- Research
- Open access
- Published:
A new bound for the spectral radius of nonnegative tensors
Journal of Inequalities and Applications volume 2017, Article number: 88 (2017)
Abstract
By estimating the ratio of the smallest component and the largest component of a Perron vector, we provide a new bound for the spectral radius of a nonnegative tensor. And it is proved that the proposed result improves the bound in (Li and Ng in Numer. Math. 130(2):315-335, 2015).
1 Introduction
Let \(\mathbb{C}\) (\(\mathbb{R}\)) be the set of all complex (real) numbers, \(\mathbb{R}_{+}\) (\(\mathbb{R}_{++}\)) be the set of all nonnegative (positive) numbers, \(\mathbb{C}^{n}\) (\(\mathbb{R}^{n}\)) be the set of all dimension n complex (real) vectors, and \(\mathbb {R}_{+}^{n}\) (\(\mathbb{R}_{++}^{n}\)) be the set of all dimension n nonnegative (positive) vectors. An order m dimension n complex (real) tensor \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\), denoted by \(\mathcal{A} \in\mathbb{C}^{[m,n]}\) (\(\mathcal{A}\in\mathbb {R}^{[m,n]}\), respectively), consists of \(n^{m}\) entries:
A tensor \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) is called nonnegative if
As the eigenvalues of matrices have many extensive applications, the H-eigenvalues [2] for higher order tensors also have a wide range of applications such as numerical multilinear algebra and higher order Markov chains [3–5].
Definition 1
[2]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{C}^{[m,n]}\). Then \((\lambda, x)\in\mathbb{C}\times\mathbb{C}^{n}\backslash\{0\}\) is called an eigenpair of \(\mathcal{A}\) if
where \(\mathcal{A}x^{m-1}\) and \(x^{[m-1]}\) are dimension n vectors with ith entries
and
Specifically, \((\lambda, x)\) is called an H-eigenpair if \((\lambda, x)\in\mathbb{R}\times\mathbb{R}^{n}\).
Recently, the spectral radius problem for nonnegative tensors has attracted special attention of researchers [6–9]. In [6], Chang et al. generalized the famous Perron-Frobenius theorem [10] for irreducible nonnegative matrices to irreducible nonnegative tensors.
Definition 2
[6]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{C}^{[m,n]}\). \(\mathcal{A}\) is called reducible if there is a nonempty proper subset \(K\subset N\) such that
If \(\mathcal{A}\) is not reducible, then we call \(\mathcal{A}\) irreducible.
Theorem 1
[6]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be an irreducible nonnegative tensor and
be the spectral radius of \(\mathcal{A}\), where
Then \(\rho(\mathcal{A})> 0\) is an eigenvalue of \(\mathcal{A}\) with a positive eigenvector x corresponding to it.
Note that \(\rho(\mathcal{A})\) and x in Theorem 1 are called the Perron root and the Perron vector of \(\mathcal{A}\), respectively, and \((\rho(\mathcal{A}), x)\) is regarded as a Perron eigenpair.
Subsequently, Friedland et al. generalized the result in Theorem 1 to weakly irreducible [7] nonnegative tensors in [8].
Definition 3
[7]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{C}^{[m,n]}\), define a matrix \(M=(M_{ij})\in\mathbb{C}^{n\times n}\) with
\(\mathcal{A}\) is called weakly irreducible if M is an irreducible matrix.
Theorem 2
[8]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a weakly irreducible nonnegative tensor. Then \(\rho(\mathcal{A})> 0\) is an eigenvalue of \(\mathcal{A}\) with a positive eigenvector x corresponding to it.
In [9], Yang and Yang generalized the result in Theorem 1 to nonnegative tensors.
Theorem 3
[9]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a nonnegative tensor, then \(\rho(\mathcal{A})\geq0\) is an eigenvalue of \(\mathcal{A}\) with a nonnegative eigenvector x corresponding to it.
For the spectral radius of a nonnegative tensor \(\mathcal{A}\), although some algorithms of calculating its value were proposed [11–13], it is not easy to choose an appropriate iterative initial value such that these iterative methods rapidly converge to its exact value. Therefore, it is necessary to give an initial estimate for the spectral radius of a nonnegative tensor. Actually, there are already some results for the bound of the nonnegative tensors’ spectral radius, for example, Yang and Yang extended the classical spectral radius bound for nonnegative matrices to nonnegative tensors in [9] and obtained the following result.
