使用l1-magic工具箱求解基追踪(BP)和基追踪降噪(BPDN)

题目:使用l1-magic工具箱求解基追踪(BP)和基追踪降噪(BPDN)

        基追踪(Basis Pursuit, BP)和基追踪降噪(Basis PursuitDe-Noising, BPDN)都不能称为是具体的算法,实际上分别是求解一个优化问题:

基追踪:

基追踪降噪:

        BP和BPDN可以由多种优化求解方法去解。为了求解这两个优化问题,基追踪可以转化为一个线性规划问题(参见《压缩感知重构算法之基追踪(Basis Pursuit, BP)》,http://blog.youkuaiyun.com/jbb0523/article/details/51986554),可以使用MATLAB自带的linprog函数求解;而基追踪降噪可以转化为一个二次规划问题(参见《压缩感知重构算法之基追踪降噪(Basis Pursuit De-Noising, BPDN)》,http://blog.youkuaiyun.com/jbb0523/article/details/52013669),可以使用MATLAB自带的quadprog函数求解。

        当然也可以用其它方法求解基追踪和基追踪降噪这两个问题。实际上同一个算法,不同的实现方式,运行效率也会有差异。

        由Emmanuel Candès and JustinRomberg等人开发的l1-magic工具箱(主页链接:http://users.ece.gatech.edu/justin/l1magic/)可以求解七类优化问题,其中(P1)问题即为基追踪,(P2)问题等价于基追踪降噪,详情可参见工具箱的说明文档[2]

(P1)问题:

(P2)问题:

1、(P1)问题

        求解(P1)问题的函数是l1eq_pd.m,位于工具箱\l1magic\Optimization目录下。该函数的使用例程参见l1eq_example.m,位于工具箱根目录\l1magic下。在例程l1eq_example中,函数l1eq_pd共有两种调用方式,如下图所示:

        默认情况是第一种,第二种为large scale(大规模),需要注意的是,当使用largescale这种调用方式时,需要额外使用工具箱中的cgsolve.m函数,位于工具箱\l1magic\Optimization目录下。若使用第二种方式,只需将第50行到54行解除注释即可(参见附录代码)。

2、(P2)问题

        求解(P2)问题的函数是l1qc_logbarrier.m,位于工具箱\l1magic\Optimization目录下。该函数的使用例程参见l1qc_example.m,位于工具箱根目录\l1magic下。需要注意的是,该函数需要调用l1qc_newton.m函数(位于工具箱\l1magic\Optimization目录下),在例程l1qc_example中,函数l1qc_logbarrier.m共有两种调用方式,如下图所示:

        默认情况是第一种,第二种为large scale(大规模),需要注意的是,当使用largescale这种调用方式时,也需要额外使用工具箱中的cgsolve.m函数。若使用第二种方式,只需将第49行到53行解除注释即可(参见附录代码)。

        先简单了解这么一点即可,若实际需要,再详细琢磨一下。

3、总结

        总结一下,对于(P1)问题,相关文件有三个:l1eq_pd.m(求解函数)、cgsolve.m(大规模方式时l1eq_pd需调用此文件)、l1eq_example.m(使用范例);对于(P2)问题,相关文件有四个:l1qc_logbarrier.m(求解函数)、l1qc_newton.m(l1qc_logbarrier需要调用此文件)、cgsolve.m(大规模方式时l1qc_logbarrier需调用此文件)、l1qc_example.m(使用范例)。

        值得注意的是,函数l1qc_logbarrier求解基追踪降噪问题的效率和重构效果均要优于本人基于MATLAB自带的quadprog函数编写的BPDN_quadprog(参见《压缩感知重构算法之基追踪降噪(BasisPursuit De-Noising, BPDN)》,http://blog.youkuaiyun.com/jbb0523/article/details/52013669),具体原因没有深究。

4、参考文献

【1】ChenS S, Donoho D L, Saunders M A.Atomicdecomposition by basis pursuit[J]. SIAM review, 2001, 43(1): 129-159.(Available at:http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.37.4272&rep=rep1&type=pdf)

