1. No category

arXiv:1411.6805v4 [math.CO] 3 Dec 2014
Positive Semi-Definiteness of Generalized Anti-Circulant
Tensors
Guoyin Li
∗
Liqun Qi† Qun Wang
‡
December 4, 2014
Abstract
Anti-circulant tensors have applications in exponential data fitting. They are special
Hankel tensors. In this paper, we extend the definition of anti-circulant tensors to
generalized anti-circulant tensors by introducing a circulant index r such that the
entries of the generating vector of a Hankel tensor are circulant with module r. For
an anti-circulant tensor, r = n, where n is the dimension of the Hankel tensor. For
generalized anti-circulant tensors, we allow r ≤ n. Hence, generalized anti-circulant
tensors are still special Hankel tensors. For r ≤ 4, we give necessary and sufficient
conditions for positive semi-definiteness of even order generalized anti-circulant tensors,
and show that in these cases, they are SOS tensors. This shows that there are no PNS
Hankel tensors in these cases. For r ≥ 5, we show that these results are still true in
the matrix case and some other cases.
Key words: Anti-circulant tensors, generalized anti-circulant tensor, generating
vectors, circulant index, positive semi-definiteness. AMS subject classifications
(2010): 15A18; 15A69
1
Introduction
∗
Department of Applied Mathematics, University of New South Wales, Sydney 2052, Australia. E-mail:
[email protected] (G. Li). This author’s work was partially supported by Australian Research Council.
†
Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon,
Hong Kong. E-mail: [email protected] (L. Qi). This author’s work was partially supported by the
Hong Kong Research Grant Council (Grant No. PolyU 502510, 502111, 501212 and 501913).
‡
Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon,
Hong Kong. Email: [email protected] (Q. Wang).
1
Anti-circulant tensors was introduced in [4]. They are extensions of anti-circulant matrices in matrix theory [3]. They have applications in exponential data fitting [4]. Anticirculant tensors are Hankel tensors. Hankel tensors arise from signal processing and some
other applications [1, 15, 17]. In this paper, we extend anti-circulant tensors to generalized
anti-circulant tensors, which are still Hankel tensors, and present conditions for positive
semi-definiteness of generalized anti-circulant tensors.
Let v = (v0 , · · · , v(n−1)m )⊤ ∈ ℜ(n−1)m+1 , where m, n ≥ 2. An mth order n dimensional
Hankel tensor A = (ai1 ···im ) is defined by
ai1 ···im = vi1 +···+im −m ,
for i1 , · · · , im = 1, · · · , n. If
vi = vi+r ,
(1)
for i = 0, · · · , (n−1)m−r, where 1 ≤ r ≤ n, then A is called a generalized anti-circulant
tensor with circulant index r. If r = n, then A is an anti-circulant tensor according to
[4].
For x = (x1 , · · · , xn )⊤ ∈ ℜn , A uniquely define a homogeneous polynomial
⊗m
f (x) ≡ Ax
=
n
X
ai1 ···im xi1 · · · xim =
n
X
vi1 +···+im −m xi1 · · · xim .
(2)
i1 ,···,im =1
i1 ,···,im =1
We call such a polynomial a Hankel polynomial.
Suppose that m = 2k is even. In (2), if f (x) ≥ 0 for all x ∈ ℜn , then f is called a
PSD (positive semi-definite) Hankel polynomial and A is called a PSD Hankel tensor
[16]. Denote 0 = (0, · · · , 0)⊤ ∈ ℜn . If f (x) > 0 for all x ∈ ℜn , x 6= 0, then f and A are
called positive definite. Clearly, the only odd order PSD Hankel tensor is O and there
are no odd order positive definite Hankel tensors. Thus, we may assume that m is even in
our discussion. If f can be decomposed to the sum of squares of polynomials of degree k,
then f is called an SOS Hankel polynomial and A is called an SOS Hankel tensor
[6, 7, 11, 14]. Clearly, an SOS Hankel tensor is a PSD Hankel tensor but not vice versa. By
[17], a necessary condition for A to be PSD is that
vjm ≥ 0,
(3)
for j = 0, · · · , n − 1. Let ei be the ith unit vector in ℜn . Substitute them to (2). Then we
get (3) directly. Vector v is called the generating vector of A. It may also generate a
(nk − k + 1) × (nk − k + 1) Hankel matrix A = (aij ) by
aij = vi+j−2
for i, j = 1, · · · , nk − k + 1. If the associated Hankel matrix A is PSD, then the Hankel
tensor A is called a strong Hankel tensor [17]. In [11], it was proved that an even order
2
strong Hankel tensor is an SOS Hankel tensor. Then, a necessary condition for A to be a
strong Hankel tensor is that
v2j ≥ 0,
(4)
for j = 0, · · · , (n − 1)k.
When r is odd, for r = 1 or 3, we show that A is PSD if and only if v0 = · · · = vr ≥ 0.
In these cases, we show that
f (x) = (x1 + · · · + xn )m ,
and A is a strong Hankel tensor.
When r is even, for r = 2 or r = 4, we show that A is PSD if and only if v0 = v2 = · · · =
vr , v1 = v3 = · · · = vr−1 , and v0 ≥ |v1 |. In these cases, we may write v1 = v0 (1 − 2t), where
t ∈ [0, 1]. We show that
m
f (x) = tv0 (x1 + · · · xn )m + (1 − t)v0 x1 − x2 + x3 − · · · + (−1)n−1 xn ,
and A is a strong Hankel tensor.
We show that these two results are true in the matrix case for all r ≥ 1. In fact, in the
matrix case, for even r, we show the result is true as long as 2 ≤ r ≡ 2p ≤ 2n − 4. We
believe that our results are new even in the matrix case.
In general, we show that these two results are true for the case that m = 4 and n ≥ r ≥ 5,
and the case that m = 4 and n ≥ r = 6, respectively. We think that these two results are
still true for any m = 2k ≥ 4, n ≥ r ≥ 5, though the proofs might be very complicated.
