Open Access

Some inequalities for nonnegative tensors

Journal of Inequalities and Applications20142014:340

https://doi.org/10.1186/1029-242X-2014-340

Received: 26 March 2014

Accepted: 14 August 2014

Published: 3 September 2014

Abstract

Let A be a nonnegative tensor and x = ( x i ) > 0 its Perron vector. We give lower bounds for x t m 1 / x i 2 x i m and upper bounds for x s m 1 / x i 2 x i m , where x s = max 1 i n x i and x t = min 1 i n x i .

MSC:15A18, 15A69, 65F15, 65F10.

Keywords

Perron vector nonnegative tensor spectral radius eigenvalues

1 Introduction

Eigenvalue problems of higher order tensors have become an important topic of study in a new applied mathematics branch, numerical multilinear algebra, and they have a wide range of practical applications [17]. The main difficulty in tensor problems is that they are generally nonlinear. Therefore, large amounts of results for matrices are never in force for higher order tensors. However, there are still some results preserved in the case of higher order tensors.

Throughout this paper we consider an m th order n-dimensional tensor A consisting of n m entries in :
A = ( a i 1 , i 2 , , i m ) , a i 1 , i 2 , , i m R , 1 i 1 , i 2 , , i m n .

The tensor A is called nonnegative (or positive) if all the entries a i 1 , i 2 , , i m 0 (or a i 1 , i 2 , , i m > 0 ). We also denote by the field of complex numbers.

Definition 1.1 A pair ( λ , x ) C × C n / { 0 } is called an eigenvalue-eigenvector pair of A , if they satisfy
A x m 1 = λ x [ m 1 ] ,
(1)
where n-dimensional column vectors A x m 1 and x [ m 1 ] are defined as
A x m 1 : = ( a i i 2 i n x i 2 x i n ) 1 i n and x [ m 1 ] : = ( x i m 1 ) 1 i n .
This definition was introduced by Qi [8] when m is even and A is symmetric. Independently, Lim [9] gave such a definition but restricted x to be a real vector and λ to be a real number. Let
ρ ( A ) = max { | λ | : λ  is an eigenvalue of  A } ,

where | λ | denotes the modulus of λ. We call ρ ( A ) the spectral radius of tensor A .

In [4], Chang et al. generalized the Perron-Frobenius theorem from nonnegative matrices to irreducible nonnegative tensors. Some further results based on this theorem are discussed by Yang [10, 11].

Definition 1.2 The tensor A is called reducible if there exists a nonempty proper index subset J { 1 , 2 , , n } such that
a i 1 , i 2 , , i m = 0 , i 1 J , i 2 , , i m J .

If A is not reducible, then we call A irreducible.

Theorem 1.3 [4]

If A is irreducible and nonnegative, then there exist a number ρ ( A ) > 0 and a vector x 0 > 0 , such that
A x 0 m 1 = ρ ( A ) x 0 [ m 1 ] .

Moreover, if λ is an eigenvalue with a nonnegative eigenvector, then λ = ρ ( A ) . If λ is an eigenvalue of A , then | λ | ρ ( A ) .

We call x 0 a Perron vector of A corresponding to its largest nonnegative eigenvalue ρ ( A ) .

In this paper, we are interested in studying some bounds for the Perron vector of A . For this purpose, we define
x s = max 1 i n x i , x t = min 1 i n x i .

In the following we first give a new and simple bound for x s / x t . Then we give some lower bounds for x t m 1 / x i 2 x i m and upper bounds for x s m 1 / x i 2 x i m , which can be used to get another bound for x s / x t .

The paper is organized as follows. In Section 2, some efforts of establishing the bounds of the nonnegative tensor are made. An application of these bounds is studied in Section 3.

2 Bounds

In [12], the authors have studied the perturbation bound for the spectral radius of A , and they show that a Perron vector x must be known in advance so that the perturbation bound can be computed. Here we cite a lemma for use below.

