Homotopy techniques for tensor decomposition and perfect identifiability

DOI10.1515/CRELLE-2016-0067MaRDI QIDQ2317007zbMATH OpenOpenAlexWikidataFDO

Authors Jonathan D. Hauenstein, Luke Oeding, Giorgio Ottaviani, Andrew J. Sommese

Publication date 7 August 2019

Published in Journal für die Reine und Angewandte Mathematik (Search for Journal in Brave)

Full work available at URL https://arxiv.org/abs/1501.00090

zbMATH Keywords

Koszul flattenings numerical algebraic geometry tensor decomposition

Mathematics Subject Classification ID

Computational aspects of higher-dimensional varieties (14Q15) Multilinear algebra, tensor calculus (15A69)

Abstract: Let T be a general complex tensor of format

(n_{1}, ..., n_{d})

. When the fraction

p r o d_{i} n_{i} / [1 + s u m_{i} (n_{i} - 1)]

is an integer, and a natural inequality (called balancedness) is satisfied, it is expected that T has finitely many minimal decomposition as a sum of decomposable tensors. We show how homotopy techniques allow us to find all the decompositions of T, starting from a given one. Computationally, this gives a guess regarding the total number of such decompositions. This guess matches exactly with all cases previously known, and predicts several unknown cases. Some surprising experiments yielded two new cases of generic identifiability: formats (3,4,5) and (2,2,2,3) which have a unique decomposition as the sum of 6 and 4 decomposable tensors, respectively. We conjecture that these two cases together with the classically known matrix pencils are the only cases where generic identifiability holds, i.e., the only identifiable cases. Building on the computational experiments, we use algebraic geometry to prove these two new cases are indeed generically identifiable.

Review text

An element of the complex vector space \(\mathbb{C}^{n_1} \otimes \dots \otimes \mathbb{C}^{n_d}\) is a tensor of format \((n_1,\dots,n_d)\). The rank of such a tensor \[ T \in \mathbb{C}^{n_1}\otimes \dots \otimes \mathbb{C}^{n_d} \] is defined to be the smallest integer \(r\) such that \[ T = \sum_{i = 1}^r v_1^i \otimes \dots \otimes v_d^i, \] where \(v_j^i \in \mathbb{C}^{n_j}\). Such an expression is called a decomposition of \(T\) into a sum of decomposable summands. A tensor of format \((n_1,\dots,n_d)\) is called perfect if the quantity \[ R(n_1,\dots,n_d) := \frac{ \prod_{i=1}^d n_i }{ 1 - d + \sum_{i=1}^d n_i } \] is an integer. The condition that a general tensor of format \((n_1,\dots,n_d)\) is perfect is necessary for it to admit finitely many decompositions into a sum of decomposable summands. Within this context, the authors study the problem of uniqueness for decompositions of perfect tensor formats. Their main results pertain to general tensors of formats \((3,4,5)\) and \((2,2,2,3)\). Specifically, for the case of general tensors of format \((3,4,5)\), the authors prove that each such general tensor admits a unique decomposition into a sum of six decomposable summands. For the case of general tensors of format \((2,2,2,3)\), they prove that such tensors admit a unique decomposition into a sum of four decomposable summands. An interesting aspect of the proof of these results is the technique of Koszul flattenings developed in [\textit{J. M. Landsberg} and \textit{G. Ottaviani}, Ann. Mat. Pura Appl. (4) 192, No. 4, 569--606 (2013; Zbl 1274.14058)]. Another aspect is the use of monodromy loops (see [\textit{A. J. Sommese} et al., NATO Sci. Ser. II, Math. Phys. Chem. 36, 297--315 (2001; Zbl 0990.65051)]). Finally, the authors formulate a conjecture about the perfect format of general tensors admitting a unique decomposition into decomposable summands.

Recommendations

Cites work

Cited in

(29)

Describes a project that uses

Uses Software

This page was built for publication: Homotopy techniques for tensor decomposition and perfect identifiability

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2317007)