈ M ,
such that the unit sphere of Hx is contained in the orthogonal group SO(Tx M ). The dimension
of M is then equal to 4m for some m > 0, and its reduced holonomy group is contained in
Sp(m)Sp(1) ⊂ SO(4m), where Sp(m) denotes the group of quaternionic unitary matrices of
order m (in particular, Sp(1) ≃ S 3 is the unit sphere in H). Such matrices act on Hm = R4m
by multiplication on the left, which commutes with the scalar multiplication on the right by
unitary quaternions: the resulting group of orthogonal transformations of R4m is Sp(m)Sp(1).
This group appears in Berger’s classification as one of the few possible reduced holonomy
groups of nonsymmetric Riemannian manifolds. The case of symmetric manifolds was discussed
in detail by J. Wolf in [84], who proved that there exists exactly one G-homogeneous symmetric
space M with a quaternionic structure for each simple compact Lie group G. When G is simply
connected, we can write M = G/K.Sp(1). Choosing a complex plane in H amounts to choosing
a circle S 1 ⊂ Sp(1), and this induces a fibration
X = G/K.S 1 −→ M = G/K.Sp(1)
whose fibers are two-spheres. Wolf proved that X is a complex variety, homogeneous under
the complexification GC of the real compact group G, and endowed with an invariant contact
structure. Such varieties were previously classified by Boothby, who showed that they are in
correspondance with simple complex Lie groups of adjoint type. In fact, X is the adjoint variety
(GC )ad , the closed orbit in PgC .
In modern terminology, X is the twistor space of M , and Wolf’s work is at the origin of this
twistor theory, which assigns to each Riemannian manifold M with a quaternionic structure a
complex variety X which is a S 2 -bundle over M , defined at each point x ∈ M by the unit sphere
in ImHx .
The adjoint varieties are endowed with natural contact structures (induced by the Kostant-
Kirillov symplectic structures on the minimal nilpotent coadjoint orbits, and of which the adjoint
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 27
varieties are the projectivizations). In fact, any compact Riemannian manifold with exceptional
holonomy Sp(n)Sp(1) and positive scalar curvature has a twistor space that is a contact Fano
manifold. Our understanding of contact Fano manifolds has recently improved very much. First
of all, there are not very many:
Theorem 3.2 (LeBrun-Salamon [69]). Up to biholomorphism, there are only finitely many con-
tact Fano manifolds of any given dimension.
This lead to the following conjecture:
Conjecture (LeBrun-Salamon). Let X be a smooth complex projective contact variety with
b2 (X) = 1. Then X must be the adjoint variety of a simple complex Lie group.
Both the theorem and conjecture can be rephrased purely on the Riemannian side using the
twistor transform.
In [52], S. Kebekus proved that the space of minimal rational curves through a fixed point of
X is a Legendre variety (except for the case of P2n−1 ), just as is the case for the adjoint varieties.
together the contributions of all its elementary pieces, one associates a morphism from g ⊗ p to
g ⊗ q to the graph. This morphism depends only on the isotopy class of the graph, and factors
through the AS and IHX relations, because of the skew-symmetry and Jacobi relation for the
Lie bracket in g. Vogel proves a stronger result for simple quadratic Lie algebras. One can
then replace the category D by another one D ′ , which is deduced from D by forcing an action
of the algebra Λ of skew-symmetric connected uni-trivalent graphs with exactly three univalent
vertices. The functor from D ′ to Modk g is then compatible with the action of Λ through a
character χg : Λ → k.
From the point of view of representation theory, the existence and the properties of the
categories D and D ′ have very interesting consequences. Indeed, each “decomposition” of [p]
in D ′ will imply the existence of a decomposition of g ⊗ p of the same type, for every simple
quadratic Lie algebra. Here, by a “decomposition” of [p], we must understand a decomposition
of the identity endomorphism of [p] as a sum of idempotents, which will correspond to projectors
onto submodules of g ⊗ p .
Describing such idempotents is far from easy and requires quite subtle computations with
uni-trivalent graphs. For p = 2, Vogel obtained the following result. (Vogel’s proof relied on
conjecture 3.6 as stated in [83], but Vogel himself realized that it is not correct as Λ, which is
conjectured to be integral, has zero divisors. Nevertheless, a case by case verification shows that
Vogel’s conclusions are indeed correct.)
Let g be a simple Lie algebra, say over the complex numbers. Then there are decompositions
∧2 g = X1 ⊕ X2 ,
S 2 g = X0 ⊕ Y2 ⊕ Y2′ ⊕ Y2′′ ,
into simple g-modules (some of which may be zero), with X0 = C and X1 = g. Moreover, there
exists α, β, γ such that the Casimir operator acts on X1 , X2 , Y2 , Y2′ and Y2′′ by multiplication by
t, 2t, 2t − α, 2t − β and 2t − γ respectively, where t = α + β + γ. (In fact the Casimir organizes
the representations in series, e.g., the representations Xk have Casimir eigenvalue kt.) Finally,
the dimensions of these g-modules are given by rational functions in α, β, γ (the dimensions of
Y2′ and Y2′′ are deduced from that of Y2 by cyclic permutations of α, β, γ):
The scalars α, β, γ are readily computed for each simple complex Lie algebra. Note that they
are defined only up to permutation and multiplication by a same scalar (this is because the
Casimir operator has not been normalized). We can therefore consider (α, β, γ) as a point in
P2 /S3 .