The significance of our results are twofold: On one hand, they enrich the theory of
positive semi-definiteness of even order symmetric tensors [2, 12, 13, 16, 17, 18, 20, 21].
On the other hand, they also contribute to the study of the Hilbert-Hankel problem. In
1988, Hilbert [5] showed that a PSD symmetric homogeneous polynomial of n variables and
degree m = 2k is always an SOS polynomial only in the following three cases: 1. m = 2;
2. n = 2; and 3. m = 4, n = 3. For the other pairs of m = 2k ≥ 4 and n ≥ 3, there are
always PSD polynomials which are not SOS polynomials. We may call such polynomials
PNS (PSD non-SOS) polynomials. Later, various PNS polynomials are found [19]. Since
each homogeneous polynomial of n variables and degree m is uniquely corresponding to an
mth order n dimensional symmetric tensor, the Hilbert discovery may also be stated in the
tensor language. In [11], a question is raised: are all the PSD Hankel tensors SOS tensors?
If the answer to this question is “yes”, then the problem for determining a given even order
Hankel tensor is a PSD tensor or not can be solved by a linear semi-definite programming
problem, i.e., solved in polynomial time. In [10], a conjecture is made that there are no
sixth order three dimensional PNS Hankel tensors. Since generalized anti-circulant tensors
are Hankel tensors, our results are related to this Hilbert-Hankel problem.
In the next two sections, we discuss the problem for odd r and even r respectively. Some
final remarks are made in Section 4.
3
The Case that r is Odd
2
2.1
The Case that r = 1
This case is trivial. However, it covers the sufficiency part of the results for the cases
that r ≥ 3 and is odd. We now present this result.
Proposition 1 Suppose that A is an mth order n dimensional generalized anti-circulant
tensor with circulant index 1, where m = 2k ≥ 2 and n ≥ 2. Then A is PSD if and only if
v0 ≥ 0. In this case, we have
f (x) = v0 (x1 + · · · + xn )m .
(5)
and
y⊤ Ay = v0 y1 + · · · + y(nk−k+1)
2
,
(6)
where A is the associated Hankel matrix, which implies that A is a strong Hankel tensor and
hence an SOS Hankel tensor.
The proof is trivial and we omit it here.
2.2
The Case that r = 3
We have the following theorem for r = 3.
Theorem 1 Suppose that A is an mth order n dimensional generalized anti-circulant tensor
with circulant index 3, where m = 2k ≥ 2 and n ≥ 3. Then A is PSD if and only if
v0 = v1 = v2 ≥ 0. In this case, we still have (5) and (6), which implies that A is a strong
Hankel tensor and hence an SOS Hankel tensor.
The “if” part of this theorem follows from Proposition 1. We will prove the “only if”
part of this theorem in the three cases (Propositions 3-5).
2.2.1
The Case that m = 6l and n ≥ r = 3
We see that the results in [10] for m = 6 and n = 3 can be extended to this case.
4
In this case, (1) and (2) have the following form:
vi = vi+3
(7)
for i = 0, · · · , (n − 1)m − 3, and for x = (x1 , · · · , xn )⊤ ∈ ℜn ,
⊗m
f (x) ≡ Ax
n
X
vi1 +···+im xi1 · · · xim = v0 f0 (x) + v1 f1 (x) + v2 f2 (x),
(8)
{xi1 · · · xim : i1 + · · · + im = j, mod(3), i1 , · · · , im = 1, · · · , n} ,
(9)
=
i1 ,···,im =1
where
fj (x) =
X
for j = 0, 1, 2. We may see that
f0 (x) + f1 (x) + f2 (x) = (x1 + · · · + xn )m .
(10)
Since we are concerned about PSD generalized anti-circulant tensors, we may assume
that (3) holds, i.e., v0 ≥ 0.
Proposition 2 Suppose that A is an mth order n dimensional generalized anti-circulant
tensor with circulant index 3, where m is a multiple of 6 and n ≥ 3. Assume that v0 ≥ 0. If
A is PSD, then
v1 + v2 = 2v0 .
(11)
Proof Suppose that A is PSD and v0 ≥ 0. Then we have f (1, −1, 0, · · · , 0) ≥ 0. Note that
X
f0 (1, −1, 0, · · · , 0) =
{xi1 · · · xim : i1 + · · · + im = 0, mod(3), i1 , · · · , im = 1, 2, x1 = 1, x2 = −1}
X
=
{xi1 · · · xim : the number of ij = 1 is m − p, the number of ij = 2 is p,
p = 0, 3, · · · , m, x1 = 1, x2 = −1}
X
m
p
: p = 0, 3, · · · , m .
(−1)
=
p
Similarly, we can prove that
f1 (1, −1, 0, · · · , 0) =
and
X
(−1)
p
m
p
: p = 1, 4, · · · , m − 2
X
m
p
: p = 2, 5, · · · , m − 1 .
(−1)
f2 (1, −1, 0, · · · , 0) =
p
We may see that
(−1)l f0 (1, −1, 0, · · · , 0) > 0.
5
(12)
Since
m
p
≡
m
m−p
, we have
f1 (1, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0).
(13)
f0 (1, −1, 0, · · · , 0) + f1 (1, −1, 0, · · · , 0) + f2 (1, −1, 0, · · · , 0) = 0.
(14)
By (10),
By (8) and (12-14), we have (−1)l (v1 + v2 − 2v0 ) ≤ 0.
On the other hand,
f0 (1, 1, −2, 0, · · · , 0)
X
=
{xi1 · · · xim : i1 + · · · + im = 0, mod(3), i1 , · · · , im = 1, 2, 3, x1 = 1, x2 = 1, x3 = −2}
X
=
{xi1 · · · xim : the number of ij = 1 is m − p − q, the number of ij = 2 is q,
the number of ij = 3 is p, 2p + q = 0, mod(3), 0 ≤ p, q ≤ m, x1 = 1, x2 = 1, x3 = −2}
m X
X
{xi1 · · · xim : the number of ij = 1 is m − p − q, the number of ij = 2 is q,
=
p=0
the number of ij = 3 is p, 2p + q = 0, mod(3), 0 ≤ q ≤ m, x1 = 1, x2 = 1, x3 = −2}
m X
X
m−p
m
p
: 2p + q = 0, mod(3), 0 ≤ q ≤ m .
