The marked length spectrum of a projective manifold or orbifold

A strictly convex real projective orbifold is equipped with a natural Finsler metric called the Hilbert metric. In the case that the projective structure is hyperbolic, the Hilbert metric and the hyperbolic metric coincide. We prove that the marked Hilbert length spectrum determines the projective structure only up to projective duality. A corollary is the existence of non-isometric diffeomorphic strictly convex projective manifolds (and orbifolds) that are isospectral. The corollary follows from work of Goldman and Choi, and Benoist.


Introduction
Let M be a Riemannian manifold. The marked length spectrum of M is the function : π 1 (M ) → R which assigns to each free homotopy class of loops the infimum of the length of loops in the class. An outstanding problem in Riemmannian geometry is the question of marked length spectrum rigidity for nonpositively curved manifolds: if a compact manifold M supports two nonpositively curved metrics g and g with the same marked length spectrum, is (M, g) isometric to (M, g )? This conjecture, which is attributed to Katok, appears in [7]. For non-positively curved surfaces, the marked length spectrum determines the metric [9], [17]. In [6] the authors prove the conjecture in the case of negatively curved locally symmetric manifolds.
In this paper we study compact, strictly convex, real projective orbifolds. We defer precise definitions to the next section. These objects include all hyperbolic manifolds. The Hilbert metric on such an n-orbifold, Q, assigns a length to every element of the orbifold fundamental group π orb 1 Q. The Hilbert-length spectrum is the resulting map : π orb 1 Q → R. Such a structure has a unique dual which is another such structure. Our main result is the following: Theorem 1.1. Let Q n be a compact, strictly convex, real projective orbifold. Two distinct structures on Q have the same marked Hilbert-length spectrum iff each structure is the projective dual of the other.
It is known that: Theorem 1.2. Let Q n be a compact, strictly convex, real projective orbifold. Then Q is self dual iff Q is real hyperbolic.
Together these provide examples of iso-spectral structures which are not isometric. However, in these cases, the Hilbert metric is not Riemannian, only Finsler.
We now sketch the proof of the main theorem, making a few simplifying assumptions without mention. We are given two monomorphisms of G = π orb 1 Q into P GL(n+1, R). Combining these gives an embedding of G onto Γ ⊂ P GL(n+1, R)× P GL(n + 1, R). Let Γ denote the Zariski closure. It is an algebraic group. By a beautiful theorem of Benoist, the projection of Γ onto each factor is either surjective or a conjugate of P O(n, 1). The latter is the hyperbolic case. In the former we now use the hypothesis on the length functions to show Γ is contained in a proper subvariety: in essence the translation length in the Hilbert metric of a matrix in P GL(n+1, R) is the logarithm of the ratio of two eigenvalues, and equality of ratios of eigenvalues of a pair of matrices is an algebraic condition. Since P GL(n + 1, R) is a simple Lie group, it follows that both coordinate projections of Γ are isomorphisms. Combining the inverse of one projection with the other projection gives an automorphism of P GL(n + 1, R) taking one embedding of G to the other. The theorem follows from the classification of automophisms of P GL(n + 1, R) namely the outer automorphism group is order two generated by the global Cartan involution given by the composition of transpose and inverse. This sends a structure to its dual. In the main proof we work over C rather than over R since this makes the algebraic geometry simpler. This completes the sketch.
The main theorem is essentially due to Inkang Kim, [12]. However he omitted the possibility of dual structures and the consequent discussion of self-dual structures. Our proof follows his in outline, but at a crucial point we use elementary algebra in place of a more sophisticated argument using buildings. We also extend Kim's result to compact orbifolds. In other work we study limits at infinity of these projective structures and we find that at infinity projective duality becomes isometry. This leads to some beautiful phenomena which are further explored in that work.
The authors thank Darren Long and Morwen Thistlethwaite for helpful discussions. This work was partially supported by NSF grant DMS-0706887. We also thank the referee for their careful reading and useful comments. The first author also thanks the mathematics department at the University of Capetown, and especially David Gay, for hospitality during the preparation of this paper.

Projective Manifolds and Orbifolds.
2.1. Projective structures. Real projective n-space RP n has automorphism group the projective general linear group P GL(n + 1, R), an element of which is called a projective transformation. If P ∼ = RP n−1 is a codimension-1 projective subspace in RP n the complement RP n \ P is called an affine patch. An affine patch can be naturally identified with R n in a way which is unique up to affine transformations of R n . A subset Ω ⊂ RP n is called a strictly convex subset if it has closure, Ω, which is a compact strictly convex subset of some affine patch R n . This means that ∂Ω does not contain any line segment of positive length.
