Introduction
One of the guiding problems in the study of properly convex projective manifolds is to understand how the geometry of their developing domains constrains the topology of compact quotients. In the divisible strictly convex case, the Hilbert geometry is Gromov hyperbolic and many features of negatively curved manifolds persist ([4 ] ). General divisible properly convex domains, however, can have nontrivial boundary segments and properly embedded simplices. In particular, neither negative sectional curvature nor hyperbolicity of the fundamental group is available in general.
Recall that a properly convex domain Ω ⊂ R P n \Omega\subset\mathbb{RP}^{n} Ω ⊂ RP n is an open convex subset that is bounded in some affine chart. It is divisible if a discrete subgroup Γ < Aut ( Ω ) \Gamma< \operatorname{Aut}(\Omega) Γ < Aut ( Ω ) acts cocompactly, and a closed properly convex projective manifold is a quotient M = Γ \ Ω M = \Gamma\backslash\Omega M = Γ\Ω with Γ \Gamma Γ torsion-free. The Hilbert metric on Ω \Omega Ω is a complete projectively invariant Finsler metric. Its geodesics include the ordinary projective segments, although uniqueness of geodesics can fail when the boundary contains segments.
For an oriented closed n n n -manifold, its simplicial volume is the ℓ 1 \ell^{1} ℓ 1 -seminorm of the real fundamental class,
∥ M ∥ = inf { ∑ j ∣ a j ∣ ∣ ∑ j a j σ j represents [ M ] } . \lVert M\rVert= \inf\left\{ \sum_j \lvert a_j\rvert\mathrel{\bigg|} \sum_j a_j\sigma_j \text{ represents } [M] \right\}. ∥ M ∥ = inf { j ∑ ∣ a j ∣ j ∑ a j σ j represents [ M ] } .
In hyperbolic geometry, positivity of ∥ M ∥ \lVert M\rVert ∥ M ∥ follows by integrating the volume form over straight simplices, whose volumes have a uniform upper bound ([12 ] ). This produces a bounded cocycle with positive evaluation on the fundamental class, which is the geometric paradigm underlying Gromov’s theory ([11 ] ). We establish an analogous bound for projective simplices in Ω \Omega Ω , measured using the Holmes–Thompson volume of the Hilbert metric.
Let C ⊂ R n + 1 C \subset\mathbb{R}^{n+1} C ⊂ R n + 1 be either of the two proper open cones projecting to Ω \Omega Ω . We say that Ω \Omega Ω is irreducible when there are no nonzero complementary subspaces L 1 , L 2 L_1,L_2 L 1 , L 2 and proper open convex cones C i ⊂ L i C_i \subset L_i C i ⊂ L i with C = C 1 × C 2 C=C_1\times C_2 C = C 1 × C 2 . Equivalently, Ω \Omega Ω is not a nontrivial projective join. This is the same irreducibility convention used in the rigidity theory of divisible properly convex domains (also called indecomposability in the sense of Vey and Benoist); it should not be confused with any connected-sum decomposition of M M M . For a divisible domain, Vey’s irreducibility theorem says that this condition implies that every finite-index subgroup of a dividing group Γ \Gamma Γ acts irreducibly on R n + 1 \mathbb{R}^{n+1} R n + 1 ; see [4 Theorem 5.1] .
Theorem A. Let Ω ⊂ R P n \Omega\subset\mathbb{RP}^n Ω ⊂ RP n , n ≥ 2 n \ge2 n ≥ 2 , be an irreducible divisible properly convex domain. There exists V Ω < ∞ V_\Omega<\infty V Ω < ∞ such that
Vol H T , Ω ( σ ) ≤ V Ω \operatorname{Vol}_{\mathrm{HT},\Omega}(\sigma) \le V_\Omega Vol HT , Ω ( σ ) ≤ V Ω for every (possibly degenerate) projective n n n -simplex σ \sigma σ with vertices in Ω \Omega Ω . The volume is the Holmes–Thompson volume of the Hilbert metric of Ω \Omega Ω .
The assertion concerns simplices in the universal cover. In particular, their vertices can approach arbitrary faces of ∂ Ω \partial\Omega ∂ Ω and their projections to a compact quotient need not be embedded. The natural Blaschke volume of the canonical affine sphere is comparable with Holmes–Thompson volume on a compact quotient. Equivariant straightening of singular simplices therefore gives the principal topological applications.
Corollary B. Every closed properly convex real projective manifold of dimension n ≥ 2 n \ge2 n ≥ 2 whose projective domain is irreducible has positive simplicial volume.
Corollary C. Let M = Γ \ Ω M=\Gamma\backslash\Omega M = Γ\Ω be a closed properly convex real projective manifold of dimension n ≥ 2 n \ge2 n ≥ 2 . Then ∥ M ∥ > 0 \lVert M\rVert>0 ∥ M ∥ > 0 if and only if Ω \Omega Ω is irreducible.
The main ingredient towards the proof of Theorem A is a characterization of irreducible divisible domains Ω \Omega Ω in terms of the geometry of the associated affine sphere asymptotic to C C C . This carries the Blaschke metric h h h , which Calabi proved to have nonpositive Ricci tensor [7 Theorem 5.1] . We give a proof of this inequality in the cocompact setting needed here in Proposition 2.2 and improve this estimate in the case of existence of directions annihilating the Ricci tensor.
Theorem D. Let Ω ⊂ R P n \Omega\subset\mathbb{RP}^n Ω ⊂ RP n , n ≥ 2 n \ge2 n ≥ 2 , be a divisible properly convex domain, C C C its positive cone, and h h h the Blaschke metric of the canonical affine sphere, normalized to affine mean curvature − 1 -1 − 1 . If
Ric h , p ( v , v ) = 0 for some p ∈ Ω and 0 ≠ v ∈ T p Ω , \operatorname{Ric}_{h,p}(v,v)=0 \qquad\text{for some }p\in\Omega\text{ and }0\ne v\in T_p\Omega, Ric h , p ( v , v ) = 0 for some p ∈ Ω and 0 = v ∈ T p Ω , then C C C is a nontrivial direct product of proper open convex cones. Consequently, if C C C is irreducible, there is ε > 0 \varepsilon>0 ε > 0 such that Ric h ≤ − ε h \operatorname{Ric}_h\le-\varepsilon h Ric h ≤ − ε h throughout $\Omega.
The two-dimensional equality case is known: see for example Li–Li–Simon [18 ] . A similar argument appears in Hu–Li–Vranc̆ken [14 ] who give characterizations of Calabi products under algebraic conditions on the cubic tensor. Section 2 shows that the null distribution is parallel and produces an ambient one-parameter automorphism whose eigenspaces split C C C .
Related work. The relation between the Euler characteristic and simplicial volume has recently advanced substantially. In [BFR26], Bucher–Francaviglia–Rago construct bounded angle cocycles using immersed straightenings. In particular, their Theorem 1.2 proves that for any closed properly convex real projective d d d -manifold there is a constant c d > 0 c_d>0 c d > 0 , with c d ≤ ( d + 2 ) / 4 c_d\le(d+2)/4 c d ≤ ( d + 2 ) /4 , such that
∣ χ ( M ) ∣ ≤ c d ∥ M ∥ . (1) |\chi(M)|\le c_d\lVert M\rVert. \tag*{(1)} ∣ χ ( M ) ∣ ≤ c d ∥ M ∥ . (1)
Thus Gromov’s question whether vanishing simplicial volume forces vanishing Euler characteristic has an affirmative answer within the class of closed properly convex projective manifolds. This result should not be conflated with positivity of simplicial volume itself: (0.1) gives no information when χ ( M ) = 0 \chi(M)=0 χ ( M ) = 0 . Indeed, Bucher–Francaviglia–Rago explicitly ask in their Question 1.4 whether closed convex real projective manifolds with irreducible universal cover necessarily have positive simplicial volume. Corollary B is an affirmative answer to precisely that question. Our method is different from the angle-cocycle approach. Rather than representing the Euler class by a bounded cocycle, we produce a uniform volume bound for projective simplices and hence a bounded representative of the volume class .
Several previously known positivity results provide useful points of comparison. For an aspherical closed manifold with word-hyperbolic fundamental group, positivity follows from bounded-cohomological results of Mineyev [20 ] . For closed locally symmetric spaces of noncompact type, positivity was established by Lafont–Schmidt [17 ] , with the relevant exceptional case treated by Bucher–Karlsson [5 ] . These cover strictly convex divisible domains and irreducible symmetric divisible domains, respectively, but not the full range of irreducible properly convex domains containing boundary segments.
Positivity is also known for closed aspherical manifolds of dimension at least two whose fundamental groups are non-elementary relatively hyperbolic. This follows from the relative bounded cohomology theorem of Mineyev–Yaman, for which Franceschini gives another proof [9 Theorem 1.1] . In convex projective geometry, this applies to certain non-strictly convex divisible domains whose dividing groups are relatively hyperbolic with respect to stabilizers of maximal properly embedded simplices, as well as to examples with peripheral subgroups arising from properly embedded cones; see [BFR26, Section 1] and the references therein. Corollary B establishes positivity for all irreducible divisible domains, without a relative hyperbolicity hypothesis.
There are also geometric examples in dimension four worth emphasizing. Riolo–Seppi–Slavich [21 ] construct closed indecomposable convex projective four-manifolds with positive Euler characteristic by gluing hyperbolic pieces. Their simplicial volume is known to be positive by the corresponding decomposition results, as noted in [BFR26]. Our theorem is not restricted to such decompositions.
Strategy of the proof. An elementary example illustrates the obstruction that our proof must exclude. Consider the open projective triangle
Ω t r i = { [ x : y : 1 ] ∈ R P 2 : x , y > 0 } . \Omega_{\mathrm{tri}}=\{[x:y:1]\in\mathbb{R}\mathbb{P}^{2}:x,y>0\}. Ω tri = {[ x : y : 1 ] ∈ R P 2 : x , y > 0 } .
Positive dilations of the two affine coordinates are projective automorphisms of this domain. Invariance under these dilations implies that its Holmes–Thompson measure has the form
d Vol H T , Ω t r i = c H T d x d y x y , c H T > 0. d\operatorname{Vol}_{\mathrm{HT},\Omega_{\mathrm{tri}}}=c_{\mathrm{HT}}\frac{dx\,dy}{xy},\qquad c_{\mathrm{HT}}>0. d Vol HT , Ω tri = c HT x y d x d y , c HT > 0.
In particular, this measure is a constant multiple of Euclidean area in the coordinates ( log x , log y ) (\log x,\log y) ( log x , log y ) . The transformations ( x , y ) ↦ ( e x , y ) (x,y)\mapsto(e x,y) ( x , y ) ↦ ( e x , y ) and ( x , y ) ↦ ( x , e y ) (x,y)\mapsto(x,e y) ( x , y ) ↦ ( x , ey ) generate a group dividing Ω t r i \Omega_{\mathrm{tri}} Ω tri : in logarithmic coordinates, their action is simply translation by Z 2 \mathbb{Z}^{2} Z 2 . Let γ ( x , y ) = ( e x , y ) \gamma(x,y)=(e x,y) γ ( x , y ) = ( e x , y ) , and consider the projective triangles
σ m = conv { ( 1 , 1 ) , ( 1 , 3 ) , ( e m , 1 ) } , m ≥ 1. \sigma_m=\operatorname{conv}\{(1,1),(1,3),(e^m,1)\},\qquad m\geq1. σ m = conv {( 1 , 1 ) , ( 1 , 3 ) , ( e m , 1 )} , m ≥ 1.
Here we keep the first two vertices fixed and move the third along the orbit γ m ( 1 , 1 ) \gamma^m(1,1) γ m ( 1 , 1 ) . Each triangle contains the rectangle
[ 1 , 1 + e m 2 ] × [ 1 , 2 ] . \left[1,\frac{1+e^m}{2}\right]\times[1,2]. [ 1 , 2 1 + e m ] × [ 1 , 2 ] .
Consequently,
Vol H T , Ω t r i ( σ m ) ≥ c H T log 2 log ( 1 + e m 2 ) ⟶ ∞ . \operatorname{Vol}_{\mathrm{HT},\Omega_{\mathrm{tri}}}(\sigma_m)\geq c_{\mathrm{HT}}\log2\log\left(\frac{1+e^m}{2}\right)\longrightarrow\infty. Vol HT , Ω tri ( σ m ) ≥ c HT log 2 log ( 2 1 + e m ) ⟶ ∞.
Geometrically, these triangles contain a strip whose width in logarithmic coordinates stays fixed while its length tends to infinity.
The singular values of the transformations producing this growth have a particularly simple form. The determinant-one representative of γ m \gamma^m γ m is
diag ( e 2 m / 3 , e − m / 3 , e − m / 3 ) . \operatorname{diag}\left(e^{2m/3},e^{-m/3},e^{-m/3}\right). diag ( e 2 m /3 , e − m /3 , e − m /3 ) .
There are only two distinct singular values: one direction expands relative to the other two, which are rescaled equally. Thus the volume growth occurs along a sequence for which the total distortion, i.e. the difference between the logarithms of the top and bottom singular values, grows without bound, while the variation within each of the two groups of singular values remains zero. Our proof obtains a quantitative obstruction to this behaviour for irreducible divisible domains and uses it to control the volumes of simplices.
We now give some details on how this is achieved. First, we strengthen Calabi’s nonpositivity of the Blaschke Ricci tensor: we show that if the Ricci tensor vanishes on a single nonzero vector then the cone splits as a nontrivial product. Irreducibility excludes this possibility, and cocompactness then gives a uniform estimate R i c h ≤ − ε h \mathrm{Ric}_h\leq-\varepsilon h Ric h ≤ − ε h for some ε > 0 \varepsilon>0 ε > 0 .
The Gauss map of the hyperbolic affine sphere converts this curvature bound into a restriction on the singular values of cone automorphisms. Put N = n + 1 N=n+1 N = n + 1 , and let μ 1 ( g ) ≥ ⋯ ≥ μ N ( g ) \mu_1(g)\geq\cdots\geq\mu_N(g) μ 1 ( g ) ≥ ⋯ ≥ μ N ( g ) be the logarithms of the singular values of a representative normalized by ∣ det g ∣ = 1 |\det g|=1 ∣ det g ∣ = 1 . Define
R ( μ ( g ) ) = μ 1 ( g ) − μ N ( g ) , R(\mu(g))=\mu_1(g)-\mu_N(g), R ( μ ( g )) = μ 1 ( g ) − μ N ( g ) ,
D k ( μ ( g ) ) = ∑ i ≤ k ( μ 1 ( g ) − μ i ( g ) ) + ∑ i > k ( μ i ( g ) − μ N ( g ) ) , 1 ≤ k < N . D_k(\mu(g))=\sum_{i\leq k}\left(\mu_1(g)-\mu_i(g)\right)+\sum_{i>k}\left(\mu_i(g)-\mu_N(g)\right),\qquad1\leq k<N. D k ( μ ( g )) = i ≤ k ∑ ( μ 1 ( g ) − μ i ( g ) ) + i > k ∑ ( μ i ( g ) − μ N ( g ) ) , 1 ≤ k < N .
Here R R R measures the total spread of the logarithmic singular values, while D k D_k D k measures their variation within the first k k k entries and within the remaining entries. In particular, D k D_k D k vanishes exactly when the entries are constant within each of these two groups. Such patterns arise naturally on a product cone, where one can rescale the two factors independently. In the example above we had R ( μ ( γ m ) ) = m R(\mu(\gamma^m))=m R ( μ ( γ m )) = m and D 1 ( μ ( γ m ) ) = 0 D_1(\mu(\gamma^m))=0 D 1 ( μ ( γ m )) = 0 , thus those are the quantities that we need to control in order to exclude the phenomenon described in the previous example. We achieve this by proving the quantitative estimate
D k ( μ ( γ ) ) ≥ δ R ( μ ( γ ) ) − c g a p , ∀ γ ∈ Γ , 1 ≤ k < N . (2) D_k(\mu(\gamma))\geq\delta R(\mu(\gamma))-c_{\mathrm{gap}},\qquad\forall\gamma\in\Gamma,\quad1\leq k<N. \tag*{(2)} D k ( μ ( γ )) ≥ δ R ( μ ( γ )) − c gap , ∀ γ ∈ Γ , 1 ≤ k < N . (2)
for some δ > 0 \delta>0 δ > 0 . This is a consequence of the relation between the Ricci curvature and the displacement of a fixed point in the symmetric space of positive definite quadratic forms. The estimate in particular says that as the distortion of a group element grows, its singular values cannot become concentrated in two groups with nearly equal values within each group.
As illustrated in the example above, this controls the volume of simplices. Consider a projective image D \mathcal{D} D of Ω \Omega Ω contained in the positive orthant, equipped with the measure
d y 1 ⋯ d y n y 1 ⋯ y n = d ( log y 1 ) ⋯ d ( log y n ) . \frac{dy_1\cdots dy_n}{y_1\cdots y_n}=d(\log y_1)\cdots d(\log y_n). y 1 ⋯ y n d y 1 ⋯ d y n = d ( log y 1 ) ⋯ d ( log y n ) .
Our intermediate goal is to bound its total mass uniformly over all such projective images. In logarithmic coordinates, this is an ordinary Euclidean volume problem. Suppose that D \mathcal{D} D contains boxes centred at a a a and b b b , with half-widths r a i ra_i r a i and r b i rb_i r b i in the i i i -th coordinate, where 0 < r ≤ 1 / 4 0<r\leq1/4 0 < r ≤ 1/4 . Taking convex combinations of these boxes shows that, for every 0 ≤ λ ≤ 1 0\leq\lambda\leq1 0 ≤ λ ≤ 1 , the domain D \mathcal{D} D contains the box centred at
p λ = ( 1 − λ ) a + λ b p^\lambda=(1-\lambda)a+\lambda b p λ = ( 1 − λ ) a + λb
with half-widths r p i λ rp_i^\lambda r p i λ . Dividing the i i i -th coordinate by p i λ p_i^\lambda p i λ sends this box to the fixed box [ 1 − r , 1 + r ] n [1-r,1+r]^n [ 1 − r , 1 + r ] n . We compare these coordinate changes as the centre moves along the segment. Write λ = e − s \lambda=e^{-s} λ = e − s and put ξ i = log ( b i / a i ) \xi_i=\log(b_i/a_i) ξ i = log ( b i / a i ) . Then
p i ( s ) = ( 1 − e − s ) a i + e − s b i . p_i(s)=(1-e^{-s})a_i+e^{-s}b_i. p i ( s ) = ( 1 − e − s ) a i + e − s b i .
For s ≥ log 2 s \ge\log2 s ≥ log 2 , the factor 1 − e − s 1-e^{-s} 1 − e − s lies between 1 / 2 1/2 1/2 and 1 1 1 . Thus the size of p i ( s ) p_i(s) p i ( s ) is determined, up to a fixed factor, by the larger of a i a_i a i and e − s b i e^{-s}b_i e − s b i . Equivalently,
log p i ( s ) = { log b i − s + O ( 1 ) , s < ξ i , log a i + O ( 1 ) , s ≥ ξ i , \log p_i(s)=
\begin{cases}
\log b_i-s+O(1), & s<\xi_i,\\
\log a_i+O(1), & s\ge\xi_i,
\end{cases} log p i ( s ) = { log b i − s + O ( 1 ) , log a i + O ( 1 ) , s < ξ i , s ≥ ξ i ,
where the errors are bounded by log 2 \log2 log 2 . Each coordinate therefore has a simple model: its logarithm decreases with slope − 1 -1 − 1 until s = ξ i s=\xi_i s = ξ i , and is constant afterwards.
Consider an interval [ s 1 , s 2 ] [s_1,s_2] [ s 1 , s 2 ] containing no transition value ξ i \xi_i ξ i in its interior, and set τ = s 2 − s 1 \tau=s_2-s_1 τ = s 2 − s 1 . Suppose that k k k coordinates are still in the decreasing part of this model. For those coordinates, p i ( s 1 ) / p i ( s 2 ) = e τ + O ( 1 ) p_i(s_1)/p_i(s_2)=e^{\tau+O(1)} p i ( s 1 ) / p i ( s 2 ) = e τ + O ( 1 ) ; for the others, this ratio is e O ( 1 ) e^{O(1)} e O ( 1 ) . The transformation comparing the two normalizations is represented in homogeneous coordinates by
diag ( p 1 ( s 1 ) p 1 ( s 2 ) , … , p n ( s 1 ) p n ( s 2 ) , 1 ) . \operatorname{diag}\left(\frac{p_1(s_1)}{p_1(s_2)},\ldots,\frac{p_n(s_1)}{p_n(s_2)},1\right). diag ( p 1 ( s 2 ) p 1 ( s 1 ) , … , p n ( s 2 ) p n ( s 1 ) , 1 ) .
Its logarithmic singular values therefore consist of k k k entries close to τ \tau τ and N − k N-k N − k entries close to zero. Subtracting a common constant to normalize the determinant does not change R R R or D k D_k D k . Consequently, this transformation has R = τ + O ( 1 ) R=\tau+O(1) R = τ + O ( 1 ) and D k = O ( 1 ) D_k=O(1) D k = O ( 1 ) .
We next relate these coordinate changes to elements of Γ \Gamma Γ . After composing all the normalizations with one fixed projective map taking the positive orthant to a bounded simplex, the normalized domains lie in a fixed bounded set and contain a ball of radius comparable to r r r around a fixed point. Cocompactness allows us to move the preimage of that point into a fixed compact subset of Ω \Omega Ω . An elementary matrix estimate implies that on an interval [ s 1 , s 2 ] [s_1,s_2] [ s 1 , s 2 ] as above this group element γ ∈ Γ \gamma\in\Gamma γ ∈ Γ satisfies
R ( μ ( γ ) ) = τ + O ( 1 + log ( 1 / r ) ) , R(\mu(\gamma))=\tau+O(1+\log(1/r)), R ( μ ( γ )) = τ + O ( 1 + log ( 1/ r )) ,
D k ( μ ( γ ) ) = O ( 1 + log ( 1 / r ) ) . D_k(\mu(\gamma))=O(1+\log(1/r)). D k ( μ ( γ )) = O ( 1 + log ( 1/ r )) .
All errors are independent of τ \tau τ and of the box centres. Inequality (2) therefore bounds τ \tau τ by a constant times 1 + log ( 1 / r ) 1+\log(1/r) 1 + log ( 1/ r ) .
Starting at s = log 2 s=\log2 s = log 2 and ending at the largest transition value, there are at most n n n such intervals with at least one changing coordinate. Summing their lengths bounds max i log ( b i / a i ) \max_i\log(b_i/a_i) max i log ( b i / a i ) ; if no transition value exceeds log 2 \log2 log 2 , this upper bound is immediate. Interchanging a a a and b b b gives the corresponding bound for the opposite ratios. We conclude that
max i ∣ log ( b i / a i ) ∣ ≤ c s e p ( 1 + log ( 1 / r ) ) . \max_i\left|\log(b_i/a_i)\right|\le c_{\mathrm{sep}}(1+\log(1/r)). i max ∣ log ( b i / a i ) ∣ ≤ c sep ( 1 + log ( 1/ r )) .
Thus boxes of a fixed relative thickness cannot have centres arbitrarily far apart in logarithmic coordinates.
The domain may still extend far by becoming increasingly thin. To control this possibility, we partition the orthant into dyadic boxes. A cell carrying mass at least ρ \rho ρ contains a coordinate box with relative widths comparable to ρ \rho ρ . The separation estimate therefore bounds the number of such cells by a constant times ( 1 + ∣ log ρ ∣ ) n (1+|\log\rho|)^n ( 1 + ∣ log ρ ∣ ) n . Integrating this counting bound over ρ \rho ρ gives a uniform bound for the total mass of D D D .
