This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

On tensor products of semistable lattices

Ehud de Shalit and Ori Parzanchevski Institute of mathematics, Hebrew University, Jerusalem 91904 ISRAEL deshalit@math.huji.ac.il, parzan@math.huji.ac.il
(Date: November 15, 2006)
Abstract.

We give an elementary proof of low rank cases of the conjecture that the tensor product of two semistable Euclidean lattices is again semistable.

The degree of a lattice LL in Euclidean space n\mathbb{R}^{n} is defined by the expression

(0.1) deg(L)=logvol(n/L),\deg(L)=-\log vol(\mathbb{R}^{n}/L),

where volvol stands for volume. Its rank is rk(L)=n,rk(L)=n, and the slope μ(L)\mu(L) is the ratio

(0.2) μ(L)=deg(L)rk(L).\mu(L)=\frac{\deg(L)}{rk(L)}.

The lattice LL is called semistable if for any (lower rank) sublattice S,S, μ(S)μ(L).\mu(S)\leq\mu(L). This notion is invariant under scaling, so if we normalize LL to be unimodular (μ(L)=0\mu(L)=0), it means that the volume of a fundamental parallelopiped of any sublattice SS is at least 1. Thus the shortest nonzero vector in LL has norm at least 1, every two independent vectors span a parallelogram of area at least 1, and so on.

The following problem seems to be open. The first author learned about it more than ten years ago from J.-B. Bost.

Question. Is LML\otimes M semistable whenever LL and MM are?

The notion of semistability and the question can be easily generalized to “metrized vector bundles over Spec(𝒪K)Spec(\mathcal{O}_{K})" for any number field K.K. They come up naturally in Arakelov geometry. However, it is already interesting, and probably not less difficult, over Spec(),Spec(\mathbb{Z)}, where it amounts to the statement made above. It is modeled on similar results for vector bundles over curves, and for filtered vector spaces (see [Fa],[To]). The corresponding result for filtered vector spaces is surprisingly difficult, although it is only a statement in linear algebra, and this leads one to suspect that the question for lattices is not easy either.

The purpose of this note is to give an affirmative answer in some special cases.

Theorem 0.1.

Let LL and MM be two semistable lattices. If SS is a sublattice of LML\otimes M and rk(S)3rk(S)\leq 3 then

(0.3) μ(S)μ(LM).\mu(S)\leq\mu(L\otimes M).

Bost had obtained other special cases using Geometric Invariant Theory (unpublished). We should also stress that our elementary methods do not generalize to “metrized vector bundles over Spec(𝒪K)Spec(\mathcal{O}_{K})" when KK is not .\mathbb{Q}.

1. The proof

1.1. Preliminaries on Euclidean spaces

If VV and WW are Euclidean spaces (finite dimensional inner product spaces over )\mathbb{R}) then so is VWV\otimes W with the inner product

(1.1) (vw,vw)VW=(v,v)V(w,w)W.(v\otimes w,v^{\prime}\otimes w^{\prime})_{V\otimes W}=(v,v^{\prime})_{V}(w,w^{\prime})_{W}.

If {ei}\left\{e_{i}\right\} is an orthonormal basis of VV and {ej}\left\{e_{j}^{\prime}\right\} an orthonormal basis of WW then {eiej}\left\{e_{i}\otimes e_{j}^{\prime}\right\} is an orthonormal basis of VW.V\otimes W. In this way VkV^{\otimes k} is endowed with a Euclidean structure.

