Matlab code for this type of factorization.A has a SR decomposition A = SR , where S ∈ R^ 2n ×2n is a symplectic matrix, i.e. S ^TJS = J

J=[0 -I;I 0] where I∈R^ n ×n means identity matrix. R=[R11 R12;R21 R22]∈ R^ 2n ×2n , is constituted by upper triangular matrices R11 , R12 , R22 and strictly upper triangular matrix R21. diag (R11 ) =|diag (R22 )| and diag (R12 ) = 0 , then the SR decomposition is unique.

Answer 1

[Christine Tobler]

I'm not very acquainted with the SR decomposition (all I know about it I found just now by googling). I would suggest to take a look at what seems to be the original paper, "Matrix factorizations for symplectic QR-like methods" by Bunse-Gerstner. This seems to suggest an algorithm using other factorizations (e.g. QR) as building blocks.

I'm also tagging this Control, because it seems that this decomposition has applications in total control.

[Christine Tobler]

Also, take a look at the acknowledgements here: "On the sensitivity of the SR decomposition", Xiao-Wen Chang. They mention some MATLAB code for computing the SR decomposition being shared between researchers. Contacting one of them directly might be your best bet (although the paper is from 1998, so the code may not be easy to find).

Answer 2

[Farooq Aamir]

if true
  % code
end
function [c, v] = optsymhouse1(a)
twon = length(a); n = twon/2;
J = [zeros(n), eye(n); -eye(n), zeros(n)];
p = sign(a(1))*norm(a); aux = a(1)-p ;
if aux == 0
c = 0; v = zeros(twon, 1); %T = eye(twon);
elseif a(n + 1) == 0
display('division by zero');
return
else
v =a/aux;
c =aux^2/(p * a(n + 1));
v(1) = 1;
%T = (eye(twon) + c * v * v' * J);
end
function [c, v] =optsymhouse2(u)
twon = length(u); n = twon/2;
J = [zeros(n), eye(n); -eye(n), zeros(n)];
if n == 1
v = zeros(twon, 1); c = 0; %T = eye(twon);
else
I = [2 : n, n + 2 : twon]; e = norm(u(I));
if e == 0
v = zeros(twon, 1); c = 0; %T = eye(twon);
else
v1 = u(n + 1);
if v1 == 0
display('division by zero')
return
else
v = -u/e; v(1) = 1; v(n + 1) = 0; c = e/v1;
T = (eye(twon) + c * v * v' * J);
end
end
end
function[S,R] = SROSH(A)
[den, dep] = size(A);
n = den/2; p = dep/2;
v = zeros(den, 1);
JJ = [zeros(n) eye(n); -eye(n) zeros(n)];
S = eye(den);
for k = 1 : p
J= [zeros(n - k + 1), eye(n - k + 1); -eye(n - k + 1), zeros(n - k + 1)];
% Computing T 2k?1:
a = A([k : n, n + k : den],[k]);
[ c3,v3] = optsymhouse1(a);
T = eye(den-2*k+2) + c3*(v3*v3')*J ; % T not formed explicitly.
% Updating A:
A([k : n, n + k : den], [k : p, p + k : dep]) = T * A([k : n, n + k : den], [k : p, p + k :dep]);
% Computing T 2k?1 :
v(k : n) = v3( 1 : n - k + 1);
v(k + n : den) = v3(n -k + 2 : 2 * (n - k + 1));
TJ = eye(den) - c3*v*v'*JJ; % T J not formed explicitly;
% Updating S:
S = S * TJ;
% Computing T 2k:
u = A([k : n, n + k : den], p + k);
[ c3,v3] = optsymhouse2(u);
T = eye(den -2*k + 2) + c3 * (v3*v3')*J ;  % T not formed explicitly.
% Updating A[:
A([k : n, n + k : den], [k : p, p + k : dep]) = T *A([k : n, n + k : den], [k : p, p + k :dep]);
% Computing T 2k:
v(k  : n) = v3(1 : n - k + 1);
v(k  + n : den) = v3(n  - k + 2 : 2 * (n - k + 1));
TJ1 = (eye(den) - c3*(v*v')*JJ); % T J not formed explicitly;
% Updating S:
S = S * TJ1;
end
R = A;
%JJ'*S'*JJ*S = eye(2n);