A strictly convex real projective orbifold, Q, is the quotient Q = Ω/Γ where Ω is an open strictly convex set in RP n and Γ is a discrete subgroup of P GL(n + 1, R) which preserves Ω. The orbifold fundamental group π orb 1 Q is defined to be Γ. The orbifold universal cover isQ = Ω and the universal orbifold-covering projection is the quotient map π :Q → Q (see [11] p 139 for a more general definition). Observe that the orbifold fundamental group, π orb 1 Q, is the group of homeomorphisms that cover the identity map. If Γ acts freely on Ω then Q is a projective manifold.
The natural map SL(n + 1, R) → P GL(n + 1, R) is an isomorphism when n is even. However, in all dimensions Aut(Ω) lifts to SL ± (n + 1, R). It is often easier for computations to work with the latter group. We include the following well known argument for the convenience of the reader. Lemma 2.1. Let Ω be a strictly convex bounded subset of an affine patch in RP n . Then there is a lift of Aut(Ω) to SL ± (n + 1, R).
Proof. Let π : S n → RP n denote the double cover. Then π −1 (Ω) = Ω + ∪ Ω − has two connected components in S n . An element of Aut(Ω) corresponds to a pair of matrices ±A ∈ SL ± (n + 1, R) which act on S n and permute these two components. It follows that exactly one of this pair acts as the identity permutation. This defines the lift.
From now on, in this paper we will use the term projective orbifold to mean strictly convex real projective orbifold. Furthermore, we will assume that it is of the form Ω/Γ with Ω a bounded strictly convex open subset of R n and Γ a discrete subgroup of either P GL(n + 1, R) or SL ± (n + 1, R).
An orbifold isomorphism between projective orbifolds Q = Ω/Γ and Q = Ω /Γ is a map f : Q → Q which is covered by a homeomorphismf : Ω → Ω . Given f, the choice off is unique up to post-composition with an element of Γ . Such a choice induces an isomorphism f * : π orb 1 Q → π orb 1 Q of groups given by f * (γ) =f • γ •f −1 . The indeterminacy of this choice means f * is only defined up to conjugacy.
The map f is a projective isomorphism iff is the restriction of a projective transformation. In this case π orb 1 Q and π orb 1 Q are conjugate (via f * ) subgroups of P GL(n + 1, R).
To define the space of projective structures on a given orbifold we choose one such structure as a base point. Given a projective orbifold Q we define the space, C(Q), of strictly convex real projective structures on Q as follows. LetC(Q) denote the set of all triples (f, Ω, Γ) where R = Ω/Γ is a projective orbifold and f : Q → R is an orbifold isomorphism called a marking which is not assumed to be a projective isomorphism.
An element ofC(Q) determines a group monomorphism (up to conjugacy) hol : π orb 1 Q → P GL(n + 1, R) called the holonomy of the structure and is defined as hol = f * . The image of the holonomy is (up to conjugacy) Γ.
We define an equivalence relation ∼ onC(Q) by  Proof. Let K denote the closure of the set of attracting fixed points of proximal (see defintion 2.7) elements of Γ. Since Ω/Γ is compact K = ∂Ω. The domain Ω is strictly convex and therefore K is contained in an affine patch of RP n . This implies K separates RP n into two connected components, one of which is Ω and the other is homeomorphic to the normal bundle of a codimension one projective subspace, thus not a cell. Thus Ω is characterized as the component of RP n \ K which is a cell.
. By 2.2 the images of hol i determines Ω i . If hol 1 and hol 2 are conjugate by A ∈ P GL(n + 1, R) using 2.2 again shows that A sends Ω 1 to Ω 2 . Thus the two structures are equivalent.
Theorem 2.4 (Benoist, [5]). Suppose that Q is a closed strictly convex real projective orbifold. Then the image of Hol in Hom(π orb 1 Q, P GL(n+1, R))/P GL(n+1, R) is a union of path components of the Euclidean topology.
A natural question is: to what extent does the orbifold fundamental group determine a convex projective structure? There is the issue of moduli: there may be different structures on a given isomorphism type of orbifold. But there is a second issue. The Borel conjecture is that two closed aspherical manifolds with isomorphic fundamental groups are homeomorphic.
The Borel Conjecture holds for word hyperbolic groups, [1]. The fundamental group of a closed strictly-convex projective manifold is word hyperbolic [4]. Hence the Borel conjecture holds for strictly-convex projective manifolds.
The results above give a weak version of the Borel conjecture for orbifolds, but with isomorphism (orbifold-diffeomorphism) in place of homeomorphism: Corollary 2.5. If G is a group then there are at most finitely many isomorphism classes of closed connected orbifolds Q that admit a strictly convex projective structure such that π orb 1 Q ∼ = G. Proof. The holonomy determines the isomorphism type of a projective orbifold. If two projective orbifolds are part of a one parameter family then they are isomorphic. Theorem 2.4 implies there is at most one isomorphism type of Q for each path component of Hom(π orb 1 Q, P GL(n + 1, R))/P GL(n + 1, R) in the Euclidean topology. Since a real algebraic variety has only finitely many path components, the result follows.

