On the third secant variety

We determine normal forms and ranks of tensors of border rank at most three. We present a differential-geometric analysis of limits of secant planes in a more general context. In particular there are at most four types of points on limiting trisecant planes for cominuscule varieties such as Grassmannians. We also show the singular locus of the first two secant varieties of all triple Segre products has codimension at least two.


Introduction
Throughout the paper we work over the field of complex numbers C. Motivated by applications, there has been a considerable amount of recent research on ranks and border ranks of tensors, see, e.g., [15,9] and references therein. In signal processing one is interested in determining ranks of tensors, see, e.g., [6] and references therein. In computational complexity, one looks for exotic algorithms via limits of tensors of a given rank, see [16]. There are adequate tests to determine the border ranks of tensors of small border rank, however the possible ranks of such tensors are not well understood. In this article we present normal forms for tensors of border rank three. Already in this case the problem becomes subtle. We work in the more general setting of secant varieties.
1.1. Definitions, notational conventions. For a projective variety X ⊂ PV not contained in a hyperplane, the X-rank of p ∈ PV , R X (p), is defined to be the smallest r such that there exist x 1 , . . . , x r ∈ X such that p is in the span of of x 1 , . . . , x r , and the X-border rank R X (p) is defined to be the smallest r such that there exist curves x 1 (t), . . . , x r (t) ∈ X such that p is in the span of the limiting plane lim t→0 x 1 (t), . . . , x r (t) . Let σ r (X) ⊂ PV denote the set of points of X-border rank at most r. When X = Seg(PA 1 × · · · × PA n ) ⊂ P(A 1 ⊗ · · · ⊗ A n ) is the set of rank one tensors in a space of tensors, the X-rank and border rank agree with the usual notions of tensor rank and border rank. The set of points of X-rank r contains a Zariski open subset of σ r (X) and we are interested in the complement of this set.
We let σ 0 r (X) denote the points of σ r (X) of rank r. The tangential variety of a smooth variety X ⊂ PV , τ (X) ⊂ PV , consists of all points on all embedded tangent P 1 's. For varieties X, Y ⊂ PV , define J(X, Y ) := {p ∈ PV | ∃x ∈ X, y ∈ Y such that p ∈ x, y }, the join of X and Y . Note that J(X, X) = σ 2 (X). For a set Z ⊂ PV ,Ẑ ⊂ V denotes the cone over it and Z its linear span. For a variety Y ⊂ PV , Y sing denotes the singular points of Y . The affine tangent space to a variety X ⊂ PV at a smooth point x is denotedT x X ⊂ V .
Throughout the paper we assume A 1 , . . . , A n , A, B, C are complex vector spaces of dimension at least 2.