【2】Emmanuel Candès and Justin Romberg.l1-magic : Recovery of Sparse Signals via Convex Programming[EB/OL].http://users.ece.gatech.edu/justin/l1magic/downloads/l1magic.pdf

5、附录

        作为一种备份,也作为一篇文档的完整性,这里将(P1)(P2)所涉及的MATLAB文件附在此,如果不愿意自己去下载,可以将这些代码保存为MATLAB文件即可。

(1)第一个是两个问题在large scale方式下都要调用的cgsolve.m

% cgsolve.m
%
% Solve a symmetric positive definite system Ax = b via conjugate gradients.
%
% Usage: [x, res, iter] = cgsolve(A, b, tol, maxiter, verbose)
%
% A - Either an NxN matrix, or a function handle.
%
% b - N vector
%
% tol - Desired precision.  Algorithm terminates when 
%    norm(Ax-b)/norm(b) < tol .
%
% maxiter - Maximum number of iterations.
%
% verbose - If 0, do not print out progress messages.
%    If and integer greater than 0, print out progress every 'verbose' iters.
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%

function [x, res, iter] = cgsolve(A, b, tol, maxiter, verbose)

if (nargin < 5), verbose = 1; end

implicit = isa(A,'function_handle');

x = zeros(length(b),1);
r = b;
d = r;
delta = r'*r;
delta0 = b'*b;
numiter = 0;
bestx = x;
bestres = sqrt(delta/delta0); 
while ((numiter < maxiter) && (delta > tol^2*delta0))

  % q = A*d
  if (implicit), q = A(d);  else  q = A*d;  end
 
  alpha = delta/(d'*q);
  x = x + alpha*d;
  
  if (mod(numiter+1,50) == 0)
    % r = b - Aux*x
    if (implicit), r = b - A(x);  else  r = b - A*x;  end
  else
    r = r - alpha*q;
  end
  
  deltaold = delta;
  delta = r'*r;
  beta = delta/deltaold;
  d = r + beta*d;
  numiter = numiter + 1;
  if (sqrt(delta/delta0) < bestres)
    bestx = x;
    bestres = sqrt(delta/delta0);
  end    
  
  if ((verbose) && (mod(numiter,verbose)==0))
    disp(sprintf('cg: Iter = %d, Best residual = %8.3e, Current residual = %8.3e', ...
      numiter, bestres, sqrt(delta/delta0)));
  end
  
end

if (verbose)
  disp(sprintf('cg: Iterations = %d, best residual = %14.8e', numiter, bestres));
end
x = bestx;
res = bestres;
iter = numiter;

(2)第二个是(P1)问题求解函数l1eq_pd.m

% l1eq_pd.m
%
% Solve
% min_x ||x||_1  s.t.  Ax = b
%
% Recast as linear program
% min_{x,u} sum(u)  s.t.  -u <= x <= u,  Ax=b
% and use primal-dual interior point method
%
% Usage: xp = l1eq_pd(x0, A, At, b, pdtol, pdmaxiter, cgtol, cgmaxiter)
%
% x0 - Nx1 vector, initial point.
%
% A - Either a handle to a function that takes a N vector and returns a K 
%     vector , or a KxN matrix.  If A is a function handle, the algorithm
%     operates in "largescale" mode, solving the Newton systems via the
%     Conjugate Gradients algorithm.
%
% At - Handle to a function that takes a K vector and returns an N vector.
%      If A is a KxN matrix, At is ignored.
%
% b - Kx1 vector of observations.
%
% pdtol - Tolerance for primal-dual algorithm (algorithm terminates if
%     the duality gap is less than pdtol).  
%     Default = 1e-3.
%
% pdmaxiter - Maximum number of primal-dual iterations.  
%     Default = 50.
%
% cgtol - Tolerance for Conjugate Gradients; ignored if A is a matrix.
%     Default = 1e-8.
%
% cgmaxiter - Maximum number of iterations for Conjugate Gradients; ignored
%     if A is a matrix.
%     Default = 200.
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%

function xp = l1eq_pd(x0, A, At, b, pdtol, pdmaxiter, cgtol, cgmaxiter)

largescale = isa(A,'function_handle');

if (nargin < 5), pdtol = 1e-3;  end
if (nargin < 6), pdmaxiter = 50;  end
if (nargin < 7), cgtol = 1e-8;  end
if (nargin < 8), cgmaxiter = 200;  end

