Tensors and Hypermatrices

15
15.1 Hypermatrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-2
15.2 Tensors and Multilinear Functionals . . . . . . . . . . . . . . . . . 15-6
15.3 Tensor Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-12
15.4 Border Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-15
15.5 Generic and Maximal Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-17
15.6 Rank-Retaining Decomposition . . . . . . . . . . . . . . . . . . . . . . 15-17
15.7 Multilinear Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-20
15.8 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-22
15.9 Hyperdeterminants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-25
Lek-Heng Lim 15.10 Odds and Ends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-28
University of Chicago References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-28

Most chapters in this handbook are concerned with various aspects and implications of lin-
earity; Chapter 14 and this chapter are unusual in that they are about multilinearity. Just as
linear operators and their coordinate representations, i.e., matrices, are the main objects of
interest in other chapters, tensors and their coordinate representations, i.e., hypermatrices,
are the main objects of interest in this chapter. The parallel is summarized in the following
schematic:

linearity → linear operators, bilinear forms, dyads → matrices

multilinearity → tensors → hypermatrices

Chapter 14, or indeed the monographs on multilinear algebra such as [Gre78, Mar23,
Nor84, Yok92], are about properties of a whole space of tensors. This chapter is about
properties of a single tensor and its coordinate representation, a hypermatrix.
The first two sections introduce (1) a hypermatrix, (2) a tensor as an element of a tensor
product of vector spaces, its coordinate representation as a hypermatrix, and a tensor as a
multilinear functional. The next sections discuss the various generalizations of well-known
linear algebraic and matrix theoretic notions, such as rank, norm, and determinant, to
tensors and hypermatrices. The realization that these notions may be defined for order-d
hypermatrices where d > 2 and that there are reasonably complete theories which parallel
and generalize those for usual 2-dimensional matrices is a recent one. However, some of these
hypermatrix notions have roots that go back as early as those for matrices. For example,
the determinant of a 2 × 2 × 2 hypermatrix can be found in Cayley’s 1845 article [Cay45];
in fact, he studied 2-dimensional matrices and d-dimensional ones on an equal footing. The
final section describes material that is omitted from this chapter for reasons of space.
In modern mathematics, there is a decided preference for coordinate-free, basis-independ-
ent ways of defining objects but we will argue here that this need not be the best strategy.
The view of tensors as hypermatrices, while strictly speaking incorrect, is nonetheless a very

15-1
15-2 Handbook of Linear Algebra

useful device. First, it gives us a concrete way to think about tensors, one that allows a
parallel to the usual matrix theory. Second, a hypermatrix is what we often get in practice:
As soon as measurements are performed in some units, bases are chosen implicitly, and the
values of the measurements are then recorded in the form of a hypermatrix. (There are of
course good reasons not to just stick to the hypermatrix view entirely.)
We have strived to keep this chapter as elementary as possible, to show the reader that
studying hypermatrices is in many instances no more difficult than studying the usual 2-
dimensional matrices. Many exciting developments have to be omitted because they require
too much background to describe.
Unless otherwise specified, everything discussed in this chapter applies to tensors or hy-
permatrices of arbitrary order d ≥ 2 and all may be regarded as appropriate generalizations
of properties of linear operators or matrices in the sense that they agree with the usual
definitions when specialized to order d = 2. When notational simplicity is desired and when
nothing essential is lost, we shall assume d = 3 and phrase our discussions in terms of
3-tensors. We will sometimes use the notation hni := {1, . . . , n} for any n ∈ N. The bases
in this chapter are always implicitly ordered according to their integer indices. All vector
spaces in this chapter are finite dimensional.
We use standard notation for groups and modules. Sd is the symmetric group of permu-
tations on d elements. An Sd -module means a C[Sd ]-module, where C[Sd ] is the set of all
formal linear combinations of elements in Sd with complex coefficients (see, e.g., [AW92]).
The general linear group of the vector space V is the group GL(V ) of linear isomorphisms
from V onto itself with operation function composition. GL(n, F ) is the general linear group
of invertible n × n matrices over F . We will however introduce a shorthand for products of
such classical groups, writing
GL(n1 , . . . , nd , F ) := GL(n1 , F ) × · · · × GL(nd , F ),
and likewise for SL(n1 , . . . , nd , F ) (where SL(n, F ) is the special linear group of n × n
matrices over F having determinant one) and U(n1 , . . . , nd , C) (where U(n, C) is the group
of n × n unitary matrices).
In this chapter, as in most other discussions of tensors in mathematics, we use ⊗ in
multiple ways: (i) When applied to abstract vector spaces U , V , W , the notation U ⊗V ⊗W
is a tensor product space as defined in Section 15.2; (ii) when applied to vectors u, v, w from
abstract vector spaces U , V , W , the notation u ⊗ v ⊗ w is a symbol for a special element of
U ⊗ V ⊗ W ; (iii) when applied to l-, m-, n-tuples in F l , F m , F n , it means the Segre outer
product as defined in Section 15.1; (iv) when applied to F l , F m , F n , F l ⊗ F m ⊗ F n means
the set of all Segre outer products that can be obtained from linear combinations of terms
like those in Eq. (15.3). Nonetheless, they are all consistent with each other.

15.1 Hypermatrices
What is the difference between an m × n matrix A ∈ Cm×n and a mn-tuple a ∈ Cmn ? The
immediate difference is a superficial one: Both are lists of mn complex numbers except that
we usually write A as a 2-dimensional array of numbers and a as a 1-dimensional array of
numbers. The more important distinction comes from consideration of the natural group
actions on Cm×n and Cmn . One may multiply A ∈ Cm×n on “two sides” independently by
an m×m matrix and an n×n matrix, whereas one may only multiply a ∈ Cmn on “one side”
by an mn × mn matrix. In algebraic parlance, Cm×n is a Cm×m × Cn×n -module whereas
Cmn is a Cmn×mn -module. This extends to any order-d hypermatrices (i.e., d-dimensional
matrices).
In Sections 15.3 to 15.9 we will be discussing various properties of hypermatrices and
tensors. Most of these are generalizations of well-known notions for matrices or order-2
Tensors and Hypermatrices 15-3

tensors. Since the multitude of indices when discussing an order-d hypermatrix can be
distracting, for many of the discussions we assume that d = 3. The main differences between
usual matrices and hypermatrices come from the transition from d = 2 to d = 3. An
advantage of emphasizing 3-hypermatrices is that these may be conveniently written down
on a 2-dimensional piece of paper as a list of usual matrices. This is illustrated in the
examples.

Definitions:
For n1 , . . . , nd ∈ N, a function f : hn1 i × · · · × hnd i → F is a (complex) hypermatrix, also called
an order-d hypermatrix or d-hypermatrix. We often just write ak1 ···kd to denote the value
f (k1 , . . . , kd ) of f at (k1 , . . . , kd ) and think of f (renamed as A) as specified by a d-dimensional
table of values, writing A = [ak1 ···kd ]kn11,...,k ,...,nd
d =1
, or A = [ak1 ···kd ].
The set of order-d hypermatrices (with domain hn1 i×· · ·×hnd i) is denoted by F n1 ×···×nd , and we
define entrywise addition and scalar multiplication: For any [ak1 ···kd ], [bk1 ···kd ] ∈ F n1 ×···×nd
and γ ∈ F , [ak1 ···kd ] + [bk1 ···kd ] := [ak1 ···kd + bk1 ···kd ] and γ[ak1 ···kd ] := [γak1 ···kd ].
The standard basis for F n1 ×···×nd is E := {Ek1 k2 ···kd : 1 ≤ k1 ≤ n1 , . . . , 1 ≤ kd ≤ nd } where
Ek1 k2 ···kd denotes the hypermatrix with 1 in the (k1 , k2 , . . . , kd )-coordinate and 0s everywhere else.
(1) (d)
Let X1 = [xij ] ∈ F m1 ×n1 , . . . , Xd = [xij ] ∈ F md ×nd and A ∈ F n1 ×···×nd . Define multilinear
matrix multiplication by A0 = (X1 , . . . , Xd ) · A ∈ F m1 ×···×md where
Xn1 ,...,nd (1) (d)
a0j1 ···jd = xj1 k1 · · · xjd kd ak1 ···kd for j1 ∈ hm1 i, . . . , jd ∈ hmd i. (15.1)
k1 ,...,kd =1

For any π ∈ Sd , the π-transpose of A = [aj1 ···jd ] ∈ F n1 ×···×nd is

Aπ := [ajπ(1) ···jπ(d) ] ∈ F nπ(1) ×···×nπ(d) . (15.2)

If n1 = · · · = nd = n, then a hypermatrix A ∈ F n×n×···×n is called cubical or hypercubical

of dimension n.
A cubical hypermatrix A = [aj1 ···jd ] ∈ F n×n×···×n is said to be symmetric if Aπ = A for every
π ∈ Sd and skew-symmetric or anti-symmetric or alternating if Aπ = sgn(π)A for every
π ∈ Sd .
The Segre outer product of a = [ai ] ∈ F ` , b = [bj ] ∈ F m , c = [ck ] ∈ F n is

a ⊗ b ⊗ c := [ai bj ck ]`,m,n
i,j,k=1 ∈ F
`×m×n
. (15.3)

Let A ∈ F n1 ×···×nd and B ∈ F m1 ×···×me be hypermatrices of orders d and e, respectively. Then

the outer product of A and B is a hypermatrix C of order d + e denoted

C = A ⊗ B ∈ F n1 ×···×nd ×m1 ×···×me (15.4)

with its (i1 , . . . , id , j1 . . . , je )-entry given by

ci1 ···id j1 ···je = ai1 ···id bj1 ···je (15.5)

for all i1 ∈ hn1 i, . . . , id ∈ hnd i and j1 ∈ hm1 i, . . . , je ∈ hme i.

Suppose A ∈ F n1 ×···×nd−1 ×n and B ∈ F n×m2 ×···×me are an order-d and an order-e hyper-
matrix, respectively, where the last index id of A and the first index j1 of B run over the same
range, i.e., id ∈ hni and j1 ∈ hni. The contraction product of A and B is an order-(d + e − 2)
hypermatrix C ∈ F n1 ×···×nd−1 ×m2 ×···×me whose entries are
Xn
ci1 ···id−1 j2 ···je = ai1 ···id−1 k bkj2 ···je , (15.6)
k=1

for i1 ∈ hn1 i, . . . , id−1 ∈ hnd−1 i and j2 ∈ hm2 i, . . . , je ∈ hme i.

15-4 Handbook of Linear Algebra

In general if A ∈ F n1 ×···×nd and B ∈ F m1 ×···×me have an index with a common range, say,
np = mq = n, then C ∈ F n1 ×···×nb p ×···×nd ×m1 ×···×m
b q ×···×me
is the hypermatrix with entries
Xn
ci1 ···bip ···i = ai1 ···k···id bj1 ···k···je . (15.7)
d j1 ···jq ···je k=1
b

where by convention a caret over any entry means that the respective entry is omitted (e.g.,
aibjk = aik and F l×m×n = F m×n ).
b

Contractions are not restricted to one pair of indices at a time. For hypermatrices A and B, the
hypermatrix
hA, Biα:λ,β:µ,...,γ:ν
is the hypermatrix obtained from contracting the αth index of A with the λth index of B, the βth
index of A with the µth index of B, . . . , the γth index of A with the νth index of B (assuming that
the indices that are contracted run over the same range and in the same order).