The kk-th exterior power kV\bigwedge^{k}V is the quotient of VkV^{\otimes k} by the subspace 𝒩\mathcal{N} spanned by tensors v1vkv_{1}\otimes\cdot\cdot\cdot\otimes v_{k} in which, for some iji\neq j, vi=vj.v_{i}=v_{j}. The orthogonal complement 𝒩\mathcal{N}^{\perp} is the space of alternating kk-tensors, spanned by

(1.2) 1k!σSksgn(σ)vσ(1)vσ(k),\frac{1}{k!}\sum_{\sigma\in S_{k}}sgn(\sigma)v_{\sigma(1)}\otimes\cdot\cdot\cdot\otimes v_{\sigma(k)},

and this tensor projects modulo 𝒩\mathcal{N} to v1vk,v_{1}\wedge\cdot\cdot\cdot\wedge v_{k}, the image modulo 𝒩\mathcal{N} of v1vk.v_{1}\otimes\cdot\cdot\cdot\otimes v_{k}. We may therefore identify kV\bigwedge^{k}V with 𝒩\mathcal{N}^{\perp}. As such it inherits from VkV^{\otimes k} a Euclidean structure. If {ei}\left\{e_{i}\right\} is an orthonormal basis of VV then eI=ei1eik,e_{I}=e_{i_{1}}\wedge\cdot\cdot\cdot\wedge e_{i_{k}}, for I=(i1<<ik),I=(i_{1}<\dots<i_{k}), is a basis of kV,\bigwedge^{k}V, which is orthogonal but not normalized: the norm of eIe_{I} is 1/k!.1/\sqrt{k!}. To correct it, we modify the inner product that kV\bigwedge^{k}V inherits from VkV^{\otimes k} by a factor of k!k! and set

(1.3) (v1vk,v1vk)=σSksgn(σ)i=1k(vi,vσ(i)).(v_{1}\wedge\cdot\cdot\cdot\wedge v_{k},v_{1}^{\prime}\wedge\cdot\cdot\cdot\wedge v_{k}^{\prime})=\sum_{\sigma\in S_{k}}sgn(\sigma)\prod_{i=1}^{k}(v_{i},v_{\sigma(i)}^{\prime}).

In this inner product the eIe_{I} form an orthonormal basis. Moreover, the volume of the kk-dimensional parallelopiped spanned by v1,,vkv_{1},\dots,v_{k} in VV is nothing but the norm |v1vk|.|v_{1}\wedge\cdot\cdot\cdot\wedge v_{k}|. To see it, choose an orthonormal basis eie_{i} of VV the first kk vectors of which span Span{v1,,vk}Span\left\{v_{1},\dots,v_{k}\right\} (assuming the viv_{i} are linearly independent) and observe that

(1.4) v1vk=det(aij)e1ekv_{1}\wedge\cdot\cdot\cdot\wedge v_{k}=\det(a_{ij})e_{1}\wedge\cdot\cdot\cdot\wedge e_{k}

where vi=aijej.v_{i}=\sum a_{ij}e_{j}.

1.2. Length and norm in LM,L\otimes M, and the rank 1 case of the theorem

To begin the proof, note that

(1.5) μ(LM)=μ(L)+μ(M).\mu(L\otimes M)=\mu(L)+\mu(M).

If α\alpha is a vector in LML\otimes M we denote by l(α)l(\alpha) the length of α,\alpha, which is the minimal number ll such that

(1.6) α=k=1lukuk.\alpha=\sum_{k=1}^{l}u_{k}\otimes u_{k}^{\prime}.

(To avoid abuse of language, |α||\alpha| will always be called the norm of α\alpha, and not its length.) In such a case, the uku_{k} and the uku_{k}^{\prime} are linearly independent vectors in LL and MM respectively. Thus the length of a vector is at most min(rk(L),rk(M)).\min(rk(L),rk(M)).

Let us write, for any lattice MM

(1.7) μmax(M)=max{μ(L);LM}.\mu_{\max}(M)=\max\left\{\mu(L);\,L\subseteq M\right\}.

The maximum is over all sublattices of all ranks (it is easily seen that the maximum is attained). A unimodular lattice MM is semistable if and only if μmax(M)=0.\mu_{\max}(M)=0.

Proposition 1.1.

Let LL and MM be any two lattices, normalized so that μmax(L)=μmax(M)=0\mu_{\max}(L)=\mu_{\max}(M)=0 and αLM.\alpha\in L\otimes M. Then

(1.8) |α|l(α).|\alpha|\geq\sqrt{l(\alpha)}.
Proof.