(−2)
=
q
p
p=0
Similarly, we can prove that
f1 (1, 1, −2, 0, · · · , 0) =
m X
X
(−2)
p=0
p
m
p
m−p
q
: 2p + q = 1, mod(3), 0 ≤ q ≤ m
and
m X
X
m−p
m
p
: 2p + q = 2, mod(3), 0 ≤ q ≤ m .
(−2)
f2 (1, 1, −2, 0, · · · , 0) =
q
p
p=0
We may see that
(−1)l+1 f0 (1, 1, −2, 0, · · · , 0) > 0.
Since
m−p
q
≡
m−p
m−p−q
(15)
, we have
f1 (1, 1, −2, 0, · · · , 0) = f2 (1, 1, −2, 0, · · · , 0).
(16)
f0 (1, 1, −2, 0, · · · , 0) + f1 (1, 1, −2, 0, · · · , 0) + f2 (1, 1, −2, 0, · · · , 0) = 0.
(17)
By (10),
by f (1, 1, −2, 0, · · · , 0) ≥ 0, (8) and (15-17) we can derive (−1)l (v1 + v2 − 2v0 ) ≥ 0. This
proves (11).
✷
We now have the following theorem.
6
Proposition 3 Theorem 1 is true if m = 6l for l ≥ 1.
Proof Suppose that A is PSD. Then v0 ≥ 0. Without loss of generality, assume that v0 > 0.
By the above discussion, v1 + v2 = 2v0 . Let v1 = v0 (1 + α). Then v2 = v0 (1 − α) and
f (x) = v0 (x1 + · · · + xn )m + v0 α(f1 (x) − f2 (x)),
(18)
where f1 and f2 are defined above. We may see that
f1 (1, 2, −3, 0, · · · , 0) − f2 (1, 2, −3, 0, · · · , 0) = f2 (1, −3, 2, 0, · · · , 0) − f1 (1, −3, 2, 0, · · · , 0) 6= 0.
Then from this, (18), f (1, 2, −3, 0, · · · , 0) ≥ 0 and f (1, −3, 2, 0, · · · , 0) ≥ 0, we have α = 0.
This proves that v0 = v1 = v2 ≥ 0.
✷
2.2.2
The Case that m = 6l + 2 and n ≥ r = 3
In this case, (1) and (3) have the form (7), and for x = (x1 , · · · , xn )⊤ ∈ ℜn ,
f (x) ≡ Ax⊗m =
n
X
vi1 +···+im +1 xi1 · · · xim = v0 f0 (x) + v1 f1 (x) + v2 f2 (x),
(19)
i1 ,···,im =1
where
f0 (1, −1, 0, · · · , 0) =
=
X
X
{xi1 · · · xim : i1 + · · · + im = 2, mod(3), i1 , · · · , im = 1, 2, x1 = 1, x2 = −1}
{xi1 · · · xim : the number of ij = 1 is m − p, the number of ij = 2 is p,
p = 0, 3, · · · , m − 2, x1 = 1, x2 = −1}
X
m
p
: p = 0, 3, · · · , m − 2 .
(−1)
=
p
Similarly, we can prove that
f1 (1, −1, 0, · · · , 0) =
X
and
f2 (1, −1, 0, · · · , 0) =
(−1)
X
We may see that (12) still holds. Since
≡
m
p
p
m
m−p
(−1)
m
p
p
m
p
: p = 1, 4, · · · , m − 1
: p = 2, 5, · · · , m .
, we have
f0 (1, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0).
(20)
By (19), (14) still holds. Similarly, we can show that
f0 (1, 0, −1, 0, · · · , 0) = f0 (1, −1, 0, · · · , 0),
7
(21)
f1 (1, 0, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0),
(22)
f2 (1, 0, −1, 0, · · · , 0) = f1 (1, −1, 0, · · · , 0),
(23)
f0 (0, 1, −1, 0, · · · , 0) = f1 (1, −1, 0, · · · , 0),
(24)
f1 (0, 1, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0)
(25)
f2 (0, 1, −1, 0, · · · , 0) = f0 (1, −1, 0, · · · , 0).
(26)
and
Since we are concerned about PSD generalized anti-circulant tensors, we may assume that
(3) holds, i.e.,
v0 ≥ 0, v8 ≥ 0, v16 ≥ 0.
(27)
From (7) and (27), we conclude that if A is PSD, then v1 = v16 ≥ 0, v2 = v8 ≥ 0, i.e.,
v0 ≥ 0, v1 ≥ 0, v2 ≥ 0,
(28)
i.e., A is a nonnegative tensor.
Proposition 4 Theorem 1 holds if m = 6l + 2 for l ≥ 0.
Proof Suppose that A is PSD. From f (1, −1, 0, · · · , 0) ≥ 0, (12), (14) and (19-20), we have
(−1)l (2v1 − v0 − v2 ) ≤ 0. From f (1, 0, −1, 0, · · · , 0) ≥ 0, (12), (14) and (19-23), we have
(−1)l (2v2 − v0 − v1 ) ≤ 0. From f (0, 1, −1, 0, · · · , 0) ≥ 0, (12), (14), (19-20) and (24-26), we
have (−1)l (2v0 − v1 − v2 ) ≤ 0. Combining them with (28), we have v0 = v1 = v2 ≥ 0.