Facts:
Facts requiring proof for which no specific reference is given can be found in [Lim] and the
references therein.
1. F n1 ×···×nd with entrywise addition and scalar multiplication is a vector space.
2. The standard basis E is a basis for F n1 ×n2 ×···×nd , and dim F n1 ×n2 ×···×nd = n1 n2 · · · nd .
3. The elements of the standard basis of F n1 ×···×nd may be written as
Ek1 k2 ···kd = ek1 ⊗ ek2 ⊗ · · · ⊗ ekd ,
using the Segre outer product (15.3).
4. Let A ∈ F n1 ×···×nd and Xk ∈ F lk ×mk , Yk ∈ F mk ×nk for k = 1, . . . , d. Then
(X1 , . . . , Xd ) · [(Y1 , . . . , Yd ) · A] = (X1 Y1 , . . . , Xd Yd ) · A.
n1 ×···×nd
5. Let A, B ∈ F , α, β ∈ F , and Xk ∈ F mk ×nk for k = 1, . . . , d. Then
(X1 , . . . , Xd ) · [αA + βB] = α(X1 , . . . , Xd ) · A + β(X1 , . . . , Xd ) · B.
6. Let A ∈ F n1 ×···×nd , α, β ∈ F , and Xk , Yk ∈ F mk ×nk for k = 1, . . . , d. Then
[α(X1 , . . . , Xd ) + β(Y1 , . . . , Yd )] · A = α(X1 , . . . , Xd ) · A + β(Y1 , . . . , Yd ) · A.
7. The Segre outer product interacts with multilinear matrix multiplication in the fol-
lowing manner
hXr i Xr
(X1 , . . . , Xd ) · βp vp(1) ⊗ · · · ⊗ vp(d) = βp (X1 vp(1) ) ⊗ · · · ⊗ (Xd vp(d) ).
p=1 p=1

8. For cubical hypermatrices, each π ∈ Sd defines a linear operator

π : F n×···×n → F n×···×n , π(A) = Aπ .
9. Any matrix A ∈ F n×n may be written as a sum of a symmetric and a skew symmetric
matrix, A = 21 (A+AT )+ 12 (A−AT ). This is not true for hypermatrices of order d > 2.
For example, for a 3-hypermatrix A ∈ F n×n×n , the equivalent of the decomposition
is
1 1
A= (A+A(1,2,3) +A(1,3,2) +A(1,2) +A(1,3) +A(2,3) )+ (A+A(1,2) −A(1,3) −A(1,2,3) )
6 3
1 1
+ (A+A(1,3) −A(1,2) −A(1,3,2) )+ (A+A(1,2,3) +A(1,3,2) −A(1,2) −A(1,3) −A(2,3) ),
3 6
where S3 = {1, (1, 2), (1, 3), (2, 3), (1, 2, 3), (1, 3, 2)} (a 3-hypermatrix has five different
“transposes” in addition to the original).
Tensors and Hypermatrices 15-5

10. When d = e = 2, the contraction of two matrices is simply the matrix-matrix

Pn product:
If A ∈ F m×n and B ∈ F n×p , then C ∈ F m×p is given by C = AB, cij = k=1 aik bkj .
n n
11. When d = e = 1, the contraction of two vectors Pan ∈ R and b ∈ R is the scalar
c ∈ R given by the Euclidean inner product c = k=1 ak bk (for complex vectors, we
contract a with b to get the Hermitian inner product).
12. The contraction product respects expressions of hypermatrices written as a sum of
decomposable hypermatrices. For example, if
Xr Xs,t
A= ai ⊗ bi ⊗ ci ∈ F l×m×n , B= wk ⊗ xj ⊗ yj ⊗ zk ∈ F n×p×m×q ,
i=1 j,k=1

then
D Xr Xs,t E
ai ⊗ bi ⊗ ci , wk ⊗ x j ⊗ y j ⊗ z k =
i=1 j,k=1 2:3,3:1
Xr,s,t
hbi , yj ihci , wk iai ⊗ xj ⊗ zk ∈ F l×p×q ,
i,j,k=1

where hbi , yj i = bTi yj and hci , wk i = cTi wk as usual.

13. Multilinear matrix multiplication may also be expressed as the contraction of A ∈
F n1 ×···×nd with matrices X1 ∈ F m1 ×n1 , . . . , Xd ∈ F md ×nd . Take d = 3 for example;
(X, Y, Z) · A for A ∈ F l×m×n and X ∈ F p×l , Y ∈ F q×m , Z ∈ F r×n is

(X, Y, Z) · A = hX, hY, hZ, Ai2:3 i2:2 i2:1 .

Note that the order of contractions does not matter, i.e.,

hX, hY, hZ, Ai2:3 i2:2 i2:1 = hY, hX, hZ, Ai2:3 i2:1 i2:2 = · · · = hZ, hY, hX, Ai2:1 i2:2 i2:3 .

14. For the special case where we contract two hypermatrices A, B ∈ Cn1 ×···×nd in all
indices to get a scalar in C, we shall drop all indices and denote it by
Xn1 ,...,nd
hA, Bi = aj1 ···jd bj1 ···jd .
j1 ,...,jd =1

If we replace B by its complex conjugate, this gives the usual Hermitian inner product.

Examples:

1. A 3-hypermatrix A ∈ Cl×m×n has “three sides” and may be multiplied by three matrices
X ∈ Cp×l , Y ∈ Cq×m , Z ∈ Cr×n . This yields another 3-hypermatrix A0 ∈ Cp×q×r where
Xl,m,n
A0 = (X, Y, Z) · A ∈ Cp×q×r , a0αβγ = xαi yβj zγk aijk .
i,j,k=1

2. A 3-hypermatrix may be conveniently written down on a (2-dimensional) piece of paper as a

4,3,2
list of usual matrices, called slices. For example A = [aijk ]i,j,k=1 ∈ C4×3×2 can be written
down as two “slices” of 4 × 3 matrices
 
a111 a121 a131 a112 a122 a132
a211 a221 a231 a212 a222 a232 
A=
  ∈ C4×3×2
a311 a321 a331 a312 a322 a332 
a411 a421 a431 a412 a422 a432

where i, j, k index the row, column, and slice, respectively.

15-6 Handbook of Linear Algebra

3. More generally a 3-hypermatrix A ∈ Cl×m×n can be written down as n slices of l × m

matrices Ak ∈ Cl×m , k = 1, . . . , n, denoted

A = [A1 | A2 | · · · | An ] ∈ Cl×m×n .

If A = [aijk ]l,m,n l,m

i,j,k=1 , then Ak = [aijk ]i,j=1 .
4. A related alternative way is to introduce indeterminates x1 , . . . , xn and represent A ∈
Cl×m×n as a matrix whose entries are linear polynomials in x1 , . . . , xn :

x1 A1 + x2 A2 + · · · + xn An ∈ C[x1 , . . . , xn ]l×m .

Clearly, we have a one-to-one correspondence between l × m × n hypermatrices in Cl×m×n

and l × m matrices in C[x1 , . . . , xn ]l×m .
5. Just like a matrix can be sliced up into rows or columns, we may of course also slice up a
3-hypermatrix in two other ways: as l slices of m × n matrices or m slices of l × n matrices.
To avoid notational clutter, we shall not introduce additional notations for these but simply
note that these correspond to looking at the slices of the π-tranposes of A (just like the rows
of a matrix A ∈ Cm×n are the columns of its transpose AT ∈ Cn×m ).

Applications:

1. [Bax78, Yan67] In statistical mechanics, the Yang–Baxter equation is given by

XN XN
R`γiα Rαβjk Rmnγβ = Rαβjk R`mαγ Rγnβk
α,β,γ=1 α,β,γ=1

where i, j, k, `, m, n = 1, . . . , N . This may be written in terms of contractions of hyperma-

trices. Let R = (Rijkl ) ∈ CN ×N ×N ×N , then we have

hhR, Ri4:1 , Ri2:3,4:4 = hR, hR, Ri4:1 i1:3,2:4 .

2. Hooke’s law in one spatial dimension, with x = extension, F = force, c = the spring constant,
is F = −cx. Hooke’s law in three spatial dimensions is given by the linear elasticity equation:
X3
σij = cijkl γkl .
k,l=1

where x = [x1 , x2 , x3 ], C = [cijkl ] ∈ R3×3×3×3 is the elasticity tensor (also called stiffness
tensor), Σ ∈ R3×3 is the stress tensor, and Γ ∈ R3×3 is strain tensor. Hooke’s law may be
expressed in terms of contraction product as

Σ = hC, Γi3:1,4:2 .

3. The observant reader might have noted that the word “tensor” was used to denote a tensor
of order 2. The stress and strain tensors are all of order 2. This is in fact the most common
use of the term “tensors” in physics, where order-2 tensors occur a lot more frequently than
those of higher orders. There are authors (cf. [Bor90], for example) who use the term “tensor”
to mean exclusively a tensor of order 2.

15.2 Tensors and Multilinear Functionals

There is a trend in modern mathematics where instead of defining a mathematical entity
(like a tensor) directly, one first defines a whole space of such entities (like a space of
tensors) and subsequently defines the entity as an element of this space. For example, a
Tensors and Hypermatrices 15-7

succinct answer to the question “What is a vector?” is, “It is an element of a vector space”
[Hal85, p. 153]. One advantage of such an approach is that it allows us to examine the
entity in the appropriate context. Depending on the context, a matrix can be an element of
a vector space F n×n , of a ring, e.g., the endomorphism ring L(F n ) of F n , of a Lie algebra,
e.g., gl(n, F ), etc. Depending on what properties of the matrix one is interested in studying,
one chooses the space it lives in accordingly.
The same philosophy applies to tensors, where one first defines a tensor product of d vector
spaces V1 ⊗ · · · ⊗ Vd and then subsequently defines an order-d tensor as an element of such
a tensor product space. Since a tensor product space is defined via a universal factorization
property (the definition used in Section 14.2), it can be interpreted in multiple ways, such as
a quotient module (the equivalent definition used here) or a space of multilinear functionals.
We will regard tensors as multilinear functionals. A perhaps unconventional aspect of our
approach is that for clarity we isolate the notion of covariance and contravariance (see
Section 15.10) from our definition of a tensor. We do not view this as an essential part of
the definition but a source of obfuscation.
Tensors can also be represented as hypermatrices by choosing a basis. Given a set of
bases, the essential information about a tensor T is captured by the coordinates aj1 ···jd ’s
(cf. Fact 3 below). We may view the coefficient aj1 ···jd as the (j1 , . . . , jd )-entry of the d-
dimensional matrix A = [aj1 ···jd ] ∈ F n1 ×···×nd , where A is a coordinate representation of T
with respect to the specified bases.
See also Section 14.2 for more information on tensors and tensor product spaces.

Definitions:
Let F be a field and let V1 , . . . , Vd be F -vector spaces.
The tensor product space V1 ⊗ · · · ⊗ Vd is the quotient module F (V1 , . . . , Vd )/R where
F (V1 , . . . , Vd ) is the free module generated by all n-tuples (v1 , . . . , vd ) ∈ V1 × · · · × Vd and R
is the submodule of F (V1 , . . . , Vd ) generated by elements of the form

(v1 , . . . , αvk + βvk0 , . . . , vd ) − α(v1 , . . . , vk , . . . , vd ) − β(v1 , . . . , vk0 , . . . , vd )

for all vk , vk0 ∈ Vk , α, β ∈ F , and k ∈ {1, . . . , d}. We write v1 ⊗· · ·⊗vd for the element (v1 , . . . , vd )+
R in the quotient space F/R.
An element of V1 ⊗· · ·⊗Vd that can be expressed in the form v1 ⊗· · ·⊗vd is called decomposable.
The symbol ⊗ is called the tensor product when applied to vectors from abstract vector spaces.
The elements of V1 ⊗ · · · ⊗ Vd are called order-d tensors or d-tensors and nk = dim Vk ,
k = 1, . . . , d are the dimensions of the tensors.
(k) (k)
Let Bk = {b1 , . . . , bnk } be a basis for Vk , k = 1, . . . , d. For a tensor T ∈ V1 ⊗ · · · ⊗ Vd , the
coordinate representation of T with respect to the specified bases is [T ]B1 ,...,Bd = [aj1 ···jd ].
where Xn1 ,...,nd (1) (d)
T = aj1 ···jd bj1 ⊗ · · · ⊗ bjd . (15.8)
j1 ,...,jd =1

The special case where V1 = · · · = Vd = V is denoted Td (V ) or V ⊗d , i.e., Td (V ) = V ⊗ · · · ⊗ V .

For any π ∈ Sd , the action of π on Td (V ) is defined by

π(v1 ⊗ · · · ⊗ vd ) := vπ(1) ⊗ · · · ⊗ vπ(d) (15.9)

for decomposable elements, and then extended linearly to all elements of Td (V ).