1.2.
Results on ranks and normal forms for tensors. The following proposition was probably "known to the experts" but we did not find it in the literature, so we include a statement and proof. Proposition 1.1. Let X = Seg(PA 1 × · · · × PA n ) ⊂ P(A 1 ⊗ · · · ⊗ A n ) be a Segre variety. There is a normal form for points x ∈σ 2 (X): (a) x = a 1 1 ⊗ · · · ⊗ a n 1 for a point of X, which has rank 1, (b) x = a 1 1 ⊗ · · · ⊗ a n 1 + a 1 2 ⊗ · · · ⊗ a n 2 for a point on a secant line to X (here we require at least two of the a i 2 to be independent of the corresponding a i 1 ), which has rank 2, (c) and for each J ⊆ {1, . . . , n}, |J| > 2, the normal form (1.1) x = j∈J a 1 1 ⊗ · · · ⊗ a j−1 1 ⊗a j 2 ⊗a j+1 1 ⊗ · · · ⊗ a n 1 where each a j 2 is independent of the corresponding a j 1 . This case has rank |J|. In particular, all ranks from 1 to n occur for elements of σ 2 (X).
We prove an analogous result for σ 3 (Seg(PA × PB × PC)): Theorem 1.4. Let p ∈ σ 3 (Seg(PA × PB × PC)). If p is a general point of type (ii) or (iii), or a general point of any component of points of type (iv), then p is a nonsingular point of σ 3 (Seg(PA × PB × PC)). Moreover, if dim A, dim B, dim C ≥ 3, and p is a general point in the set of the points contained in some P(C 2 ⊗ C 3 ⊗ C 3 ), then p is a nonsingular point of σ 3 (Seg(PA × PB × PC)), and similarly for permuted statements.
For type (iv) there are two other normal forms, where the role of a is switched with that of b and c. These normal forms are depicted in terms of "slices" in Table 1 on page 18. (In the tensor literature, 3-way tensors T ∈ A⊗B⊗C are often studied by their images T (A * ) ⊂ B⊗C etc... and these images are studied in terms of bases, resulting in a parametrized subspace of a space of matrices. These parametrized spaces of matrices are called slices.) Here a j , b j , c j need not be independent vectors, so to parametrize the spaces, fix bases of each space and write the a j , b j , c j as arbitrary linear combinations of basis vectors. (However there are some independence requirements.) Here are normal forms for all n: p (i) = a 1 1 ⊗ · · · ⊗ a n 1 + a 1 2 ⊗ · · · ⊗ a n 2 + a 1 3 ⊗ · · · ⊗ a n 3 (1.2) Again, (1.5) has n − 1 other normal forms, where the role of a 1 * is exchanged with a i * . Also, the vectors need not all be linearly independent.
Remark 1.5. In contrast to case (iv) above, already with four points on a three factor Segre spanning a three dimensional vector space, one can obtain new limits by taking a second derivative, even when the limiting points are distinct. Consider the points x 1 = a 1 ⊗b 1 ⊗c 1 , x 2 = a 2 ⊗b 2 ⊗c 1 , Here both first and second derivatives of curves give new points. More generally, consider Any four points lying on Seg(v 2 (P 1 ) × P 0 × · · · × P 0 ) will be linearly dependent. Exceptional limit points turn out to be important -an exceptional limit in σ 5 (Seg(PA × PB × PC)) is used in Bini's approximate algorithm to multiply 2 × 2 matrices with an entry zero, and an exceptional limit in σ 7 (Seg(PA × PB × PC)) is used in Schönhage's approximate algorithm to multiply 3 × 3 matrices using 21 multiplications, see [3, §4.4].
Since there are only finitely many configurations of triples of points in A i up to the action of GL(A i ), we conclude: Corollary 1.6. There are only finitely many orbits of the action of GL(A 1 ) × · · · × GL(A n ) on σ 3 (Seg(PA 1 × · · · × PA n )).
In the three factor case, there are 39 orbits, see §6.
Remark 1.7. Points of the form y + y ′ + y ′′ where y(t) is a curve onŜeg(PA 1 ⊗ · · · ⊗ PA n ) have rank at most n+1 2 because all such points are of the form (1.4) (perhaps with linearly dependent variables). The bound R Seg(PA 1 ⊗···⊗ PAn) (y + y ′ + y ′′ ) ≤ n+1 2 is not tight, as for n = 3 the following theorem shows R Seg(PA×PB×PC) (y + y ′ + y ′′ ) is at most five.  5. All other points of σ 3 (Seg(PA × PB × PC)) have rank less than five, so in particular, the maximum rank of any point of σ 3 (Seg(PA × PB × PC)) is 5. Remark 1.9. Theorem 1.8 seems to have been a "folklore" theorem in the tensor literature. For example, in [9], Table 3.2 the result is stated and refers to [10], but in that paper the result is stated and a paper that never appeared is referred to. Also, there appear to have been privately circulating proofs, one due to R. Rocci from 1993 has been shown to us. We thank M. Mohlenkamp for these historical remarks.
The Comon conjecture on ranks says that for T ∈ S d V ⊂ V ⊗d the symmetric tensor rank of T equals the tensor rank of T . Corollary 1.10. The Comon conjecture holds for T ∈σ 3 (v 3 (PV )). Corollary 1.10 follows by comparing the normal forms and ranks of this paper with those of [12].
In §3 we generalize Theorem 1.2 to generalized cominuscule varieties, a class of homogeneous varieties which includes Grassmannians and spinor varieties. See §3 for the definition of a generalized cominuscule variety, and §2 for the definition of the second fundamental form II. Theorem 1.11. Let X ⊂ PV be generalized cominuscule. Then [p] ∈ σ 3 (X) if and only if at least one of the following situations occurs: To make sense of elements of the tangent and normal spaces as elements of V we have chosen a splitting V =x ⊕ T ⊕ N as described in §2.1.

1.3.
Overview. In §2 we review facts from projective differential geometry. In §3 we prove Theorem 1.11. In §4 we apply Theorem 1.11 to cominuscule varieties, including Grassmannians and spinor varieties. In §5 we analyze the case of the Segre variety in detail, and we give two proofs of Theorem 1.2, a short proof by computing the Lie algebras of the stabilizers of the points p ( * ) , and a longer proof that contains more precise information which is of interest in its own right. In §6 we restrict attention to the three-factor Segre variety, and prove Theorems 1.3, 1.4 and 1.8.
1.4. Acknowledgments. We thank M. Mohlenkamp for pointing out an error in an earlier version of this article, related to the rank of y + y ′ + y ′′ in Theorem 1.8. This paper grew out of questions raised at the 2008 AIM workshop Geometry and representation theory of tensors for computer science, statistics and other areas, and the authors thank AIM and the conference participants for inspiration. The mathematics in this paper was finally completed while the authors were guests at the Mittag-Leffler Institut in Spring 2011 and we gratefully thank the institute for providing a wonderful environment for doing mathematics. We truly appreciate the help of the referee, his careful proof reading of the article, and his many thoughtful comments.
2. Curves in submanifolds of projective space 2.1. Fubini forms, fundamental forms, and the prolongation property. Let X n ⊂ PV be a subvariety and let o ∈ X be a smooth point. We may choose a splitting such thatô ≃ C is the one dimensional linear subspace corresponding to o ∈ PV , andô ⊕ T is the affine tangent spaceT o X. We will abuse notation and identify T with the Zariski tangent space T o X =ô * ⊗(T o X/ô) and N with the normal space N o X := T o PV /T o X. Since we are working at a point, the twist by the line bundle will not matter. Our choice of splitting will not effect the end results of the calculations.
Any point [v] ∈ PV has a lift to a point v ∈ V of the form (ō,ṽ, v N ) or (0,ṽ, v N ), where 0 and o are points inô ≃ C, andṽ ∈ T , v N ∈ N . In an analytic neighborhood of o we may write X as a graph, that is, for x ∈ X near o, the vector x N depends holomorphically on the vectorx and we expand this holomorphic map into a Taylor series: Herex ∈ T andx s ∈ S s T . Later we will study curves x(t) ∈ X, and express the whole curve using (2.2), writingx(t) to be the curve in T ,x s (t) ∈ S s T . is well-defined (independent of the choice of splitting (2.1)) and called the s-th fundamental form of X at o. Fundamental forms satisfy a prolongation property (see [8,Chap. 3]): if o ∈ X is a general point, then for all f 1 ∈ S s 1 T and f 2 ∈ S s 2 T we have We write III o = F 3,o . If there is no risk of confusion, we will often omit the base point and write II := II o , F s := F s,o , etc.