N = length(x0);

alpha = 0.01;
beta = 0.5;
mu = 10;

gradf0 = [zeros(N,1); ones(N,1)];

% starting point --- make sure that it is feasible
if (largescale)
  if (norm(A(x0)-b)/norm(b) > cgtol)
    disp('Starting point infeasible; using x0 = At*inv(AAt)*y.');
    AAt = @(z) A(At(z));
    [w, cgres, cgiter] = cgsolve(AAt, b, cgtol, cgmaxiter, 0);
    if (cgres > 1/2)
      disp('A*At is ill-conditioned: cannot find starting point');
      xp = x0;
      return;
    end
    x0 = At(w);
  end
else
  if (norm(A*x0-b)/norm(b) > cgtol)
    disp('Starting point infeasible; using x0 = At*inv(AAt)*y.');
    opts.POSDEF = true; opts.SYM = true;
    [w, hcond] = linsolve(A*A', b, opts);
    if (hcond < 1e-14)
      disp('A*At is ill-conditioned: cannot find starting point');
      xp = x0;
      return;
    end
    x0 = A'*w;
  end  
end
x = x0;
u = (0.95)*abs(x0) + (0.10)*max(abs(x0));

% set up for the first iteration
fu1 = x - u;
fu2 = -x - u;
lamu1 = -1./fu1;
lamu2 = -1./fu2;
if (largescale)
  v = -A(lamu1-lamu2);
  Atv = At(v);
  rpri = A(x) - b;
else
  v = -A*(lamu1-lamu2);
  Atv = A'*v;
  rpri = A*x - b;
end

sdg = -(fu1'*lamu1 + fu2'*lamu2);
tau = mu*2*N/sdg;

rcent = [-lamu1.*fu1; -lamu2.*fu2] - (1/tau);
rdual = gradf0 + [lamu1-lamu2; -lamu1-lamu2] + [Atv; zeros(N,1)];
resnorm = norm([rdual; rcent; rpri]);

pditer = 0;
done = (sdg < pdtol) | (pditer >= pdmaxiter);
while (~done)
  
  pditer = pditer + 1;
  
  w1 = -1/tau*(-1./fu1 + 1./fu2) - Atv;
  w2 = -1 - 1/tau*(1./fu1 + 1./fu2);
  w3 = -rpri;
  
  sig1 = -lamu1./fu1 - lamu2./fu2;
  sig2 = lamu1./fu1 - lamu2./fu2;
  sigx = sig1 - sig2.^2./sig1;
  
  if (largescale)
    w1p = w3 - A(w1./sigx - w2.*sig2./(sigx.*sig1));
    h11pfun = @(z) -A(1./sigx.*At(z));
    [dv, cgres, cgiter] = cgsolve(h11pfun, w1p, cgtol, cgmaxiter, 0);
    if (cgres > 1/2)
      disp('Cannot solve system.  Returning previous iterate.  (See Section 4 of notes for more information.)');
      xp = x;
      return
    end
    dx = (w1 - w2.*sig2./sig1 - At(dv))./sigx;
    Adx = A(dx);
    Atdv = At(dv);
  else
    w1p = -(w3 - A*(w1./sigx - w2.*sig2./(sigx.*sig1)));
    H11p = A*(sparse(diag(1./sigx))*A');
    opts.POSDEF = true; opts.SYM = true;
    [dv,hcond] = linsolve(H11p, w1p, opts);
    if (hcond < 1e-14)
      disp('Matrix ill-conditioned.  Returning previous iterate.  (See Section 4 of notes for more information.)');
      xp = x;
      return
    end
    dx = (w1 - w2.*sig2./sig1 - A'*dv)./sigx;
    Adx = A*dx;
    Atdv = A'*dv;
  end
  
  du = (w2 - sig2.*dx)./sig1;
  
  dlamu1 = (lamu1./fu1).*(-dx+du) - lamu1 - (1/tau)*1./fu1;
  dlamu2 = (lamu2./fu2).*(dx+du) - lamu2 - 1/tau*1./fu2;
  
