RANK-(Mr, Nr, ·) TERMS 2
IGNAT DOMANOV†, NICO VERVLIET†, AND LIEVEN DE LATHAUWER† 3
Key words. multilinear algebra, third-order tensor, block term decomposition, multilinear rank 4
AMS subject classifications. 15A23, 15A69 5
1. Introduction.
6
1.1. Terminology and problem statement. Throughout the paper F denotes
7
the field of real or complex numbers; bold-face lower case, bold-face capital case, and
8
capital calligraphic letters refer to vectors, matrices, and tensors, respectively (e.g.,
9
a1, B2, and T ); O and 0 denote a zero matrix and a zero vector, respectively; Inand 10
Om×n denote the n × n identity matrix and the m × n zero matrix, respectively; On 11
a is shorthand for On×n; Null (·) denotes the null space of a matrix; “∗” , “T”, and 12
“H” denote the conjugation, transpose, and hermitian transpose, respectively.
13
Let tensor T ∈ FI×J ×K and let T1, . . . .TK ∈ FI×J denote the frontal slices of 14 T . The matrices 15 T(1)= [T1 . . . TK] ∈ FI×J K, 16 T(2)= [TT1 . . . TTK] ∈ FJ ×IK, 17
T(3)= [vec(T1) . . . vec(TK)]T ∈ FK×IJ, 18
19
are called the matrix unfoldings of T and the triplet of their ranks (rT(1), rT(2), rT(3))
20
is called the multiLinear (ML) rank of T .
21
Tensor T is ML rank-(M, N, ·) if rT(1) = M , rT(2) = N and rT(3) is not specified.
22
It can easily be shown that T is ML rank-(M, N, ·) if and only if its frontal slices can
23
be simultaneously factorized as
24
(1) Tk = ADkBT, Dk∈ FM ×N, k = 1, . . . , K, 25
where the factors A ∈ FI×M
and B ∈ FJ ×N have full column rank and the matrices
D(1):= [D1 . . . DK] ∈ FM ×N K and D(2):= [DT1 . . . D T K] ∈ F
N ×M K
have full row rank. (One can, for instance, take A and B equal to the “U” factors in
26
the compact SVDs of T(1)and T(2), respectively, and set Dk = AHTkB∗.) It’s worth 27
to be noted that the factors A, Dk, and B in(1)are not unique. Moreover, a generic 28
∗Submitted to the editors DATE.
Funding: This work was funded by (1) Research Council KU Leuven: C1 project c16/15/059-nD; (2) F.W.O.: project G.0830.14N, G.0881.14N; (3) the Belgian Federal Science Policy Office: IUAP P7 (DYSCO II, Dynamical systems, control and optimization, 2012-2017); (4) EU: The re-search leading to these results has received funding from the European Rere-search Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC Advanced Grant: BIOTENSORS (no. 339804). This paper reflects only the authors’ views and the Union is not liable for any use that may be made of the contained informationTODO: Nico
† Group Science, Engineering and Technology, KU Leuven - Kulak, E. Sabbelaan 53, 8500
Kortrijk, Belgium and Dept. of Electrical Engineering ESAT/STADIUS KU Leuven, Kasteel-park Arenberg 10, bus 2446, B-3001 Leuven-Heverlee, Belgium (ignat.domanov@kuleuven.be,
ML rank-(M, N, ·) tensor can parameterized with (I − M )M + (J − N )N + M N K
29
parameters. Indeed, if X ∈ FI×I
and Y ∈ FJ ×J are nonsingular matrices, then 30
T1, . . . , TK can also be factorized as 31
(2) Tk = (AX) X−1DkY−T (BY)T = bA bDkBbT, k = 1, . . . , K.
32
Assuming that A, B, and Dk in (2)are generic we can set X (resp. Y) equal to the 33
inverse of the matrix formed by the first M (resp. N ) rows of A (resp. B), implying
34
that the right-hand side of the K equations
35 Tk= bA bDkBbT = IM ∗ ∗IN ∗ T , k = 1, . . . , K. 36
can be parameterized with (I − M )M + (J − N )N + M N · K parameters.
37
We will write the K factorizations in(1)as T = D •1A •2B,
where D denotes an M × N × K tensor with frontal slices D1, . . . , DK. 38
In this paper we study a decomposition of T into a sum of ML rank-(Mr, Nr, ·) 39
tensors (seeFigure 1)
40 (3) T = R X r=1 Dr•1Ar•2Br, Dr∈ FMr×Nr×K, Ar∈ FI×Mr, Br∈ FJ ×Nr. 41 T = A1 B1 D1 + · · · + AR BR DR
Fig. 1. ML rank-(Mr, Nr, ·) decomposition of a third-order tensor
42
Definition 1.1. ML rank-(Mr, Nr, ·) decomposition of tensor T is called unique 43
if for any two decompositions of the form (3)one can be obtained from another by a
44
permutation of summands.
45
Definition 1.2. ML rank-(Mr, Nr, ·) decomposition is called generically unique 46 if 47 48 µ{(A1, . . . , AR, B1, . . . , BR, D1, . . . , DR) : 49 the decomposition of T = R X r=1 Dr•1Ar•2Br is not unique} = 0, 50 51
where µ denotes a measure on FIP Mr+JP Nr+KP MrNr absolutely continuous with
52
respect to the Lebesgue measure.
