Tensor rank decomposition
In multilinear algebra, the tensor rank decomposition or canonical polyadic decomposition (CPD) is one generalization of the matrix singular value decomposition (SVD) to tensors, which have found application in statistics, signal processing, computer vision, computer graphics, psychometrics, linguistics and chemometrics. The tensor rank decomposition was introduced by Hitchcock in 1927 and later rediscovered several times, notably in psychometrics. For this reason, the tensor rank decomposition is often referred to as CANDECOMP, PARAFAC, or CANDECOMP/PARAFAC (CP).
Another popular generalization of the matrix SVD is known as the higher-order singular value decomposition.
A scalar variable is denoted by lower case italic letters, and a constant scalar is denoted by an upper case italic letter, .
Indices are denoted by a combination of lowercase and upper case italic letters, . Multiple indices that one might encounter when referring to the multiple modes of a tensor are conveniently denoted by where .
A vector is denoted by a lower case bold Times Roman, and a matrix is denoted by bold upper case lettering .
A higher order tensor is denoted by calligraphic letters,. An element of an -order tensor is denoted by or .
A tensor is a multilinear transformation that maps a set of vector spaces to another vector space. A data tensor is a collection of multivariate observations organized into a M-way array.
Consider a data tensor , where is either the real field or the complex field . Every (order-, refers to the number of modes) tensor in this space may then be represented with a suitably large as a linear combination of rank-1 tensors:
where and where . When the number of terms is minimal in the above expression, then is called the rank of the tensor, and the decomposition is often referred to as a (tensor) rank decomposition, minimal CP decomposition, or Canonical Polyadic Decomposition (CPD). Contrariwise, if the number of terms is not minimal, then the above decomposition is often referred to as -term decomposition, CANDECOMP/PARAFAC or Polyadic decomposition.
Contrary to the case of matrices, the rank of a tensor is presently not understood well. It is known that the problem of computing the rank of a tensor is NP-hard. The only notable well-understood case consists of tensors in , whose rank can be obtained from the Kronecker–Weierstrass normal form of the linear matrix pencil that the tensor represents. A simple polynomial-time algorithm exists for certifying that a tensor is of rank 1, namely the higher-order singular value decomposition.
The rank of the tensor of zeros is zero by convention. The rank of a tensor is one, provided that .
The rank of a tensor depends on the field over which the tensor is decomposed. It is known that some real tensors may admit a complex decomposition whose rank is strictly less than the rank of a real decomposition of the same tensor. As an example, consider the following real tensor
where . The rank of this tensor over the reals is known to be 3, while its complex rank is only 2 because it is the sum of a complex rank-1 tensor with its complex conjugate, namely
In contrast, the rank of real matrices will never decrease under a field extension to : real matrix rank and complex matrix rank coincide for real matrices.
The generic rank is defined as the least rank such that the closure in the Zariski topology of the set of tensors of rank at most is the entire space . In the case of complex tensors, tensors of rank at most form a dense set : every tensor in the aforementioned space is either of rank less than the generic rank, or it is the limit in the Euclidean topology of a sequence of tensors from . In the case of real tensors, the set of tensors of rank at most only forms an open set of positive measure in the Euclidean topology. There may exist Euclidean-open sets of tensors of rank strictly higher than the generic rank. All ranks appearing on open sets in the Euclidean topology are called typical ranks. The smallest typical rank is called the generic rank; this definition applies to both complex and real tensors. The generic rank of tensor spaces was initially studied in 1983 by Volker Strassen.
As an illustration of the above concepts, it is known that both 2 and 3 are typical ranks of while the generic rank of is 2. Practically, this means that a randomly sampled real tensor (from a continuous probability measure on the space of tensors) of size will be a rank-1 tensor with probability zero, a rank-2 tensor with positive probability, and rank-3 with positive probability. On the other hand, a randomly sampled complex tensor of the same size will be a rank-1 tensor with probability zero, a rank-2 tensor with probability one, and a rank-3 tensor with probability zero. It is even known that the generic rank-3 real tensor in will be of complex rank equal to 2.