✷
2.2.3
The Case that m = 6l + 4 and n ≥ r = 3
In this case, (1) and (3) have the form (7), and for x = (x1 , · · · , xn )⊤ ∈ ℜn ,
f (x) ≡ Ax⊗m =
n
X
vi1 +···+im +2 xi1 · · · xim = v0 f0 (x) + v1 f1 (x) + v2 f2 (x),
(29)
i1 ,···,im =1
where
f0 (1, −1, 0, · · · , 0) =
=
X
X
{xi1 · · · xim : i1 + · · · + im = 1, mod(3), i1 , · · · , im = 1, 2, x1 = 1, x2 = −1}
{xi1 · · · xim : the number of ij = 1 is m − p, the number of ij = 2 is p,
p = 0, 3, · · · , m − 1, x1 = 1, x2 = −1}
X
m
p
: p = 0, 3, · · · , m − 1 .
(−1)
=
p
We may see that
(−1)l f0 (1, −1, 0, · · · , 0) < 0.
8
(30)
Similarly, we can prove that
X
m
p
: p = 1, 4, · · · , m
(−1)
f1 (1, −1, 0, · · · , 0) =
p
and
X
m
p
: p = 2, 5, · · · , m − 2 .
(−1)
f2 (1, −1, 0, · · · , 0) =
p
m
Since mp ≡ m−p
, we have
f0 (1, −1, 0, · · · , 0) = f1 (1, −1, 0, · · · , 0).
(31)
By (29), (14) still holds. Similarly, we can show that
f0 (1, 0, −1, 0, · · · , 0) = f0 (1, −1, 0, · · · , 0),
(32)
f1 (1, 0, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0),
(33)
f2 (1, 0, −1, 0, · · · , 0) = f1 (1, −1, 0, · · · , 0),
(34)
f0 (0, 1, −1, 0, · · · , 0) = f2 (1, −1, 0, · · · , 0),
(35)
f1 (0, 1, −1, 0, · · · , 0) = f1 (1, −1, 0, · · · , 0)
(36)
f2 (0, 1, −1, 0, · · · , 0) = f0 (1, −1, 0, · · · , 0).
(37)
and
Proposition 5 Theorem 1 holds if m = 6l + 4 for l ≥ 0.
Proof Suppose that A is PSD. From f (1, −1, 0, · · · , 0) ≥ 0, (12), (14), (29) and (31), we
have (−1)l (v0 + v1 − 2v2 ) ≤ 0. From f (1, 0, −1, 0, · · · , 0) ≥ 0, (12), (14) and (30-34), we have
(−1)l (v0 + v2 − 2v1 ) ≤ 0. From f (0, 1, −1, 0, · · · , 0) ≥ 0, (12), (14), (29), (31) and (35-37),
we have (−1)l (v1 + v2 − 2v0 ) ≤ 0. Combining them with (3), we have v0 = v1 = v2 ≥ 0. ✷
Combining Propositions 1, 3-5, Theorem 1 is proved.
2.3
The Case that m = 2 and 1 ≤ r = 2p + 1 ≤ n: The Matrix Case
We now consider the matrix case with r to be odd. We have the following theorem.
Theorem 2 Suppose that A is an n dimensional generalized anti-circulant matrix with 1 ≤
r = 2p + 1 ≤ n. Then A is PSD if and only if v0 = · · · = vr ≥ 0. In this case, we have
f (x) ≡ x⊤ Ax = v0 (x1 + · · · + xn )2 .
9
(38)
Proof Let the ith unit vector in ℜn be ei . Suppose that A is PSD. Then v0 ≥ 0. Let
x = ej − ej+1 for j = 1, · · · , n − 1. From f (x) ≥ 0, we have
v2j−2 − 2v2j−1 + v2j ≥ 0,
(39)
for j = 1, · · · , n − 1.
When r < n, by (39) and (1), we have
vj−1 − 2vj + vj+1 ≥ 0,
(40)
for j = 1, · · · , r.
When r = n = 2p + 1, let x = en − e1 . From f (x) ≥ 0, we have
v2p−1 − 2v2p + v0 ≥ 0.
(41)
By (39), (41) and (1), we still have (40) for j = 1, · · · , r.
From v0 ≥ 0 and (40) for j = 1, · · · , r, we have v0 = · · · = vr ≥ 0.
The remaining conclusions follow from Proposition 1.
2.4
✷
The Case that m = 4 and n ≥ r = 5
In this case, (1) and (3) have the form
vi = vi+5 ,
(42)
v4j ≥ 0
(43)
for i = 0, · · · , 4n − 9, and
for j = 0, · · · , n − 1. From (42) and (43), we conclude that if A is PSD, then v1 = v16 ≥ 0,
v2 = v12 ≥ 0, v3 = v8 ≥ 0, i.e.,
v0 ≥ 0, v1 ≥ 0, v2 ≥ 0, v3 ≥ 0, v4 ≥ 0
(44)
i.e., A is a nonnegative tensor.
In this case, by (2), we have
f (x1 , · · · , x5 , 0, · · · , 0)
= v0 x41 + 4(x2 x34 + x32 x3 + x33 x5 + x4 x35 ) + 6(x22 x25 + x23 x24 )
+12(x1 x2 x23 + x1 x22 x4 + x1 x3 x25 + x1 x24 x5 + x21 x2 x5 + x21 x3 x4 ) + 24x2 x3 x4 x5
+v1 x45 + 4(x1 x33 + x31 x2 + x32 x4 + x3 x34 ) + 6(x21 x24 + x22 x23 )
+12(x1 x22 x5 + x1 x4 x25 + x21 x3 x5 + x2 x3 x25 + x2 x24 x5 + x23 x4 x5 ) + 24x1 x2 x3 x4
10
+v2 x44 + 4(x31 x3 + x1 x35 + x2 x33 + x32 x5 ) + 6(x21 x22 + x23 x25 )
+12(x1 x2 x24 + x1 x23 x4 + x21 x4 x5 + x2 x4 x25 + x22 x3 x4 + x3 x24 x5 ) + 24x1 x2 x3 x5
+v3 x43 + 4(x1 x32 + x31 x4 + x2 x35 + x34 x5 ) + 6(x21 x25 + x22 x24 )
+12(x1 x22 x3 + x1 x2 x25 + x21 x2 x4 + x2 x3 x24 + x2 x23 x5 + x22 x4 x5 ) + 24x1 x3 x4 x5
+12(x1 x3 x24 + x1 x23 x5 + x21 x2 x3 + x2 x23 x4 + x22 x3 x5 + x3 x4 x25 ) + 24x1 x2 x4 x5
+v4 x42 + 4(x1 x34 + x31 x5 + x3 x35 + x33 x4 ) + 6(x21 x23 + x24 x25 )
(45)
Theorem 3 Suppose that A is a fourth order n dimensional generalized anti-circulant tensor defined above. Then A is PSD if and only if v0 = v1 = v2 = v3 = v4 ≥ 0. In this case,
we still have (5) and (6), which implies that A is a strong Hankel tensor and hence an SOS
Hankel tensor.
