Canonical Polyadic Decomposition of
Third-Order Tensors of Which a Mode Rank is
Equal to the Rank
∗
Lieven De Lathauwer
December 1, 2011
1
Introduction
We consider the Canonical Polyadic Decomposition (CPD) of a rank-R tensor T ∈ KI1×I2×I3, with R = rank 3(T ) 6 I3: T = R X r=1 ar ⊗br ⊗cr. (1) Contributions:
• A new, relaxed, algebraic condition on ar, br, 1 6 r 6 R that is
sufficient for essential uniqueness of the CPD.
• Under this condition the CP factors can be obtained by means of stan-dard linear algebra (set of linear equations and matrix Eigenvalue De-composition (EVD)).
∗Available in SISTA publication engine as: L. De Lathauwer, CPD Study, Part III,
• Algorithm for the estimation of the CP factors in the case of an ap-proximate CPD.
• Generic uniqueness for a given set (I1, I2, R) is established by checking
the algebraic condition for one random sample. • Verification for (3 × 7 × 12) tensors of rank 12.
• Similar approach for ar = br, 1 6 r 6 R. (Generic bound will be
different.)
Preceding work. In matrix format we have:
T[1,2;3] = (A ⊙ B) · CT. (2)
A necessary condition for uniqueness of the CPD is that A ⊙ B has full column rank. Our mode-3 rank assumption implies that rank(C) = R. In [1] we showed that (1) can be obtained by means of standard linear algebra (set of linear equations and matrix Eigenvalue Decomposition (EVD)) if the tensors in the set U = {Φ(as1b
T s1, as2b
T
s2)|1 6 s1 < s2 6 R} are linearly
independent. In the case of an approximate decomposition, after solving a set of linear equations in least-squares sense, the computation of (1) can be reduced to the computation of a CPD of the form
M =
R
X
r=1
wr ⊗wr ⊗lr, (3)
in which M ∈ KR×R×R and in which W ∈ KR×R and L ∈ KR×R are
nonsin-gular. Taking into account symmetries [2], the tensors in the set U can be represented by the columns of a matrix U ∈ KCI12 C
2 I2×C
2 R
, which needs to have full rank. Consequently, in this approach R is bounded by C2
R 6CI21C
2 I2. In
the next section we propose a technique that is less restrictive.
2
Symmetrized mapping
Let F ∈ KI1I2×R
and G ∈ KI3×R be such that
is a rank-revealing decomposition of T[1,2;3]. Then we have:
F = (A ⊙ B) · W−1
, (5)
GT = W · CT, (6)
in which W ∈ KR×R is nonsingular.
Theorem 1 Consider the mapping Φ(3) : (X(r1), X(r2), X(r3)
) ∈ KI×J ×
KI×J × KI×J → Φ(3)(X(r1), X(r2), X(r3)) ∈ KI×I×I×J ×J ×J defined by
(Φ(3)(X(r1), X(r2), X(r3))) i1i2i3j1j2j3 = P σ∈P(r1,r2,r3)(x (σ(r1)) i1j1 x (σ(r2)) i2j2 − x (σ(r1)) i1j2 x (σ(r2)) i2j1 )x (σ(r3)) i3j3 ,
in which P(r1, r2, r3) is the set of permutations of (r1, r2, r3). Then Φ(3)(X, X, X) =
O if and only if X has at most rank 1.
Note that Φ(3)(·, ·, ·) is symmetric in its arguments. Define
Pr1r2r3 = Φ (3)((f r1)[1;2],(fr2)[1;2],(fr3)[1;2]), 1 6 r1 6r2 6r3 6R. (7) We have that Pr1r2r3 = R X s1,s2,s3=1 (W−1 )s1r1(W −1 )s2r2(W −1 )s3r3Φ (3)(a s1b T s1, as2b T s2, as3b T s3). (8) We look for symmetric tensors X ∈ KR×R×R that satisfy
R X r1,r2,r3=1 xr1r2r3Pr1r2r3 = O. (9) We stack Pr1r2r3, 1 6 r1 6 r2 6 r3 6 R, in a matrix ¯P ∈ K I3 1I 3 2×C 3 R+2.
We determine the vectors x in the null space of ¯P, which can then be mapped to tensors X . We make the crucial assumption that the set of tensors {Φ(3)(a s1b T s1, as2b T s2, as3b T s3)|1 6 s1 < s2 6 s3 6 R} is linearly independent.
Combination of (7)–(9) shows that, under this assumption, the null space of ¯P has dimension R. Moreover, the vectors in the null space represent
tensors that are linear combinations of wr ⊗wr ⊗wr, 1 6 r 6 R. If we stack
the tensors obtained from a basis of the null space of ¯P in a super-tensor Y ∈ KR×R×R×R, then this super-tensor admits the CPD
Y =
R
X
r=1
wr ⊗wr ⊗wr ⊗vr. (10)
The vectors vr, 1 6 r 6 R, are linearly independent since they generate a
basis of the null space of ¯P. Since additionally W is nonsingular, CPD (10) is unique. In the case of an exact decomposition, (10) follows from a matrix EVD [3, 4]. In the case of an approximate decomposition, the EVD may be used to initialize an algorithm for fitting (10).
Remark 1. Instead of working with ¯P∈ KI13I 3 2×C
3
R+2, one may also estimate
the null space of ¯PH · ¯P∈ KC3 R+2×C
3
R+2. The latter matrix can be computed
efficiently by computing (fr1⊗fr2⊗fr3) H(f s1⊗fs2⊗fs3) as (f H r1fs1)(f H r2fs2)(f H r3fs3).