  % make sure that the step is feasible: keeps lamu1,lamu2 > 0, fu1,fu2 < 0
  indp = find(dlamu1 < 0);  indn = find(dlamu2 < 0);
  s = min([1; -lamu1(indp)./dlamu1(indp); -lamu2(indn)./dlamu2(indn)]);
  indp = find((dx-du) > 0);  indn = find((-dx-du) > 0);
  s = (0.99)*min([s; -fu1(indp)./(dx(indp)-du(indp)); -fu2(indn)./(-dx(indn)-du(indn))]);
  
  % backtracking line search
  suffdec = 0;
  backiter = 0;
  while (~suffdec)
    xp = x + s*dx;  up = u + s*du; 
    vp = v + s*dv;  Atvp = Atv + s*Atdv; 
    lamu1p = lamu1 + s*dlamu1;  lamu2p = lamu2 + s*dlamu2;
    fu1p = xp - up;  fu2p = -xp - up;  
    rdp = gradf0 + [lamu1p-lamu2p; -lamu1p-lamu2p] + [Atvp; zeros(N,1)];
    rcp = [-lamu1p.*fu1p; -lamu2p.*fu2p] - (1/tau);
    rpp = rpri + s*Adx;
    suffdec = (norm([rdp; rcp; rpp]) <= (1-alpha*s)*resnorm);
    s = beta*s;
    backiter = backiter + 1;
    if (backiter > 32)
      disp('Stuck backtracking, returning last iterate.  (See Section 4 of notes for more information.)')
      xp = x;
      return
    end
  end
  
  
  % next iteration
  x = xp;  u = up;
  v = vp;  Atv = Atvp; 
  lamu1 = lamu1p;  lamu2 = lamu2p;
  fu1 = fu1p;  fu2 = fu2p;
  
  % surrogate duality gap
  sdg = -(fu1'*lamu1 + fu2'*lamu2);
  tau = mu*2*N/sdg;
  rpri = rpp;
  rcent = [-lamu1.*fu1; -lamu2.*fu2] - (1/tau);
  rdual = gradf0 + [lamu1-lamu2; -lamu1-lamu2] + [Atv; zeros(N,1)];
  resnorm = norm([rdual; rcent; rpri]);
  
  done = (sdg < pdtol) | (pditer >= pdmaxiter);
  
  disp(sprintf('Iteration = %d, tau = %8.3e, Primal = %8.3e, PDGap = %8.3e, Dual res = %8.3e, Primal res = %8.3e',...
    pditer, tau, sum(u), sdg, norm(rdual), norm(rpri)));
  if (largescale)
    disp(sprintf('                  CG Res = %8.3e, CG Iter = %d', cgres, cgiter));
  else
    disp(sprintf('                  H11p condition number = %8.3e', hcond));
  end
  
end

(3)第三个是 (P1)问题使用范例l1eq_example.m(处于large scale方式下)

% l1eq_example.m
%
% Test out l1eq code (l1 minimization with equality constraints).
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%

% put key subdirectories in path if not already there
% path(path, './Optimization');
% path(path, './Data');

% To reproduce the example in the documentation, uncomment the 
% two lines below
%load RandomStates
%rand('state', rand_state);
%randn('state', randn_state);

% signal length
N = 512;
% number of spikes in the signal
T = 20;
% number of observations to make
K = 120;

% random +/- 1 signal
x = zeros(N,1);
q = randperm(N);
x(q(1:T)) = sign(randn(T,1));

% measurement matrix
disp('Creating measurment matrix...');
A = randn(K,N);
A = orth(A')';
disp('Done.');
	