2.2.
When taking limits, we may assume one curve is stationary.
Lemma 2.1. Let G be a connected algebraic group and P a parabolic subgroup. Let X = G/P ⊂ PV be a homogeneously embedded homogeneous variety and let p ∈ σ r (X). Then there exist a point ξ ∈X and r − 1 curves y j (t) ∈X such that p ∈ lim t→0 ξ, y 1 (t), . . . , y r−1 (t) .
Set ξ = x 0 and appropriately modify the y j (t) to complete the proof.
We remark, that for non-homogeneous X, an analogous statement is rarely true. If r = 2, and X is smooth, then it is true, see Proposition 2.3. But already if r = 2 and X is singular, one often needs both curves moving (a cuspidical rational curve embedded in P 3 is an example). Also if r = 3, and X has a trisecant line (for example X is a high degree rational normal curve projected from a general point on a trisecant plane), then one also needs three curves moving to obtain some of the points on the third secant variety.
2.3. Dimension counting and higher order invariants. Since dim σ r (X) ≤ r dim X +r −1, one can use a parameter count to see what one expects in choosing a point of the boundary. Suppose dim X > 1, X is not a cone and the third fundamental form is nonzero -for example X = Seg(PA × PB × PC). One can predict that the third fundamental form does not arise when computing a point of σ 3 (X) which is on a plane obtained as a limit of spans of 3 points converging to the same general point of X. This is because the third fundamental form is only well defined modulo the second osculating space, which will have dimension greater than dim X. In the case of the three factor Segre variety the second osculating space has dimension ab + ac + bc, and the third fundamental form is only well defined modulo the second osculating space. So were there a term III(v 3 ) appearing in an expression for a point on σ 3 (Seg(PA × PB × PC)), with no restrictions on v, then the resulting variety would have to have dimension at least ab + ac + bc for the term to be well defined. If the dimensions of the vector spaces are sufficiently large, this contradicts the dimension count. Such heuristics can be useful in calculations.
The following lemma will allow us to eliminate higher fundamental forms from our considerations when studying σ 3 (X). It illustrates the dimension counting principle.
, the power series ring with coefficients in S, and let J k be the ideal generated by I and t k . The curveṽ =ṽ(t) =ṽ 0 + tṽ 1 + t 2ṽ 2 + · · · is naturally an element in S[[t]]. In this interpretation, F s (ṽ(t) s ) = t k (· · · ) if and only ifṽ(t) s ∈ J k . In particular, our assumptions are: •ṽ(t) 2 ∈ J m and • the constant coefficientṽ 2 0 ∈ I (because m > 0), thus alsoṽ s 0 ∈ I for s ≥ 2. To show thatṽ(t) s ∈ J m+s−2 for s ≥ 2, we argue by induction on s.
Since the constant coefficientṽ 0 s ∈ I, it follows, thatṽ(t) s ∈ J m+s−2 as claimed.
2.4. Points on σ 2 (X). We reprove the standard fact that a point on a secant variety to a smooth variety X is either on X, on an honest secant line, or on a tangent line to X. The proof we present prepares the way for new results. Recall that if a point of σ 2 (X) is not on an honest secant line, it must arise from a point on a limiting P 1 which is obtained by a curve of P 1 's, Proposition 2.3. Let X ⊂ PV be a smooth variety and let [z] ∈ σ 2 (X)\σ 2 (X) 0 . Then z may be obtained from first order information, that is, Proof. There exist curves [ may be obtained as a point of the limiting P 1 = P(lim t→0 x(t), y(t) ).
Then the same affine plane can be obtained as lim t→0 ō, z(t) , thus one point is fixed and the other approaches the first one from the direction ofṽ 0 .
3. Generalized cominuscule varieties: proof of theorem 1.11 Following [13], a homogeneously embedded homogeneous variety G/P ⊂ PV is called generalized cominuscule if there is a choice of splitting (at any point) such that the Fubini forms reduce to fundamental forms, that is: . . , f }, and F s = F s = 0 for all s > f . Generalized cominuscule varieties may be characterized intrinsically as the homogeneously embedded G/P where the unipotent radical of P is abelian. A generalized cominuscule variety is cominuscule if and only if G is simple and the embedding is the minimal homogeneous one. For those familiar with representation theory, a homogeneously embedded homogeneous variety weight, that is, the highest root of g has coefficient one on the simple root α i . Generalized cominuscule varieties are Segre-Veronese embeddings of products of cominuscule varieties. Grassmannians G(k, W ), projective spaces P n and products of projective spaces in any homogeneous embedding (in particular, respectively, G(k, W ) in the Plücker embedding, Veronese varieties, and Segre varieties) are generalized cominuscule.