Lemma 2.1 [12]

Suppose A is nonnegative such that m 3 and x is its Perron vector x. Then
(2)
Define the i th row sum of A as
R i ( A ) = i 2 , , i m = 1 n a i i 2 i m ,
and denote the largest, the smallest, and the average row sums of A by
R max ( A ) = max i = 1 , , n R i ( A ) , R min ( A ) = min i = 1 , , n R i ( A ) .
Let
l = min i 1 , , i m a i 1 i m , L = max i 1 , , i m a i 1 i m .
Theorem 2.2 Suppose A is nonnegative with Perron vector x. Then
(3)
Proof Since, for any i = 1 , 2 , , n ,
ρ ( A ) x t m 1 ρ ( A ) x i m 1 = i 2 , , i m = 1 n a i i 2 i m x i 2 x i m i 2 , , i m = 1 n a i i 2 i m x s m 1 ,
we have
ρ ( A ) x t m 1 R min ( A ) x s m 1 .
(4)
Similarly,
ρ ( A ) x s m 1 R max ( A ) x t m 1 .
(5)
It follows from (4) and (5) that
x s m 1 x t m 1 x t m 1 x s m 1 R max ( A ) R min ( A )
and therefore
x s x t { R max ( A ) R min ( A ) } 1 2 ( m 1 ) .
If we assume x 1 = 1 , then, for any i = 1 , 2 , , n ,
(6)
Similarly,
and therefore

The result follows. □

Remark By the obvious inequality

we see that our new lower bound in (3) is sharper than that in (2).

If the tensor A is positive, we derive a new upper bound for x s / x t in terms of the entries of A and the spectral radius ρ ( A ) .

Theorem 2.3 If A is positive with Perron vector x, then
l ρ ( A ) R min ( A ) + l ( m 1 ) n x t m 1 i 2 , , i m = 1 n x i 2 x i m L ρ ( A ) R max ( A ) + L ( m 1 ) n ,
(7)
ρ ( A ) R max ( A ) + [ ( m 1 ) n 1 ] L ρ ( A ) R max ( A ) + L ( m 1 ) n 1 ( m 1 ) n 1 x s m 1 i 2 , , i m = 1 n x i 2 x i m ρ ( A ) R min ( A ) + l ρ ( A ) R min ( A ) + l ( m 1 ) n ,
(8)
{ ρ ( A ) R max ( A ) + [ ( m 1 ) n 1 ] L L 1 ( m 1 ) n 1 } 1 m 1 x s x t { ρ ( A ) R min ( A ) + l l } 1 m 1 .
(9)
Proof First, we prove the right side of (9). Now we consider
ρ ( A ) x t m 1 l i 2 , , i m = 1 n x i 2 x i m = i 2 , , i m = 1 n a t i 2 i m x i 2 x i m l i 2 , , i m = 1 n x i 2 x i m x t m 1 [ R min ( A ) l ( m 1 ) n ]
(10)
and
x s m 1 i 2 , , i m = 1 n x i 2 x i m = 1 i 2 , , i m = 1 n except = = = = except i 2 = = i m = s x i 2 x i m i 2 , , i m = 1 n x i 2 x i m 1 [ ( m 1 ) n 1 ] i 2 , , i m = 1 n x i 2 x i m x t m 1 ,
(11)
so we can get
x t m 1 i 2 , , i m = 1 n x i 2 x i m l ρ ( A ) R min ( A ) + l ( m 1 ) n , x s m 1 i 2 , , i m = 1 n x i 2 x i m ρ ( A ) R min ( A ) + l ρ ( A ) R min ( A ) + l ( m 1 ) n ,
then
x s x t { ρ ( A ) R min ( A ) + l l } 1 m 1 .
On the other hand,
L i 2 , , i m = 1 n x i 2 x i m ρ ( A ) x t m 1 = L i 2 , , i m = 1 n x i 2 x i m i 2 , , i m = 1 n a t i 2 i m x i 2 x i m x t m 1 [ L ( m 1 ) n R max ( A ) ]
(12)
and
x t m 1 i 2 , , i m = 1 n x i 2 x i m = 1 i 2 , , i m = 1 n except = = = = except i 2 = = i m = t x i 2 x i m i 2 , , i m = 1 n x i 2 x i m 1 [ ( m 1 ) n 1 ] i 2 , , i m = 1 n x i 2 x i m x s m 1 ,
(13)
so we can get
x t m 1 i 2 , , i m = 1 n x i 2 x i m L ρ ( A ) R max ( A ) + L ( m 1 ) n , x s m 1 i 2 , , i m = 1 n x i 2 x i m ρ ( A ) R max ( A ) + [ ( m 1 ) n 1 ] L ρ ( A ) R max ( A ) + L ( m 1 ) n 1 ( m 1 ) n 1 ,
then
x s x t { ρ ( A ) R max ( A ) + [ ( m 1 ) n 1 ] L L 1 ( m 1 ) n 1 } 1 m 1 .