% observations
y = A*x;

% initial guess = min energy
x0 = A'*y;

% solve the LP
% tic
% xp = l1eq_pd(x0, A, [], y, 1e-3);
% toc

% large scale
Afun = @(z) A*z;
Atfun = @(z) A'*z;
tic
xp = l1eq_pd(x0, Afun, Atfun, y, 1e-3, 30, 1e-8, 200);
toc

(4)第四个是 (P2)问题求解函数l1qc_logbarrier.m

% l1qc_logbarrier.m
%
% Solve quadratically constrained l1 minimization:
% min ||x||_1   s.t.  ||Ax - b||_2 <= \epsilon
%
% Reformulate as the second-order cone program
% min_{x,u}  sum(u)   s.t.    x - u <= 0,
%                            -x - u <= 0,
%      1/2(||Ax-b||^2 - \epsilon^2) <= 0
% and use a log barrier algorithm.
%
% Usage:  xp = l1qc_logbarrier(x0, A, At, b, epsilon, lbtol, mu, cgtol, cgmaxiter)
%
% x0 - Nx1 vector, initial point.
%
% A - Either a handle to a function that takes a N vector and returns a K 
%     vector , or a KxN matrix.  If A is a function handle, the algorithm
%     operates in "largescale" mode, solving the Newton systems via the
%     Conjugate Gradients algorithm.
%
% At - Handle to a function that takes a K vector and returns an N vector.
%      If A is a KxN matrix, At is ignored.
%
% b - Kx1 vector of observations.
%
% epsilon - scalar, constraint relaxation parameter
%
% lbtol - The log barrier algorithm terminates when the duality gap <= lbtol.
%         Also, the number of log barrier iterations is completely
%         determined by lbtol.
%         Default = 1e-3.
%
% mu - Factor by which to increase the barrier constant at each iteration.
%      Default = 10.
%
% cgtol - Tolerance for Conjugate Gradients; ignored if A is a matrix.
%     Default = 1e-8.
%
% cgmaxiter - Maximum number of iterations for Conjugate Gradients; ignored
%     if A is a matrix.
%     Default = 200.
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%

function xp = l1qc_logbarrier(x0, A, At, b, epsilon, lbtol, mu, cgtol, cgmaxiter)  

largescale = isa(A,'function_handle');

if (nargin < 6), lbtol = 1e-3; end
if (nargin < 7), mu = 10; end
if (nargin < 8), cgtol = 1e-8; end
if (nargin < 9), cgmaxiter = 200; end

newtontol = lbtol;
newtonmaxiter = 50;

N = length(x0);

% starting point --- make sure that it is feasible
if (largescale)
  if (norm(A(x0)-b) > epsilon)
    disp('Starting point infeasible; using x0 = At*inv(AAt)*y.');
    AAt = @(z) A(At(z));
    [w, cgres] = cgsolve(AAt, b, cgtol, cgmaxiter, 0);
    if (cgres > 1/2)
      disp('A*At is ill-conditioned: cannot find starting point');
      xp = x0;
      return;
    end
    x0 = At(w);
  end
else
  if (norm(A*x0-b) > epsilon)
    disp('Starting point infeasible; using x0 = At*inv(AAt)*y.');
    opts.POSDEF = true; opts.SYM = true;
    [w, hcond] = linsolve(A*A', b, opts);
    if (hcond < 1e-14)
      disp('A*At is ill-conditioned: cannot find starting point');
      xp = x0;
      return;
    end
    x0 = A'*w;
  end  
end
x = x0;
u = (0.95)*abs(x0) + (0.10)*max(abs(x0));

disp(sprintf('Original l1 norm = %.3f, original functional = %.3f', sum(abs(x0)), sum(u)));