The Hilbert Metric.
Definition 2.6 (Hilbert metric). Let Ω be a bounded convex open subset of R n . Let p, q be distinct points in Ω and let p ∞ and q ∞ be the points on the boundary of Ω colinear with p and q and ordered, so that p separates p ∞ from q. The Hilbert distance between p and q is given by the logarithm of the cross ratio of the points (p ∞ , p, q, q ∞ ) namely: The usual argument adapts to show that the straight line with endpoints p, q has minimal length, and in the case that Ω is strictly convex, this is the unique path of minimal length with these endpoints [10].
Let Aut(Ω) denote the group of projective transformations which map Ω onto itself. Following Benoist we say a discrete subgroup Γ ⊂ Aut(Ω) divides Ω if the quotient Q = Ω/Γ is compact. Since cross ratio is preserved by projective transformations, Aut(Ω) (and thus the subgroup Γ) acts on Ω by isometries of the Hilbert metric. It follows that the projective orbifold Q inherits a metric from d Ω called the Hilbert metric on Q. The holonomy provides an identification π orb 1 Q ≡ Γ. Given g ∈ Aut(Ω) the Hilbert translation length of g is Since Aut(Ω) acts by isometries it follows that conjugate elements have the same translation length. If Q = Ω/Γ is a strictly convex real projective orbifold we define the Hilbert translation length function : π orb 1 Q → R where (g) is the Hilbert length of g. If Q is a compact manifold then (g) is the length of the unique geodesic loop representative in the free homotopy class.
From now on we assume Ω is strictly convex. The geodesics in the Hilbert metric on Ω are then (subarcs of projective) lines. An isometry g ∈ Aut(Ω) is called hyperbolic if it preserves a line in Ω called the axis, and otherwise γ fixes some point in Ω and is called elliptic. When g is hyperbolic we can compute (g) from the eigenvalues of g.
has an attracting fixed point p in RP n ; algebraically this just means M has a unique eigenvalue of largest modulus, it has algebraic multiplicity 1, and is real. The matrix M is proximal if both M and M −1 are semi-proximal and, in addition, the eigenvalues of smallest and largest modulus have the same sign. Theorem 1.1 of [3] implies that hyperbolic isometries are proximal.
Lemma 2.8. Suppose g ∈ Aut(Ω) is hyperbolic and let M ∈ SL ± (n + 1, R) be the proximal matrix which represents g. Let λ + and λ − be the eigenvalues of M with largest and smallest absolute value. Then the Hilbert translation length of g is given by log(λ + /λ − ).
Proof. Let p ± be the attracting fixed points of g ±1 . These two points are in the boundary of Ω in RP n and correspond to the eigenvectors v ± ∈ R n+1 for the eigenvalues λ ± . Let A be the line segment in Ω with endpoints p ± . Then A is the axis of g. The Hilbert metric on the interior of A equals the restriction of the Hilbert metric on Ω to A. It follows that the translation length of g in Ω equals the translation length of the restricton of g to A, which is d A (q, gq) for any choice of a point q in the interior of A. This reduces the problem of computing the translation length of g to the case of RP 1 . The action of g on A is determined by the restriction of g to the invariant 2-dimensional subspace of R n+1 spanned by v ± . This action is given by a mobius transformation fixing two points. The Hilbert metric on A is the one dimensional hyperbolic metric, and an easy calculation (or see [2]) gives that the translation length of a Möbius transformation in this metric is as stated.
It follows from the next result that the notion of (projective) equivalence used in the definition of the space of strictly convex projective structures, C(Q), on a specified orbifold Q can be taken to be isometry of Hilbert metrics which preserve the marking. Proposition 2.9 (Isometric implies equivalent). If Q and Q are strictly convex projective orbifolds with isometric Hilbert metrics then they are projectively equivalent.
Proof. Let Q = Ω/Γ and Q = Ω /Γ . First we show that an isometry f : Q → Q is an orbifold isomorphism and therefore covered by a map h : Ω → Ω so the following diagram commutes.
Let ΣQ denote the singular set of Q and choose a small ball U ⊂ Q − ΣQ so that f (U ) ⊂ Q − ΣQ and the quotient map ρ : Ω → Ω/Γ restricted to a liftŨ of U is an isometry. Choose a liftŨ of f (U ). The map h = (ρ |Ũ ) −1 • f • ρ is an isometry fromŨ toŨ . Letx be a point inŨ andỹ = h(x). For any line throughx there is a corresponding line in Ω which passes throughỹ and contains h(Ũ ∩ ). An isometry between lines is determined by its restriction to an interval, so for every throughx, extend h isometrically from to . Since Ω is a convex ball in R n , the set of all lines −x foliate Ω −x. Therefore, we have a continuous extension of h to all of Ω that, when restricted to each line throughx, is an isometry.
