The document summarizes a presentation on minimizing tensor estimation error using alternating minimization. It begins with an introduction to tensor decompositions including CP, Tucker, and tensor train decompositions. It then discusses nonparametric tensor estimation using an alternating minimization method. The method iteratively updates components while holding other components fixed, achieving efficient computation. The analysis shows that after t iterations, the estimation error is bounded by the sum of a statistical error term and an optimization error term decaying exponentially in t. Real data analysis uses the method for multitask learning.

1. Minimax optimal alternating minimization
for kernel nonparametric tensor learning
†‡
Taiji Suzuki
joint work with †
Heishiro Kanagawa, ⋄
Hayato Kobayashi, ⋄
Nobuyuki Shimizu
and ⋄
Yukihiro Tagami
†
Tokyo Institute of Technology
Department of Mathematical Computing Sciences
‡
JST, PRESTO and AIP, RIKEN
⋄
Yahoo! Japan.
19th/Jan/2017
PFN 主催 NIPS2016 読み会
1 / 41

2. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
2 / 41

3. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
3 / 41

4. High dimensional parameter estimation
Vector
Sparsity
Method
Lasso
Sure Screening
Application
Feature selection
Gene data
analysis
4 / 41

5. High dimensional parameter estimation
Vector
Sparsity
Matrix
Low rank
Method
Lasso
Sure Screening
Application
Feature selection
Gene data
analysis
Method
PCA
Trace norm reg.
Application
Dim. Reduction
Recommendation
system
Three layer NN
4 / 41

6. High dimensional parameter estimation
Vector
Sparsity
Matrix
Low rank
Tensor
Low rank
Method
Lasso
Sure Screening
Application
Feature selection
Gene data
analysis
Method
PCA
Trace norm reg.
Application
Dim. Reduction
Recommendation
system
Three layer NN
This study
Higher order relation
4 / 41

7. “Tensors” in NIPS2016
Zhao Song, David Woodruff, Huan Zhang:
“Sublinear Time Orthogonal Tensor Decomposition”
Shandian Zhe, Kai Zhang, Pengyuan Wang, Kuang-chih Lee, Zenglin Xu,
Yuan Qi, Zoubin Ghahramani
“Distributed Flexible Nonlinear Tensor Factorization”
Guillaume Rabusseau, Hachem Kadri:
“Low-Rank Regression with Tensor Responses”
Chuan-Yung Tsai, Andrew M. Saxe, Andrew M. Saxe, David Cox:
“Tensor Switching Networks”
Tao Wu, Austin R. Benson, David F. Gleich:
“General Tensor Spectral Co-clustering for Higher-Order Data”
Yining Wang, Anima Anandkumar:
“Online and Differentially-Private Tensor Decomposition”
Edwin Stoudenmire, David J. Schwab:
“Supervised Learning with Tensor Networks”
5 / 41

8. Tensor workshop
Amnon Shashua: On depth efficiency of convolutional networks: the use of
hierarchical tensor decomposition for network design and analysis.
Deep neural network can be formulated as hierarchical Tucker
decomposition. (Cohen et al., 2016; Cohen & Shashua, 2016)
Three layer NN corresponds to (generalized) CP-decomposition.
DNN truly has more expressive power than shallow ones.
A(hy ) =
∑Z
z=1 ay
z (Faz,1
) ⊗g · · · ⊗g (Faz,N
)
Lek-Heng Lim: Tensor network ranks
and other interesting talks.
6 / 41

9. This presentation
Suzuki, Kanagawa, Kobayashi, Shimizu and Tagami: Minimax optimal alternating
minimization for kernel nonparametric tensor learning. NIPS2016, pp. 3783–3791.
Nonparametric low rank tensor estimation
Alternating minimization method:
efficient computation + nice statistical property.
After t iterations, the estimation erro is bounded by
˜O
(
dKn− 1
1+s + dK
(
3
4
)t
)
.
Related papers:
Suzuki: Convergence rate of Bayesian tensor estimator and its minimax optimality.
ICML2015, pp. 1273–1282, 2015.
Kanagawa, Suzuki, Kobayashi, Shimizu and Tagami: Gaussian process
nonparametric tensor estimator and its minimax optimality. ICML2016, pp.
1632–1641, 2016.
7 / 41