A tensor T ∈ Td (V ) is symmetric if π(T ) = T for all π ∈ Sd and is alternating if
π(T ) = sgn(π)T for all π ∈ Sd .
For a vector space V , V ∗ denotes the dual space of linear functionals of V (cf. Section 3.6).
A multilinear functional on V1 , . . . , Vd is a function T : V1 × V2 × · · · × Vd → F , i.e., for α, β ∈ F ,
15-8 Handbook of Linear Algebra

T (x1 , . . . , αyk + βzk , . . . , xd ) = αT (x1 , . . . , yk , . . . , xd ) + βT (x1 , . . . , zk , . . . , xd ), (15.10)

for every k = 1, . . . , d.
The vector space of (F -valued) multilinear functionals on V1 , . . . , Vd is denoted by L(V1 , . . . , Vd ; F ).
A multilinear functional T ∈ L(V1 , . . . , Vd ; F ) is decomposable if T = θ1 · · · θd where θi ∈ Vi∗
and θ1 · · · θd (v1 , . . . , vd ) = θ1 (v1 ) · · · θd (vd ).
For any π ∈ Sd , the action of π on L(V, . . . , V, F ) is defined by

π(T )(v1 , . . . , vd ) = T (vπ(1) , . . . , vπ(d) ). (15.11)

A multilinear functional T ∈ L(V1 , . . . , Vd ; F ) is symmetric if π(T ) = T for all π ∈ Sd and is

alternating if π(T ) = sgn(π)T for all π ∈ Sd .

Facts:
Facts requiring proof for which no specific reference is given can be found in [Bou98,
Chap. II], [KM97, Chap. 4], [Lan02, Chap. XVI], and [Yok92, Chap. 1]. Additional facts
about tensors can be found in Section 14.2.
1. The tensor product space V1 ⊗ · · · ⊗ Vd with ν : V1 × · · · × Vm → V1 ⊗ · · · ⊗ Vd defined
by
ν(v1 , . . . , vd ) = v1 ⊗ · · · ⊗ vd = (v1 , . . . , vd ) + R ∈ F (V1 , . . . , Vd )/R
and extended linearly satisfies the Universal Factorization Property that can be used
to define tensor product spaces (cf. Section 14.2), namely:
If ϕ is a multilinear map from V1 × · · · × Vd into the vector space U , then there
exists a unique linear map ψ from V1 ⊗ · · · ⊗ Vd into U , that makes the following
diagram commutative:

V1 × · · · × Vd
ν / V1 ⊗ · · · ⊗ Vd

ψ
ϕ
( 
U

i.e., ψν = ϕ.
2. If U = F l , V = F m , W = F n , we may identify

F l ⊗ F m ⊗ F n = F l×m×n

through the interpretation of the tensor product of vectors as a hypermatrix via the
Segre outer product (cf. Eq. (15.3)),

[a1 , . . . , al ]T ⊗ [b1 , . . . , bm ]T ⊗ [c1 , . . . , cn ]T = [ai bj ck ]l,m,n

i,j,k=1 .

This is a model of the universal definition of ⊗ given in Section 14.2.

(k) (k)
3. Given bases Bk = {b1 , . . . , bnk } for Vk , k = 1, . . . , d, any tensor T in V1 ⊗ · · · ⊗ Vd ,
can be expressed as a linear combination
Xn1 ,...,nd (1) (d)
T = aj1 ···jd bj1 ⊗ · · · ⊗ bjd .
j1 ,...,jd =1

In older literature, the aj1 ···jd ’s are often called the components of T .
4. One loses information when going from the tensor to its hypermatrix representation,
in the sense that the bases B1 , . . . , Bd must be specified in addition to the hypermatrix
A in order to recover the tensor T .
Tensors and Hypermatrices 15-9

5. Every choice of bases on V1 , . . . , Vd gives a (usually) different hypermatrix represen-

tation of the same tensor in V1 ⊗ · · · ⊗ Vd .
6. Given two sets of bases B1 , . . . , Bd and B10 , . . . , Bd0 for V1 , . . . , Vd , the same tensor T
has two coordinate representations as a hypermatrix,

A = [T ]B1 ,...,Bd and A0 = [T ]B10 ,...,Bd0

where A = [ak1 ···kd ], A0 = [a0k1 ···kd ] ∈ F n1 ×···×nd . The relationship between A and A0
is given by the multilinear matrix multiplication

A0 = (X1 , . . . , Xd ) · A

where Xk ∈ GL(nk , F ) is the change-of-basis matrix transforming Bk0 to Bk for k =

1, . . . , d. We shall call this the change-of-basis rule. Explicitly, the entries of A0 and
A are related by
Xn1 ,...,nd
a0j1 ···jd = xj1 k1 · · · xjd kd ak1 ···kd for j1 ∈ hn1 i, . . . , jd ∈ hnd i,
k1 ,...,kd =1

where X1 = [xj1 k1 ] ∈ GL(n1 , F ), . . . , Xd = [xjd kd ] ∈ GL(nd , F ).

7. Cubical hypermatrices arise from a natural coordinate representation of tensors T ∈
Td (V ), i.e.,
T : V × ··· × V → F
where by “natural” we mean that we make the same choice of basis B for every copy
of V , i.e.,
[T ]B,...,B = A.
8. Td (V ) is an Sd -module.
9. Symmetric and alternating hypermatrices are natural coordinate representations of
symmetric and alternating tensors.
10. For finite-dimensional vector spaces V1 , . . . , Vd , the space L(V1∗ , . . . , Vd∗ ; F ) of multi-
linear functionals on the dual spaces Vi∗ is naturally isomorphic to the tensor product
space V1 ⊗ · · · ⊗ Vd , with v1 ⊗ · · · ⊗ vd ↔ v̂1 · · · v̂d extended by linearity, where
v̂ ∈ V ∗∗ is defined by v̂(f ) = f (v) for f ∈ V ∗ . Since V is naturally isomorphic to
V ∗∗ via v ↔ v̂ (cf. Section 3.6), every vector space of multilinear functionals is a
tensor product space (of the dual spaces), and a multilinear functional is a tensor:
A decomposable multilinear functional T = θ1 · · · θd ∈ L(V1 , . . . , Vd ; F ) with θi ∈ Vi∗
is naturally associated with θ1 ⊗ · · · ⊗ θd ∈ V1∗ ⊗ · · · ⊗ Vd∗ , and this relationship is
extended by linearity.
11. A multilinear functional is decomposable as a multilinear functional if and only if it
is a decomposable tensor.
(k) (k)
12. Let Bk = {b1 , . . . , bnk } be a basis for the vector space Vk for k = 1, . . . , d,
so Vk ∼ n
= F k where nk = dim Vk , with the isomorphism x 7→ [x] where [x] de-
notes the coordinate vector with respect to basis Bk . For a multilinear functional
(1) (d)
T : V1 × · · · × Vd → F , define aj1 ···jd := T (bj1 , . . . , bjd ) for ji ∈ hni i. Then T has
the explicit formula
Xn1 ,...,nd (1) (d)
T (x1 , . . . , xd ) = aj1 ···jd xj1 · · · xjd ,
j1 ,...,jd =1

(k) (k)
in terms of the coordinates of the coordinate vector [xk ] = [x1 , . . . , xnk ]T ∈ F nk for
k = 1, . . . , d. In older literature, the aj1 ···jd ’s are also often called the components of
T as in Fact 3.
15-10 Handbook of Linear Algebra

(k) (k)
13. Let Bk = {b1 , . . . , bnk } be a basis for Vk for k = 1, . . . , d.
(a) For each i1 ∈ hn1 i, . . . , id ∈ hnd i, define the multilinear functional ϕi1 ···id :
V1 × · · · × Vd → F by
(
(1) (d) 1 if i1 = j1 , . . . , id = jd ,
ϕi1 ···id (bj1 , . . . , bjd ) =
0 otherwise,

and extend the definition to all of V1 × · · · × Vd via (15.10). The set

B ∗ := {ϕi1 ···id : i1 ∈ hn1 i, . . . , id ∈ hnd i}
is a basis for L(V1 , . . . , Vd ; F ).
(b) The set
n o
(1) (d)
B := bj1 ⊗ · · · ⊗ bjd : j1 ∈ hn1 i, . . . , jd ∈ hnd i

is a basis for V1 ⊗ · · · ⊗ Vd .
(c) For a multilinear functional T : V1 × · · · × Vd → F with aj1 ···jd as defined in
Pn ,...,n
Fact 12, T = j11,...,jdd=1 aj1 ···jd ϕj1 ···jd .
(d) dim V1 ⊗ · · · ⊗ Vd = dim V1 · · · dim Vd since |B ∗ | = |B| = n1 · · · nd .
14. Td (V ) is an End(V )-module (where End(V ) is the algebra of linear operators on V )
with the natural action defined on decomposable elements via
g(v1 ⊗ · · · ⊗ vd ) = g(v1 ) ⊗ · · · ⊗ g(vd )
for any g ∈ End(V ) and then extended linearly to all of Td (V ).

Examples:
For notational convenience, let d = 3.
1. Explicitly, the definition of a tensor product space above simply means that
nXn o
U ⊗ V ⊗ W := αi ui ⊗ vi ⊗ wi : ui ∈ U, vi ∈ V, wi ∈ W, n ∈ N
i=1

where ⊗ satisfies

(αv1 + βv10 ) ⊗ v2 ⊗ v3 = αv1 ⊗ v2 ⊗ v3 + βv10 ⊗ v2 ⊗ v3 ,

v1 ⊗ (αv2 + βv20 ) ⊗ v3 = αv1 ⊗ v2 ⊗ v3 + βv1 ⊗ v20 ⊗ v3 ,
v1 ⊗ v2 ⊗ (αv3 + βv30 ) = αv1 ⊗ v2 ⊗ v3 + βv1 ⊗ v2 ⊗ v30 .

The “modulo relation” simply means that ⊗ obeys these rules. The statement that u⊗v ⊗w
are generators of U ⊗ V ⊗ W simply means that U ⊗ V ⊗ W is the set of all possible linear
combinations of the form u ⊗ v ⊗ w where u ∈ U , v ∈ V , w ∈ W .
2. We emphasize here that a tensor and a hypermatrix are quite different. To specify a tensor
T ∈ V1 ⊗ · · · ⊗ Vd , we need both the hypermatrix [T ]B1 ,...,Bd ∈ Cn1 ×···×nd and the bases
B1 , . . . , Bd that we chose for V1 , . . . , Vd .
3. Each hypermatrix in F n1 ×···×nd has a unique, natural tensor associated with it: the tensor
in the standard basis of F n1 ⊗· · ·⊗F nd . (The same is true for matrices and linear operators.)

Applications:

1. In physics parlance, a decomposable tensor represents factorizable or pure states. In general,

a tensor in U ⊗ V ⊗ W will not be decomposable.
Tensors and Hypermatrices 15-11

2. [Cor84] In the standard model of particle physics, a proton is made up of two up quarks and
one down quark. A more precise statement is that the state of a proton is a 3-tensor

1
√ (e1 ⊗ e1 ⊗ e2 − e2 ⊗ e1 ⊗ e1 ) ∈ V ⊗ V ⊗ V
2
where V is a 3-dimensional inner product space spanned by orthonormal vectors e1 , e2 , e3
that have the following interpretation:

e1 = state of the up quark,

e2 = state of the down quark,
e3 = state of the strange quark,

and thus

ei ⊗ ej ⊗ ek = composite state of the three quark states ei , ej , ek .