Proof Suppose that A is PSD. From f (1, −1, 0, −1, 1, 0, · · · , 0) ≥ 0, f (0, 1, −1, −1, 1, 0, · · · , 0) ≥
0 and f (1, −1, −1, 1, 0, 0, · · · , 0) ≥ 0, we have that
4v3 ≥ v0 + v1 + v2 + v4 ,
4v0 ≥ v1 + v2 + v3 + v4 ,
(46)
4v1 ≥ v0 + v2 + v3 + v4 .
Since f (1, −1, 0, · · · , 0) ≥ 0, f (1, 0, −1, 0, · · · , 0) ≥ 0, f (0, 1, −1, 0, · · · , 0) ≥ 0, f (0, 0, 1, −1, 0, · · · , 0) ≥
0, f (0, 1, 0, 0, −1, 0, · · · , 0) ≥ 0, f (0, 0, 1, 0, −1, 0, · · · , 0) ≥ 0 and f (0, 0, 0, 1, −1, 0, · · · , 0) ≥
0, we have that
v0 − 4v1 + 6v2 − 4v3 + v4 ≥ 0,
v0 − 4v1 − 4v2 + v3 + 6v4 ≥ 0,
−4v0 + 6v1 − 4v2 + v3 + v4 ≥ 0,
6v0 − 4v1 + v2 + v3 − 4v4 ≥ 0,
(47)
6v0 + v1 − 4v2 − 4v3 + v4 ≥ 0,
−4v0 + v1 + 6v2 + v3 − 4v4 ≥ 0,
−4v0 + v1 + v2 − 4v3 + 6v4 ≥ 0.
Combining (46) with (47), we have
v0 + v2 = v1 + v3 ,
v0 + v3 = v1 + v4 ,
v0 + v4 = v1 + v2 ,
v0 + v3 ≥ v1 + v4 ,
v2 + v3 ≥ v0 + v4 .
11
(48)
From them and (44), we have v0 = v1 = v2 = v3 = v4 ≥ 0.
The “if” part follows from Proposition 1.
✷
The Case that r is Even
3
3.1
The Case that r = 2
We see that the results in [10] for m = 6 and n = 3 can be extended to this case.
In this case, (1) and (2) have the following form:
vi = vi+2
(49)
for i = 0, · · · , (n − 1)m − 2, and for x = (x1 , · · · , xn )⊤ ∈ ℜn ,
⊗m
f (x) ≡ Ax
n
X
=
vi1 +···+im −m xi1 · · · xim = v0 f0 (x) + v1 f1 (x),
(50)
i1 ,···,im =1
where
fj (x) =
X
{xi1 · · · xim : i1 + · · · + im = j, mod(2), i1 , · · · , im = 1, · · · , n} ,
(51)
for j = 0, 1. We may see that
f0 (x) + f1 (x) = (x1 + · · · + xn )m .
(52)
Since we are concerned about PSD generalized anti-circulant tensors, we may assume
that (3) holds, i.e., v0 ≥ 0.
Theorem 4 Suppose that A is an mth order n dimensional generalized anti-circulant tensor
with circulant index r = 2, where m = 2k ≥ 2 and n ≥ 2. Then A is PSD if and only if
|v1 | ≤ v0 . In these cases, we may write v1 = v0 (1 − 2t), where t ∈ [0, 1]. We have that
m
f (x) = tv0 (x1 + · · · xn )m + (1 − t)v0 x1 − x2 + x3 − · · · + (−1)n−1 xn ,
and A is a strong Hankel tensor.
Proof Suppose that A is PSD and v0 ≥ 0. Then we have f (1, 1, 0, · · · , 0) ≥ 0. Note that
X
f0 (1, 1, 0, · · · , 0) =
{xi1 · · · xim : i1 + · · · + im = 0, mod(2), i1 , · · · , im = 1, 2, x1 = 1, x2 = 1}
X
=
{xi1 · · · xim : the number of ij = 1 is m − p, the number of ij = 2 is p,
p = 0, 2, · · · , m, x1 = 1, x2 = 1}
X m : p = 0, 2, · · · , m .
=
p
12
Similarly, we can prove that
f1 (1, 1, 0, · · · , 0) =
X m f0 (1, −1, 0, · · · , 0) =
and
p
X
: p = 1, 3, · · · , m − 1 ,
m
p
: p = 0, 2, · · · , m
X
m
p
: p = 1, 3, · · · , m − 1 .
(−1)
f1 (1, −1, 0, · · · , 0) =
p
We may see that
f0 (1, −1, 0, · · · , 0) + f1 (1, −1, 0, · · · , 0) = 0,
f0 (1, 1, 0, · · · , 0) = f0 (1, −1, 0, · · · , 0) > 0,
f1 (1, 1, 0, · · · , 0) = −f1 (1, −1, 0, · · · , 0) > 0.