Throughout this section we assume X is generalized cominuscule. When studying points of σ 3 (X), one has to take into account curves limiting to points on a trisecant line of X. When X is cut out by quadrics, as with homogeneous varieties, any trisecant line of X will be contained in X. Theorem 1.11 shows such points are already accounted for by curves with just one or two limit points, and that higher order differential invariants do not appear, as was hinted at in Lemma 2.2.
We commence the proof of Theorem 1.11 with an observation about the freedom of choice of splitting as in (3.1).
Lemma 3.1. Let X be generalized cominuscule and let x, y 1 , . . . , y r−1 be r points on X. Then there exists a choice of splitting as in (3.1) (so F s (S s T ) ⊂ N s for all s), such that x = o is the center of this splitting and none of the points y 1 , . . . , y r−1 lies on the hyperplane T ⊕N 2 ⊕N 3 ⊕· · · .
Proof. Let G be the automorphism group of X and P ⊂ G be the parabolic subgroup preserving It is also non-empty, because G acts on X transitively, and it is P -invariant. Since the Lie algebra of P contains all positive root spaces, andx is the highest weight space, the linex is contained in every P -invariant linear subspace of V (see, e.g., [7,Prop. 14.13]).
Fix H 0 ∈ Y x and consider the intersection B := p∈P p · H 0 . This is a linear subspace of V , which is invariant under P . So either B = 0 orx ⊂ B. The latter is however impossible, asx ∩ H 0 = 0 by our assumptions. So B = 0. The set of hyperplanes {p · H 0 ∈ PV * | p ∈ P } is non-empty, irreducible with trivial base locus, so its dimension is positive and by a trivial instance of Bertini's Theorem there exists at least one hyperplane H in this set that avoids all points y 1 , . . . , y r−1 .
Since there are only finitely many non-zero Fubini forms, the parameterization: , and a polynomial map v N : T → N . Then every point y ∈ X is either on the hyperplane P(T ⊕ N ), or is in the image of the parameterization φ.
Proof. We use the following elementary topological statement: Let P be a topological space, let I ⊂ U ⊂ P with I closed in U , and letĪ be the closure of I in P . ThenĪ ∩ U = I. To prove this, let J ⊂ P be a closed subset such that U ∩ J = I, which exists from the definition of subspace topology. ThenĪ ⊂ J, from the definition of the closure, and so We use the statement with P = PV , U the affine piece of PV , which is the complement of the hyperplane P(T ⊕ N ), and I = φ(T ). Note that φ(T ) is closed in U ≃ T ⊕ N , because it is the graph of v N (which is a polynomial map by our assumption). Moreover,Ī = X, and so X ∩ U = I, and X ⊂ I ∪ P(T ⊕ N ) as claimed.
This implies the following property of tangent spaces on X. Proof. Fix o := [ξ] ∈ ℓ. By Lemma 3.1 we may choose a splitting (3.1) such that [η] / ∈ T ⊕ N . Thus [η] is in the image of the parameterization by Remark 3.2. Consider a curve y(t) ∈X with y(0) = η. Note thatỹ(0) ∈ T is in the tangent direction to ℓ. Then in the splitting (3.1): Here (⋆) holds by the prolongation property (2.3), because II(ỹ(0) 2 ) = 0. Thus letting ν ′ be any non-zero vector in T ξ ℓ ⊂ T we have: This formula is independent of η, so we can vary η ∈ ℓ freely. Exchanging the roles of ξ, and η, we can also vary ξ. Thus, T ℓ is determined by the geometry of ℓ ⊂ X. But the group of automorphisms of X acts transitively on each irreducible component of the space parameterizing lines on X. When X = G/P with G simple, this is [11,Thm. 4.3] and [5]. (This is true for any minimally embedded homogeneous variety G/P I , with G simple, where I indexes the deleted simple roots, as long as I does not contain an "exposed short root" in the language of [11].) When X = Seg(v d 1 (G 1 /P 1 ) × · · · × v dn (G n /P n )) is generalized cominuscule (with each G i /P i cominuscule), the set of lines on X is the disjoint union of the variety of lines on each G i /P i such that d i = 1. Thus dim T ℓ must be constant over these irreducible components.
To see that Z(X) is a closed subset of PV , note Z(X) is the image of a projective space bundle over the variety parameterizing lines on X, whose fiber over ℓ ⊂ X is P(T ℓ ). Since dim T ℓ is locally constant by Lemma 3.3, this bundle is a projective variety, and thus Z(X) is an image of a projective variety, hence projective.
In the following lemma, we provide an uniform interpretation of the points of types (iii)-(iv).