3. In physics, the question “What is a tensor?” is often taken to mean “What kind of physical
quantities should be represented by tensors?” It is often cast in the form of questions such as
“Is elasticity a tensor?”, “Is gravity a tensor?”, etc. The answer is that the physical quantity
in question is a tensor if it obeys the change-of-bases rule in Fact 15.2.6: A d-tensor is an
object represented by a list of numbers aj1 ···jd ∈ C, jk = 1, . . . , nk , k = 1, . . . , d, once a basis
is chosen, but only if these numbers transform themselves as expected when one changes the
basis.
4. Elasticity is an order-4 tensor and may be represented by a hypermatrix C ∈ R3×3×3×3 . If
we measure stress using a different choice of coordinates (i.e., different basis), then the new
hypermatrix representation C 0 ∈ R3×3×3×3 must be related to C via

C 0 = (X, X, X, X) · C (15.12)

where X ∈ GL(3, R) is the change-of-basis matrix, and Eq. (15.12) is defined according to
Fact 6 X3
c0pqrs = xpi xqj xrk xsl cijkl , p, q, r, s = 1, 2, 3. (15.13)
i,j,k,l=1

5. Let A be an algebra over a field F, i.e., a vector space on which a notion of vector multipli-
cation · : A × A → A, (a, b) 7→ a · b is defined. Let B = {e1 , . . . , en } be a basis for A. Then
A is completely determined by the hypermatrix C = [cijk ] ∈ Fn×n×n where
n
X
ei · ej = cijk ek .
k=1

The n3 entries of C are often called the structure constants of A. This hypermatrix is the
coordinate representation of a tensor in A ⊗ A ⊗ A with respect to the basis B. If we had
chosen a new basis B0 , then the new coordinate representation C 0 would be related to C as
in Fact 6 — in this case C 0 = (X, X, X) · C where X is the change of basis matrix from
B to B0 . Note that this says that the entries of the hypermatrix of structure constants are
coordinates of a tensor with respect to a basis.
6. For an explicit example, the Lie algebra so(3) is the set of all skew-symmetric matrices in
R3×3 . A basis is given by
     
0 0 0 0 0 −1 0 −1 0
Z1 = 0 0 −1 , Z2 =  0 0 0 , Z3 =  1 0 0 .
0 1 0 1 0 0 0 0 0
15-12 Handbook of Linear Algebra

The product · for so(3) is the commutator product [X, Y ] = XY − Y X. Note that [Y, X] =
−[X, Y ] and [X, X] = 0. Since [Z1 , Z2 ] = Z3 , [Z2 , Z3 ] = Z1 , [Z3 , Z1 ] = Z2 , the structure
constants of so(3) are given by the hypermatrix ε = (εijk ) ∈ R3×3×3 defined by

+1 if (i, j, k) = (1, 2, 3), (2, 3, 1), (3, 1, 2),

εijk = −1 if (i, j, k) = (1, 3, 2), (2, 1, 3), (3, 2, 1),


0 if i = j, j = k, k = i,
(i − j)(j − k)(k − i)
= .
2
ε is often called the Levi-Civita symbol.

15.3 Tensor Rank

There are several equivalent ways to define the rank of a matrix which yield non-equivalent
definitions on hypermatrices of higher order. We will examine two of the most common ones
in this chapter: tensor rank as defined below and multilinear rank as defined in Section 15.7.
Both notions are due to Frank L. Hitchcock [Hit27a, Hit27b].

Definitions:
Let F be a field.
A hypermatrix A ∈ F n1 ×···×nd has rank one or rank-1 if there exist non-zero v(i) ∈ F n ,
i = 1, . . . , d, so that A = v(1) ⊗ · · · ⊗ v(d) and v(1) ⊗ · · · ⊗ v(d) is the Segre outer product defined
in Eq. (15.3).
The rank of a hypermatrix A ∈ F n1 ×···×nd is defined to be the smallest r such that it may be
written as a sum of r rank-1 hypermatrices, i.e.,
n Xr (1) (d)
o
rank(A) := min r : A = vp ⊗ · · · ⊗ vp . (15.14)
p=1

For vector spaces V1 , . . . , Vd , the rank or tensor rank of T ∈ V1 ⊗ · · · ⊗ Vd is

n Xr (1) (d)
o
rank(T ) = min r : T = vp ⊗ · · · ⊗ vp . (15.15)
p=1

(k)
Here vp is a vector in the abstract vector space Vk and ⊗ denotes tensor product as defined in
Section 15.2.
A hypermatrix or a tensor has rank zero if and only if it is zero (in accordance with the
convention that the sum over the empty set is zero).
A minimum length decomposition of a tensor or hypermatrix, i.e.,
Xrank(T ) (1) (d)
T = vp ⊗ · · · ⊗ vp , (15.16)
p=1

is called a rank-retaining decomposition or simply rank decomposition.

Facts:
Facts requiring proof for which no specific reference is given can be found in [BCS96,
Chap. 19], [Lan12, Chap. 3], [Lim], and the references therein.

1. Let A = [aijk ]l,m,n l,m

i,j,k=1 and Ak = [aijk ]i,j=1 . The following are equivalent statements
characterize rank(A) ≤ r.
Tensors and Hypermatrices 15-13

(a) there exist x1 , . . . , xr ∈ F l , y1 , . . . , yr ∈ F m , and z1 , . . . , zr ∈ F m ,

A = x1 ⊗ y1 ⊗ z1 + · · · + xr ⊗ yr ⊗ zr ;
(b) there exist x1 , . . . , xr ∈ F l and y1 , . . . , yr ∈ F m with
span{A1 , . . . , An } ⊆ span{x1 y1T , . . . , xr yrT };
(c) there exist diagonal D1 , . . . , Dn ∈ F r×r and X ∈ F l×r , Y ∈ F m×r ,
Ak = XDk Y T , k = 1, . . . , n.
The statements analogous to (1a), (1b), and (1c) with r required to be minimal
characterize rank(A) = r. A more general form of this fact is valid for for d-tensors.
2. For A ∈ F n1 ×···×nd and (X1 , . . . , Xd ) ∈ GL(n1 , . . . , nd , F ),
rank((X1 , . . . , Xd ) · A) = rank(A).
3. If T ∈ V1 ⊗ · · · ⊗ Vd , B1 , . . . , Bd are bases for V1 , . . . , Vd and A = [T ]B1 ,...,Bd ∈
F n1 ×···×nd (where nk = dim Vk , k = 1, . . . , d), then rank(A) = rank(T ).
4. Since Rl×m×n ⊆ Cl×m×n , given A ∈ Rl×m×n , we may consider its rank over F = R
or F = C, n Xr o
rankF (A) = r : A = x i ⊗ y i ⊗ zi , x i ∈ F l , y i ∈ F m , zi ∈ F n .
i=1
Clearly rankC (A) ≤ rankR (A). However, strict inequality can occur (see Example 1
next). Note for a matrix A ∈ Rm×n , this does not happen; we always have rankC (A) =
rankR (A).
5. When d = 2, Eq. (15.14) agrees with the usual definition of matrix rank and Eq.
(15.15) agrees with the usual definition of rank for linear operators and bilinear forms
on finite-dimensional vector spaces.
6. In certain literature, the term “rank” is often used to mean what we have called
“order” in Section 15.2. We avoid such usage for several reasons, among which the
fact that it does not agree with the usual meaning of rank for linear operators or
matrices.

Examples:

1. The phenomenon of rank dependence on field was first observed by Bergman [Ber69]. Take
linearly independent pairs of vectors x1 , y1 ∈ Rl , x2 , y2 ∈ Rm , x3 , y3 ∈ Rn and set zk =
xk + iyk and z̄k = xk − iyk , then
A = x1 ⊗ x2 ⊗ x3 + x1 ⊗ y2 ⊗ y3 − y1 ⊗ x2 ⊗ y3 + y1 ⊗ y2 ⊗ x3 (15.17)
1
= (z̄1 ⊗ z2 ⊗ z̄3 + z̄1 ⊗ z̄2 ⊗ z3 ).
2
One may in fact show that rankC (A) = 2 < 3 = rankR (A).
2. An example where rankR (A) ≤ 2 < rankQ (A) is given by
z ⊗ z ⊗ z + z ⊗ z ⊗ z = 2x ⊗ x ⊗ x − 4y ⊗ y ⊗ x + 4y ⊗ x ⊗ y − 4x ⊗ y ⊗ y ∈ Q2×2×2 ,
√ √
where z = x + 2y and z = x − 2y.

Applications:

1. Let M, C, K ∈ Rn×n be the mass, damping, and stiffness matrices of a viscously damped
linear system in free vibration
M ẍ(t) + C ẋ(t) + Kx(t) = 0.
where M, C, K are all symmetric positive definite. The system may be decoupled using
classical modal analysis [CO65] if and only if
CM −1 K = KM −1 C.
Formulated in hypermatrix language, this asks when A = [M | C | K] ∈ Rn×n×3 has
rank(A) ≤ 3.
15-14 Handbook of Linear Algebra

2. The notion of tensor rank arises in several areas and a well-known one is algebraic com-
putational complexity [BCS96], notably the complexity of matrix multiplications. This is
surprisingly easy to explain. For matrices X = [xij ], Y = [yjk ] ∈ Cn×n , observe that the
product may be expressed as
Xn Xn
XY = xik ykj Eij = ϕik (X)ϕkj (Y )Eij (15.18)
i,j,k=1 i,j,k=1

where Eij = ei e∗j ∈ Cn×n has all entries 0 except 1 in the (i, j)-entry and ϕij (X) =
∗
tr(Eij X) = xij is the linear functional ϕij : Cn×n → C dual to Eij . Let Tn : Cn×n ×
n×n
C → Cn×n be the map that takes a pair of matrices (X, Y ) ∈ Cn×n × Cn×n to their
product T (X, Y ) = XY ∈ Cn×n . Then by Eq. (15.18), Tn is given by the tensor
Xn
Tn = ϕik ⊗ ϕkj ⊗ Eij ∈ (Cn×n )∗ ⊗ (Cn×n )∗ ⊗ Cn×n . (15.19)
i,j,k=1

The exponent of matrix multiplication is then a positive number ω defined in terms of tensor
rank,
ω := inf{α : rank(Tn ) = O(nα ), n ∈ N}.

It is not hard to see that whatever the value of ω > 0, there must exist O(nω ) algorithms for
multiplying n×n matrices. In fact, every r-term decomposition (15.16) of Tn yields an explicit
algorithm for multiplying two n × n matrices in O(nlog2 r ) complexity. Via elementary row
operations, we may deduce that any O(nω ) algorithm for computing matrix multiplications
(A, B) 7→ AB would also yield a corresponding O(nω ) algorithm for matrix inversion A 7→
A−1 (and thus for solving linear systems Ax = b). If we choose the standard bases {Eij :
i, j ∈ hni} on Cn×n and its dual bases {ϕij : i, j ∈ hni} on the dual space (Cn×n )∗ , then
the 3-tensor Tn may be represented by a 3-hypermatrix
2
×n2 ×n2
Mn ∈ Cn .

The connection between the exponent of matrix multiplication and tensor rank was first
noted by Strassen in [Str73]. We refer the reader to Chapter 61 and [Knu98] for very readable
accounts and to [BCS96] for an extensive in-depth discussion.
3. The special case n = 2 is behind Strassen’s algorithm for matrix multiplication and inversion
with O(nlog2 7 ) time complexity [Str69]. We shall present it in the modern language of
hypermatrices. We write M2 = [A1 | A2 | A3 | A4 ] ∈ C4×4×4 where the matrix slices of M2
are        
I O O O O I O O
A1 = , A2 = , A3 = , A4 = ,
O O I O O O O I

I and O are 2 × 2 identity and zero matrices. Define

   
1 1 0 0 1 0 1 1 1 1 0 0 1 0
 0 0 0 1 0 0 1 0 0 −1 1 0 0 0
X=
 0
, Y = ,
1 0 0 0 1 0 0 1 0 0 1 0 0
−1 0 1 1 0 1 0 1 0 0 1 −1 0 1
D1 = diag(1, 0, 1, 1, 0, 0, −1), D2 = diag(0, 0, −1, 0, 0, 1, 0),
D3 = diag(0, 0, 0, 0, 1, 0, 1), D4 = diag(−1, 1, 0, 0, −1, −1, 0).

One may check that XDj Y T = Aj for j = 1, 2, 3, 4. In other words, this is a simultaneous
diagaonlization of A1 , A2 , A3 , A4 by X and Y in the sense of Fact 1 and so we conclude that
rank(M2 ) ≤ 7. In fact, it has been shown that rank(M2 ) = 7 [HK71, Win71] and much more
recently, the border rank (see Section 15.4) of M2 is 7 [Lan06].
Tensors and Hypermatrices 15-15

4. In quantum computing, a pure state, also known as a completely separable state, corresponds
to a rank-1 tensor. A quantum state that is not pure is called entangled. A natural, but not
the most commonly used, measure of the degree of entanglement is therefore tensor rank
[Bry02], i.e., the minimal number of pure states it can be written as a sum of. For example,
the well-known Greenberger–Horne–Zeilinger state [GHZ89] may be regarded as a 2 × 2 × 2
hypermatrix of rank 2:

1
|GHZi = √ (|0i ⊗ |0i ⊗ |0i + |1i ⊗ |1i ⊗ |1i) ∈ C2×2×2 ,
2

while the W state [DVC00] may be regarded as a 2 × 2 × 2 hypermatrix of rank 3:

1
|W i = √ (|0i ⊗ |0i ⊗ |1i + |0i ⊗ |1i ⊗ |0i + |1i ⊗ |0i ⊗ |0i) ∈ C2×2×2 .
3

15.4 Border Rank

We now discuss a phenomenon that may appear peculiar at first since one does not encounter
this for usual matrices or tensors of order 2. As one will see from the following examples, we
may get a sequence of 3-hypermatrices of rank not more than 2 converging to a limit that has
rank 3, which is somewhat surprising since for matrices, this can never happen. Another way
to say this is that the set {A ∈ Cm×n : rank(A) ≤ r} is closed. What we deduce from this
example is the same does not hold for hypermatrices, the set {A ∈ Cl×m×n : rank(A) ≤ r}
is not a closed set in general.

