Citation: Ming Yang, Dunren Che, Wen Liu, Zhao Kang, Chong Peng, Mingqing Xiao, Qiang Cheng. On identifiability of 3-tensors of multilinear rank (1; Lr; Lr)[J]. Big Data and Information Analytics, 2016, 1(4): 391-401. doi: 10.3934/bdia.2016017
[1] | Zhouchen Lin . A Review on Low-Rank Models in Data Analysis. Big Data and Information Analytics, 2016, 1(2): 139-161. doi: 10.3934/bdia.2016001 |
[2] | Ugo Avila-Ponce de León, Ángel G. C. Pérez, Eric Avila-Vales . A data driven analysis and forecast of an SEIARD epidemic model for COVID-19 in Mexico. Big Data and Information Analytics, 2020, 5(1): 14-28. doi: 10.3934/bdia.2020002 |
[3] | Subrata Dasgupta . Disentangling data, information and knowledge. Big Data and Information Analytics, 2016, 1(4): 377-390. doi: 10.3934/bdia.2016016 |
[4] | Guojun Gan, Qiujun Lan, Shiyang Sima . Scalable Clustering by Truncated Fuzzy c-means. Big Data and Information Analytics, 2016, 1(2): 247-259. doi: 10.3934/bdia.2016007 |
[5] | Elnaz Delpisheh, Aijun An, Heidar Davoudi, Emad Gohari Boroujerdi . Time aware topic based recommender System. Big Data and Information Analytics, 2016, 1(2): 261-274. doi: 10.3934/bdia.2016008 |
[6] |
Hamzeh Khazaei, Marios Fokaefs, Saeed Zareian, Nasim Beigi-Mohammadi, Brian Ramprasad, Mark Shtern, Purwa Gaikwad, Marin Litoiu .
How do I choose the right NoSQL solution? A comprehensive theoretical and experimental survey . Big Data and Information Analytics, 2016, 1(2): 185-216.
doi: 10.3934/bdia.2016004
|
[7] | Bill Huajian Yang . Resolutions to flip-over credit risk and beyond-least squares estimates and maximum likelihood estimates with monotonic constraints. Big Data and Information Analytics, 2018, 3(2): 54-67. doi: 10.3934/bdia.2018007 |
[8] | Tao Wu, Yu Lei, Jiao Shi, Maoguo Gong . An evolutionary multiobjective method for low-rank and sparse matrix decomposition. Big Data and Information Analytics, 2017, 2(1): 23-37. doi: 10.3934/bdia.2017006 |
[9] | Yaguang Huangfu, Guanqing Liang, Jiannong Cao . MatrixMap: Programming abstraction and implementation of matrix computation for big data analytics. Big Data and Information Analytics, 2016, 1(4): 349-376. doi: 10.3934/bdia.2016015 |
[10] | Ruiqi Li, Yifan Chen, Xiang Zhao, Yanli Hu, Weidong Xiao . TIME SERIES BASED URBAN AIR QUALITY PREDICATION. Big Data and Information Analytics, 2016, 1(2): 171-183. doi: 10.3934/bdia.2016003 |
The importance and usefulness of tensors that are characterized by multiway arrays for big data sets, has been increasingly recognized in the last decades, as testified by a number of surveys [15,20,14,5,17] and among others. Identifiability property (see [3,10,2,9]), including both exact and generic identifiability, is critical for tensor models in various applications, and widely used in many areas, such as signal processing, statistics, computer science, and so on. For instance, in signal processing, the tensor encodes data from received signals and one needs to decompose the tensor to obtain the transmitted signals. If the uniqueness does not hold, one may not recover the transmitted signals. Therefore, to establish the uniqueness property of appropriate tensor decomposition is not only mathematically interest but also necessary in various real applications. Extensive studies under the framework of algebraic geometry have provided various characteristics involving tensor rank and dimensions to ensure generic identifiability.