The generic rank of tensor spaces depends on the distinction between balanced and unbalanced tensor spaces. A tensor space , where , is called unbalanced whenever
and it is called balanced otherwise.
Unbalanced tensor spaces
When the first factor is very large with respect to the other factors in the tensor product, then the tensor space essentially behaves as a matrix space. The generic rank of tensors living in an unbalanced tensor spaces is known to equal
Balanced tensor spaces
The generic rank of tensors living in a balanced tensor space is expected to equal
almost everywhere for complex tensors and on a Euclidean-open set for real tensors, where
More precisely, the rank of every tensor in , where is some indeterminate closed set in the Zariski topology, is expected to equal the above value. For real tensors, is the least rank that is expected to occur on a set of positive Euclidean measure. The value is often referred to as the expected generic rank of the tensor space because it is only conjecturally correct. It is known that the true generic rank always satisfies
The Abo–Ottaviani–Peterson conjecture states that equality is expected, i.e., , with the following exceptional cases:
In each of these exceptional cases, the generic rank is known to be . Note that while the set of tensors of rank 3 in is defective (13 and not the expected 14), the generic rank in that space is still the expected one, 4.
The AOP conjecture has been proved completely in a number of special cases. Lickteig showed already in 1985 that , provided that . In 2011, a major breakthrough was established by Catalisano, Geramita, and Gimigliano who proved that the expected dimension of the set of rank tensors of format is the expected one except for rank 3 tensors in the 4 factor case, yet the expected rank in that case is still 4. As a consequence, for all binary tensors.
The maximum rank that can be admitted by any of the tensors in a tensor space is unknown in general; even a conjecture about this maximum rank is missing. Presently, the best general upper bound states that the maximum rank of , where , satisfies
where is the (least) generic rank of . It is well-known that the foregoing inequality may be strict. For instance, the generic rank of tensors in is two, so that the above bound yields , while it is known that the maximum rank equals 3.
A rank- tensor is called a border tensor if there exists a sequence of tensors of rank at most whose limit is . If is the least value for which such a convergent sequence exists, then it is called the border rank of . For order-2 tensors, i.e., matrices, rank and border rank always coincide, however, for tensors of order they may differ. Border tensors were first studied in the context of fast approximate matrix multiplication algorithms by Bini, Lotti, and Romani in 1980.
A classic example of a border tensor is the rank-3 tensor
It can be approximated arbitrarily well by the following sequence of rank-2 tensors
as . Therefore, its border rank is 2, which is strictly less than its rank. When the two vectors are orthogonal, this example is also known as a W state.
It follows from the definition of a pure tensor that if and only if there exist such that and for all m. For this reason, the parameters of a rank-1 tensor are called identifiable or essentially unique. A rank- tensor is called identifiable if every of its tensor rank decompositions is the sum of the same set of distinct tensors where the 's are of rank 1. An identifiable rank- thus has only one essentially unique decomposition
Order-2 tensors in , i.e., matrices, are not identifiable for . This follows essentially from the observation
The situation changes completely for higher-order tensors in with and all . For simplicity in notation, assume without loss of generality that the factors are ordered such that . Let denote the set of tensors of rank bounded by . Then, the following statement was proved to be correct using a computer-assisted proof for all spaces of dimension , and it is conjectured to be valid in general:
There exists a closed set in the Zariski topology such that every tensor is identifiable ( is called generically identifiable in this case), unless either one of the following exceptional cases holds:
- The rank is too large: ;
- The space is identifiability-unbalanced, i.e., , and the rank is too large: ;
- The space is the defective case and the rank is ;
- The space is the defective case , where , and the rank is ;
- The space is and the rank is ;
- The space is and the rank is ; or
- The space is and the rank is .
