PaperlayerView exact revision on arXiv

Source-linked AI summary

On generic identifiability of 3-tensors of small rank

Luca Chiantini, Giorgio Ottaviani

arXiv:1103.2696v1math.AG

TL;DR

The paper addresses how to determine when tensor decompositions are unique for general tensors. It introduces an inductive geometric method based on weak defectivity and proves broad identifiability bounds, while also identifying low-dimensional failures such as rank-6 4×4×4 tensors.

  • Problem

    The paper studies the limited known range of generic uniqueness for decompositions of tensors into rank-1 tensors.

  • Method

    The paper combines induction on Segre varieties with Terracini’s Lemma and weak defectivity to establish identifiability conditions.

  • Results

    For a ≤ b ≤ c, general rank-k tensors are uniquely decomposable when k ≤ 2^α+β−2; additionally, general 4×4×4 rank-6 tensors have exactly two decompositions.

  • Takeaways & Limitations

    The method substantially expands the known generic identifiability range while exposing specific small-format exceptions relevant to DNA-string models.

  • Takeaways & Limitations

    The inductive criterion cannot be transferred directly to lower-dimensional spaces, and identifiability is impossible above a maximal rank.

Abstract

from arXiv · show

We introduce an inductive method for the study of the uniqueness of decompositions of tensors, by means of tensors of rank 1. The method is based on the geometric notion of weak defectivity. For three-dimensional tensors of type (a, b, c), a\le b\le c, our method proves that the decomposition is unique (i.e. k-identifiability holds) for general tensors of rank k, as soon as k\le (a+1)(b+1)/16. This improves considerably the known range for identifiability. The method applies also to tensor of higher dimension. For tensors of small size, we give a complete list of situations where identifiability does not hold. Among them, there are 4\times4\times4 tensors of rank 6, an interesting case because of its connection with the study of DNA strings.

1. Introduction

The paper develops a geometric, inductive method for generic tensor identifiability and proves substantially broader uniqueness ranges. It also characterizes notable low-dimensional exceptions, including rank-6 tensors of format 4×4×4.

  • Main result: For a ≤ b ≤ c, general tensors of rank k have unique decompositions when k ≤ 2^α+β−2, where 2^α ≤ a and 2^β ≤ b are maximal.This is the paper’s main general bound for three-dimensional tensors.
  • Main result: When a and b are powers of 2, the uniqueness range becomes k ≤ ab/4.The theorem yields this specialized bound, with a better bound possible when a is near a power of 2.
  • Cubic tensors: In the cubic case, general tensors have unique decompositions through a quadratic rank range, with the theorem providing a stronger bound near powers of 2.The cited passage states the cubic specialization as k ≤ a^2/16 and notes an improved bound in nearby cases.
  • Limits and refinements: The bound is log-asymptotically sharp, although it is not sharp for all small tensor formats.The paper notes both an upper obstruction beyond k_max and computer-assisted improvements in initial small cases.
  • Low-dimensional exceptions: Generic identifiability fails for general 4×4×4 tensors of rank 6, which have exactly two decompositions and connect to phylogenetic DNA models.The 4-dimensional factors can be indexed by the nucleotides A, C, G, and T.
  • Method: The method combines an inductive approach with weak defectivity to study uniqueness through tangent spaces of Segre varieties.Terracini’s Lemma connects non-weak defectivity with unique decompositions, while induction splits one vector space into lower-dimensional summands.

2. Preliminaries on Segre varieties

The section models tensors through Segre varieties and connects unique decompositions to tangent-space criteria for k-identifiability. It also distinguishes tangential weak defectivity from weak defectivity and records computational and dimensional boundaries of the criterion.

  • X=P(A)×P(B)×P(C) is embedded by the Segre map into P^N, with N=abc−1.
  • A general rank-k tensor is k-identifiable when it has a unique decomposition into k decomposable tensors, up to scalar multiplication.
  • Terracini’s Lemma identifies the tangent space to the k-th secant variety with the span of tangent spaces at the decomposition points.
  • The criterion requires that the span of tangent spaces at k general points contain no additional tangent space, and this implies k-identifiability when N≥k(n+1).
  • Tangential weak defectivity is weaker than the relevant tangent-space condition: failure of the condition implies weak defectivity, but the converse need not hold.
  • The criterion has a dimensional boundary: if N+1<(k+1)n, k-identifiability is excluded, while computations establish initial cases through a≤7.

3. The inductive statement

The inductive statement strengthens weak-defectivity conditions by tracking auxiliary subspaces and reduces higher-dimensional identifiability checks to lower-dimensional ones. Its inductive step combines certified conditions on complementary factors, while computations and specialization determine practical scope.

  • For a Segre point x=u⊗v⊗w, its tangent space is the projectivization of A⊗v⊗w+u⊗B⊗w+u⊗v⊗C.
  • If complementary products X′ and X′′ satisfy the appropriate not-weakly-defective conditions, Proposition 3.6 combines them into a condition on X with k=k1+k2 and summed auxiliary parameters.
  • The inductive procedure can stop because specialized points make a condition fail, which does not by itself prove that the original tensor is non-identifiable.
  • The (k,p,q,r)-not weakly defective condition tests whether tangent spaces at k points, together with auxiliary spaces A⊗u_i, B⊗v_i, and C⊗w_i, contain any other tangent space.
  • When p=q=r=0, this generalized condition coincides with k-tangential weak defectivity, and its failure implies k-identifiability.
  • The reduction lemma transfers a (k,p,q,r)-not weakly defective condition from P^(a−1)×P^(b−1)×P^(c−1) to larger factor dimensions.
  • Macaulay2 computations verify reduced cases, including P^2×P^2×P^2 examples that are not weakly defective despite being 3-defective.