53
We say that T is a sum of R generic ML rank-(Mr, Nr, ·) terms if the entries of Ar, 54
Br, and Dr in(3)are randomly sampled from an absolutely continuous distribution. 55
Thus, the generic uniqueness means uniqueness that holds with probability one.
Let
57
(4) A = [A1 . . . AR] ∈ FI×P Mr and B = [B1 . . . BR] ∈ FJ ×P Nr. 58
Throughout the paper we assume that the matrices
59
(5) A and B have full column rank.
60
1.2. Previous results. Let Dr,k∈ FMr×Nr denote the k-th frontal slice of the 61
tensor Dr and let 62
(6) Dk = blockdiag(D1,k, . . . , DR,k) ∈ FP Mr×P Nr, k = 1, . . . , K 63
denote a block-diagonal matrix with the matrices D1,k, . . . , DR,kon the diagonal. By 64
(1), tensor decomposition(3)can be interpreted as the problem of (possibly
nonsym-65
metric and nonsquare) Joint Block Diagonalization of the frontal slices of T :
66
(7) Tk= ADkBT = A blockdiag(D1,k, . . . , DR,k)BT, k = 1, . . . , K, 67
where A and B are defined in(4).
68
Problem(7)(both uniqueness and algorithms) has been relatively well studied in
69
the following cases.
70
1. JBD by congruence and ∗-congruence, i.e., the cases B = A and B = A∗,
71
implying the constraint Mr= Nr for all r (see, for instance, [1,2,8] and the 72
references therein).
73
2. All blocks are row vectors, i.e., M1= · · · = MR= 1, implying that ML rank-74
(Mr, Nr, ·) terms in (3) are actually ML rank-(1, Nr, Nr) (see, for instance, 75
[3,4,10,9]);
76
3. All blocks are column vectors, i.e., N1 = · · · = NR = 1 (follows from 2. by 77
taking transpose of the identities in(7)).
78
Apart from the items 1.–3. above the generic uniqueness of (7) was proved in [3]
79
under the assumptions Mr= Nr for all r and K ≥ 3. 80
1.3. Our contribution. We obtain results on uniqueness and present
algo-81
rithms for the computation of ML rank-(Mr, Nr, ·) tensors in (3) in the case where 82
Mr6= Nr for some r. 83
1.3.1. Exact decomposition.
84
Theorem 1.3. Let T admit the ML rank-(Mr, Nr, ·) decomposition (3), the ma-85
trices A, B and D1, . . . , DK be defined in (4) and (6), respectively, and 86 (8) MD = DT1 ⊗ II −IJ⊗ D1 .. . ... DTK⊗ II −IJ⊗ DK . 87 Assume that 88
1. A and B have full column rank and
89
2. dim Null (MD) = R. 90
Then the ML rank-(Mr, Nr, ·) decomposition of T is unique and can be computed by 91
means of Coupled Simultaneous Eigenvalue Decomposition (CS-EVD).
92
Proof. Seesection 2.
Assumption 2) inTheorem 1.3expresses the fact that the system of matrix equations
94
(9) XDk= DkY, k = 1, . . . , K
95
has exactly R linearly independent solutions (X, Y) ∈ FP Mr×P Mr× FP Nr×P Nr. If
96
PI,r = blockdiag(OM1, . . . , OMr−1, IMr, OMr+1, . . . , OMR) and
PJ,r = blockdiag(ON1, . . . , ONr−1, INr, ONr+1, . . . , ONR), r = 1, . . . , R, (10)
97
then (X, Y) = (PI,r, PJ,r) is, obviously, a solution of (9), 98 99 PI,rDk = DkPJ,r= 100 blockdiag(OM1×N1, . . . , OMr−1×Nr−1, Dr,k, OMr+1×Nr+1, . . . , OMR×NR). 101 102
Thus, assumption 2) in Theorem 1.3 means that the dimension of the subspace of
103
solutions of (9)is minimal.
104
The algebraic procedure related toTheorem 1.3 is summarized in Algorithm 1.
105
By assumption 1) inTheorem 1.3, the number of columns of A and B cannot exceed
106
the number of rows, i.e.,P Mr≤ I andP Nr≤ J . To simplify the presentation we 107
assume inAlgorithm 1 that P Mr= I and P Nr= J , implying that A and B are 108
square nonsingular matrices and T ∈ FP Mr×P Nr×K. Since, by Lemma 2.1 below,
109
T is ML rank-(P Mr,P Nr, ·) tensor, such a compression can be done, for instance, 110
using ML Singular Value Decomposition [5]. Note that Algorithm 1 also recovers
111
the dimensions of the blocks, i.e., the values Mr and Nr. Indeed, in steps 1–4 of 112
Algorithm 1we construct matrices UI,1, . . . , UI,R∈ FI×I and UJ,1, . . . , UJ,R∈ FJ ×J 113
that can be simultaneously diagonalized by A and B−T, respectively. In steps 5 of 114
Algorithm 1we use the fact that for each r the spectra of UI,rand UJ,r relate to each 115
other as in (12)–(13). Thus, if the values α1, . . . , αR are generic, then the distinct 116
eigenvalues of the matrices F = α1UI,1+· · ·+αRUI,Rand G = α1UJ,1+· · ·+αRUJ,R 117
are the same and their multiplicities are M1, . . . , MR and N1, . . . , NR respectively. 118
Hence, the values Mr, Nr and the corresponding blocks Ar, Br can be recovered 119
from the EVDs of F and G.