Lemma 3.5. [p] is of type (iii) or (iv), if and only if
Moreover, for u ∈ V , the following conditions are equivalent: (1) u ∈ II; (2) There exist a curveṽ(t) ∈ T and an integer m, such that II(ṽ(t) 2 ) = t m u + t m+1 (. . . ); (3) There exist an integer m and vectorsṽ 0 ,ṽ 1 , . . . ,ṽ m ∈ T , such that Note that PII is the closure of the image of the rational map ii : PT PN given by Proof of Lemma 3.5. The equivalence of (1)-(3) is clear. In the notation of (3), a point p is of type (iii) if and only if it is of type (iii-iv) with m = 0, and it is of type (iv') if and only if it is of type (iii-iv) with m = 1. So suppose p is of type (iii-iv) with m > 1. Then it is in the closure of Z(X), the set of points of type (iv'). But Z(X) is closed by Lemma 3.4, so p is of type (iv).
Proof of Theorem 1.11. Suppose p ∈ σ 3 (X), so there exist ξ and y(t) := y 1 (t), z(t) := y 2 (t) as in Lemma 2.1. Write ξ =ō, and by Lemma 3.1 we may choose the splitting (3.1) such that for small values of t, we have y(t), z(t) ∈ T ⊕ N . So y(t) = (ō,ỹ(t), y N (t)) by Remark 3.2 and similarly for z(t). Consider the curvesỹ(t),z(t) ∈ T . Exchanging the roles of y and z if necessary, pick maximal integers k, l, with l ≥ k ≥ 0 and such that: for some holomorphic function λ(t) ∈ C and curvesṽ(t),w(t) ∈ T . From now on, we write y for y(t), etc. We adopt the convention l = ∞ ifw = 0.
If l = 0, then 0,ỹ 0 ,z 0 are three distinct and non-collinear points in T . This implies that p is on an honest 3-secant, and we are in case (i). So from now on suppose l > 0.
Our goal is to understand the leading term (in t) of Expanding out terms we obtain: First consider the case k ≥ 1, so that the three limit points coincide:ō = y 0 = z 0 . In this case, the terms in (3.4) with t (s−1)k+l are of order higher than l. By Lemma 2.2, the higher fundamental forms F s with s ≥ 3 will always have higher degree leading term than II. Thus: We conclude that any point p in the limiting space, which is spanned byō,ṽ 0 , and the leading term of t lw 0 + t 2k λ(λ − 1)II(ṽ 2 ) , is of the form (iii-iv). In the remainder of the argument assume k = 0 and we still assume l > 0. If λ 0 = 0, 1, the three limit points 0,ỹ 0 ,z 0 are distinct, but they lie on a line in T . Also suppose that II(ṽ 0 2 ) = 0. This means (e.g. by (2.2)) that the projective line from o in the direction ofṽ 0 is not contained in X. It follows thatō, y 0 , z 0 are linearly independent, because any line trisecant to X is entirely contained in X. This leads to case (i). Now say λ 0 = 0 or 1, and II(ṽ 0 2 ) = 0. If λ 0 = 0, thenō = z(0). If λ 0 = 1, then y(0) = z(0). Swapping the roles of x and y if necessary, we may assume λ 0 = 0 and write λ = t m λ m +t m+1 (. . . ) with m ≥ 1 and λ m = 0. Note alsoỹ =ṽ in this case (because k = 0). Then the leading term of (3.4) is the leading term of t lw 0 + f s=2 (λ s − λ)F s ỹ 0 s or it is of order at least l + 1. Therefore: + terms of higher order =ō ∧ y 0 ∧ λỹ 0 + t lw 0 + terms of higher order =ō ∧ y 0 ∧ λ m t mỹ 0 + t lw 0 + terms of higher order.
Note y 0 is linearly independent from T , because II(ỹ 0 2 ) = 0. We cannot have m = l and w 0 = −λ mỹ0 , because then the choice of l would not be maximal. Thus we have non-zero terms of degrees l or m, and the limiting space is spanned byō, y 0 and a tangent vector to o (which is a linear combination ofỹ 0 andw 0 ). Therefore we are in case (ii).
Finally, suppose II(ṽ 0 2 ) = 0 (so the line o, y(0) is contained in X). Hence (3.4) becomes: We claim that the summands with F s for s ≥ 3 are irrelevant to the leading term. First note for s ≥ 3 the fundamental form F s (ṽ s−1w ) vanishes at t = 0 by the prolongation property (2.3). So t l F s ṽ s−1w ) has order of vanishing at least l + 1, unless s = 2. Next we treat (1 + λ + · · · + λ s−2 )F s ṽ s By Lemma 2.2, for s ≥ 3 the leading term of F s (ṽ s ) is of higher order than that of II(ṽ s ). Thus the leading term of (3.4) can only come from the leading term of (3.5) t lw 0 + (λ 2 − λ)II ṽ 2 + 2λt l II ṽw 0 . Suppose µ is a holomorphic function in one variable, and m is the maximal integer such that λ − 1 = t m µ 2 for sufficiently small values of t. Note that µ has invertible values near t = 0. If m ≥ l, then only t lw 0 + 2λt l II(ṽw 0 ) contributes to the leading term of (3.4), and p is of type (iii-iv). Suppose m < l, and rewrite (3.5), up to terms of order > l: Thus there exists u ∈ II (either u = 0 or u is the leading coefficient of II (µṽ + t l−m µw 0 ) 2 up to scale, compare with Lemma 3.5(2)), such that the limiting space lim t→0 ō, y(t), z(t) is spanned by eitherō, y 0 , u orō, y 0 ,w 0 + u. Since y 0 ∈ô ⊕ T , in either case we have p = ξ ′ + u for some ξ ′ ∈ô ⊕ T , a linear combination ofō, y 0 andw 0 , and also after possible rescaling of u. That is, p is a point of type (iii-iv).
It remains to prove that any point p of the form (i), (ii), or (iii-iv) is in σ 3 (X). Case (i) is clear, case (ii) follows as σ 3 (X) = J(X, σ 2 (X)) ⊃ J(X, τ (X)) and points on tangent lines are handled by Proposition 2.3.