Writing α=k=1lukuk\alpha=\sum_{k=1}^{l}u_{k}\otimes u_{k}^{\prime}, a direct computation gives

(1.9) |α|2=k,m(uk,um)(uk,um).|\alpha|^{2}=\sum_{k,m}(u_{k},u_{m})(u_{k}^{\prime},u_{m}^{\prime}).

Consider the Gramians A=((uk,um))A=\left((u_{k},u_{m})\right) and A=((uk,um)).A^{\prime}=\left((u_{k}^{\prime},u_{m}^{\prime})\right). These are symmetric positive definite ll by ll matrices with det(A)1\det(A)\geq 1 and det(A)1\det(A^{\prime})\geq 1 because we normalized LL and MM so that the volume of the parallelopiped formed by the uku_{k} or the uku_{k}^{\prime} is at least 1.

We are indebted to Assaf Goldberger for pointing out the following fact.

Lemma 1.2.

Let AA and AA^{\prime} be two symmetric positive definite ll by ll matrices. Then the spectrum of AAAA^{\prime} is real and positive.

The matrix AA,AA^{\prime}, although not necessarily symmetric, is conjugate to

(1.10) A1/2AA1/2=(A1/2)tAA1/2,A^{1/2}A^{\prime}A^{1/2}=(A^{1/2})^{t}A^{\prime}A^{1/2},

which is symmetric positive definite since AA^{\prime} is.

We now continue the proof of the proposition. By our assumption, the product of the eigenvalues of AAAA^{\prime} is at least 1, and the lemma implies that they are real and positive. By the inequality between the geometric and arithmetic means, Tr(AA)l.Tr(AA^{\prime})\geq l. However,

(1.11) Tr(AA)=|α|2,Tr(AA^{\prime})=|\alpha|^{2},

so |α|l|\alpha|\geq\sqrt{l} as needed. ∎

The proposition proves the case 𝐫𝐤(𝐒)=𝟏\mathbf{rk(S)=1} of the theorem, but to prove the cases rk(S)=2rk(S)=2 or 33 we shall need its full strength.

1.3. The rank 2 case of the theorem

Let now SS be a sublattice of rank 22 in LM.L\otimes M. Let α\alpha be a nonzero vector of shortest norm in S,S, and β\beta another vector in SS such that S=α+βS=\mathbb{Z}\alpha+\mathbb{Z}\beta. Subtracting a multiple of α\alpha from β\beta we may assume that the angle between α\alpha and β\beta is between 6060^{\circ} and 120.120^{\circ}. It follows that

(1.12) |αβ|32|α||β|.|\alpha\wedge\beta|\geq\frac{\sqrt{3}}{2}|\alpha||\beta|.

If either of α\alpha or β\beta is indecomposable (of length 2\geq 2) we are done, since its norm is then at least 2\sqrt{2} and the other’s norm is at least 1, but 23>2.\sqrt{2}\sqrt{3}>2. We may therefore assume that

(1.13) α=v1w1,β=v2w2\alpha=v_{1}\otimes w_{1},\,\,\beta=v_{2}\otimes w_{2}

are both decomposable tensors. We now use the identity

(1.14) |(v1w1)(v2w2)|2\displaystyle|(v_{1}\otimes w_{1})\wedge(v_{2}\otimes w_{2})|^{2} =\displaystyle= |v1|2|v2|2|w1w2|2+|v1v2|2|w1|2|w2|2\displaystyle|v_{1}|^{2}|v_{2}|^{2}|w_{1}\wedge w_{2}|^{2}+|v_{1}\wedge v_{2}|^{2}|w_{1}|^{2}|w_{2}|^{2}
|v1v2|2|w1w2|2,\displaystyle-|v_{1}\wedge v_{2}|^{2}|w_{1}\wedge w_{2}|^{2},

which follows at once from the fact that

(1.15) |αβ|2=|α|2|β|2(α,β)2.|\alpha\wedge\beta|^{2}=|\alpha|^{2}|\beta|^{2}-(\alpha,\beta)^{2}.