(53)
By (50), we have v0 +v1 ≥ 0. From f (1, −1, 0, · · · , 0) ≥ 0 and (50), we have v0 −v1 ≥ 0. This
implies that v0 ≥ |v1 |. On the other hand, suppose that v0 ≥ |v1 |. We may write v1 = v0 (2t−
1), where t ∈ [0, 1]. Write f (x) = v0 f0 (x)+v1 f1 (x) such that f0 (x)+f1 (x) = (x1 +· · ·+xn )m
and f0 (x) − f1 (x) = (x1 − x2 + x3 − · · · + (−1)n−1 xn )m for all x = (x1 , · · · , xn )⊤ ∈ ℜn . It
then follows from (50) that
f (x) = v0 f0 (x) + v1 f1 (x) = v0 f0 (x) + (2t − 1)v0 f1 (x)
= tv0 (f0 (x) + f1 (x)) + (1 − t)v0 (f0 (x) − f1 (x))
= tv0 (x1 + · · · + xn )m + (1 − t)v0 (x1 − x2 + x3 − · · · + (−1)n−1 xn )m .
Similarly, we have
g(y) = tv0 (y1 + · · · + ynk−k+1)2 + (1 − t)v0 (y1 − y2 + y3 − · · · + (−1)nk−k ynk−k+1)2 .
The conclusions now follow from the definitions of PSD, SOS and strong Hankel tensors.
✷
3.2
The Case that r = 4
We have the following theorem for r = 4.
Theorem 5 Suppose that A is an mth order n dimensional generalized anti-circulant tensor
with circulant index r = 4, where m = 2k ≥ 2 and n ≥ 4. Then A is PSD if and only if
v0 = v2 , v1 = v3 and |v1 | ≤ v0 . In these cases, we may write v1 = v0 (1 − 2t), where t ∈ [0, 1].
We have that
m
f (x) = tv0 (x1 + · · · xn )m + (1 − t)v0 x1 − x2 + x3 − · · · + (−1)n−1 xn ,
and A is a strong Hankel tensor.
13
The “if” part of this theorem follows from Theorem 4. We will prove the “only” if part
of this theorem in two cases (Propositions 6-7).
3.2.1
The Case that m = 4l, n ≥ r = 4
In this case, (1) and (2) have the following form:
vi = vi+4
(54)
for i = 0, · · · , 4((n − 1)l − 1). For x = (x1 , · · · , xn )⊤ ∈ ℜn ,
⊗m
f (x) ≡ Ax
=
n
X
vi1 +···+im xi1 · · · xim = v0 f0 (x) + v1 f1 (x) + v2 f2 (x) + v3 f3 (x), (55)
i1 ,···,im =1
where
fj (x) =
X
{xi1 · · · xim : i1 + · · · + im = j, mod(4), i1 , · · · , im = 1, · · · , n} ,
(56)
for j = 0, 1, 2, 3.
We have the following proposition.
Proposition 6 Theorem 5 is true if m = 4l for l ≥ 1.
Proof First suppose that A is PSD. From (56), we see that
f1 (1, 0, −1, 0, · · · , 0) = f3 (1, 0, −1, 0, · · · , 0) = 0.
From (55) and (56), we have
f0 (1, 0, −1, 0, · · · , 0) = −f2 (1, 0, −1, 0, · · · , 0) > 0.
Then by f (1, 0, −1, 0, · · · , 0) ≥ 0, we have v0 ≥ v2 .
From (56), we may also show that
f1 (1, −1, −1, 1, 0, · · · , 0) = f3 (1, −1, −1, 1, 0, 0, · · · , 0) = 0.
From (55) and (56), we have
f0 (1, −1, −1, 1, 0, · · · , 0) = −f2 (1, −1, −1, 1, 0, · · · , 0) < 0.
Then by f (1, −1, −1, 1, 0, · · · , 0) ≥ 0, we have v0 ≤ v2 . Thus, we derive that v0 = v2 .
From f (α, 1, −α, 0, · · · , 0) ≥ 0, f (α, −1, −α, 0, · · · , 0) ≥ 0 and (56), we derive that v0 ≥
φ(α)|v3 − v1 |, where φ(α) → ∞ if α → ∞. Letting α tend to ∞, we have v1 = v3 . The
remaining conclusions now follow from Theorem 4.
✷
14
3.2.2
The Case that m = 4l + 2 and n ≥ r = 4
In this case, (54) holds. For x = (x1 , · · · , xn )⊤ ∈ ℜn ,
⊗m
f (x) ≡ Ax
=
n
X
vi1 +···+im +2 xi1 · · · xim = v0 f0 (x)+v1 f1 (x)+v2 f2 (x)+v3 f3 (x), (57)
i1 ,···,im =1
where
fj (x) =
X
{xi1 · · · xim : i1 + · · · + im = j + 2, mod(4), i1 , · · · , im = 1, · · · , n} ,
(58)
for j = 0, 1, 2, 3.
We have the following proposition.
Proposition 7 Theorem 5 is true if m = 4l + 2 for l ≥ 1.
Proof First suppose that A is PSD. From (58), we see that
f1 (1, 0, −1, 0, · · · , 0) = f3 (1, 0, −1, 0, · · · , 0) = 0.
From (57) and (58), we have
f0 (1, 0, −1, 0, · · · , 0) = −f2 (1, 0, −1, 0, · · · , 0) > 0.
Then by f (1, 0, −1, 0, · · · , 0) ≥ 0, we have v0 ≥ v2 .
From (58), we may also show that
f1 (0, −1, 0, 1, 0, · · · , 0) = f3 (0, −1, 0, 1, 0, 0, · · · , 0) = 0.
From (55) and (56), we have
f0 (0, −1, 0, 1, 0, · · · , 0) = −f2 (0, −1, 0, 1, 0, · · · , 0) < 0.
Then by f (0, −1, 0, 1, 0, · · · , 0) ≥ 0, we have v0 ≤ v2 . Thus, we derive that v0 = v2 .
From f (α, 1, −α, 0, · · · , 0) ≥ 0, f (α, −1, −α, 0, · · · , 0) ≥ 0 and (58), we derive that v0 ≥
ψ(α)|v3 − v1 |, where ψ(α) → ∞ if α → ∞. Letting α tend to ∞, we have v1 = v3 . The
remaining conclusions now follow from Theorem 4.