10. Error bound comparison
Parametric tensor model (CP-decomposition)
Method Least squares Convex reg. Bayes
via matricization
Error bound
∏K
k=1 Mk
n
dK/2
√∏
k Mk
n
d(
∑K
k=1 Mk ) log(n)
n
K: dimension of the tensor, d: rank, Mk : size
Convex reg.: Tomioka et al. (2011); Tomioka and Suzuki (2013); Zheng and
Tomioka (2015); Mu et al. (2014)
Bayes: Suzuki (2015)
Nonparametric tensor model (CP-decomposition)
Method Naive method Bayes/
Alternating min.
Error bound n− 1
1+Ks dKn− 1
1+s
K: size, s: complexity of the model space
Bayes: Kanagawa et al. (2016)
Alternating minimization: This paper.
8 / 41

11. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
9 / 41

12. Tensor decompositions
CP-decomposition
Tucker decomposition
Tensor train
Tensor network
10 / 41

13. Tensor rank: CP-rank
=
A
B
C
+ +…+
a1
b1
c1
a2
b2
c2
ad
bd
cd
=
CP-decomposition
Canonical Polyadic decomp. (Hitchcock, 1927; Hitchcock, 1927)
CANDECOMP/PARAFAC (Carroll Chang, 1970; Harshman, 1970)
Xijk =
∑d
r=1 air bjr ckr =: [[A, B, C]].
CP-decomposition defines CP-rank of a tensor.
CP-decomposition is NP-hard.
(But under a mild assumption, it can be solved efficiently (De Lathauwer,
2006; De Lathauwer et al., 2004; Leurgans et al., 1993))
Orthogonal decomposition does not necessary exist (even for symmetric
tensor).
11 / 41

14. Tensor rank: Tucker-rank
=
X
G
A
B
C
Tucker-decomposition (Tucker, 1966)
Xijk =
∑r1
l=1
∑r2
m=1
∑r3
n=1 glmnail bjmckn =: [[G; A, B, C]].
G is called core tensor.
Tucker-rank = (r1, r2, r3)
12 / 41

15. Matricization
Mode-k unfolding:
A(k)
∈ RMk ×N/Mk
, (N =
K∏
k=1
Mk ).
A A
(k)
G
A
B
C
A
Gx2Bx3C
rk = rank(A(k)
) gives the Tucker-rank.
13 / 41

16. Other tensor decomposition models
Tensor train (Oseledets, 2011)
Ti1,i2,...,iK
=
∑
α1,...,αK−1
G1(i1, α1)G2(α1, i2, α2) · · ·
GK−1(αK−2, iK−1, αK−1)GK (αK−1, iK )
i2 i3 i4
i1
i5
G1 G2 G3 G4 G5
Tensor network
14 / 41

17. Applications
Recommendation system
Relational data
Multi-task learning
Signal processing (space (2D) × time)
Natural language processing (vector representation of words)
1
13
1
2
2
2
4
2
4
21
3
2
4
1
2
3
2
3
4
2
1
3
2
1
4
1
13
2
4 41
4
1
3
3
2
1
3
2
User
Item
Context
Rating
Prediction
Tensor completion
15 / 41

18. Other applications
EEG analysis (De Vos et al., 2007)
time × frequency × space
EEG monitoring: Epileptic seizure onset localization
Denoising by tensor train (Phien et al., 2016)
Recovery by different tensor learning methods Casting an image
into a higher-order tensor 16 / 41

19. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
17 / 41

20. Nonparametric tensor regression model
Nonparametric regression model
yi = f (xi) + ϵi.
Goal: Estimate f from the data Dn = {xi , yi }n
i=1.
Nonparametric tensor model:
f (x(1)
, . . . , x(K)
) =
d∑
r=1
f (1)
r (x(1)
) × · · · × f (K)
r (x(K)
)
We suppose that f
(k)
r ∈ H where H is an RKHS.
Parametric tensor model:
f (x(1)
, . . . , x(K)
) =
d∑
r=1
⟨x(1)
, u(1)
r ⟩ × · · · × ⟨x(K)
, u(K)
r ⟩
Matrix case:
∑d
r=1⟨x(1)
, u
(1)
r ⟩⟨x(2)
, u
(2)
r ⟩ = (x(1)
)⊤
(
d∑
r=1
u(1)
r u(2)
r
⊤
)
matrix
x(2)
.
18 / 41

21. Application: Nonlinear recommendation
f (x(1)
, x(2)
) = x(1)⊤
Ax(2)
=
d∑
r=1
⟨x(1)
, u(1)
r ⟩⟨u(2)
r , x(2)
⟩
x(1)
: User feature，x(2)
: Movie feature．
19 / 41

22. Application: Nonlinear recommendation
f (x(1)
, x(2)
) =
d∑
r=1
f (1)
r (x(1)
)f (2)
r (x(2)
)
x(1)
: User feature，x(2)
: Movie feature．
19 / 41

23. Application: Smoothing
(De Vos et al., 2007) (Cao et al., 2016)
min
{u
(k)
r }r,k
X −
d∑
r=1
u(1)
r ⊗ u(2)
r ⊗ u(3)
r
2
+
d∑
r=1
3∑
k=1
u(k)⊤
r Gu(k)
r
Smoothing
t
u1
u2 u3 u4
u5
u⊤
Gu =
∑
j (uj − uj+1)2
Smoothing ⇒ Kernel method
(Zdunek, 2012; Yokota et al., 2015a; Yokota et al., 2015b)
20 / 41

24. Application: Multi-task learning
Tasktype1
Task type 2
Function (f*)
Related tasks aligned with two indexes (s, t).
f(s,t): the regression function for task (s, t).
fr (x) (r = 1, . . . , d): factors behind tasks that give an expression of f(s,t) as
f(s,t)(x) =
d∑
r=1
βr,(s,t) fr (x)
Latent factor
=
d∑
r=1
αr,sαr,tfr (x)
We estimate αr,s ∈ R, αr,t ∈ R, fr ∈ Hr by using Gaussian process prior.
21 / 41

25. Estimation methods
f (x(1)
, . . . , x(K)
) =
d∑
r=1
f (1)
r (x(1)
) × · · · × f (K)
r (x(K)
)
1. Alternating minimization (MAP estimator) (NIPS2016)
Repeating convex optimization. Fast computation.
Stronger assumptions are required for minimax optimality.
Local optimality is still problematic.
2. Bayes estimator (ICML2016)
Nice statistical performance. Minimax optimal.
Heavy computation.
3. Convex regularization (Signoretto et al., 2013)
Question
Estimation error guarantee: How does the error decrease? Is it optimal?
Computational complexity?
Performance on real data?
22 / 41

26. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
23 / 41

27. Alternating minimization method
Update f
(k)
r for a chosen (r, k) while other components are fixed:
F({f (k)
r }r,k ) :=
1
n
n∑
i=1
(
yi −
d∗
∑
r=1
K∏
k=1
f (k)
r (x
(k)
i )
)2
(Empirical error)
ˆf (k)
r ← arg min
f
(k)
r ∈H
{
F(f (k)
r |{ˆf
(k′
)
r′ }(r′,k′)̸=(r,k)) + λ∥f (k)
r ∥2
H
}
.
The objective function is non-convex.
But it is convex w.r.t. one component f
(k)
r (kernel ridge regression).
It should converge to a local optimal (e.g. coordinate descent).
24 / 41