120
The following corollary completes the results on uniqueness and computation of
121
the ML rank-(1, Nr, Nr) decomposition in [3,4,10]. 122
Corollary 1.4. Let T admit the ML rank-(1, Nr, Nr) decomposition (3), the 123
matrices A, B and D1, . . . , DK be defined in (4)and (6), respectively, and let 124 (14) Cr:= [DTr,1 . . . D T r,K] T ∈ FK×Nr, r = 1, . . . , R. 125 Assume that 126
1. A and B have full column rank and
127
2. r[Ci Cj] ≥ max(Ni, Nj) + 1 for all 1 ≤ i < j ≤ R.
128
Then the ML rank-(1, Nr, Nr) decomposition of T is unique and can be computed by 129
means of CS-EVD.
130
Proof. Seesection 2.
131
The following theorem contains bounds on the tensor dimensions that
guaran-132
tee that assumptions 1) and 2) in Theorem 1.3 hold for generic matrices A, B and
133
D1, . . . , DK. 134
Algorithm 1 Computation of BTD(3)by CS-EVD (seeTheorem 1.3) Require: tensor T ∈ FI×J ×K such that I =P Mr and J =P Nr
1: Construct the IJ K-by-(I2+ J2) matrix MT as
(11) MT = TT 1 ⊗ II −IJ⊗ T1 .. . ... TT K⊗ II −IJ⊗ TK
2: Compute ur= [u1,r . . . uI2+J2,r]T, r = 1, . . . , R that form a basis of Null (MT) 3: For r = 1, . . . , R reshape [u1,r . . . uI2,r]T into the I × I matrix UI,r
4: For r = 1, . . . , R reshape [uI2+1,r . . . uI2+J2,r]T into the J × J matrix UJ,r
5: Solve CS-EVD problem
UI,r= A blockdiag(g1rIM1, . . . , gRrIMR)A −1, (12) UJ,r = B−Tblockdiag(g1rIN1, . . . , gRrINR)B T, r = 1, . . . , R (13)
with respect to the matrices A, B and the values M1, . . . , MR, N1, . . . , NR 6: Solve(7)with respect to D1, . . . , DK
7: For r = 1, . . . , R stack Dr,1, . . . , Dr,K into the Mr× Nr× K tensor Dr 8: return Ar∈ FI×Mr, Br∈ FJ ×Nr and Dr∈ FMr×Nr×K such that(3)holds
Theorem 1.5. Let T be a sum of R generic ML rank-(Mr, Nr, ·) terms. Assume 135 that 136 (15) I ≥XMr, J ≥ X Nr, K ≥ max max Mr Nr + max Nr Mr , 3 , 137
where dxe denotes the smallest integer not less than x. Then the decomposition of T
138
into a sum of ML rank-(Mr, Nr, ·) terms is unique and can be computed by means of 139
CS-EVD.
140
Proof. Seesection 3.
141
The following corollaries easily follow fromTheorem 1.5.
142
Corollary 1.6. Let T be a sum of R1 ≥ 1 generic ML rank-(1, Nr, Nr) terms
and R2≥ 1 generic ML rank-(Mr, 1, Mr) terms. Assume that
I ≥ R1+
X
Mr, J ≥ R2+
X
Nr, and K ≥ max Mr+ max Nr.
Then the decomposition of T into a sum of R1ML rank-(1, Nr, Nr) terms and R2 ML 143
rank-(Mr, 1, Mr) terms is unique and can be computed by means of CS-EVD. 144
Corollary 1.7. Let T be a sum of R generic ML rank-(1, Nr, Nr) terms. As-145
sume that
146
(16) I ≥ R, J ≥XNr, and K ≥ max Nr+ 1. 147
Then the decomposition of T into a sum of ML rank-(1, Nr, Nr) terms is unique and 148
can be computed by means of CS-EVD.
Corollary 1.8. Let T be a sum of R generic ML rank-(M, N, ·) terms. Assume that I ≥ RM, J ≥ RN, K ≥ max( M N + 1, N M + 1, 3).
Then the decomposition of T into a sum of ML rank-(M, N, ·) terms is unique and
150
can be computed by means of CS-EVD.
151
The bounds inCorollary 1.6,Corollary 1.7, andCorollary 1.8are new and guarantee
152
that a generic ML rank (Mr, Nr, ·) decomposition can be computed by Algorithm 1. 153
The bound on K in(16)complements the bounds for generic uniqueness of the ML
154
rank-(1, N, N ) decomposition in [3, Section 4] and is more relaxed than the bound for
155
generic uniqueness of ML rank-(1, Nr, Nr) decomposition in [9, Theorem 1.7 (E)]. 156
In the remaining part of this section we apply Theorem 1.3 to (generic)
con-157
strained JBD problems in Table 1. Uniqueness of the constrained JBD problem
Table 1
Some examples of constrained JBD problems
Type of JBD problem Constraints in(7) F
JBD by congruence B = A R or C
symmetric JBD B = A and Dk = DTk, k = 1, . . . , K R or C
JBD by ∗-congruence B = A∗ C
Hermitian JBD B = A∗ and Dk= DHk , k = 1, . . . , K C 158
means uniqueness of the corresponding constrained ML rank-(Nr, Nr, 1) decomposi-159
tion. Generic uniqueness of the constrained JBD problem can be defined in a similar
160
way as inDefinition 1.2. By way of example, we resort to the following definition of
161
generic uniqueness of the symmetric JBD problem.