✷
Combining Theorem 4, Propositions 6-7, Theorem 5 is proved.
15
3.3
The Case that m = 2 and 2 ≤ r = 2p ≤ 2n − 4: The Matrix Case
We now consider the matrix case with r to be even. Here we allow r = 2p ≤ 2n − 4
instead of r ≤ n, and still call such a matrix a generalized anti-circulant matrix. We have
the following theorem.
Theorem 6 Suppose that A is an n dimensional generalized anti-circulant matrix with 2 ≤
r = 2p ≤ 2n − 4. Then A is PSD if and only if v0 = v2 = · · · = vr , v1 = v3 = · · · = vr−1
and v0 ≥ |v1 |. In this case, we may write v1 = v0 (1 − 2t), where t ∈ [0, 1]. Then we have
f (x) ≡ x⊤ Ax = tv0 (x1 + · · · xn )2 + (1 − t)v0 x1 − x2 + x3 − · · · + (−1)n−1 xn
2
.
Proof Denote the ith unit vector in ℜn by ei . Suppose that A is PSD. Then v0 ≥ 0. Let
x = ej − ej+2 for j = 1, · · · p. From f (x) ≥ 0, we have
v2j−2 − 2v2j + v2j+2 ≥ 0
for j = 1, · · · , p. From these, we have v0 = v2 = · · · = vr .
Let x = αej−1 + ej − αej+1 for j = 2, · · · , p. From f (x) ≥ 0, we have
α2 v2j−4 + 2αv2j−3 + v2j−2 (1 − 2α2) − 2αv2j−1 + α2 v2j ≥ 0
for j = 2, · · · , p. Since v2j−4 = v2j−2 = v2j = v0 , we have
2α(v2j−3 − v2j−1 ) + v0 ≥ 0
for j = 2, · · · , p. Similarly, letting x = αej−1 − ej − αej+1 for j = 2, · · · , p, we have
−2α(v2j−3 − v2j−1 ) + v0 ≥ 0
for j = 2, · · · , p. Letting α → ∞, we have v2j−3 = v2j−1 for j = 2, · · · , p, i.e., v1 = v3 =
· · · = vr−1 . The remaining conclusions now follow from Theorem 4.
✷
3.4
The Case that m = 4 and n ≥ r = 6
In this case, (1) and (3) have the form
vi = vi+6 ,
(59)
vjm ≥ 0
(60)
for i = 0, · · · , (n − 1)m − 6, and
16
for j = 0, · · · , n − 1. From (59) and (60), we conclude that if A is PSD, then v2 = v8 ≥ 0
and v4 = v16 ≥ 0, i.e.,
v0 ≥ 0, v2 ≥ 0, v4 ≥ 0.
(61)
In this case, (2) has the form
f (x1 , x2 , x3 , x4 , x5 , x6 , 0 · · · , 0)
= v0 x41 + x44 + 4(x1 x33 + x1 x35 + x32 x4 + x4 x36 ) + 6(x21 x24 + x22 x23 + x22 x26 + x23 x25 + x25 x26 )
+12(x1 x22 x5 + x1 x3 x26 + x21 x2 x6 + x21 x3 x5 + x2 x4 x25 + x2 x24 x6 + x3 x24 x5 + x23 x4 x6 )
+24(x1 x2 x3 x4 + x1 x4 x5 x6 + x2 x3 x5 x6 )]
+v1 4(x31 x2 + x2 x33 + x2 x35 + x32 x5 + x34 x5 + x5 x36 ) + 12(x1 x2 x24 + x21 x3 x6 + x1 x22 x6
+x1 x23 x4 + x1 x4 x26 + x1 x25 x6 + x21 x4 x5 + x2 x3 x26 + x22 x3 x4 + x3 x4 x25 + x3 x24 x6 + x23 x5 x6 )
+24(x1 x2 x3 x5 + x2 x4 x5 x6 )]
+v2 x43 + x46 + 4(x31 x3 + x32 x6 + x3 x35 + x34 x6 ) + 6(x21 x22 + x21 x25 + x22 x24 + x23 x26 + x24 x25 )
+12(x1 x3 x24 + x1 x23 x5 + x1 x5 x26 + x21 x4 x6 + x2 x23 x4 + x2 x4 x26 + x2 x25 x6 + x22 x3 x5 )
+24(x1 x2 x3 x6 + x1 x2 x4 x5 + x3 x4 x5 x6 )]
+v3 4(x1 x32 + x1 x34 + x1 x36 + x31 x4 + x33 x4 + x4 x35 ) + 12(x1 x2 x25 + x1 x23 x6 + x21 x2 x3
+x21 x5 x6 + x2 x3 x24 + x2 x23 x5 + x2 x5 x26 + x22 x3 x6 + x22 x4 x5 + x3 x4 x26 + x3 x25 x6 + x24 x5 x6 )
+24(x1 x2 x4 x6 + x1 x3 x4 x5 )]
+v4 x42 + x45 + 4(x31 x5 + x2 x34 + x2 x36 + x33 x5 ) + 6(x21 x23 + x21 x26 + x22 x25 + x23 x24 + x24 x26 )
+12(x1 x22 x3 + x1 x3 x25 + x1 x24 x5 + x21 x2 x4 + x2 x23 x6 + x22 x4 x6 + x3 x5 x26 + x4 x25 x6 )
+24(x1 x2 x5 x6 + x1 x3 x4 x6 + x2 x3 x4 x5 )]
+v5 4(x31 x6 + x32 x3 + x3 x34 + x3 x36 + x33 x6 + x35 x6 ) + 12(x1 x2 x23 + x1 x2 x26 + x1 x22 x4
+x1 x4 x25 + x1 x24 x6 + x21 x2 x5 + x21 x3 x4 + x2 x3 x25 + x2 x24 x5 + x22 x5 x6 + x23 x4 x5 + x4 x5 x26 )
+24(x1 x3 x5 x6 + x2 x3 x4 x6 )] .
(62)
We have the following theorem.