Theorem 4
[9]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a nonnegative tensor, then
where
In [1], by estimating the ratio of the smallest component and the largest component of a Perron vector, Li and Ng gave the following bound for the spectral radius of a nonnegative tensor and proved it is better than the bound in (1).
Theorem 5
[1]
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a nonnegative tensor, then
where
Furthermore, \(r\leq\nu(\mathcal{A})\leq\omega(\mathcal{A})\leq R\).
In this paper, we continue to study this problem and present a new lower bound and a new upper bound for the spectral radius of a nonnegative tensor by giving a new ratio of the smallest component and the largest component of a Perron vector. It is proved that this bound is better than the bound in (2). Numerical examples are also given to illustrate the efficiency of the proposed results.
2 Bounds for the spectral radius of nonnegative tensors
In this section, we first give a lemma to estimate the ratio of the smallest component and the largest component of a Perron vector, and then we give a bound for the spectral radius of nonnegative tensors.
Lemma 1
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a weakly irreducible nonnegative tensor with a Perron vector x, and let \(x_{s}=\min_{i\in N}\{x_{i}\}\), \(x_{l}=\max_{i\in N}\{x_{i}\}\). Then
where
Proof
Since \(\mathcal{A}\) is a weakly irreducible nonnegative tensor, according to Theorem 2, we have \((\rho(\mathcal{A}), x)\in \mathbb{R}_{++}\times\mathbb{R}_{++}^{n}\) is a Perron eigenpair of \(\mathcal{A}\). Without loss of generality, suppose that \(r_{p}(\mathcal {A})=R\), \(r_{q}(\mathcal{A})=r\). By \(\mathcal{A}x^{m-1}=\rho(\mathcal {A}) x^{[m-1]}\), we have that for each \(i\in N\),
Taking \(i=p\) in (3), we obtain that
Multiplying \(x_{p}^{-(m-1)}\) on both sides of (4) gives
Similarly, we have that for each \(i\in N\),
Taking \(i=q\) in (6), we have that
Multiplying \(x_{q}^{-(m-1)}\) on both sides of (7) gives
Multiplying \((\frac{x_{s}}{x_{l}} )^{m-1}\) on both sides of (9) gives
Note that it is not easy to get the bound of \(\frac{x_{s}}{x_{l}}\) simply from (10); however, we can overcome this difficulty by using the fact that \(0\leq\frac{x_{s}}{x_{l}}\leq1\) for the right-hand side of (10). Hence by (10) we have that
that is, \(\frac{x_{s}}{x_{l}} \leq (\frac{r-\beta_{0}(\mathcal{A})}{R-\beta_{0}(\mathcal {A})} )^{\frac{1}{2(m-1)}}\), which together with (10) yields
The conclusion follows. □
Theorem 6
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a weakly irreducible nonnegative tensor. Then
where
and
Proof
Since \(\mathcal{A}\) is a weakly irreducible nonnegative tensor, there is a Perron vector \(x\in R_{++}^{n}\) such that \(\mathcal{A}x^{m-1}=\rho (\mathcal{A})x^{[m-1]}\). Suppose that \(x_{s}=\min_{i\in N}\{x_{i}\}\), \(x_{l}=\max_{i\in N}\{x_{i}\}\). By (3) we have that for each \(i\in N\),
Taking \(i=s\) and multiplying \(x_{s}^{-(m-1)}\) on both sides of (13), we obtain that
Combining (14) with Lemma 1 gives
Similarly, by the first inequality of (6), we have that for each \(i\in N\),
Taking \(i=l\) and multiplying \(x_{l}^{-(m-1)}\) on both sides of (16), we obtain that
Combining (17) with Lemma 1 gives
The proof is completed. □
Remark 1
It is easy to see that the bound in (12) also holds for general nonnegative tensors. In fact, if \(\mathcal {A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) is a nonnegative tensor, and \(\mathcal{F}=(f_{i_{1}i_{2}\cdots i_{m}})\in \mathbb{R}^{[m,n]}\) with \(f_{i_{1}i_{2}\cdots i_{m}}=1\) for all \(i_{r}\in N\), \(r=1,2,\ldots,m\), then \(\mathcal{A}+\varepsilon\mathcal {F}\) is a weakly irreducible tensor for any \(\varepsilon>0\). Hence by Theorem 6 we can give the bound of \(\rho(\mathcal{A}+\varepsilon \mathcal{F})\). Since the spectral radius of a nonnegative tensor is a continuous function of its entries, the bound for \(\rho(\mathcal{A})\) can be obtained when \(\varepsilon\rightarrow0\), which is exactly the bound in (12).
