Matrix Theory, Math6304 Lecture Notes from November 20, 2012 taken by Andy Chang Last Time: • Example for induced matrix norms • Matrix norms and spectral radius 5.3 Equivalence of norms and Gelfand formula 5.3.1 Theorem. Any two norms � · � and |||·||| on Cn are equivalent, that is, there exists c > 1 such that for every x ∈ Cn , 1 �x� ≤ |||x||| ≤ c�x�. c Proof. For later. 5.3.2 Theorem (Gelfand formula). If |||·||| is any matrix norm on Mn , A ∈ Mn , then ��� ��� 1 ρ(A) = lim ���Ak ��� k . k→∞ For the proof, we want to show double-sided inequalities. Proof. We have ρ(A)k as the maximum modulus occurring amongst the eigenvalues raised to the k-power so this is equivalent to taking the spectral radius of Ak , thus we obtain ρ(Ak ) = ρ(Ak ). By the usual inequality of the spectral radius and the matrix norm, we get the inequality ��� ��� ρ(Ak ) = ρ(Ak ) ≤ ���Ak ��� so taking the k-th root, ��� ��� 1 ρ(A) ≤ ���Ak ��� k . Conversely, we prove that for every � > 0, there exists K ∈ N such that for all for all k, k ≥ K, ��� k ��� ���A ��� ≤ ρ(A) + �. � � We recall that there is a norm � · � on Mn with �A� ≤ ρ(A) + 2� and there is a c ≥ 1 such that |||x||| ≤ c�x� for all x ∈ Mn . 1 Utilizing the previous theorem and submultiplicativity, we obtain, � �k ��� k ��� ���A ��� ≤ c�Ak � ≤ c�A�k ≤ c ρ(A) + � 2 and taking the k-th root � � ��� k ��� k1 ���A ��� ≤ c k1 ρ(A) + � 2 1/k Note that as k → ∞, c → 1. Thus, choosing K sufficiently large gives that, for all k ≥ K, ��� k ��� k1 ���A ��� ≤ ρ(A) + �. The first statement gives that there is matrix norm that is epsilon close to the spectral radius. The second statement says no matter the norm, raising it to a power, evaluating this result, then taking the kth root results in an expression which, as k → ∞, converges to the spectral radius. Thus, you can always find the largest eigenvalue for a matrix by looking at the powers and applying any matrix norm. This is useful in obtaining explicit formulas for matrix norms with difficult expressions. 5.4 Eigenvalue Perturbation Question: What can we say about the spectrum of a matrix A with Ai,i = 1 and ”small” off diagonal? Essentially, we have something that is close to the identity matrix, so we expect that all eigenvalues are close to 1. Answer: Let A = I + B, Bj,j = 0 for all j ∈ {1, 2, · · · , n}. Assume we have an eigenvalue λ, Ax = λx then, (I + B)x = λx ⇒ Bx = (λ − I)x so we have λ ∈ {z ∈ C : |z − 1| ≤ ρ(B)}. From ρ ≤ �B�∞→∞ = max j In terms of A, n � l=1 |Bj,l | |λ − 1| ≤ max j n � l=1 l�=j |Aj,l | Later, we will see how to relate all eigenvalues and diagonal entries with a similar expression. In order to prepare this, we will start with a qualitative result on the eigenvalue perturbation. 2 5.4.3 Theorem. Let {Ak } be a sequence in Mn , Ak → A entry-wise with some A ∈ Mn and λj (Ak ) are the eigenvalues of Ak . Then there exists permutations {πk } on {1, 2, · · · , n} such that lim λπk (Ak ) = λj (Ak ) k→+∞ Proof. Assume the converse. If the eigenvalues of a sequence {Ak }, where Ak → A, do not converge, then there exists � > 0, and eigenvalue λ of multiplicity m of A and subsequence {Aki } such that the number of eigenvalues of Aki in D� (λ) = {z ∈ K : |z − λ| ≤ �} is smaller than m. Using Schur’s theorem, we triangularize {Aki }, so Ti = Ui∗ Aki Ui and choose Ti and the columns of Ui such that the first n − m + 1 diagonal entries of Ti are outside of the epsilon ball D� (λ). (Schur triangularization allows to put the eigenvalues in any order.) By compactness of unitaries, we can choose a subsequence Uil → U with U unitary and Til = Ui∗l Akil Uil −→ U ∗ AU = T with T upper triangular (because each of Til is upper triangular). But A has eigenvalue λ on first m entries of its diagonal, which contradicts entry-wise convergence and the property of prelimiting Til . Next, we arrive at a quantitative result on pertubations. 5.4.4 Theorem (Geršgorin). Let A = [Ai,j ]ni,j=1 ∈ Mn . For j ∈ {1, 2, · · · , n}, Rj (A) = n � l=1 l�=j |Aj,l |, Gj (A) = {z ∈ C : |z − Aj,j | ≤ Rj (A), then if λ is an eigenvalue of A, we have, λ∈ n � Gj (A). j=1 Furthermore, if the union of k such discs Gj (A) is disconnected from the others, then it contains k-eigenvalues. Proof. Take λ as an eigenvalue, Ax = λx, where x �= 0. Pick p such that |xp | = �x�∞ then, n � λxp = [Ax]p = Ap,j xj j=1 and xp (λ − Ap,p ) = 3 n � j=1 j�=p Ap,j xj We estimate |xp ||λ − Ap,p | ≤ n � j=1 j�=p |Ap,j ||xj | = Rp (A) and λ is in a disc centered at some Ap,p with radius Rp (A). To prove eigenvalue count, we use continuity. Sketch: Figure 1: Write A = D + B, with D diagonal, Dj,j, = Aj,j for all j, and B = A − D. Set A� = D + �B, for 0 ≤ � ≤ 1. Note that Rj (A� ) = �Rj (A). WLOG, assume that the first k discs G1 (A), G2 (A), · · · , Gk (A) form a connected region, disjoint from other discs. Then, k � H(�) = Gj (A� ) ⊂ H(1) j=1 is also disjoint from Gj (A� ), j > k. Furthermore, by continuity of eigenvalues, � �→ λj (A� ) is continuous starting at λj (A0 ) = Aj,j and all λj (A� ) remain in the H(1). This implies that there are at least k eigenvalues of A1 ≡ A in H(1). Moreover, by the same argument for j > k and � j>k most k eigenvalues of A are in H(1). 4 Gj (A� ) � H(1) = ∅, we have that at