% choose initial value of tau so that the duality gap after the first
% step will be about the origial norm
tau = max((2*N+1)/sum(abs(x0)), 1);
                                                                                                                          
lbiter = ceil((log(2*N+1)-log(lbtol)-log(tau))/log(mu));
disp(sprintf('Number of log barrier iterations = %d\n', lbiter));

totaliter = 0;

for ii = 1:lbiter

  [xp, up, ntiter] = l1qc_newton(x, u, A, At, b, epsilon, tau, newtontol, newtonmaxiter, cgtol, cgmaxiter);
  totaliter = totaliter + ntiter;
  
  disp(sprintf('\nLog barrier iter = %d, l1 = %.3f, functional = %8.3f, tau = %8.3e, total newton iter = %d\n', ...
    ii, sum(abs(xp)), sum(up), tau, totaliter));
  
  x = xp;
  u = up;
 
  tau = mu*tau;
  
end

(5)第五个是 (P2)问题求解函数须调用的l1qc_newton.m

% l1qc_newton.m
%
% Newton algorithm for log-barrier subproblems for l1 minimization
% with quadratic constraints.
%
% Usage: 
% [xp,up,niter] = l1qc_newton(x0, u0, A, At, b, epsilon, tau, 
%                             newtontol, newtonmaxiter, cgtol, cgmaxiter)
%
% x0,u0 - starting points
%
% A - Either a handle to a function that takes a N vector and returns a K 
%     vector , or a KxN matrix.  If A is a function handle, the algorithm
%     operates in "largescale" mode, solving the Newton systems via the
%     Conjugate Gradients algorithm.
%
% At - Handle to a function that takes a K vector and returns an N vector.
%      If A is a KxN matrix, At is ignored.
%
% b - Kx1 vector of observations.
%
% epsilon - scalar, constraint relaxation parameter
%
% tau - Log barrier parameter.
%
% newtontol - Terminate when the Newton decrement is <= newtontol.
%         Default = 1e-3.
%
% newtonmaxiter - Maximum number of iterations.
%         Default = 50.
%
% cgtol - Tolerance for Conjugate Gradients; ignored if A is a matrix.
%     Default = 1e-8.
%
% cgmaxiter - Maximum number of iterations for Conjugate Gradients; ignored
%     if A is a matrix.
%     Default = 200.
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%


function [xp, up, niter] = l1qc_newton(x0, u0, A, At, b, epsilon, tau, newtontol, newtonmaxiter, cgtol, cgmaxiter) 

% check if the matrix A is implicit or explicit
largescale = isa(A,'function_handle');

% line search parameters
alpha = 0.01;
beta = 0.5;  

if (~largescale), AtA = A'*A; end

% initial point
x = x0;
u = u0;
if (largescale), r = A(x) - b; else  r = A*x - b; end
fu1 = x - u;
fu2 = -x - u;
fe = 1/2*(r'*r - epsilon^2);
f = sum(u) - (1/tau)*(sum(log(-fu1)) + sum(log(-fu2)) + log(-fe));

niter = 0;
done = 0;
while (~done)
  
  if (largescale), atr = At(r); else  atr = A'*r; end
  
  ntgz = 1./fu1 - 1./fu2 + 1/fe*atr;
  ntgu = -tau - 1./fu1 - 1./fu2;
  gradf = -(1/tau)*[ntgz; ntgu];
  
  sig11 = 1./fu1.^2 + 1./fu2.^2;
  sig12 = -1./fu1.^2 + 1./fu2.^2;
  sigx = sig11 - sig12.^2./sig11;
    
  w1p = ntgz - sig12./sig11.*ntgu;
  if (largescale)
    h11pfun = @(z) sigx.*z - (1/fe)*At(A(z)) + 1/fe^2*(atr'*z)*atr;
    [dx, cgres, cgiter] = cgsolve(h11pfun, w1p, cgtol, cgmaxiter, 0);
    if (cgres > 1/2)
      disp('Cannot solve system.  Returning previous iterate.  (See Section 4 of notes for more information.)');
      xp = x;  up = u;
      return
    end
    Adx = A(dx);
  else
    H11p = diag(sigx) - (1/fe)*AtA + (1/fe)^2*atr*atr';
    opts.POSDEF = true; opts.SYM = true;
    [dx,hcond] = linsolve(H11p, w1p, opts);
    if (hcond < 1e-14)
      disp('Matrix ill-conditioned.  Returning previous iterate.  (See Section 4 of notes for more information.)');
      xp = x;  up = u;
      return
    end
    Adx = A*dx;
  end
  du = (1./sig11).*ntgu - (sig12./sig11).*dx;  
 