The singular-value estimate also holds for the dual action. We can therefore apply the same integral bound to projective images of Ω ∗ \Omega^* Ω ∗ . Faifman’s projective duality [8 Corollary 3.11] converts it into a uniform Funk volume bound for simplices, which we then promote to a bound on the Holmes–Thompson volume. Finally, equivariant straightening and integration of the Blaschke volume form produce a bounded cocycle whose evaluation on the fundamental class is positive. This proves positivity of simplicial volume.
Organization of the paper. Section 1 fixes the cone, affine-sphere, singular-value, and Holmes–Thompson conventions. Section 2 proves Theorem D. Section 3 establishes the Cartan gap; Section 4 proves the integral bound estimate. Section 5 applies duality to obtain the uniform bound in Theorem A. Section 6 concludes with an explicit equivariant straightening and the proof of the corollaries regarding the simplicial volume.
AI declaration. The main strategy of the proof was motivated by the example, explained in the introduction, of a family of simplices with unbounded Holmes–Thomson volume, which was an original observation of the author, who thus realized that the Cartan gap (0.2) could be the key to prove the main theorem. The Ricci estimate and the Cartan gap were proved by the author without any AI tool. Chat GPT Sol 6.1 was later used to deduce the integral bound in Section 4 from the Cartan gap. In other words, all mathematical ideas outside Section 4 are original of the author. The author also used Chat GPT Sol 6.1 for literature review and editing.
Acknowledgment. The author acknowledges support from the Italian Ministry of University and Research (MUR) through the FIS3 project “Geometry and Topology of higher Teichmüller spaces”, grant no. FIS-2024-05142.
Preliminaries
We write n = dim Ω ≥ 2 n = \dim\Omega\ge2 n = dim Ω ≥ 2 and N = n + 1 N = n + 1 N = n + 1 . We fix the standard ambient volume density on R N \mathbb{R}^{N} R N . The standard Euclidean operator norm of a linear map is denoted by ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ .
Cones, duality, and canonical lifts
An open convex cone C ⊂ R N C \subset\mathbb{R}^{N} C ⊂ R N is proper if it is nonempty and full-dimensional and its closure is pointed: C ‾ ∩ ( − C ‾ ) = { 0 } \overline{C} \cap(-\overline{C}) = \{0\} C ∩ ( − C ) = { 0 } . Its projectivization is a properly convex domain. Conversely, a properly convex domain has two opposite proper cones; we fix one and call it C C C . Define its open dual cone and projective dual domain by
C ∗ = { ℓ ∈ ( R N ) ∗ : ℓ ( v ) > 0 for every v ∈ C ‾ ∖ { 0 } } , Ω ∗ = P ( C ∗ ) . C^{*} = \{\ell\in(\mathbb{R}^{N})^{*} : \ell(v) > 0 \text{ for every } v \in\overline{C} \setminus\{0\}\}, \qquad\Omega^{*} = \mathbb{P}(C^{*}). C ∗ = { ℓ ∈ ( R N ) ∗ : ℓ ( v ) > 0 for every v ∈ C ∖ { 0 }} , Ω ∗ = P ( C ∗ ) .
In particular, an element of C ∗ C^{*} C ∗ is strictly positive on every nonzero boundary ray of C C C .
A product decomposition of C C C consists of a direct sum R N = L 1 ⊕ L 2 \mathbb{R}^{N} = L_{1} \oplus L_{2} R N = L 1 ⊕ L 2 and proper open convex cones C i ⊂ L i C_{i} \subset L_{i} C i ⊂ L i such that C = C 1 × C 2 C = C_{1} \times C_{2} C = C 1 × C 2 . It is nontrivial when both L i L_{i} L i are nonzero; a one-dimensional factor is allowed. Projectivizing gives the join of P ( C 1 ) \mathbb{P}(C_{1}) P ( C 1 ) and P ( C 2 ) \mathbb{P}(C_{2}) P ( C 2 ) .
Every g ∈ Aut ( Ω ) g \in\operatorname{Aut}(\Omega) g ∈ Aut ( Ω ) has a unique linear representative preserving C C C and satisfying ∣ det g ∣ = 1 |\det g| = 1 ∣ det g ∣ = 1 ([BDL18]). We use the same letter for a projective automorphism and this representative. The contragredient group acts on C ∗ C^{*} C ∗ by g − T g^{-T} g − T .
The canonical affine sphere
We use the existence and uniqueness theorem for complete hyperbolic affine spheres, together with its conormal duality; see [19 Theorem 3 and Proposition 8] . We record the exact normalizations needed here.
Lemma 1.1. There is a unique radial lift f : Ω → C f : \Omega\to C f : Ω → C , with [ f ( x ) ] = x [f(x)] = x [ f ( x )] = x , whose image is a hyperbolic affine sphere centered at the origin, of affine mean curvature − 1 -1 − 1 , meeting every ray of C C C once and having complete Blaschke metric h h h . It is equivariant for the cone-preserving lifts just defined. Its conormal map ν : Ω → C ∗ \nu: \Omega\to C^{*} ν : Ω → C ∗ , characterized by
ν x ( f ( x ) ) = 1 , ν x ( f ∗ X ) = 0 , \nu_{x}(f(x)) = 1, \qquad\nu_{x}(f_{*}X) = 0, ν x ( f ( x )) = 1 , ν x ( f ∗ X ) = 0 , projectivizes to a diffeomorphism Ω → Ω ∗ \Omega\to\Omega^{*} Ω → Ω ∗ and is equivariant for the contragredient action. In particular, if Γ \Gamma Γ divides Ω \Omega Ω , its contragredient group divides Ω ∗ \Omega^{*} Ω ∗ .
Let D D D be the flat ambient connection and let ∇ \nabla ∇ be the Levi–Civita connection of h h h . If ∇ a f f \nabla^{\mathrm{aff}} ∇ aff is the induced affine connection, set A = ∇ a f f − ∇ A = \nabla^{\mathrm{aff}} - \nabla A = ∇ aff − ∇ . The structure equation is
D X f ∗ Y = f ∗ ( ∇ X Y + A X Y ) + h ( X , Y ) f . (3) D_{X}f_{*}Y = f_{*}(\nabla_{X}Y + A_{X}Y) + h(X,Y)f. \tag*{(3)} D X f ∗ Y = f ∗ ( ∇ X Y + A X Y ) + h ( X , Y ) f . (3)
The tensor A ( X , Y , Z ) = h ( A X Y , Z ) A(X,Y,Z) = h(A_{X}Y,Z) A ( X , Y , Z ) = h ( A X Y , Z ) is symmetric in all three entries, each endomorphism A X A_{X} A X is trace free, and
( ∇ X A ) ( Y , Z , W ) is symmetric in X , Y , Z , W . (4) (\nabla_{X}A)(Y,Z,W) \text{ is symmetric in } X,Y,Z,W. \tag*{(4)} ( ∇ X A ) ( Y , Z , W ) is symmetric in X , Y , Z , W . (4)
We use the conventions
Rm ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z , \operatorname{Rm}(X,Y)Z = \nabla_{X}\nabla_{Y}Z - \nabla_{Y}\nabla_{X}Z - \nabla_{[X,Y]}Z, Rm ( X , Y ) Z = ∇ X ∇ Y Z − ∇ Y ∇ X Z − ∇ [ X , Y ] Z ,
Ric ( Y , Z ) = ∑ a h ( Rm ( e a , Y ) Z , e a ) , Δ = tr h ∇ 2 . \operatorname{Ric}(Y,Z) = \sum_{a} h(\operatorname{Rm}(e_{a},Y)Z,e_{a}), \qquad\Delta= \operatorname{tr}_{h}\nabla^{2}. Ric ( Y , Z ) = a ∑ h ( Rm ( e a , Y ) Z , e a ) , Δ = tr h ∇ 2 .
Thus Δ \Delta Δ has positive principal part ∑ a ∂ a 2 \sum_a \partial_a^2 ∑ a ∂ a 2 in a normal frame. The Gauss equation, with these conventions, is
Rm ( X , Y ) Z = − h ( Y , Z ) X + h ( X , Z ) Y − [ A X , A Y ] Z . (5) \operatorname{Rm}(X,Y)Z=-h(Y,Z)X+h(X,Z)Y-[A_X,A_Y]Z. \tag*{(5)} Rm ( X , Y ) Z = − h ( Y , Z ) X + h ( X , Z ) Y − [ A X , A Y ] Z . (5)
In an h h h -orthonormal frame put A i = A e i A_i=A_{e_i} A i = A e i and B i j = tr ( A i A j ) B_{ij}=\operatorname{tr}(A_iA_j) B ij = tr ( A i A j ) . Tracing (5) gives
Ric h = − ( n − 1 ) h + B . (6) \operatorname{Ric}_h=-(n-1)h+B. \tag*{(6)} Ric h = − ( n − 1 ) h + B . (6)
The tensor B B B is symmetric and nonnegative. We freely identify symmetric two-tensors with self-adjoint endomorphisms using h h h .
Cartan projections
For a linear map g g g with ∣ det g ∣ = 1 |\det g|=1 ∣ det g ∣ = 1 , let s 1 ( g ) ≥ ⋯ ≥ s N ( g ) > 0 s_1(g)\ge\cdots\ge s_N(g)>0 s 1 ( g ) ≥ ⋯ ≥ s N ( g ) > 0 be its singular values. Its Cartan vector is
μ ( g ) = ( μ 1 ( g ) , … , μ N ( g ) ) , μ i ( g ) = log s i ( g ) . \mu(g)=(\mu_1(g),\ldots,\mu_N(g)),\qquad\mu_i(g)=\log s_i(g). μ ( g ) = ( μ 1 ( g ) , … , μ N ( g )) , μ i ( g ) = log s i ( g ) .
It lies in the closed ordered chamber of the trace-zero hyperplane. For an arbitrary projective transformation, we use any representative normalized to absolute determinant one; the two sign choices have the same singular values. On an ordered trace-zero vector μ \mu μ set
R ( μ ) = μ 1 − μ N , R(\mu)=\mu_1-\mu_N, R ( μ ) = μ 1 − μ N ,
D k ( μ ) = ∑ i ≤ k ( μ 1 − μ i ) + ∑ i > k ( μ i − μ N ) , 1 ≤ k < N . (7) D_k(\mu)=\sum_{i\le k}(\mu_1-\mu_i)+\sum_{i>k}(\mu_i-\mu_N),\qquad1\le k<N. \tag*{(7)} D k ( μ ) = i ≤ k ∑ ( μ 1 − μ i ) + i > k ∑ ( μ i − μ N ) , 1 ≤ k < N . (7)
Thus D k D_k D k vanishes exactly on the ray of vectors with their first k k k entries equal and their remaining entries equal.
We record the elementary perturbation estimate used below.
Lemma 1.2. Suppose g = U g 0 V g=Ug_0V g = U g 0 V , where g , U , g 0 , V g,U,g_0,V g , U , g 0 , V have determinant of absolute value one. If
∥ U ∥ , ∥ U − 1 ∥ ≤ e η U , ∥ V ∥ , ∥ V − 1 ∥ ≤ e η V , \lVert U\rVert,\lVert U^{-1}\rVert\le e^{\eta_U},\qquad\lVert V\rVert,\lVert V^{-1}\rVert\le e^{\eta_V}, ∥ U ∥ , ∥ U − 1 ∥ ≤ e η U , ∥ V ∥ , ∥ V − 1 ∥ ≤ e η V , then ∣ μ i ( g ) − μ i ( g 0 ) ∣ ≤ η U + η V |\mu_i(g)-\mu_i(g_0)|\le\eta_U+\eta_V ∣ μ i ( g ) − μ i ( g 0 ) ∣ ≤ η U + η V for every i i i . In particular, a fixed change of ambient basis changes all Cartan coordinates by a uniformly bounded amount. Moreover, the dual Cartan vector satisfies
μ ~ i = μ i ( g − T ) = − μ N + 1 − i ( g ) , R ( μ ~ ) = R ( μ ) , D k ( μ ~ ) = D N − k ( μ ) . (8) \widetilde{\mu}_i=\mu_i(g^{-\mathsf{T}})=-\mu_{N+1-i}(g),\qquad R(\widetilde{\mu})=R(\mu),\qquad D_k(\widetilde{\mu})=D_{N-k}(\mu). \tag*{(8)} μ i = μ i ( g − T ) = − μ N + 1 − i ( g ) , R ( μ ) = R ( μ ) , D k ( μ ) = D N − k ( μ ) . (8)
Proof. The min–max characterization of singular values gives s i ( U g 0 V ) ≤ ∥ U ∥ ∥ V ∥ s i ( g 0 ) s_i(Ug_0V)\le\lVert U\rVert\lVert V\rVert s_i(g_0) s i ( U g 0 V ) ≤ ∥ U ∥ ∥ V ∥ s i ( g 0 ) . Applying the same inequality to g 0 = U − 1 g V − 1 g_0=U^{-1}gV^{-1} g 0 = U − 1 g V − 1 gives the reverse bound. Taking logarithms proves the first assertion. A basis change is conjugation by a fixed matrix; normalizing that matrix to absolute determinant one leaves the conjugation unchanged and gives the second assertion. Inversion reverses and reciprocates the singular values, whereas transposition does not change them. This proves the first equality in (8) ; substitution into (7) proves the other two. □ \square □
If two ordered trace-zero vectors differ by at most ϵ 0 \epsilon_0 ϵ 0 in every coordinate, their R R R values differ by at most 2 ϵ 0 2\epsilon_0 2 ϵ 0 , and their D k D_k D k values differ by at most 2 N ϵ 0 2N\epsilon_0 2 N ϵ 0 . These bounds also follow immediately from (7) .
Funk and Holmes–Thompson volumes
Choose a bounded affine realization of a properly convex domain U U U . For x ∈ U x\in U x ∈ U and v ≠ 0 v\ne0 v = 0 , the Funk norm is defined by F U ( x , v ) = 1 / t F_U(x,v)=1/t F U ( x , v ) = 1/ t , where t > 0 t>0 t > 0 is the unique parameter such that x + t v ∈ ∂ U x+tv\in\partial U x + t v ∈ ∂ U ; put F U ( x , 0 ) = 0 F_U(x,0)=0 F U ( x , 0 ) = 0 . Its tangent unit ball is U − x U-x U − x . Its cotangent unit ball is the polar convex body
K U , x ∗ = { ξ ∣ ξ ( z − x ) ≤ 1 for every z ∈ U } . K^*_{U,x}=\{\xi\mid\xi(z-x)\le1\text{ for every }z\in U\}. K U , x ∗ = { ξ ∣ ξ ( z − x ) ≤ 1 for every z ∈ U } .
The Hilbert norm is H U ( x , v ) = 1 2 ( F U ( x , v ) + F U ( x , − v ) ) H_U(x,v)=\frac{1}{2}(F_U(x,v)+F_U(x,-v)) H U ( x , v ) = 2 1 ( F U ( x , v ) + F U ( x , − v )) . In terms of cotangent unit balls,
K H U , x ∗ = 1 2 ( K U , x ∗ − K U , x ∗ ) . (9) K^*_{H_U,x}=\frac{1}{2}(K^*_{U,x}-K^*_{U,x}). \tag*{(9)} K H U , x ∗ = 2 1 ( K U , x ∗ − K U , x ∗ ) . (9)
This follows because a norm is the support function of its cotangent unit ball and support functions add under Minkowski addition.
Let ω n \omega_n ω n be the Euclidean volume of the unit ball in R n \mathbb{R}^n R n . For a possibly asymmetric Finsler metric f f f , its Holmes–Thompson volume density is
d Vol H T , f ( x ) = Vol E u c l ( K f , x ∗ ) ω n d x 1 ⋯ d x n . (10) d\operatorname{Vol}_{\mathrm{HT},f}(x)=\frac{\operatorname{Vol}_{\mathrm{Eucl}}(K^*_{f,x})}{\omega_n}\,dx_1\cdots dx_n. \tag*{(10)} d Vol HT , f ( x ) = ω n Vol Eucl ( K f , x ∗ ) d x 1 ⋯ d x n . (10)
We abbreviate the Hilbert Holmes–Thompson measure by Vol H T , U \operatorname{Vol}_{\mathrm{HT},U} Vol HT , U and the Funk Holmes–Thompson measure by Vol F , U \operatorname{Vol}_{F,U} Vol F , U . Unlike the Hilbert norm, the Funk norm depends on the affine chart. Its total volume on a convex subdomain has the projective invariance and duality stated precisely in Section 5.
For later use, both Hilbert and Funk cotangent unit balls vary continuously on compact subsets of a bounded convex domain, in the Hausdorff topology. One can see this by first translating the closed domain by − x -x − x and then taking polars; on a compact set of basepoints these bodies contain a fixed ball and lie in a fixed ball, so polarity is continuous. Their Euclidean volumes are therefore positive and continuous. In particular the Holmes–Thompson density is a positive continuous density, invariant under projective automorphisms.
Ricci curvature and cone splitting
Calabi’s nonpositivity theorem [7 Theorem 5.1] applies to complete hyperbolic affine spheres. We reprove it in the present cocompact setting, using a differential identity for the negative of the Ricci tensor. We then analyze the null case: this identity and a strong minimum principle show that the Ricci kernel has constant rank and is parallel. A nonzero null direction consequently produces a parallel field and a nontrivial product decomposition of the cone, proving Theorem D.
We identify symmetric covariant two-tensors with self-adjoint endomorphisms by means of h h h . At a point, choose a normal h h h -orthonormal frame e 1 , … , e n e_1,\ldots,e_n e 1 , … , e n , and write
A i = A e i , A i j k = h ( A e i e j , e k ) , B i j = tr ( A i A j ) , P = ( n − 1 ) id − B . A_i=Ae_i,\qquad A_{ijk}=h(A_{e_i}e_j,e_k),\qquad B_{ij}=\operatorname{tr}(A_iA_j),\qquad P=(n-1)\operatorname{id}-B. A i = A e i , A ij k = h ( A e i e j , e k ) , B ij = tr ( A i A j ) , P = ( n − 1 ) id − B .
Thus P = − Ric h P=-\operatorname{Ric}_h P = − Ric h by (6) . Repeated indices are summed from 1 1 1 to n n n . The inner product of endomorphisms is the Frobenius inner product ⟨ U , V ⟩ H S = tr ( U ∗ V ) \langle U,V\rangle_{\mathrm{HS}}=\operatorname{tr}(U^*V) ⟨ U , V ⟩ HS = tr ( U ∗ V ) , where the adjoint is taken with respect to h h h . Define the symmetric nonnegative two-tensors
H i j = ∑ k , a , b ( ∇ k A ) i a b ( ∇ k A ) j a b , \mathcal{H}_{ij}=\sum_{k,a,b}(\nabla_kA)_{iab}(\nabla_kA)_{jab}, H ij = k , a , b ∑ ( ∇ k A ) iab ( ∇ k A ) j ab ,
Q i j = ∑ a ⟨ [ A i , A a ] , [ A j , A a ] ⟩ H S , Q_{ij}=\sum_a\langle[A_i,A_a],[A_j,A_a]\rangle_{\mathrm{HS}}, Q ij = a ∑ ⟨[ A i , A a ] , [ A j , A a ] ⟩ HS ,
R i j 2 = ∑ a ⟨ Rm ( e i , e a ) , Rm ( e j , e a ) ⟩ H S . \mathcal{R}^2_{ij}=\sum_a\langle\operatorname{Rm}(e_i,e_a),\operatorname{Rm}(e_j,e_a)\rangle_{\mathrm{HS}}. R ij 2 = a ∑ ⟨ Rm ( e i , e a ) , Rm ( e j , e a ) ⟩ HS .
The tensors A A A , B B B and P P P satisfy the following differential equations. These computations are classical and can be found for instance in [13 ] and [16 Proposition 3.4] . We re-derive all three formulas here in order to avoid painful book-keeping of the different conventions used in the aforementioned references.
Lemma 2.1. The cubic form satisfies
Δ A i j k = − ( n + 1 ) A i j k + B i p A p j k + B j p A i p k + B k p A i j p − 2 tr ( A i A j A k ) . (11) \Delta A_{ijk}=-(n+1)A_{ijk}+B_{ip}A_{pjk}+B_{jp}A_{ipk}+B_{kp}A_{ijp}-2\operatorname{tr}(A_iA_jA_k). \tag*{(11)} Δ A ij k = − ( n + 1 ) A ij k + B i p A p j k + B j p A i p k + B k p A ij p − 2 tr ( A i A j A k ) . (11) Its quadratic contraction satisfies
1 2 Δ B = H − ( n + 1 ) B + B 2 + Q , (12) \frac{1}{2}\Delta B=\mathcal{H}-(n+1)B+B^2+Q, \tag*{(12)} 2 1 Δ B = H − ( n + 1 ) B + B 2 + Q , (12) and
1 2 Δ P = ( n + 1 ) P − P 2 − R 2 − H . (13) \frac{1}{2}\Delta P=(n+1)P-P^2-\mathcal{R}^2-\mathcal{H}. \tag*{(13)} 2 1 Δ P = ( n + 1 ) P − P 2 − R 2 − H . (13)
Proof. Since A A A is tracefree and ∇ A \nabla A ∇ A is full symmetric, we have
∑ a ∇ a A a j k = ∇ j ∑ a A a a k = 0. \sum_a \nabla_a A_{ajk}=\nabla_j\sum_a A_{aak}=0. a ∑ ∇ a A aj k = ∇ j a ∑ A aak = 0.
Writing R m a i j p \mathrm{Rm}_{aij}{}^p Rm aij p for the coefficients of R m ( e a , e i ) e j \mathrm{Rm}(e_a,e_i)e_j Rm ( e a , e i ) e j , and using the preceding divergence identity, we obtain
Δ A i j k = ∑ a ( ∇ a ∇ i − ∇ i ∇ a ) A a j k = − ∑ a , p ( R m a i a p A p j k + R m a i j p A a p k + R m a i k p A a j p ) = R i c i p A p j k − ∑ a , p ( R m a i j p A a p k + R m a i k p A a j p ) . \begin{aligned}
\Delta A_{ijk}=\sum_a(\nabla_a\nabla_i-\nabla_i\nabla_a)A_{ajk} \\
&=-\sum_{a,p}(\mathrm{Rm}_{aia}{}^p A_{pjk}+\mathrm{Rm}_{aij}{}^p A_{apk}+\mathrm{Rm}_{aik}{}^p A_{ajp}) \\
&=\mathrm{Ric}_{ip}A_{pjk}-\sum_{a,p}(\mathrm{Rm}_{aij}{}^p A_{apk}+\mathrm{Rm}_{aik}{}^p A_{ajp}).
\end{aligned} Δ A ij k = a ∑ ( ∇ a ∇ i − ∇ i ∇ a ) A aj k = − a , p ∑ ( Rm aia p A p j k + Rm aij p A a p k + Rm aik p A aj p ) = Ric i p A p j k − a , p ∑ ( Rm aij p A a p k + Rm aik p A aj p ) .
The Gauss equation (5) yields
− ∑ a , p R m a i j p A a p k = − A i j k + ∑ a , p [ A a , A i ] p j A a p k = − A i j k + B k p A i j p − tr ( A i A j A k ) . \begin{aligned}
-\sum_{a,p}\mathrm{Rm}_{aij}{}^p A_{apk}=-A_{ijk}+\sum_{a,p}[A_a,A_i]^p{}_jA_{apk} \\
&=-A_{ijk}+B_{kp}A_{ijp}-\operatorname{tr}(A_iA_jA_k).
\end{aligned} − a , p ∑ Rm aij p A a p k = − A ij k + a , p ∑ [ A a , A i ] p j A a p k = − A ij k + B k p A ij p − tr ( A i A j A k ) .
The other curvature contraction interchanges j j j and k k k . The trace of a product of three symmetric matrices is symmetric in those matrices, by cyclicity of the trace and transposition. Substituting R i c i p = − ( n − 1 ) δ i p + B i p \mathrm{Ric}_{ip}=-(n-1)\delta_{ip}+B_{ip} Ric i p = − ( n − 1 ) δ i p + B i p proves (11) .