Assuming this, it follows that h covers f and therefore, by definition, f is an orbifold isomorphism. Moreover, f is an isometry which implies the map h is an isometry of the Hilbert domains Ω and Ω . Since the domains are strictly convex, an isometry is the restriction of a projective transformation (see proposition 3 of [10]). Therefore the orbifolds Q and Q are projectively equivalent. It only remains to prove the claim.
We first discuss the local structure of the singular locus when there are reflections. Locally a projective orbifold is V /G where V is a open set in projective space and G is a finite group of projective transformations that preserve V. The singular locus in V /G is the image of the union of the subsets of fixed points of non-trivial elements in G. This is the intersection with V of a finite set of projective subspaces. The fixed subset of a projective transformation T has codimension 1 iff T is a reflection which means it is conjugate to ±diag(−1, 1, 1, · · · , 1). A mirror point in an orbifold is a point with local group of order two generated by a reflection. The mirror of an orbifold is the collection of all mirror points. The mirror is an open subset of the singular locus. If the mirror is not empty it has codimension 1. The singular locus minus the mirror has codimension at least 2.
A mirror geodesic in Q is a map : R → Q that is a composition = ρ •˜ wherẽ : R → Ω is a Hilbert geodesic such that is a path in Q which only intersects the singular locus at mirror points (where it bounces off the mirror like a light ray.) In addition we require that a mirror geodesic is not contained in the singular locus. By considering the local structure of V /G it follows that the set of singular points on a mirror geodesic is discrete and that mirror geodesics are dense in Q. In particular a geodesic which is disjoint from the singular locus is a mirror geodesic.
If γ is a mirror geodesic in Q then f • γ is a mirror geodesic in Q . The reason is that in a Hilbert domain there is a unique geodesic connecting any two points, and it is a segment of a projective line. Thus, away from the singular locus, mirror geodesics are sent to mirror geodesics by an isometry. This property also holds true near a mirror point x on a mirror geodesic γ. This is seen by working in an orbifold neighborhood V /G. Near x there is a small segment δ of γ containing x in its interior which is length minimizing among all paths with with the same endpoints as δ subject to the constraint that the path contains a mirror point.
Suppose is a line throughx such that ρ • is a mirror geodesic in Q and ρ • h • is mirror geodesic in Q . The set of all such is dense in Ω. Since ρ • is a mirror geodesic it follows from the above discussion that f • ρ • is a mirror geodesic. This mirror geodesic coincides with ρ • h • in U and therefore they are equal everywhere. It follows that the claim holds on the dense subset of Ω consisting of all such . By continuity the claim holds everywhere.

The Root Ratio polynomial
The main result of this section is 3.12 where we show the subset of SL ± (n, R) × SL ± (n, R) consisting of pairs of proximal matrices with the same Hilbert translation length is not Zariski dense. In view of 2.8, accomplishing this involves a discussion of ratios of pairs of roots of a polynomial.
be a degree n polynomial with roots {α i ∈ C : 1 ≤ i ≤ n } counted with multiplicity. A complex number r is a root ratio of p(x) if r = α i /α j , where α j is a non-zero root and i = j.
Let p(x) = x n + a 1 x n−1 + · · · + a n be a monic polynomial of degree n ≥ 2 in the ring C[x] with roots α i counted with multiplicity. The root ratio polynomial of p is: Proposition 3.2. If p is a polynomial of degree at least 2 and p(0) = 0 then the zeros of R p are exactly the root ratios of p.
Since the product of the roots of p(x) is (−1) n a n it follows that, R p (r) = a n−1 n r n(n−1) + lower order terms.
Furthermore, since the coefficients of R p are symmetric polynomials in the α i , they are polynomials in the coefficients a i .
By a general polynomial we mean a polynomial whose coefficients are independent transcendentals. Subsequently we may replace these by certain complex numbers. To this end we will let K denote a polynomial ring over C and consider polynomials in K[x] thought of as polynomials in the variable x with coefficients in K. A C-module homomorphism θ : K → C will be called a specialization. Given a specialization, there is a unique extension to a C-module homomorphism between the polynomial rings K[x] and C[x], which for convenience we also denote by θ. ), B(x); x) ∈ K is a polynomial in the m + n + 2 coefficients of these two polynomials obtained by eliminating x, see [13]. If

Given two general polynomials
Then From this one sees that the coefficients of the resultant are symmetric polynomials of the α i and the β j . The elementary symmetric functions generate the ring of all symmetric polynomials, therefore the terms in the resultant are polynomial functions of the coefficients of A(x) and B(x). The crucial property for us is: (1) there exists an x 0 ∈ C which is a root of both θ(A(x)) and θ(B(x)).
(2) the specialization reduces the degree of both A(x) and B(x).