162
Definition 1.9. Let Lr denote a linear bijection between FK
Nr (Nr +1)
2 and the
163
subspace of Nr× Nr× K tensors whose frontal slices are symmetric matrices. Let 164
also µ be a measure on FIP Nr+KPNr (Nr +1)2 absolutely continuous with respect to the
165
Lebesgue measure. A solution of the generic symmetric JBD problem is called unique
166 if 167 168 µ{(A1, . . . , AR, d1, . . . , dR) : 169
the constraint decomposition of T =
R X r=1 Lr(dr) •1Ar•2Ar is not unique} = 0. 170 171
Theorem 1.5 guarantees that if K ≥ 3 and the diagonal blocks Dr,k in JBD 172
problem (7) are square and generic, then the decomposition is unique and can be
173
computed by means of CS-EVD. The following theorem states that the bound K ≥
174
3 remains valid for all constrained JBD problems in Table 1 as well and that the
175
computation relies on S-EVD.
176
Theorem 1.10. Assume that I ≥ P Nr, and K ≥ 3. Then the solution of a 177
generic constrained JBD problems inTable 1is unique and can be computed by means
178
of S-EVD.
179
Steps 1–4 and modified step 5 of Algorithm 1 constitute the algebraic procedure
180
related toTheorem 1.10. In step 5, because of the constraints B = A and B = A∗,
181
the CS-EVD problem(12)–(13)turns into the S-EVD problem. It can be shown that
in the case of the JBD by congruence problem the proposed reduction to S-EVD
183
essentially coincides with the one in [1, Subsection 2.3].
184
Proof. Seesection 3.
185
2. Proof ofTheorem 1.3andCorollary 1.4. We need the following lemmas.
186 187
Lemma 2.1. Let the conditions of Theorem 1.3 hold and let U1 and U2 denote 188
the “U” factors in the compact SVDs of T(1) and T(2), respectively. Then 189
1. tensor T ∈ FI×J ×K is ML rank-(P M
r,P Nr, ·); 190
2. the “compressed” matrices Ac := UH
1A and Bc := UH2 B are square and 191
nonsingular;
192
3. A = U1Ac and B = U2Bc; 193
4. the “compressed” tensor Tc := T •
1UH1 •2 UH2 ∈ FP Mr×P Nr×K is ML 194
rank-(P Mr,P Nr, ·); 195
5. tensor Tc admits the ML rank-(M
r, Nr, ·) decomposition Tc = R X r=1 Dr•1(UH1 Ar) •2(UH2Br),
where the tensors Dr∈ FMr×Nr×K are the same as in decomposition (3). 196
Proof. 1. First we show that the matrices
197
(17) D(1)= [D1 . . . DK] and D(2)= [DT1 . . . DTK] have full row rank. 198
Assume that fTD
(1) = 0. Then 1fTDk = DkOP Nr, where 1 denotes theP Mr× 1
199
vector of ones. Hencef ⊗ 1 0
∈ Null (MD). Since dim Null (MD) = R, it follows that 200
the rank-1 matrix 1fT is a linear combination of the matrices P
I,1, . . . , PI,R. Hence 201
f = 0. Thus D(1) has full row rank. In a similar way one can prove that D(2) has 202
also full row rank. Now, from assumption 1 inTheorem 1.3,(17), and the identities
203
(18) T(1)= AD(1)blockdiag(BT, . . . , BT), T(2) = BD(2)blockdiag(AT, . . . , AT) 204
it follows that
205
(19) col(T(1)) = col(A) and col(T(2)) = col(B), 206
where col(·) denotes the column space of a matrix. Hence rT(1) = rA = P Mr
207
and rT(2) = rB = P Nr, implying that the tensor T ∈ F
I×J ×K is ML rank-208
(P Mr,P Nr, ·). 209
2. and 3. follow from (19)and the fact that U1 and U2 are the “U” factors in 210
the compact SVDs of T(1) and T(2), respectively. 211
4. From the construction of Tc
(1) and(18)it follows that 212 213 (20) Tc(1) = [UH1T1U∗2 . . . U H 1 TKU∗2] = U H 1T(1)blockdiag(U∗2, . . . , U ∗ 2) = 214 UH1 AD(1)blockdiag((UH2B) T, . . . , (UH 2B) T) = 215 AcD(1)blockdiag((Bc)T, . . . , (Bc)T). 216 217
