% "FIR Filter Design via Spectral Factorization and Convex Optimization"
% by S.-P. Wu, S. Boyd, and L. Vandenberghe
% (figures are generated)
%
% Designs an FIR lowpass filter using spectral factorization method where we:
% - minimize the filter order
% - have a constraint on the maximum passband ripple
% - have a constraint on the maximum stopband attenuation
%
%   minimize   filter order n
%       s.t.   1/delta <= |H(w)| <= delta   for w in the passband
%              |H(w)| <= atten_level        for w in the stopband
%
% We change variables via spectral factorization method and get:
%
%   minimize   filter order n
%       s.t.   (1/delta)^2 <= R(w) <= delta^2  for w in the passband
%              R(w) <= atten_level^2           for w in the stopband
%              R(w) >= 0                       for all w
%
% where R(w) is the squared magnited of the frequency response
% (and the Fourier transform of the autocorrelation coefficients r).
% Variables are coeffients r. delta is the allowed passband ripple
% and atten_level is the max allowed level in the stopband.
%
% This is a quasiconvex problem and can be solved using a bisection.
%
% Written for CVX by Almir Mutapcic 02/02/06

%*********************************************************************
% user's filter specs (for a low-pass filter example)
%*********************************************************************
% filter order that is used to start the bisection (has to be feasible)
max_order = 20;

wpass = 0.12*pi;        % passband cutoff freq (in radians)
wstop = 0.24*pi;        % stopband start freq (in radians)
delta = 1;              % max (+/-) passband ripple in dB
atten = -30;      % stopband attenuation level in dB

%********************************************************************
% create optimization parameters
%********************************************************************
m = 15*(max_order);   % freq samples (rule-of-thumb)
w = linspace(0,pi,m);

%*********************************************************************
% use bisection algorithm to solve the problem
%*********************************************************************

n_bot = 1;
n_top = max_order;
n_best = Inf;

while( n_top - n_bot > 1)
    % try to find a feasible design for given specs
    n_cur = ceil( (n_top + n_bot)/2 );

    % create optimization matrices
    % A is the matrix used to compute the power spectrum
    % A(w,:) = [1 2*cos(w) 2*cos(2*w) ... 2*cos(n*w)]
    A = [ones(m,1) 2*cos(kron(w',[1:n_cur-1]))];

    % passband 0 <= w <= w_pass
    ind = find((0 <= w) & (w <= wpass));    % passband
    Ap  = A(ind,:);

    % transition band is not constrained (w_pass <= w <= w_stop)

    % stopband (w_stop <= w)
    ind = find((wstop <= w) & (w <= pi));   % stopband
    As  = A(ind,:);

    % This is the feasiblity problem:
    % cvx_begin quiet
    %     variable r_cur(n_cur+1,1);
    %     10^(-delta/10) <= Ap * r_cur <=  10^(+delta/10);
    %     abs( As * r_cur ) <= +10^(+atten/10);
    %     A * r >= 0;
    % cvx_end
    % Unfortunately it seems to be a bit unreliable to solve. So we have
    % reformulated it as a stopband minimization. If the optimum stopband
    % attenuation is smaller than 10^(atten/10), then we have feasibility.
    % formulate and solve the feasibility linear-phase lp filter design
    cvx_begin quiet
        variables r_cur(n_cur,1);
        minimize( max( abs( As * r_cur ) ) );
        10^(-delta/10) <= Ap * r_cur <= 10^(+delta/10);
        A * r_cur >= 0;
    cvx_end

    % bisection
    if isnan( cvx_optval ),
        fprintf( 1, 'Solver failed for n = %d taps, assuming infeasible\n', n_cur );
        n_bot = n_cur;
    elseif cvx_optval <= 10^(atten/10), % strfind(cvx_status,'Solved') % feasible
        fprintf(1,'Problem is feasible for n = %d taps\n',n_cur);
        n_top = n_cur;
        if n_best > n_cur, r = r_cur; end
    else % not feasible
        fprintf(1,'Problem not feasible for n = %d taps\n',n_cur);
        n_bot = n_cur;
    end
end

h = spectral_fact(r);
n = length(r);
fprintf(1,'\nOptimum number of filter taps for given specs is %d.\n',n);


%********************************************************************
% plots
%********************************************************************
figure(1)
% FIR impulse response
plot([0:n-1],h','o',[0:n-1],h','b:')
xlabel('t'), ylabel('h(t)')

figure(2)
% frequency response
H = exp(-j*kron(w',[0:n-1]))*h;
% magnitude
subplot(2,1,1)
plot(w,20*log10(abs(H)),...
    [wstop pi],[atten atten],'r--',...
    [0 wpass],[delta delta],'r--',...
    [0 wpass],[-delta -delta],'r--');
axis([0,pi,-40,10])
xlabel('w'), ylabel('mag H(w) in dB')
% phase
subplot(2,1,2)
plot(w,angle(H))
axis([0,pi,-pi,pi])
xlabel('w'), ylabel('phase H(w)')
Problem not feasible for n = 11 taps
Problem not feasible for n = 16 taps
Problem is feasible for n = 18 taps
Problem is feasible for n = 17 taps

Optimum number of filter taps for given specs is 17.