Definitions:
The border rank of a hypermatrix A ∈ Cn1 ×···×nd is
 
rank(A) = min r : inf kA − Bk = 0 ,
rank(B)≤r

where k · k is any norm on hypermatrices (including one obtained by identifying A ∈ Cn1 ×···×nd
with Cn1 ···nd ; see Section 15.8).

Facts:
Facts requiring proof for which no specific reference is given can be found in [BCS96,
Chap. 19], [Lan12, Chap. 3], [Lim], and the references therein.
1. For A ∈ Cn1 ×···×nd , rank(A) ≤ rank(A).
2. For a matrix A ∈ Cm×n , rank(A) = rank(A).
3. There exist examples of a sequence of 3-hypermatrices of rank not more than 2 that
converges to a limit that has rank 3 [BLR80] (see Examples 1 and 2). In fact, the gap
between rank and border rank can be arbitrarily large.
4. Let A = [A1 | A2 | A3 ] ∈ Cn×n×3 where A2 is invertible. Then
1
rank(A) ≥ n + rank(A1 A−1 2 A3 − A3 A2 A1 ).
−1
2
5. Let A ∈ Cn1 ×···×nd and (X1 , . . . , Xd ) ∈ GL(n1 , . . . , nd , C). Then
rank((X1 , . . . , Xd ) · A) = rank(A).
6. While border rank is defined here for hypermatrices, the definition of border rank
extends to tensors via coordinate representation because of the invariance of rank
under change of basis (cf. Fact 15.3.2).
15-16 Handbook of Linear Algebra

7. Suppose rank(T ) = r and rank(T ) = s. If s < r, then T has no best rank-s approxi-
mation.

Examples:

1. [BLR80] The original example of a hypermatrix whose border rank is less than its rank is a
2 × 2 × 2 hypermatrix. In the notation of Section 15.1, if we choose xp = (1, 0), yp = (0, 1) ∈
C2 for p = 1, 2, 3, then
                   
1 1 0 1 0 1 0 1 1 0 1 1 0
A= ⊗ ⊗ + ⊗ ⊗ + ⊗ ⊗ = ∈ C2×2×2
0 0 1 0 1 0 1 0 0 1 0 0 0

and
           
1 1 1 1 1 1 1 1
Aε = ⊗ ⊗ − ⊗ ⊗
ε ε ε ε ε 0 0 0
ε2
     
1 1 ε ε 1 1 0 0 0 0 1 1 ε
= − = ∈ C2×2×2 ,
ε ε ε 2 ε2 ε3 ε 0 0 0 0 1 ε ε ε2

from which it is clear that limε→0 Aε = A. We note here that the hypermatrix A is actu-
ally the same hypermatrix that represents the W state in quantum computing (cf. Exam-
ple 15.3.4).
2. Choose linearly independent pairs of vectors x1 , y1 ∈ Cl , x2 , y2 ∈ Cm , x3 , y3 ∈ Cn (so
l, m, n ≥ 2 necessarily). Define the 3-hypermatrix,

A := x1 ⊗ x2 ⊗ y3 + x1 ⊗ y2 ⊗ x3 + y1 ⊗ x2 ⊗ x3 ∈ Cl×m×n ,

and, for any ε 6= 0, a family of 3-hypermatrices parameterized by ε,

(x1 + εy1 ) ⊗ (x2 + εy2 ) ⊗ (x3 + εy3 ) − x1 ⊗ x2 ⊗ x3
Aε := .
ε
Now it is easy to verify that

A − Aε = ε(y1 ⊗ y2 ⊗ x3 + y1 ⊗ x2 ⊗ y3 + x1 ⊗ y2 ⊗ y3 ) (15.20)

and so
kA − Aε k = O(ε)
In other words, A can be approximated arbitrarily closely by Aε . As a result,

lim Aε = A.
ε→0

One may also check that rank(A) = 3 while it is clear that rank(Aε ) ≤ 2. By definition, the
border rank of A is not more than 2. In fact, rank(A) = 2 because a rank-1 hypermatrix
always has border rank equals to 1 and as such the border rank of a hypermatrix whose rank
and border rank differ must have border rank at least 2 and rank at least 3.
3. There are various results regarding the border rank of hypermatrices of specific dimensions.
If either l, m, or n = 2, then a hypermatrix A = [A1 , A2 ] ∈ Cm×n×2 may be viewed as a
matrix pencil and we may depend on the Kronecker-Weierstraß canonical form to deduce
results about rank and border rank.
4. One may define border rank algebraically over arbitrary fields without involving norm by
considering expressions like those in Eq. (15.20) modulo ε. We refer the reader to [BCS96,
Knu98] for more details on this.
Tensors and Hypermatrices 15-17

15.5 Generic and Maximal Rank

Unlike tensor rank, border rank, and multilinear rank, the notions of rank discussed in this
section apply to the whole space rather than an individual hypermatrix or tensor.

Definitions:
Let F be a field.
The maximum rank of F n1 ×···×nd is

maxrankF (n1 , . . . , nd ) = max{rank(A) : A ∈ F n1 ×···×nd }.

The generic rank of Cn1 ×···×nd as

genrankC (n1 , . . . , nd ) = max{rank(A) : A ∈ Cn1 ×···×nd }.

Facts:
Facts requiring proof for which no specific reference is given can be found in [Lan12, Chap. 3].
1. The concepts of maxrank and genrank are uninteresting for matrices (i.e., 2-hypermatrices)
since
genrankC (m, n) = maxrankC (m, n) = max{m, n},
but for d > 2, the exact values of genrankC (l, m, n) and maxrankC (l, m, n) are mostly
unknown.
2. Since rank(A) ≤ rank(A), genrankC (n1 , . . . , nd ) ≤ maxrankC (n1 , . . . , nd ).
3. A simple dimension count yields the following lower bound
 
n1 · · · nd
genrankC (n1 , . . . , nd ) ≥ .
n1 + · · · + nd − d + 1
Note that strict inequaltiy can occur when d > 2. For example, for d = 3,

genrankC (2, 2, 2) = 2 whereas maxrankC (2, 2, 2) = 3

and genrankC (3, 3, 3) = 5 [Str83] while d33 /(3 × 3 − 3 + 1)e = 4.

4. The generic rank is a generic property in Cn1 ×···×nd , i.e., every hypermatrix would
have rank equal to genrankC (n1 , . . . , nd ) except for those in some proper Zariski-closed
subset. In particular, hypermatrices that do not have rank genrankC (n1 , . . . , nd ) are
contained in a measure-zero subset of Cn1 ×···×nd (with respect to, say, the Lebesgue
or Gaussian measure).
5. There is a notion of typical rank that is analogous to generic rank for fields that are
not algebraically closed.

15.6 Rank-Retaining Decomposition

Matrices have a notion of rank-retaining decomposition, i.e., given A ∈ Cm×n , there are
various ways to write it down as a product of two matrices A = GH where G ∈ Cm×r and
H ∈ Cr×n where r = rank(A). Examples include the LU, QR, SVD, and others.
There is a rank-retaining decomposition that may be stated in a form that is almost
identical to the singular value decomposition of a matrix, with one caveat — the analogue
of the singular vectors would in general not be orthogonal (cf. Fact 1). This decomposition
is often unique, a result due to Kruskal that depends on the notion of Kruskal rank, a
15-18 Handbook of Linear Algebra

concept that has been reinvented multiple times under different names (originally as girth
in matroid theory [Oxl11], as krank in tensor decompositions [SB00], and most recently as
spark in compressive sensing [DE03]). The term “Kruskal rank” was not coined by Kruskal
himself who simply denoted it as kS . It is an unfortunate nomenclature, since it is not a
notion of rank. “Kruskal dimension” would have been more appropriate.

Definitions:
Let V be a vector space, v1 , . . . , vn ∈ V , and S = {v1 , . . . , vn }.
The Kruskal rank or krank of S, denoted krank(S) or krank(v1 , . . . , vr ), is the largest k ∈ N
so that every k-element subset of S is linearly independent.
The girth or spark of S is the smallest s ∈ N so that there exists an s-element subset of S that
is linear dependent.
We say that a decomposition of the form (15.21) is essentially unique if given another such
decomposition, Xr Xr
(1) (d) (1) (d)
αp vp ⊗ · · · ⊗ vp = βp wp ⊗ · · · ⊗ wp ,
p=1 p=1
(1) (d)
we must have (i) the coefficients αp = βp for all p = 1, . . . , r; and (ii) the factors vp , . . . , vp and
(1) (d)
wp , . . . , wp differ at most via unimodulus scaling, i.e.,

(1) (1) (d) (d)

vp = eiθ1p wp , . . . , vp = eiθdp wp

where θ1p + · · · + θdp ≡ 0 mod 2π, for all p = 1, . . . , r. In the event when successive coefficients are
equal, σp−1 > σp = σp+1 = · · · = σp+q > σp+q+1 , the uniqueness of the factors in (ii) is only up
to relabeling of indices p, . . . , p + q.

Facts:
Facts requiring proof for which no specific reference is given can be found in [Lim] and the
references therein.
1. Let A ∈ Cn1 ×···×nd . Then A has a rank-retaining decomposition
Xr
A= σp vp(1) ⊗ · · · ⊗ vp(d) , (15.21)
p=1

with r = rank(A) and σ1 , . . . , σr ∈ R, which can be chosen to be all positive, and

σ1 ≥ σ2 ≥ · · · ≥ σr ,
(k) (k)
and vp ∈ Cnk has unit norm kvp k2 = 1 all k = 1, . . . , d and all p = 1, . . . , r.
2. Fact 1 has a coordinate-free counterpart: If T ∈ V1 ⊗ · · · ⊗ Vd , then we may also write
Xr
T = σp vp(1) ⊗ · · · ⊗ vp(d) ,
p=1

(k)
where r = rank(T ), vp ∈ Vk are vectors in an abstract C-vector space. Furthermore
(k)
if the Vk ’s are all equipped with norms, then vp may all be chosen to be unit vectors
and σp may all be chosen to be positive real numbers. As such, we may rightly call
a decomposition of the form (15.21) a tensor decomposition, with Fact 1 being the
special case when Vk = Cnk .
(1) (1) (2) (2)
3. For the special case d = 2, the unit-norm vectors {v1 , . . . , vr } and {v1 , . . . , vr }
(1) (1) n1 ×r
may in fact be chosen to be orthonormal. If we write U := [v1 , . . . , vr ] ∈ C
(2) (2)
and V := [v1 , . . . , vr ] ∈ Cn2 ×r , Fact 1 yields the singular value decomposition
Tensors and Hypermatrices 15-19