Hence, by statement 2. and (17), rTc
(1) = rAc = P Mr. In a similar way one can
218 prove that rTc (2) = rBc=P Nr. Thus, T c is ML rank-(P M r,P Nr, ·). 219
5. Follows from(20)and 2.
Lemma 2.2. Let I = P Mr, J = P Nr, A and B be nonsingular, and let the 221
matrices MD and MT be defined in (8) and (11), respectively. Then 222
1) the matrices MD and MT are related by 223 MT = blockdiag(B ⊗ A, . . . , B ⊗ A) · MD· AT ⊗ A−1 O O B−1⊗ BT ; 224 225
2) the matrices X ∈ FI×I
and Y ∈ FJ ×J satisfy the equations XD
k = DkY, 226 k = 1, . . . , K if and only if MD vec(X) vec(Y) = 0; 227
3) the matrices X ∈ FI×I
and Y ∈ FJ ×J satisfy the equations XT
k = TkY, 228 k = 1, . . . , K if and only if MT vec(X) vec(Y) = 0; 229 4) Null (MD) ⊇ span ( vec(PI,r) vec(PJ,r) : r = 1, . . . , R )
, where the projectors PI,r 230
and PJ,r are defined in (10); 231 5) Null (MT) = A−T ⊗ A O O B ⊗ B−T Null (MD); 232
6) dim Null (MT) = dim Null (MD) ≥ R; 233
7) if dim Null (MD) = R, then the matrices 234
[AD1BT . . . ADKBT] ∈ FI×J K and [BDT1A
T . . . BDT KA T ] ∈ FJ ×IK 235 236
have full row rank.
237
Proof. 1) follows from(7), (8), and(11); 2)–4) are trivial; 5) follows from 1); 6)
238
follows from 5) and 4); 7) follows fromLemma 2.11. and(18).
239
Now we are ready to proveTheorem 1.3.
240
Proof of Theorem 1.3. ByLemma 2.1, WLOG, we can assume that I = P Mr 241
and J =P Nr. 242
1) First we show that decomposition (3) can be computed algebraically. From assumption 2) andLemma 2.2 4), 6) we have that
Null (MD) = span ( vec(PI,r) vec(PJ,r) : r = 1, . . . , R )
and dim Null (MT) = R.
Hence, byLemma 2.2 5), 243 (21) Null (MT) = A−T ⊗ A O O B ⊗ B−T span ( vec(PI,r) vec(PJ,r) : r = 1, . . . , R ) . 244
Let the columns of U ∈ F(I2+J2)×R
form a basis of Null (MT). We represent U as 245
(22) U =vec(UI,1) . . . vec(UI,R) vec(UJ,1) . . . vec(NJ,R)
,
246
where UI,r ∈ FI×I and UJ,r ∈ FJ ×J, r = 1, . . . , R. From (21) and (22), it follows 247
that there exists a unique nonsingular R × R matrix G such that
248
vec(UI,1) . . . vec(UI,R) = A−T ⊗ A [vec(PI,1) . . . vec(PI,R)]G, 249
vec(UJ,1) . . . vec(UJ,R) = B ⊗ B−T [vec(PJ,1) . . . vec(PJ,R)]G. 250
Hence, (12)–(13) hold. Since the matrices A, B and G are nonsingular, it follows
252
that the matrices UI,1, . . . , UI,R are linearly independent. Thus, we can recover A 253
from(12). Similarly, we can recover B from(13).
254
2) Now we show that decomposition(3)is unique. Assume that(3)holds for Ar, 255
Br, etc. replaced by eAr, eBr, etc. To prove the uniqueness it is sufficient to show that 256
e
A, eB and M
e
D satisfy assumptions 1) and 2), i.e., that eA and eB are nonsingular and 257 dim Null M e D = R. By (7), 258 [AD1BT . . . ADKBT] =[T1 . . . TK] = [ eA eD1BeT . . . eA eDKBeT] = eA[ eD1BeT . . . eDKBeT], (23) 259 [BDT1AT . . . BDTKAT] =[TT1 . . . TTK] = [ eB eDT1AeT . . . eB eDTKAeT] = eB[ eDT1AeT . . . eDTKAeT]. (24) 260 261
Since, byLemma 2.27), the matrices in the left-hand side of (23)and(24)have full
262
row rank, it follows that the matrices eA and eB are nonsingular. Since, by assumption
263
2), dim Null (MD) = R, and, byLemma 2.25), 264 Null (MT) = A−T ⊗ A O O B ⊗ B−T Null (MD) = " e A−T ⊗ eA O O B ⊗ ee B−T # Null M e D , 265
it follows that dim Null M
e D = R. 266
Proof of Corollary 1.4. It is sufficient to prove that assumption2inCorollary 1.4
267
implies assumption2 in Theorem 1.3. Thus, we need to show that the subspace of
268
solution of (9)is spanned by the trivial solutions (X, Y) = (PI,r, PJ,r), r = 1, . . . , R, 269
where PI,r and PJ,r are defined in(10)and M1= · · · = MR= 1. 270
Let X = (xij)Ri,j=1 and let Y = (Yij)i,j=1R consists of the blocks Yij ∈ FNi×Nj. 271
Since the matrices Dk are block-diagonal, we can rewrite (9)as 272 xijDj,k= Di,kYij, 1 ≤ i, j ≤ R, k = 1, . . . , K, 273 which, by(14), is equivalent to 274 (25) xijCj = CiYij, 1 ≤ i, j ≤ R. 275
We need to show that
276 Yii = xiiINi, 1 ≤ i ≤ R, (26) 277 Yij = O, xij = 0, 1 ≤ i 6= j ≤ R. (27) 278 279
(i) Let us prove(26). By(25),
280
(28) Ci(xiiINi− Yii) = O, 1 ≤ i ≤ R.
