subspacea.m

function [theta,varargout] = subspacea(F,G,A)
%SUBSPACEA angles between subspaces
%  subspacea(F,G,A)
%  Finds all min(size(orth(F),2),size(orth(G),2)) principal angles
%  between two subspaces spanned by the columns of matrices F and G 
%  in the A-based scalar product x'*A*y, where A
%  is Hermitian and positive definite. 
%  COS of principal angles is called canonical correlations in statistics.  
%  [theta,U,V] = subspacea(F,G,A) also computes left and right
%  principal (canonical) vectors - columns of U and V, respectively.
%
%  If F and G are vectors of unit length and A=I, 
%  the angle is ACOS(F'*G) in exact arithmetic. 
%  If A is not provided as a third argument, than A=I and 
%  the function gives the same largest angle as SUBSPACE.m by Andrew Knyazev,
%  see
%  http://www.mathworks.com/matlabcentral/fileexchange/Files.jsp?type=category&id=&fileId=54
%  MATLAB's SUBSPACE.m function is still badly designed and fails to compute 
%  some angles accurately.
%
%  The optional parameter A is a Hermitian and positive definite matrix,
%  or a corresponding function. When A is a function, it must accept a
%  matrix as an argument. 
%  This code requires ORTHA.m, Revision 1.5.8 or above,
%  which is included. The standard MATLAB version of ORTH.m
%  is used for orthonormalization, but could be replaced by QR.m.
%  
%  Examples: 
%  F=rand(10,4); G=randn(10,6); theta = subspacea(F,G);
%  computes 4 angles between F and G, while in addition 
%  A=hilb(10); [theta,U,V] = subspacea(F,G,A);
%  computes angles relative to A and corresponding vectors U and V. 
%  
%  The algorithm is described in A. V. Knyazev and M. E. Argentati,
%  Principal Angles between Subspaces in an A-Based Scalar Product: 
%  Algorithms and Perturbation Estimates. SIAM Journal on Scientific Computing, 
%  23 (2002), no. 6, 2009-2041.
%  http://epubs.siam.org/sam-bin/dbq/article/37733

%  Tested under MATLAB R10-14
%  Copyright (c) 2000 Andrew Knyazev, Rico Argentati
%  Contact email: knyazev@na-net.ornl.gov
%  License: free software (BSD)
%  $Revision: 4.5 $  $Date: 2005/6/27


threshold=sqrt(2)/2; % Define threshold for determining when an angle is small

if size(F,1) ~= size(G,1)
   subspaceaError(['The row dimension ' int2str(size(F,1)) ...
         ' of the matrix F is not the same as ' int2str(size(G,1)) ...
         ' the row dimension of G'])
end

if nargin<3  % Compute angles using standard inner product
   
   % Trivial column scaling first, if ORTH.m is used later 
   for i=1:size(F,2),
     normi=norm(F(:,i),inf);
     %Adjustment makes tol consistent with experimental results
     if normi > eps^.981
       F(:,i)=F(:,i)/normi;
       % Else orth will take care of this
     end
   end
   for i=1:size(G,2),
     normi=norm(G(:,i),inf);
     %Adjustment makes tol consistent with experimental results
     if normi > eps^.981
       G(:,i)=G(:,i)/normi;
       % Else orth will take care of this
     end
   end

  % Compute angle using standard inner product
  
  QF = orth(F);      %This can also be done using QR.m, in which case
  QG = orth(G);      %the column scaling above is not needed 
  