Series Lie algebra α β γ
SL sln −2 2 n
OSP son , sp−n −2 4 n−4
EXC sl3 −2 3 2
g2 −3 5 4
so8 −2 6 4
f4 −2 5 6
e6 −2 6 8
e7 −2 8 12
e8 −2 12 20
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 29
Note that up in P2 , the points of the three series SL, OSP and EXC are located on three
lines, respectively α + β = 0, 2α + β = 0, 2α + 2β − γ = 0. Hence the slogan that there exists
only three complex simple Lie algebras, SL, OSP, and EXC. At the level of representations, it is
a classical and well-known fact that the categories of modules over the sln , or the son and sp2n ,
have close and precise relationships. This is more surprising for the exceptional series. In §4.6
we discuss another collection of Lie algebras that form a “line” in Vogel’s plane.
Remark 4.1. In a spirit similar to Vogel’s, Cvitanovic [21] has proposed a proof of the Killing-
Cartan classification. Using notation inspired by the Feynman diagrams of quantum field theory,
he represents the invariant tensors of a Lie algebra by diagrams, and by requring that irreducible
representations be of integer dimension, he deduces severe limits on what simple groups could
possibly exist. His approach dates back to 1976 [19, 20]. The method provides the correct
Killing-Cartan list of all possible simple Lie algebras, but fails to prove existence.
4.2. Vogel’s decompositions and Tits correspondences. As remarked above, Tits corre-
spondences aid one in decomposing the tensor powers of nice (e.g., fundamental) representations
into irreducible factors. We illustrate this by showing how to recover Vogel’s decompositions
of S 2 g and Λ2 g. In fact we can recover Vogel’s higher decompositions as well, but we need to
use a slightly more general technique which we call diagram induction, see [63]. Using diagram
induction, we are also able to recover the Casimir functions for the representations occuring in
series such as the Xk .
If V is a fundamental representation corresponding to a root that is not short (e.g., the
fundamental adjoint representations) and X = G/Pα ⊂ PV the closed orbit, then the Fano
variety F1 (X) of P1 ’s on X is G-homogeneous by Theorem 2.5, and according to Tits fibrations,
F1 (X) = G/PS where S is the set of roots joined to α in D(g). Let V2 = hF1 (X)i denote
the linear span of the cone over F1 (X) in its minimal homogeneous embedding. Note that
F1 (X) ⊂ G(2, V ) ⊂ Λ2 V , and we conclude that V2 ⊂ Λ2 V . (To do this, one needs to make sure
the embedding of F1 (X) is indeed the minimal one, which is one reason we need the stronger
technique of diagram induction.)
In the case of the adjoint representation, we also know, because of the cobracket, that g ⊂ Λ2 g.
In fact, we have Λ2 g = g ⊕ g2 . (In the case of an , (an )2 is the direct sum of two dual irreducible
representations, which is also predicted by Tits fibrations.)
Similiarly, inside S 2 g, we know there is a trivial representation X0 corresponding to the Killing
form, and the Cartan power g(2) , which, breaking Vogel’s symmetry, we set Y2 = g(2) so the
problem becomes to recover the remaining modules, which we think of as the “primitive” ones.
These are provided in most cases by Tits fibrations of homogeneous quadrics contained in X.
There are up to two such families, we call the larger one gQ = Y2′ and if there is a second we
call it gQ′ = Y2′′ , see [63]. Thus, using Tit’s correspondences to find the primitive factors, one
recovers the Vogel decomposition of g ⊗ 2 for g simple (except for cn , which is tractible using
diagram induction). Even when our methods predict all factors, we would like to emphasize
that they do not explain why these are the only factors.
While our observations enable one to recover uniform decompositions of plethysms in series,
the questions of why such series exist and how to find them systematically without looking at
the list of simple Lie algebras is mysterious using our present techniques. It would be wonderful
and challenging to generalize our perspective sufficiently to approach Vogel’s. One indication
that such a generalization might be possible is that, with the above choices, we obtain a geo-
metric interpretation of Vogel’s parameter β, namely β is the dimension of the largest quadric
hypersurface contained in Xad (see [63]). If there is a second unextendable quadric in Xad , then
γ is its dimension.
4.3. The exceptional series. Inspired by Vogel’s work, Deligne [24], investigated the decom-
positions of the tensor powers of the exceptional Lie algebras into irreducible components, going
up to degree four (with the help of Cohen, deMan and the computer program LiE, [14]) and
giving more explicit dimension formulas.
