| 1 |
function [phiKM, AscendingLambda] = Lanczos(K, M, sigma, Jmax) |
| 2 |
[rows,cols] = size(K); |
| 3 |
if (rows ~= cols) |
| 4 |
fprintf('Lanczos needs square matrices'); |
| 5 |
end |
| 6 |
Z = K - sigma*M; % initialize some |
| 7 |
Q = zeros(rows,Jmax+1); % variables |
| 8 |
T = zeros(Jmax,Jmax); % |
| 9 |
rRand = randn(rows,1); % |
| 10 |
rOld = rRand; |
| 11 |
betaOld = sqrt(rOld'*M*rOld); % |
| 12 |
for j = 1:Jmax, % |
| 13 |
Q(:,j+1) = rOld/betaOld; % |
| 14 |
u = Z \ (M*Q(:,j+1) - Z*Q(:,j)*betaOld); % D.S.Scott's formulation |
| 15 |
alpha = Q(:,j+1)'*M*u; % of the recurrence |
| 16 |
r = u - alpha*Q(:,j+1); % |
| 17 |
for i=1:3, % |
| 18 |
sum = zeros(rows,1); % repeat a full orhto- |
| 19 |
for k=2:j+1, % gonalization three |
| 20 |
sum = sum + (Q(:,k)'*M*r)*Q(:,k); % times to ensure |
| 21 |
end; % high quality |
| 22 |
r = r - sum; % solutions |
| 23 |
end; % |
| 24 |
beta = sqrt(r'*M*r); % |
| 25 |
T(j,j) = alpha; % |
| 26 |
if (j ~= Jmax) % augment [T] with new |
| 27 |
T(j+1,j) = beta; % alpha_i, beta_i+1 |
| 28 |
T(j,j+1) = beta; % |
| 29 |
end; % |
| 30 |
Jactual = j; % |
| 31 |
if (abs(beta) < 1.0e-12) % singular beta; going |
| 32 |
break % any more will introduce |
| 33 |
end % spurious modes |
| 34 |
betaOld = beta; % |
| 35 |
rOld = r; % |
| 36 |
end % |
| 37 |
[phiT,lambdaT] = eig(T(1:Jactual,1:Jactual)); % solve [T]{y} = L{y} |
| 38 |
lambdaKM = zeros(Jactual,1); % |
| 39 |
for j = 1:Jactual, % invert and shift the |
| 40 |
lambdaKM(j) = sigma + 1/lambdaT(j,j); % eigenvalues to the |
| 41 |
end % user's domain |
| 42 |
[AscendingLambda, ordering] = sort(lambdaKM); % |
| 43 |
phiKM = zeros(rows,Jactual); % sort the eigenvalues |
| 44 |
UnOrdphiKM = zeros(rows,Jactual); % in ascending order |
| 45 |
UnOrdphiKM = Q(:,2:Jactual+1)*phiT; % |
| 46 |
for j = 1:Jactual, % resequence the e-vectors |
| 47 |
phiKM(:,j) = UnOrdphiKM(:,ordering(j)); % to correspond to the |
| 48 |
end % e-values |