For the quadratic contraction, introduce
B i j ( 1 ) = tr ( B A i A j ) , B i j ( 2 ) = ∑ a tr ( A i A a A j A a ) . B^{(1)}_{ij}=\operatorname{tr}(BA_iA_j),\qquad B^{(2)}_{ij}=\sum_a\operatorname{tr}(A_iA_aA_jA_a). B ij ( 1 ) = tr ( B A i A j ) , B ij ( 2 ) = a ∑ tr ( A i A a A j A a ) .
Both tensors are symmetric. Cubic symmetry gives
( ∑ a A a 2 ) i j = ∑ a , p A a i p A a p j = B i j . \left(\sum_a A_a^2\right)_{ij}=\sum_{a,p}A_{aip}A_{apj}=B_{ij}. ( a ∑ A a 2 ) ij = a , p ∑ A ai p A a p j = B ij .
Since every commutator [ A i , A a ] [A_i,A_a] [ A i , A a ] is skew-adjoint, its expansion gives
Q i j = − ∑ a tr ( ( A i A a − A a A i ) ( A j A a − A a A j ) ) = 2 B i j ( 1 ) − 2 B i j ( 2 ) . Q_{ij}=-\sum_a\operatorname{tr}\bigl((A_iA_a-A_aA_i)(A_jA_a-A_aA_j)\bigr)=2B^{(1)}_{ij}-2B^{(2)}_{ij}. Q ij = − a ∑ tr ( ( A i A a − A a A i ) ( A j A a − A a A j ) ) = 2 B ij ( 1 ) − 2 B ij ( 2 ) .
The cubic term of (11) has contraction
∑ a , b A j a b tr ( A i A a A b ) = ∑ a , b , p , q , r A j a b A i p q A a q r A b r p = ∑ r tr ( A j A r A i A r ) = B i j ( 2 ) . \begin{aligned}
\sum_{a,b}A_{jab}\operatorname{tr}(A_iA_aA_b)=\sum_{a,b,p,q,r}A_{jab}A_{ipq}A_{aqr}A_{brp} \\
&=\sum_r\operatorname{tr}(A_jA_rA_iA_r)=B^{(2)}_{ij}.
\end{aligned} a , b ∑ A j ab tr ( A i A a A b ) = a , b , p , q , r ∑ A j ab A i pq A a q r A b r p = r ∑ tr ( A j A r A i A r ) = B ij ( 2 ) .
Here the second equality uses A a q r = A r a q A_{aqr}=A_{raq} A a q r = A r a q and A b r p = A r b p A_{brp}=A_{rbp} A b r p = A r b p . The product rule therefore gives
1 2 Δ B i j = H i j + 1 2 ∑ a , b ( ( Δ A i a b ) A j a b + A i a b ( Δ A j a b ) ) = H i j − ( n + 1 ) B i j + ( B 2 ) i j + 2 B i j ( 1 ) − 2 B i j ( 2 ) . \begin{aligned}
\frac{1}{2}\Delta B_{ij}=H_{ij}+\frac{1}{2}\sum_{a,b}\bigl((\Delta A_{iab})A_{jab}+A_{iab}(\Delta A_{jab})\bigr) \\
&=H_{ij}-(n+1)B_{ij}+(B^2)_{ij}+2B^{(1)}_{ij}-2B^{(2)}_{ij}.
\end{aligned} 2 1 Δ B ij = H ij + 2 1 a , b ∑ ( ( Δ A iab ) A j ab + A iab ( Δ A j ab ) ) = H ij − ( n + 1 ) B ij + ( B 2 ) ij + 2 B ij ( 1 ) − 2 B ij ( 2 ) .
This proves (12) .
We next contract the Gauss equation. With J i j X = h ( e j , X ) e i − h ( e i , X ) e j J_{ij}X=h(e_j,X)e_i-h(e_i,X)e_j J ij X = h ( e j , X ) e i − h ( e i , X ) e j , one has
∑ a ⟨ J i a , J j a ⟩ H S = 2 ( n − 1 ) δ i j . \sum_a\langle J_{ia},J_{ja}\rangle_{\mathrm{HS}}=2(n-1)\delta_{ij}. a ∑ ⟨ J ia , J j a ⟩ HS = 2 ( n − 1 ) δ ij .
The mixed contraction is
∑ a ⟨ J i a , [ A j , A a ] ⟩ H S = 2 ∑ a [ A j , A a ] i a = 2 ∑ a , p ( A j i p A a p a − A a i p A j p a ) = − 2 B i j . \begin{aligned}
\sum_a\langle J_{ia},[A_j,A_a]\rangle_{\mathrm{HS}}=2\sum_a[A_j,A_a]_{ia} \\
&=2\sum_{a,p}(A_{jip}A_{apa}-A_{aip}A_{jpa})=-2B_{ij}.
\end{aligned} a ∑ ⟨ J ia , [ A j , A a ] ⟩ HS = 2 a ∑ [ A j , A a ] ia = 2 a , p ∑ ( A j i p A a p a − A ai p A j p a ) = − 2 B ij .
The first term vanishes by trace freeness. Consequently
R 2 = Q + 2 ( n − 1 ) id − 4 B . (14) \mathcal{R}^{2}=\mathcal{Q}+2(n-1)\operatorname{id}-4B. \tag*{(14)} R 2 = Q + 2 ( n − 1 ) id − 4 B . (14)
Combining this with (12) , and substituting B = ( n − 1 ) id − P B=(n-1)\operatorname{id}-P B = ( n − 1 ) id − P , proves (13) .
Proposition 2.2 (Calabi). The Blaschke Ricci tensor satisfies
− ( n − 1 ) h ≤ Ric h ≤ 0. -(n-1)h\leq\operatorname{Ric}_{h}\leq0. − ( n − 1 ) h ≤ Ric h ≤ 0. Equivalently, 0 ≤ B ≤ ( n − 1 ) id 0\leq B\leq(n-1)\operatorname{id} 0 ≤ B ≤ ( n − 1 ) id and P ≥ 0 P\geq0 P ≥ 0 .
Proof. Equations (12) and (14) give
1 2 Δ B = H + R 2 + ( B − ( n − 1 ) id ) ( B + 2 id ) . (15) \frac{1}{2}\Delta B=\mathcal{H}+\mathcal{R}^{2}+\left(B-(n-1)\operatorname{id}\right)(B+2\operatorname{id}). \tag*{(15)} 2 1 Δ B = H + R 2 + ( B − ( n − 1 ) id ) ( B + 2 id ) . (15)
Choose a compact fundamental domain F ⊂ Ω \mathcal{F}\subset\Omega F ⊂ Ω for the action of Γ \Gamma Γ . The largest eigenvalue of B B B is continuous and Γ \Gamma Γ -invariant, so its global maximum λ max \lambda_{\max} λ m a x is attained at a point p ∈ F p\in\mathcal{F} p ∈ F . Moreover, λ max ≥ 0 \lambda_{\max}\geq0 λ m a x ≥ 0 because B B B is symmetric and non-negative definite.
Choose a unit eigenvector v v v at p p p relative to λ max \lambda_{\max} λ m a x and extend it to a local unit field with ∇ v ( p ) = 0 \nabla v(p)=0 ∇ v ( p ) = 0 . The scalar function b = B ( v , v ) b=B(v,v) b = B ( v , v ) has a local maximum at p p p . At that point,
Δ b = ( Δ B ) ( v , v ) + 2 B ( Δ v , v ) = ( Δ B ) ( v , v ) . \Delta b=(\Delta B)(v,v)+2B(\Delta v,v)=(\Delta B)(v,v). Δ b = ( Δ B ) ( v , v ) + 2 B ( Δ v , v ) = ( Δ B ) ( v , v ) .
The last equality follows from B v = λ max v Bv=\lambda_{\max}v B v = λ m a x v and h ( Δ v , v ) = 0 h(\Delta v,v)=0 h ( Δ v , v ) = 0 , the latter being the second derivative of the unit-length condition at a point where ∇ v = 0 \nabla v=0 ∇ v = 0 . Hence (15) implies
0 ≥ 1 2 ( Δ B ) ( v , v ) = H ( v , v ) + R 2 ( v , v ) + ( λ max − ( n − 1 ) ) ( λ max + 2 ) . 0\geq\frac{1}{2}(\Delta B)(v,v)=\mathcal{H}(v,v)+\mathcal{R}^{2}(v,v)+(\lambda_{\max}-(n-1))(\lambda_{\max}+2). 0 ≥ 2 1 ( Δ B ) ( v , v ) = H ( v , v ) + R 2 ( v , v ) + ( λ m a x − ( n − 1 )) ( λ m a x + 2 ) .
Both H ( v , v ) \mathcal{H}(v,v) H ( v , v ) and R 2 ( v , v ) \mathcal{R}^{2}(v,v) R 2 ( v , v ) are nonnegative and λ max + 2 > 0 \lambda_{\max}+2>0 λ m a x + 2 > 0 , so λ max ≤ n − 1 \lambda_{\max}\leq n-1 λ m a x ≤ n − 1 . Thus P ≥ 0 P\geq0 P ≥ 0 . The remaining inequality Ric h ≥ − ( n − 1 ) h \operatorname{Ric}_{h}\geq-(n-1)h Ric h ≥ − ( n − 1 ) h follows directly from B ≥ 0 B\geq0 B ≥ 0 and (6) .
We shall use the following scalar minimum principle for continuous functions.
Lemma 2.3. Let ( X , h ) (X,h) ( X , h ) be a connected smooth Riemannian manifold and let c ≥ 0 c\geq0 c ≥ 0 be a constant. Suppose that u : X → [ 0 , ∞ ) u:X\to[0,\infty) u : X → [ 0 , ∞ ) is continuous and satisfies
− Δ u + c u ≥ 0 -\Delta u+cu\geq0 − Δ u + c u ≥ 0 in the following sense: whenever a C 2 C^{2} C 2 function ϕ \phi ϕ touches u u u from below at p p p , one has Δ ϕ ( p ) ≤ c u ( p ) \Delta\phi(p)\leq cu(p) Δ ϕ ( p ) ≤ c u ( p ) . If u u u vanishes at one point, then u u u vanishes identically.
Proof. This is a special case of the viscosity strong maximum principle in [10 Theorem 4.7, Example 4.12] , applied to the Laplacian operator. We give a direct proof in this case for the convenience of the reader.
Suppose that both the positive set and the zero set of u u u are nonempty. In a coordinate chart choose a Euclidean ball B r 0 ( x ) B_{r_{0}}(x) B r 0 ( x ) contained in the positive set and tangent to its zero set at a point q q q of the chart. Such a ball is obtained by taking its center in the positive set sufficiently close to its boundary and using the distance to the zero set as radius. It can be chosen with its closure contained in the chart. In these coordinates write
Δ = a i j ( y ) ∂ i ∂ j + b i ( y ) ∂ i , \Delta=a^{ij}(y)\partial_{i}\partial_{j}+b^{i}(y)\partial_{i}, Δ = a ij ( y ) ∂ i ∂ j + b i ( y ) ∂ i ,
where ( a i j ) (a^{ij}) ( a ij ) is uniformly positive definite on a neighborhood of B r 0 ( x ) ‾ \overline{B_{r_{0}}(x)} B r 0 ( x ) . On the annulus r 0 / 2 ≤ ∣ y − x ∣ ≤ r 0 r_{0}/2\leq|y-x|\leq r_{0} r 0 /2 ≤ ∣ y − x ∣ ≤ r 0 , set
w ( y ) = e − α ∣ y − x ∣ 2 − e − α r 0 2 . w(y)=e^{-\alpha|y-x|^{2}}-e^{-\alpha r_{0}^{2}}. w ( y ) = e − α ∣ y − x ∣ 2 − e − α r 0 2 .
For sufficiently large α \alpha α ,
Δ w − c w > 0 (16) \Delta w-cw>0 \tag*{(16)} Δ w − c w > 0 (16)
on a neighborhood of this closed annulus. Indeed, writing ξ = y − x \xi=y-x ξ = y − x , the left side is
e − α ∣ ξ ∣ 2 ( 4 α 2 a i j ξ i ξ j − 2 α ∑ i a i i − 2 α b i ξ i − c ) + c e − α r 0 2 . e^{-\alpha|\xi|^{2}}\left(4\alpha^{2}a^{ij}\xi_{i}\xi_{j}-2\alpha\sum_{i}a^{ii}-2\alpha b^{i}\xi_{i}-c\right)+ce^{-\alpha r_{0}^{2}}. e − α ∣ ξ ∣ 2 ( 4 α 2 a ij ξ i ξ j − 2 α i ∑ a ii − 2 α b i ξ i − c ) + c e − α r 0 2 .
uniform ellipticity and ∣ ξ ∣ ≥ r 0 / 2 |\xi| \ge r_{0}/2 ∣ ξ ∣ ≥ r 0 /2 make the quadratic term in α \alpha α dominate the other terms. Choose δ > 0 \delta> 0 δ > 0 so small that u ≥ δ w u \ge\delta w u ≥ δ w on the inner sphere. The same inequality holds on the outer sphere, where w = 0 w = 0 w = 0 . If u − δ w u - \delta w u − δ w had a negative minimum m m m inside the annulus, then δ w + m \delta w + m δ w + m would touch u u u from below there. The assumed viscosity inequality and (16) would give
0 ≤ − δ Δ w + c ( δ w + m ) < 0 , 0 \le-\delta\Delta w + c(\delta w + m) < 0, 0 ≤ − δ Δ w + c ( δ w + m ) < 0 ,
a contradiction. Thus u ≥ δ w u \ge\delta w u ≥ δ w on the annulus. Immediately outside the outer sphere one has w < 0 ≤ u w < 0 \le u w < 0 ≤ u , so δ w \delta w δ w touches u u u from below at q q q . Since u ( q ) = w ( q ) = 0 u(q) = w(q) = 0 u ( q ) = w ( q ) = 0 , (16) contradicts the viscosity inequality at q q q as well. This proves the lemma. □
Lemma 2.4. The kernel N = ker P \mathcal{N} = \ker P N = ker P is a smooth parallel subbundle. Moreover, for every local section v v v of N \mathcal{N} N ,
Rm ( v , X ) = 0 , ( ∇ X A ) ( v , Y , Z ) = 0 (17) \operatorname{Rm}(v,X) = 0,\qquad(\nabla_X A)(v,Y,Z) = 0 \tag*{(17)} Rm ( v , X ) = 0 , ( ∇ X A ) ( v , Y , Z ) = 0 (17) for all tangent fields X , Y , Z X,Y,Z X , Y , Z .
Proof. Let s k s_k s k be the sum of the k k k smallest eigenvalues of P P P , for 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n . These functions are continuous and nonnegative by Proposition 2.2 . We claim that they are viscosity supersolutions of
− Δ s k + 2 ( n + 1 ) s k ≥ 0. (18) -\Delta s_k + 2(n+1)s_k \ge0. \tag*{(18)} − Δ s k + 2 ( n + 1 ) s k ≥ 0. (18)
Suppose that ϕ \phi ϕ touches s k s_k s k from below at p p p . Choose orthonormal eigenvectors e 1 , … , e k e_1,\ldots,e_k e 1 , … , e k for the k k k smallest eigenvalues at p p p , and extend them orthonormally with ∇ e i ( p ) = 0 \nabla e_i(p) = 0 ∇ e i ( p ) = 0 . The variational characterization of the sum of the smallest eigenvalues gives
s f r = ∑ i = 1 k P ( e i , e i ) ≥ s k , s f r ( p ) = s k ( p ) . s_{\mathrm{fr}} = \sum_{i=1}^{k} P(e_i,e_i) \ge s_k,\qquad s_{\mathrm{fr}}(p) = s_k(p). s fr = i = 1 ∑ k P ( e i , e i ) ≥ s k , s fr ( p ) = s k ( p ) .
Thus ϕ \phi ϕ also touches the smooth function s f r s_{\mathrm{fr}} s fr from below at p p p . All frame-derivative terms in its Laplacian vanish there: the terms with first derivatives vanish by construction, while
P ( Δ e i , e i ) = λ i ( P ) h ( Δ e i , e i ) = 0. P(\Delta e_i,e_i) = \lambda_i(P)h(\Delta e_i,e_i) = 0. P ( Δ e i , e i ) = λ i ( P ) h ( Δ e i , e i ) = 0.
It follows from (13) that
Δ ϕ ( p ) ≤ Δ s f r ( p ) = ∑ i = 1 k ( Δ P ) ( e i , e i ) ≤ 2 ( n + 1 ) s k ( p ) , \Delta\phi(p) \le\Delta s_{\mathrm{fr}}(p) = \sum_{i=1}^{k}(\Delta P)(e_i,e_i) \le2(n+1)s_k(p), Δ ϕ ( p ) ≤ Δ s fr ( p ) = i = 1 ∑ k ( Δ P ) ( e i , e i ) ≤ 2 ( n + 1 ) s k ( p ) ,
because P 2 P^2 P 2 , R 2 \mathcal{R}^2 R 2 , and H \mathcal{H} H are nonnegative. This proves (18) with exactly the lower-test interpretation in Lemma 2.3.
Let d n u l l d_{\mathrm{null}} d null be the largest nullity attained by P P P . If d n u l l = 0 d_{\mathrm{null}} = 0 d null = 0 , its nullity is already constant. Otherwise s d n u l l s_{d_{\mathrm{null}}} s d null vanishes somewhere, and Lemma 2.3 gives s d n u l l ≡ 0 s_{d_{\mathrm{null}}} \equiv0 s d null ≡ 0 . The nullity is therefore at least d n u l l d_{\mathrm{null}} d null everywhere, and maximality makes it exactly d n u l l d_{\mathrm{null}} d null . The constant-rank theorem now makes N = ker P \mathcal{N} = \ker P N = ker P a smooth subbundle.
For a local section v v v of N \mathcal{N} N , differentiating P ( v , ⋅ ) = 0 P(v,\cdot) = 0 P ( v , ⋅ ) = 0 gives
( ∇ a P ) ( v , W ) = − P ( ∇ a v , W ) . (\nabla_a P)(v,W) = -P(\nabla_a v,W). ( ∇ a P ) ( v , W ) = − P ( ∇ a v , W ) .
Twice differentiating the identically zero scalar P ( v , v ) P(v,v) P ( v , v ) , in a normal orthonormal frame at the point of calculation, gives
0 = ( Δ P ) ( v , v ) + 4 ∑ a ( ∇ a P ) ( ∇ a v , v ) + 2 P ( Δ v , v ) + 2 ∑ a P ( ∇ a v , ∇ a v ) = ( Δ P ) ( v , v ) − 2 ∑ a P ( ∇ a v , ∇ a v ) . \begin{aligned}
0 &= (\Delta P)(v,v) + 4\sum_a(\nabla_a P)(\nabla_a v,v) + 2P(\Delta v,v) + 2\sum_a P(\nabla_a v,\nabla_a v) \\
&= (\Delta P)(v,v) - 2\sum_a P(\nabla_a v,\nabla_a v).
\end{aligned} 0 = ( Δ P ) ( v , v ) + 4 a ∑ ( ∇ a P ) ( ∇ a v , v ) + 2 P ( Δ v , v ) + 2 a ∑ P ( ∇ a v , ∇ a v ) = ( Δ P ) ( v , v ) − 2 a ∑ P ( ∇ a v , ∇ a v ) .
On the other hand, P v = 0 Pv = 0 P v = 0 in (13) gives the exact identity
∑ a P ( ∇ a v , ∇ a v ) + R 2 ( v , v ) + H ( v , v ) = 0. (19) \sum_a P(\nabla_a v,\nabla_a v) + \mathcal{R}^2(v,v) + \mathcal{H}(v,v) = 0. \tag*{(19)} a ∑ P ( ∇ a v , ∇ a v ) + R 2 ( v , v ) + H ( v , v ) = 0. (19)
All summands are nonnegative. Hence ∇ a v ∈ N \nabla_a v \in\mathcal{N} ∇ a v ∈ N , which proves that N \mathcal{N} N is parallel. The equality R 2 ( v , v ) = 0 \mathcal{R}^2(v,v)=0 R 2 ( v , v ) = 0 means
∑ a ∥ Rm ( v , e a ) ∥ H S 2 = 0 \sum_a \lVert\operatorname{Rm}(v,e_a)\rVert_{\mathrm{HS}}^2=0 a ∑ ∥ Rm ( v , e a ) ∥ HS 2 = 0
and the equality H ( v , v ) = 0 \mathcal{H}(v,v)=0 H ( v , v ) = 0 means
∑ k , a , b ( ( ∇ e k A ) ( v , e a , e b ) ) 2 = 0. \sum_{k,a,b}\bigl((\nabla_{e_k}A)(v,e_a,e_b)\bigr)^2=0. k , a , b ∑ ( ( ∇ e k A ) ( v , e a , e b ) ) 2 = 0.
These are precisely (17) . □ \square □
Proposition 2.5. If the Blaschke Ricci tensor has a nonzero null vector at one point, then the cone C C C is a nontrivial direct product of proper open convex cones.
Proof. By Lemma 2.4, the parallel bundle N = ker P \mathcal{N}=\ker P N = ker P has positive rank and satisfies Rm ( v , X ) = 0 \operatorname{Rm}(v,X)=0 Rm ( v , X ) = 0 for v ∈ N v\in\mathcal{N} v ∈ N . The first Bianchi identity gives
Rm ( X , Y ) v = − Rm ( Y , v ) X − Rm ( v , X ) Y = 0. \operatorname{Rm}(X,Y)v=-\operatorname{Rm}(Y,v)X-\operatorname{Rm}(v,X)Y=0. Rm ( X , Y ) v = − Rm ( Y , v ) X − Rm ( v , X ) Y = 0.
Thus the induced connection on N \mathcal{N} N is flat. Since Ω \Omega Ω is simply connected, every vector in a fibre of N \mathcal{N} N extends to a unique global parallel section of N \mathcal{N} N .
Let V \mathcal{V} V be the full vector space of global parallel vector fields on ( Ω , h ) (\Omega,h) ( Ω , h ) . Every such field W W W belongs to N \mathcal{N} N : parallelism gives Rm ( X , Y ) W = 0 \operatorname{Rm}(X,Y)W=0 Rm ( X , Y ) W = 0 , and hence Ric ( W , W ) = 0 \operatorname{Ric}(W,W)=0 Ric ( W , W ) = 0 ; since P ≥ 0 P\geq0 P ≥ 0 , this implies P W = 0 PW=0 P W = 0 . Consequently evaluation at any point identifies V \mathcal{V} V with the corresponding fibre of N \mathcal{N} N . In particular V \mathcal{V} V is a nonzero finite-dimensional Euclidean space, with the inner product given by h h h on parallel fields.
For X , Y ∈ V X,Y\in\mathcal{V} X , Y ∈ V , equation (17) gives, for arbitrary Z , W Z,W Z , W ,
h ( ( ∇ Z ( A X Y ) ) , W ) = ( ∇ Z A ) ( X , Y , W ) = 0. h\bigl((\nabla_Z(A_{X}Y)),W\bigr)=(\nabla_ZA)(X,Y,W)=0. h ( ( ∇ Z ( A X Y )) , W ) = ( ∇ Z A ) ( X , Y , W ) = 0.
Therefore A X Y A_XY A X Y is a global parallel field, proving A ( V , V ) ⊂ V A(\mathcal{V},\mathcal{V})\subset\mathcal{V} A ( V , V ) ⊂ V . The same calculation shows that A ( X , Y , Z ) A(X,Y,Z) A ( X , Y , Z ) is constant for X , Y , Z ∈ V X,Y,Z\in\mathcal{V} X , Y , Z ∈ V . We may thus regard the restriction of A A A as a fixed symmetric cubic on V \mathcal{V} V .