Remark 2
Note that the first inequality of (6) can be replaced by
then, similar to the proof of Lemma 1, we can obtain that
where
And hence, by the similar proof of Theorem 6, we can give another bound of spectral radius for a nonnegative weakly irreducible tensor \(\mathcal{A}\) as follows.
Corollary 1
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a nonnegative weakly irreducible tensor. Then
where
and
Remark 3
From the expression of \(\mathcal{L}(\mathcal{A})\), \(\mathcal {U}(\mathcal{A})\), \(\mathcal{L}_{1}(\mathcal{A})\) and \(\mathcal {U}_{1}(\mathcal{A})\), it can be easily obtained that
Although the bound in (19) is not better than the bound in (12), it needs less computations.
Next is a comparison result for the bound in (12) and the bound in (2).
Theorem 7
Let \(\mathcal{A}=(a_{i_{1}i_{2}\cdots i_{m}})\in\mathbb{R}^{[m,n]}\) be a nonnegative tensor. Then
Proof
We only prove \(\nu(\mathcal{A})\leq\mathcal{L}(\mathcal{A})\), \(\mathcal {U}(\mathcal{A})\leq\omega(\mathcal{A})\) can be similarly proved. Note that
and
Since \(r\leq R\), we have
Hence we can obtain \(\zeta(\mathcal{A})\leq\tau(\mathcal{A})\), consequently, \(\nu(\mathcal{A})\leq\mathcal{L}(\mathcal{A})\) because
and
The proof is completed. □
Remark 4
Note that \(\delta(\mathcal{A})\leq\tau(\mathcal{A})\) is also obvious. Therefore, combining the proof of Theorem 7 and Remark 3, we have that
3 Numerical examples
In this section, we use two examples to illustrate the effectiveness of our proposed results.
Example 1
Let \(\mathcal{A}=(a_{i_{1}i_{2}i_{3}i_{4}})\) be an order 4 dimension 3 tensor, where
By the bound in (1), we have
By the bound in (2), we have
By the bound in (19), we have
By the bound in (12), we have
In fact, \(\rho(\mathcal{A})\approx14.265484618202352\).
And the relative errors for the exact value and the lower and upper bound of (1) respectively are \(\frac{\rho(\mathcal {A})-r}{\rho(\mathcal{A})}=0.115052846932948\), and \(\frac{R-\rho (\mathcal{A})}{\rho(\mathcal{A})}=0.206653714240879\).
The relative errors for the exact value and the lower bound of (2) respectively are \(\frac{\rho(\mathcal{A})-\nu(\mathcal {A})}{\rho(\mathcal{A})}=0.114262489050269\) and \(\frac{\omega(\mathcal {A})-\rho(\mathcal{A})}{\rho(\mathcal{A})}=0.200044721377158\).
The relative errors for the exact value and the lower and upper bound of (19) respectively are \(\frac{\rho(\mathcal {A})-\mathcal{L}_{1}(\mathcal{A})}{\rho(\mathcal {A})}=0.114259245307614\) and \(\frac{\mathcal{U}_{1}(\mathcal{A})-\rho (\mathcal{A})}{\rho(\mathcal{A})}=0.176901265065847\).
The relative errors for the exact value and the lower and upper bound of (12) respectively are \(\frac{\rho(\mathcal {A})-\mathcal{L}(\mathcal{A})}{\rho(\mathcal{A})}=0.114256376642732\) and \(\frac{\mathcal{U}(\mathcal{A})-\rho(\mathcal{A})}{\rho(\mathcal {A})}=0.150125147551858\).
This example shows that the bound in (12) is better.
Example 2
Consider the ninth tensor generated by the MATLAB code
By the bound in (1), we have
By the bound in (2), we have
By the bound in (19), we have
By the bound in (12), we have
In fact, \(\rho(\mathcal{A})\approx3.997739793470586\mathrm{e}{+}03\).
And the relative errors for the exact value and the lower and upper bound of (1) respectively are \(\frac{\rho(\mathcal {A})-r}{\rho(\mathcal{A})}=0.013096756471712\) and \(\frac{R-\rho(\mathcal {A})}{\rho(\mathcal{A})}=0.012756697389729\).