If v1v_{1} and v2v_{2} are proportional, then w1w_{1} and w2w_{2} are independent, and |v1|2|v2|2|w1w2|2|v_{1}|^{2}|v_{2}|^{2}|w_{1}\wedge w_{2}|^{2} 1,\geq 1, while the second and the third terms vanish. Otherwise, the first term is larger than the third, and the second is 1\geq 1 by our assumption. This concludes the rank 2 case.

1.4. The rank 3 case of the theorem

The rank 3 case is handled similarly, but the details are more complicated. Let rk(S)=3.rk(S)=3. Let α\alpha be a shortest nonzero vector in SS and β\beta a second shortest vector independent from α,\alpha, and assume as before that the angle between α\alpha and β\beta is between 6060^{\circ} and 120.120^{\circ}. Complete to a basis of SS by a vector γ,\gamma, |γ||β||α|1.|\gamma|\geq|\beta|\geq|\alpha|\geq 1. Using orthonormal coordinates x,y,zx,y,z in the real subspace spanned by S,S, such that the xx-axis is in the α\alpha-direction and the (x,y)(x,y)-plane is the plane spanned by α\alpha and β,\beta, we may assume, subtracting from γ\gamma a suitable integral linear combination of α\alpha and β\beta that

(1.16) γ=(x,y,z)\gamma=(x,y,z)

with |x||α|/2|x|\leq|\alpha|/2 and |y||β|/2.|y|\leq|\beta|/2. This is because the rectangle

(1.17) {|x||α|/2,|y||β|sinθ/2},\left\{|x|\leq|\alpha|/2,\,\,|y|\leq|\beta|\sin\theta/2\right\},

where θ\theta is the angle between α\alpha and β,\beta, is a fundamental domain for α+β.\mathbb{Z}\alpha+\mathbb{Z}\beta. We now have

(1.18) z2|γ|2(|α|2+|β|2)/4|γ|2/2z^{2}\geq|\gamma|^{2}-(|\alpha|^{2}+|\beta|^{2})/4\geq|\gamma|^{2}/2

and

(1.19) |αβ|1|\alpha\wedge\beta|\geq 1

(by the rank 2 case), so

(1.20) |αβγ|2=|αβ|2z2|γ|2/2.|\alpha\wedge\beta\wedge\gamma|^{2}=|\alpha\wedge\beta|^{2}z^{2}\geq|\gamma|^{2}/2.

If l(α)l(\alpha), l(β)l(\beta) or l(γ)2,l(\gamma)\geq 2, then by the proposition and the fact that |γ||β||α|,|\gamma|\geq|\beta|\geq|\alpha|, we must have |γ|22,|\gamma|^{2}\geq 2, and we are done.

There remains the case α=v1w1,\alpha=v_{1}\otimes w_{1}, β=v2w2\beta=v_{2}\otimes w_{2} and γ=v3w3,\gamma=v_{3}\otimes w_{3}, in which the proposition does not help us. The key in this case is the following identity.

Lemma 1.3.