See [8] for details on the above definition, and justifications of the above facts about the resultant. We are primarily interested in the case when A(x) and B(x) are both monic general polynomials, that is a polynomial with transcendental coefficients of the form x n + a 1 x n−1 + · · · + a n . In this case a specialization does not reduce the degree.
Resultants may be used to eliminate variables in systems of polynomial equations. This leads to a method to calculate the root ratio polynomial. It is straightforward from the definitions to show: Proposition 3.4. The root ratio polynomial, R p (r), of p(x) = x n +a 1 x n−1 +· · ·+a n satisifies Res(p(rx), p(x); x) = a n (r − 1) n R p (r).
Definition 3.5. Given two monic polynomials p(x) and q(x) the common root ratio polynomial is defined by C p,q = Res(R p (r), R q (r); r).
The coefficients of R p (r) are polynomial in the coefficients of p, therefore C p,q is polynomial in the coefficients of p and q. Example 3. The common root ratio polynomial for p(x) = x 2 + ax + b and q(x) = x 2 + cx + d is

THE MARKED LENGTH SPECTRUM OF A PROJECTIVE MANIFOLD OR ORBIFOLD. 9
The common root ratio polynomial for p(x) = x 3 + x 2 + x + 1 and q( The following is immediate: are monic and both have degree at least 2. If α ∈ C satisfies R p (α) = R q (α) = 0 then C p,q = 0.
Proposition 3.7. Suppose p(x), q(x) ∈ K[x] are monic polynomials and C p,q ∈ K is their common root ratio polynomial. Suppose θ : K → C is a specialization and θ(C p,q ) = 0. Then one of the following occurs (a) there is r ∈ C \ 0 which is a root ratio of both θ(p(x)) and θ(q(x)).
Proof. Applying 3.3 to θ(C p,q ) = θ(Res[R p (r), R q (r); r]) yields two possibilities. The first possibility, case (1), is that there is r ∈ C such that θ(R p (r)) = 0 = θ(R q (r)). Since p is monic, the specialization does not reduce the degree of p(x), and θ(R p ) = R θ(p) . Similarly for q(x), therefore R θ(p) (r) = R θ(q) (r) = 0. Case (b) results if either θ(p(x)) or θ(q(x)) has a root of 0. The remaining possibility is case (2) of 3.3; namely that the specialization reduces the r-degree of both R p (r) and R q (r). Since p is monic, the coefficient of the highest power of r in R p (r) is a n−1 n where a n = p(0) and n = deg x (p(x)). Therefore specialization reduces the r-degree of R p (r) iff θ(p(0)) = 0 which implies conclusion (b).
Definition 3.8. Given a matrix M ∈ GL(n, C) let λ 1 , · · · , λ n be the eigenvalues of M counted with multiplicity. For 1 ≤ i = j ≤ n the numbers λ i /λ j are called the eigenvalue ratios of M.
Note that for any non-zero complex number k, the matrices M and kM have the same eigenvalue ratios, hence eigenvalue ratio is well-defined for elements of P GL(n, C). Lemma 3.9. Let P be the characteristic polynomial of a matrix M ∈ GL(n, C) with n ≥ 2. Then the set of eigenvalue ratios of M equals the set of zeroes of the root-ratio polynomial R P .
Proof. The characteristic polynomial of P is monic and zero is not a root. The result now follows from 3.2. Proof. Since SL ± (n, R) is defined by the polynomial equation (det M ) 2 = 1 it follows that SL ± (n, R) × SL ± (n, R) is a real algebraic variety. We will show A is a subvariety by exhibiting it as the zero locus of one additional polynomial. It is clear that A = SL ± (n, R) × SL ± (n, R) thus the subvariety is proper.
Let P (x) and Q(x) be the characteristic polynomials of M and N . The common root ratio polynomial, C P,Q , is a polynomial function of the entries of M and N . Indeed, since P and Q are monic, the common root-ratio polynomial C P,Q is a polynomial in the coefficients of P and Q. These, in turn, are polynomial functions of the entries of M and N .
We claim that A is the subvariety V ⊂ SL ± (n, R)×SL ± (n, R) defined by C P,Q = 0. First assume that (M, N ) ∈ V. We apply proposition 3.7 to C P,Q . Since M and N are in SL ± (n, R) it follows that P and Q have constant terms ± det(M ) and ± det(N ) that are non-zero, therefore case (b) does not occur. It now follows from case (a) that P (x) and Q(x) have a common root-ratio, which in turn implies the matrices M and N have a common eigenvalue ratio, thus V ⊂ A.