30 J.M. LANDSBERG AND L. MANIVEL
4.3.1. Decomposition formulas. As with the Vogel decompositions, these decomposition formulas
are all uniform, in the sense, for example, that the numbers of irreducible components are always
the same when the algebra varies in the exceptional series. This assertion has to be understood
with some care: it happens that some components vanish, or should be taken with a minus sign.
In addition, Deligne and Vogel deal with the algebra twisted by the symmetry of the marked
(for the adjoint representation) diagram.
In degree two, one has the decomposition of Vogel, but the decomposition of S 2 g simplifies
because the equation 2α + 2β − γ = 0 implies Y2′′ is zero. In degree three we have
S 3 g = g ⊕ X2 ⊕ A ⊕ Y3 ⊕ Y3′ ,
∧3 g = C ⊕ X2 ⊕ Y2 ⊕ Y2′ ⊕ X3 ,
S21 g = 2g ⊕ X2 ⊕ Y2 ⊕ Y2′ ⊕ A ⊕ C ⊕ C ′
4.3.2. Dimension formulas. For each exceptional Lie algebra sl2 , sl3 , so8 g2 , f4 , e6 , e7 , e8 , let
λ = −3, −2, −1, − 32 , − 32 , − 12 , − 13 , − 51 respectively. λ is a linear function of the length of the
longest root (with the Casimir normalized to act as the identity on g). In Vogel’s parametrization
this gives (α, β, γ) = (λ, 1 − λ, 2).
From the decomposition formulas stated above and the knowledge of the Casimir eigenvalue
on each irreducible component, it is easy (at least with a computer) to calculate the dimensions
of these components as rational functions of λ, if one knows the dimension of g itself in terms
of λ. For example,
(λ + 5)(λ − 6)
dimg = −2 ,
λ(λ − 1)
(λ + 5)(λ − 4)
dimY2 = −90 ,
λ2 (λ− 1)(2λ − 1)
4.3.3. Deligne dimension formulas via triality. Deligne’s conjecture is elegant while his proof is
brute force computer based. In [62] we reprove and extend his formulas using methods whose
level of elegance is somewhere between the conjecture and proof of Deligne. The starting point
2
was our observation that the parameter λ could be written as λ = − a+2 , with a = − 23 , 1, 2, 4, 8.
(This implies that (α, β, γ) = (−2, a + 4, 2a + 4).) This indicated that the relation between
exceptional Lie algebras and normed division algebras should be exploited. The idea was to use
this relation to find a good description of the exceptional root systems, and then apply the Weyl
dimension formula.
The Tits construction is not convenient to describe the exceptional root systems and we were
led to rediscover the triality model. Recall that for the exceptional series,
g(A, O) = t(A) × t(O) ⊕ (A1 ⊗ O1 ) ⊕ (A2 ⊗ O2 ) ⊕ (A3 ⊗ O3 ),
can be identified with the compact form of f4 , e6 , e7 , e8 respectively for A = R, C, H, O. In what
follows we complexify this construction, without changing notation. In particular t(O) is now
the complex orthogonal Lie algebra so8 (C). The point is that t(A) × t(O) is now a maximal rank
reductive subalgebra of g(A, O). In particular, if we choose Cartan subalgebras h(A) ⊂ t(A)
and h(O) ⊂ t(O), we obtain a Cartan subalgebra h(A) × h(O) ⊂ g(A, O). This provides a nice
description of the root system of g(A, O). There are three kinds of roots:
• the heart of the root system is the set of roots of t(O) = so8 (C),
• the linear part is made of the roots of the form µ + ν, where µ is a weight of Oi and ν a
weight of Ai for some i,
• the residue is the set of roots of t(A).
When A = 0, the root system is of type D4 . When A = R, the root system is of type F4 .
There is no residue in this case. The heart is the set of long roots, and the linear part the set of
short roots. Both are root systems of type D4 . When we change A into C, H, O, the linear part
expands, is size being each time multiplied by two, and its roots become long. The residue also
increases, but in limited proportions.
Next we specify the positive roots inside the exceptional root systems. They are the roots on
which some linear form takes positive values, and we choose a linear form that takes very large
values on the set ∆+ of positive roots of so8 (C). Then the positive roots in the linear part of the
root system are the µ + ν for µ belonging to an explicit set Σ of weights. This implies that the
half-sum of the positive roots can be expressed as ρ = ρt(A) + ρt(O) + aγt(O) , where ρt(A) ∈ h(A)∗
denotes the half sum of the positive roots in t(A) and γt(O) ∈ h(O)∗ is the sum of the weights
belonging to Σ.
We are now close to being able to apply the Weyl dimension formula to g(A, O)-modules.
What remains to do is to describe the dominant integral weights in h(A)∗ × h(O)∗ . At least we
can describe the set C(O) ⊂ h(O)∗ of weights that are dominant and integral for each g(A, O).
Of course, such weights are be dominant and integral in so8 (C), but this is not sufficient.