Choose a unit vector u ∈ V u\in\mathcal{V} u ∈ V at which A ( u , u , u ) A(u,u,u) A ( u , u , u ) is maximal on the unit sphere of V \mathcal{V} V . The Lagrange multiplier equation, together with A u u ∈ V A_{uu}\in\mathcal{V} A uu ∈ V , gives
A u u = a u , a = A ( u , u , u ) , (20) A_{uu}=au,\qquad a=A(u,u,u), \tag*{(20)} A uu = a u , a = A ( u , u , u ) , (20)
where a a a is constant. By self-adjointness A u A_u A u preserves u ⊥ u^\perp u ⊥ . For X ⊥ u X\perp u X ⊥ u , the Gauss equation and (17) give
0 = Rm ( u , X ) u = X − A u 2 X + a A u X . 0=\operatorname{Rm}(u,X)u=X-A_u^2X+aA_uX. 0 = Rm ( u , X ) u = X − A u 2 X + a A u X .
Thus
A u 2 − a A u − id = 0 on u ⊥ . (21) A_u^2-aA_u-\operatorname{id}=0\quad\text{on }u^\perp. \tag*{(21)} A u 2 − a A u − id = 0 on u ⊥ . (21)
Moreover u u u and A u A_u A u are parallel, the latter because
h ( ( ∇ X A u ) Y , Z ) = ( ∇ X A ) ( u , Y , Z ) = 0. h\bigl((\nabla_XA_u)Y,Z\bigr)=(\nabla_XA)(u,Y,Z)=0. h ( ( ∇ X A u ) Y , Z ) = ( ∇ X A ) ( u , Y , Z ) = 0.
The radial vector f f f is transverse to the tangent space of the affine sphere. Define an ambient endomorphism Θ \Theta Θ at each point by
Θ f = f ∗ u , Θ f ∗ Y = f ∗ A u Y + h ( u , Y ) f . (22) \Theta f=f_*u,\qquad\Theta f_*Y=f_*A_uY+h(u,Y)f. \tag*{(22)} Θ f = f ∗ u , Θ f ∗ Y = f ∗ A u Y + h ( u , Y ) f . (22)
We show that it is a constant matrix. Using the structure equation (3) and the parallelism just proved gives
( D X Θ ) f = f ∗ ( A X u − A u X ) = 0. (D_X\Theta)f=f_*(A_Xu-A_uX)=0. ( D X Θ ) f = f ∗ ( A X u − A u X ) = 0.
On tangent vectors, the same differentiation gives
( D X Θ ) f ∗ Y = f ∗ ( [ A X , A u ] Y + h ( u , Y ) X − h ( X , Y ) u ) + ( h ( X , A u Y ) − h ( u , A X Y ) ) f = − f ∗ Rm ( X , u ) Y = 0. \begin{aligned}
(D_X\Theta)f_*Y
&=f_*\bigl([A_X,A_u]Y+h(u,Y)X-h(X,Y)u\bigr)\\
&\quad+\bigl(h(X,A_uY)-h(u,A_XY)\bigr)f\\
&=-f_*\operatorname{Rm}(X,u)Y=0.
\end{aligned} ( D X Θ ) f ∗ Y = f ∗ ( [ A X , A u ] Y + h ( u , Y ) X − h ( X , Y ) u ) + ( h ( X , A u Y ) − h ( u , A X Y ) ) f = − f ∗ Rm ( X , u ) Y = 0.
The radial coefficient vanishes by cubic symmetry, and the last equality follows from (17) . These equalities hold on a basis of the ambient space, so D X Θ = 0 D_X\Theta=0 D X Θ = 0 for every X X X . Connectedness makes Θ \Theta Θ a single constant endomorphism of R N \mathbb{R}^N R N . On span ( f , f ∗ u ) \operatorname{span}(f,f_*u) span ( f , f ∗ u ) , its matrix is
( 0 1 1 a ) \begin{pmatrix} 0 & 1 \\ 1 & a \end{pmatrix} ( 0 1 1 a )
The complementary invariant space f ∗ ( u ⊥ ) f_*(u^\perp) f ∗ ( u ⊥ ) has Θ = A u \Theta=A_u Θ = A u and satisfies (21) . Consequently Θ \Theta Θ is diagonalizable and has exactly the two distinct real eigenvalues
λ ± = a ± a 2 + 4 2 . \lambda_\pm=\frac{a\pm\sqrt{a^2+4}}{2}. λ ± = 2 a ± a 2 + 4 .
Both eigenvalues occur already on the displayed two-dimensional space. Denote their nonzero ambient eigenspaces by L + L_+ L + and L − L_- L − .
The parallel unit field u u u is complete, since its integral curves are unit geodesics of the complete metric h h h . Let ϕ t \phi_t ϕ t be its global flow. Equation (22) gives
d d t f ( ϕ t ( x ) ) = Θ f ( ϕ t ( x ) ) , f ∘ ϕ t = e t Θ f . \frac{d}{dt}f(\phi_t(x))=\Theta f(\phi_t(x)),\qquad f\circ\phi_t=e^{t\Theta}f. d t d f ( ϕ t ( x )) = Θ f ( ϕ t ( x )) , f ∘ ϕ t = e t Θ f .
The affine sphere meets every ray of C C C once, so e t Θ C = C e^{t\Theta}C=C e t Θ C = C for all t ∈ R t\in\mathbb{R} t ∈ R . Combining this action with radial multiplication allows independent positive scaling on L + L_+ L + and L − L_- L − . Indeed, prescribed factors s + , s − > 0 s_+,s_->0 s + , s − > 0 are obtained from r e t Θ re^{t\Theta} r e t Θ by taking
t = log s + − log s − λ + − λ − , log r = log s + − t λ + . t=\frac{\log s_+-\log s_-}{\lambda_+-\lambda_-},\qquad\log r=\log s_+-t\lambda_+. t = λ + − λ − log s + − log s − , log r = log s + − t λ + .
Set C ± = pr L ± C C_\pm=\operatorname{pr}_{L_\pm}C C ± = pr L ± C . These are nonempty open full-dimensional convex cones in L ± L_\pm L ± . We claim that
C = C + × C − in R N = L + ⊕ L − . (23) C=C_+\times C_-\quad\text{in }\mathbb{R}^N=L_+\oplus L_-. \tag*{(23)} C = C + × C − in R N = L + ⊕ L − . (23)
For x + ∈ C + x_+\in C_+ x + ∈ C + and x − ∈ C − x_-\in C_- x − ∈ C − choose ( x + , y − ) ∈ C (x_+,y_-)\in C ( x + , y − ) ∈ C and ( y + , x − ) ∈ C (y_+,x_-)\in C ( y + , x − ) ∈ C . Independent positive scaling and addition give
( x + + η y + , x − + η y − ) ∈ C ( η > 0 ) . (x_++\eta y_+,x_-+\eta y_-)\in C\qquad(\eta>0). ( x + + η y + , x − + η y − ) ∈ C ( η > 0 ) .
Letting η \eta η tend to zero proves C + ‾ × C − ‾ ⊂ C ‾ \overline{C_+}\times\overline{C_-}\subset\overline{C} C + × C − ⊂ C . Since the product is open, it is contained in int C ‾ = C \operatorname{int}\overline{C}=C int C = C . The reverse inclusion is immediate from the definitions, proving (23) . Finally C ‾ = C + ‾ × C − ‾ \overline{C}=\overline{C_+}\times\overline{C_-} C = C + × C − is pointed, so each closed factor is pointed. Thus both open factors are proper. Their ambient spaces are nonzero, making this product decomposition nontrivial. □ \square □
Proposition 2.6. If C C C is irreducible, there is a constant ε > 0 \varepsilon>0 ε > 0 such that
− ( n − 1 ) h ≤ Ric h ≤ − ε h on Ω . (24) -(n-1)h\leq\operatorname{Ric}_h\leq-\varepsilon h\quad\text{on }\Omega. \tag*{(24)} − ( n − 1 ) h ≤ Ric h ≤ − ε h on Ω. (24)
Proof. Proposition 2.2 gives nonpositive Ricci curvature. Proposition 2.5 and irreducibility exclude a null direction at every point, so P = − Ric h P=-\operatorname{Ric}_h P = − Ric h is positive definite. Its smallest eigenvalue is a continuous positive Γ \Gamma Γ -invariant function. It has a positive minimum ε \varepsilon ε on a compact set F \mathcal{F} F with Γ F = Ω \Gamma\mathcal{F}=\Omega Γ F = Ω . This proves the upper bound in (24) ; the lower bound is Proposition 2.2 . Together with Proposition 2.5 , this proves Theorem D. □ \square □
The affine-sphere Gauss map and Cartan projections
In this section we assume that C C C is irreducible. By Proposition 2.6, there is ε > 0 \varepsilon>0 ε > 0 such that
Ric h ≤ − ε h on Ω . (25) \operatorname{Ric}_h\leq-\varepsilon h\quad\text{on }\Omega. \tag*{(25)} Ric h ≤ − ε h on Ω. (25)
We write ∇ \nabla ∇ for the Levi–Civita connection of h h h . Our aim is to convert this curvature bound into a restriction on the Cartan projections of cone automorphisms. The comparison will use the Blaschke length of projective segments.
Fix p ∈ Ω p\in\Omega p ∈ Ω . For each x ∈ Ω x\in\Omega x ∈ Ω , let G x G_x G x be the positive definite form on R N \mathbb{R}^N R N characterized by
G x ( f ( x ) , f ( x ) ) = 1 , G x ( f ( x ) , f ∗ T x Ω ) = 0 , G x ( f ∗ X , f ∗ Y ) = h x ( X , Y ) . G_x(f(x),f(x))=1,\qquad G_x(f(x),f_*T_x\Omega)=0,\qquad G_x(f_*X,f_*Y)=h_x(X,Y). G x ( f ( x ) , f ( x )) = 1 , G x ( f ( x ) , f ∗ T x Ω ) = 0 , G x ( f ∗ X , f ∗ Y ) = h x ( X , Y ) .
We denote the norm defined by G p G_p G p by ∥ ⋅ ∥ p \lVert\cdot\rVert_p ∥ ⋅ ∥ p , and its dual norm by ∥ ⋅ ∥ p , ∗ \lVert\cdot\rVert_{p,*} ∥ ⋅ ∥ p , ∗ . For a cone-preserving matrix g g g normalized by ∣ det g ∣ = 1 \lvert\det g\rvert=1 ∣ det g ∣ = 1 , write
μ 1 p ( g ) ≥ ⋯ ≥ μ N p ( g ) , ∑ i μ i p ( g ) = 0 , \mu_1^p(g)\ge\cdots\ge\mu_N^p(g),\qquad\sum_i\mu_i^p(g)=0, μ 1 p ( g ) ≥ ⋯ ≥ μ N p ( g ) , i ∑ μ i p ( g ) = 0 ,
for its logarithmic singular values in this norm. We first work with μ p \mu^p μ p and return to the fixed ambient Euclidean norm at the end of the section.
The ambient positive definite forms
Let P \mathcal{P} P be the space of positive definite forms on R N \mathbb{R}^N R N having determinant det G p \det G_p det G p . Its Riemannian metric is normalized by
g G ( U , V ) = 1 4 tr ( G − 1 U G − 1 V ) . g_G(U,V)=\frac{1}{4}\operatorname{tr}(G^{-1}UG^{-1}V). g G ( U , V ) = 4 1 tr ( G − 1 U G − 1 V ) .
Here a tangent vector to P \mathcal{P} P is viewed as a symmetric matrix in any fixed ambient basis.
Lemma 3.1. The map G : x ↦ G x G:x\mapsto G_x G : x ↦ G x takes values in P \mathcal{P} P and satisfies
G ∗ g = N h + Ric h . (26) G^*g=Nh+\operatorname{Ric}_h. \tag*{(26)} G ∗ g = N h + Ric h . (26) It is equivariant for the action of { g ∈ GL ( N , R ) : ∣ det g ∣ = 1 } \{g\in\operatorname{GL}(N,\mathbb{R}):\lvert\det g\rvert=1\} { g ∈ GL ( N , R ) : ∣ det g ∣ = 1 } given by g ⋅ G = g − T G g − 1 g\cdot G=g^{-T}Gg^{-1} g ⋅ G = g − T G g − 1 . Moreover, for every cone automorphism g g g with ∣ det g ∣ = 1 \lvert\det g\rvert=1 ∣ det g ∣ = 1 , we have
d P ( G p , G g p ) 2 = ∑ i = 1 N μ i p ( g ) 2 . (27) d_{\mathcal{P}}(G_p,G_{gp})^2=\sum_{i=1}^{N}\mu_i^p(g)^2. \tag*{(27)} d P ( G p , G g p ) 2 = i = 1 ∑ N μ i p ( g ) 2 . (27)
Proof. Use the bundle identification
R ⊕ T Ω ⟶ Ω × R N , ( α , Y ) ⟼ α f + f ∗ Y . \mathbb{R}\oplus T\Omega\longrightarrow\Omega\times\mathbb{R}^N,\qquad(\alpha,Y)\longmapsto\alpha f+f_*Y. R ⊕ T Ω ⟶ Ω × R N , ( α , Y ) ⟼ α f + f ∗ Y .
In this identification, G G G is the product of the standard metric on R \mathbb{R} R and h h h on T Ω T\Omega T Ω . The structure equation (3) gives
D X ( α , Y ) = ( X α + h ( X , Y ) , ∇ X Y + A X Y + α X ) . D_X(\alpha,Y)=\left(X\alpha+h(X,Y),\nabla_XY+A_XY+\alpha X\right). D X ( α , Y ) = ( X α + h ( X , Y ) , ∇ X Y + A X Y + α X ) .
Consequently
D = ∇ ˉ + Ψ , ∇ ˉ X ( α , Y ) = ( X α , ∇ X Y ) , Ψ X = ( 0 X ♭ X A X ) . D=\bar{\nabla}+\Psi,\qquad\bar{\nabla}_X(\alpha,Y)=(X\alpha,\nabla_XY),\qquad\Psi_X=\begin{pmatrix}0&X^\flat\\X&A_X\end{pmatrix}. D = ∇ ˉ + Ψ , ∇ ˉ X ( α , Y ) = ( X α , ∇ X Y ) , Ψ X = ( 0 X X ♭ A X ) .
The connection ∇ ˉ \bar{\nabla} ∇ ˉ preserves G G G . Cubic symmetry makes Ψ X \Psi_X Ψ X self-adjoint for G G G , and tr Ψ X = tr A X = 0 \operatorname{tr}\Psi_X=\operatorname{tr}A_X=0 tr Ψ X = tr A X = 0 . Thus, in a flat ambient trivialization,
D X G = − 2 G Ψ X , X ( log det G ) = − 2 tr Ψ X = 0. D_XG=-2G\Psi_X,\qquad X(\log\det G)=-2\operatorname{tr}\Psi_X=0. D X G = − 2 G Ψ X , X ( log det G ) = − 2 tr Ψ X = 0.
Therefore, for every tangent vector X X X ,
∥ d G ( X ) ∥ g 2 = tr ( Ψ X 2 ) = 2 h ( X , X ) + tr ( A X 2 ) = N h ( X , X ) + Ric h ( X , X ) , \lVert dG(X)\rVert_g^2=\operatorname{tr}(\Psi_X^2)
=2h(X,X)+\operatorname{tr}(A_X^2)
=Nh(X,X)+\operatorname{Ric}_h(X,X), ∥ d G ( X ) ∥ g 2 = tr ( Ψ X 2 ) = 2 h ( X , X ) + tr ( A X 2 ) = N h ( X , X ) + Ric h ( X , X ) ,
where the last equality is (6) . Polarization proves (26) . In particular,
∥ d G ( X ) ∥ g 2 ≤ ( N − ε ) h ( X , X ) . (28) \lVert dG(X)\rVert_g^2\le(N-\varepsilon)h(X,X). \tag*{(28)} ∥ d G ( X ) ∥ g 2 ≤ ( N − ε ) h ( X , X ) . (28)
The canonical affine sphere and its Blaschke metric are preserved by every cone-preserving matrix of absolute determinant one. The definition of G x G_x G x therefore gives
G g x = g − T G x g − 1 . G_{gx}=g^{-T}G_xg^{-1}. G g x = g − T G x g − 1 .
This is the asserted equivariance.
We recall the distance formula with our normalization. After a linear change of basis taking one positive definite form to the identity, the distance from id \operatorname{id} id to a positive definite matrix G 0 G_0 G 0 is
d P ( id , G 0 ) 2 = 1 4 ∑ i ( log λ i ( G 0 ) ) 2 . d_{\mathcal{P}}(\operatorname{id},G_0)^2=\frac{1}{4}\sum_i\left(\log\lambda_i(G_0)\right)^2. d P ( id , G 0 ) 2 = 4 1 i ∑ ( log λ i ( G 0 ) ) 2 .
Choose a G p G_p G p -orthonormal ambient basis. In this basis G p = i d G_p=\mathrm{id} G p = id and G g p = g − T g − 1 G_{gp}=g^{-\mathsf{T}}g^{-1} G g p = g − T g − 1 . The eigenvalues of the latter matrix are the inverse squared singular values of g g g . The distance formula hence gives
d P ( G p , G g p ) 2 = 1 4 ∑ i ( − 2 μ i p ( g ) ) 2 , d_P(G_p,G_{gp})^2=\frac{1}{4}\sum_i\left(-2\mu_i^p(g)\right)^2, d P ( G p , G g p ) 2 = 4 1 i ∑ ( − 2 μ i p ( g ) ) 2 ,
which proves (27) .
Complete projective chords
We next control the length of a projective segment from p p p to g p gp g p . The first step records why completeness supplies a differential equation on the whole real line.
Lemma 3.2. Every complete projective chord in Ω \Omega Ω admits a Blaschke arclength parametrization c : R → Ω c:\mathbb{R}\to\Omega c : R → Ω . Given two distinct points p , q p,q p , q on that chord, choose its orientation so that p p p precedes q q q . Let [ a ] [a] [ a ] be the endpoint of the chord beyond q q q in the forward direction and [ b ] [b] [ b ] its other endpoint, beyond p p p ; choose unit lifts a , b ∈ ∂ C ∖ { 0 } a,b\in\partial C\setminus\{0\} a , b ∈ ∂ C ∖ { 0 } of these endpoints. There are smooth positive functions x , y x,y x , y on R \mathbb{R} R such that
f ( c ( s ) ) = x ( s ) a + y ( s ) b , x ′ ( s ) > 0 , y ′ ( s ) < 0. f(c(s))=x(s)a+y(s)b,\qquad x'(s)>0,\qquad y'(s)<0. f ( c ( s )) = x ( s ) a + y ( s ) b , x ′ ( s ) > 0 , y ′ ( s ) < 0. Writing T = c ˙ T=\dot{c} T = c ˙ , the affine-sphere lift satisfies
d 2 d s 2 ( f ∘ c ) = ϑ ( s ) d d s ( f ∘ c ) + f ∘ c , ϑ ( s ) = A ( T ( s ) , T ( s ) , T ( s ) ) , ∣ ϑ ( s ) ∣ ≤ n − 1 . (29) \frac{d^2}{ds^2}(f\circ c)=\vartheta(s)\frac{d}{ds}(f\circ c)+f\circ c,\qquad\vartheta(s)=A(T(s),T(s),T(s)),\qquad|\vartheta(s)|\le\sqrt{n-1}. \tag*{(29)} d s 2 d 2 ( f ∘ c ) = ϑ ( s ) d s d ( f ∘ c ) + f ∘ c , ϑ ( s ) = A ( T ( s ) , T ( s ) , T ( s )) , ∣ ϑ ( s ) ∣ ≤ n − 1 . (29)
Proof. In a bounded affine realization of Ω \Omega Ω , the chord is an open interval whose endpoints lie on ∂ Ω \partial\Omega ∂ Ω . A regular parametrization can be reparametrized by h h h -arclength on an interval ( s − , s + ) (s_-,s_+) ( s − , s + ) . Suppose, for example, that s + < ∞ s_+<\infty s + < ∞ . The curve is then Cauchy as s s s tends to s + s_+ s + , since its Riemannian distance between two points is at most their intervening arclength. Completeness of h h h would give a limit in Ω \Omega Ω . This contradicts convergence of the projective chord to its boundary endpoint. Thus s + = + ∞ s_+=+\infty s + = + ∞ , and the same argument gives s − = − ∞ s_-=-\infty s − = − ∞ .
Every point of the chord has a unique positive expression in the rays of a a a and b b b , so f ∘ c = x a + y b f\circ c=xa+yb f ∘ c = x a + y b for smooth x , y > 0 x,y>0 x , y > 0 . The canonical conormal belongs to the interior dual cone. In particular,
ν c ( s ) ( a ) > 0 , ν c ( s ) ( b ) > 0. \nu_{c(s)}(a)>0,\qquad\nu_{c(s)}(b)>0. ν c ( s ) ( a ) > 0 , ν c ( s ) ( b ) > 0.
The conormal identities give
0 = ν c ( s ) ( ( f ∘ c ) ′ ) = x ′ ν c ( s ) ( a ) + y ′ ν c ( s ) ( b ) . 0=\nu_{c(s)}\left((f\circ c)'\right)=x'\nu_{c(s)}(a)+y'\nu_{c(s)}(b). 0 = ν c ( s ) ( ( f ∘ c ) ′ ) = x ′ ν c ( s ) ( a ) + y ′ ν c ( s ) ( b ) .
The derivative ( f ∘ c ) ′ (f\circ c)' ( f ∘ c ) ′ is nonzero because c c c has unit speed. Thus x ′ , y ′ x',y' x ′ , y ′ are nonzero and have opposite signs. The projective coordinate x / y x/y x / y increases in the direction toward [ a ] [a] [ a ] , and its derivative cannot vanish for a regular chord. Therefore
d d s ( x / y ) = x ′ y − x y ′ y 2 > 0 \frac{d}{ds}(x/y)=\frac{x'y-xy'}{y^2}>0 d s d ( x / y ) = y 2 x ′ y − x y ′ > 0
selects x ′ > 0 x'>0 x ′ > 0 and y ′ < 0 y'<0 y ′ < 0 .
The vectors f ∘ c f\circ c f ∘ c and ( f ∘ c ) ′ (f\circ c)' ( f ∘ c ) ′ span the fixed ambient two-plane of the chord. The structure equation gives
( f ∘ c ) ′ ′ = f ∗ ( ∇ T T + A T T ) + f ∘ c . (f\circ c)''=f_*(\nabla_TT+A_TT)+f\circ c. ( f ∘ c ) ′′ = f ∗ ( ∇ T T + A T T ) + f ∘ c .
Since f f f is transverse to the affine-sphere tangent space, the intersection of this two-plane with that tangent space is precisely the line spanned by ( f ∘ c ) ′ (f\circ c)' ( f ∘ c ) ′ . It follows that
∇ T T + A T T = ϑ T . \nabla_TT+A_TT=\vartheta T. ∇ T T + A T T = ϑT .
Pairing with the unit vector T T T gives ϑ = A ( T , T , T ) \vartheta=A(T,T,T) ϑ = A ( T , T , T ) , because h ( ∇ T T , T ) = 0 h(\nabla_TT,T)=0 h ( ∇ T T , T ) = 0 . Finally Proposition 2.2 gives
∣ ϑ ∣ ≤ ∥ A T ∥ H S = tr ( A T 2 ) ≤ n − 1 . |\vartheta|\le\lVert A_T\rVert_{\mathrm{HS}}=\sqrt{\operatorname{tr}(A_T^2)}\le\sqrt{n-1}. ∣ ϑ ∣ ≤ ∥ A T ∥ HS = tr ( A T 2 ) ≤ n − 1 .
This proves (29) .
The next estimate uses only the bound on the cubic coefficient. In particular, the strict Ricci inequality is not needed to control the following scalar differential equations.