281
Since the 1 × Ni× K tensor Di in(3) is ML rank-(1, Ni, Ni) and the second matrix 282
unfolding of Di coincides with CTi , it follows that CTi has full row rank or that Ci 283
has full column rank. Hence(26)follows from (28).
(ii) Let us prove(27). It is clear that(25)can be rewritten as 285 (29) O = [Ci Cj] Yij xijINj . 286
Assume to the contrary that xij 6= 0. Then from (29), Sylvester’s rank inequality,
and assumption2 it follows that 0 ≥ r[Ci Cj]+ r[YT
ij xijINj]T − (Ni+ Nj) ≥ max(Ni, Nj) + 1 + Nj− (Ni+ Nj) ≥ 1, which is a contradiction. Thus, xij = 0. Hence, by (25), CiYij = O. Since Ci has 287
full column rank, it follows that Yij = O. 288
3. Proofs of Theorem 1.5 and Theorem 1.10 . We need the following two
289 lemmas. 290 Lemma 3.1. Let 291 (30) K = max( M N + 1, N M + 1, 3). 292
Then there exist matrices E1, . . . , EK ∈ FM ×N such that the system of matrix equa-293
tions
294
(31) XEk= EkY, X ∈ FM ×M, Y ∈ FN ×N k = 1, . . . , K 295
has only the trivial solutions X = cIM and Y = cIN, where c ∈ F. 296
Proof. It is clear that by taking the transpose of (31) we obtain an equivalent
297
problem of the form (31) in which M and N are switched. Thus, w.l.o.g., we can
298
always assume that M ≥ N . We consider three cases: M = N , N < M ≤ 2N , and
299
2N < M .
300
Case M = N . By assumption(30), K = 3. Let a ∈ FN be a vector with distinct 301
entries and let D ∈ FN ×N be a matrix with nonzero entries. We show that if E1= IN, 302
E2= diag(a), and E3= D, then(31)has only trivial solutions. 303
The equation XE1= E1Y implies that X = Y. Hence,
Y diag(a) = YE2= XE2= E2Y = diag(a)Y.
Since the entries of a are distinct, it follows that Y is a diagonal matrix, Y = diag(y1, . . . , yN). Finally, since the entries of E3= D are nonzero, the equation
diag(y1, . . . , yN)D = YE3= XE3= E3Y = D diag(y1, . . . , yN)
implies that y1= · · · = yN =: c. Thus, X = Y = cIN . 304
Case N < M ≤ 2N . By assumption(30), K = 3. Let a ∈ FN be a vector with
distinct entries, D ∈ F(M −N )×(2N −M )be a full row rank matrix with nonzero entries,
and let bIM −N denote the matrix formed by the first M − N rows of IN. We show
that if E1= IN O , E2= O D , E3= diag (a) bIM −N ,
then(31)has only trivial solutions. Let X =X1 X3 X2 X4
with X1∈ FN ×N, X2, XT3 ∈ 305
F(M −N )×N and X4 ∈ F(M −N )×(M −N ). Then the equations XE1 = E1Y, XE2 = 306
E2Y, and XE3= E3Y imply 307 X1= Y and X2= 0, (32) 308 X3D = O and X4D = DY, 309
X1diag (a) + X3bIM −N = diag (a)Y and X2diag (a) + X4bIM −N = bIM −NY,
(33)
310 311
respectively. Since X3D = O and D has full row rank, it follows that X3= O. Hence,
by(32)and(33),
Y diag (a) = X1diag (a) + X3bIM −N = diag (a)Y.
Since all entries of a are distinct, it follows that Y is a diagonal matrix, Y =
312
diag(y1, . . . , yN). From the second identities in (32) and (33) it follows that X4 = 313
diag(y1, . . . , yM −N). Finally, since the entries of D are nonzero, it follows from X4D = 314
DY that y1= · · · = yN =: c. Hence, Y = cIN and X = blockdiag(X1, X4) = cIM. 315
Case 2N < M . By assumption (30), K =M
N + 1. Let r = M − (K − 2)N . 316
Then 0 < r ≤ N . We define E1, . . . , EK−2 ∈ FM ×N and bEK−1∈ FM ×r as blocks in 317
the partition IM = [E1 . . . EK−2 EbK−1] and set EK−1= [ bEK−1 O], where the zero 318
matrix has N − r columns. We also set Ek =
diag(a) D O , where a ∈ FN is a vector 319
with distinct entries and D ∈ FN ×N is a matrix with nonzero entries. We show that 320
for such choice of the matrices Ek system(31) has only the trivial solution. From 321
the first K − 1 equations in(31)it follows that X = blockdiag(Y, . . . , Y, bY), where
322
b
Y denotes a matrix located at the intersection of the first r rows and r columns of
323
Y. The remaining equation XEK= EKY in(31)implies that Y diag(a) = diag(a)Y 324
and YD = DY. Now, in exactly the same way as “Case M = N ” one can obtain
325
that Y = cIN for some c ∈ F. Hence, X = cIM. 326 Lemma 3.2. Let 327 (34) K = 2, if M1> N1 and M2< N2, dM2/N2e + 1, if M1> N1 and M2≥ N2, dN1/M1e + 1, if M1≤ N1 and M2< N2, dM2/N2e + dN1/M1e , if M1≤ N1 and M2≥ N2. 328
Then there exist matrices E1, . . . , EK∈ FM2×N2 and F1, . . . , FK ∈ FM1×N1 such that 329
the system of matrix equations
330
(35) XEk= FkY, X ∈ FM1×M2, Y ∈ FN1×N2 k = 1, . . . , K 331
has only the zero solution.