The relative errors for the exact value and the lower bound of (2) respectively are \(\frac{\rho(\mathcal{A})-\nu(\mathcal {A})}{\rho(\mathcal{A})}=0.013096636820587\) and \(\frac{\omega(\mathcal {A})-\rho(\mathcal{A})}{\rho(\mathcal{A})}=0.012756674214456\).
The relative errors for the exact value and the lower and upper bound of (19) respectively are \(\frac{\rho(\mathcal {A})-\mathcal{L}_{1}(\mathcal{A})}{\rho(\mathcal {A})}=0.013096636820547\) and \(\frac{\mathcal{U}_{1}(\mathcal{A})-\rho (\mathcal{A})}{\rho(\mathcal{A})}=0.012702068983443\).
The relative errors for the exact value and the lower and upper bound of (12) respectively are \(\frac{\rho(\mathcal {A})-\mathcal{L}(\mathcal{A})}{\rho(\mathcal{A})}=0.013096636820545\) and \(\frac{\mathcal{U}(\mathcal{A})-\rho(\mathcal{A})}{\rho(\mathcal {A})}=0.012114596987186\).
This example shows that the bound in (12) is better.
4 Results and discussion
The main result of this paper is Theorem 6. From Remark 2 and the proof of Lemma 1, it is not difficult to see that the right expressions of last inequality (6) can also be replaced by many similar expressions according to the extent of magnifying inequality. Therefore, we can also obtain other bounds for the spectral radius of a nonnegative tensor. Furthermore, we notice that the bound in (12) is the best of them as the last inequality (6) reaches the optimum for all those possible expressions, which can be shown by two numerical examples above.
5 Conclusions
In this paper, we propose a new bound for the spectral radius of a nonnegative tensor by estimating the ratio of the smallest component and the largest component of a Perron vector. And we prove that the proposed result improves the bound in [1].
References
Li, W, Ng, MK: Some bounds for the spectral radius of nonnegative tensors. Numer. Math. 130(2), 315-335 (2015)
Qi, L: Eigenvalues of a real supersymmetric tensor. J. Symb. Comput. 40(6), 1302-1324 (2005)
Qi, L, Sun, W, Wang, Y: Numerical multilinear algebra and its applications. Front. Math. China 2(4), 501-526 (2007)
Norris, J: Markov Chains. Cambridge University Press, Cambridge (1997)
Adke, S, Deshmukh, S: Limit distribution of a high order Markov chain. J. R. Stat. Soc., Ser. B, Methodol. 50(1), 105-108 (1988)
Chang, K, Pearson, K, Zhang, T: Perron-Frobenius theorem for nonnegative tensors. Commun. Math. Sci. 6(2), 507-520 (2008)
Hu, S, Huang, Z, Qi, L: Strictly nonnegative tensors and nonnegative tensor partition. Sci. China Math. 57(1), 181-195 (2014)
Friedlandv, S, Gaubert, S, Han, L: Perron-Frobenius theorem for nonnegative multilinear forms and extensions. Linear Algebra Appl. 438(2), 738-749 (2013)
Yang, Y, Yang, Q: Further results for Perron-Frobenius theorem for nonnegative tensors. SIAM J. Matrix Anal. Appl. 31(5), 2517-2530 (2010)
Horn, R, Johnson, C: Matrix Analysis. Cambridge University Press, Cambridge (1986)
Ng, M, Qi, L Zhou, G: Finding the largest eigenvalue of a nonnegative tensor. SIAM J. Matrix Anal. Appl. 31(3), 1090-1099 (2009)
Liu, Y, Zhou, G, Ibrahim, N: An always convergent algorithm for the largest eigenvalue of an irreducible nonnegative tensor. J. Comput. Appl. Math. 235(1), 286-292 (2010)
Zhang, L, Qi, L: Linear convergence of an algorithm for computing the largest eigenvalue of a nonnegative tensor. Numer. Linear Algebra Appl. 19(5), 830-841 (2012)
Acknowledgements
This work was supported by the National Natural Science Foundation of China (11361074, 11601473), and CAS ‘Light of West China’ Program.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
All authors contributed equally to this work. All authors read and approved the final manuscript.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Li, S., Li, C. & Li, Y. A new bound for the spectral radius of nonnegative tensors. J Inequal Appl 2017, 88 (2017). https://doi.org/10.1186/s13660-017-1362-7
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13660-017-1362-7