Lemma 3.3. Let ϑ : R → R \vartheta:\mathbb{R}\to\mathbb{R} ϑ : R → R be continuous with ∣ ϑ ∣ ≤ κ |\vartheta|\leq\kappa ∣ ϑ ∣ ≤ κ . Suppose that positive C 1 C^1 C 1 functions u , v u,v u , v on R \mathbb{R} R satisfy
u ′ = 1 + ϑ u − u 2 , v ′ = − 1 + ϑ v + v 2 . u' = 1+\vartheta u-u^2,\qquad v'=-1+\vartheta v+v^2. u ′ = 1 + ϑ u − u 2 , v ′ = − 1 + ϑ v + v 2 . Then
u − ≤ u , v ≤ u + , u + = κ + κ 2 + 4 2 , u − = u + − 1 . u_-\leq u,v\leq u_+,\qquad u_+=\frac{\kappa+\sqrt{\kappa^2+4}}{2},\qquad u_-=u_+^{-1}. u − ≤ u , v ≤ u + , u + = 2 κ + κ 2 + 4 , u − = u + − 1 . Consequently
c κ ≤ u u + v ≤ 1 − c κ , c κ = u − u − + u + > 0. (30) c_\kappa\leq\frac{u}{u+v}\leq1-c_\kappa,\qquad c_\kappa=\frac{u_-}{u_-+u_+}>0. \tag*{(30)} c κ ≤ u + v u ≤ 1 − c κ , c κ = u − + u + u − > 0. (30)
Proof. The constants satisfy u + u − = 1 u_+u_-=1 u + u − = 1 and u + − u − = κ u_+-u_-=\kappa u + − u − = κ , so
1 + κ z − z 2 = − ( z − u + ) ( z + u − ) , 1 − κ z − z 2 = ( u − − z ) ( z + u + ) . 1+\kappa z-z^2=-(z-u_+)(z+u_-),\qquad1-\kappa z-z^2=(u_--z)(z+u_+). 1 + κ z − z 2 = − ( z − u + ) ( z + u − ) , 1 − κ z − z 2 = ( u − − z ) ( z + u + ) .
Suppose that u ( s 0 ) > u + u(s_0)>u_+ u ( s 0 ) > u + , and reverse time by setting z ( τ ) = u ( s 0 − τ ) z(\tau)=u(s_0-\tau) z ( τ ) = u ( s 0 − τ ) . The differential inequality for u u u gives
z ′ ≥ ( z − u + ) ( z + u − ) > 0 z'\geq(z-u_+)(z+u_-)>0 z ′ ≥ ( z − u + ) ( z + u − ) > 0
as long as z > u + z>u_+ z > u + . In particular z z z increases and blows up in finite time, contradicting its existence on all of R \mathbb{R} R .
If 0 < u ( s 0 ) < u − 0<u(s_0)<u_- 0 < u ( s 0 ) < u − , the reversed function satisfies
z ′ ≤ − ( u − − z ) ( z + u + ) . z'\leq-(u_--z)(z+u_+). z ′ ≤ − ( u − − z ) ( z + u + ) .
It decreases while 0 < z < u − 0<z<u_- 0 < z < u − . On 0 < z ≤ u ( s 0 ) 0<z\leq u(s_0) 0 < z ≤ u ( s 0 ) its derivative is at most − ( u − − u ( s 0 ) ) u + < 0 -(u_--u(s_0))u_+<0 − ( u − − u ( s 0 )) u + < 0 , so it reaches zero in finite time. This contradicts positivity. These two arguments prove u − ≤ u ≤ u + u_-\leq u\leq u_+ u − ≤ u ≤ u + .
The corresponding contradictions for v v v occur in forward time. If v ( s 0 ) > u + v(s_0)>u_+ v ( s 0 ) > u + , then
v ′ ≥ ( v − u + ) ( v + u − ) > 0 , v'\geq(v-u_+)(v+u_-)>0, v ′ ≥ ( v − u + ) ( v + u − ) > 0 ,
which forces finite-time blow-up. If 0 < v ( s 0 ) < u − 0<v(s_0)<u_- 0 < v ( s 0 ) < u − , then
v ′ ≤ − ( u − − v ) ( v + u + ) , v'\leq-(u_--v)(v+u_+), v ′ ≤ − ( u − − v ) ( v + u + ) ,
which forces a finite-time zero. Thus u − ≤ v ≤ u + u_-\leq v\leq u_+ u − ≤ v ≤ u + as well. The bounds in (30) follow by taking the minimum and maximum of u / ( u + v ) u/(u+v) u / ( u + v ) on [ u − , u + ] 2 [u_-,u_+]^2 [ u − , u + ] 2 .
For p , q ∈ Ω p,q\in\Omega p , q ∈ Ω , denote by L ( p , q ) L(p,q) L ( p , q ) the Blaschke length of their projective segment. We emphasize that this is the length of that particular segment, rather than the Blaschke distance.
Lemma 3.4. There is a positive constant c p c_p c p , depending only on the fixed cone and basepoint, such that every cone-preserving automorphism g g g with ∣ det g ∣ = 1 |\det g|=1 ∣ det g ∣ = 1 satisfies
L ( p , g p ) 2 ≤ ( μ 1 p ( g ) + c p ) ( − μ N p ( g ) + c p ) . (31) L(p,gp)^2\leq\left(\mu_1^p(g)+c_p\right)\left(-\mu_N^p(g)+c_p\right). \tag*{(31)} L ( p , g p ) 2 ≤ ( μ 1 p ( g ) + c p ) ( − μ N p ( g ) + c p ) . (31)
Proof. Assume first that q = g p ≠ p q=gp\neq p q = g p = p . Orient the complete chord from p p p toward q q q , so that its arclength parameter satisfies s ( p ) < s ( q ) s(p)<s(q) s ( p ) < s ( q ) . In Lemma 3.2 , [ a ] [a] [ a ] is then the boundary endpoint beyond q q q , while [ b ] [b] [ b ] lies beyond $p; normalize a , b a,b a , b to have unit length in ∥ ⋅ ∥ p \|\cdot\|_p ∥ ⋅ ∥ p . All integrals ∫ p q \int_p^q ∫ p q below are consequently taken in the positive s s s -direction. Set
u = x ′ x > 0 , v = − y ′ y > 0. u=\frac{x'}{x}>0,\qquad v=-\frac{y'}{y}>0. u = x x ′ > 0 , v = − y y ′ > 0.
Equation (29) gives the system in Lemma 3.3 , with κ = n − 1 \kappa=\sqrt{n-1} κ = n − 1 . In particular, ζ = u / ( u + v ) \zeta=u/(u+v) ζ = u / ( u + v ) satisfies (30) . Direct differentiation of the two equations gives
( log ζ ) ′ = u ′ u − u ′ + v ′ u + v = ( ϑ + 1 u − u ) − ( ϑ + v − u ) = 1 u − v . (\log\zeta)'=\frac{u'}{u}-\frac{u'+v'}{u+v}=\left(\vartheta+\frac{1}{u}-u\right)-(\vartheta+v-u)=\frac{1}{u}-v. ( log ζ ) ′ = u u ′ − u + v u ′ + v ′ = ( ϑ + u 1 − u ) − ( ϑ + v − u ) = u 1 − v .
Define
α = log x ( q ) x ( p ) = ∫ p q u d s > 0 , β = log y ( q ) y ( p ) = − ∫ p q v d s < 0. \alpha=\log\frac{x(q)}{x(p)}=\int_p^q u\,ds>0,\qquad\beta=\log\frac{y(q)}{y(p)}=-\int_p^q v\,ds<0. α = log x ( p ) x ( q ) = ∫ p q u d s > 0 , β = log y ( p ) y ( q ) = − ∫ p q v d s < 0.
It follows that
∫ p q d s u = − β + log ζ ( q ) ζ ( p ) ≤ − β + c c h , c c h = log 1 − c κ c κ . \int_p^q\frac{ds}{u}=-\beta+\log\frac{\zeta(q)}{\zeta(p)}\leq-\beta+c_{\mathrm{ch}},\qquad c_{\mathrm{ch}}=\log\frac{1-c_\kappa}{c_\kappa}. ∫ p q u d s = − β + log ζ ( p ) ζ ( q ) ≤ − β + c ch , c ch = log c κ 1 − c κ .
Cauchy–Schwarz therefore gives
L ( p , q ) 2 ≤ ( ∫ p q u d s ) ( ∫ p q d s u ) ≤ α ( − β + c c h ) . (32) L(p,q)^2\leq\left(\int_p^q u\,ds\right)\left(\int_p^q\frac{ds}{u}\right)\leq\alpha(-\beta+c_{\mathrm{ch}}). \tag*{(32)} L ( p , q ) 2 ≤ ( ∫ p q u d s ) ( ∫ p q u d s ) ≤ α ( − β + c ch ) . (32)
We next give bounds on the coefficients α \alpha α and β \beta β at the fixed basepoint p p p . Choose ℓ ∈ C ∗ \ell\in C^* ℓ ∈ C ∗ , and let
S p = { w ∈ R N : ∥ w ∥ p = 1 } , m ℓ = min C ‾ ∩ S p ℓ > 0 , d p = dist ∥ ⋅ ∥ p ( f ( p ) , ∂ C ) > 0. S_p=\{w\in\mathbb{R}^N:\|w\|_p=1\},\qquad m_\ell=\min_{\overline{C}\cap S_p}\ell>0,\qquad d_p=\operatorname{dist}_{\|\cdot\|_p}(f(p),\partial C)>0. S p = { w ∈ R N : ∥ w ∥ p = 1 } , m ℓ = C ∩ S p min ℓ > 0 , d p = dist ∥ ⋅ ∥ p ( f ( p ) , ∂ C ) > 0.
Since C ‾ \overline{C} C is pointed and full-dimensional, its dual cone has nonempty interior. Thus a strictly positive ℓ \ell ℓ can be chosen; C ‾ ∩ S p \overline{C}\cap S_p C ∩ S p is compact, and ℓ \ell ℓ has a positive minimum there. Applying ℓ \ell ℓ to f ( p ) = x ( p ) a + y ( p ) b f(p)=x(p)a+y(p)b f ( p ) = x ( p ) a + y ( p ) b gives
x ( p ) , y ( p ) ≤ U p : = ℓ ( f ( p ) ) m ℓ . x(p),y(p)\leq U_p:=\frac{\ell(f(p))}{m_\ell}. x ( p ) , y ( p ) ≤ U p := m ℓ ℓ ( f ( p )) .
For a lower bound, the vector y ( p ) b y(p)b y ( p ) b lies in ∂ C \partial C ∂ C , and its distance from f ( p ) f(p) f ( p ) is x ( p ) x(p) x ( p ) , since ∥ a ∥ p = 1 \|a\|_p=1 ∥ a ∥ p = 1 . Interchanging the endpoints gives
d p ≤ x ( p ) , y ( p ) ≤ U p . (33) d_p\leq x(p),y(p)\leq U_p. \tag*{(33)} d p ≤ x ( p ) , y ( p ) ≤ U p . (33)
At q q q , positivity gives
m ℓ x ( q ) ≤ ℓ ( f ( q ) ) ≤ ∥ ℓ ∥ p , ∗ ∥ g f ( p ) ∥ p ≤ ∥ ℓ ∥ p , ∗ e μ 1 p ( g ) . m_\ell x(q)\leq\ell(f(q))\leq\|\ell\|_{p,*}\|g f(p)\|_p\leq\|\ell\|_{p,*}e^{\mu_1^p(g)}. m ℓ x ( q ) ≤ ℓ ( f ( q )) ≤ ∥ ℓ ∥ p , ∗ ∥ g f ( p ) ∥ p ≤ ∥ ℓ ∥ p , ∗ e μ 1 p ( g ) .
Here ∥ f ( p ) ∥ p = 1 \|f(p)\|_p=1 ∥ f ( p ) ∥ p = 1 by definition. Combining this with (33) yields
α ≤ μ 1 p ( g ) + log ∥ ℓ ∥ p , ∗ m ℓ d p . (34) \alpha\leq\mu_1^p(g)+\log\frac{\|\ell\|_{p,*}}{m_\ell d_p}. \tag*{(34)} α ≤ μ 1 p ( g ) + log m ℓ d p ∥ ℓ ∥ p , ∗ . (34)
For the other endpoint, the two conormal equations are
x ( q ) ν q ( a ) + y ( q ) ν q ( b ) = 1 , u ( q ) x ( q ) ν q ( a ) − v ( q ) y ( q ) ν q ( b ) = 0. x(q)\nu_q(a)+y(q)\nu_q(b)=1,\qquad u(q)x(q)\nu_q(a)-v(q)y(q)\nu_q(b)=0. x ( q ) ν q ( a ) + y ( q ) ν q ( b ) = 1 , u ( q ) x ( q ) ν q ( a ) − v ( q ) y ( q ) ν q ( b ) = 0.
Solving them gives
y ( q ) ν q ( b ) = u ( q ) u ( q ) + v ( q ) ≥ c κ . y(q)\nu_q(b)=\frac{u(q)}{u(q)+v(q)}\geq c_\kappa. y ( q ) ν q ( b ) = u ( q ) + v ( q ) u ( q ) ≥ c κ .
The functional ν p \nu_p ν p is the G p G_p G p -dual of the unit vector f ( p ) f(p) f ( p ) , so ∥ ν p ∥ p , ∗ = 1 \|\nu_p\|_{p,*}=1 ∥ ν p ∥ p , ∗ = 1 . Conormal equivariance and ∥ b ∥ p = 1 \|b\|_p=1 ∥ b ∥ p = 1 therefore imply
ν q ( b ) ≤ ∥ ν q ∥ p , ∗ = ∥ g − T ν p ∥ p , ∗ ≤ e − μ N p ( g ) . \nu_q(b)\leq\|\nu_q\|_{p,*}=\|g^{-T}\nu_p\|_{p,*}\leq e^{-\mu_N^p(g)}. ν q ( b ) ≤ ∥ ν q ∥ p , ∗ = ∥ g − T ν p ∥ p , ∗ ≤ e − μ N p ( g ) .
Consequently y ( q ) ≥ c κ e μ N p ( g ) y(q)\geq c_\kappa e^{\mu_N^p(g)} y ( q ) ≥ c κ e μ N p ( g ) . Using y ( p ) ≤ U p y(p)\leq U_p y ( p ) ≤ U p gives
− β ≤ − μ N p ( g ) + log ( U p / c κ ) . (35) -\beta\leq-\mu_N^p(g)+\log(U_p/c_\kappa). \tag*{(35)} − β ≤ − μ N p ( g ) + log ( U p / c κ ) . (35)
The determinant normalization ensures μ 1 p ( g ) ≥ 0 \mu_1^p(g)\geq0 μ 1 p ( g ) ≥ 0 and − μ N p ( g ) ≥ 0 -\mu_N^p(g)\geq0 − μ N p ( g ) ≥ 0 . We may therefore choose one positive constant c p c_p c p large enough to absorb the additive constants in (34) and (35) , together with c c h c_{\mathrm{ch}} c ch in (32) . This proves (31) . If g p = p g p=p g p = p , its left side is zero, and the same positive choice of c p c_p c p proves the inequality directly. □
The Cartan inequality
Combining the two length comparisons proves the main estimate of the section.
Proposition 3.5. There is a constant c p > 0 c_p > 0 c p > 0 such that every cone-automorphism g g g normalized by ∣ det g ∣ = 1 |\det g| = 1 ∣ det g ∣ = 1 satisfies
∑ i = 1 N ( μ i p ( g ) ) 2 ≤ ( N − ε ) ( μ 1 p ( g ) + c p ) ( − μ N p ( g ) + c p ) . (36) \sum_{i=1}^{N} (\mu_i^p(g))^2 \le(N-\varepsilon)(\mu_1^p(g)+c_p)(-\mu_N^p(g)+c_p). \tag*{(36)} i = 1 ∑ N ( μ i p ( g ) ) 2 ≤ ( N − ε ) ( μ 1 p ( g ) + c p ) ( − μ N p ( g ) + c p ) . (36)
The constants depend only on the fixed domain, the basepoint, and the Ricci bound, and not on g g g .
Proof. Suppose that g p ≠ p gp \ne p g p = p , and apply Lemma 3.1 to the projective segment from p p p to g p gp g p . Equation (28) bounds the length of its Gauss image by N − ε L ( p , g p ) \sqrt{N-\varepsilon}\,L(p,gp) N − ε L ( p , g p ) . Distance is at most the length of this image, so (27) gives
∑ i ( μ i p ( g ) ) 2 ≤ ( N − ε ) L ( p , g p ) 2 . \sum_i (\mu_i^p(g))^2 \le(N-\varepsilon)L(p,gp)^2. i ∑ ( μ i p ( g ) ) 2 ≤ ( N − ε ) L ( p , g p ) 2 .
Lemma 3.4 proves (36) .
If g p = p gp = p g p = p , equivariance gives G p = g − T G p g − 1 G_p = g^{-T}G_pg^{-1} G p = g − T G p g − 1 , so g g g is G p G_p G p -orthogonal. All its logarithmic singular values vanish, and the result follows. □
We finally express this estimate as a linear gap from all two-level Cartan rays.
Proposition 3.6. Let μ ( γ ) \mu(\gamma) μ ( γ ) be the vector of the logarithms of the singular values of the cone-automorphism γ \gamma γ , computed in any fixed ambient Euclidean norm, arranged in decreasing order. There are constants δ > 0 \delta> 0 δ > 0 and c g a p < ∞ c_{\mathrm{gap}} < \infty c gap < ∞ such that
D k ( μ ( γ ) ) ≥ δ R ( μ ( γ ) ) − c g a p for all γ ∈ Γ , 1 ≤ k < N . (37) D_k(\mu(\gamma)) \ge\delta R(\mu(\gamma)) - c_{\mathrm{gap}} \qquad\text{for all } \gamma\in\Gamma,\ 1 \le k < N. \tag*{(37)} D k ( μ ( γ )) ≥ δ R ( μ ( γ )) − c gap for all γ ∈ Γ , 1 ≤ k < N . (37) One may take δ = ( N − 1 ) ε / N 2 \delta= (N-1)\varepsilon/N^2 δ = ( N − 1 ) ε / N 2 , with the additive constant depending on the chosen norm.
Proof. We first work in the norm G p G_p G p . Write
a = μ 1 p ≥ 0 , b = − μ N p ≥ 0 , R = a + b . a = \mu_1^p \ge0,\qquad b = -\mu_N^p \ge0,\qquad R = a+b. a = μ 1 p ≥ 0 , b = − μ N p ≥ 0 , R = a + b .
If R > 0 R > 0 R > 0 , trace zero gives
a ≤ ( N − 1 ) b , b ≤ ( N − 1 ) a . a \le(N-1)b,\qquad b \le(N-1)a. a ≤ ( N − 1 ) b , b ≤ ( N − 1 ) a .
Hence a / R a/R a / R lies in [ 1 / N , ( N − 1 ) / N ] [1/N,(N-1)/N] [ 1/ N , ( N − 1 ) / N ] , and
a b ≥ η R 2 , η = N − 1 N 2 . (38) ab \ge\eta R^2,\qquad\eta= \frac{N-1}{N^2}. \tag*{(38)} ab ≥ η R 2 , η = N 2 N − 1 . (38)
The following identity uses ∑ i μ i p = 0 \sum_i \mu_i^p = 0 ∑ i μ i p = 0 :
N a b − ∑ i ( μ i p ) 2 = ∑ i ( a − μ i p ) ( μ i p + b ) . Nab-\sum_i(\mu_i^p)^2=\sum_i(a-\mu_i^p)(\mu_i^p+b). N ab − i ∑ ( μ i p ) 2 = i ∑ ( a − μ i p ) ( μ i p + b ) .
Every summand is nonnegative. For i ≤ k i \le k i ≤ k it is at most R ( a − μ i p ) R(a-\mu_i^p) R ( a − μ i p ) , whereas for i > k i > k i > k it is at most R ( μ i p + b ) R(\mu_i^p+b) R ( μ i p + b ) . It follows that
N a b − ∑ i ( μ i p ) 2 ≤ R D k ( μ p ) . (39) Nab-\sum_i(\mu_i^p)^2 \le RD_k(\mu^p). \tag*{(39)} N ab − i ∑ ( μ i p ) 2 ≤ R D k ( μ p ) . (39)
Expanding (36) gives constants k 1 , k 0 ≥ 0 k_1,k_0 \ge0 k 1 , k 0 ≥ 0 , independent of γ \gamma γ and k k k , such that
∑ i ( μ i p ) 2 ≤ ( N − ε ) a b + k 1 R + k 0 . \sum_i(\mu_i^p)^2 \le(N-\varepsilon)ab+k_1R+k_0. i ∑ ( μ i p ) 2 ≤ ( N − ε ) ab + k 1 R + k 0 .
Together with (38) and (39) , this yields
R D k ( μ p ) ≥ ε a b − k 1 R − k 0 ≥ ε η R 2 − k 1 R − k 0 . RD_k(\mu^p) \ge\varepsilon ab-k_1R-k_0 \ge\varepsilon\eta R^2-k_1R-k_0. R D k ( μ p ) ≥ ε ab − k 1 R − k 0 ≥ ε η R 2 − k 1 R − k 0 .
For R ≥ 1 R \ge1 R ≥ 1 , division gives
D k ( μ p ) ≥ ε η R − ( k 1 + k 0 ) . D_k(\mu^p) \ge\varepsilon\eta R-(k_1+k_0). D k ( μ p ) ≥ ε η R − ( k 1 + k 0 ) .
For 0 ≤ R < 1 0 \le R < 1 0 ≤ R < 1 , the nonnegativity of D k D_k D k gives the same bound after enlarging one additive constant. Thus (37) holds in G p G_p G p .
To transfer the conclusion to a fixed ambient norm ∥ ⋅ ∥ 0 \lVert\cdot\rVert_0 ∥ ⋅ ∥ 0 , choose a linear isomorphism S 0 S_0 S 0 such that ∥ v ∥ 0 = ∥ S 0 v ∥ p \lVert v\rVert_0 = \lVert S_0v\rVert_p ∥ v ∥ 0 = ∥ S 0 v ∥ p . The Cartan vector in ∥ ⋅ ∥ 0 \lVert\cdot\rVert_0 ∥ ⋅ ∥ 0 is the Cartan vector of S 0 γ S 0 − 1 S_0\gamma S_0^{-1} S 0 γ S 0 − 1 in ∥ ⋅ ∥ p \lVert\cdot\rVert_p ∥ ⋅ ∥ p . The singular-value inequalities for left and right multiplication give
∣ μ i 0 ( γ ) − μ i p ( γ ) ∣ ≤ e b a s , e b a s = log ( ∥ S 0 ∥ p ∥ S 0 − 1 ∥ p ) . \left|\mu_i^0(\gamma)-\mu_i^p(\gamma)\right|\le e_{\mathrm{bas}},\qquad e_{\mathrm{bas}}=\log\left(\lVert S_0\rVert_p\lVert S_0^{-1}\rVert_p\right). μ i 0 ( γ ) − μ i p ( γ ) ≤ e bas , e bas = log ( ∥ S 0 ∥ p ∥ S 0 − 1 ∥ p ) .
In particular,
∣ R ( μ 0 ) − R ( μ p ) ∣ ≤ 2 e b a s , ∣ D k ( μ 0 ) − D k ( μ p ) ∣ ≤ 2 N e b a s . \left|R(\mu^0)-R(\mu^p)\right|\le2e_{\mathrm{bas}},\qquad\left|D_k(\mu^0)-D_k(\mu^p)\right|\le2Ne_{\mathrm{bas}}. R ( μ 0 ) − R ( μ p ) ≤ 2 e bas , D k ( μ 0 ) − D k ( μ p ) ≤ 2 N e bas .
The gap in G p G_p G p therefore implies
D k ( μ 0 ( γ ) ) ≥ δ R ( μ 0 ( γ ) ) − ( c g a p + 2 N e b a s + 2 δ e b a s ) . D_k(\mu^0(\gamma))\ge\delta R(\mu^0(\gamma))-\left(c_{\mathrm{gap}}+2Ne_{\mathrm{bas}}+2\delta e_{\mathrm{bas}}\right). D k ( μ 0 ( γ )) ≥ δ R ( μ 0 ( γ )) − ( c gap + 2 N e bas + 2 δ e bas ) .