A = U ΣV ∗ . Hence, Fact 1 may be viewed as a generalization of matrix singular value

decomposition. The one thing that we lost in going from order d = 2 to d ≥ 3 is the
(k) (k)
orthonormality of the “singular vectors” v1 , . . . , vr . In fact, this is entirely to be
expected because when d ≥ 2, it is often the case that rank(A) > max{n1 , . . . , nd } and
(k) (k)
so it is impossible for v1 , . . . , vr to be orthonormal or even linearly independent.
4. A decomposition of the form in Fact 1 is in general not unique if we do not impose
(k)
orthogonality on the factors {vp : p = 1, . . . , r} but this is easily seen to be im-
possible via a simple dimension count: Suppose n1 = · · · = nd = n, an orthonormal
collection of n vectors has dimension n(n − 1)/2 and so the right-hand side of Eq.
(15.21) has dimension at most n + dn(n − 1)/2 but the left-hand side has dimension
nd . Fortunately, there is an amazing result due to Kruskal that guarantees a slightly
weaker form of uniqueness of Eq. (15.21) for d ≥ 3 without requiring orthogonality.
5. It is clear that girth (= spark) and krank are one and the same notion, related by
girth(S) = krank(S) + 1.
6. The krank of v1 , . . . , vr ∈ V is GL(V )-invariant: If X ∈ GL(V ), then
krank(Xv1 , . . . , Xvr ) = krank(v1 , . . . , vr ).
7. krank(v1 , . . . , vr ) ≤ dim span{v1 , . . . , vr }.
8. [Kru77] Let A ∈ Cl×m×n . Then a decomposition of the form
Xr
A= σp up ⊗ vp ⊗ wp
p=1

is both rank-retaining, i.e., r = rank(A), and essentially unique if the following com-
binatorial condition is met:
krank(u1 , . . . , ur ) + krank(v1 , . . . , vr ) + krank(w1 , . . . , wr ) ≥ 2r + 2.
This is the Kruskal uniqueness theorem.
9. [SB00] Fact 8 has been generalized to arbitrary order d ≥ 3. Let A ∈ Cn1 ×···×nd where
d ≥ 3. Then a decomposition of the form
Xr
A= σp vp(1) ⊗ · · · ⊗ vp(d)
p=1

is both rank-retaining, i.e., r = rank A, and essentially unique if the following condi-
tion is satisfied: Xd (k)
krank(v1 , . . . , vr(k) ) ≥ 2r + d − 1. (15.22)
k=1
10. A result analogous to the Kruskal uniqueness theorem (Fact 8) does not hold for
d = 2, since a decomposition of a matrix A ∈ Cn×n can be written in infinitely
many different ways A = U V > = (U X −1 )(XV ) for any X ∈ GL(n, C). This is not
surprising since Eq. (15.22) can never be true for d = 2 because of Fact 7.
11. [Der13] For d ≥ 3, the condition (15.22) is sharp in the sense that the right-hand side
cannot be further reduced.
12. One may also write the decomposition in Fact 1 in the form of multilinear matrix
multiplication,
Xr Xr
A= λi xi ⊗ yi ⊗ zi = (X, Y, Z) · Λ, aαβγ = λi xαi yβi zγi ,
i=1 i=1

of three matrices X = [x1 , . . . , xr ] ∈ C , Y = [y1 , . . . , yr ] ∈ Cm×r , Z = [z1 , . . . , zr ] ∈

l×r

Cn×r with a diagonal hypermatrix Λ = diag[λ1 , . . . , λr ] ∈ Cr×r×r , i.e.,

(
λi i = j = k,
λijk =
0 otherwise.
15-20 Handbook of Linear Algebra

Note however that this is not a multilinear rank-retaining decomposition in the sense
of the next section, since µ rank(A) 6= (r, r, r) in general.

15.7 Multilinear Rank

For a matrix A = [aij ] ∈ F m×n we have a nice equality between three numerical invariants
associated with A:
rank(A) = dim spanF {A•1 , . . . , A•n } (15.23)
= dim spanF {A1• , . . . , Am• } (15.24)
= min{r : A = x1 y1> + ··· + xr yr> }. (15.25)

Here we let Ai• = [ai1 , . . . , aim ]T ∈ F m and A•j = [a1j , . . . , anj ]T ∈ F n denote the ith
row and jth column vectors of A. The numbers in Eqs. (15.24) and (15.23) are the row
and column ranks of A. Their equality is a standard fact in linear algebra and the common
value is called the rank of A. The number in Eq. (15.25) is also easily seen to be equal to
rank(A).
For a hypermatrix A = [aijk ] ∈ F l×m×n , one may also define analogous numbers. How-
ever, we generally have four distinct numbers, with the first three, called multilinear rank
associated with the three different ways to slice A, and the analog of Eq. (15.25) being the
tensor rank of A (cf. Section 15.3). The multilinear rank is essentially matrix rank and so
inherits many of the latter’s properties, so we do not see the sort of anomalies discussed in
Sections 15.4 and 15.5.
Like the tensor rank, the notion of multilinear rank was due to Hitchcock [Hit27b], as a
special case (2-plex rank) of his multiplex rank.

Definitions:
For a hypermatrix A = [aijk ] ∈ F l×m×n ,

r1 = dim spanF {A1•• , . . . , Al•• }, (15.26)

r2 = dim spanF {A•1• , . . . , A•m• }, (15.27)
r3 = dim spanF {A••1 , . . . , A••n }. (15.28)

Here Ai•• = [aijk ]m,n

j,k=1 ∈ F
m×n
, A•j• = [aijk ]l,n
i,k=1 ∈ F
l×n
, A••k = [aijk ]l,m
i,j=1 ∈ F
l×m
.
The multilinear rank of A ∈ F l×m×n is µ rank(A) := (r1 , r2 , r3 ), with r1 , r2 , r3 defined in
Eqs. (15.26), (15.27), and (15.28). The definition for a d-hypermatrix with d > 3 is analogous.
The kth flattening map on F n1 ×···×nd is the function

[k : F n1 ×···×nd → F nk ×(n1 ...n

b k ...nd )

defined by
([k (A))ij = (A)sk (i,j)
where sk (i, j) is the jth element in lexicographic order in the subset of hn1 i × · · · × hnd i consisting
of elements that have kth coordinate equal to i, and by convention a caret over any entry of a
d-tuple means that the respective entry is omitted. See Example 1.
The multilinear kernels or nullspaces of T ∈ U ⊗ V ⊗ W are

ker1 (T ) = {u ∈ U : T (u, v, w) = 0 for all v ∈ V, w ∈ W },

ker2 (T ) = {v ∈ V : T (u, v, w) = 0 for all u ∈ U, w ∈ W },
ker3 (T ) = {w ∈ W : T (u, v, w) = 0 for all u ∈ U, v ∈ V }.
Tensors and Hypermatrices 15-21

The multilinear images or range spaces of T ∈ U ⊗ V ⊗ W are

im1 (T ) = {A ∈ V ⊗ W : A(v, w) = T (u0 , v, w) for some u0 ∈ U },

im2 (T ) = {B ∈ U ⊗ W : B(u, w) = T (u, v0 , w) for some v0 ∈ V },
im3 (T ) = {C ∈ U ⊗ V : C(u, v) = T (u, v, w0 ) for some w0 ∈ W }.

The multilinear nullity of T ∈ U ⊗ V ⊗ W is

µ nullity(T ) = (dim ker1 (T ), dim ker2 (T ), dim ker3 (T )).

The multilinear rank of T ∈ U ⊗ V ⊗ W is

µ rank(T ) = (dim im1 (T ), dim im2 (T ), dim im3 (T )).

Facts:
Facts requiring proof for which no specific reference is given can be found in [Lan12, Chap. 2]
as well as [Lim] and the references therein.
1. For A ∈ F l×m×n , the multilinear rank µ rank(A) = (r1 , r2 , r3 ) is given by

r1 = rank([1 (A)), r2 = rank([2 (A)), r3 = rank([3 (A)),

where rank here is of course the usual matrix rank of the matrices [1 (A), [2 (A), [3 (A).
2. Let A ∈ F l×m×n . For any X ∈ GL(l, F ), Y ∈ GL(m, F ), Z ∈ GL(n, F ),

µ rank((X, Y, Z) · A) = µ rank(A).

3. Let T ∈ V1 ⊗ · · · ⊗ Vd . For bases Bk for Vk , k = 1, . . . , d,

µ rank([T ]B1 ,...,Bd ) = µ rank(T ).

4. Let T ∈ U ⊗ V ⊗ W . The rank-nullity theorem for multilinear rank is:

µ nullity(T ) + µ rank(T ) = (dim U, dim V, dim W ).

5. If A ∈ F l×m×n has µ rank(A) = (p, q, r), then there exist matrices X ∈ F l×p , Y ∈
F m×q , Z ∈ F n×r of full column rank and C ∈ Cp×q×r such that

A = (X, Y, Z) · C,

i.e., A may be expressed as a multilinear matrix product of C by X, Y, Z, defined

according to Eq. (15.1). We shall call this a multilinear rank-retaining decomposition
or just multilinear decomposition for short.
6. One may also write Fact 5 in the form of a multiply indexed sum of decomposable
hypermatrices Xp,q,r
A= cijk xi ⊗ yj ⊗ zk
i,j,k=1
l m n
where xi ∈ F , yj ∈ F , zk ∈ F are the respective columns of the matrices X, Y, Z.
7. Let A ∈ Cl×m×n and µ rank(A) = (p, q, r). There exist L1 ∈ Cl×p , L2 ∈ Cm×q ,
L3 ∈ Cn×r unit lower triangular and U ∈ Cp×q×r such that

A = (L1 , L2 , L3 ) · U.

This may be viewed as the analogues of the LU decomposition for hypermatrices with
respect to the notion of multilinear rank.
15-22 Handbook of Linear Algebra

8. Let A ∈ Cl×m×n and µ rank(A) = (p, q, r). There exist Q1 ∈ Cl×p , Q2 ∈ Cm×q ,
Q3 ∈ Cn×r with orthonormal columns and R ∈ Cp×q×r such that
A = (Q1 , Q2 , Q3 ) · R.
This may be viewed as the analogues of the QR decomposition for hypermatrices with
respect to the notion of multilinear rank.
9. The LU and QR decompositions in Facts 7 and 8 are among the few things that
can be computed for hypermatrices, primarily because multilinear rank is essen-
tially a matrix notion. For example, one may apply usual Gaussian elimination or
Householder/Givens QR to the flattenings of A ∈ F l×m×n successively, reducing all
computations to standard matrix computations.

Examples:

1. Let  
a111 a121 a131 a112 a122 a132
a211 a221 a231 a212 a222 a232 
A=  ∈ C4×3×2 ,
a311 a321 a331 a312 a322 a332 
a411 a421 a431 a412 a422 a432
then
 
a111 a112 a121 a122 a131 a132
a211 a212 a221 a222 a231 a232 
[1 (A) =   ∈ C4×6 ,
a311 a312 a321 a322 a331 a332 
a411 a412 a421 a422 a431 a432
 
a111 a112 a211 a212 a311 a312 a411 a412
[2 (A) = a121 a122 a221 a222 a321 a322 a421 a422  ∈ C3×8 ,
a131 a132 a231 a232 a331 a332 a431 a432
 
a a a a a a a a a a a a
[3 (A) = 111 121 131 211 221 231 311 321 331 411 421 431 ∈ C2×12 .
a112 a122 a132 a212 a222 a232 a312 a322 a332 a412 a422 a432

2. Note that if we had a bilinear form represented by a matrix M ∈ Cm×n , the analogues of
these subspaces would be the four fundamental subspaces of the matrix: ker1 (M ) = ker(M ),
im1 (M ) = im(M ), ker2 (M ) = ker(M T ), im2 (M ) = im(M T ). The rank-nullity theorem
reduces to the usual one:

(nullity(M ) + rank(M ), nullity(M T ) + rank(M T )) = (n, m).

15.8 Norms
In this section we will discuss the Hölder, induced, and nuclear norms of hypermatrices.
When discussing multiple variety of norms, one has to introduce different notation to dis-
tinguish them and here we follow essentially the notation and terminology for matrix norms
in Chapter 24, adapted as needed for hypermatrices. For the induced and nuclear norms,
we assume d = 3 to avoid notational clutter.

Definitions:
For A = [aj1 ···jd ]n 1 ,...,nd
j1 ,...,jd =1 ∈ C
n1 ×···×nd
and p ∈ [1, ∞], the Hölder p-norm is defined by
Xn1 ,...,nd 1/p
kAkH,p := |aj1 ···jd |p , (15.29)
j1 ,...,jd =1
Tensors and Hypermatrices 15-23

with the usual alternate definition for p = ∞,

kAkH,∞ := max{|aj1 ···jd | : j1 = 1, . . . , n1 ; . . . ; jd = 1, . . . , nd }.

kAkH,2 is often denoted kAkF and called the Frobenius norm or Hilbert–Schmidt norm of
the hypermatrix A ∈ Cn1 ×···×nd .
For hypermatrices, the induced norm, natural, or operator norm is defined for p, q, r ∈
[1, ∞], by the quotient
|A(x, y, z)|
kAkp,q,r := max (15.30)
x,y,z6=0 kxk p kykq kzkr

where
Xl,m,n
A(x, y, z) = aijk xi yj zjk .
i,j,k=1

The special case p = q = r = 2, i.e., k · k2,2,2 is the spectral norm.

