三模Tucker积张量秩的一些性质
2021-01-04张双韩乐
张双,韩乐
三模Tucker积张量秩的一些性质
张双,韩乐
(华南理工大学 数学学院,广东 广州 510640)
张量Tubal秩的定义不止一种,但本质上是用离散傅立叶变换矩阵对原始张量做三模Tucker积得到一个复张量,这个复张量所有前片秩的最大值就是张量Tubal秩.借助三模Tucker积从代数角度研究三阶张量Tubal秩的计算,并给出原始张量与变换后的复张量之间CP秩、Tucker秩的关系.
Tucker积;Tubal秩;CP秩;Tucker秩
1 引言及预备知识
在计算机视觉和信号处理领域,有大量的多模态数据需要被分析和处理.作为向量和矩阵的高阶推广,张量可以更便利地对多模态数据进行数学建模和分析.张量学习、张量优化受到越来越多国内外学者的关注,在计算机视觉、机器学习、信号处理和模式识别等领域得到广泛的应用[1-4].
张量的主要应用之一是张量低秩恢复,它离不开张量秩的定义.为了避免矩阵展开破坏张量数据的内在结构,Kilmer[5]等对三阶张量沿着第三维做离散傅立叶变换,提出了张量奇异值分解(t-SVD),并衍生出了一些新的张量秩[6-7].对于一个原始的三阶张量,沿张量的第三维做离散傅立叶变换可以得到一个同样大小的复张量,这个复张量所有前片的秩形成的向量是原始张量的Multi秩,所有前片秩的最大值称为原始张量的Tubal秩.自然地,可以用傅立叶域上每个前片的核范数近似它的秩,文献[8-9]相应地给出了一种张量Tubal秩的凸代理,一经提出就受到机器学习领域研究学者的关注,在张量完全、子空间聚类和张量低秩稀疏分解等方面取得了一些研究成果[10-14].
然而,对于张量Tubal秩的直接计算的研究,目前基本处于空白状态.本文从代数的角度研究了三阶张量Tubal秩的计算,利用三模Tucker积与张量第三维上离散傅立叶变换的关系,刻画了一些计算张量Tubal秩的条件,在这些条件下,不需要对原始张量第三维做离散傅立叶变换就可以确定三阶张量的Tubal秩.另外,还讨论了离散傅里叶变换前后张量的Tucker秩和CP秩的关系.虽然对于多个前片的三阶张量未能给出理论分析,但对公开数据集waterSurface和hall进行了测试,从数值上验证了多个前片的三阶张量有类似于2个前片的三阶张量的性质.
2 主要结果及证明
3 数值模拟
表1 实验结果
4结语
本文利用张量第三维上离散傅立叶变换与张量三模Tucker积的关系,主要刻画了一些张量Tubal秩计算的充分条件.在这些条件下,不需要对原始张量的第三维做离散傅立叶变换就可以确定三阶张量的Tubal秩.对于多个前片的三阶张量虽然未能给出理论分析,但对张量数据waterSurface和hall的数值测试显示多个前片的三阶张量也有类似于2个前片的三阶张量的性质.
[1] Cichocki A,Mandic D,Lathauwer L,et al.Tensor decompositions for signal processing applications:from two-way to multiway component analysis[J].IEEE Signal Processing Magazine,2015,32(2):145-163
[2] Qi L,Chen Y,Bakshi M,et al.Triple Decomposition and Tensor Recovery of Third Order Tensors[J].arXiv,2020:2002.02259
[3] Kamal M H,Heshmat B,Raskar R,et al.Tensor low-rank and sparse light field photography[J].Computer vision and image understanding,2016,145(4):172-181
[4] Liu Y,Long Z,Huang H,et al.Low CP rank and Tucker rank tensor completion for estimating missing components in image data[J].IEEE Transactions on Circuits and Systems for Video Technology,2020,30(4):944- 954
[5] Kilmer M E,Martin C D.Factorization strategies for third-order tensor[J].Linear Algebra and Its Applications,2011,435(3): 641-658
[6] Kilmer M E,Braman K,Hao N.Third order tensors as operators on matrices:a theoretical and computational framework with applications in imaging[J].SIAM Journal on Matrix Analysis & Applications,2013,34(1):148-172
[7] Lu C,Feng J,Chen Y,et al.Tensor Robust Principal Component Analysis:Exact Recovery of Corrupted Low-Rank Tensors via Convex Optimization[C]//IEEE International Conference on Computer Vision and Pattern Recognition,Las Vegas:IEEE,2016:249-257
[8] Zhang Z,Ely G,Aeron S,et al.Novel methods for multilinear data completion and de-noising based on tensor-SVD[C]// IEEE Conference on Computer Vision and Pattern Recognition,Columbus:IEEE,2014:42–49
[9] Hao N,Kilmer M E,Braman K,et al.Facial recognition using tensor-tensor decompositions[J].SIAM Journal of Imaging Sciences, 2013,6(1):437-463
[10] Lu C,Feng J,Chen Y,et al.Tensor Robust Principal Component Analysis with A New Tensor Nuclear Norm[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2020,42(4):925-938
[11] Zhang Z,Aeron S.Exact tensor completion using t-SVD[J].IEEE Transactions on Signal Processing,2017,65(6):1511-1526
[12] Xie Y,Tao D,Zhang W,et al.On unifying multi-view self-representations for clustering by tensor multi-rank minimization[J].International Journal of Computer Vision,2018,126:1157-1179
[13] Kolda T,Bader B.Tensor decompositions and applications[J].SIAM Review,2009,51(3):455-500
[14] Wang A,Song X,Wu X,et al.Robust low-tubal-rank tensor completion[C]//IEEE International Conference on Acoustics,Speech and Signal Processing,Brighton:IEEE,2019:32-36
[15] Jiang B,Yang F,Zhang S.Tensor and its tucker core:The invariance relationships[J].Numerical Linear Algebra with Applications,2017,24(3):1-13
Some properties of the tensor rank under 3-mode Tucker product
ZHANG Shuang,HAN Le
(School of Mathematics,South China University of Technology,Guangzhou 510640,China)
Although there are some definitions of the Tubal rank,in essence,the Tubal rank of tensor is the maximum of the ranks of the frontal slices of a complex tensor,that is obtained by the 3-mode Tucker product with the discrete Fourier Transformation.The calculation of the Tubal rank of the third-order tensors is researched from the algebraic perspective by means of the 3-mode Tucker product,and the relationships of CP rank and Tucker rank between the original tensor and the complex tensor was given.
Tucker product;Tubal rank;CP rank;Tucker rank
O151
A
10.3969/j.issn.1007-9831.2020.11.004
1007-9831(2020)11-0014-05
2020-06-27
教育部人文社科青年基金资助项目(17YJC630026);华南理工大学研究生全英文课程建设项目(Y2181421)
张双(1995-),女,四川达州人,在读硕士研究生,从事张量优化研究.E-mail:zhshuangs@163.com