This proves the proposition with the same positive δ \delta δ .
Remark 3.7. The estimate passes directly to the contragredient action. In the dual norm, the ordered Cartan vector of γ − T \gamma^{-T} γ − T is
μ ~ i = − μ N + 1 − i . \widetilde{\mu}_i=-\mu_{N+1-i}. μ i = − μ N + 1 − i .
Consequently
R ( μ ~ ) = R ( μ ) , D k ( μ ~ ) = D N − k ( μ ) . R(\widetilde{\mu})=R(\mu),\qquad D_k(\widetilde{\mu})=D_{N-k}(\mu). R ( μ ) = R ( μ ) , D k ( μ ) = D N − k ( μ ) .
Thus the dual group satisfies (37) with the same constants when the norms are dual, and with an adjusted additive constant in any other fixed norm.
Projective packing in logarithmic coordinates
Let n ≥ 2 n\ge2 n ≥ 2 and N = n + 1 N=n+1 N = n + 1 . We use the Cartan vector μ \mu μ , its spread R R R , and the functions D k D_k D k defined above. Matrix norms in this section are taken in the fixed standard basis. In particular, the singular values of a positive diagonal matrix are its diagonal entries, arranged in decreasing order.
Theorem 4.1. Let K ⊂ R P n K\subset\mathbb{R}\mathrm{P}^n K ⊂ R P n be a properly convex domain and let Γ < P G L ( N , R ) \Gamma<\mathrm{PGL}(N,\mathbb{R}) Γ < PGL ( N , R ) be a discrete group acting cocompactly on K K K . Fix an affine chart R n ⊂ R P n \mathbb{R}^n\subset\mathbb{R}\mathrm{P}^n R n ⊂ R P n with positive orthant R > 0 n \mathbb{R}_{>0}^n R > 0 n . Assume the Cartan gap (37) : there are constants δ > 0 \delta>0 δ > 0 and c g a p ≥ 0 c_{\mathrm{gap}}\ge0 c gap ≥ 0 such that
D k ( μ ( γ ) ) ≥ δ R ( μ ( γ ) ) − c g a p for all γ ∈ Γ , 1 ≤ k < N . D_k(\mu(\gamma))\ge\delta R(\mu(\gamma))-c_{\mathrm{gap}}\qquad\text{for all }\gamma\in\Gamma,\ 1\le k<N. D k ( μ ( γ )) ≥ δ R ( μ ( γ )) − c gap for all γ ∈ Γ , 1 ≤ k < N . Then there is a constant c p a c k c_{\mathrm{pack}} c pack , depending only on these fixed data, such that
sup S ∈ P G L N ( R ) S K ⊂ R > 0 n ∫ S K d y 1 ⋯ d y n y 1 ⋯ y n ≤ c p a c k . (40) \sup_{\substack{S\in\mathrm{PGL}_N(\mathbb{R})\\ SK\subset\mathbb{R}_{>0}^n}}\int_{SK}\frac{dy_1\cdots dy_n}{y_1\cdots y_n}\le c_{\mathrm{pack}}. \tag*{(40)} S ∈ PGL N ( R ) S K ⊂ R > 0 n sup ∫ S K y 1 ⋯ y n d y 1 ⋯ d y n ≤ c pack . (40)
We first make one fixed choice of a bounded affine realization of K K K . This entails no restriction on the theorem. Indeed, after a fixed projective change S b S_b S b , replace K , Γ , S K,\Gamma,S K , Γ , S by S b K , S b Γ S b − 1 , S S b − 1 S_bK,S_b\Gamma S_b^{-1},SS_b^{-1} S b K , S b Γ S b − 1 , S S b − 1 respectively. The target image S K SK S K is unchanged. Lemma 1.2 shows that conjugation by the fixed matrix S b S_b S b changes every Cartan coordinate by a bounded amount. Since R R R and D k D_k D k are Lipschitz functions of these coordinates, the gap retains the same positive δ \delta δ after increasing c g a p c_{\mathrm{gap}} c gap . We therefore work with a fixed bounded source K ⊂ R n K\subset\mathbb{R}^n K ⊂ R n , and rename the constants. All constants below may depend on n n n , this bounded realization of K K K , a fixed compact orbit section F \mathcal{F} F , and the constants in the gap, but never on S S S .
Every image D = S K D=SK D = S K contained in the positive affine orthant is ordinarily convex, including when it is unbounded. For a ∈ R > 0 n a\in\mathbb{R}_{>0}^n a ∈ R > 0 n and 0 < r ≤ 1 / 4 0<r\le1/4 0 < r ≤ 1/4 , write
B ( a , r ) = ∏ i = 1 n [ a i ( 1 − r ) , a i ( 1 + r ) ] \mathcal{B}(a,r)=\prod_{i=1}^{n}[a_i(1-r),a_i(1+r)] B ( a , r ) = i = 1 ∏ n [ a i ( 1 − r ) , a i ( 1 + r )]
for the closed relative coordinate box. The following elementary consequence of John’s ellipsoid theorem ([Joh48]) gives such boxes from logarithmic mass.
Lemma 4.2. There is c ∗ > 0 c_* > 0 c ∗ > 0 , depending only on n n n , with the following property. For m ∈ Z n m \in\mathbb{Z}^n m ∈ Z n , let
t i = 2 − m i , Q m = ∏ i = 1 n ( t i , 2 t i ) , q m = ∫ D ∩ Q m d y 1 ⋯ d y n y 1 ⋯ y n . t_i = 2^{-m_i}, \qquad Q_m = \prod_{i=1}^{n}(t_i,2t_i), \qquad q_m = \int_{D \cap Q_m} \frac{\mathrm{d}y_1 \cdots\mathrm{d}y_n}{y_1 \cdots y_n}. t i = 2 − m i , Q m = i = 1 ∏ n ( t i , 2 t i ) , q m = ∫ D ∩ Q m y 1 ⋯ y n d y 1 ⋯ d y n . where D ⊂ R > 0 n D \subset\mathbb{R}^n_{>0} D ⊂ R > 0 n is any convex open set. Put q max = ( log 2 ) n < 1 q_{\max} = (\log2)^n < 1 q m a x = ( log 2 ) n < 1 . Then q m ≤ q max q_m \leq q_{\max} q m ≤ q m a x , and if q m ≥ ρ > 0 q_m \geq\rho> 0 q m ≥ ρ > 0 , there is a ∈ Q m a \in Q_m a ∈ Q m such that
B ( a , r ) ⊂ D , r = c ∗ ρ ≤ 1 / 4. (41) B(a,r) \subset D, \qquad r = c_*\rho\leq1/4. \tag*{(41)} B ( a , r ) ⊂ D , r = c ∗ ρ ≤ 1/4. (41)
Proof. The upper bound is the integral over the entire dyadic box:
q m ≤ ∏ i = 1 n ∫ t i 2 t i d y i y i = ( log 2 ) n . q_m \leq\prod_{i=1}^{n} \int_{t_i}^{2t_i} \frac{\mathrm{d}y_i}{y_i} = (\log2)^n. q m ≤ i = 1 ∏ n ∫ t i 2 t i y i d y i = ( log 2 ) n .
Assume now q m ≥ ρ > 0 q_m \geq\rho> 0 q m ≥ ρ > 0 . Scale y i = t i Y i y_i = t_iY_i y i = t i Y i , and let W m = diag ( t 1 − 1 , … , t n − 1 ) ( D ∩ Q m ) W_m = \operatorname{diag}(t_1^{-1},\ldots,t_n^{-1})(D \cap Q_m) W m = diag ( t 1 − 1 , … , t n − 1 ) ( D ∩ Q m ) . It is convex and open, is contained in ( 1 , 2 ) n (1,2)^n ( 1 , 2 ) n , and satisfies
ρ ≤ q m = ∫ W m d Y 1 ⋯ d Y n Y 1 ⋯ Y n ≤ Vol E u c ( W m ) . \rho\leq q_m = \int_{W_m} \frac{\mathrm{d}Y_1 \cdots\mathrm{d}Y_n}{Y_1 \cdots Y_n} \leq\operatorname{Vol}_{\mathrm{Euc}}(W_m). ρ ≤ q m = ∫ W m Y 1 ⋯ Y n d Y 1 ⋯ d Y n ≤ Vol Euc ( W m ) .
Since W m W_m W m is an open convex set of positive volume, its closure is a convex body with int W m ‾ = W m \operatorname{int}\overline{W_m} = W_m int W m = W m and Vol E u c ( W m ‾ ) = Vol E u c ( W m ) \operatorname{Vol}_{\mathrm{Euc}}(\overline{W_m}) = \operatorname{Vol}_{\mathrm{Euc}}(W_m) Vol Euc ( W m ) = Vol Euc ( W m ) . The first identity is the standard fact that an open convex set is the interior of its closure; the second follows because the boundary of a full-dimensional convex body has Lebesgue measure zero. Apply John’s ellipsoid theorem [Joh48] to the compact convex body W m ‾ \overline{W_m} W m . Its maximal inscribed ellipsoid E m E_m E m , with semiaxes a 1 ′ , … , a n ′ a'_1,\ldots,a'_n a 1 ′ , … , a n ′ , satisfies
Vol E u c ( E m ) ≥ n − n Vol E u c ( W m ‾ ) ≥ n − n ρ . \operatorname{Vol}_{\mathrm{Euc}}(E_m) \geq n^{-n}\operatorname{Vol}_{\mathrm{Euc}}(\overline{W_m}) \geq n^{-n}\rho. Vol Euc ( E m ) ≥ n − n Vol Euc ( W m ) ≥ n − n ρ .
Every semiaxis is at most n / 2 \sqrt{n}/2 n /2 , since E m ⊂ [ 1 , 2 ] n E_m \subset[1,2]^n E m ⊂ [ 1 , 2 ] n . Set
c n = n − n ω n ( n / 2 ) n − 1 , c ∗ = min { 1 4 , c n 4 n } , c_n = \frac{n^{-n}}{\omega_n(\sqrt{n}/2)^{n-1}}, \qquad c_* = \min\left\{\frac{1}{4},\frac{c_n}{4\sqrt{n}}\right\}, c n = ω n ( n /2 ) n − 1 n − n , c ∗ = min { 4 1 , 4 n c n } ,
with ω n \omega_n ω n being the Euclidean unit-ball volume. Then, we have
min i a i ′ ≥ n − n ρ ω n ( n / 2 ) n − 1 = c n ρ . \min_i a'_i \geq\frac{n^{-n}\rho}{\omega_n(\sqrt{n}/2)^{n-1}} = c_n\rho. i min a i ′ ≥ ω n ( n /2 ) n − 1 n − n ρ = c n ρ .
To obtain a closed set contained in the open cell, shrink E m E_m E m by the factor 1 / 2 1/2 1/2 about its centre z z z . The resulting closed ellipsoid lies in int E m ⊂ int W m ‾ = W m \operatorname{int} E_m \subset\operatorname{int}\overline{W_m} = W_m int E m ⊂ int W m = W m . It contains a closed Euclidean ball of radius c n ρ / 2 c_n\rho/2 c n ρ /2 , and hence a closed coordinate cube about z z z with half-width c n ρ / ( 2 n ) c_n\rho/(2\sqrt{n}) c n ρ / ( 2 n ) . Set a i = t i z i a_i = t_iz_i a i = t i z i , so t i < a i < 2 t i t_i < a_i < 2t_i t i < a i < 2 t i . Scaling the cube back gives coordinate half-widths c n ρ t i / ( 2 n ) c_n\rho t_i/(2\sqrt{n}) c n ρ t i / ( 2 n ) . With c ∗ c_* c ∗ fixed as above, set r = c ∗ ρ r = c_*\rho r = c ∗ ρ . Then r a i ≤ 2 c ∗ ρ t i ≤ c n ρ t i / ( 2 n ) ra_i \leq2c_*\rho t_i \leq c_n\rho t_i/(2\sqrt{n}) r a i ≤ 2 c ∗ ρ t i ≤ c n ρ t i / ( 2 n ) , which proves the box inclusion. Since ρ ≤ q max < 1 \rho\leq q_{\max} < 1 ρ ≤ q m a x < 1 , we also have r ≤ 1 / 4 r \leq1/4 r ≤ 1/4 . □
Lemma 4.3 (Relative-box normalization). Suppose D = S K ⊂ R > 0 n D = SK \subset\mathbb{R}^n_{>0} D = S K ⊂ R > 0 n contains B ( a , r ) B(a,r) B ( a , r ) and B ( b , r ) B(b,r) B ( b , r ) , where 0 < r ≤ 1 / 4 0 < r \leq1/4 0 < r ≤ 1/4 . For 0 ≤ λ ≤ 1 0 \leq\lambda\leq1 0 ≤ λ ≤ 1 put
p λ = ( 1 − λ ) a + λ b , D λ = diag ( ( p 1 λ ) − 1 , … , ( p n λ ) − 1 , 1 ) , p^\lambda= (1-\lambda)a + \lambda b, \qquad D_\lambda= \operatorname{diag}((p_1^\lambda)^{-1},\ldots,(p_n^\lambda)^{-1},1), p λ = ( 1 − λ ) a + λb , D λ = diag (( p 1 λ ) − 1 , … , ( p n λ ) − 1 , 1 ) , and define
J ( Y ) = Y 1 + ∑ i Y i , Φ λ = J D λ S . (42) \mathcal{J}(Y) = \frac{Y}{1+\sum_iY_i}, \qquad\Phi_\lambda= \mathcal{J}D_\lambda S. \tag*{(42)} J ( Y ) = 1 + ∑ i Y i Y , Φ λ = J D λ S . (42) Let
T = { z i > 0 , ∑ i z i < 1 } , b ∗ = 1 n ( 1 , … , 1 ) , β n = 1 8 n ( 1 + n ) . \mathcal{T} = \{z_i > 0,\ \sum_i z_i < 1\}, \qquad b_* = \frac{1}{n}(1,\ldots,1), \qquad\beta_n = \frac{1}{8n(1+\sqrt{n})}. T = { z i > 0 , i ∑ z i < 1 } , b ∗ = n 1 ( 1 , … , 1 ) , β n = 8 n ( 1 + n ) 1 . Then for every λ \lambda λ ,
B ‾ E u c ( b ∗ , β n r ) ⊂ Φ λ K ⊂ T . (43) \overline{B}_{\mathrm{Euc}}(b_*,\beta_nr) \subset\Phi_\lambda K \subset\mathcal{T}. \tag*{(43)} B Euc ( b ∗ , β n r ) ⊂ Φ λ K ⊂ T . (43)
Proof. Minkowski addition of coordinate intervals gives the exact identity
( 1 − λ ) B ( a , r ) + λ B ( b , r ) = B ( p λ , r ) . (1-\lambda)\mathcal{B}(a,r)+\lambda\mathcal{B}(b,r)=\mathcal{B}(p^\lambda,r). ( 1 − λ ) B ( a , r ) + λ B ( b , r ) = B ( p λ , r ) .
Thus convexity of D \mathcal{D} D implies B ( p λ , r ) ⊂ D \mathcal{B}(p^\lambda,r)\subset\mathcal{D} B ( p λ , r ) ⊂ D . After applying D λ \mathcal{D}_\lambda D λ , this is the same box [ 1 − r , 1 + r ] n [1-r,1+r]^n [ 1 − r , 1 + r ] n for every parameter. The entire positive orthant is mapped by J \mathcal{J} J into T \mathcal{T} T , which proves the outer inclusion even if D \mathcal{D} D is unbounded.
For the inner inclusion use the explicit inverse
J − 1 ( z ) = z 1 − ∑ i z i . \mathcal{J}^{-1}(z)=\frac{z}{1-\sum_i z_i}. J − 1 ( z ) = 1 − ∑ i z i z .
Writing 1 = ( 1 , … , 1 ) \mathbf{1}=(1,\ldots,1) 1 = ( 1 , … , 1 ) and z = b ∗ + w z=b_*+w z = b ∗ + w , we obtain
J − 1 ( b ∗ + w ) − 1 = w + 1 ∑ i w i 1 / N − ∑ i w i . \mathcal{J}^{-1}(b_*+w)-\mathbf{1}
=\frac{w+\mathbf{1}\sum_i w_i}{1/N-\sum_i w_i}. J − 1 ( b ∗ + w ) − 1 = 1/ N − ∑ i w i w + 1 ∑ i w i .
If ∥ w ∥ 2 ≤ β n r \lVert w\rVert_2\le\beta_n r ∥ w ∥ 2 ≤ β n r and r ≤ 1 / 4 r\le1/4 r ≤ 1/4 , then w i > − 1 / N w_i>-1/N w i > − 1/ N and 1 − ∑ i ( b ∗ , i + w i ) = 1 / N − ∑ i w i > 0 1-\sum_i(b_{*,i}+w_i)=1/N-\sum_i w_i>0 1 − ∑ i ( b ∗ , i + w i ) = 1/ N − ∑ i w i > 0 . Hence b ∗ + w ∈ T b_*+w\in\mathcal{T} b ∗ + w ∈ T , and J − 1 \mathcal{J}^{-1} J − 1 is defined there. Moreover the denominator in the displayed formula is at least 1 / ( 2 N ) 1/(2N) 1/ ( 2 N ) , while the numerator in the sup norm is at most ( 1 + n ) ∥ w ∥ 2 (1+\sqrt{n})\lVert w\rVert_2 ( 1 + n ) ∥ w ∥ 2 . Therefore
∥ J − 1 ( b ∗ + w ) − 1 ∥ ∞ ≤ 2 N ( 1 + n ) β n r = r / 4 < r . \lVert\mathcal{J}^{-1}(b_*+w)-\mathbf{1}\rVert_\infty
\le2N(1+\sqrt{n})\beta_n r=r/4<r. ∥ J − 1 ( b ∗ + w ) − 1 ∥ ∞ ≤ 2 N ( 1 + n ) β n r = r /4 < r .
The asserted ball is consequently contained in the image of [ 1 − r , 1 + r ] n [1-r,1+r]^n [ 1 − r , 1 + r ] n , and hence in Φ λ K \Phi_\lambda K Φ λ K . □
Lemma 4.4. There is a constant c c o r r ≥ 1 c_{\mathrm{corr}}\ge1 c corr ≥ 1 , depending only on n n n , the bounded realization of K K K , a compact fundamental domain F \mathcal{F} F and R V = diam T R_V=\operatorname{diam}\mathcal{T} R V = diam T such that the following holds. If 0 < r ≤ 1 / 4 0<r\le1/4 0 < r ≤ 1/4 and a projective map Φ ′ \Phi' Φ ′ satisfies
B ‾ E u c ( b ∗ , β n r ) ⊂ Φ ′ K ⊂ T , \overline{B}_{\mathrm{Euc}}(b_*,\beta_n r)\subset\Phi'K\subset\mathcal{T}, B Euc ( b ∗ , β n r ) ⊂ Φ ′ K ⊂ T , one can choose γ ′ ∈ Γ \gamma'\in\Gamma γ ′ ∈ Γ so that the projective map W ′ = Φ ′ ( γ ′ ) − 1 W'=\Phi'(\gamma')^{-1} W ′ = Φ ′ ( γ ′ ) − 1 has condition number
cond ( W ′ ) : = ∥ W ′ ∥ ∥ ( W ′ ) − 1 ∥ ≤ c c o r r / r . (44) \operatorname{cond}(W'):=\lVert W'\rVert\lVert(W')^{-1}\rVert\le c_{\mathrm{corr}}/r. \tag*{(44)} cond ( W ′ ) := ∥ W ′ ∥ ∥( W ′ ) − 1 ∥ ≤ c corr / r . (44) Its absolute-determinant-one representative and its inverse both have operator norm at most c c o r r / r c_{\mathrm{corr}}/r c corr / r . Moreover, if Φ ′ , Φ ′ ′ \Phi',\Phi'' Φ ′ , Φ ′′ satisfy the same hypothesis, then
∥ μ ( Φ ′ ( Φ ′ ′ ) − 1 ) − μ ( γ ′ ( γ ′ ′ ) − 1 ) ∥ ∞ ≤ 2 log ( c c o r r / r ) . (45) \left\lVert\mu\left(\Phi'(\Phi'')^{-1}\right)-\mu\left(\gamma'(\gamma'')^{-1}\right)\right\rVert_\infty
\le2\log(c_{\mathrm{corr}}/r). \tag*{(45)} μ ( Φ ′ ( Φ ′′ ) − 1 ) − μ ( γ ′ ( γ ′′ ) − 1 ) ∞ ≤ 2 log ( c corr / r ) . (45)
Proof. Fix a compact fundamental domain F ⊂ int ( K ) \mathcal{F}\subset\operatorname{int}(K) F ⊂ int ( K ) . Since b ∗ ∈ Φ ′ K b_*\in\Phi'K b ∗ ∈ Φ ′ K , the point ( Φ ′ ) − 1 b ∗ (\Phi')^{-1}b_* ( Φ ′ ) − 1 b ∗ belongs to K K K . Cocompactness therefore supplies γ ′ ∈ Γ \gamma'\in\Gamma γ ′ ∈ Γ such that x = γ ′ ( Φ ′ ) − 1 b ∗ ∈ F x=\gamma'(\Phi')^{-1}b_*\in\mathcal{F} x = γ ′ ( Φ ′ ) − 1 b ∗ ∈ F . Then W ′ W' W ′ maps K K K to Φ ′ K \Phi'K Φ ′ K and sends x x x to b ∗ b_* b ∗ . There are fixed positive constants r U r_U r U , R U R_U R U , R V R_V R V such that the source and target translates
U = K − x , V = Φ ′ K − b ∗ U=K-x,\qquad V=\Phi'K-b_* U = K − x , V = Φ ′ K − b ∗
satisfy
B E u c ( 0 , r U ) ⊂ U ⊂ B E u c ( 0 , R U ) , B_{\mathrm{Euc}}(0,r_U)\subset U\subset B_{\mathrm{Euc}}(0,R_U), B Euc ( 0 , r U ) ⊂ U ⊂ B Euc ( 0 , R U ) ,
B E u c ( 0 , β n r ) ⊂ V ⊂ B E u c ( 0 , R V ) . B_{\mathrm{Euc}}(0,\beta_n r)\subset V\subset B_{\mathrm{Euc}}(0,R_V). B Euc ( 0 , β n r ) ⊂ V ⊂ B Euc ( 0 , R V ) .
We may take R V = diam T R_V=\operatorname{diam}\mathcal{T} R V = diam T after translating by b ∗ b_* b ∗ .
We compute condition numbers using any invertible matrix representative: scalar multiplication does not change them. Translate x x x and b ∗ b_* b ∗ to zero. An invertible matrix representing the resulting projective map has blocks
( L a c T d ) . \begin{pmatrix}L&a\\ c^{T}&d\end{pmatrix}. ( L c T a d ) .
Since zero maps to the finite affine point zero, a = 0 a=0 a = 0 and d ≠ 0 d\ne0 d = 0 . After dividing the whole matrix by d d d and renaming its blocks, we obtain the representative normalized by its bottom-right entry d = 1 d=1 d = 1 :
W ^ = ( L 0 c T 1 ) , u ⟼ L u 1 + c ⋅ u . \widehat{W}=\begin{pmatrix}L&0\\ c^{T}&1\end{pmatrix},
\qquad u\longmapsto\frac{Lu}{1+c\cdot u}. W = ( L c T 0 1 ) , u ⟼ 1 + c ⋅ u Lu .
The denominator is positive on U U U : it equals one at zero, it has no zero in U U U because the image is entirely finite, and U U U is connected. Its positivity on B E u c ( 0 , r U ) B_{\mathrm{Euc}}(0,r_U) B Euc ( 0 , r U ) implies ∥ c ∥ ≤ r U − 1 \lVert c\rVert\le r_U^{-1} ∥ c ∥ ≤ r U − 1 . For a unit vector v v v , evaluate at u = ( r U / 2 ) v u=(r_U/2)v u = ( r U /2 ) v . The denominator belongs to [ 1 / 2 , 3 / 2 ] [1/2,3/2] [ 1/2 , 3/2 ] and the image has norm at most R V R_V R V , so
∥ L v ∥ ≤ 3 R V / r U . \lVert Lv\rVert\leq3R_V/r_U. ∥ Lv ∥ ≤ 3 R V / r U .
