Spectral radius

From Wikipedia, the free encyclopedia
  (Redirected from Spectral radius formula)
Jump to: navigation, search

In mathematics, the spectral radius of a square matrix or a bounded linear operator is the largest absolute value of its eigenvalues (i.e. supremum among the absolute values of the elements in its spectrum). It is sometimes denoted by ρ(·).


Let λ1, ..., λn be the (real or complex) eigenvalues of a matrix ACn×n. Then its spectral radius ρ(A) is defined as:

The condition number of can be expressed using the spectral radius as .

The spectral radius is a sort of infimum of all norms of a matrix. On the one hand, for every induced matrix norm , and on the other hand, Gelfand's formula states that ; both these results are shown below. However, the spectral radius does not necessarily satisfy for arbitrary vectors . To see why, let be arbitrary and consider the matrix . The characteristic polynomial of is , hence its eigenvalues are , and thus . However , so for being any norm on . What still allows as is that , making as .

for all

does hold when is a Hermitian matrix and is the Euclidean norm.


The spectral radius of a finite graph is defined to be the spectral radius of its adjacency matrix.

This definition extends to the case of infinite graphs with bounded degrees of vertices (i.e. there exists some real number C such that the degree of every vertex of the graph is smaller than C). In this case, for the graph G define:

Let γ be the adjacency operator of G:

The spectral radius of G is defined to be the spectral radius of the bounded linear operator γ.

Machine learning[edit]

To avoid the vanishing gradient problem and the exploding gradient problem in a recurrent neural network (RNN), it is desired to keep the spectral radius of the Jacobian of the hidden-to-hidden transition function near 1[1].

Upper bound[edit]

The following proposition shows a simple yet useful upper bound for the spectral radius of a matrix:


Proposition. Let ACn×n with spectral radius ρ(A) and a consistent matrix norm ||⋅||. Then for each integer :

Proof of proposition[edit]

Let (v, λ) be an eigenvector-eigenvalue pair for a matrix A. By the sub-multiplicative property of the matrix norm, we get:

and since v ≠ 0 we have

and therefore


The spectral radius is closely related to the behaviour of the convergence of the power sequence of a matrix; namely, the following theorem holds:

Theorem. Let ACn×n with spectral radius ρ(A). Then ρ(A) < 1 if and only if
On the other hand, if ρ(A) > 1, .

The statement holds for any choice of matrix norm on Cn×n.

Proof of theorem[edit]

Assume the limit in question is zero, we will show that ρ(A) < 1. Let (v, λ) be an eigenvector-eigenvalue pair for A. Since Akv = λkv we have:

and, since by hypothesis v ≠ 0, we must have

which implies |λ| < 1. Since this must be true for any eigenvalue λ, we can conclude ρ(A) < 1.

Now assume the radius of A is less than 1. From the Jordan normal form theorem, we know that for all ACn×n, there exist V, JCn×n with V non-singular and J block diagonal such that:



It is easy to see that

and, since J is block-diagonal,

Now, a standard result on the k-power of an Jordan block states that, for :

Thus, if then for all i . Hence for all i we have:

which implies


On the other side, if , there is at least one element in J which doesn't remain bounded as k increases, so proving the second part of the statement.

Gelfand's formula[edit]

Theorem (Gelfand's Formula; 1941). For any matrix norm ||⋅||, we have


For any ε > 0, first we construct the following two matrices:


First we apply the previous theorem to A+:

That means, by the sequence limit definition, there exists N+N such that for all k ≥ N+,


Applying the previous theorem to A implies is not bounded and there exists NN such that for all k ≥ N,


Let N = max{N+, N}, then we have:

which, by definition, is

Gelfand corollaries[edit]

Gelfand's formula leads directly to a bound on the spectral radius of a product of finitely many matrices, namely assuming that they all commute we obtain

Actually, in case the norm is consistent, the proof shows more than the thesis; in fact, using the previous lemma, we can replace in the limit definition the left lower bound with the spectral radius itself and write more precisely:

which, by definition, is

where the + means that the limit is approached from above.


Consider the matrix

whose eigenvalues are 5, 10, 10; by definition, ρ(A) = 10. In the following table, the values of for the four most used norms are listed versus several increasing values of k (note that, due to the particular form of this matrix,):

1 14 15.362291496 10.681145748
2 12.649110641 12.328294348 10.595665162
3 11.934831919 11.532450664 10.500980846
4 11.501633169 11.151002986 10.418165779
5 11.216043151 10.921242235 10.351918183
10 10.604944422 10.455910430 10.183690042
11 10.548677680 10.413702213 10.166990229
12 10.501921835 10.378620930 10.153031596
20 10.298254399 10.225504447 10.091577411
30 10.197860892 10.149776921 10.060958900
40 10.148031640 10.112123681 10.045684426
50 10.118251035 10.089598820 10.036530875
100 10.058951752 10.044699508 10.018248786
200 10.029432562 10.022324834 10.009120234
300 10.019612095 10.014877690 10.006079232
400 10.014705469 10.011156194 10.004559078
1000 10.005879594 10.004460985 10.001823382
2000 10.002939365 10.002230244 10.000911649
3000 10.001959481 10.001486774 10.000607757
10000 10.000587804 10.000446009 10.000182323
20000 10.000293898 10.000223002 10.000091161
30000 10.000195931 10.000148667 10.000060774
100000 10.000058779 10.000044600 10.000018232

Bounded Linear Operators[edit]

For a bounded linear operator A and the operator norm ||·||, again we have

A bounded operator (on a complex Hilbert space) is called a spectraloid operator if its spectral radius coincides with its numerical radius. An example of such an operator is a normal operator.

Notes and References[edit]

  1. ^ Pascanu, Razvan; Mikolov, Tomas; Bengio, Yoshua. "On the difficulty of training recurrent neural networks". Proceedings of the 30th International Conference on Machine Learning, PMLR. 28 (3): 1310–1318. Retrieved 8 July 2017. 
  2. ^ The formula holds for any Banach algebra; see Lax 2002, pp. 195-197

See also[edit]