Theorem 7 Suppose that A is a fourth order n dimensional generalized anti-circulant tensor defined above. Then A is PSD if and only if v0 = v2 = v4 , v1 = v3 = v5 and v0 ≥ |v1 |.
In this case, we may write v1 = v0 (1 − 2t), where t ∈ [0, 1]. Then we have
f (x) = tv0 (x1 + · · · xn )4 + (1 − t)v0 x1 − x2 + x3 − · · · + (−1)n−1 xn
4
.
This implies that A is PSD if only if it is SOS. Furthermore, in this case, A is a strong
Hankel tensor.
17
Proof First suppose that A is PSD. Then f (1, 0, −1, 0, · · · , 0) ≥ 0, f (1, 0, 0, 0, −1, 0, · · · , 0) ≥
0 and f (0, 1, 0, 0, 0, −1, 0, · · · , 0) ≥ 0. From (62), we derive that v0 = v2 = v4 . From
f (α, 1, −α, 0, · · · , 0) ≥ 0, f (α, −1, −α, 0, · · · , 0) ≥ 0 and (62), we derive that v0 ≥ η(α)|v3 −
v5 |, where η(α) → ∞ if α → ∞. Letting α tend to ∞, we have v3 = v5 . From f (0, 1, −1, −1, 1, 0, · · · , 0)
≥ 0 and f (0, 0, 1, −1, −1, 1, 0, · · · , 0) ≥ 0, we derive that v1 = v3 . Thus, we have v1 = v3 =
v5 . From f (1, 0, 0, 1, 0, · · · , 0) ≥ 0, f (1, 0, 0, −1, 0, · · · , 0) ≥ 0 and (62), we derive that
v0 ≥ |v1 |. The remaining conclusions now follow from Theorem 4.
✷
4
Final Remarks
We put the matrix case in Subsections 2.3 and 3.3, as the proofs of Theorems 2 and 6
need to use Proposition 1 (r = 1) and Theorem 4 (r = 2), respectively. This shows that the
matrix case (m = 2) and the higher order tensor case (m ≥ 3) are connected to each other.
We believe that the results are still true for any m = 2k ≥ 4, n ≥ r ≥ 5, but some new
proof techniques may be needed.
References
[1] R. Badeau and R. Boyer, “Fast multilinear singular value decomposition for
structured tensors”, SIAM Journal on Matrix Analysis and Applications 30
(2008) 1008-1021.
[2] H. Chen and L. Qi, “Positive definiteness and semi-definiteness of even order symmetric Cauchy tensors”, to appear in: J. Indust. Management Optim.,
arXiv:1405.6363.
[3] P. J. Davis, Circulant Matrices, John Wiley & Sons, New York, 1979.
[4] W. Ding, L. Qi and Y. Wei, “Fast Hankel tensor-vector products and application
to exponential data fitting”, January 2014. arXiv:1401.6238
¨
[5] D. Hilbert, “Uber
die Darstellung definiter Formen als Summe von Formenquadraten”, Mathematical Annals, 32 (1888) 342-350.
[6] S. Hu, G. Li, L. Qi and Y. Song, “Finding the maximum eigenvalue of essentially
nonnegative symmetric tensors via sum of squares programming”, Journal of
Optimization Theory and Applications, 158 (2013) 717-738.
18
[7] S. Hu, G. Li and L. Qi, “A tensor analogy of Yuan’s alternative theorem and
polynomial optimization with sign structure”, Journal of Optimization Theory
and Applications, (2014), DOI 10.1007/s10957-014-0652-1.
[8] J.B. Lasserre, “Global optimization with polynomials and the problem of moments”, SIAM Journal on Optimization 11 (2001) 796-817.
[9] M. Laurent, “Sums of squares, moment matrices and optimization over polynomials”, Emerging Applications of Algebraic Geometry, Vol. 149 of IMA Volumes
in Mathematics and its Applications, M. Putinar and S. Sullivant eds., Springer,
(2009) pp. 157-270.
[10] G. Li, L. Qi and Q. Wang, “Are there sixth order three dimensional Hankel
tensors?”, arXiv:1411.2368, November 2014.
[11] G. Li, L. Qi and Y. Xu, “SOS-Hankel Tensors: Theory and Application”,
arXiv:1410.6989, October 2014.
[12] C. Li and Y. Li, “Double B-tensor and quasi-double B-tensor”, Linear Algebra
and Its Applications 466 (2015) 343-356.
[13] C. Li, F. Wang, J. Zhao, Y. Zhu and Y. Li, “Criterions for the positive definiteness of real supersymmetric tensors”, Journal of Computational and Applied
Mathematics 255 (2014) 1-14.
[14] Z. Luo, L. Qi and Y. Ye, “Linear operators and positive semidefiniteness of symmetric tensor spaces”, Science China Mathematics (2015), doi: 10.1007/s11425000-0000-0.
[15] J.M. Papy, L. De Lauauwer and S. Van Huffel, “Exponential data fitting using multilinear algebra: The single-channel and multi-channel case”, Numerical
Linear Algebra with Applications 12 (2005) 809-826.
[16] L. Qi, “Eigenvalues of a real supersymmetric tensor”, Journal of Symbolic Computation 40 (2005) 1302-1324.
[17] L. Qi, “Hankel tensors: Associated Hankel matrices and Vandermonde decomposition”, Communications in Mathematical Sciences 13 (2015) 113-125.
[18] L. Qi and Y. Song, “An even order symmetric B tensor is positive definite”,
Linear Algebra and Its Applications 457 (2014) 303-312.
[19] B. Reznick, “Some concrete aspects of Hilbert’s 17th problem”, Contemporary
Mathematics 253 (2000) 251-272.
19
[20] Y. Song and L. Qi, “Infinite and finite dimensional Hilbert tensors”, Linear
Algebra and Its Applications 451 (2014) 1-14.
[21] L. Zhang, L. Qi and G. Zhou, “M-tensors and some applications”, SIAM Journal
on Matrix Analysis and Applications 35 (2014) 437-452.
20