- The space is perfect, i.e., is an integer, and the rank is .
In these exceptional cases, the generic (and also minimum) number of complex decompositions is
- proved to be in the first 4 cases;
- proved to be two in case 5;
- expected to be six in case 6;
- proved to be two in case 7; and
- expected to be at least two in case 8 with exception of the two identifiable cases and .
In summary, the generic tensor of order and rank that is not identifiability-unbalanced is expected to be identifiable (modulo the exceptional cases in small spaces).
Ill-posedness of the standard approximation problem
The rank approximation problem asks for the rank- decomposition closest (in the usual Euclidean topology) to some rank- tensor , where . That is, one seeks to solve
where is the Frobenius norm.
It was shown in a 2008 paper by de Silva and Lim that the above standard approximation problem may be ill-posed. A solution to aforementioned problem may sometimes not exist because the set over which one optimizes is not closed. As such, a minimizer may not exist, even though an infimum would exist. In particular, it is known that certain so-called border tensors may be approximated arbitrarily well by a sequence of tensor of rank at most , even though the limit of the sequence converges to a tensor of rank strictly higher than . The rank-3 tensor
can be approximated arbitrarily well by the following sequence of rank-2 tensors
as . This example neatly illustrates the general principle that a sequence of rank- tensors that converges to a tensor of strictly higher rank needs to admit at least two individual rank-1 terms whose norms become unbounded. Stated formally, whenever a sequence
has the property that (in the Euclidean topology) as , then there should exist at least such that
as . This phenomenon is often encountered when attempting to approximate a tensor using numerical optimization algorithms. It is sometimes called the problem of diverging components. It was, in addition, shown that a random low-rank tensor over the reals may not admit a rank-2 approximation with positive probability, leading to the understanding that the ill-posedness problem is an important consideration when employing the tensor rank decomposition.
A common partial solution to the ill-posedness problem consists of imposing an additional inequality constraint that bounds the norm of the individual rank-1 terms by some constant. Other constraints that result in a closed set, and, thus, well-posed optimization problem, include imposing positivity or a bounded inner product strictly less than unity between the rank-1 terms appearing in the sought decomposition.
Calculating the CPD
General optimization algorithms:
- simultaneous diagonalization (SD)
- simultaneous generalized Schur decomposition (SGSD)
- Levenberg–Marquardt (LM)
- nonlinear conjugate gradient (NCG)
- limited memory BFGS (L-BFGS)
General polynomial system solving algorithms:
In machine learning, the CP-decomposition is the central ingredient in learning probabilistic latent variables models via the technique of moment-matching. For example, consider the multi-view model which is a probabilistic latent variable model. In this model, the generation of samples are posited as follows: there exists a hidden random variable that is not observed directly, given which, there are several conditionally independent random variables known as the different "views" of the hidden variable. For simplicity, assume there are three symmetrical views of a -state categorical hidden variable . Then the empirical third moment of this latent variable model can be written as: .
In applications such as topic modeling, this can be interpreted as the co-occurrence of words in a document. Then the eigenvalues of this empirical moment tensor can be interpreted as the probability of choosing a specific topic and each column of the factor matrix corresponds to probabilities of words in the vocabulary in the corresponding topic.
- Latent class analysis
- Multilinear subspace learning
- Singular value decomposition
- Tucker decomposition
- Higher-order singular value decomposition
- Tensor decomposition
- F. L. Hitchcock (1927). "The expression of a tensor or a polyadic as a sum of products". Journal of Mathematics and Physics. 6: 164–189.
- Carroll, J. D.; Chang, J. (1970). "Analysis of individual differences in multidimensional scaling via an n-way generalization of 'Eckart–Young' decomposition". Psychometrika. 35 (3): 283–319. doi:10.1007/BF02310791.
- Harshman, Richard A. (1970). "Foundations of the PARAFAC procedure: Models and conditions for an "explanatory" multi-modal factor analysis" (PDF). UCLA Working Papers in Phonetics. 16: 84. No. 10,085. Archived from the original (PDF) on October 10, 2004.
