Contents

function [z, history] = regressor_sel(A, b, K, rho)
% regressor_sel  Solve lasso problem via ADMM
%
% [x, history] = regressor_sel(A, b, K, rho, alpha)
%
% Attempts to solve the following problem via ADMM:
%
%   minimize || Ax - b ||_2^2
%   subject to card(x) <= K
%
% where card() is the number of nonzero entries.
%
% The solution is returned in the vector x.
%
% history is a structure that contains the objective value, the primal and
% dual residual norms, and the tolerances for the primal and dual residual
% norms at each iteration.
%
% rho is the augmented Lagrangian parameter.
%
% alpha is the over-relaxation parameter (typical values for alpha are
% between 1.0 and 1.8).
%
%
% More information can be found in the paper linked at:
% http://www.stanford.edu/~boyd/papers/distr_opt_stat_learning_admm.html
%


t_start = tic;

Global constants and defaults

QUIET    = 0;
MAX_ITER = 1000;
ABSTOL   = 1e-4;
RELTOL   = 1e-2;

Data preprocessing

[m, n] = size(A);

% save a matrix-vector multiply
Atb = A'*b;

ADMM solver

x = zeros(n,1);
z = zeros(n,1);
u = zeros(n,1);

% cache the factorization
[L U] = factor(A, rho);

if ~QUIET
    fprintf('%3s\t%10s\t%10s\t%10s\t%10s\t%10s\n', 'iter', ...
      'r norm', 'eps pri', 's norm', 'eps dual', 'objective');
end

for k = 1:MAX_ITER

    % x-update
    q = Atb + rho*(z - u);    % temporary value
    if( m >= n )    % if skinny
       x = U \ (L \ q);
    else            % if fat
       x = q/rho - (A'*(U \ ( L \ (A*q) )))/rho^2;
    end

    % z-update with relaxation
    zold = z;
    z = keep_largest(x + u, K);

    % u-update
    u = u + (x - z);

    % diagnostics, reporting, termination checks
    history.objval(k)  = objective(A, b, x);

    history.r_norm(k)  = norm(x - z);
    history.s_norm(k)  = norm(-rho*(z - zold));

    history.eps_pri(k) = sqrt(n)*ABSTOL + RELTOL*max(norm(x), norm(-z));
    history.eps_dual(k)= sqrt(n)*ABSTOL + RELTOL*norm(rho*u);

    if ~QUIET
        fprintf('%3d\t%10.4f\t%10.4f\t%10.4f\t%10.4f\t%10.2f\n', k, ...
            history.r_norm(k), history.eps_pri(k), ...
            history.s_norm(k), history.eps_dual(k), history.objval(k));
    end

    if (history.r_norm(k) < history.eps_pri(k) && ...
       history.s_norm(k) < history.eps_dual(k))
         break;
    end

end

if ~QUIET
    toc(t_start);
end
end

function p = objective(A, b, x)
    p = sum_square(A*x - b);
end

function z = keep_largest(z, K)
    [val pos] = sort(abs(z), 'descend');
    z(pos(K+1:end)) = 0;
end

function [L U] = factor(A, rho)
    [m, n] = size(A);
    if ( m >= n )    % if skinny
       L = chol( A'*A + rho*speye(n), 'lower' );
    else            % if fat
       L = chol( speye(m) + 1/rho*(A*A'), 'lower' );
    end

    % force matlab to recognize the upper / lower triangular structure
    L = sparse(L);
    U = sparse(L');
end