This completes the proof. □

In the following corollary, we get some bounds for x t / j = 1 n x j and x s / j = 1 n x j in terms of the entries of A and the spectral radius ρ ( A ) .

Corollary 2.4 If A is positive with Perron vector x. Then
x t i k = 1 n x i k l ρ ( A ) R min ( A ) + n l ,
(14)
x s i k = 1 n x i k ρ ( A ) R min ( A ) + l ρ ( A ) R min ( A ) + n l ,
(15)
x s x t ρ ( A ) R min ( A ) + l l .
(16)
Proof From
ρ ( A ) x t m 1 = i 2 , , i m = 1 n a t i 2 i m x i 2 x i m ( i 2 , , i m = 1 n a t i 2 i m x i k ) x t m 2 ,
we have
ρ ( A ) x t l i k = 1 n x i k i 2 , , i m = 1 n a t i 2 i m x i k l i k = 1 n x i k x t [ R min ( A ) n l ] .
(17)

Similar to the proof of Theorem 2.3, we can get the results. □

Remark Similar to the proof of Theorem 2.3, we can get the lower bound of x s x t .

3 Application to the perturbation bound

Suppose A 0 and is another tensor satisfying A B . In this section we are interested in the bound of ρ ( B ) ρ ( A ) .

Lemma 3.1 [10]

0 A B , then ρ ( A ) ρ ( B ) .

We let P = { ( y 1 , y 2 , , y n ) | y i 0 } be the positive cone, and let the interior of be denoted by int ( P ) = { ( y 1 , y 2 , , y n ) | y i > 0 } .

Theorem 3.2 Let A be weakly irreducible. For a nonzero x P , we define
μ ( x ) = min x i > 0 ( A x m 1 ) i x i m 1 , ν ( x ) = max x i > 0 ( A x m 1 ) i x i m 1 .
Then
μ ( x ) ρ ( A ) ν ( x ) .
(18)
Proof Since
min x i > 0 ( A x m 1 ) i x i m 1 ( A x m 1 ) i x i m 1 max x i > 0 ( A x m 1 ) i x i m 1 , x i 0 , μ ( x ) x m 1 A x m 1 ν ( x ) x m 1 .
By the result of Lemma 5.4 in [10], we have
μ ( x ) ρ ( A ) ν ( x ) .

 □

Theorem 3.3 Let A , be irreducible and positive such that ω = min i 1 , , i m ( b i 1 , , i m a i 1 , , i m ) 0 and μ = max i 1 , , i m ( b i 1 , , i m a i 1 , , i m ) . Then
ω [ ρ ( A ) R min ( A ) + l ( m 1 ) n ] ρ ( A ) R min ( A ) + l ρ ( B ) ρ ( A ) μ [ ρ ( A ) R min ( A ) + l ( m 1 ) n ] l .
Proof Let x be the Perron vector of A . Define i as follows:
( B x m 1 ) i x i m 1 = max k ( B x m 1 ) k x k m 1 .
Then, by Theorem 3.2, we have
max k ( B x m 1 ) k x k m 1 ρ ( B ) .
From the simple equality
( B A ) x m 1 + ρ ( A ) x [ m 1 ] = B x m 1 ,
and by considering the i th coordinate,
i 2 , , i m = 1 n ( b i i 2 i m a i i 2 i m ) x i 2 x i m x i m 1 + ρ ( A ) = ( B x m 1 ) i x i m 1 ρ ( B ) .
Since μ = max i 1 , , i m ( b i 1 , , i m a i 1 , , i m ) and x t = min 1 i n x i ,
ρ ( B ) ρ ( A ) μ i 2 , , i m = 1 n x i 2 x i m x t m 1 μ [ ρ ( A ) R min ( A ) + l ( m 1 ) n ] l .
Similarly, we can get
ρ ( B ) ρ ( A ) ω i 2 , , i m = 1 n x i 2 x i m x s m 1 ω [ ρ ( A ) R min ( A ) + l ( m 1 ) n ] ρ ( A ) R min ( A ) + l .