The disavantage is that the condition number is squared.
Example 1. Consider tensors of dimension (3 × 7 × R) and rank R. The number of free parameters in the CPD equals (3+7+R−2)R = (8+R)R. The number of tensor entries is 21R. Consequently, for R = 13 there generically exist a finite number of decompositions. Moreover, we expect that for R 6 12 the CPD is generically unique. For R = 12 uniqueness is not guaranteed by the result in [1]. However, it can be verified that the 352 tensors in the set {Φ(3)(a s1b T s1, as2b T s2, as3b T
s3)|1 6 s1 < s2 6s3 612} are linearly independent
for a random example. Hence, for R = 12 the decomposition is generically unique and it can be computed as explained above.
Remark 2. It is possible to generalize the trilinear mapping Φ(3) to a
K-linear mapping Φ(K) that is supposed to yield a matrix ¯P ∈ KIK 1 I
K 2 ×C
K R+K−1
that has an R-dimensional null space.
(The trick here is that by increasing K we work with a matrix of which the number of rows increases faster than the number of columns. Since the columns have a multiplicative rather than a linear structure, this will hopefully lead to more relaxed conditions and algorithms.)
3
Alternative mapping
Theorem 2 Consider the mappingΨ(3) : (X, Y, Z) ∈ KI×J
×KI×J ×KI×J → Ψ(3)(X, Y, Z) ∈ KI×I×J ×J ×6 defined by (Ψ(3)(X, Y, Z)) i1i2j1j21 = xi1j1yi2j2 − xi1j2yi2j1, (Ψ(3)(X, Y, Z)) i1i2j1j22 = yi1j1xi2j2 − yi1j2xi2j1, (Ψ(3)(X, Y, Z)) i1i2j1j23 = xi1j1zi2j2 − xi1j2zi2j1, (Ψ(3)(X, Y, Z)) i1i2j1j24 = zi1j1xi2j2 − zi1j2xi2j1, (Ψ(3)(X, Y, Z)) i1i2j1j25 = yi1j1zi2j2 − yi1j2zi2j1, (Ψ(3)(X, Y, Z)) i1i2j1j26 = zi1j1yi2j2 − zi1j2yi2j1.
Then Ψ(3)(X, X, X) = O if and only if X has at most rank 1.
Define Pr1r2r3 = Ψ (3)((f r1)[1;2],(fr2)[1;2],(fr3)[1;2]), 1 6 r1 6r2 6r3 6R. (11) We have that Pr1r2r3 = R X s1,s2,s3=1 (W−1) s1r1(W −1) s2r2(W −1) s3r3Ψ (3)(a s1b T s1, as2b T s2, as3b T s3). (12) We look for tensors X ∈ KR×R×R that satisfy
R X r1,r2,r3=1 xr1r2r3Pr1r2r3 = O. (13) We stack Pr1r2r3, 1 6 r1, r2, r3 6 R, in a matrix ¯P ∈ K 6I2 1I 2 2×R 3 . We de-termine the vectors x in the null space of ¯P, which can then be mapped to tensors X . We make the crucial assumption that the set of tensors {Ψ(3)(a s1b T s1, as2b T s2, as3b T s3)|1 6 s1, s2, s3 6 R,¬(s1 = s2 = s3)} is linearly
independent. Combination of (11)–(13) shows that, under this assumption, the null space of ¯Phas dimension R. Moreover, the vectors in the null space represent tensors that are linear combinations of wr ⊗wr ⊗wr, 1 6 r 6 R.
If we stack the tensors obtained from a basis of the null space of ¯P in a super-tensor Y ∈ KR×R×R×R, then this super-tensor admits the CPD
Y =
R
X
r=1
wr ⊗wr ⊗wr ⊗vr. (14)
The vectors vr, 1 6 r 6 R, are linearly independent since they generate a
basis of the null space of ¯P. Since additionally W is nonsingular, CPD (14) is unique. In the case of an exact decomposition, (14) follows from a matrix EVD [3, 4]. In the case of an approximate decomposition, the EVD may be used to initialize an algorithm for fitting (14).
Remark 3. Since all tensors X are symmetric, we might as well have worked with a matrix ¯¯P ∈ K6I12I22×CR+23 in which the columns corresponding to the
same entry are added together. In other words, we might have symmetrized Ψ(3).
Remark 4. It seems that the mapping Ψ(3) does not yield a uniqueness
condition that is more relaxed than the one in [1]. However, Ψ(3) might
perhaps be used as a starting point for further research.
References
[1] L. De Lathauwer, “A Link between the Canonical Decomposition in Mul-tilinear Algebra and Simultaneous Matrix Diagonalization”, SIAM J. Ma-trix Anal. Appl., Vol. 28, No. 3, 2006, pp. 642–666.
[2] A link between the decomposition of a third-order tensor in rank-(L,L,1) terms and simultaneous matrix diagonalization, Tech. report.
[3] R.A. Harshman, Foundations of the PARAFAC procedure: Model and conditions for an“explanatory” multi-mode factor analysis, UCLA Work-ing Papers in Phonetics, 16 (1970), pp. 1–84.
[4] S.E. Leurgans, R.T. Ross, and R.B. Abel, A decomposition for three-way arrays, SIAM J. Matrix Anal. Appl., 14 (1993), pp. 1064–1083.