Examples
In the next sections we treat the case of Segre product with at least 3 factors in detail. Here we briefly review some other cases.

4.1.
Known results. We record the following known results: Example 4.1. Let X ⊂ PV be one of v 2 (P n ) (symmetric matrices of rank one), G(2, n) (skewsymmetric matrices of rank two), Seg(PA × PB) (matrices of rank one), or the Cayley plane OP 2 . Then any point on σ r (X) for any r is on an honest secant P r−1 .

Example 4.2. [12] Let
Normal forms for σ 3 (v d (PV ))\σ 2 (v d (PV )) of these types are respectively x d + y d + z d , x d−1 y + z d and x d−1 y + x d−2 z 2 , where x, y, z ∈ V . Thus the points of type (iv) do not occur in this case.

4.2.
Grassmannians in Plücker embedding. Let X := G(k, n) ⊂ P( k C n ), and suppose 3 ≤ k ≤ n − k and n − k > 3. The tangent space at E ∈ G(k, n) can be identified with the space of k × (n − k)-matrices k−1 E ⊗ F ≃ E * ⊗ F , where F = C n /E. The local parametrization in this case comes from a choice of splitting C n ≃ E ⊕ F and the determined splitting: The parametrization has the following form: In (iv'), ν = ǫ i+1 − ǫ i is a rank 1 matrix, so II(ν 2 ) = 0, and 2 ǫ i+1 − 2 ǫ i = 1 2 II(u, ǫ i ). In all normal forms, we can pick M to be in some normal form. For example, if i = k = n − k, then M may be (at least) assumed to be in Jordan normal form.
is the minimal homogeneous embedding. In this case the local parametrization is identical, but with T ≃ S 2 C k and M a symmetric k × k matrix, see [1, §5]. The normal forms are also identical.

Spinor varieties.
Let X be the spinor variety S k = D k /P k for k ≥ 7 in its minimal homogeneous embedding P( even C k ). In this case T ≃ 2 C k and M is a skew-symmetric k × k matrix, and the parameterization is similar to the previous cases: where Pf 2s M ∈ 2s C k , expressed in linear coordinates, is the collection of all 2s × 2s sub-Pfaffians of M . 5. The Segre product Seg(PA 1 × · · · × PA n )) Recall that for any smooth variety X, if x ∈ σ 2 (X), then either x ∈ X, x ∈ σ 0 2 (X) or x lies on an embedded tangent line to X, see Proposition 2.3.

5.1.
Proof of Proposition 1.1. All the assertions except for the rank of x in (1.1) are immediate. The rank of x is at most |J| because there are |J| terms in the summation.
Remark 5.1. The case n = 3 was previously established by Grigoriev, Ja'Ja' and Teichert.

5.2.
Parameterization in the Segre case. Suppose X = Seg(PA 1 × · · · × PA n ). Letō = a 1 1 ⊗ · · · ⊗ a n 1 , and let A ′ j = a 1 1 ⊗ · · · ⊗ a j−1 Thus i for some i and if further II(ṽ 0ṽ1 ) = 0 thenṽ 1 ∈ A ′ i for the same i. In particular, if a line ℓ ⊂ X contains o and is tangent toṽ 0 , then by (3.2) we have: Now we prove Theorem 1.2. The normal forms follow from the discussion in the previous sections. Now suppose dim A i ≥ 3. To see that the general points of each type do not belong to the other types, note that for any type and for any i, in the normal forms (1.2)-(1.5) either a i 1 , a i 2 , a i 3 are linearly independent, or the point is contained in a subspace variety, i.e., a closed subvariety consisting of tensors in some A 1 ⊗ · · · ⊗ A i−1 ⊗ C 2 ⊗ A i+1 ⊗ · · · ⊗ A n . Thus the general points of each type form a single orbit (or n orbits for type (iv)) of the action of GL(A 1 ) × · · · × GL(A n ). Therefore the only possible way that they could overlap, is if one of the orbits were equal to the other. But the orbits are distinct by the dimension count below, which we present in two different forms.