  q = min(size(QF,2),size(QG,2));
  [Ys,s,Zs] = svd(QF'*QG,0);
  if size(s,1)==1
    % make sure s is column for output
    s=s(1);
  end
  s = min(diag(s),1);
  theta = max(acos(s),0);
  U = QF*Ys;
  V = QG*Zs;
  indexsmall = s > threshold;
  if max(indexsmall) % Check for small angles and recompute only small   
    RF = U(:,indexsmall); 
    RG = V(:,indexsmall); 
    %[Yx,x,Zx] = svd(RG-RF*(RF'*RG),0);
    [Yx,x,Zx] = svd(RG-QF*(QF'*RG),0); % Provides more accurate results
    if size(x,1)==1
      % make sure x is column for output
      x=x(1);
    end
    Tmp = fliplr(RG*Zx);
    V(:,indexsmall) = Tmp(:,indexsmall);
    U(:,indexsmall) = RF*(RF'*V(:,indexsmall))*...   
    diag(1./s(indexsmall)); 
    x = diag(x);               
    thetasmall=flipud(max(asin(min(x,1)),0));
    theta(indexsmall) = thetasmall(indexsmall);
  end     
  
  % Compute angle using inner product relative to A
  else 
    [m,n] = size(F);
    if ~isstr(A)
      [mA,mA] = size(A);
      if any(size(A) ~= mA)
        subspaceaError('Matrix A must be a square matrix or a string.')
      end
    if size(A) ~= m
       subspaceaError(['The size ' int2str(size(A)) ...
             ' of the matrix A is not the same as ' int2str(m) ...
             ' - the number of rows of F'])
    end
  end
  
  [QF,AQF]=ortha(A,F);
  [QG,AQG]=ortha(A,G);
  q = min(size(QF,2),size(QG,2));
  [Ys,s,Zs] = svd(QF'*AQG,0);
  if size(s,1)==1
    % make sure s is column for output
    s=s(1);
  end
  s=min(diag(s),1);
  theta = max(acos(s),0);
  U = QF*Ys;
  V = QG*Zs;
  indexsmall = s > threshold;  
  if max(indexsmall) % Check for small angles and recompute only small     
    RG = V(:,indexsmall); 
    AV = AQG*Zs;
    ARG = AV(:,indexsmall);  
    RF = U(:,indexsmall);
    %S=RG-RF*(RF'*(ARG));
    S=RG-QF*(QF'*(ARG));% A bit more cost, but seems more accurate
      
    % Normalize, so ortha would not delete wanted vectors
    for i=1:size(S,2),
      normSi=norm(S(:,i),inf);
      %Adjustment makes tol consistent with experimental results
      if normSi > eps^1.981
        QS(:,i)=S(:,i)/normSi;
        % Else ortha will take care of this
      end
    end

    [QS,AQS]=ortha(A,QS);
    [Yx,x,Zx] = svd(AQS'*S);
    if size(x,1)==1
      % make sure x is column for output
      x=x(1);
    end
    x = max(diag(x),0);
     
    Tmp  = fliplr(RG*Zx);
    ATmp = fliplr(ARG*Zx);
    V(:,indexsmall) = Tmp(:,indexsmall);
    AVindexsmall = ATmp(:,indexsmall);   
    U(:,indexsmall) = RF*(RF'*AVindexsmall)*...
                       diag(1./s(indexsmall)); 
    thetasmall=flipud(max(asin(min(x,1)),0));
    
    %Add zeros if necessary
    if sum(indexsmall)-size(thetasmall,1)>0
      thetasmall=[zeros(sum(indexsmall)-size(thetasmall,1),1)',...
           thetasmall']';
    end
    
    theta(indexsmall) = thetasmall(indexsmall);
  end
end
varargout(1)={U(:,1:q)};
varargout(2)={V(:,1:q)};