Proposition 4.1. The set C(O) ⊂ h(O)∗ is the simplicial cone of nonnegative integer linear
combinations of the four following weights:
1 2 2 1
The lattice in t(O)∗ generated by C(O) is not the weight lattice, but the root lattice of so8 (C).
These weights, that we expressed above in terms of simple roots, occur respectively, ω(g) = ω2
as the highest weight of g, ω(X2 ) = ω1 + ω3 + ω4 as the highest weight of Λ2 g, ω(X3 ) = 2ω1 + 2ω3
as the highest weight of Λ3 g, and ω(Y2′ ) = 2ω1 as the highest weight of S 2 g − g(2) .
We can now apply the Weyl dimension formula to weights in C(O), considered for each choice
of A as integral dominant weights of g(A, O). It is essential that the residue of the root system
will not contribute. We obtain (see [62] for details):
32 J.M. LANDSBERG AND L. MANIVEL
Theorem 4.2. The dimension of the irreducible g(A, O)-module of highest weight ω ∈ C(O) is
given by the following formula:
(aγt(O) + ρt(O) + ω, β ∨ ) + a2 − 1
Y (aγt(O) + ρt(O) + ω, α∨ ) Y (ω, β ∨ )
dim Vω = ∨) − a
.
(aγt(O) + ρt(O) , α∨ )
α∈∆+ ∪Σ
(aγ t(O) + ρ t(O) + ω, β 2
β∈Σ
(ω, β ∨ )
If ω = pω(g) + qω(X2 ) + rω(X3 ) + sω(Y2∗ ), this formula gives a rational function of a, whose
numerator and denominator are products of 6p + 12q + 16r + 10s + 24 linear forms in a with
2
integer coefficients. Since λ = − a+2 , we obtain formulas of the type of those of Deligne, and
an infinite family of such. For example, the k-th Cartan power g(A, O)(k) is Yk in Deligne’s
notations, and we get
k
(2k − 1)λ − 6 Y ((j − 1)λ − 4)((j − 2)λ − 5)((j − 2)λ − 6)
dim Yk = .
k!λk (λ + 6) (jλ − 1)((j − 1)λ − 2)
j=1
Also we can understand why we can expect good dimension formulas for all irreducible compo-
nents of g(A, O) ⊗ k only for small k. This is because when k increases, we will unavoidably get
components whose highest weight does not come only from so8 (C), but has a contribution from
h(A)∗ . Then the Weyl dimension formula does not give a well behaved expression.
While this takes some of the mystery out of the dimension formulas, the remarkable symmetry
λ 7→ 1 − λ noticed by Deligne, which in our parameter a is a 7→ −4 a+3 a+4 remains beyond our
understanding.
4.3.4. Numerology. In the formula above for dim Yk , there are very few values of λ for which
dimYk is an integer for all k. We are currently, with B. Westbury [65] investigating these extra
numbers that produce integers and Lie algebras that go with them. At least in the case a = 6,
one is led to an algebra of dimension six, the sextonions, that leads to a nonreductive row of the
magic square between the third and fourth. Like the odd symplectic groups of Proctor [73] and
Gelfand and Zelevinsky [36], this series has certain behavior as if it were a series of reductive
Lie algebras.
4.4. Freudenthal geometries. The discovery of the Cayley plane OP2 and its automorphism
group by Chevalley, Schafer and others led to a period of intense activity around the excep-
tional groups and their geometric interpretations in the 1950’s involving Freudenthal, Tits and
Rozenfeld.
Freudenthal and Rozenfeld defined explicit varieties whose automorphism groups (or groups
related to the automorphism groups) were the exceptional groups. The starting point is the
Jordan algebras J3 (A). The projective plane AP2 is the space of rank one idempotents of
J3 (A). The subgroup of GL(J3 (A)) preserving the determinant of order three matrices (which
is well defined even for A = O) then acts transitively on AP2 . In particular, one recovers
the transitive action of E6 on OP2 . The subgroup of GL(J3 (A)) preserving the determinant
and the quadratic form (A, B) = trace(AB) is the automorphism group of the Jordan algebra
J3 (A). It acts irreducibly on J3 (A)0 , the subspace of trace zero matrices. Tits and Freudenthal
defined elliptic and projective geometries on AP2 with these groups of isometries. In the elliptic
geometry, a point is the same as a line, and is defined by an element of AP20 . In the projective
geometry, a point is defined to be an element of AP2 and a line is determined by an element [α]
of the dual projective plane (the rank one idempotents in the dual Jordan algebra corresponding
to the dual vector space), namely {X ∈ AP2 | α(X) = 0}.