- Hillar, C. J.; Lim, L. (2013). "Most tensor problems are NP-Hard". Journal of the ACM. 60 (6): 1–39. arXiv:0911.1393. doi:10.1145/2512329.
- Landsberg, J. M. (2012). Tensors: Geometry and Applications. AMS.
- de Silva, V.; Lim, L. (2008). "Tensor Rank and the Ill-Posedness of the Best Low-Rank Approximation Problem". SIAM Journal on Matrix Analysis and Applications. 30 (3): 1084–1127. arXiv:math/0607647. doi:10.1137/06066518x.
- Strassen, V. (1983). "Rank and optimal computation of generic tensors". Linear Algebra and its Applications. 52/53: 645–685. doi:10.1016/0024-3795(83)80041-x.
- Catalisano, M. V.; Geramita, A. V.; Gimigliano, A. (2002). "Ranks of tensors, secant varieties of Segre varieties and fat points". Linear Algebra and its Applications. 355: 263–285. doi:10.1016/s0024-3795(02)00352-x.
- Abo, H.; Ottaviani, G.; Peterson, C. (2009). "Induction for secant varieties of Segre varieties". Transactions of the American Mathematical Society. 361 (2): 767–792. arXiv:math/0607191. doi:10.1090/s0002-9947-08-04725-9.
- Lickteig, Thomas (1985). "Typical tensorial rank". Linear Algebra and its Applications. 69: 95–120. doi:10.1016/0024-3795(85)90070-9.
- Catalisano, M. V.; Geramita, A. V.; Gimigliano, A. (2011). "Secant varieties of ℙ1 × ··· × ℙ1 (n-times) are not defective for n ≥ 5". Journal of Algebraic Geometry. 20 (2): 295–327. doi:10.1090/s1056-3911-10-00537-0.
- Blehkerman, G.; Teitler, Z. (2014). "On maximum, typical and generic ranks". Mathematische Annalen. In press. (3–4): 1–11. arXiv:1402.2371. doi:10.1007/s00208-014-1150-3.
- Bini, D.; Lotti, G.; Romani, F. (1980). "Approximate solutions for the bilinear form computational problem". SIAM Journal on Scientific Computing. 9 (4): 692–697. doi:10.1137/0209053.
- Harris, Joe (1992). Algebraic Geometry SpringerLink. Graduate Texts in Mathematics. 133. doi:10.1007/978-1-4757-2189-8. ISBN 978-1-4419-3099-6.
- Chiantini, L.; Ottaviani, G.; Vannieuwenhoven, N. (2014-01-01). "An Algorithm For Generic and Low-Rank Specific Identifiability of Complex Tensors". SIAM Journal on Matrix Analysis and Applications. 35 (4): 1265–1287. arXiv:1403.4157. doi:10.1137/140961389. ISSN 0895-4798.
- Bocci, Cristiano; Chiantini, Luca; Ottaviani, Giorgio (2014-12-01). "Refined methods for the identifiability of tensors". Annali di Matematica Pura ed Applicata. 193 (6): 1691–1702. arXiv:1303.6915. doi:10.1007/s10231-013-0352-8. ISSN 0373-3114.
- Chiantini, L.; Ottaviani, G.; Vannieuwenhoven, N. (2017-01-01). "Effective Criteria for Specific Identifiability of Tensors and Forms". SIAM Journal on Matrix Analysis and Applications. 38 (2): 656–681. arXiv:1609.00123. doi:10.1137/16m1090132. ISSN 0895-4798.
- Chiantini, L.; Ottaviani, G. (2012-01-01). "On Generic Identifiability of 3-Tensors of Small Rank". SIAM Journal on Matrix Analysis and Applications. 33 (3): 1018–1037. arXiv:1103.2696. doi:10.1137/110829180. ISSN 0895-4798.