Let A ∈ Cl×m×n . The nuclear norm or Schatten 1-norm of A is defined as
X
r Xr
kAkS,1 := min |λi | : A = λi ui ⊗ vi ⊗ wi ,
i=1 i=1

kui k2 = kvi k2 = kwi k2 = 1, r ∈ N . (15.31)

For hypermatrices A ∈ Cn1 ×···×nd of arbitrary order, use the obvious generalizations, and of
course for p = ∞, replace the sum by max{|λ1 |, . . . , |λr |}.
The inner product on Cn1 ×···×nd obtained from the dot product on Cn1 ···nd by viewing hyper-
matrices as complex vectors of dimension n1 · · · nd , the dot product will be denoted by
Xn1 ,...,nd
hA, Bi := aj1 ···jd b̄j1 ···jd . (15.32)
j1 ,...,jd =1

Facts:
Facts requiring proof for which no specific reference is given can be found in [DF93, Chap. I]
as well as [Lim] and the references therein.

1. The hypermatrix Hölder p-norm kAkH,p of A ∈ Cn1 ×···×nd is the (vector Hölder)
p-norm of A when regarded as a vector of dimension n1 · · · nd (see Section 50.1).
2. The Hölder p-norm has the property of being multiplicative on rank-1 hypermatrices
in the following sense:

ku ⊗ v ⊗ · · · ⊗ zkH,p = kukp kvkp · · · kzkp ,

where u ∈ Cn1 , v ∈ Cn2 , . . . , z ∈ Cnd and the norms on these are the usual p-norms
on vectors.
3. The Frobenius norm is the norm (length) in the inner product space Cn1 ×···×nd with
Eq. (15.32).
4. The inner product (15.32) (and thus the Frobenius norm) is invariant under multi-
linear matrix multiplication by unitary matrices,

h(Q1 , . . . , Qd ) · A, (Q1 , . . . , Qd ) · Bi = hA, Bi, k(Q1 , . . . , Qd ) · AkF = kAkF

for any Q1 ∈ U(n1 ), . . . , Qd ∈ U(nd ).

5. The inner product (15.32) satisfies a Cauchy-Schwarz inequality

|hA, Bi| ≤ kAkF kBkF ,

15-24 Handbook of Linear Algebra

and more generally a Hölder inequality

1 1
|hA, Bi| ≤ kAkH,p kBkH,q , + = 1, p, q ∈ [1, ∞].
p q
6. The induced (p, q, r)-norm is a norm on Cl×m×n (with norm as defined in Sec-
tion 50.1).
7. Alternate expressions for the induced norms include
kAkp,q,r = max{|A(u, v, w)| : kukp = kvkq = kwkr = 1}
= max{|A(x, y, z)| : kxkp ≤ 1, kykq ≤ 1, kzkr ≤ 1}.
8. The induced norm is multiplicative on rank-1 hypermatrices
ku ⊗ v ⊗ wkp,q,r = kukp kvkq kwkr
for u ∈ Cl , v ∈ Cm , w ∈ Cn .
1 1
9. For square matrices A ∈ Cn×n and p + q = 1,

|A(x, y)| |xT Ay| kAxkq

kAkp,q = max = max = max ,
x,y6=0 kxkp kykq x,y6=0 kxkp kykq x6=0 kxkp

the matrix (p, q)-norm.

10. The spectral norm is invariant under mutilinear matrix multiplication
k(Q1 , . . . , Qd ) · Ak2,...,2 = kAk2,...,2
for any Q1 ∈ U(n1 ), . . . , Qd ∈ U(nd ).
11. When d = 2, the nuclear norm defined above reduces to the nuclear norm (i.e.,
Schatten 1-norm) of a matrix A ∈ Cm×n defined more commonly by
Xmax{m,n}
kAkS,1 = σi (A).
i=1

12. [DF93] The nuclear norm defines a norm on Cn1 ×···×nd (with norm as defined in
Section 50.1).
13. The nuclear norm is invariant under mutilinear unitary matrix multiplication, i.e.,
k(Q1 , . . . , Qd ) · AkS,1 = kAkS,1
for any Q1 ∈ U(n1 ), . . . , Qd ∈ U(nd ).
14. The nuclear norm and the spectral norm are dual norms to each other, i.e.,
kAkS,1 = max{|hA, Bi| : kBk2,...,2 = 1},
kAk2,...,2 = max{|hA, Bi| : kBkS,1 = 1}.

15. Since Cn1 ×···×nd is finite dimensional, all norms are necessarily equivalent (see Sec-
tion 50.1). Given that all norms induce the same topology, questions involving con-
vergence of sequences of hypermatrices, whether a set of hypermatrices is closed, etc.,
are independent of the choice of norms.

Examples:

1. [Der13a] Let Tn be the matrix multiplication tensor in Eq. (15.19). Then kTn kF = n3/2 ,
kTn kS,1 = n3 , and kTn k2,2,2 = 1.
Tensors and Hypermatrices 15-25

15.9 Hyperdeterminants
There are two ways to extend the determinant of a matrix to hypermatrices of higher
order. One is to simply extend the usual expression of an n × n matrix determinant as
a sum of n! monomials in the entries of the matrix, which we will call the combinatorial
hyperdeterminant, and the other, which we will call the geometric hyperdeterminant, is by
using the characterization that a matrix has det A = 0 if and only if Ax = 0 has nonzero
solutions. Both approaches were proposed by Cayley [Cay49, Cay45], who also gave the
explicit expression of a 2 × 2 × 2 geometric hyperdeterminant. Gelfand, Kapranov, and
Zelevinsky [GKZ94, GKZ92] have a vast generalization of Cayley’s result describing the
dimensions of hypermatrices for which geometric hyperdeterminants exist.

Definitions:
The combinatorial hyperdeterminant of a cubical d-hypermatrix A = [ai1 i2 ···id ] ∈ F n×···×n is
defined as
n
1 X Y
det(A) = sgn π1 · · · sgn πd aπ1 (i)···πd (i) . (15.33)
n! π1 ,...,πd ∈Sn
i=1

If it exists for a given set of dimensions (n1 , . . . , nd ), a geometric hyperdeterminant or hy-

perdeterminant Detn1 ,...,nd (A) of A ∈ Cn1 ×···×nd is a homogeneous polynomial in the entries of
A such that Detn1 ,...,nd (A) = 0 if and only if the system of multilinear equations ∇A(x1 , . . . , xd ) =
0 has a nontrivial solution, i.e., x1 , . . . , xd all nonzero. Here ∇A(x1 , . . . , xd ) is the gradient of
A(x1 , . . . , xd ) viewed as a function of the coordinates of the vector variables x1 , . . . , xd (see Pre-
liminaries for the definition of gradient). See Fact 4 for conditions that describe the n1 , . . . , nd for
which Detn1 ,...,nd exists.

Facts:
Facts requiring proof for which no specific reference is given may be found in [GKZ94,
Chap. 13].

1. For odd order d, the combinatorial hyperdeterminant of a cubical d-hypermatrix is

identically zero.
2. For even order d, the combinatorial hyperdeterminant of a cubical d-hypermatrix
A = [ai1 i2 ···id ] is

X n
Y
det(A) = sgn(π2 · · · πd ) aiπ2 (i)···πd (i) .
π2 ,...,πd ∈Sn
i=1

For d = 2, this reduces to the usual expression for the determinant of an n×n matrix.
3. Taking the π-transpose of an order d-hypermatrix A ∈ F n×···×n leaves the combina-
torial hyperdetermiant invariant, i.e., det Aπ = det A for and π ∈ Sd .
4. (Gelfand–Kapranov–Zelevinsky) A geometric hyperdeterminant exists for Cn1 ×···×nd
if and only if for all k = 1, . . . , d, the following dimension condition is satisfied:
X
nk − 1 ≤ (nj − 1).
j6=k

5. When the dimension condition in Fact 4 is met, a geometric hyperdeterminant

Detn1 ,...,nd (A) is a multivariate homogeneous polynomial in the entries of A that
is unique up to a scalar multiple, and can be scaled to have integer coefficients with
greatest common divisor one (the latter is referred to as the geometric hyperdetermi-
nant and is unique up to ±1 multiple).
15-26 Handbook of Linear Algebra

6. The geometric hyperdeterminant exists for all cubical hypermatrices (and by defi-
nition is nontrivial when it exists, including for odd order cubical hypermatrices).
The geometric hyperdeterminant also exists for some non-cubical hypermatrices (see
Fact 3).
7. For Cm×n , the dimension condition in Fact 4 is m ≤ n and n ≤ m, which may be
viewed as a reason why matrix determinants are only defined for square matrices.
8. It is in general nontrivial to find an explicit expression for Detn1 ,...,nd (A). While
systematic methods for finding it exists, notably one due to Schläfli [GKZ94], an
expression may nonetheless contain a large number of terms when expressed as a sum
of monomials. For example, Det2,2,2,2 (A) has more than 2.8 million monomial terms
[GHS08] even though a hypermatrix A ∈ C2×2×2×2 has only 16 entries.
9. Unlike rank and border rank (cf. Facts 15.3.2 and 15.4.5), the geometric hyperde-
terminant is not invariant under multilinear matrix multiplication by nonsingular
matrices (which is expected since ordinary matrix determinant is not invariant under
left and right multiplications by nonsingular matrices either). However, it is relatively
invariant in the following sense [GKZ94].
Let n1 , . . . , nd satisfy the dimension condition in Fact 4. Then for any A ∈
Cn1 ×···×nd and any X1 ∈ GL(n1 , C), . . . , Xd ∈ GL(nd , C),
Detn1 ,...,nd ((X1 , . . . , Xd ) · A) = det(X1 )m/n1 · · · det(Xd )m/nd Detn1 ,...,nd (A) (15.34)
where m is the degree of Detn1 ,...,nd . Hence for X1 ∈ SL(n1 , C), . . . , Xd ∈ SL(nd , C),
we get
Detn1 ,...,nd ((X1 , . . . , Xd ) · A) = Detn1 ,...,nd (A).
10. A consequence of Fact 9 is the properties of the usual matrix determinant under
row/column interchanges, addition of a scalar multiple of row/column to another,
etc., are also true for the hyperdeterminant. For notational convenience, we shall
just state the following corollary for a 3-hypermatrix although it is true in general for
any d-hypermatrix [GKZ94]. We use the notation in Example 15.1.3 where A1 , . . . , An
denote the slices of the 3-hypermatrix A, analogues of rows or columns of a matrix. Let
l, m, n satisfy the dimension condition in Fact 4 and let A = [A1 | . . . |An ] ∈ Cl×m×n .
Then
(a) interchanging two slices leaves the hyperdeterminant invariant up to sign:
Detl,m,n ([A1 |...|Ai |...|Aj |...|An ]) = ± Detl,m,n ([A1 |...|Aj |...|Ai |...|An ]);

(b) adding a scalar multiple of a slice to another leaves the hyperdeterminant in-
variant:
Detl,m,n ([A1 |...|Ai |...|Aj |...|An ]) = Detl,m,n ([A1 |...|Ai + αAj |...|Aj |...|An ])
for any α ∈ C;
(c) having two proportional slices implies that the hyperdetermiant vanishes:
Detl,m,n ([A1 |...|Ai |...|αAi |...|An ]) = 0 for any α ∈ C;

(d) taking π-transpose leaves the hyperdetermiant invariant

Detl,m,n (Aπ ) = Detl,m,n (A) for any π ∈ S3 .

In particular, the last property implies that the other three properties hold for slices
of A in any other fixed index.
11. Detn1 ,...,nd (A) = 0 if and only if 0 is a singular value of A.
Tensors and Hypermatrices 15-27

Examples:

1. For A = [aijk` ] ∈ F 2×2×2×2 , the combinatorial hyperdeterminant of A is

det A = a1111 a2222 − a1112 a2221 − a1121 a2212 + a1122 a2211

− a1211 a2122 + a1122 a2211 + a1221 a2112 + a1212 a2121 .