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [Q,varargout]=ortha(A,X)
%ORTHA Orthonormalization Relative to matrix A
%  Q=ortha(A,X)
%  Q=ortha('Afunc',X)
%  Computes an orthonormal basis Q for the range of X, relative to the 
%  scalar product using a positive definite and selfadjoint matrix A.
%  That is, Q'*A*Q = I, the columns of Q span the same space as
%  columns of X, and rank(Q)=rank(X).
%
%  [Q,AQ]=ortha(A,X) also gives AQ = A*Q.
%
%  Required input arguments:
%  A : either an m x m positive definite and selfadjoint matrix A
%  or a linear operator A=A(v) that is positive definite selfadjoint;
%  X : m x n matrix containing vectors to be orthonormalized relative 
%  to A.
%
%  ortha(eye(m),X) spans the same space as orth(X)
%
%  Examples:
%  [q,Aq]=ortha(hilb(20),eye(20,5))
%  computes 5 column-vectors q spanned by the first 5 coordinate vectors,
%  and orthonormal with respect to the scalar product given by the
%  20x20 Hilbert matrix,
%  while an attempt to orthogonalize (in the same scalar product)
%  all 20 coordinate vectors using
%  [q,Aq]=ortha(hilb(20),eye(20))
%  gives 14 column-vectors out of 20. 
%  Note that rank(hilb(20)) = 13 in double precision. 
%
%  Algorithm:
%  X=orth(X), [U,S,V]=SVD(X'*A*X), then Q=X*U*S^(-1/2)
%  If A is ill conditioned an extra step is performed to
%  improve the result. This extra step is performed only
%  if a test indicates that the program is running on a
%  machine that supports higher precison arithmetic
%  (greater than 64 bit precision).
%
%  See also ORTH, SVD
%
%  Copyright (c) 2000 Andrew Knyazev, Rico Argentati
%  Contact email: knyazev@na-net.ornl.gov
%  License: free software (BSD)
%  $Revision: 1.5.8 $  $Date: 2001/8/28
%  Tested under MATLAB R10-12.1

% Check input parameter A
[m,n] = size(X);
if ~isstr(A)
  [mA,mA] = size(A);
  if any(size(A) ~= mA)
    subspaceaError('Matrix A must be a square matrix or a string.')
  end
  if size(A) ~= m
    subspaceaError(['The size ' int2str(size(A)) ...
           ' of the matrix A does not match with ' int2str(m) ...
           ' - the number of rows of X'])
  end
end

% Normalize, so ORTH below would not delete wanted vectors
for i=1:size(X,2),
  normXi=norm(X(:,i),inf);
  %Adjustment makes tol consistent with experimental results
  if normXi > eps^.981
    X(:,i)=X(:,i)/normXi;
    % Else orth will take care of this
  end
end

% Make sure X is full rank and orthonormalize 
X=orth(X); %This can also be done using QR.m, in which case
           %the column scaling above is not needed 

%Set tolerance           
[m,n]=size(X);
tol=max(m,n)*eps;

% Compute an A-orthonormal basis
if ~isstr(A)
  AX = A*X;
else
  AX = feval(A,X);
end
XAX = X'*AX;

XAX = 0.5.*(XAX' + XAX);
[U,S,V]=svd(XAX);

if n>1 s=diag(S);
  elseif n==1, s=S(1);
  else s=0;
end

%Adjustment makes tol consistent with experimental results  
threshold1=max(m,n)*max(s)*eps^1.1;

r=sum(s>threshold1);
s(r+1:size(s,1))=1;
S=diag(1./sqrt(s),0);
X=X*U*S;
AX=AX*U*S;
XAX = X'*AX;

% Check subspaceaError against tolerance 
subspaceaError=normest(XAX(1:r,1:r)-eye(r));
% Check internal precision, e.g., 80bit FPU registers of P3/P4
precision_test=[1 eps/1024 -1]*[1 1 1]'; 
if subspaceaError<tol | precision_test==0;
  Q=X(:,1:r);
  varargout(1)={AX(:,1:r)};
  return
end

% Complete another iteration to improve accuracy
% if this machine supports higher internal precision 
if ~isstr(A)
  AX = A*X;
else
  AX = feval(A,X);
end
XAX = X'*AX;

XAX = 0.5.*(XAX' + XAX);
[U,S,V]=svd(XAX);

if n>1 s=diag(S);
  elseif n==1, s=S(1);
  else s=0;
end
   
threshold2=max(m,n)*max(s)*eps;
r=sum(s>threshold2);
S=diag(1./sqrt(s(1:r)),0);
Q=X*U(:,1:r)*S(1:r,1:r);
varargout(1)={AX*U(:,1:r)*S(1:r,1:r)};