In this paper, we consider the model of low multilinear rank tensor decomposition (
Definition 1.1. (see Chapter Ⅲ in [19]) Let
(a1,…,αak+βa′k,…,an)−α(a1,…,ak,…,an)−β(a1,…,a′k,…,an) |
for all
An element of
The elements of
If
Kl⊗Km⊗Kn=Kl×m×n |
through the interpretation of the tensor product of vectors as a tensor via the Segre outer product,
[u1,…,ul]T⊗[v1,…,vm]T⊗[w1,…,wn]T=[uivjwk]l,m,ni,j,k=1. |
Definition 1.2. The Khatria-Rao Product is the "matching columnwise" Segre outer product. Given matrices
A⊙B=[a1⊗b1a2⊗b2⋯aK⊗bK]. |
If
Given standard orthnormal bases
X=I1,…,IN∑i1,…,iN=1ti1⋯iNe(1)i1⊗⋯⊗e(N)iN. |
In older literature, the
The rank of
rank(X)=min{r:X=r∑p=1a(1)p⊗⋯⊗a(N)p}. |
Definition 1.3. The
♭n:KI1×⋯×IN→KIn×(I1…ˆIn…IN) |
defined by
(♭n(X))ij=(X)sn(i,j), |
where
For a tensor
r1=dimspanK{X1∙∙,…,Xl∙∙},r2=dimspanK{X∙1∙,…,X∙m∙},r3=dimspanK{X∙∙1,…,X∙∙n}. |
Here
Xi∙∙=[tijk]m,nj,k=1∈Km×n,X∙j∙=[tijk]l,ni,k=1∈Kl×n,X∙∙k=[tijk]l,mi,j=1∈Kl×m. |
The multilinear rank of
Definition 1.4. (see Definition 11 in [4]) A decomposition of a tensor
X=R∑r=1ar⊗Xr, |
in which the
It is clear that in
Definition 1.5. Let
μK{(KI×R×KJ×K×R):X=R∑r=1ar⊗Xr is not unique for ar∈KI,Xr∈KJ×K}. |
Note that in Definition 1.4, we could require
Definition 1.6. A decomposition of a tensor
X=R∑r=1ar⊗Xr. |
As in Definition 1.4,
The main results of the paper are the following, and their proofs will be given in the following sections:
Theorem 1.7. Assume
spanK{Xj1, …, Xjs}∩Σ≤Ljt(KJ×K)⊂{Xj1, …, Xjs}, 1≤t≤s, |
where
Remark 1. In reasonably small cases, one can use tools from numerical algebraic geometry such as those described in [18,12,13].
Remark 2. A generic
p=b′1⊗c′1+⋯+b′L⊗c′L, |
where
We now establish a simpler condition related to the uniqueness of
Theorem 1.8.
I≥2, J=K≠one of {2L1+L22,2L2+L12,L1,L2}. |
Theorem 1.9.
I≥R, K≥R∑r=1Lr, J≥2max{Li}, (Jmax{Li})≥R,Li+Lj>Lk |
for all
For low multilinear rank decomposition in orthogonal frame, we have the following theorem.
Theorem 1.10. A tensor decomposition of
X=R∑r=1ar⊗Xr, |
as in Definition 1.6 is essentially unique if and only if for any non-identity special orthogonal matrix
rank (εk1X1+⋯+εkRXR)≠L1,…,LR. |
In this paper, we first provide some known and preliminary results related to the tensor decompositions of multilinear rank
Definition 2.1. For a vector space
Proof.