Conversely, suppose (M, N ) ∈ A, then M and N have a common eigenvalue ratio α. By 3.9 it follows that α is a zero of the root-ratio polynomials R P and R Q . Proposition 3.6 implies that C P,Q is zero at (M, N ). Thus (M, N ) ∈ V and A ⊂ V.  Proof. The Hilbert length of a proximal matrix is determined by the ratio of the largest and smallest eigenvalue by 2.8. Thus the variety A from proposition 3.10 contains Q. There are pairs of elements in SO(n, 1) with no common eigenvalue ratios, hence A does not contain SO(n, 1) × SO(n, 1).

Dual and Self-Dual Projective Structures
In this section we define the notion of dual projective structure and show that a structure is self-dual iff it is a hyperbolic structure.
An ellipsoid in RP n is the image under a projective map of the open unit ball, B, in an affine patch. The subgroup of P GL(n + 1, R) stabilizing B is Aut(B) ∼ = P O(n, 1). The Hilbert metric on B is isometric to hyperbolic space H n and if Γ is a discrete subgroup of Aut(B) then B/Γ is a (real) hyperbolic orbifold. Theorem 4.1 (Benoist,Thm 3.6 [3]). Let Γ be a discrete subgroup of SL(n + 1, R) that divides a properly convex open set Ω in RP n . The Zariski closure Γ is SO(n, 1) iff Ω is an ellipsoid and otherwise Γ = SL(n + 1, R).
To allow for non-orientable orbifolds, we need the following corollary to Benoist's Theorem.
Proof. Let Γ + = Γ ∩ SL(n + 1, R). Then Γ + is at most an index two subgroup of Γ, and therefore divides Ω. By 4.1, the Zariski closure of Γ + is SO(n, 1) when Ω is an ellipsoid and SL(n + 1, R) otherwise. When Ω is an ellipsoid the full stabilizer of Ω in SL ± (n + 1, R) is O(n, 1). Note that SO(n, 1) and SL(n + 1, R) are respectively index two subgroups of O(n, 1) and SL ± (n + 1, R). Since the Zariski closure of a subgroup of an algebraic group is an algebraic subgroup (pg. 99 of [16]), the result follows.
Choose an inner product , on R n+1 . Let Ω ⊂ RP n be a strictly convex open set. Let Ω + ⊂ S n be one of the pre-images of Ω. Then { v ∈ R n+1 : ∀x ∈ Ω + v, x > 0 } is a convex set and its image in RP n is called the dual of Ω and is denoted by Ω * . A different choice of inner product results in a projectively equivalent domain. For convenience we will use the standard inner product.
It is immediate that if A ∈ P GL(n + 1, R) preserves Ω then the transpose A t preserves Ω * . The map d : P GL(n + 1, R) → P GL(n + 1, R) given by d(A) = (A t ) −1 is called the global Cartan involution or the dual map and d(A) is the dual matrix. It is an isomorphism and a different choice of inner product changes d by composition with an inner automorphism.

Proposition 4.3.
There is an orbifold isomorphism f : Q → Q * such that Proof. The duality map provides an isomorphism π orb 1 Q → π orb 1 Q * , however we must show they are orbifold isomorphic (if they are manifolds this means diffeomorphic). There is a natural diffeomorphism f : Ω → Ω * defined using affine spheres, see for example [14]. Naturality ensures that it is equivariant with respect to the group actions.
It follows that duality induces an involution on C(Q) the space of projective structures on an orbifold Q. A projective structure is called self-dual if it is projectively equivalent to is dual. Thus the equivalence classes of the self-dual structures are just the fixed points of this involution. Proof. It is easy to see that hyperbolic orbifolds are self-dual. For the converse, let Q = Ω/Γ be a self-dual projective orbifold with dual Q * = Ω * /Γ * . We may assume Γ, Γ * ⊂ SL ± (n + 1, R). Self-duality implies there is an element in P ∈ SL(n + 1, R) such that for all We claim that Γ + = Γ ∩ SL(n + 1, R) is not Zariski dense in SL(n + 1, R). Indeed, this equation implies f (A) = trace(A) − trace(A −1 ) is zero on Γ + and this is a non-trivial polynomial condition on SL(n + 1, R) provided n ≥ 2. This proves the claim. We now apply theorem 4.1 to deduce that Ω is an ellipsoid. An ellipsoid can be mapped to the unit ball by a projective map, thus Γ is conjugate by this map to a group which stabilizes the unit ball. The stabilizer of the unit ball is O(n, 1) and therefore Q is equivalent to a hyperbolic orbifold.
Projective duality is due to the presence of the duality map on SL(n + 1, R). Since we have been unable to find a reference for the following theorem, we provide a proof.  SL(n, C)). For all n ≥ 2 the group of holomorphic outer automorphisms of SL(n, C) has order 2 and is generated by the duality map, d, that sends A → (A t ) −1 . If σ is an automorphism of SL(n, C) then either σ = φ g where φ g is some inner automorphism, or σ = φ g • d.