A synthetic geometry Freudenthal terms symplectic may be associated to each of the groups
in the third row. A point is an element [X] ∈ Pg(A, H), such that ad(X)2 = 0. This condition
is equivalent to requiring that [X] is in the adjoint variety Xad ⊂ Pg(A, H). A plane is is
determined by an element [P ] ∈ PZ3 (A), P × P = 0. (See §4.6 for the definition of the cross
product.) This condition is equivalent to saying that [P ] ∈ Gw (A3 , A6 ) ⊂ PZ3 (A), the closed
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 33
orbit. The corresponding plane is {[X] ∈ Xad | XP = 0}. A line is an intersection of two
planes with at least two points. It is also determined by a point [α] ∈ F1 (X) ⊂ PV2 , where
V2 is the representation defined in §4.6 below and F1 (Gw (A3 , A6 )) is the closed orbit (which, as
the notation suggests, parametrizes the lines on the closed orbit Gw (A3 , A6 ) ⊂ PZ3 (A)). One
can then define incidence rules for these geometric elements which generalize the symplectic
geometry in P5 .
For the fourth line of Freudenthal’s magic square Freudenthal defines a metasymplectic geom-
etry. There are now four types of elements, points, lines, planes and symplecta, with incidence
rules explained in detail by Freudenthal. As above, each geometric element is a point in the
closed orbit of the projectivization of a g(A, O)-module. Here is the table of types of elements:
We see the four generators of the cone C(O) in Proposition 4.1 define the four types of
elements of Freudenthal’s metasymplectic geometry! The analogous results hold for the second
and third rows of the magic chart, that is, Freudenthals elements are the closed orbits in the
projectivizations of the modules whose highest weights define the fixed cones for our dimension
formulas in [62].
4.5. A geometric magic square. Consider the following square of complex homogeneous
varieties, which we investigate in [61]:
R C H O
R v2 (Q1 ) P(T P2 ) Gω (2, 6) OP2C,0 section of Severi
C v2 (P2 ) P2 × P2 G(2, 6) OP2C Severi
H Gω (3, 6) G(3, 6) S12 E7 /P7 Legendre
O F4ad E6ad E7ad E8ad adjoint
This geometric magic square is obtained as follows. One begins with the adjoint varieties for
the exceptional groups: this is the fourth line of the square. Taking the varieties of lines through
a point and applying Theorem 2.5, one obtains the third line. The second line is deduced from
the third by the same process. Then take hyperplane sections to get the first line. We obtain
projective varieties that are homogeneous under groups given by Freudenthal’s magic square.
4.6. The subexceptional series. In this final section we consider the series corresponding to
the third row of the extended Freudenthal chart:
4.6.1. Relation to the universal Lie algebra. This series also corresponds to a new line in Vogel’s
plane, namely (α, β, γ) = (−2, a, a + 4), where a = − 23 , 0, 1, 2, 4, 8. Moreover, Vogel’s formulas
are valid for the semi-simple Lie algebra sl2 × sl2 × sl2 despite the plane being defined only for
absolutely simple objects. Unlike the three lines discovered by Vogel, this line is generic to order
three among actual Lie algebras, in the sense that no Vogel space is zero except X3′′ , which is
zero for all actual Lie algebras. We have dim g(a) = 3(2a+3)(3a+4)
(a+4) .
34 J.M. LANDSBERG AND L. MANIVEL
4.6.2. Freudenthalia. For this series there are three preferred representations from Freudenthal’s
perspective, corresponding to the ambient spaces for points, lines, and planes in his incidence
geometries. The points representation is just g = g(A, H) and has a nice model thanks to the
triality construction. For a discussion of the line space, which we will denote V2 = V2 (a) see
[61]. We have dimV2 (a) = 9(a + 1)(2a + 3). The most preferred representation is for the planes,
which we denote V = V (a) and has dimV (a) = 6a + 8. This space is the complexification of the
algebra of Zorn matrices we encountered in §3.8:
( )
a X
Z3 (A) = , a, b ∈ R, X, Y ∈ J3 (A) .
Y b
The closed orbit Xplanes ⊂ PZ3 (A) is a natural compactification of the Jordan algebra J3 (A),
and its embedding inside PZ3 (A) is given by the translates of the determinant on J3 (A). This is
the conformal compactification considered by Faraut and Gindikin in [30]. In [61] we propose a
geometric interpretation of Xplanes as the Grassmannian of O3 ’s in O6 isotropic for a Hermitian
symplectic form and use the notation Gw (O3 , O6 ).
4.6.3. Decomposition formulas. The plethysms of V are extraodinarily well behaved in series.
For example, here is a decomposition of S k V into irreducible components for each k.
M
tk S k V = (1 − tV )−1 (1 − t2 g)−1 (1 − t3 V )−1 (1 − t4 )−1 (1 − t4 V2 )−1 .
k≥0
The right hand side in the formula is to be expanded out in geometric series and multiplication
of representations is taken in terms of Cartan products Vλ Vµ = Vλ+µ . The case of sl2 × sl2 × sl2
is somewhat special, since the formula above holds true only if the natural action of S3 is taken
into account. See [63].
Each of the factors in the rational function above can be accounted for in terms of diagram
induction. In particular, g = VQ , the ambient space for the variety parametrizing the G-
homogeneous quadrics on Xplanes and V2 = hF1 (Xplanes )i is the ambient space for the variety
parametrizing the lines on Xplanes .