28. Reproducing Kernel Hilbert Space (RKHS)
kernel function ⇔ Reproducing Kernel Hilbert Space (RKHS)
k(x, x′
) ⇔ Hk
Reproducibility: for f ∈ Hk, the function value at x is recovered as
f (x) = ⟨f , k(x, ·)⟩Hk
.
Representer theorem:
min
f ∈H
1
n
n∑
i=1
(yi − f (xi ))2
+ C∥f ∥2
H
⇐⇒ min
α∈Rn
1
n
n∑
i=1
(
yi −
n∑
j=1
αj k(xj , xi )
)2
+ Cα⊤
Kα,
where Ki,j = k(xi , xj ). ˆf =
∑n
i=1 αi k(xi , ·).
Gaussian kernel
Polynomial kernel
Graph kernel, time series
kernel, ... 25 / 41

29. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
26 / 41

30. Complexity of the RKHS
0 s 1: representing complexity of the model.
Spectrum decomposition:
k(x, x′
) =
∑∞
ℓ=1 µℓϕℓ(x)ϕℓ(x′
),
where {ϕℓ}∞
ℓ=1 is ONS in L2(P).
Spectrum Condition (s)
There exists 0 s 1 such that
µℓ ≤ Cℓ− 1
s (∀ℓ).
s represents the complexity of RKHS.
Large s means complex. Small s means simple.
The optimal learning rate in a single kernel learning setting is
∥ˆf − f ∗
∥2
L2(P) = Op(n− 1
1+s ).
27 / 41

31. Convergence of alternating minimization method
Assumption
f ∗
satisfies the incoherence condition (its definition is in the next slide).
P(X) = P(X1) × · · · × P(XK ).
Some other technical conditions.
ˆf [t]
: the estimator after the t-th step.
Theorem (Main result)
There exit constants C1, C2 such that, if d(ˆf [0]
, f ∗
) ≤ C1, then with probability
1 − δ, we have
∥ˆf [t]
− f ∗
∥2
L2(P) ≤ C2
(
d∗
K (3/4)
t
Optimization error
+ d∗
Kn− 1
1+s
Estimation error
log(1/δ)
)
.
Linear convergence to a local optimal (log(n) times update is sufficient)
d(ˆf [0]
, f ∗
) = Op(1) =⇒ ∥ˆf [log(n)]
− f ∗
∥2
L2
= Op
(
d∗
Kn− 1
1+s
)
.
If we start from a good initial point, then it achieves the minimax
optimal error.
Naive method: O(n− 1
1+Ks ) (curse of dimensionality).
28 / 41

32. Details of technical conditions
Incoherence: ∃µ∗
1 s.t.
|⟨f ∗(k)
r , f
∗(k)
r′ ⟩| ≤ µ∗
∥f ∗(k)
r ∥L2
∥f
∗(k)
r′ ∥L2
(r ̸= r′
).
fr*(k)
fr'*(k)
Lower and upper bound of f ∗
:
0 vmin ≤ ∥f ∗(k)
r ∥L2
≤ vmax (∀r, k).
sup-norm condition: 0 ∃s2 1 s.t.
∥f ∥∞ ≤ C∥f ∥1−s2
L2
∥f ∥s2
H (∀f ∈ H)
29 / 41

33. Details of technical conditions
Incoherence: ∃µ∗
1 s.t.
|⟨f ∗(k)
r , f
∗(k)
r′ ⟩| ≤ µ∗
∥f ∗(k)
r ∥L2
∥f
∗(k)
r′ ∥L2
(r ̸= r′
).
fr*(k)
fr'*(k)
Lower and upper bound of f ∗
:
0 vmin ≤ ∥f ∗(k)
r ∥L2
≤ vmax (∀r, k).
sup-norm condition: 0 ∃s2 1 s.t.
∥f ∥∞ ≤ C∥f ∥1−s2
L2
∥f ∥s2
H (∀f ∈ H)
For vr = ∥
∏K
k=1 f
∗(k)
r ∥L2 , ˆvr = ∥
∏K
k=1
ˆf
(k)
r ∥L2 , f
∗∗(k)
r =
f ∗(k)
r
∥f
∗(k)
r ∥L2
,
ˆˆf
(k)
r =
ˆf (k)
r
∥ˆf
(k)
r ∥L2
,
d(ˆf , f ∗
) = max
(r,k)
{|ˆvr − vr | + vr ∥ˆˆf (k)
r − f ∗∗(k)
r ∥L2
}.
29 / 41