 □

Remark If we let x be the Perron vector of , then by a similar method, we can get the following bound:
ω [ ρ ( B ) R min ( B ) + l ( m 1 ) n ] ρ ( B ) R min ( B ) + l ρ ( B ) ρ ( A ) μ [ ρ ( B ) R min ( B ) + l ( m 1 ) n ] l .

4 Conclusion

We have obtained a new and sharper bound of x s / x t , lower bounds for x t m 1 / x i 2 x i m , and upper bounds for x s m 1 / x i 2 x i m , where x = ( x i ) > 0 is the Perron vector of a positive tensor A with x s = max 1 i n x i and x t = min 1 i n x i . In addition we have given bounds of ρ ( B ) ρ ( A ) when 0 < A B by these bounds.

Declarations

Acknowledgements

This research is supported by NSFC (61170311), Chinese Universities Specialized Research Fund for the Doctoral Program (20110185110020), Sichuan Province Sci. & Tech. Research Project (12ZC1802). The first author is supported by the Fundamental Research Funds for Central Universities.

Authors’ Affiliations

(1)
School of Mathematical Sciences, University of Electronic Science and Technology of China

References

  1. Chang KC, Pearson K, Zhang T: Primitivity, the convergence of the NZQ method, and the largest eigenvalue for nonnegative tensors. SIAM J. Matrix Anal. Appl. 2011, 32: 806-819. 10.1137/100807120MathSciNetView ArticleMATHGoogle Scholar
  2. Qi L: Eigenvalues and invariants of tensor. J. Math. Anal. Appl. 2007, 325: 1363-1377. 10.1016/j.jmaa.2006.02.071MathSciNetView ArticleMATHGoogle Scholar
  3. Ng M, Qi L, Zhou G: Finding the largest eigenvalue of a non-negative tensor. SIAM J. Matrix Anal. Appl. 2009, 31: 1090-1099.MathSciNetView ArticleMATHGoogle Scholar
  4. Chang KC, Zhang T: Perron-Frobenius theorem for nonnegative tensors. Commun. Math. Sci. 2008, 6: 507-520. 10.4310/CMS.2008.v6.n2.a12MathSciNetView ArticleMATHGoogle Scholar
  5. Chang KC, Pearson K, Zhang T: On eigenvalue problems of real symmetric tensors. J. Math. Anal. Appl. 2009, 350: 416-422. 10.1016/j.jmaa.2008.09.067MathSciNetView ArticleMATHGoogle Scholar
  6. Liu Y, Zhou G, Ibrahim NF: An always convergent algorithm for the largest eigenvalue of an irreducible nonnegative tensor. J. Comput. Appl. Math. 2010, 235: 286-292. 10.1016/j.cam.2010.06.002MathSciNetView ArticleMATHGoogle Scholar
  7. Zhou G, Caccetta L, Qi L: Convergence of an algorithm for the largest singular value of a nonnegative rectangular tensor. Linear Algebra Appl. 2013, 438: 959-968. 10.1016/j.laa.2011.06.038MathSciNetView ArticleMATHGoogle Scholar
  8. Qi L: Eigenvalues of a real supersymmetric tensor. J. Symb. Comput. 2005, 40: 1302-1324. 10.1016/j.jsc.2005.05.007View ArticleMathSciNetMATHGoogle Scholar
  9. Lim LH: Singular values and eigenvalues of tensors: a variational approach. Proceedings of the IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing 2005 CAMSAP 05, 129-132. CAMSAP 05Google Scholar
  10. Yang Y, Yang Q: Further results for Perron-Frobenius theorem for nonnegative tensors. SIAM J. Matrix Anal. Appl. 2010, 31: 2517-2530. 10.1137/090778766MathSciNetView ArticleMATHGoogle Scholar
  11. Yang Y, Yang Q: Further results for Perron-Frobenius theorem for nonnegative tensors II. SIAM J. Matrix Anal. Appl. 2011, 32: 1236-1250. 10.1137/100813671MathSciNetView ArticleMATHGoogle Scholar
  12. Wen, L, Ng, M: The perturbation bound for the spectral radius of a non-negative tensor. http://www.math.hkbu.edu.hk/mng/tensor-research/tensorGoogle Scholar

Copyright

© He et al.; licensee Springer. 2014

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.