4. Proof of Theorem 1.1

The proof establishes identifiability by recursively splitting powers-of-two factor dimensions and applying non-weak-defectivity criteria. It yields k-identifiability bounds that are log-asymptotically sharp and improve Kruskal’s bound.

  • Inductive base: P1 × P1 × P1 provides the induction base: it is both (1, 0, 0, 0)-not and (0, 1, 1, 1)-not weakly defective.These base cases initiate the recursive argument for larger powers-of-two dimensions.
  • Recursive reduction: Powers of 2 enable recursive reductions that split one vector space into two equal subspaces while balancing point and auxiliary-space counts.The reduction repeatedly lowers a factor dimension and transforms the weak-defectivity parameters.
  • Main bound: For ordered dimensions 1 ≤ α ≤ β ≤ γ, P2^α−1 × P2^β−1 × P2^γ−1 is (k, 0, 0, 0)-not weakly defective for k ≤ 2^(α+β−2).The balanced case gives k-identifiability for k ≤ 2^(2α−2).
  • Main bound: For general dimensions a ≤ b ≤ c, choosing maximal α and β with 2^α ≤ a and 2^β ≤ b gives k-identifiability for k ≤ 2^α2^β/4.The argument transfers non-weak-defectivity from powers-of-two subspaces to the original factors.
  • Comparison: The resulting bound is at least log-asymptotically sharp and improves Kruskal’s identifiability bound.The sharpness comparison uses ab/3 ≤ kmax ≤ ab.
  • Variants: Replacing powers of 2 by powers of another integer p can strengthen the bound in specific cases.For P^26 × P^26 × P^26, the bound rises from 64 using powers of 2 to 81 using p = 3, while kmax = 249.

5. Some examples in low dimension

The low-dimensional analysis identifies both positive and negative cases for unique tensor decomposition. In particular, general 4 × 4 × 4 tensors of rank 6 have exactly two decompositions, while unbalanced cases admit theoretical uniqueness results.

  • DNA-tensor example: A general 4 × 4 × 4 tensor of rank 6 has exactly two decompositions into six rank-1 tensors, up to scalar multiplication and permutations.Geometrically, a general point lies in exactly two 6-secant 5-planes.
  • Geometric explanation: The rank-6 phenomenon is connected to elliptic normal curves of degree 12 passing through six general points of P3 × P3 × P3.Such a curve has 6-secant order 2, producing the two decompositions.
  • Geometric explanation: The elliptic-curve argument initially includes a probabilistic step, so the authors provide a theoretical parameter-count argument yielding finitely many curves through six general points.The parameter count gives 34 algebraic conditions on 34 parameters.
  • Unbalanced products: In unbalanced cases with c ≥ (a−1)(b−1), a general tensor of rank (a−1)(b−1) has a unique decomposition.The proof uses a flattening whose projectivized image meets a lower-dimensional Segre variety in exactly the decomposition points.
  • Unbalanced products: When c ≥ (a−1)(b−1)+2, the paper gives a general uniqueness theorem for rank-k tensors, while rank (a−1)(b−1)+1 can have multiple decompositions.The number of decompositions is always greater than one except when a = b = 2.

6. Products with many factors

The paper extends its weak-defectivity method from three factors to products of many projective spaces. An inductive splitting criterion and powers-of-two analysis yield general identifiability bounds that are log-asymptotically sharp.

  • Generalization: The weak-defectivity framework and its main theorem extend to products of many vector spaces.The authors note that the generalization increases notation but preserves the method’s scope.
  • Definitions and criterion: A product is (k, p1, ..., pn)-not weakly defective when the prescribed tangent-space span contains TxX only at the selected points.The definition introduces auxiliary points on factor-deleted products.
  • Inductive method: The inductive step splits one factor Ai into two subspaces and combines non-weak-defectivity statements for the resulting products.If the two pieces satisfy the required parameter conditions, the original product is (k1 + k2, p1, ..., pn)-not weakly defective.
  • Powers-of-two analysis: For dimensions dim(Ai) = 2^αi, Lemma 6.5 supplies non-weak-defectivity under the constraints ui ≤ α1 + ... + α̂i + ... + αn − (n−1).It yields both (0, 2u1, ..., 2un)- and (1, 2u1−1, 2u2−1, 2u3−1)-non-weak-defectivity.
  • Main result: For ordered αi, the product is not k-weakly defective for k ≤ 2^(α1+...+αn−1) − (n−1).This is the many-factor numerical bound derived from the inductive criteria.
  • Main result: Theorem 6.7 converts the many-factor non-weak-defectivity result into k-identifiability for general tensors with arbitrary factor dimensions.The theorem chooses maximal powers of two not exceeding each dimension.
  • Sharpness and comparison: The many-factor bound is log-asymptotically sharp, and for a ≥ 4 it also admits a weaker but more convenient inequality.The paper compares the result with the maximal meaningful identifiability rank and with Kruskal-type bounds.
Loading 1103.2696v1…