Proof. Let φ be a holomorphic automorphism of SL(n, C). Differentiating φ induces an automorphism on the tangent space at the identity, the Lie algebra sl(n, C). Furthermore, since SL(n, C) is connected, the differential map D : Aut(SL(n, C)) → Aut(sl(n, C)) is injective ( §2 Theorem 4 [16]). The inner automorphisms of SL(n, C) map onto the inner automorphisms, called the inner derivations, of sl(n, C).
For a complex Lie algebra G, the group Inn(G) forms a normal subgroup of the group of automorphisms, and Aut(G)/Inn(G) is isomorphic to the automorphisms group of the Dynkin diagram for G (pg. 202 [16]). In the case of sl(n, C), the Dynkin diagram is A n , which is a linear graph with n vertices and therefore has one nontrivial automorphism. Since the map Aut(SL(n, C)) → Aut(sl(n, C)) is injective, Aut(SL(n, C))/Inn(SL(n, C)) has at most two elements. The map A → (A T ) −1 , after differentiating, induces the non-trivial automorphism of the dynkin diagram.
Recall that GL(n, C) is a connected Lie group and therefore P GL(n, C) is also connected. The natural map SL(n, C) → P GL(n, C) is an n-fold cyclic covering: the kernel is the center of SL(n, C), which is isomorphic to the cyclic subgroup of order n in C generated by a primitive n'th root of unity.
Corollary 4.6 (Automorphisms of P GL(n, C)). For all n ≥ 2 the group of holomorphic outer automorphisms Out(P GL(n, C)) ∼ = Z 2 is generated by the duality map.
Proof. The group SL(n, C) is a finite cover of P GL(n, C) and therefore the Lie algebras of P GL(n, C) and SL(n, C) are isomorphic. Arguing as in 4.5 shows that Out(P GL(n, C)) is cyclic of order two, generated by the duality map.
We record some facts about normal subgroups in Lie groups in the following lemma.
Lemma 4.7. Every nontrivial normal subgroup of P GL(n, C), P O(n, C), or P SO(n, C) contains P SO(n, C).
The group O(n, C) has two connected components, the one containing the identity is SO(n, C) (pg. 43 [16]). Since O(n, C) is a simple Lie group with finite center, P O(n, C) is a centerless simple Lie group with identity component equal to P SO(n, C). The only proper normal subgroups in simple Lie groups are discrete. By [16] pg. 45, any discrete normal subgroup of a connected Lie group is contained in the center. Therefore, in all three cases, any nontrivial normal subgroup contains P SO(n, C).

Proof of Main Theorem
The main theorem of this paper is that the Hilbert translation length function determines a strictly convex real projective structure up to projective duality. We now state the main theorem more precisely.
Suppose Q is a closed strictly convex projective orbifold. The translation length function defined in section 2 is a map : π orb 1 Q → R. An element α ∈ π orb 1 Q has finite order iff it is elliptic iff it has translation length zero. Otherwise the holonomy of α is a proximal element [A] ∈ P GL(n + 1, R) and (α) = log (λ + /λ − ) where λ ± are the eigenvalues of largest and smallest absolute value for any representative matrix A.
be the map which sends a strictly convex projective structure to its Hilbert translation length function. For any x ∈ C(Q), we have L −1 (L(x)) = { x, x * }, where x * denotes the dual of x. Moreover x = x * iff x is projectively self dual iff x is a hyperbolic structure.
Proof. For i = 1, 2 suppose that x i ∈ C(Q) are two strictly convex projective structures on Q. These determine two holonomy representations r i : π orb 1 Q → Γ i ⊂ P GL(n + 1, R) up to conjugacy in P GL(n + 1, R).
We will show that there is an automorphism φ of P GL(n + 1, R) such that r 2 = φ • r 1 . By 4.5 it follows that r 2 is conjugate to either r 1 or to the dual d • r 1 . By 2.3 the conjugacy class of the holonomy determines the projective structure and the result then follows.
By 2.1 we can always lift the holonomy to SL ± (n+1, R). To prevent proliferation of notation, for the next few paragraphs we shall use r i for the lifted holonomy and Γ i for its image.
Since Q is a compact orbifold, there are only finitely many different orders of torsion elements. We define k to be the least common multiple of these orders. It follows that for every g ∈ π orb 1 Q that (r i (g)) k is either I or proximal. Let P ⊂ SL ± (n + 1, R) × SL ± (n + 1, R) be the proximal pair variety defined in 3.11. Then P contains all pairs of proximal matrices with the same translation length. It also contains (I, I). The map τ : SL ± (n + 1, R) × SL ± (n + 1, R) → SL ± (n + 1, R) × SL ± (n + 1, R) given by τ (M, N ) = (M k , N k ) is a polynomial map therefore A = τ −1 (P ) is an affine algebraic variety.