This is immediate from Fact 2, or by applying the definition (15.33) and simplifying.
2. [Cay45] The geometric hyperdeterminant of A = [A1 |A2 ] = [aijk ] ∈ C2×2×2 (where A1 and
A2 slices of A, cf. Section 15.1) is

1
Det2,2,2 (A) = [det(A1 + A2 ) − det(A1 − A2 )]2 − 4 det[A1 ] det[A2 ]
4
        2
1 a111 a112 a a221 a111 a121 a a221
= det + 211 − det − 211
4 a121 a122 a212 a222 a112 a122 a212 a222
   
a a121 a a221
− 4 det 111 det 211 .
a112 a122 a212 a222
The statement that the gradient ∇A(x, y, z) of the trilinear functional defined by A vanishes
for some nonzero vectors x, y, z is the statement that the system of bilinear equations,

a111 x1 y1 + a121 x1 y2 + a211 x2 y1 + a221 x2 y2 = 0,

a112 x1 y1 + a122 x1 y2 + a212 x2 y1 + a222 x2 y2 = 0,
a111 x1 z1 + a112 x1 z2 + a211 x2 z1 + a212 x2 z2 = 0,
a121 x1 z1 + a122 x1 z2 + a221 x2 z1 + a222 x2 z2 = 0,
a111 y1 z1 + a112 y1 z2 + a121 y2 z1 + a122 y2 z2 = 0,
a211 y1 z1 + a212 y1 z2 + a221 y2 z1 + a222 y2 z2 = 0,

has a non-trivial solution (x, y, z all non-zero).

3. The geometric hyperdeterminant of A = [aijk ] ∈ C2×2×3 is
   
a111 a112 a113 a211 a212 a213
Det2,2,3 (A) = det a211 a212 a213  det a121 a122 a123 
a121 a122 a123 a221 a222 a223
   
a111 a112 a113 a111 a112 a113
− det a211 a212 a213  det a121 a122 a123  .
a221 a222 a223 a221 a222 a223
The statement that the gradient ∇A(x, y, z) of the trilinear functional defined by A vanishes
for some nonzero vectors x, y, z is the statement that the system of bilinear equations,

a111 x1 y1 + a121 x1 y2 + a211 x2 y1 + a221 x2 y2 = 0,

a112 x1 y1 + a122 x1 y2 + a212 x2 y1 + a222 x2 y2 = 0,
a113 x1 y1 + a123 x1 y2 + a213 x2 y1 + a223 x2 y2 = 0,
a111 x1 z1 + a112 x1 z2 + a113 x1 z3 + a211 x2 z1 + a212 x2 z2 + a213 x2 z3 = 0,
a121 x1 z1 + a122 x1 z2 + a123 x1 z3 + a221 x2 z1 + a222 x2 z2 + a223 x2 z3 = 0,
a111 y1 z1 + a112 y1 z2 + a113 y1 z3 + a121 y2 z1 + a122 y2 z2 + a123 y2 z3 = 0,
a211 y1 z1 + a212 y1 z2 + a213 y1 z3 + a221 y2 z1 + a222 y2 z2 + a223 y2 z3 = 0,

has a non-trivial solution (x, y, z all non-zero).

15-28 Handbook of Linear Algebra

Applications:

1. In Example 15.3.4, we saw how one may characterize the notion of a pure state in terms of
the rank of a hypermatrix, namely rank(A) = 1. The hyperdeterminant characterizes the
dual notion — the degenerate entangled states are precisely the hypermatrices A ∈ C2×2×2
with Det2,2,2 (A) = 0 [MW02]. More generally, when Det2,2,2 (A) 6= 0, the magnitude of the
hyperdeterminant |Det2,2,2 (A)| gives another way to measure the amount of entanglement
called the concurrence [HW97].

15.10 Odds and Ends

Because of page constraints, many topics have to be omitted from this chapter: covariance
and contravariance, tensor fields and hypermatrix-valued functions, symmetric tensors and
symmetric hypermatrices, alternating tensors and skew-symmetric hypermatrices, eigen-
values and singular values of hypermatrices, positive definite hypermatrices and Cholesky
decomposition, nonnegative hypermatrices and the Perron-Frobenius theorem, the Dirac
and Einstein notations, tensor products of modules, of Hilbert and Banach spaces, of func-
tions, of operators, of representations, etc.
Also, topics that are best presented in a usual narrative format (as opposed to a hand-
book format) are not included in the chapter. These include discussions of the difference
between hypermatrices and tensors, the existence of canonical forms for hypermatrices, the
duality between the geometric hyperdeterminant and tensor rank, questions relating to com-
putability and complexity, as well as connections to algebraic geometry and representation
theory.
As a consequence, there are many additonal interesting applications and examples that are
not included in this chapter because they require one or more of these omitted topics. These
include self-concordance and higher optimality conditions in optimization theory, Riemann
curvature tensor and the Einstein field equations in general relativity, electromagnetic field
tensor in gauge theory, linear piezoelectric equations in continuum mechanics, Fourier coeffi-
cients of triply-periodic functions in X-ray crystallography, other forms of the Yang–Baxter
equation, density matrix renormalization group (dmrg) techniques in quantum chemistry,
the Salmon conjecture (now a theorem) in phylogenetics, moments and cumulants in statis-
tics, polynomial Mercer kernels and naı̈ve Bayes model in machine learning, and blind source
separation and independent components analysis in signal processing.
We refer the reader to [Lim] for an expanded treatment that covers these and other topics.

Acknowledgment
The author would like to express his heartfelt gratitude to Leslie Hogben for her tireless
editorial efforts, transforming a manuscript that is an earlier version of [Lim] into the HLA
format that one sees here. It is an arduous undertaking without which this chapter would
not have made it in time for inclusion in this volume. The author would also like to thank
Harm Derksen, Shenglong Hu, J. M. Landsberg, Peter McCullagh, and Ke Ye for their very
helpful comments.

References
[AW92] W. A. Adkins and S. H. Weintraub. Algebra: An Approach Via Module Theory.
Springer-Verlag, New York, 1992.
[Bax78] R. J. Baxter. Solvable eight-vertex model on an arbitrary planar lattice. Philos. Trans.
Tensors and Hypermatrices 15-29

Roy. Soc. London Ser. A, 289 (1359): 315–346, 1978.

[Ber69] G. M. Bergman, Ranks of tensors and change of base field, J. Algebra, 11: 613–621,
1969.
[BLR80] D. Bini, G. Lotti, and F. Romani. Approximate solutions for the bilinear form compu-
tational problem. SIAM J. Comput., 9 (4): 692–697, 1980.
[Bor90] S. F. Borg. Matrix-Tensor Methods in Continuum Mechanics, 2nd ed. World Scientific,
Singapore, 1990.
[Bou98] N. Bourbaki. Algebra I: Chapters 1–3. Springer-Verlag, Berlin, 1998.
[Bry02] J.-L. Brylinski. Algebraic measures of entanglement. In R. K. Brylinski and G. Chen
(Eds.), Mathematics of Quantum Computation, pp. 3–23, CRC Press, Boca Raton,
FL, 2002.
[BCS96] P. Bürgisser, M. Clausen, and M. A. Shokrollahi. Algebraic Complexity Theory,
Springer-Verlag, Berlin, 1996.
[CO65] T. K. Caughey, M. E. J. O’Kelly. Classical normal modes in damped linear dynamic
systems. ASME J. Applied Mechanics, 32: 583—588, 1965.
[Cay45] A. Cayley. On the theory of linear transformations. Cambridge Math. J., 4: 193–209,
1854.
[Cay49] A. Cayley. On the theory of determinants. Trans. Cambridge Philos. Soc., 8, no. 7: 75–
88, 1849.
[Cor84] J. F. Cornwell. Group Theory in Physics, Vol. II. Academic Press, London, 1984.
[DF93] A. Defant and K. Floret. Tensor Norms and Operator Ideals. North-Holland, Amster-
dam, 1993.
[Der13] H. Derksen. Sharpness of Kruskal’s theorem. Linear Algebra Appl., 438(2): 708–712,
2013.
[Der13a] H. Derksen. Personal communication, 2013.
[DE03] D. L. Donoho and M. Elad. Optimally sparse representation in general (nonorthogonal)
dictionaries via `1 minimization. Proc. Nat. Acad. Sci., 100(5): 2197–2202, 2003.
[DVC00] W. Dür, G. Vidal, and J. I. Cirac. Three qubits can be entangled in two inequivalent
ways. Phys. Rev. A, 62(6), 062314, 12 pp., 2000.
[GKZ94] I. M. Gelfand, M. M. Kapranov, and A. V. Zelevinsky. Discriminants, Resultants,
and Multidimensional Determinants. Birkhäuser Publishing, Boston, MA, 1994.
[GKZ92] I. M. Gelfand, M. M. Kapranov, and A. V. Zelevinsky. Hyperdeterminants. Adv. Math.,
96(2): 226–263, 1992.
[GHZ89] D. M. Greenberger, M. A. Horne, and A. Zeilinger. Going beyond Bell’s theorem, In M.
Kafatos (Ed.), Bell’s Theorem, Quantum Theory and Conceptions of the Universe,
pp. 69–72, Kluwer, Dordrecht, Netherlands, 1989.
[Gre78] W. Greub. Multilinear Algebra, 2nd ed. Springer-Verlag, New York, 1978.
[GHS08] D. Grier, P. Huggins, B. Sturmfels, and J. Yu. The hyperdeterminant and triangulations
of the 4-cube. Math. Comp., 77(263): 1653–1679, 2008.
[Hal85] P. Halmos. I Want to Be a Mathematician. . . An Automathography. Springer-Verlag,
New York, 1985.
[HW97] S. Hill and W. K. Wootters. Entanglement of a pair of quantum bits. Phys. Rev. Lett.,
78(26): 5022–5025, 1997.
[Hit27a] F. L. Hitchcock. The expression of a tensor or a polyadic as a sum of products. J. Math.
Phys., 6(1): 164–189, 1927.
[Hit27b] F. L. Hitchcock. Multiple invariants and generalized rank of a p-way matrix or tensor.
J. Math. Phys., 7(1): 39–79, 1927.
15-30 Handbook of Linear Algebra

[HK71] J. E. Hopcroft and L. R. Kerr. On minimizing the number of multiplications necessary

for matrix multiplication. SIAM J. Appl. Math., 20(1): 30–36, 1971.
[Knu98] D. Knuth. The Art of Computer Programming 2: Seminumerical Algorithms, 3rd ed.
Addison–Wesley, Reading, MA, 1998.
[KM97] A. I. Kostrikin and Y. I. Manin. Linear Algebra and Geometry. Gordon and Breach,
Amsterdam, 1997.
[Kru77] J. B. Kruskal. Three-way arrays: rank and uniqueness of trilinear decompositions, with
application to arithmetic complexity and statistics. Lin. Alg. Appl., 18(2): 95–138,
1977.
[Lan06] J. M. Landsberg. The border rank of the multiplication of 2 × 2 matrices is seven. J.
Amer. Math. Soc., 19(2): 447–459, 2006.
[Lan12] J. M. Landsberg. Tensors: Geometry and Applications. AMS, Providence, RI, 2012.
[Lan02] S. Lang. Algebra, Rev. 3rd ed. Springer-Verlag, New York, 2002.
[Lim] L.-H. Lim. Tensors and hypermatrices. Preprint, http://www.stat.uchicago.edu/
~lekheng/work/tensors.pdf.
[Mar23] M. Marcus. Finite Dimensional Multilinear Algebra, Parts I and II. Marcel Dekker,
New York, 1973 and 1975.
[MW02] A. Miyake and M. Wadati. Multipartite entanglement and hyperdeterminant. Quantum
Inf. Comput., 2: 540–555, 2002.
[Nor84] D. G. Northcott. Multilinear Algebra. Cambridge University Press, Cambridge, 1984.
[Oxl11] J. G. Oxley. Matroid Theory, 2nd ed. Oxford University Press, Oxford, 2011.
[SB00] N. D. Sidiropoulos and R. Bro. On the uniqueness of multilinear decomposition of N -way
arrays. J. Chemometrics, 14(3): 229–239, 2000.
[Str69] V. Strassen. Gaussian elimination is not optimal. Numer. Math., 13(4): 354–356, 1969.
[Str83] V. Strassen. Rank and optimal computation of generic tensors. Lin. Alg. Appl.,
52/53: 645–685, 1983.
[Str73] V. Strassen. Vermeidung von Divisionen. J. Reine Angew. Math., 264: 184–202, 1973.
[Win71] S. Winograd. On multiplication of 2 × 2 matrices, Lin. Alg. Appl., 4(4): 381–388, 1971.
[Yan67] C. N. Yang. Some exact results for the many-body problem in one dimension with
repulsive delta-function interaction. Phys. Rev. Lett., 19(23): 1312–1315, 1967.
[Yok92] T. Yokonuma. Tensor Spaces and Exterior Algebra. AMS, Providence, RI, 1992.