4.6.4. Dimension formulas. Our method for applying the Weyl dimension formula in series also
works for the three distinguished representations. For example, we have
3a
k+2a+1 k+ 2 +1
2a + 2k + 2 2a+1 3a
+1
dim V (k) = 2
k+ a2 +1
,
a+1 a
2
+1
k+x
where the binomial coefficients are defined by k = (1 + x) · · · (k + x)/k! and thus are rational
polynomials of degre k in x.
Note that since V (k) is the complement of Ik (Xplanes ) in S k V ∗ , the above formula also gives
the Hilbert functions of the varieties Xplanes in a uniform manner.
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 35
4.6.5. Zakology. The varieties Xplanes ⊂ PV are important for the following geometric classifi-
cation problem: Given a smooth variety X ⊂ PV , one defines its dual variety X ∗ ⊂ PV ∗ to be
the set of hyperplanes tangent to X. Usually the degree of X ∗ is quite large with respect to X
and Zak has proposed the problem of classifying smooth varieties whose duals are of low degree
[86].
Here the duals of Xplanes ⊂ PV are of degree four. Moreover, they are the tangential varieties
∗
of the closed orbits in the dual projective space, i.e., we have Xplanes ≃ τ (Xplanes ).
1
The first variety in the series is Xplanes (−2/3) = v3 (P ) and the equation of its dual is the
classical discriminant of a cubic polynomial. Using A-valued variables, we write the equations
for the duals in a uniform fashion, see [61]. In particular, we characterize the other quartics by
their restriction to the preferred subspace. The second variety in the series is X = Seg(P1 ×
P1 × P1 ) and its dual is Cayley’s hyperdeterminant, see [35]. Our restriction result gives a new
characterization of the hyperdeterminant.
4.6.6. Orbits. The orbit structure is also uniform.
Proposition 4.3. [61] For each of the varieties Xplanes = Gw (A3 , A6 ) ⊂ PV there are exactly
four orbits, the closures of which are ordered by inclusion:
Gw (A3 , A6 ) ⊂ σ+ (Gw (A3 , A6 )) ⊂ τ (Gw (A3 , A6 )) ⊂ PV.
The dimensions are respectively 3a + 3, 5a + 3 and 6a + 6.
Needless to say, the singular orbit closures have uniform desingularizations by Kempf’s
method.
Once again, the triality model plays a unifying role for geometric and representation geometric
phenomena which at first glance seem sporadic. Another striking example of this role, which
is currently under investigation, concerns nilpotent orbits in exceptional Lie algebras and the
associated unipotent characters of exceptional Chevalley groups [66].
References
[1] Alexander J. , Hirschowitz A., Polynomial interpolation in several variables, J. Algebraic Geom. 4 no. 2,
201–222 (1995).
[2] Allison B. N., A class of nonassociative algebras with involution containing the class of Jordan algebras,
Math. Ann. 237 no. 2, 133–156 (1978).
[3] Allison B. N., Faulkner J. R., Nonassociative coefficient algebras for Steinberg unitary Lie algebras, J. Algebra
161 no. 1, 1–19 (1993).
[4] Arnol’d V.I., Normal forms of functions near degenerate critical points, the Weyl groups of Ak , Dk , Ek and
Lagrangian singularities, Funk. Anal 6, 3–25 (1972).
[5] Baez J., The octonions, arXiv: math.RA/0105155.
[6] Bar-Natan D., On the Vassiliev knot invariants, Topology 34 no. 2, 423-472 (1995).
[7] Barton C.H., Sudbery A., Magic squares of Lie algebras, arXiv:math.RA/0001083.
[8] Beauville A., Fano contact manifolds and nilpotent orbits, Comment. Math. Helv. 73 no. 4, 566–583 (1998).
[9] Bourbaki N., Groupes et algèbres de Lie, Hermann, Paris 1968.
[10] Cartan E., Le principe de dualité et la théorie des groupes simples et semi-simples, Bull. Sc. math. 49, 361-374
(1925).
[11] Cartan E., Sur les variétés de courbure constante d’un espace euclidien ou non-euclidien, Collected Works.
III, 321-432.
[12] Cartan E., Sur des familles remarquables d’hypersurfaces isoparamétriques dans les espaces sphériques. Math.
Z. 45, (1939). 335–367
[13] Carlini E., Chipalkatti J., On Waring’s problem for several algebraic forms, arXiv:math.ag/0112110.
[14] Cohen A., de Man R., Computational evidence for Deligne’s conjecture regarding exceptional groups, C.R.A.S.
322, 427-432 (1996).
[15] Chaput P. E., Severi varieties, arXiv:math.AG/0102042, to appear in Math. Zeitschrift.
[16] Chaput P. E., Variétés de Scorza et algèbres de Jordan, preprint 2002.
[17] Clerc J.L., Special prehomogeneous vector spaces associated to F4 , E6 , E7 , E8 and simple Jordan algebras of
rank 3, preprint 2001.