34. Illustration of the theoretical result
True
Pred. Error
Emp. Error
True
Pred. Risk
Emp. Risk
True
Pred. Risk
Emp. Risk
Small sample Large sample
The predictive risk shapes like a convex function locally around the true
function.
The empirical risk gets closer to the predictive one as the sample size
increases.
Technique: Local Rademacher complexity
30 / 41

35. Tools used in the proof
Rademacher complexity
H: function space
Ex
[
sup
h∈H
1
n
n∑
i=1
h(xi )
Empirical error
− E[h]
Predictive error
]
≤2Ex,σ
[
sup
h∈H
1
n
n∑
i=1
σi h(xi )
]
≤
C
√
n
where {σi }n
i=1 are i.i.d. Rademacher random va-
riables (P(σi = 1) = P(σi = −1)).
Pred. Risk
Emp. Risk
Uniform bound
Local Rademacher complexity + peeling device:
Utilize strong convexity of the squared loss
Ex,σ
[
sup
h∈H
|1
n
∑n
i=1 σi h(xi )|
∥h∥L2 + λ
]
≤ C
λ− s
2
√
n
∨
λ− 1
2
n− 1
1+s
Tighter around the true function
Pred. Risk
Emp. Risk
Uniform bound
31 / 41

36. Convergence of alternating minimization
0 5 10 15 20 25
Number of iterations
10-4
10-3
10-2
10-1
100relativeMSE
n=400
n=800
n=1200
n=1600
n=2000
n=2400
n=2800
Relative MSE E[∥ˆf [t]
− f ∗
∥2
] v.s. the number of iteration t
for different sample sizes n.
32 / 41

37. Minimax optimality
The derived upper bound is minimax optimal (up to a constant).
A set of tensors with rank d∗
:
H(d∗,K)(R) :=
{
f =
d∗
∑
r=1
K∏
k=1
f (k)
r f (k)
r ∈ H(r,k)(R)
}
.
Theorem (Minimax risk)
inf
ˆf
sup
f ∗∈H(d∗,K)(R)
E[∥f ∗
− ˆf ∥2
L2(PX )] ≳ d∗
Kn− 1
1+s ,
where inf is taken over all estimators ˆf .
The Bayes estimator attains the minimax risk.
33 / 41

38. Issue of local optimality
The convergence is only proven for a good initial solution that is sufficiently close
to the optimal one.
True
Pred. Risk
Emp. Risk
Question: Does the algorithm converge to the global optimal?
→ Open question.
34 / 41

39. NIPS2016 papers about local optimality
■ Every local minimum of the matrix completion problem is the global minimum
(with high probability).
min
U∈RM×k
∑
(i,j)∈E
(Yi,j − (UU⊤
)i,j )2
Rong Ge, Jason D. Lee, Tengyu Ma: “Matrix Completion has No Spurious
Local Minimum.”
Srinadh Bhojanapalli, Behnam Neyshabur, Nati Srebro: “Global Optimality of
Local Search for Low Rank Matrix Recovery.”
■ Deep NN also satisfies a similar property.
Kenji Kawaguchi: “Deep Learning without Poor Local Minima.”
(Essentially, the proof is valid only for linear deep neural network.)
Strictly saddle function: Every critical point has a negative curvature direction
or is the global optimum.
Trust region method (Conn et al., 2000), noisy stochastic gradient (Ge et al.,
2015) can reach the global optimum for strictly saddle obj (Sun et al., 2015).
35 / 41

40. Outline
1 Introduction
2 Basics of low rank tensor decomposition
3 Nonparametric tensor estimation
Alternating minimization
Convergence analysis
Real data analysis: multitask learning
36 / 41

41. Numerical experiments on Real data
Multi-task learnnig [Nonlinear regression]
Restaurant data: multi-task learning, 138 customers × 3 aspects (138 × 3
tasks)
We want to predict the rating (3 level) of the restaurant for each customer
and each aspect.
Each restaurant is described by a 44-dimensional feature vector.
37 / 41