- Hauenstein, J. D.; Oeding, L.; Ottaviani, G.; Sommese, A. J. (2016). "Homotopy techniques for tensor decomposition and perfect identifiability". J. Reine Angew. Math. arXiv:1501.00090. doi:10.1515/crelle-2016-0067.
- Bocci, Cristiano; Chiantini, Luca (2013). "On the identifiability of binary Segre products". Journal of Algebraic Geometry. 22 (1): 1–11. arXiv:1105.3643. doi:10.1090/s1056-3911-2011-00592-4. ISSN 1056-3911.
- Domanov, Ignat; Lathauwer, Lieven De (January 2014). "Canonical Polyadic Decomposition of Third-Order Tensors: Reduction to Generalized Eigenvalue Decomposition". SIAM Journal on Matrix Analysis and Applications. 35 (2): 636–660. arXiv:1312.2848. doi:10.1137/130916084. ISSN 0895-4798.
- Domanov, Ignat; De Lathauwer, Lieven (January 2017). "Canonical polyadic decomposition of third-order tensors: Relaxed uniqueness conditions and algebraic algorithm". Linear Algebra and its Applications. 513: 342–375. arXiv:1501.07251. doi:10.1016/j.laa.2016.10.019. ISSN 0024-3795.
- Faber, Nicolaas (Klaas) M.; Ferré, Joan; Boqué, Ricard (January 2001). "Iteratively reweighted generalized rank annihilation method". Chemometrics and Intelligent Laboratory Systems. 55 (1–2): 67–90. doi:10.1016/s0169-7439(00)00117-9. ISSN 0169-7439.
- Leurgans, S. E.; Ross, R. T.; Abel, R. B. (October 1993). "A Decomposition for Three-Way Arrays". SIAM Journal on Matrix Analysis and Applications. 14 (4): 1064–1083. doi:10.1137/0614071. ISSN 0895-4798.
- Lorber, Avraham. (October 1985). "Features of quantifying chemical composition from two-dimensional data array by the rank annihilation factor analysis method". Analytical Chemistry. 57 (12): 2395–2397. doi:10.1021/ac00289a052. ISSN 0003-2700.
- Sanchez, Eugenio; Kowalski, Bruce R. (January 1990). "Tensorial resolution: A direct trilinear decomposition". Journal of Chemometrics. 4 (1): 29–45. doi:10.1002/cem.1180040105. ISSN 0886-9383.
- Sands, Richard; Young, Forrest W. (March 1980). "Component models for three-way data: An alternating least squares algorithm with optimal scaling features". Psychometrika. 45 (1): 39–67. doi:10.1007/bf02293598. ISSN 0033-3123.
- Bernardi, A.; Brachat, J.; Comon, P.; Mourrain, B. (May 2013). "General tensor decomposition, moment matrices and applications". Journal of Symbolic Computation. 52: 51–71. arXiv:1105.1229. doi:10.1016/j.jsc.2012.05.012. ISSN 0747-7171.
- Bernardi, Alessandra; Daleo, Noah S.; Hauenstein, Jonathan D.; Mourrain, Bernard (December 2017). "Tensor decomposition and homotopy continuation". Differential Geometry and its Applications. 55: 78–105. arXiv:1512.04312. doi:10.1016/j.difgeo.2017.07.009. ISSN 0926-2245.
- Anandkumar, Animashree; Ge, Rong; Hsu, Daniel; Kakade, Sham M; Telgarsky, Matus (2014). "Tensor decompositions for learning latent variable models". The Journal of Machine Learning Research. 15 (1): 2773–2832.
- Kolda, Tamara G.; Bader, Brett W. (2009). "Tensor Decompositions and Applications". SIAM Rev. 51 (3): 455–500. CiteSeerX 10.1.1.153.2059. doi:10.1137/07070111X.
- Landsberg, Joseph M. (2012). Tensors: Geometry and Applications. AMS.