We have an 18-term relation

(1.21) |αβγ|2\displaystyle|\alpha\wedge\beta\wedge\gamma|^{2} =\displaystyle= |v1|2|v2|2|v3|2|w1w2w3|2+|v1v2v3|2|w1|2|w2|2|w3|2\displaystyle|v_{1}|^{2}|v_{2}|^{2}|v_{3}|^{2}|w_{1}\wedge w_{2}\wedge w_{3}|^{2}+|v_{1}\wedge v_{2}\wedge v_{3}|^{2}|w_{1}|^{2}|w_{2}|^{2}|w_{3}|^{2}
12i=13|vi|2|vivi′′|2|wi|2|wiwi′′|2\displaystyle-\frac{1}{2}\sum_{i=1}^{3}|v_{i}|^{2}|v_{i^{\prime}}\wedge v_{i^{\prime\prime}}|^{2}|w_{i}|^{2}|w_{i^{\prime}}\wedge w_{i^{\prime\prime}}|^{2}
+12ij|vi|2|vivi′′|2|wj|2|wjwj′′|2\displaystyle+\frac{1}{2}\sum_{i\neq j}|v_{i}|^{2}|v_{i^{\prime}}\wedge v_{i^{\prime\prime}}|^{2}|w_{j}|^{2}|w_{j^{\prime}}\wedge w_{j^{\prime\prime}}|^{2}
12i=13|vi|2|vivi′′|2|w1w2w3|2\displaystyle-\frac{1}{2}\sum_{i=1}^{3}|v_{i}|^{2}|v_{i^{\prime}}\wedge v_{i^{\prime\prime}}|^{2}|w_{1}\wedge w_{2}\wedge w_{3}|^{2}
12i=13|v1v2v3|2|wi|2|wiwi′′|2\displaystyle-\frac{1}{2}\sum_{i=1}^{3}|v_{1}\wedge v_{2}\wedge v_{3}|^{2}|w_{i}|^{2}|w_{i^{\prime}}\wedge w_{i^{\prime\prime}}|^{2}
+12|v1v2v3|2|w1w2w3|2,\displaystyle+\frac{1}{2}|v_{1}\wedge v_{2}\wedge v_{3}|^{2}|w_{1}\wedge w_{2}\wedge w_{3}|^{2},

where we have used the convention that {i,i′′}\{i^{\prime},i^{\prime\prime}\} are the two indices complementary to ii.

Proof.

 Expanding the grammian of three vectors in Euclidean space we get

(1.22) |α1α2α3|2\displaystyle|\alpha_{1}\wedge\alpha_{2}\wedge\alpha_{3}|^{2} =\displaystyle= i=13|αi|2i=13|αi|2(αi,αi′′)2\displaystyle\prod_{i=1}^{3}|\alpha_{i}|^{2}-\sum_{i=1}^{3}|\alpha_{i}|^{2}(\alpha_{i^{\prime}},\alpha_{i^{\prime\prime}})^{2}
+2(α1,α2)(α2,α3)(α3,α1).\displaystyle+2(\alpha_{1},\alpha_{2})(\alpha_{2},\alpha_{3})(\alpha_{3},\alpha_{1}).

For our decomposable tensors we get the expression

(1.23) |αβγ|2\displaystyle|\alpha\wedge\beta\wedge\gamma|^{2} =\displaystyle= i=13|vi|2|wi|2i=13|vi|2(vi,vi′′)2|wi|2(wi,wi′′)2\displaystyle\prod_{i=1}^{3}|v_{i}|^{2}|w_{i}|^{2}-\sum_{i=1}^{3}|v_{i}|^{2}(v_{i^{\prime}},v_{i^{\prime\prime}})^{2}|w_{i}|^{2}(w_{i^{\prime}},w_{i^{\prime\prime}})^{2}
+2(v1,v2)(v2,v3)(v3,v1)(w1,w2)(w2,w3)(w3,w1).\displaystyle+2(v_{1},v_{2})(v_{2},v_{3})(v_{3},v_{1})(w_{1},w_{2})(w_{2},w_{3})(w_{3},w_{1}).

We proceed to replace all the inner products by expressions involving only wedge-products (sines instead of cosines). The elements of the form (x,y)2(x,y)^{2} can be replaced by |x|2|y|2|xy|2|x|^{2}|y|^{2}-|x\wedge y|^{2} as in the case of rank 2. For (x,y)(y,z)(z,x)(x,y)(y,z)(z,x) use

(1.24) 2(x,y)(y,z)(z,x)\displaystyle 2(x,y)(y,z)(z,x) =\displaystyle= |xyz|2|x|2|y|2|z|2+|x|2(y,z)2+\displaystyle|x\wedge y\wedge z|^{2}-|x|^{2}|y|^{2}|z|^{2}+|x|^{2}(y,z)^{2}+\cdot\cdot\cdot
=\displaystyle= |xyz|2|x|2|y|2|z|2+|x|2(|y|2|z|2|yz|2)+\displaystyle|x\wedge y\wedge z|^{2}-|x|^{2}|y|^{2}|z|^{2}+|x|^{2}(|y|^{2}|z|^{2}-|y\wedge z|^{2})+\cdot\cdot\cdot