5.3.
First proof of dimensions in Theorem 1.2. We compute the Lie algebras of the stabilizers of each type of point. Without loss of generality (for computing codimension), assume dim A j = 3. Write Γ = (x 1 , . . . , x n ) where x α = (x i j,α ), 1 ≤ i, j ≤ 3. We calculate the Γ such that Γ.p ( * ) = 0 in each case * = i, ii, iii, iv and denote this algebra by g p ( * ) . In each case one has a system of 3 n = dim(A 1 ⊗ · · · ⊗ A n ) linear equations, many of which are zero or redundant.
We show the assertion about the codimensions of types (ii),(iii),(iv). Type (ii) is immediate as its closure is J(X, τ (X)) which is easily seen to have the expected dimension via Terracini's lemma.
We will use the following lemma: Proof. Any such R is isomorphic either to 3 distinct reduced points, or a double point and a reduced point, or one of the two kinds of triple points: Spec C[x]/x 3 , or Spec C[x, y]/ x 2 , xy, y 2 . If n = 2, without loss of generality, we may suppose dim A 1 = dim A 2 = 3. We can write down explicitly R ⊂ P(A 1 ⊗ A 2 ) for each of the schemes as, respectively:  The claim may be verified explicitly for each case, by calculating the scheme defined by 2 × 2 minors of each of the matrices. If n ≥ 3, let B i = A 1 ⊗ · · · ⊗ A i−1 ⊗ A i+1 ⊗ · · · ⊗ A n . Then X = n i=1 PA i × PB i , and the claim easily follows from the n = 2 statement. Lemma 5.3. Suppose n ≥ 2 and dim A i ≥ 3 for all i ∈ {1, . . . , n}. Let X = Seg(PA 1 ×· · ·×PA n ) and let ℓ ⊂ X be a line spanned by x, y ∈ X. Let v ∈T x X +T y X be general and consider P 2 spanned by ℓ and [v]. Then P 2 ∩ X = ℓ.
Consider a linear combination sv + tx + uy. The intersection P 2 ∩ X is contained in the zero locus of the 2 × 2 minors of the following matrix: which can be identified with the line s = 0, that is the line spanned by x and y.
Let Osc(X) be the closure of the set of points of type (iii). Let [p] ∈ Osc(X) be a general point. We claim such p uniquely determines [x] such that p = x + x ′ + x ′′ . Suppose without loss of generality dim A 1 = 3. Write p = p (iii) of (1.4), and consider the underlying map p (iii) : A 1 * → A 2 ⊗ · · · ⊗ A n : The projectivization of the image is a P 2 containing a degree 3 scheme R ⊂ Seg(PA 2 ×· · ·×PA n ) in general position, which is isomorphic to the triple point Spec C[x]/x 3 point supported at [p(a 1 3 * )]. By Lemma 5.2, R is determined by R = P(p(A 1 * )), so it is independent of the choice of normal form. Therefore a 1 2 * , a 1 3 * , which is the linear span of the unique degree 2 subscheme of R, is determined by p, and so is a 1 1 (up to scale). Similarly, a i 1 are determined by p up to scale.
Finally consider Z(X), the set of points of type (iv), which is closed by Lemma 3.4. Let [p] ∈ Z(X) be a general point of any of the irreducible components. We claim p uniquely determines the line P x, y such that p = x + x ′ + y + y ′ . Suppose without loss of generality dim A i = 3 for all i. Possibly permuting the factors, write p = p (iv) of (1.5). First consider the underlying map p (iv) : A 1 * :→ A 2 ⊗ · · · ⊗ A n : a 2 1 ⊗ · · · ⊗ a i−1 1 ⊗a i 2 ⊗a i+1 1 ⊗ · · · ⊗ a n 1 , p (iv) (a 1 3 * ) = a 2 1 ⊗ · · · ⊗ a n 1 .
By Lemma 5.3 the projective line P p (iv) (a n 2 * ), p (iv) (a n 3 * ) is determined by p (iv) . Thus a n 1 (and similarly a i 1 for i ∈ {2, . . . , n}) is determined (up to scale) by p (iv) . Therefore, the line P( a 1 1 , a 1 2 ⊗ a 2 1 ⊗ · · · ⊗ a n 1 ) ⊂ X is uniquely determined by p (iv) . The lines on X are parametrized by n irreducible varieties: By the argument above we have a rational dominant map χ : Z(X) L 1 ⊔ · · · ⊔ L n . A general fiber over ℓ ∈ L i is PT ℓ in the notation of Lemma 3.3, the linear span of projective tangent spaces to X at points of ℓ. By (5.1) dim T ℓ = 2dim X + 2 − dim A i , and the dimension of each irreducible component of Z(X) is equal to 3 (dim A i − 1) − 2.

Orbits of tensors in
This subspace variety admits a desingularization as follows. A) is the tautological rank a ′ subspace bundle and similarly for B, C. Then PE → Sub a ′ ,b ′ ,c ′ (A⊗B⊗C) is a desingularization and using it one can see that Sub a ′ ,b ′ ,c ′ (A⊗B⊗C) sing = Sub a ′ −1,b ′ ,c ′ ∪ Sub a ′ ,b ′ −1,c ′ ∪ Sub a ′ ,b ′ ,c ′ −1 , whenever a ′ < bc, and similarly for permuted statements. In [4, §6], normal forms for tensors in Sub 233 ∪ Sub 323 ∪ Sub 332 are given. There are 33 such.
We present the list of remaining orbits in σ 3 (Seg(PA×PB ×PC)) under the action of GL(A)× GL(B) × GL(C).
Each orbit is uniquely determined by its closure, which is an algebraic variety listed in the second column of the table. The orbit itself is an open dense subset of this variety. The dimension of the algebraic variety is in the third column. The fourth column is the normal form of the underlying tensor, the distinct variables are assumed to be linearly independent. The normal form is also given as a slice. The border rank and rank are given in the next columns.
Z(X) A , Z(X) B , Z(X) C denote the three components of Z(X), the set points of type (iv) in Theorem 1.2. Osc(X) denotes the the closure of the set points of type (iii), while J(X, τ (X)) denotes the the closure of the set points of type (ii).
The ranks of cases 34-37 in Table 1 are calculated in §6.1. The rank of case 39 is obvious, while the rank of case 38 is at most 4, due to the normal form expression. If it were 3, then a general point of type (ii), would be expressible as a point of type (i), a contradiction with Theorem 1.2.
Osc(X) 3a + 3b + 3c − 9 a 1 ⊗(b 1 ⊗c 3 + b 2 ⊗c 2 + b 3 ⊗c 1 ) +a 2 ⊗(b 1 ⊗c 2 + b 2 c 1 ) + a 3 ⊗b 1 ⊗c 1 u t s t s s 3 5 38 J(X, τ (X)) 3a + 3b + 3c − 8 a 1 ⊗(b 1 ⊗c 2 + b 2 ⊗c 1 ) +a 2 ⊗b 1 ⊗c 1 + a 3 ⊗b 3 ⊗c 3 t s s u 3 4 39 σ 3 (X) 3a + 3b + 3c − 7 a 1 ⊗b 1 ⊗c 1 + a 2 ⊗b 2 ⊗c 2 + a 3 ⊗b 3 ⊗c 3 s t u 3 3  To see the ranks are at least five, were it four in the first case, we would be able to find a  of rank 1, such that the 4-plane spanned by: is spanned by matrices of rank 1. In particular, T 1 would be in the span of T 2 , T 3 , T and another matrix of rank 1. Thus we would be able to find constants β, γ, f 1 , f 2 , f 3 , g 1 , g 2 , g 3 , such that the rank of   γ β 1 β 1 0 is one. There are two cases: if g 3 = 0, then we can subtract g 1 g 3 times the third column from the first, and g 2 g 3 times the third column from the second to obtain  which has rank at least two. If g 3 = 0 the matrix already visibly has rank at least two. Thus it is impossible to find such constants β, γ, f i , g i and the rank in question is necessarily at least 5.
The second case is more delicate. Write all 2 × 2 minors of   t s u s 0 0 and consider f i and g j as parameters of degree 0, and remaining variables α 1 , α 2 , α 3 , x of degree 1. We claim (sf 3 − uf 2 ) 2 and (sg 3 − ug 2 ) 2 are in the ideal I generated by minors. This can be verified by patient calculation, or using a computer algebra system, such as Magma [2]. Thus f 2 = f 3 = g 2 = g 3 = 0, for otherwise we have a degree 1 equation in the radical ideal √ I, and then the rank 1 matrices do not span the four dimensional linear space. But in such a case u 2 and s 2 are among the minors, giving u and s as linear equations in √ I, a contradiction.