  % minimum step size that stays in the interior
  ifu1 = find((dx-du) > 0); ifu2 = find((-dx-du) > 0);
  aqe = Adx'*Adx;   bqe = 2*r'*Adx;   cqe = r'*r - epsilon^2;
  smax = min(1,min([...
    -fu1(ifu1)./(dx(ifu1)-du(ifu1)); -fu2(ifu2)./(-dx(ifu2)-du(ifu2)); ...
    (-bqe+sqrt(bqe^2-4*aqe*cqe))/(2*aqe)
    ]));
  s = (0.99)*smax;
  
  % backtracking line search
  suffdec = 0;
  backiter = 0;
  while (~suffdec)
    xp = x + s*dx;  up = u + s*du;  rp = r + s*Adx;
    fu1p = xp - up;  fu2p = -xp - up;  fep = 1/2*(rp'*rp - epsilon^2);
    fp = sum(up) - (1/tau)*(sum(log(-fu1p)) + sum(log(-fu2p)) + log(-fep));
    flin = f + alpha*s*(gradf'*[dx; du]);
    suffdec = (fp <= flin);
    s = beta*s;
    backiter = backiter + 1;
    if (backiter > 32)
      disp('Stuck on backtracking line search, returning previous iterate.  (See Section 4 of notes for more information.)');
      xp = x;  up = u;
      return
    end
  end
  
  % set up for next iteration
  x = xp; u = up;  r = rp;
  fu1 = fu1p;  fu2 = fu2p;  fe = fep;  f = fp;
  
  lambda2 = -(gradf'*[dx; du]);
  stepsize = s*norm([dx; du]);
  niter = niter + 1;
  done = (lambda2/2 < newtontol) | (niter >= newtonmaxiter);
  
  disp(sprintf('Newton iter = %d, Functional = %8.3f, Newton decrement = %8.3f, Stepsize = %8.3e', ...
    niter, f, lambda2/2, stepsize));
  if (largescale)
    disp(sprintf('                CG Res = %8.3e, CG Iter = %d', cgres, cgiter));
  else
    disp(sprintf('                  H11p condition number = %8.3e', hcond));
  end
      
end

(6)第六个是(P2)问题使用范例l1qc_example.m(处于large scale方式下)

% l1qc_example.m
%
% Test out l1qc code (l1 minimization with quadratic constraint).
%
% Written by: Justin Romberg, Caltech
% Email: jrom@acm.caltech.edu
% Created: October 2005
%

% put optimization code in path if not already there
path(path, './Optimization');

% signal length
N = 512;

% number of spikes to put down
T = 20;

% number of observations to make
K = 120;

% random +/- 1 signal
x = zeros(N,1);
q = randperm(N);
x(q(1:T)) = sign(randn(T,1));

% measurement matrix
disp('Creating measurment matrix...');
A = randn(K,N);
A = orth(A')';
disp('Done.');
	
% noisy observations
sigma = 0.005;
e = sigma*randn(K,1);
y = A*x + e;

% initial guess = min energy
x0 = A'*y;

% take epsilon a little bigger than sigma*sqrt(K)
epsilon =  sigma*sqrt(K)*sqrt(1 + 2*sqrt(2)/sqrt(K));
                                                                                                                           
% tic
% xp = l1qc_logbarrier(x0, A, [], y, epsilon, 1e-3);
% toc

% large scale
Afun = @(z) A*z;
Atfun = @(z) A'*z;
tic
xp = l1qc_logbarrier(x0, Afun, Atfun, y, epsilon, 1e-3, 50, 1e-8, 500);
toc


评论 21
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值