This is an 5-term expression and when we multiply the one for the visv_{i}^{\prime}s with the one for the wisw_{i}^{\prime}s we get 25 terms, and a total of 38 terms for |αβγ|2.|\alpha\wedge\beta\wedge\gamma|^{2}. A direct computation (carried out by a computer) reveals that there are many cancellations, and the 18-term relation falls out. ∎

We continue the proof of the theorem, when α\alpha, β\beta and γ\gamma are decomposable as above. Choose an orthonormal basis e1,e2,e3,e_{1},e_{2},e_{3},... of LL_{\mathbb{R}} so that v1e1,v_{1}\in\langle e_{1}\rangle, v2e1,e2,v_{2}\in\langle e_{1},e_{2}\rangle, v3e1,e2,e3v_{3}\in\langle e_{1},e_{2},e_{3}\rangle and similarly an orthonormal basis e1,e2,e3,e_{1}^{\prime},e_{2}^{\prime},e_{3}^{\prime},... of MM_{\mathbb{R}} putting the wiw_{i} in a triangular form. Assume first that the viv_{i} are linearly independent. Expanding in the orthonormal basis eieje_{i}\otimes e_{j}^{\prime} of LML_{\mathbb{R}}\otimes M_{\mathbb{R}} and in the corresponding orthonormal basis of of 3(LM)\bigwedge^{3}(L_{\mathbb{R}}\otimes M_{\mathbb{R}}) (recall the normalization from 1.1)

(1.25) |αβγ|2\displaystyle|\alpha\wedge\beta\wedge\gamma|^{2} \displaystyle\geq i=12j=13coef2((e1e1)(e2ei)(e3ej))\displaystyle\sum_{i=1}^{2}\sum_{j=1}^{3}coef^{2}((e_{1}\otimes e_{1}^{\prime})\wedge(e_{2}\otimes e_{i}^{\prime})\wedge(e_{3}\otimes e_{j}^{\prime}))
=\displaystyle= |v1v2v3|2|w1|2|w2|2|w3|21.\displaystyle|v_{1}\wedge v_{2}\wedge v_{3}|^{2}|w_{1}|^{2}|w_{2}|^{2}|w_{3}|^{2}\geq 1.

Assume next that both the viv_{i} and the wiw_{i} are linearly dependent. If v1v_{1} and v2v_{2} are proportional then w1w_{1} and w2w_{2} can not be proportional, otherwise rk(S)2.rk(S)\leq 2. In this case we may assume that v1,v2e1v_{1},v_{2}\in\langle e_{1}\rangle and v3e1,e2,v_{3}\in\langle e_{1},e_{2}\rangle, while w1e1w_{1}\in\langle e_{1}^{\prime}\rangle and w2,w3e1,e2.w_{2},w_{3}\in\langle e_{1}^{\prime},e_{2}^{\prime}\rangle. We find

(1.26) |αβγ|2\displaystyle|\alpha\wedge\beta\wedge\gamma|^{2} =\displaystyle= i=12coef2((e1e1)(e1e2)(e2ei))\displaystyle\sum_{i=1}^{2}coef^{2}((e_{1}\otimes e_{1}^{\prime})\wedge(e_{1}\otimes e_{2}^{\prime})\wedge(e_{2}\otimes e_{i}^{\prime}))
=\displaystyle= |v1||v2||v1v3||v2v3||w1w2|2|w3|21.\displaystyle|v_{1}||v_{2}||v_{1}\wedge v_{3}||v_{2}\wedge v_{3}||w_{1}\wedge w_{2}|^{2}|w_{3}|^{2}\geq 1.

There remains the case where the viv_{i} and the wiw_{i} are linearly dependent, but no two vectors in each triplet are proportional.