It follows, for this particular representative, that
∥ W ^ ∥ ≤ 1 + ( 1 + 3 R V ) / r U ≤ c f w d . (46) \lVert\widehat{W}\rVert\leq1+(1+3R_V)/r_U\leq c_{\mathrm{fwd}}. \tag*{(46)} ∥ W ∥ ≤ 1 + ( 1 + 3 R V ) / r U ≤ c fwd . (46)
The inverse is
W ^ − 1 = ( L − 1 0 − c T L − 1 1 ) . \widehat{W}^{-1}=
\begin{pmatrix}
L^{-1} & 0\\
-c^{T}L^{-1} & 1
\end{pmatrix}. W − 1 = ( L − 1 − c T L − 1 0 1 ) .
Its own denominator is positive on V \mathcal{V} V , it fixes zero, and its image is contained in U \mathcal{U} U . Applying positivity on its source ball of radius β n r \beta_n r β n r gives ∥ c T L − 1 ∥ ≤ ( β n r ) − 1 \lVert c^{T}L^{-1}\rVert\leq(\beta_n r)^{-1} ∥ c T L − 1 ∥ ≤ ( β n r ) − 1 . Evaluation at v β n r / 2 v\beta_n r/2 v β n r /2 , for a unit vector v v v , gives ∥ L − 1 v ∥ ≤ 3 R U / ( β n r ) \lVert L^{-1}v\rVert\leq3R_U/(\beta_n r) ∥ L − 1 v ∥ ≤ 3 R U / ( β n r ) . Consequently
∥ W ^ − 1 ∥ ≤ 1 + ( 1 + 3 R U ) / ( β n r ) ≤ c i n v / r . (47) \lVert\widehat{W}^{-1}\rVert\leq1+(1+3R_U)/(\beta_n r)\leq c_{\mathrm{inv}}/r. \tag*{(47)} ∥ W − 1 ∥ ≤ 1 + ( 1 + 3 R U ) / ( β n r ) ≤ c inv / r . (47)
The two translations and their inverses have uniformly bounded matrix representatives, because x ∈ F x\in\mathcal{F} x ∈ F and b ∗ b_\ast b ∗ is fixed. Undoing these translations gives a representative W W W of W ′ W' W ′ with ∥ W ∥ ≤ c f w d ′ \lVert W\rVert\leq c'_{\mathrm{fwd}} ∥ W ∥ ≤ c fwd ′ and ∥ W − 1 ∥ ≤ c i n v ′ / r \lVert W^{-1}\rVert\leq c'_{\mathrm{inv}}/r ∥ W − 1 ∥ ≤ c inv ′ / r . In particular,
cond ( W ) = ∥ W ∥ ∥ W − 1 ∥ ≤ c f w d ′ c i n v ′ r . \operatorname{cond}(W)=\lVert W\rVert\lVert W^{-1}\rVert\leq\frac{c'_{\mathrm{fwd}}c'_{\mathrm{inv}}}{r}. cond ( W ) = ∥ W ∥ ∥ W − 1 ∥ ≤ r c fwd ′ c inv ′ .
This estimate is projectively well defined, since condition number is unchanged by scalar rescaling. Normalize W W W by W ~ = ∣ det W ∣ − 1 / N W \widetilde{W}=|\det W|^{-1/N}W W = ∣ det W ∣ − 1/ N W . The singular values of W ~ \widetilde{W} W have geometric mean one, so s 1 ( W ~ ) ≥ 1 ≥ s N ( W ~ ) s_1(\widetilde{W})\geq1\geq s_N(\widetilde{W}) s 1 ( W ) ≥ 1 ≥ s N ( W ) . Hence
∥ W ~ ∥ = s 1 ( W ~ ) ≤ cond ( W ) , ∥ W ~ − 1 ∥ = s N ( W ~ ) − 1 ≤ cond ( W ) . \lVert\widetilde{W}\rVert=s_1(\widetilde{W})\leq\operatorname{cond}(W),\qquad
\lVert\widetilde{W}^{-1}\rVert=s_N(\widetilde{W})^{-1}\leq\operatorname{cond}(W). ∥ W ∥ = s 1 ( W ) ≤ cond ( W ) , ∥ W − 1 ∥ = s N ( W ) − 1 ≤ cond ( W ) .
The existence of the constant c c o r r ≥ 1 c_{\mathrm{corr}}\geq1 c corr ≥ 1 as in the statement follows.
Finally choose absolute-determinant-one representatives of Φ ′ \Phi' Φ ′ , Φ ′ ′ \Phi'' Φ ′′ and γ ′ \gamma' γ ′ , γ ′ ′ \gamma'' γ ′′ , and define W ′ = Φ ′ ( γ ′ ) − 1 W'=\Phi'(\gamma')^{-1} W ′ = Φ ′ ( γ ′ ) − 1 and W ′ ′ = Φ ′ ′ ( γ ′ ′ ) − 1 W''=\Phi''(\gamma'')^{-1} W ′′ = Φ ′′ ( γ ′′ ) − 1 as matrices. All these matrices have absolute determinant one, and
Φ ′ ( Φ ′ ′ ) − 1 = W ′ ( γ ′ ( γ ′ ′ ) − 1 ) ( W ′ ′ ) − 1 . \Phi'(\Phi'')^{-1}=W'(\gamma'(\gamma'')^{-1})(W'')^{-1}. Φ ′ ( Φ ′′ ) − 1 = W ′ ( γ ′ ( γ ′′ ) − 1 ) ( W ′′ ) − 1 .
The singular-value inequalities used in Lemma 1.2 state that
s i ( g 0 ) ∥ U − 1 ∥ ∥ V − 1 ∥ ≤ s i ( U g 0 V ) ≤ ∥ U ∥ ∥ V ∥ s i ( g 0 ) . \frac{s_i(g_0)}{\lVert U^{-1}\rVert\lVert V^{-1}\rVert}\leq s_i(Ug_0V)\leq\lVert U\rVert\lVert V\rVert s_i(g_0). ∥ U − 1 ∥ ∥ V − 1 ∥ s i ( g 0 ) ≤ s i ( U g 0 V ) ≤ ∥ U ∥ ∥ V ∥ s i ( g 0 ) .
Use U = W ′ U=W' U = W ′ , V = ( W ′ ′ ) − 1 V=(W'')^{-1} V = ( W ′′ ) − 1 and the normalized matrix bounds. This gives (45) . □
Lemma 4.5. There is a constant c s e p c_{\mathrm{sep}} c sep such that, whenever a projective image D = S K ⊂ R > 0 n D=SK\subset\mathbb{R}_{>0}^{n} D = S K ⊂ R > 0 n contains B ( a , r ) \mathcal{B}(a,r) B ( a , r ) and B ( b , r ) \mathcal{B}(b,r) B ( b , r ) with 0 < r ≤ 1 / 4 0<r\leq1/4 0 < r ≤ 1/4 , one has
max i ∣ log ( b i / a i ) ∣ ≤ c s e p ( 1 + log ( 1 / r ) ) . (48) \max_i\left|\log(b_i/a_i)\right|\leq c_{\mathrm{sep}}(1+\log(1/r)). \tag*{(48)} i max ∣ log ( b i / a i ) ∣ ≤ c sep ( 1 + log ( 1/ r )) . (48)
Proof. Use the maps Φ λ \Phi_\lambda Φ λ from Lemma 4.3. For each parameter, apply Lemma 4.4 with the same value of r r r to Φ λ \Phi_\lambda Φ λ , obtaining an element γ λ \gamma_\lambda γ λ . The relative Cartan comparison (45) then holds uniformly for any pair of parameters. Put ξ i = log ( b i / a i ) \xi_i=\log(b_i/a_i) ξ i = log ( b i / a i ) and set λ = e − s \lambda=e^{-s} λ = e − s for s ≥ log 2 s\geq\log2 s ≥ log 2 . Write Φ s = Φ e − s \Phi_s=\Phi_{e^{-s}} Φ s = Φ e − s and γ s = γ e − s \gamma_s=\gamma_{e^{-s}} γ s = γ e − s . Then
p i ( s ) = a i ( 1 − e − s + e ξ i − s ) . p_i(s)=a_i(1-e^{-s}+e^{\xi_i-s}). p i ( s ) = a i ( 1 − e − s + e ξ i − s ) .
Since 1 − e − s ∈ [ 1 / 2 , 1 ) 1-e^{-s}\in[1/2,1) 1 − e − s ∈ [ 1/2 , 1 ) , the elementary bounds for the sum of the two positive terms give
∣ log p i ( s ) − log a i − ( ξ i − s ) + ∣ ≤ log 2. (49) \left|\log p_i(s)-\log a_i-(\xi_i-s)_+\right|\leq\log2. \tag*{(49)} ∣ log p i ( s ) − log a i − ( ξ i − s ) + ∣ ≤ log 2. (49)
We first control the ratios b i / a i b_i/a_i b i / a i that are large, namely ξ i = log ( b i / a i ) > log 2 \xi_i=\log(b_i/a_i)>\log2 ξ i = log ( b i / a i ) > log 2 ; the opposite ratios are handled at the end by interchanging a a a and b b b . Suppose that ξ max = max i ξ i > log 2 \xi_{\max}=\max_i\xi_i>\log2 ξ m a x = max i ξ i > log 2 . Subdivide the interval [ log 2 , ξ max ] [\log2,\xi_{\max}] [ log 2 , ξ m a x ] at ξ i \xi_i ξ i . There are at most n n n positive-length pieces. On a piece [ s 1 , s 2 ] [s_1,s_2] [ s 1 , s 2 ] let I = { i : ξ i > s } I=\{i:\xi_i>s\} I = { i : ξ i > s } for an interior point s s s , and put τ = s 2 − s 1 \tau=s_2-s_1 τ = s 2 − s 1 . The set I I I is constant and nonempty on the interior. On this piece every coordinate satisfies the dichotomy ξ i ≤ s 1 \xi_i \le s_1 ξ i ≤ s 1 or ξ i ≥ s 2 \xi_i \ge s_2 ξ i ≥ s 2 ; the second alternative holds exactly when i ∈ I i \in I i ∈ I . Consequently,
( ξ i − s 1 ) + − ( ξ i − s 2 ) + = τ 1 I ( i ) (\xi_i-s_1)_+-(\xi_i-s_2)_+=\tau1_I(i) ( ξ i − s 1 ) + − ( ξ i − s 2 ) + = τ 1 I ( i )
holds also at the endpoints. Applying (49) there gives
ℓ i : = log p i ( s 1 ) p i ( s 2 ) = τ 1 I ( i ) + e i , ∣ e i ∣ ≤ 2 log 2. (50) \ell_i:=\log\frac{p_i(s_1)}{p_i(s_2)}=\tau1_I(i)+e_i,\qquad|e_i|\le2\log2. \tag*{(50)} ℓ i := log p i ( s 2 ) p i ( s 1 ) = τ 1 I ( i ) + e i , ∣ e i ∣ ≤ 2 log 2. (50)
Write j = ∣ I ∣ j=|I| j = ∣ I ∣ , so 1 ≤ j ≤ n < N 1\le j\le n<N 1 ≤ j ≤ n < N . By definition of Φ s 2 \Phi_{s_2} Φ s 2 , we have
Φ s 2 Φ s 1 − 1 = J diag ( p 1 ( s 1 ) p 1 ( s 2 ) , … , p n ( s 1 ) p n ( s 2 ) , 1 ) J − 1 . (51) \Phi_{s_2}\Phi_{s_1}^{-1}=J\operatorname{diag}\left(\frac{p_1(s_1)}{p_1(s_2)},\ldots,\frac{p_n(s_1)}{p_n(s_2)},1\right)J^{-1}. \tag*{(51)} Φ s 2 Φ s 1 − 1 = J diag ( p 1 ( s 2 ) p 1 ( s 1 ) , … , p n ( s 2 ) p n ( s 1 ) , 1 ) J − 1 . (51)
with
J = ( I n 0 1 n T 1 ) . J=\begin{pmatrix} I_n & 0 \\ 1_n^{T} & 1 \end{pmatrix}. J = ( I n 1 n T 0 1 ) .
Append ℓ N = e N = 0 \ell_N=e_N=0 ℓ N = e N = 0 in (50) . Divide the diagonal matrix by the N N N -th root of its determinant and let
Z j = ( ( N − j ) / N , … , ( N − j ) / N , − j / N , … , − j / N ) . Z_j=((N-j)/N,\ldots,(N-j)/N,-j/N,\ldots,-j/N). Z j = (( N − j ) / N , … , ( N − j ) / N , − j / N , … , − j / N ) .
Then the Cartan vector of the normalized diagonal matrix is within 4 log 2 4\log2 4 log 2 of τ Z j \tau Z_j τ Z j . To bound the Cartan vector of the conjugated matrix, we let F = diag ( 1 , … , 1 , − 1 ) F=\operatorname{diag}(1,\ldots,1,-1) F = diag ( 1 , … , 1 , − 1 ) . The matrix J J J has determinant one and satisfies J − 1 = F J F J^{-1}=FJF J − 1 = F J F . Since F F F is orthogonal, ∥ J − 1 ∥ = ∥ J ∥ \lVert J^{-1}\rVert=\lVert J\rVert ∥ J − 1 ∥ = ∥ J ∥ . For any absolute-determinant-one matrix g 0 g_0 g 0 , the min–max singular-value inequalities, applied to U = J U=J U = J and V = J − 1 V=J^{-1} V = J − 1 , give
∣ μ i ( J g 0 J − 1 ) − μ i ( g 0 ) ∣ ≤ log ( ∥ J ∥ ⋅ ∥ J − 1 ∥ ) = log cond J . \left|\mu_i(Jg_0J^{-1})-\mu_i(g_0)\right|\le\log(\lVert J\rVert\cdot\lVert J^{-1}\rVert)=\log\operatorname{cond}J. μ i ( J g 0 J − 1 ) − μ i ( g 0 ) ≤ log (∥ J ∥ ⋅ ∥ J − 1 ∥) = log cond J .
Using (45) , we conclude that
∥ μ ( γ s 2 γ s 1 − 1 ) − τ Z j ∥ ∞ ≤ E ( r ) : = 4 log 2 + log cond J + 2 log ( c c o r r / r ) . (52) \left\|\mu(\gamma_{s_2}\gamma_{s_1}^{-1})-\tau Z_j\right\|_\infty\le E(r):=4\log2+\log\operatorname{cond}J+2\log(c_{\mathrm{corr}}/r). \tag*{(52)} μ ( γ s 2 γ s 1 − 1 ) − τ Z j ∞ ≤ E ( r ) := 4 log 2 + log cond J + 2 log ( c corr / r ) . (52)
The vector Z j Z_j Z j satisfies R ( Z j ) = 1 R(Z_j)=1 R ( Z j ) = 1 and D j ( Z j ) = 0 D_j(Z_j)=0 D j ( Z j ) = 0 . If an ordered vector v v v is within E ( r ) E(r) E ( r ) of τ Z j \tau Z_j τ Z j , then R ( v ) ≥ τ − 2 E ( r ) R(v)\ge\tau-2E(r) R ( v ) ≥ τ − 2 E ( r ) . Every summand in D j ( v ) D_j(v) D j ( v ) is at most 2 E ( r ) 2E(r) 2 E ( r ) , since the corresponding summand for τ Z j \tau Z_j τ Z j is zero; hence D j ( v ) ≤ 2 N E ( r ) D_j(v)\le2NE(r) D j ( v ) ≤ 2 N E ( r ) . Apply the Cartan gap (37) to the group element γ s 2 γ s 1 − 1 \gamma_{s_2}\gamma_{s_1}^{-1} γ s 2 γ s 1 − 1 . It gives
δ ( τ − 2 E ( r ) ) − c g a p ≤ 2 N E ( r ) , \delta(\tau-2E(r))-c_{\mathrm{gap}}\le2NE(r), δ ( τ − 2 E ( r )) − c gap ≤ 2 N E ( r ) ,
and thus
τ ≤ ( 2 + 2 N δ ) E ( r ) + c g a p δ = : L p i e c e ( r ) . (53) \tau\le\left(2+\frac{2N}{\delta}\right)E(r)+\frac{c_{\mathrm{gap}}}{\delta}=:L_{\mathrm{piece}}(r). \tag*{(53)} τ ≤ ( 2 + δ 2 N ) E ( r ) + δ c gap =: L piece ( r ) . (53)
This is uniform over all active sets and all pieces. Summing at most n n n pieces yields
max i ξ i ≤ log 2 + n L p i e c e ( r ) . \max_i \xi_i\le\log2+nL_{\mathrm{piece}}(r). i max ξ i ≤ log 2 + n L piece ( r ) .
The same inequality holds directly when ξ max ≤ log 2 \xi_{\max}\le\log2 ξ m a x ≤ log 2 . Interchanging a a a and b b b and repeating the construction bounds max i ( − ξ i ) \max_i(-\xi_i) max i ( − ξ i ) by the same expression. Consequently
max i ∣ ξ i ∣ ≤ log 2 + n L p i e c e ( r ) ≤ c s e p ( 1 + log ( 1 / r ) ) , \max_i|\xi_i|\le\log2+nL_{\mathrm{piece}}(r)\le c_{\mathrm{sep}}(1+\log(1/r)), i max ∣ ξ i ∣ ≤ log 2 + n L piece ( r ) ≤ c sep ( 1 + log ( 1/ r )) ,
as required. Breakpoints with equal values create no positive-length piece, and the endpoint identities used above remain valid there. □ \square □
Lemma 4.6. Let ( q m ) m ∈ Z n (q_m)_{m\in\mathbb{Z}^n} ( q m ) m ∈ Z n be nonnegative numbers with q m ≤ q max = ( log 2 ) n q_m\le q_{\max}=(\log2)^n q m ≤ q m a x = ( log 2 ) n . If
# { m : q m ≥ ρ } ≤ c c n t ( 1 + ∣ log ρ ∣ ) n (with 0 < ρ ≤ q max ), \#\{m:q_m\ge\rho\}\le c_{\mathrm{cnt}}(1+|\log\rho|)^n\qquad\text{(with }0<\rho\le q_{\max}\text{),} # { m : q m ≥ ρ } ≤ c cnt ( 1 + ∣ log ρ ∣ ) n (with 0 < ρ ≤ q m a x ), then
∑ m ∈ Z n q m ≤ c c n t ∫ 0 q max ( 1 + ∣ log ρ ∣ ) n d ρ < ∞ . \sum_{m\in\mathbb{Z}^n}q_m\le c_{\mathrm{cnt}}\int_0^{q_{\max}}(1+|\log\rho|)^n\,d\rho<\infty. m ∈ Z n ∑ q m ≤ c cnt ∫ 0 q m a x ( 1 + ∣ log ρ ∣ ) n d ρ < ∞.
Proof. Tonelli’s theorem for nonnegative functions gives
∑ m ∈ Z n q m = ∑ m ∫ 0 q max 1 { ρ < q m } d ρ = ∫ 0 q max # { m : q m > ρ } d ρ ≤ c c n t ∫ 0 q max ( 1 + ∣ log ρ ∣ ) n d ρ . \sum_{m\in\mathbb{Z}^{n}}q_m=\sum_m\int_0^{q_{\max}}\mathbf{1}_{\{\rho<q_m\}}\,d\rho=\int_0^{q_{\max}}\#\{m:q_m>\rho\}\,d\rho\le c_{\mathrm{cnt}}\int_0^{q_{\max}}(1+|\log\rho|)^n\,d\rho. m ∈ Z n ∑ q m = m ∑ ∫ 0 q m a x 1 { ρ < q m } d ρ = ∫ 0 q m a x # { m : q m > ρ } d ρ ≤ c cnt ∫ 0 q m a x ( 1 + ∣ log ρ ∣ ) n d ρ .
by assumption. Since q max < 1 q_{\max}<1 q m a x < 1 , we conclude
∫ 0 q max ( 1 + ∣ log ρ ∣ ) n d ρ ≤ ∫ 0 1 ( 1 − log ρ ) n d ρ = ∫ 0 ∞ ( 1 + t ) n e − t d t < ∞ . \int_0^{q_{\max}}(1+|\log\rho|)^n\,d\rho\le\int_0^1(1-\log\rho)^n\,d\rho=\int_0^\infty(1+t)^n e^{-t}\,dt<\infty. ∫ 0 q m a x ( 1 + ∣ log ρ ∣ ) n d ρ ≤ ∫ 0 1 ( 1 − log ρ ) n d ρ = ∫ 0 ∞ ( 1 + t ) n e − t d t < ∞.
Proof of Theorem 4.1 . Fix S S S and put D = S K ⊂ R > 0 n \mathcal{D}=SK\subset\mathbb{R}^n_{>0} D = S K ⊂ R > 0 n . Use the dyadic boxes Q m Q_m Q m ; they partition D \mathcal{D} D up to subsets of measure zero. Fix 0 < ρ ≤ q max 0<\rho\le q_{\max} 0 < ρ ≤ q m a x . For every index with q m ≥ ρ q_m\ge\rho q m ≥ ρ , choose a centre a ( m ) ∈ Q m a(m)\in Q_m a ( m ) ∈ Q m such that B ( a ( m ) , c ∗ ρ ) ⊂ D B(a(m),c_*\rho)\subset\mathcal{D} B ( a ( m ) , c ∗ ρ ) ⊂ D . If there are two such indices m , m ′ m,m' m , m ′ , then Lemma 4.5 gives
max i ∣ log a i ( m ) − log a i ( m ′ ) ∣ ≤ c s e p ( 1 + log ( 1 / ( c ∗ ρ ) ) ) . \max_i|\log a_i(m)-\log a_i(m')|\le c_{\mathrm{sep}}(1+\log(1/(c_*\rho))). i max ∣ log a i ( m ) − log a i ( m ′ ) ∣ ≤ c sep ( 1 + log ( 1/ ( c ∗ ρ ))) .
Write a i ( m ) = 2 − m i z i ( m ) a_i(m)=2^{-m_i}z_i(m) a i ( m ) = 2 − m i z i ( m ) , where 1 < z i ( m ) < 2 1<z_i(m)<2 1 < z i ( m ) < 2 . It follows that
∣ m i − m i ′ ∣ log 2 ≤ ∣ log a i ( m ) − log a i ( m ′ ) ∣ + ∣ log z i ( m ) − log z i ( m ′ ) ∣ ≤ c s e p ( 1 + log ( 1 / ( c ∗ ρ ) ) ) + log 2. \begin{aligned}
|m_i-m_i'|\log2\le|\log a_i(m)-\log a_i(m')|+|\log z_i(m)-\log z_i(m')| \\
&\le c_{\mathrm{sep}}(1+\log(1/(c_*\rho)))+\log2.
\end{aligned} ∣ m i − m i ′ ∣ log 2 ≤ ∣ log a i ( m ) − log a i ( m ′ ) ∣ + ∣ log z i ( m ) − log z i ( m ′ ) ∣ ≤ c sep ( 1 + log ( 1/ ( c ∗ ρ ))) + log 2.