42. Numerical experiments on Real data
Multi-task learnnig [Nonlinear regression]
Restaurant data: multi-task learning, 138 customers × 3 aspects (138 × 3
tasks)
We want to predict the rating (3 level) of the restaurant for each customer
and each aspect.
Each restaurant is described by a 44-dimensional feature vector.
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
500 1000 1500 2000 2500
MSE
Sample size
GRBF
GRBF(2)+lin(1)
GRBF(1)+lin(2)
linear
scaled latent
ALS (Best)
ALS (50)
Nonprametric methods (Bayes and alternating minimization) achieved the best
performance.
37 / 41

43. Multi-task learnnig [Nonlinear regression]
Restaurant data: multi-task learning, 138 customers × 3 aspects (138 × 3
tasks) with a different kernel between tasks.
500 1000 1500 2000 2500
Sample size
0.35
0.40
0.45
0.50
0.55
0.60
MSE
AMP(RBF)
AMP(Linear)
Lin(2)+RBF(1)
Lin(1)+RBF(2)
GP-MTL
GP-MTL: Gaussian process method (Bayes).
AMP: alternating minimization method.
The Bayes method is slightly better.
38 / 41

44. Numerical experiments on Real data
Multi-task learnnig [Nonlinear regression]
School data: multi-task learning, 139 school × 3 years (139 × 3 tasks)
28
30
32
34
36
38
40
2000 3000 4000 5000 6000 7000 8000 900010000
Explainedvariance
Sample size
GRBF
GRBF(2)+lin(1)
GRBF(1)+lin(2)
linear
scaled latent
ALS (50)
ALS (Best)
Explained variance = 100 ×
Var(Y ) − MSE
Var(Y )
Nonprametric Bayes method and the alternating minimization method achieved
the best performance.
39 / 41

45. Online shopping sales prediction
Predict the online shopping (Yahoo! shopping) sales.
shop × item × customer (508 shops, 100 items)
Predict the number of certain items that a customer will buy in a shop.
A customer is represented by a feature vector.
We construct a kernel defined by the nearest neighbor graph between shops.
10
11
12
13
14
15
16
4000 6000 8000 10000 12000 14000
MSE
Sample size
GP-MTL(cosdis)
GP-MTL(cossim)
AMP(cosdis)
AMP(cossim)
Figuur: Sales prediction of online shop. Comparison between different metrics between
shops.
40 / 41

46. Summary
Convergence rate of nonlinear tensor estimator was given.
The alternating minimization method achieve the minimax optimality.
The theoretical analysis requires some strong assumptions, for example, on
the choice initial guess.
Estimation error of the alternating minimization procedure
∥ˆf [t]
− f ∗
∥2
L2(Π) ≤ C
(
dKn− 1
1+s + dK(3/4)t
)
.
where ˆf [t]
is the solution of the t-th update during the procedure.
41 / 41

47. Cao, W., Wang, Y., Sun, J., Meng, D., Yang, C., Cichocki, A., Xu, Z. (2016).
Total variation regularized tensor RPCA for background subtraction from
compressive measurements. IEEE Transactions on Image Processing, 25,
4075–4090.
Carroll, J. D., Chang, J.-J. (1970). Analysis of individual differences in
multidimensional scaling via an n-way generalization of“ eckart-young ”
decomposition. Psychometrika, 35, 283–319.
Cohen, N., Sharir, O., Shashua, A. (2016). On the expressive power of deep
learning: A tensor analysis. The 29th Annual Conference on Learning Theory
(pp. 698–728).
Cohen, N., Shashua, A. (2016). Convolutional rectifier networks as generalized
tensor decompositions. Proceedings of the 33th International Conference on
Machine Learning (pp. 955–963).
Conn, A. R., Gould, N. I., Toint, P. L. (2000). Trust region methods, vol. 1.
Siam.
De Lathauwer, L. (2006). A link between the canonical decomposition in
multilinear algebra and simultaneous matrix diagonalization. SIAM journal on
Matrix Analysis and Applications, 28, 642–666.
De Lathauwer, L., De Moor, B., Vandewalle, J. (2004). Computation of the
canonical decomposition by means of a simultaneous generalized schur
decomposition. SIAM journal on Matrix Analysis and Applications, 26, 295–327.
41 / 41