Introduce the notation

(1.27) λi=|vi||vivi′′|\lambda_{i}=|v_{i}||v_{i^{\prime}}\wedge v_{i^{\prime\prime}}|

and similarly μi=|wi||wiwi′′|.\mu_{i}=|w_{i}||w_{i^{\prime}}\wedge w_{i^{\prime\prime}}|. Note that

λi=|v1||v2||v3||sinθi|\lambda_{i}=|v_{1}||v_{2}||v_{3}||\sin\theta_{i}|

where θi\theta_{i} is the angle between viv_{i^{\prime}} and vi′′v_{i^{\prime\prime}} in the plane spanned by the vis.v_{i}^{\prime}s. We may assume that θi=θi+θi′′.\theta_{i}=\theta_{i^{\prime}}+\theta_{i^{\prime\prime}}. A direct computation shows then that

(1.28) λi2λi2λi′′2=±2λiλi′′cosθi.\lambda_{i}^{2}-\lambda_{i^{\prime}}^{2}-\lambda_{i^{\prime\prime}}^{2}=\pm 2\lambda_{i^{\prime}}\lambda_{i^{\prime\prime}}\cos\theta_{i}.

Similarly denote by ωi\omega_{i} the angle between wiw_{i^{\prime}} and wi′′.w_{i^{\prime\prime}}. We now observe that in the 18-term relation, nine terms drop out and the remaining nine give

(1.32) |αβγ|2\displaystyle|\alpha\wedge\beta\wedge\gamma|^{2} =\displaystyle= 12{λ12μ12+λ12μ22+λ12μ32+λ22μ12λ22μ22+λ22μ32+λ32μ12+λ32μ22λ32μ32}\displaystyle\frac{1}{2}\left\{\begin{array}[]{lll}-\lambda_{1}^{2}\mu_{1}^{2}&+\lambda_{1}^{2}\mu_{2}^{2}&+\lambda_{1}^{2}\mu_{3}^{2}\\ +\lambda_{2}^{2}\mu_{1}^{2}&-\lambda_{2}^{2}\mu_{2}^{2}&+\lambda_{2}^{2}\mu_{3}^{2}\\ +\lambda_{3}^{2}\mu_{1}^{2}&+\lambda_{3}^{2}\mu_{2}^{2}&-\lambda_{3}^{2}\mu_{3}^{2}\end{array}\right\}
=\displaystyle= 12{λ12(μ12+μ22+μ32)+λ22(μ12μ22+μ32)+(λ12+λ22±2λ1λ2cosθ3)(μ12+μ22μ32)}\displaystyle\frac{1}{2}\left\{\begin{array}[]{c}\lambda_{1}^{2}(-\mu_{1}^{2}+\mu_{2}^{2}+\mu_{3}^{2})+\lambda_{2}^{2}(\mu_{1}^{2}-\mu_{2}^{2}+\mu_{3}^{2})\\ +(\lambda_{1}^{2}+\lambda_{2}^{2}\pm 2\lambda_{1}\lambda_{2}\cos\theta_{3})(\mu_{1}^{2}+\mu_{2}^{2}-\mu_{3}^{2})\end{array}\right\}
=\displaystyle= λ12μ22+λ22μ12±λ1λ2cosθ3(μ12+μ22μ32)\displaystyle\lambda_{1}^{2}\mu_{2}^{2}+\lambda_{2}^{2}\mu_{1}^{2}\pm\lambda_{1}\lambda_{2}\cos\theta_{3}(\mu_{1}^{2}+\mu_{2}^{2}-\mu_{3}^{2})
=\displaystyle= λ12μ22+λ22μ12±2λ1λ2μ1μ2cosθ3cosω3\displaystyle\lambda_{1}^{2}\mu_{2}^{2}+\lambda_{2}^{2}\mu_{1}^{2}\pm 2\lambda_{1}\lambda_{2}\mu_{1}\mu_{2}\cos\theta_{3}\cos\omega_{3}
\displaystyle\geq λ12μ22+λ22μ12λ12μ22cos2θ3λ22μ12cos2ω3\displaystyle\lambda_{1}^{2}\mu_{2}^{2}+\lambda_{2}^{2}\mu_{1}^{2}-\lambda_{1}^{2}\mu_{2}^{2}\cos^{2}\theta_{3}-\lambda_{2}^{2}\mu_{1}^{2}\cos^{2}\omega_{3}
=\displaystyle= λ12μ22sin2θ3+λ22μ12sin2ω3\displaystyle\lambda_{1}^{2}\mu_{2}^{2}\sin^{2}\theta_{3}+\lambda_{2}^{2}\mu_{1}^{2}\sin^{2}\omega_{3}
\displaystyle\geq λ12μ22|v1|2|v2|2+λ22μ12|w1|2|w2|2\displaystyle\frac{\lambda_{1}^{2}\mu_{2}^{2}}{|v_{1}|^{2}|v_{2}|^{2}}+\frac{\lambda_{2}^{2}\mu_{1}^{2}}{|w_{1}|^{2}|w_{2}|^{2}}
=\displaystyle= |v2v3|2|w2|2|w1w3|2|v2|2+|w2w3|2|v2|2|v1v3|2|w2|2\displaystyle\frac{|v_{2}\wedge v_{3}|^{2}|w_{2}|^{2}|w_{1}\wedge w_{3}|^{2}}{|v_{2}|^{2}}+\frac{|w_{2}\wedge w_{3}|^{2}|v_{2}|^{2}|v_{1}\wedge v_{3}|^{2}}{|w_{2}|^{2}}
(1.36) \displaystyle\geq |w2|2|v2|2+|v2|2|w2|22.\displaystyle\frac{|w_{2}|^{2}}{|v_{2}|^{2}}+\frac{|v_{2}|^{2}}{|w_{2}|^{2}}\geq 2.