[18] Cohen A., Cooperstein B., Lie incidence systems from projective varieties, Proc. Am. Math. Soc. 126, 2095-
2102 (1998).
36 J.M. LANDSBERG AND L. MANIVEL
[19] Cvitanović P., Group theory for Feynman diagrams in non-Abelian gauge theories, Phys. Rev. D14, 1536
(1976).
[20] Cvitanović P., Negative dimensions and E7 symmetry, Nucl. Phys. B188, 373 (1981).
[21] Cvitanović P., Group theory, available at www.nbi.dk/GroupTheory/, a monograph in preparation (1984-
2000).
[22] Cohen A.M., van Leeuwen M.A., Lisser B., LiE, a package for Lie group computations, CAN, Amsterdam,
1992.
[23] Dadok J., Harvey R., A triality unification of Lie algebras, preprint.
[24] Deligne P., La série exceptionnelle des groupes de Lie, C.R.A.S. 322, 321-326 (1996).
[25] Deligne P., de Man R., The exceptional series of Lie groups, C.R.A.S. 323, 577-582 (1996).
[26] C. De Concini, J. Weyman, A formula with nonnegative terms for the degree of the dual variaety of homoge-
neous space, Proc. AMS, 125:1 (1997), pp 1-8 .
[27] Dixmier J., Algèbres enveloppantes, Gauthier-Villars, Paris 1974.
[28] Dynkin E.B., Maximal subgroups of the classical groups, Amer. Math. Soc. Trans. Series 6, 245-378 (1957).
[29] Ein L., Shepherd-Barron N., Some special Cremona transformations, Amer. J. Math. 111 no. 5, 783–800
(1989).
[30] Faraut J., Gindikin S., Pseudo-Hermitian symmetric spaces of tube type, in Topics in geometry, In memory
of Joseph D’Atri, Gindikin S. ed., Birkhauser, pp. 123-154 (1996).
[31] Freudenthal H., Oktaven, Ausnahmegruppen und Oktavengeometrie (notes from 1951), Geom. Dedicata 19,
7-63 (1985).
[32] Freudenthal H., Lie groups in the foundations of geometry, Advances in Math. 1, 145-190 (1964).
[33] Garibaldi R.S., Structurable algebras and groups of type E6 and E7 , J. Algebra 236, 651-691 (2001).
[34] Gray A., Wolf J., Homogeneous spaces defined by Lie group automorphisms, J. Diff. Geometry 2, 77-114 and
115-159 (1968).
[35] Gelfand I. M., Kapranov M. M., Zelevinsky A. V., Discriminants, resultants, and multidimensional determi-
nants, Mathematics: Theory and Applications, Birkhuser 1994.
[36] Gelfand I. M., Zelevinsky, A. V., Representation models for classical groups and their higher symmetries, The
mathematical heritage of Elie Cartan (Lyon, 1984), Astérisque 1985, Numéro Hors Série, 117–128.
[37] Griffiths P.A., Harris J., Algebraic Geometry and Local Differential Geometry , Ann. scient. Ec. Norm. Sup.
12, 355-432 (1979).
[38] Gross B., Wallach R., On quaternionic discrete series representations and their continuations, J. reine angew.
Math. 481, 73-123 (196).
[39] Harvey F.R., Spinors and calibrations, Perspectives in Math. Vol. 9, Academic Press, 1990.
[40] Hawkins, “Emergence of the theory of Lie groups. An essay in the history of mathematics 1869–1926 ”,
Sources and Studies in the History of Mathematics and Physical Sciences. Springer-Verlag, New York, 2000.
xiv+564 pp.
[41] Holweck F., Singularities of hyperplane sections of homogeneous varieties, in preparation.
[42] Hong J., Fano manifolds with geometric structures modeled after homogeneous contact manifolds Internat. J.
Math. 11 no. 9, 1203–1230 (2000).
[43] Howe R., Transcending classical invariant theory, J. Am. Math. Soc. 2, 535-552 (1989).
[44] Hwang J-M, Mok N., Varieties of minimal rational tangents on uniruled projective manifolds, Several complex
variables (Berkeley, CA, 1995–1996), 351–389, Math. Sci. Res. Inst. Publ. 37, Cambridge University Press
1999.
[45] Hwang J-M, Mok N., Uniruled projective manifolds with irreducible reductive G-structures, J. reine Angew.
Math. 490, 55–64 (1997).
[46] Hwang J-M, Mok N., Rigidity of irreducible Hermitian symmetric spaces of the compact type under Khler
deformation, Invent. Math. 131 no. 2, 393–418 (1998).
[47] Iarrobino A., Kanev V., Power sums, Gorenstein algebras and Determinantal Loci, Springer LNM 1721,
1999.
[48] Kac V., Infinite dimensional Lie algebras, Third edition, Cambridge University Press 1990.
[49] Kac V., Some remarks on nilpotent orbits J. Algebra 64 (1980), no. 1, 190–213.