a′r∈spanK{a1,…,aR}. |
Since if not, we have
a′∗r∈span⊥K{a1,…,aR}. |
This implies
⟨X, a′∗r⟩=0=X′r, |
which is a contradiction. Therefore, we have
X=R∑r=1a′r⊗X′r=R∑r=1(R∑j=1αrjar⊗X′j), |
we know that
X′r∈spanK{Xj1,…,Xjs}∩Σ≤Lr(KJ×K). |
But
a1⊗X1+⋯+aR⊗XR=a1⊗X′jt−χ2a1⊗X2−⋯−χRa1⊗XR+a2⊗X2+⋯+aR⊗XR=a1⊗X′jt+(a2−χ2a1)⊗X2+⋯+(aR−χRa1)⊗XR=a1⊗X′jt+a′2⊗X2+⋯+a′R⊗XR. |
So
Example 1. A tensor decomposition of
X=R∑r=1ar⊗Xr, |
as in Definition 1.4 is essentially unique if the singular vectors of
Proof. Assume the contrary that
X′r=χ1X1+⋯+χRXR. |
Let
Ur=(|||ur1ur2⋯urJ|||) |
Vr=(|||vr1vr2⋯vrK|||) |
and
Xr=σr1ur1⊗vr1+⋯+σrLrurLr⊗vrLr, |
then we can see the rank of
Proof. It is sufficient to prove the case
Consider
X1=b1,1⊗c1,1+⋯+b1,L1−lb⊗c1,L1−lb+b0,1⊗c1,L1−lb+1+⋯+b0,lb⊗c0,lc∈(B1⊕B0)⊗(C1⊕C0)≅KL1⊗KL1,X2=b2,1⊗c2,1+⋯+b2,L1−lb⊗c2,L1−lb+b0,1⊗c2,L1−lb+1+⋯+b0,lb⊗c0,lc∈(B2⊕B0)⊗(C2⊕C0)≅KL2⊗KL2, |
where
{b0,1,…,b0,lb},{b1,1,…,b1,L1−lb},{b2,1,…,b2,L2−lb},{c0,1,…,c0,lc},{c1,1,…,c1,L1−lc},{c2,1,…,c2,L2−lc}, |
are bases for
(χ1⋱χ1χ1+χ2⋱χ1+χ2χ2⋱χ2) |
has rank
rank (χ1X1+χ2X2)≠L1 or L2 |
if and only if
J,K≠one of {2L1+L22,2L2+L12,L1,L2}. |
Then Theorem 1.8 follows from Theorem 1.7.
Example 2. For
X=a1⊗(b1⊗c1+b2⊗c2)+a2⊗(b2⊗c2+b3⊗c3)=a1⊗(b1⊗c1−b3⊗c3)+(a1+a2)⊗(b2⊗c2+b3⊗c3), |
where
Example 3. For
X=a1⊗(b1⊗c1+b2⊗c2)+a2⊗(b3⊗c1+b4⊗c2)=a1⊗((b1+b3)⊗c1+(b2+b4)⊗c2)+(a2−a1)⊗(b3⊗c1+b4⊗c2), |
where
Example 4. There are explicit Weierstrass canonical forms (see Chapter 10 in [16]) of tensors in
a1⊗(b1⊗c1+⋯+bL⊗cL)+a2⊗(λ1b1⊗c1+⋯+λLbL⊗cL), |
but it is obviously not unique.
Proof. It is sufficient to prove the case
Without loss of generality, for
Ejp=bjp,1⊗cjp,1+bjp,2⊗cjp,2+⋯+bjp,Ljp⊗cjp,Ljp∈Bjp⊗Cjp, |
where
E′jt=b′1⊗c′1+⋯+b′Ljt⊗c′Ljt |
be a general point of
E′jt=∑1≤p≤sχpEjp=∑1≤p≤sχp(bjp,1⊗cjp,1+bjp,2⊗cjp,2+⋯+bjp,Ljp⊗cjp,Ljp). |
If there exist
(xμ⋱xμxμ+xν⋱xμ+xνxν⋱xν) |
has rank at least
The following Remark can be easily obtained using elementary combinatorics.
Remark 3. When
I≥R, J, K≥R∑r=1Lr, Li+Lj>Lk ∀1≤i,j,k≤R, |
a low multilinear rank tensor decomposition of
X=R∑r=1ar⊗[(∑ru=1Lu∑r′=1+∑r−1u=1Lubr′⊗cr′)]. |
Proof.
[X1⋯XR]⊙[e1⋮eR]=[X′1⋯X′R]⊙[e′1⋮e′R]=[X′1⋯X′R]⊙Q[e1⋮eR]. |
Since
X′r=εr1X1+⋯+εrRXR, 1≤r≤R. |
However
rank X′r=rank (εr1X1+⋯+εrRXR)≠L1,…,LR, |
which is a contradiction. Therefore
X′i=εi1X1+⋯+εiRXR, 1≤i≤R, |
we then have
Remark 4. Since the rotation matrix in the plane is
(cos θ−sin θsin θ+cos θ), |
a tensor decomposition of
rank (cosθ X1+sinθ X2)≠L1 or L2, |
and same for
Different from most current approach in the analysis of big data sets, in this paper, some uniqueness characteristics of low multilinear rank tensor decomposition
The first author Ming Yang is grateful to Mingqing Xiao for his insight and for clarity of proofs. The first author is also grateful to the Qiang Cheng's machine learning lab, where this paper was mainly written.