332
Proof. We consider each of the four cases in(34)separately.
333
Case M1> N1 and M2< N2. By(34), K = 2. By definition put
E1= O, E2= [IM2 O], F1= IN1
O
, E2= O.
Then the equations XE1 = F1Y and XE2 = F2Y imply that Y = O and X = O, 334
respectively.
Case M1> N1 and M2≥ N2. By(34), K = dM2/N2e + 1. Let E1, . . . , EK−1 be 336
matrices such that [E1 . . . EK−1] has full row rank, let EK be an arbitrary matrix, 337
F1, . . . , FK−1be zero matrices, and FK=
IN1 O
. Then from the first K − 1 equations
338
in (35) it follows that X[E1 . . . EK−1] = 0. Hence X = O. Now the remaining 339
equation XEK = FKY implies that Y = O. 340
Case M1 ≤ N1 and M2 < N2. By (34), K = dN1/M1e + 1. The problem can 341
be reduced to the setting used in the previous case by taking the transpose of the
342
equations in(35).
343
Case M1 ≤ N1 and M2 ≥ N2. By (34), K = dM2/N2e + dN1/M1e. Let
E1, . . . , EdM2/N2e be matrices such that [E1 . . . EdM2/N2e] has full row rank and let EdM2/N2e+1= EK = O. Let also F1= · · · = FdM2/N2e = 0 and FdM2/N2e+1, . . . , FK be matrices such that [FT
dM1/N2e+1 . . . F
T K]
T has full column rank. Then from the
first dM2/N2e and the last dN1/M1e equations in(35)it follows that
X[E1 . . . EdM2/N2e] = O and O = [F T dM1/N2e+1 . . . F T K] TY,
respectively. Hence, X = O and Y = O.
344
Now we are ready to proveTheorem 1.5.
345
Proof of Theorem 1.5. We show that assumptions 1) and 2) inTheorem 1.3hold
346
for generic matrices A, B, and D1, . . . , DK. The first two inequalities in(15)express 347
the fact that the number of columns of A and B cannot exceed the number of rows.
348
Since A and B are generic this means that assumption 1) inTheorem 1.3 holds. To
349
prove that assumption 2) inTheorem 1.3 also holds it is sufficient to show that the
350
subspace of solutions of (9)has dimension R. In the remaining part of the proof we
351
show that for generic matrices D1, . . . , DK the subspace of solutions of (9)is spanned 352
by (PI,1, PJ,1), . . . , (PI,R, PJ,R), where the projectors PI,r, PJ,r are defined in(10). 353
Let X = (Xij)Ri,j=1 and Y = (Yij)i,j=1R consist of the blocks Xij ∈ FMi×Mj and 354
Yij ∈ FNi×Nj, respectively. Since the matrices Dkare block-diagonal, we can rewrite 355
(9)in terms of the blocks Xij and Yij as 356
(36) XijDj,k= Di,kYij, k = 1, . . . , K, 357
where i, j ∈ {1, . . . , R}. Let the subsets
Ωij, Ωi⊆ FM1×N1× · · · × FMR×NR × · · · × FM1×N1× · · · × FMR×NR
consist of the elements (D1,1, . . . , DR,1, . . . , D1,K, . . . , DR,K) and be defined as fol-358
lows: the elements in Ωij (i 6= j) are such that (36)has only the zero solution and 359
the elements in Ωi are such that (36)has only the trivial solution, i.e., Xii = ciIMi
360
and Yii= ciINi for some ci∈ F.
361
To prove that (X, Y) belongs to the linear subspace spanned by (PI,1, PJ,1), . . . , 362
(PI,R, PJ,R) we have to show that (∩Ωij)T(∩Ωi) is a set of full measure. Since the 363
intersection of a finite number of sets of full measure is also a set of full measure it
364
is sufficient to prove that all sets in the intersection have full measure. This can be
365
done as follows.
366
1) Let i 6= j. W.l.o.g. we can assume that i = 1 and j = 2. Then, by(36),
367
X12D2,k= D1,kY12, k = 1, . . . , K.
(37)
368 369
It is clear that system(37)has only the zero solution if and only if the matrix M = DT 2,1⊗ IM1 − IN2⊗ D1,1 .. . DT 2,K⊗ IM1 − IN2⊗ D1,K
has full column rank. Hence, from Lemma 3.2 and assumption (15) it follows that
370
there exist matrices D1,1, . . . , D1,K and D2,1. . . D2,K such that M has full column 371
rank. Now we assume that D1,1, . . . , D1,K and D2,1. . . D2,K in (37) are generic. 372
Then, by [6, Lemma 6.3], M has full column rank, implying that system(37)has only
373
the zero solution. Hence, by Fubini’s theorem [7, Theorem C, p.148], Ω12 is a set of 374
full measure.