6.2.
Singularities. In this subsection we prove Theorems 1.3 and 1.4. The strategy is uniform to most cases: using the desingularization PE → Sub i,j,k as in the beginning paragraph of §6, which is birational away from the locus Sub i−1,j,k ∪ Sub i,j−1,k ∪ Sub i,j,k−1 , we reduce statements to properties of secant varieties of low dimensional Segre products.
To argue for the other components, i.e., when T is of the form 34 or 36 in Table 1, one can permute the factors A, B, and C.
Proof of Theorem 1.4. Assume dim A, dim B, dim C ≥ 3. Since the map P(E) → Sub 3,3,3 is an isomorphism near a general point of type (iv), the Lemma implies that such a point is a smooth point of σ 3 (Seg(PA × PB × PC)) for any A, B, C (each of dimension at least 3). But orbits 34-36 from Table 1 are in the closure of orbits 37 and 38. So σ 3 (Seg(PA × PB × PC)) is non-singular at a general point of each type (ii)-(iv).
The final thing to prove is that σ 3 (Seg(PA × PB × PC)) is non-singular at a general point of Sub 233 . Let p be such a point. Since σ 3 (Seg(PA × PB × PC)) ⊂ Sub 333 , we may assume dim A = dim B = dim C = 3. First note that Sub 233 is not contained in J(X, τ (X)), as they are both irreducible, have the same dimension and J(X, τ (X)) ⊂ Sub 233 . So p is not in J(X, τ (X)). By Theorem 1. Let W := φ −1 (U ). Then φ| W : W → U is a regular surjective map. The aim is to calculate the tangent map at any point in φ −1 (p). We commence with identifying φ −1 (p). Since R X (p) = 3, any point in φ −1 (p) will be contained in a fixed (A ′ × B ′ × C ′ ) ×3 with dim A ′ = 2, dim B ′ = dim C ′ = 3 by [4, Cor. 2.2]. Write p = [a 1 ⊗b 1 ⊗c 1 +(a 1 +a 2 )⊗b 2 ⊗c 2 +a 2 ⊗b 3 ⊗c 3 ] (see [4, §6]). We claim that this normal form is unique up to trivialities such as 7-dimensions worth of rescalings, and permutations of summands. By writing p : (A ′ ) * → B ′ ⊗ C ′ , we obtain the slice s s+t t . The set of rank 2 elements in this linear space is given by the determinant of the matrix. This set consists of three lines in (A ′ ) * spanned by a * 1 , a * 1 − a * 2 , and a * 2 . Thus the triple a 1 , (a 1 + a 2 ), a 2 is (up to order and scale) determined by p. In a similar way we consider the other slices, and 2 × 2 minors of the resulting matrices, to conclude, that triples b 1 , b 2 , b 3 and c 1 , c 2 , c 3 are determined by p, up to order and scale. It is easy to see, that any meaningfully different choice of orders, or scaling will give a different tensor, so the preimage of p consists of 6 components, each of dimension 7, isomorphic to (C * ) 7 .
3c + 11 3c + 9 C 3 ⊗ C b ⊗ C c 3b + 3c + 2 3b + 3c C a ⊗ C b ⊗ C c 3a + 3b + 3c − 7 2a + 3b + 3c − 6 Table 2. Singularities of σ 3 (Seg(PA × PB × PC)). In the first column we list the tensor space, assuming 4 ≤ a ≤ b ≤ c. In the second column we write the dimension of the secant variety. In the third column we present the upper bound on the dimension of the singular locus of the secant variety, which follows from our results in this section.
This space is independent of the choice of the order or scalings in q. Also the linear space above has dimension 3(dim A + dim B + dim C) − 7 = dim σ 3 (PA × PB × PC), because there are 3(dim A + dim B + dim C) tensors listed above, and each a 1 ⊗ b 1 ⊗ c 1 , (a 1 + a 2 ) ⊗ b 2 ⊗ c 2 , a 2 ⊗ b 3 ⊗ c 3 is listed three times and p is a sum of those three tensors. One can check there no other linear dependencies.
Thus φ : W → P(A ⊗ B ⊗ C) is a map with constant rank on an open subset containing φ −1 (p). Therefore the image is non-singular at p as claimed.
We summarize our results in Table 2. In particular, it follows that σ 3 (Seg(PA × PB × PC)) is always non-singular in codimension 1, that is, codimension of the singular locus is at least 2. Moreover, it is of codimension 2 if and only if, one of the factors is C 3 , and the others have dimension at least 3.