48. De Vos, M., Vergult, A., De Lathauwer, L., De Clercq, W., Van Huffel, S.,
Dupont, P., Palmini, A., Van Paesschen, W. (2007). Canonical
decomposition of ictal scalp eeg reliably detects the seizure onset zone.
NeuroImage, 37, 844–854.
Ge, R., Huang, F., Jin, C., Yuan, Y. (2015). Escaping from saddle
points—online stochastic gradient for tensor decomposition. Proceedings of
The 28th Conference on Learning Theory (pp. 797–842).
Harshman, R. A. (1970). Foundations of the PARAFAC procedure: Models and
conditions for an “explanatory” multi-modal factor analysis. UCLA Working
Papers in Phonetics, 16, 1–84.
Hitchcock, F. L. (1927). Multilple invariants and generalized rank of a p-way
matrix or tensor. Journal of Mathematics and Physics, 7, 39–79.
Kanagawa, H., Suzuki, T., Kobayashi, H., Shimizu, N., Tagami, Y. (2016).
Gaussian process nonparametric tensor estimator and its minimax optimality.
Proceedings of the 33rd International Conference on Machine Learning
(ICML2016) (pp. 1632–1641).
Leurgans, S., Ross, R., Abel, R. (1993). A decomposition for three-way arrays.
SIAM Journal on Matrix Analysis and Applications, 14, 1064–1083.
Mu, C., Huang, B., Wright, J., Goldfarb, D. (2014). Square deal: Lower
bounds and improved relaxations for tensor recovery. Proceedings of the 31th
International Conference on Machine Learning (pp. 73–81).
41 / 41

49. Oseledets, I. V. (2011). Tensor-train decomposition. SIAM Journal on Scientific
Computing, 33, 2295–2317.
Phien, H. N., Tuan, H. D., Bengua, J. A., Do, M. N. (2016). Efficient tensor
completion: Low-rank tensor train. arXiv preprint arXiv:1601.01083.
Signoretto, M., Lathauwer, L. D., Suykens, J. A. K. (2013). Learning tensors in
reproducing kernel Hilbert spaces with multilinear spectral penalties. CoRR,
abs/1310.4977.
Sun, J., Qu, Q., Wright, J. (2015). When are nonconvex problems not scary?
arXiv preprint arXiv:1510.06096.
Suzuki, T. (2015). Convergence rate of Bayesian tensor estimator and its minimax
optimality. Proceedings of the 32nd International Conference on Machine
Learning (ICML2015) (pp. 1273–1282).
Tomioka, R., Suzuki, T. (2013). Convex tensor decomposition via structured
schatten norm regularization. Advances in Neural Information Processing
Systems 26 (pp. 1331–1339). NIPS2013.
Tomioka, R., Suzuki, T., Hayashi, K., Kashima, H. (2011). Statistical
performance of convex tensor decomposition. Advances in Neural Information
Processing Systems 24 (pp. 972–980). NIPS2011.
Tucker, L. R. (1966). Some mathematical notes on three-mode factor analysis.
Psychometrika, 31, 279–311.
41 / 41

50. Yokota, T., Zdunek, R., Cichocki, A., Yamashita, Y. (2015a). Smooth
nonnegative matrix and tensor factorizations for robust multi-way data analysis.
Signal Processing, 113, 234–249.
Yokota, T., Zhao, Q., Cichocki, A. (2015b). Smooth parafac decomposition for
tensor completion. arXiv preprint arXiv:1505.06611.
Zdunek, R. (2012). Approximation of feature vectors in nonnegative matrix
factorization with gaussian radial basis functions. International Conference on
Neural Information Processing (pp. 616–623).
Zheng, Q., Tomioka, R. (2015). Interpolating convex and non-convex tensor
decompositions via the subspace norm. Advances in Neural Information
Processing Systems (pp. 3088–3095).
41 / 41