375
2) Let i = j. W.l.o.g. we can assume that i = 1. We have to prove that the
376
system
377
(38) X11D1,k= D1,kY11, k = 1, . . . , K, 378
has only the trivial solution X11 = cIM1 and Y11 = cIN1, where c ∈ F. It is clear that system(38)has only the trivial solution if and only if the dimension of the null space of the matrix
M = DT 1,1⊗ IM1 − IN1⊗ D1,1 .. . DT 1,K⊗ IM1 − IN1⊗ D1,1
is one. Hence, fromLemma 3.1and assumption(15)it follows that there exist matrices
379
D1,1, . . . , D1,K such that M12+ N12− 1 columns of M are linearly independent. Now 380
we assume that D1,1, . . . , D1,K are generic. Then, by [6, Lemma 6.3], the same 381
M12+ N12− 1 columns of M are linearly independent, implying that system(38)has
382
only the trivial solution. Hence, by Fubini’s theorem [7, Theorem C, p.148], Ω11 is a 383
set of full measure.
384
Proof of Theorem 1.10. We show that assumptions 1) and 2) inTheorem 1.3hold
385
for generic matrix A, B, and D1, . . . , DK, where B = A or B = A∗ and D1, . . . , DK 386
may be symmetric, Hermitian or unconstrained.
387
1) Since A is generic and I ≥ P Nr, it follows that A and B has full column 388
rank.
389
2) WLOG we assume that K = 3. We proceed as in the proof ofTheorem 1.5.
390
First we show that the subspace of the solutions of (9) has dimension R for a
spe-391
cific choice of D1, D2, and D3. Let D1 = IP Nr, D2 be a diagonal matrix with
392
distinct real values on the main diagonal and let D3 = blockdiag(D1,3, . . . , DR,3), 393
where the entries of D1,3, . . . , DR,3 are nonzero. Then XD1 = D1Y implies that 394
X = Y. From XD2 = D2Y = D2X, it follows that X = diag(x1, . . . , xP Nr).
395
Finally, from diag(x1, . . . , xP Nr)D3 = XD3 = D3Y = D3diag(x1, . . . , xP Nr) it
396
follows that X = Y = blockdiag(c1IN1, . . . , cRINR) for some c1, . . . , cR ∈ F. Thus,
397
assumption 2) in Theorem 1.3holds for a specific choice of D1, D2, and D3. Hence 398
some 2(P Nr)2− R columns of MD are linearly independent. From [6, Lemma 6.3] 399
it follows that if D1, D2, and D3 are generic symmetric, Hermitian or unconstrained 400
matrices, then the same 2(P Nr)2− R columns of MD are linearly independent, 401
implying that dim Null (MD) = R. 402
REFERENCES 403
[1] Y. Cai and C. Liu, An algebraic approach to nonorthogonal general joint block diagonaliza-404
tion, SIAM J. Matrix Anal. Appl., 38(1) (2017), pp. 50–71. 405
[2] G. Chabriel, M. Kleinsteuber, E. Moreau, H. Shen, P. Tichavsky, and A. Yere-406
dor, Joint matrices decompositions and blind source separation: A survey of methods, 407
identification, and applications, IEEE Signal Processing Magazine, 31 (2014), pp. 34–43. 408
[3] L. De Lathauwer, Decompositions of a higher-order tensor in block terms — Part II: Defi-409
nitions and uniqueness, SIAM J. Matrix Anal. Appl., 30 (2008), pp. 1033–1066. 410
[4] L. De Lathauwer, Blind separation of exponential polynomials and the decomposition of a 411
tensor in rank-(Lr, Lr, 1) terms, SIAM J. Matrix Anal. Appl., 32 (2011), pp. 1451–1474.
412
[5] L. De Lathauwer, B. De Moor, and J. Vandewalle, A multilinear singular value decom-413
position, SIAM J. Matrix Anal. Appl., 21 (2000), pp. 1253–1278. 414
[6] I. Domanov and L. De Lathauwer, On the uniqueness of the canonical polyadic decompo-415
sition of third-order tensors — Part II: Overall uniqueness, SIAM J. Matrix Anal. Appl., 416
34 (2013), pp. 876–903. 417
[7] P. R. Halmos, Measure theory, Springer-Verlag, New-York, 1974. 418
[8] T. Maehara and K. Murota, Algorithm for error-controlled simultaneous block-419
diagonalization of matrices, SIAM J. Matrix Anal. Appl., 32 (2011), pp. 605–620. 420
[9] Y. Ming, On partial and generic uniqueness of block term tensor decompositions, Annali 421
Dell’Universita’Di Ferrara, 60 (2014), pp. 465–493. 422
[10] M. Sørensen and L. De Lathauwer, Coupled canonical polyadic decompositions and (cou-423
pled) decompositions in multilinear rank-(Lrn, Lrn, 1) terms—Part I: Uniqueness, SIAM
424
J. Matrix Anal. Appl., 36 (2015), pp. 496–522. 425