This concludes the proof of the theorem.

1.5. An argument from duality

Duality allows us to conclude more. Quite generally, if

(1.37) 0PQR00\rightarrow P\rightarrow Q\rightarrow R\rightarrow 0

is an exact sequence of Euclidean lattices (this means that Q/PQ/P is torsion-free and that the metrics on PP_{\mathbb{R}} and RR_{\mathbb{R}} are the ones induced from the metric on QQ_{\mathbb{R}}) then

(1.38) μ(Q)=rkPrkQμ(P)+rkRrkQμ(R).\mu(Q)=\frac{rkP}{rkQ}\mu(P)+\frac{rkR}{rkQ}\mu(R).

If RR^{\vee} is the dual lattice, μ(R)=μ(R).\mu(R^{\vee})=-\mu(R). From these two observations it follows at once that if LL is semistable, so is L.L^{\vee}.

Let LL and MM be two unimodular semistable lattices of ranks nn and mm. Then LL^{\vee} and MM^{\vee} are unimodular and semistable too. Suppose we know the desired results for all SS of some rank s.s. Let PP be a rank nmsnm-s (we shall say it has corank ss) sublattice of LML\otimes M for which we want to prove that μ(P)0.\mu(P)\leq 0. We may assume that (LM)/P\,(L\otimes M)/P is torsion free. We equip it with the quotient metric and call it RR. Then RR is a rank ss lattice and R=PR^{\vee}=P^{\perp} is a rank ss sublattice of LM.L^{\vee}\otimes M^{\vee}. By what we have seen already, μ(R)0,\mu(R^{\vee})\leq 0, but

(1.39) μ(P)=snmsμ(R)\mu(P)=-\frac{s}{nm-s}\mu(R)

since μ(LM)=0,\mu(L\otimes M)=0, and μ(R)=μ(R).\mu(R)=-\mu(R^{\vee}). This shows that if we know the desired result for all rank ss lattices, we also know it for all corank ss lattices.

Corollary 1.4.

If LL and MM are semistable and they are both of rank 2, or one of them is of rank 2 and the other is of rank 3, then LML\otimes M is semistable.

References

[Fa] Faltings, G.: Mumford-Stabilität in der algebraischen Geometrie, Proceedings of the 1994 ICM.

[To] Totaro, B.: Tensor products in pp-adic Hodge theory, Duke Math.J. 83 (1996), 79-104.