[1] | [ L. Chiantini and C. Ciliberto, Weakly defective varieties, Transactions of the American Mathematical Society, 354(2002), 151-178. |
[2] | [ L. Chiantini and G. Ottaviani, On generic identifiability of 3-tensors of small rank, SIAM Journal on Matrix Analysis and Applications, 33(2012), 1018-1037. |
[3] | [ L. Chiantini, G. Ottaviani and N. Vannieuwenhoven, An algorithm for generic and low-rank specific identifiability of complex tensors, SIAM Journal on Matrix Analysis and Applications, 35(2014), 1265-1287. |
[4] | [ A. Cichocki, D. Mandic, L. De Lathauwer, G. Zhou, Q. Zhao, C. Caiafa and H. Phan, Tensor decompositions for signal processing applications:From two-way to multiway component analysis, Signal Processing Magazine, IEEE, 32(2015), 145-163. |
[5] | [ P. Comon, Tensor decompositions, State of the art and applications, J. G. McWhirter and I. K. Proudler. Mathematics in Signal Processing V, Clarendon Press, Oxford, 71(2002), 1-24. |
[6] | [ L. DeLathauwer, Decompositions of a higher-order tensor in block terms-part I:Lemmas for partitioned matrices, SIAM Journal on Matrix Analysis and Applications, 30(2008), 1022-1032. |
[7] | [ L. DeLathauwer, Decompositions of a higher-order tensor in block terms-part Ⅱ:Definitions and uniqueness, SIAM Journal on Matrix Analysis and Applications, 30(2008), 1033-1066. |
[8] | [ L. DeLathauwer and D. Nion, Decompositions of a higher-order tensor in block terms-part Ⅲ:Alternating least squares algorithms, SIAM Journal on Matrix Analysis and Applications, 30(2008), 1067-1083. |
[9] | [ I. Domanov and L. DeLathauwer, Generic uniqueness of a structured matrix factorization and applications in blind source separation, IEEE Journal of Selected Topics in Signal Processing, 10(2016), 701-711. |
[10] | [ I. Domanov and L. De Lathauwer, Generic uniqueness conditions for the canonical polyadic decomposition and INDSCAL, SIAM Journal on Matrix Analysis and Applications, 36(2015), 1567-1589. |
[11] | [ J. Draisma and J. Kuttler, Bounded-rank tensors are defined in bounded degree, Duke Mathematical Journal, 163(2014), 35-63. |
[12] | [ J. D. Hauenstein and A. J. Sommese, Witness sets of projections, Applied Mathematics and Computation, Elsevier, 217(2010), 3349-3354. |
[13] | [ J. D. Hauenstein and A. J. Sommese, Membership tests for images of algebraic sets by linear projections, Applied Mathematics and Computation, Elsevier, 219(2013), 6809-6818. |
[14] | [ R. Ke, W. Li and M. Xiao, Characterization of extreme points of multi-stochastic tensors, Computational Methods in Applied Mathematics, 16(2016), 459-474. |
[15] | [ T. G. Kolda and B. W. Bader, Tensor decompositions and applications, SIAM Review, 51(2009), 455-500. |
[16] | [ J. M. Landsberg, Tensors:Geometry and Applications, AMS, Providence, Rhode Island, USA, 2012. |
[17] | [ M. W. Mahoney, L.-H. Lim and G. E. Carlsson, Algorithmic and statistical challenges in modern largescale data analysis are the focus of MMDS 2008, ACM SIGKDD Explorations Newsletter, 10(2008), 57-60. |
[18] | [ F. Malgouyres and J. Landsberg, Stable recovery of the factors from a deep matrix product and application to convolutional network, preprint, arXiv:1703.08044. |
[19] | [ Y. Matsushima (E. Kobayashi, translator), Differentiable Manifolds, Marcel Dekker, Inc., North-Holland Publishing Co., North Miami Beach, FL, U.S.A., 1972. |
[20] | [ E. E. Papalexakis, C. Faloutsos and N. D. Sidiropoulos, Tensors for data mining and data fusion:Models, applications, and scalable algorithms, ACM Transactions on Intelligent Systems and Technology (TIST), 8(2017), 1-44. |