[50] Kantor, I. Certain generalizations of Jordan algebras (Russian) Trudy Sem. Vektor. Tenzor. Anal. 16 (1972),
407–499.
[51] Kebekus S., Peternell Th., Sommese A., Wisniewski J., Projective contact manifolds, Invent. Math. 142, 1-15
(2000).
[52] Kebekus S., Lines on contact manifolds, J. reine Angew. Math. 539, 167–177 (2001).
[53] Kempf G., On the collapsing of homogeneous bundles, Invent. Math. 37, 229-239 (1976).
[54] F. Knop, Ein neuer Zusammenhang zwischen eifachen Gruppen und einfachen Singularitaten, Invent. Math
90, 579–604 (1987).
[55] Kostant B., The principle of triality and a distinguished unitary representation of SO(4, 4), Differential
geometrical methods in theoretical physics, Proc. 16th Int. Conf., Como/Italy 1987, NATO ASI Ser., Ser. C
250, 65-108 (1988).
REPRESENTATION THEORY AND PROJECTIVE GEOMETRY 37
[56] Kostant B., A cubic Dirac operator and the emergence of Euler number multiplets of representations for equal
rank subgroups, Duke Math. J. 100 no. 3, 447–501 (1999).
[57] Landsberg J.M., On degenerate secant and tangential varieties and local differential geometry, Duke Math.
J. 85, 605-634 (1996).
[58] Landsberg J.M., On the infinitesimal rigidity of homogeneous varieties, Compositio Math. 118, 189-201
(1999).
[59] Landsberg J.M., Manivel L., On the projective geometry of homogeneous varieties, arXiv:math.AG/9810140.
[60] Landsberg J.M., Manivel L., Classification of simple Lie algebras via projective geometry, to appear in Selecta
Mathematica.
[61] Landsberg J.M., Manivel L., The projective geometry of Freudenthal’s magic square, J. Algebra 239, 477-512
(2001).
[62] Landsberg J.M., Manivel L., Triality, exceptional Lie algebras, and Deligne dimension formulas, to appear
in Advances Math.
[63] Landsberg J.M., Manivel L., Series of Lie algebras, preprint.
[64] Landsberg J.M., Manivel L., Legendrian and contact Fano varieties, in preparation.
[65] Landsberg J.M., Manivel L., Westbury B., Sextonions and exceptional analogues of the odd symplectic groups,
in preparation.
[66] Landsberg J.M., Manivel L., Westbury B., Series of nilpotent orbits, in preparation.
[67] Lascoux A., Degree of the dual of a Grassmann variety, Comm. Algebra 9, 1215–1225 (1981).
[68] LeBrun C., Fano manifolds, contact structures, and quaternionic geometry, Internat. J. Math. 6, 419–437
(1995).
[69] LeBrun C., Salamon S., Strong rigidity of positive quaternion-Kahler manifolds, Invent. Math. 118, 109–132
(1994).
[70] Loos O., Symmetric spaces I and II, Benjamin 1969.
[71] Mukai S., Simple Lie algebra and Legendre variety, preprint 1998.
[72] Lie groups and Lie algebras, III. Structure of Lie groups and Lie algebras, Encyclopaedia of Mathematical
Sciences 41. Springer-Verlag 1994. iv+248 pp.
[73] Proctor I., Odd symplectic groups, Invent. Math. 92, 307–332 (1988).
[74] Postnikov M., Groupes et algèbres de Lie, Editions Mir, Moscou 1985.
[75] Ranestad K., Schreyer F.-O., Varieties of sums of powers, J. reine Angew. Math. 525, 1455-1468 (2000).
[76] Rosenfeld B., Geometry of Lie groups, Mathematics and its Applications, vol. 393, Kluwer 1997.
[77] Rubenthaler H., Les paires duales dans les algèbres de Lie réductives, Astérisque 219, 1994.
[78] Tevelev, E., Projectively dual varieties, preprint math.AG/0112028.
[79] Tits J., Algèbres alternatives, algèbres de Jordan et algèbres de Lie exceptionnelles I. Construction, Indag.
Math. 28, 223–237 (1966).
[80] Tits J., Les groupes de Lie exceptionnels et leur interprétation géométrique, Bull. Soc. Math. Belg. 8, 48-81
(1956).
[81] Tits J., Groupes semi-simples complexes et géométrie projective, Séminaire Bourbaki 7 (1954/1955), exposé
112, 11 pages.
[82] Vogel P., Algebraic structures on modules of diagrams, preprint 1996.
[83] Vogel P., The universal Lie algebra, preprint 1999.
[84] Wolf J., Complex homogeneous contact manifolds and quaternionic symmetric spaces, Journal of Math. and
Mech. 14, 1033-1047 (1965).
[85] Zak F., Tangents and secants of algebraic varieties, Translations of Math. Monographs 127, AMS 1993.
[86] Zak F., Some properties of dual varieties and their applications in projective geometry, in Algebraic geometry,
Proceedings, Chicago 1989, Springer LNM 1479. 273–280.