Because c ∗ c_* c ∗ is fixed and 0 < ρ ≤ q max < 1 0<\rho\le q_{\max}<1 0 < ρ ≤ q m a x < 1 , all high-mass indices have pairwise lattice distance at most c l a t ( 1 + ∣ log ρ ∣ ) c_{\mathrm{lat}}(1+|\log\rho|) c lat ( 1 + ∣ log ρ ∣ ) in the sup norm. If the set is nonempty, fix one of its indices m 0 m^0 m 0 ; all the others lie in an integer-coordinate cube about m 0 m^0 m 0 of this radius. Consequently
# { m : q m ≥ ρ } ≤ ( 2 ⌊ c l a t ( 1 + ∣ log ρ ∣ ) ⌋ + 1 ) n ≤ c c n t ( 1 + ∣ log ρ ∣ ) n . (54) \#\{m:q_m\ge\rho\}\le(2\lfloor c_{\mathrm{lat}}(1+|\log\rho|)\rfloor+1)^n\le c_{\mathrm{cnt}}(1+|\log\rho|)^n. \tag*{(54)} # { m : q m ≥ ρ } ≤ ( 2 ⌊ c lat ( 1 + ∣ log ρ ∣ )⌋ + 1 ) n ≤ c cnt ( 1 + ∣ log ρ ∣ ) n . (54)
If the set of high-mass indices is empty, the same bound is clearly satisfied. Finally, countable additivity on the dyadic cells and Lemma 4.6 give
∫ D d y 1 ⋯ d y n y 1 ⋯ y n = ∑ m ∈ Z n q m ≤ c c n t ∫ 0 q max ( 1 + ∣ log ρ ∣ ) n d ρ < ∞ . \int_{\mathcal{D}}\frac{dy_1\cdots dy_n}{y_1\cdots y_n}=\sum_{m\in\mathbb{Z}^n}q_m\le c_{\mathrm{cnt}}\int_0^{q_{\max}}(1+|\log\rho|)^n\,d\rho<\infty. ∫ D y 1 ⋯ y n d y 1 ⋯ d y n = m ∈ Z n ∑ q m ≤ c cnt ∫ 0 q m a x ( 1 + ∣ log ρ ∣ ) n d ρ < ∞.
All constants are independent of S S S , which proves the asserted supremum.
Uniform volumes of projective simplices
We now convert the estimate in the previous section into a bound for projective simplices. Recall that ω n \omega_n ω n denotes the Euclidean volume of the unit ball in R n \mathbb{R}^n R n . The Holmes–Thompson normalization is the one in (10) : the density is the cotangent unit-ball volume divided by ω n \omega_n ω n . We write Vol F , L ( E ) \operatorname{Vol}_{F,L}(E) Vol F , L ( E ) for the volume of E E E computed with the Funk metric of the enclosing convex domain int L \operatorname{int}L int L .
First, Theorem 4.1 applies to the dual domain. Indeed, the canonical conormal map induces a Γ \Gamma Γ -equivariant diffeomorphism Ω ⟶ Ω ∗ \Omega\longrightarrow\Omega^* Ω ⟶ Ω ∗ , where the action on Ω ∗ \Omega^* Ω ∗ is contragredient; see [19 Theorem 3 and Proposition 8] . Consequently this action is cocompact. In the dual ambient norm, its Cartan vector is
μ ~ i ( γ ) = − μ N + 1 − i ( γ ) , 1 ≤ i ≤ N . \widetilde{\mu}_i(\gamma)=-\mu_{N+1-i}(\gamma),\qquad1\le i\le N. μ i ( γ ) = − μ N + 1 − i ( γ ) , 1 ≤ i ≤ N .
For the spread and the defects used in the Cartan gap this gives
R ( μ ~ ) = R ( μ ) , D k ( μ ~ ) = D N − k ( μ ) . R(\widetilde{\mu})=R(\mu),\qquad D_k(\widetilde{\mu})=D_{N-k}(\mu). R ( μ ) = R ( μ ) , D k ( μ ) = D N − k ( μ ) .
Thus the same gap holds for the contragredient action. We may therefore apply Theorem 4.1 to every projective image of Ω ∗ \Omega^* Ω ∗ contained in a positive orthant.
We will use the following precise form of Funk volume duality. For a properly convex body K ⊂ R P n K\subset\mathbb{RP}^n K ⊂ RP n , let K ∨ K^\vee K ∨ be its projective dual body, the closure of the set of hyperplanes disjoint from K K K .
Proposition 5.1. [8 Corollary 3.11] Let K , L ⊂ R P n K,L \subset\mathbb{R}P^n K , L ⊂ R P n be properly convex bodies with nonempty interiors, and suppose K ⊂ int L K \subset\operatorname{int} L K ⊂ int L . Then
Vol F , L ( K ) = Vol F , K ∨ ( L ∨ ) . (55) \operatorname{Vol}_{F,L}(K)=\operatorname{Vol}_{F,K^\vee}(L^\vee). \tag*{(55)} Vol F , L ( K ) = Vol F , K ∨ ( L ∨ ) . (55) These volumes are independent of the affine charts used to define the Funk metrics.
The density on the enclosing dual simplex can be computed directly.
Lemma 5.2. Let
S = { ( a 0 , … , a n ) : a i > 0 , ∑ i = 0 n a i = 1 } . S=\{(a_0,\ldots,a_n):a_i>0,\ \sum_{i=0}^{n}a_i=1\}. S = {( a 0 , … , a n ) : a i > 0 , i = 0 ∑ n a i = 1 } . In the coordinates a 1 , … , a n a_1,\ldots,a_n a 1 , … , a n , with a 0 = 1 − ∑ i = 1 n a i a_0=1-\sum_{i=1}^{n}a_i a 0 = 1 − ∑ i = 1 n a i , the Funk–Holmes–Thompson density of S S S is
d Vol F , S = 1 n ! ω n d a 1 ⋯ d a n a 0 a 1 ⋯ a n . (56) d\operatorname{Vol}_{F,S}=\frac{1}{n!\omega_n}\frac{da_1\cdots da_n}{a_0a_1\cdots a_n}. \tag*{(56)} d Vol F , S = n ! ω n 1 a 0 a 1 ⋯ a n d a 1 ⋯ d a n . (56)
Proof. At a ∈ S a\in S a ∈ S , the tangent unit ball is S − a S-a S − a , using Minkowski-sum notation. Its polar, the cotangent unit ball, is the simplex with vertices
ϖ i = − d a i a i ( 1 ≤ i ≤ n ) , ϖ 0 = d a 1 + ⋯ + d a n a 0 . \varpi_i=-\frac{da_i}{a_i}\quad(1\leq i\leq n),\qquad\varpi_0=\frac{da_1+\cdots+da_n}{a_0}. ϖ i = − a i d a i ( 1 ≤ i ≤ n ) , ϖ 0 = a 0 d a 1 + ⋯ + d a n .
In the cotangent basis d a 1 , … , d a n da_1,\ldots,da_n d a 1 , … , d a n , the volume of this simplex is ∣ det W a ∣ / n ! |\det W_a|/n! ∣ det W a ∣/ n ! , where − W a -W_a − W a is the matrix whose columns are ϖ i − ϖ 0 \varpi_i-\varpi_0 ϖ i − ϖ 0 . Hence,
W a = diag ( a 1 − 1 , … , a n − 1 ) + a 0 − 1 11 T W_a=\operatorname{diag}(a_1^{-1},\ldots,a_n^{-1})+a_0^{-1}\mathbf{1}\mathbf{1}^{T} W a = diag ( a 1 − 1 , … , a n − 1 ) + a 0 − 1 1 1 T
and
∣ det W a ∣ = 1 a 1 ⋯ a n ( 1 + a 1 + ⋯ + a n a 0 ) = 1 a 0 ⋯ a n . |\det W_a|=\frac{1}{a_1\cdots a_n}\left(1+\frac{a_1+\cdots+a_n}{a_0}\right)=\frac{1}{a_0\cdots a_n}. ∣ det W a ∣ = a 1 ⋯ a n 1 ( 1 + a 0 a 1 + ⋯ + a n ) = a 0 ⋯ a n 1 .
Dividing by ω n \omega_n ω n proves the formula.
Proof of Theorem A. Let σ ⊂ Ω \sigma\subset\Omega σ ⊂ Ω be a nondegenerate projective n n n -simplex with interior vertices. Choose positive independent cone lifts v 0 , … , v n ∈ C v_0,\ldots,v_n\in C v 0 , … , v n ∈ C and put v σ = ∑ i v i v_\sigma=\sum_i v_i v σ = ∑ i v i . On the dual affine section ℓ ( v σ ) = 1 \ell(v_\sigma)=1 ℓ ( v σ ) = 1 , set
a i = ℓ ( v i ) , 0 ≤ i ≤ n . a_i=\ell(v_i),\qquad0\leq i\leq n. a i = ℓ ( v i ) , 0 ≤ i ≤ n .
These coordinates identify the dual domain with a convex domain D σ ⊂ S D_\sigma\subset S D σ ⊂ S . The enclosing simplex S S S is the interior of σ ∨ \sigma^\vee σ ∨ . Since σ \sigma σ is compactly contained in Ω \Omega Ω , the dual body D σ ‾ \overline{D_\sigma} D σ is compactly contained in S S S . Proposition 5.1 and Lemma 5.2 give the exact identity
Vol F , Ω ( σ ) = Vol F , S ( D σ ) = 1 n ! ω n ∫ D σ d a 1 ⋯ d a n a 0 a 1 ⋯ a n . (57) \operatorname{Vol}_{F,\Omega}(\sigma)=\operatorname{Vol}_{F,S}(D_\sigma)=\frac{1}{n!\omega_n}\int_{D_\sigma}\frac{da_1\cdots da_n}{a_0a_1\cdots a_n}. \tag*{(57)} Vol F , Ω ( σ ) = Vol F , S ( D σ ) = n ! ω n 1 ∫ D σ a 0 a 1 ⋯ a n d a 1 ⋯ d a n . (57)
Partition D σ D_\sigma D σ into Borel sets E j E_j E j , 0 ≤ j ≤ n 0\leq j\leq n 0 ≤ j ≤ n , on which a j a_j a j is a largest coordinate, resolving ties by the smallest index. On E j E_j E j we have a j ≥ 1 / N a_j\geq1/N a j ≥ 1/ N , so
∫ E j d a 1 ⋯ d a n ∏ i = 0 n a i ≤ N ∫ D σ ∏ i ≠ j d a i ∏ i ≠ j a i . \int_{E_j}\frac{da_1\cdots da_n}{\prod_{i=0}^{n}a_i}\leq N\int_{D_\sigma}\frac{\prod_{i\ne j}da_i}{\prod_{i\ne j}a_i}. ∫ E j ∏ i = 0 n a i d a 1 ⋯ d a n ≤ N ∫ D σ ∏ i = j a i ∏ i = j d a i .
Theorem 4.1 bounds the integrals by one constant c d u a l c_{\mathrm{dual}} c dual , independent of σ \sigma σ and j j j . Summing yields
∫ D σ d a 1 ⋯ d a n ∏ i = 0 n a i ≤ N ∑ j = 0 n ∫ D σ ∏ i ≠ j d a i ∏ i ≠ j a i ≤ N 2 c d u a l . \int_{D_\sigma}\frac{da_1\cdots da_n}{\prod_{i=0}^{n}a_i}\leq N\sum_{j=0}^{n}\int_{D_\sigma}\frac{\prod_{i\ne j}da_i}{\prod_{i\ne j}a_i}\leq N^2c_{\mathrm{dual}}. ∫ D σ ∏ i = 0 n a i d a 1 ⋯ d a n ≤ N j = 0 ∑ n ∫ D σ ∏ i = j a i ∏ i = j d a i ≤ N 2 c dual .
The Funk volume bound is therefore
Vol F , Ω ( σ ) ≤ N 2 c d u a l n ! ω n . \operatorname{Vol}_{F,\Omega}(\sigma)\leq\frac{N^2c_{\mathrm{dual}}}{n!\omega_n}. Vol F , Ω ( σ ) ≤ n ! ω n N 2 c dual .
To pass to the Hilbert metric, we apply the Rogers–Shephard inequality [22 Theorem 1]
d Vol H T , Ω ≤ 2 − n ( 2 n n ) d Vol F , Ω . d\operatorname{Vol}_{HT,\Omega}\leq2^{-n}\binom{2n}{n}d\operatorname{Vol}_{F,\Omega}. d Vol H T , Ω ≤ 2 − n ( n 2 n ) d Vol F , Ω .
Hence
Vol H T , Ω ( σ ) ≤ 2 − n ( 2 n n ) N 2 c d u a l n ! ω n . \operatorname{Vol}_{\mathrm{HT},\Omega}(\sigma) \le2^{-n}\binom{2n}{n}\frac{N^{2}c_{\mathrm{dual}}}{n!\omega_n}. Vol HT , Ω ( σ ) ≤ 2 − n ( n 2 n ) n ! ω n N 2 c dual .
The bound is uniform over ordinary projective simplices, as required. A degenerate n n n -simplex has dimension less than n n n and hence zero n n n -dimensional volume, so the same bound holds in that case.
Straightening and simplicial volume
The uniform bound established above is an estimate on simplices in the convex domain Ω \Omega Ω . A projected simplex is only a singular simplex and may fail to be embedded; however, neither embeddedness nor non-overlap is needed. We use the canonical affine sphere to define an equivariant straightening homotopic to the identity, from which we will conclude the proof of Corollaries B and C. The final argument is similar to the proof of the lower bound of the simplicial volume for closed hyperbolic manifolds in [3 ] . We, therefore, only explain the main ideas and refer the interested reader to Benedetti–Petronio’s book for details.
First suppose that M = Γ \ Ω M=\Gamma\backslash\Omega M = Γ\Ω is oriented, and let ω h \omega_h ω h be its positive smooth Blaschke volume form. The ratio of the Blaschke density and the Holmes–Thompson density is continuous and Γ \Gamma Γ -invariant. It descends to the compact quotient, so there is b Ω < ∞ b_\Omega<\infty b Ω < ∞ such that
d Vol h ≤ b Ω d Vol H T , Ω on Ω . d\operatorname{Vol}_h \le b_\Omega\,d\operatorname{Vol}_{\mathrm{HT},\Omega}\quad\text{on }\Omega. d Vol h ≤ b Ω d Vol HT , Ω on Ω.
By Theorem A, there is consequently V h < ∞ V_h<\infty V h < ∞ bounding the Blaschke volumes of all ordinary projective n n n -simplices in Ω \Omega Ω with interior vertices.
Proposition 6.1. For every integer 0 ≤ k ≤ n + 1 0\le k\le n+1 0 ≤ k ≤ n + 1 and points ( x i ) i = 0 , … , k ∈ Ω (x_i)_{i=0,\ldots,k}\in\Omega ( x i ) i = 0 , … , k ∈ Ω , the formula
st ( x 0 , … , x k ) ( t 0 , … , t k ) = [ ∑ i = 0 k t i f ( x i ) ] (58) \operatorname{st}(x_0,\ldots,x_k)(t_0,\ldots,t_k)=\left[\sum_{i=0}^{k}t_i f(x_i)\right] \tag*{(58)} st ( x 0 , … , x k ) ( t 0 , … , t k ) = [ i = 0 ∑ k t i f ( x i ) ] (58) defines a smooth simplex in Ω \Omega Ω . It is Γ \Gamma Γ -equivariant and compatible with faces. It induces a chain map on the continuous singular chains of M M M , chain-homotopic to the identity. If the input singular simplex is smooth, all simplices in the homotopy may be taken smooth. For k = n k=n k = n and projectively independent vertices, the straightened simplex is the ordinary projective convex hull of its vertices. For projectively dependent vertices its n n n -volume vanishes.
Proof. The sum in (58) belongs to the open convex cone C C C . Equivariance follows from the equivariance of f f f , and setting one coordinate t i t_i t i equal to zero gives the formula for the corresponding face.
Let π : Ω → M \pi:\Omega\to M π : Ω → M be the covering map. For a continuous singular simplex σ : Δ k → M \sigma:\Delta^k\to M σ : Δ k → M , choose a lift σ ~ \widetilde{\sigma} σ , set x i = σ ~ ( e i ) x_i=\widetilde{\sigma}(e_i) x i = σ ( e i ) , and define
st k σ = π ∘ st ( x 0 , … , x k ) . \operatorname{st}_k\sigma=\pi\circ\operatorname{st}(x_0,\ldots,x_k). st k σ = π ∘ st ( x 0 , … , x k ) . A different lift differs by a deck transformation, so equivariance makes this definition independent of the lift. Face compatibility proves that it is a chain map. On each lifted simplex the map
H σ ~ ( t , u ) = [ ( 1 − u ) f ( σ ~ ( t ) ) + u ∑ j = 0 k t j f ( x j ) ] , ( t , u ) ∈ Δ k × [ 0 , 1 ] , \mathcal{H}_{\widetilde{\sigma}}(t,u)=\left[(1-u)f(\widetilde{\sigma}(t))+u\sum_{j=0}^{k}t_jf(x_j)\right],\qquad(t,u)\in\Delta^k\times[0,1], H σ ( t , u ) = [ ( 1 − u ) f ( σ ( t )) + u j = 0 ∑ k t j f ( x j ) ] , ( t , u ) ∈ Δ k × [ 0 , 1 ] , is an equivariant homotopy from σ ~ \widetilde{\sigma} σ to st ( x 0 , … , x k ) \operatorname{st}(x_0,\ldots,x_k) st ( x 0 , … , x k ) . If σ \sigma σ is smooth, then so is this homotopy.
Proposition 6.2. On continuous real singular n n n -chains of M M M , the formula
c v o l ( σ ) = ∫ st n σ ω h c_{\mathrm{vol}}(\sigma)=\int_{\operatorname{st}_n\sigma}\omega_h c vol ( σ ) = ∫ st n σ ω h
defines a bounded cocycle with ∥ c v o l ∥ ∞ ≤ V h \lVert c_{\mathrm{vol}}\rVert_\infty\le V_h ∥ c vol ∥ ∞ ≤ V h . Its evaluation on the fundamental class is
c v o l ( [ M ] ) = Vol h ( M ) > 0. c_{\mathrm{vol}}([M]) = \operatorname{Vol}_h(M) > 0. c vol ([ M ]) = Vol h ( M ) > 0.
Proof. The integral is defined for every continuous singular simplex, because its straightening is smooth. Compatibility with faces and Stokes’ theorem show that c v o l c_{\mathrm{vol}} c vol is a cocycle. Proposition 6.1 and the uniform bound V h V_h V h show ∣ c v o l ( σ ) ∣ ≤ V h \lvert c_{\mathrm{vol}}(\sigma)\rvert\le V_h ∣ c vol ( σ )∣ ≤ V h , including degenerate cases, hence c v o l c_{\mathrm{vol}} c vol is a bounded cocycle. Choose an oriented smooth triangulation of M M M , and let z 0 z_0 z 0 be its smooth singular fundamental cycle. It satisfies
∫ z 0 ω h = ∫ M ω h = Vol h ( M ) . \int_{z_0}\omega_h=\int_M\omega_h=\operatorname{Vol}_h(M). ∫ z 0 ω h = ∫ M ω h = Vol h ( M ) .
Every real continuous singular fundamental cycle z z z is homologous to z 0 z_0 z 0 . Their difference is a singular boundary, on which the cocycle c v o l c_{\mathrm{vol}} c vol vanishes. Hence c v o l ( [ M ] ) = c v o l ( z 0 ) = Vol h ( M ) . □ c_{\mathrm{vol}}([M])=c_{\mathrm{vol}}(z_0)=\operatorname{Vol}_h(M). \quad\square c vol ([ M ]) = c vol ( z 0 ) = Vol h ( M ) . □
Proof of Corollary B. If M M M is oriented, for every real singular fundamental cycle z z z , Proposition 6.2 gives
Vol h ( M ) = ∣ c v o l ( z ) ∣ ≤ ∥ c v o l ∥ ∞ ∥ z ∥ 1 . \operatorname{Vol}_h(M)=\lvert c_{\mathrm{vol}}(z)\rvert\le\lVert c_{\mathrm{vol}}\rVert_\infty\lVert z\rVert_1. Vol h ( M ) = ∣ c vol ( z )∣ ≤ ∥ c vol ∥ ∞ ∥ z ∥ 1 .
Taking the infimum over all such cycles yields the usual bounded-cocycle lower bound for simplicial volume [11 ] :
∥ M ∥ ≥ Vol h ( M ) ∥ c v o l ∥ ∞ > 0. (59) \lVert M\rVert\ge\frac{\operatorname{Vol}_h(M)}{\lVert c_{\mathrm{vol}}\rVert_\infty}>0. \tag*{(59)} ∥ M ∥ ≥ ∥ c vol ∥ ∞ Vol h ( M ) > 0. (59)
If the original manifold is nonorientable, apply the argument to its orientation double cover M ^ \widehat{M} M , which has the same properly convex universal cover and irreducible cone. We use the standard convention for nonorientable simplicial volume, under which ∥ M ∥ = 1 2 ∥ M ^ ∥ \lVert M\rVert=\frac{1}{2}\lVert\widehat{M}\rVert ∥ M ∥ = 2 1 ∥ M ∥ . Thus the positivity conclusion holds in this case as well. □ \quad\square □
Proof of Corollary C. We only need to prove that if the cone is reducible then ∥ M ∥ = 0 \lVert M\rVert=0 ∥ M ∥ = 0 . The decomposition of a proper convex cone into indecomposable direct factors is unique up to permutation; see [1 Chapter II, Remark 1.11] . Consequently, cone automorphisms permute these factors. After passing to a finite-index subgroup Γ 0 < Γ \Gamma_0<\Gamma Γ 0 < Γ , we may therefore assume that Γ 0 \Gamma_0 Γ 0 preserves every factor and the orientation of Ω \Omega Ω . We obtain a Γ 0 \Gamma_0 Γ 0 -invariant decomposition
R N = L 1 ⊕ L 2 , C = C 1 × C 2 , \mathbb{R}^{N}=L_1\oplus L_2,\qquad C=C_1\times C_2, R N = L 1 ⊕ L 2 , C = C 1 × C 2 ,
where each C i ⊂ L i C_i\subset L_i C i ⊂ L i is a proper open convex cone. Set d i = dim L i > 0 d_i=\dim L_i>0 d i = dim L i > 0 , and consider
g t = e d 2 t id L 1 ⊕ e − d 1 t id L 2 , t ∈ R . g_t=e^{d_2t}\operatorname{id}_{L_1}\oplus e^{-d_1t}\operatorname{id}_{L_2},\qquad t\in\mathbb{R}. g t = e d 2 t id L 1 ⊕ e − d 1 t id L 2 , t ∈ R .
These transformations preserve C C C , have determinant one, and commute with Γ 0 \Gamma_0 Γ 0 . By the naturality of the canonical affine sphere, they preserve its Blaschke metric h h h . Their projective action therefore descends to an isometric flow on the closed manifold M ^ = Γ 0 \ Ω \widehat{M}=\Gamma_0\backslash\Omega M = Γ 0 \Ω .
This flow has a nowhere vanishing generating vector field. Indeed, every vector in C C C has nonzero components in both L 1 L_1 L 1 and L 2 L_2 L 2 , and the two components are rescaled with different weights. Thus no point of Ω \Omega Ω is stationary under the projective flow, and the same holds infinitesimally on M ^ \widehat{M} M .
The isometry group of the compact Riemannian manifold ( M ^ , h ) (\widehat{M},h) ( M , h ) is a compact Lie group. The closure of the descended one-parameter subgroup is therefore a positive-dimensional compact connected abelian Lie group, hence a torus. Any circle subgroup of this torus gives a nontrivial smooth S 1 S^1 S 1 -action on M ^ \widehat{M} M . Yano’s vanishing theorem [23 p. 493] now gives
∥ M ^ ∥ = 0. \lVert\widehat{M}\rVert=0. ∥ M ∥ = 0.
Finally, multiplicativity of simplicial volume under finite coverings yields
0 = ∥ M ^ ∥ = [ Γ : Γ 0 ] ∥ M ∥ , 0=\lVert\widehat{M}\rVert=[\Gamma:\Gamma_0]\lVert M\rVert, 0 = ∥ M ∥ = [ Γ : Γ 0 ] ∥ M ∥ ,
so ∥ M ∥ = 0 \lVert M\rVert=0 ∥ M ∥ = 0 . This also covers nonorientable M M M , using the orientation-cover convention adopted above. □ \quad\square □
References