Given g ∈ π orb 1 Q the elements r 1 (g) and r 2 (g) have the same translation length. If the translation length is zero then g is torsion, so r 1 (g) k and r 2 (g) k are both the identity. Otherwise, lemma 2.8 implies that r 1 (g) and r 2 (g) have a common eigenvalue ratio r and therefore r 1 (g) k and r 2 (g) k have a common eigenvalue ratio of r k . In either case, (r 1 (g), r 2 (g)) ∈ A. Thus Im(r 1 × r 2 ) = {(r 1 (g), r 2 (g)) | g ∈ π orb 1 Q} ⊆ A ⊆ SL ± (n + 1, R) × SL ± (n + 1, R). Algebraic geometry is easier over C than over R. For this reason we will now complexify everything. The complexification of a real affine algebraic variety, V ⊂ R n , is the complex affine variety V C ⊂ C n obtained by taking the complex zeros of the real polynomials defining V. Observe that V = R n ∩ V C .
Define D ⊂ SL ± (n + 1, C) × SL ± (n + 1, C) to be the Zariski closure over C of Im(r 1 ×r 2 ). Since D ⊂ A C it is a proper subvariety of SL ± (n+1, C)×SL ± (n+1, C). Let Γ i be the Zariski closure over C of Γ i in SL ± (n + 1, C). By corollary 4.2, Γ i is either conjugate to SO(n, 1, C) or O(n, 1, C), or else equals SL(n + 1, C) or SL ± (n + 1, C). In the former cases, these groups are isomorphic to SO(n + 1, C) and O(n + 1, C), and we adjust by a conjugacy so that Γ i = SO(n + 1, C) or O(n + 1, C). Henceforth, in all cases we may assume that SO(n + 1, C) ⊆ Γ i .
The Zariski closure of a subgroup of an algebraic group is an algebraic subgroup (Lemma 2.2.4 of [19]) therefore D is a complex algebraic subgroup. For i = 1, 2 let p i : D → SL ± (n+1, C) be the projection map onto the i'th coordinate. By (theorem 3, page 102 of [16]) the image under an algebraic homomorphism of a complex algebraic group is also a complex algebraic group. It follows that G i = p i (D) is a complex algebraic subgroup of SL ± (n + 1, C). Now G i contains Γ i and G i is algebraic so Γ i ⊆ G i . We claim that This is because p −1 i (Γ i ) is a complex algebraic group which contains Im(r 1 × r 2 ), thus D ⊂ p −1 i (Γ i ) and G i = p i (D) ⊆ p i (p −1 i (Γ i )) = Γ i . Now we projectivize everything. Let P : SL ± (n + 1, C) → P GL(n + 1, C) be the natural projection. Define P D = P × P )(D) ⊂ P GL(n + 1, C) × P GL(n + 1, C).
We have the two coordinate projections π i : P D → P GL(n + 1, C). Define P G i = π i (P D) = P (G i ) = P (Γ i ) ⊂ P GL(n + 1, C).
We have shown that P G i is either P GL(n + 1, C), P O(n + 1, C), or P SO(n + 1, C).
If G 1 = P GL(n + 1, C) then φ is either an inner automorphism, or the composition of the duality map and an inner automorphism by 4.6. The assertion of the theorem involves P GL(n + 1, R) rather than P GL(n + 1, C). We need to prove the inner automorphism of P GL(n + 1, C) actually extends an inner automorphism of P GL(n + 1, R).
Suppose an element A ∈ GL(n + 1, C) conjugates Γ 1 to Γ 2 . Since the latter subgroups are real A may be chosen to have real entries. Then we obtain an inner automorphism of P GL(n + 1, R). A similar remark applies in the case of the dual map composed with an inner automorphism. In either case, since Γ i ⊂ G i , restriction gives r 2 = φ • r 1 .
Corollary 5.3. There exist two diffeomorphic strictly convex real projective surfaces M and N with equivalent marked length spectra that are not isometric.
Proof. Goldman shows that the dimension of the deformation space of strictly convex projective structures on a surface S is −8χ(S). This implies that there exists non-hyperbolic strictly convex projective structures. Together theorems 4.4 and 5.1 imply there are non-equivalent projective structures on S with the same marked length spectra. The corollary then follows from proposition 2.9.
As an example consider the 2-orbifold Q = S 2 (3, 3, 4). According to Goldman and Choi, C(Q) ∼ = R 2 . There is a unique hyperbolic structure on this orbifold, which we may assume is 0 ∈ R 2 , and the duality map induces an involution on C(Q) which is a rotation of order 2 fixing 0. All the non-zero points correspond to projective structures that are not self dual.
A closed hyperbolic n-manifold which contains an embedded, totally geodesic codimension-1 submanifold admits a 1 parameter family of strictly convex projective structures obtained by a projective version of the bending construction. Thus isospectral examples also exist in dimensions 3 and 4.