corner.m 8.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
function [k_corner,info] = corner(rho,eta,fig)
%CORNER Find corner of discrete L-curve via adaptive pruning algorithm.
%
% [k_corner,info] = corner(rho,eta,fig)
%
% Returns the integer k_corner such that the corner of the log-log
% L-curve is located at ( log(rho(k_corner)) , log(eta(k_corner)) ).
%
% The vectors rho and eta must contain corresponding values of the
% residual norm || A x - b || and the solution's (semi)norm || x ||
% or || L x || for a sequence of regularized solutions, ordered such
% that rho and eta are monotonic and such that the amount of
% regularization decreases as k increases.
%
% The second output argument describes possible warnings.
% Any combination of zeros and ones is possible.
%   info = 000 : No warnings - rho and eta describe a discrete 
%                   L-curve with a corner.
%   info = 001 : Bad data - some elements of rho and/or eta are
%                   Inf, NaN, or zero.
%   info = 010 : Lack of monotonicity - rho and/or eta are not
%                   strictly monotonic.
%   info = 100 : Lack of convexity - the L-curve described by rho
%                   and eta is concave and has no corner.
%
% The warnings described above will also result in text warnings on the
% command line. Type 'warning off Corner:warnings' to disable all
% command line warnings from this function.
%
% If a third input argument is present, then a figure will show the discrete
% L-curve in log-log scale and also indicate the found corner.

% Reference: P. C. Hansen, T. K. Jensen and G. Rodriguez, "An adaptive
% pruning algorithm for the discrete L-curve criterion," J. Comp. Appl.
% Math., 198 (2007), 483-492.
  
% Per Christian Hansen and Toke Koldborg Jensen, IMM, DTU;
% Giuseppe Rodriguez, University of Cagliari, Italy; March 22, 2006.

% Initialization of data
rho = rho(:);       % Make rho and eta column vectors.
eta = eta(:);

if (nargin < 3) | isempty(fig)
    fig = 0;        % Default is no figure.
elseif fig < 0,
    fig = 0;
end

info = 0;

fin = isfinite(rho+eta);    % NaN or Inf will cause trouble.
nzr = rho.*eta~=0;          % A zero will cause trouble.
kept = find(fin & nzr);
if isempty(kept)
	error('Too many Inf/NaN/zeros found in data')
end
if length(kept) < length(rho)
	info = info + 1;
	warning('Corner:warnings', ...
           ['Bad data - Inf, NaN or zeros found in data\n' ...
            '         Continuing with the remaining data'])
end
rho = rho(kept);            % rho and eta with bad data removed.
eta = eta(kept);

if any(rho(1:end-1)<rho(2:end)) | any(eta(1:end-1)>eta(2:end))
	info = info + 10;
	warning('Corner:warnings', 'Lack of monotonicity')
end

% Prepare for adaptive algorithm.
nP = length(rho);           % Number of points.
P = log10([rho eta]);       % Coordinates of the loglog L-curve.
V = P(2:nP,:)-P(1:nP-1,:);  % The vectors defined by these coordinates.
v = sqrt(sum(V.^2,2));      % The length of the vectors.
W = V./repmat(v,1,2);       % Normalized vectors.
clist = [];                 % List of candidates.
p = min(5, nP-1);           % Number of vectors in pruned L-curve.
convex = 0;                 % Are the pruned L-curves convex?

% Sort the vectors according to the length, the longest first.
[Y,I] = sort(v);
I = flipud(I);

% Main loop -- use a series of pruned L-curves. The two functions
% 'Angles' and 'Global_Behavior' are used to locate corners of the
% pruned L-curves. Put all the corner candidates in the clist vector.
while p < (nP-1)*2
	elmts = sort(I(1:min(p, nP-1)));
    
    % First corner location algorithm
	candidate = Angles( W(elmts,:), elmts);
	if candidate>0,
		convex = 1;
	end
	if candidate & ~any(clist==candidate)
		clist = [clist;candidate];
	end
    
    % Second corner location algorithm
	candidate = Global_Behavior(P, W(elmts,:), elmts);
	if ~any(clist==candidate)
		clist = [clist; candidate];
	end
    
	p = p*2;
end

% Issue a warning and return if none of the pruned L-curves are convex.
if convex==0
	k_corner = [];
	info = info + 100;
	warning('Corner:warnings', 'Lack of convexity')
	return
end

% Put rightmost L-curve point in clist if not already there; this is
% used below to select the corner among the corner candidates.
if sum(clist==1) == 0
	clist = [1;clist];
end

% Sort the corner candidates in increasing order.
clist = sort(clist);

% Select the best corner among the corner candidates in clist. 
% The philosophy is: select the corner as the rightmost corner candidate
% in the sorted list for which going to the next corner candidate yields
% a larger increase in solution (semi)norm than decrease in residual norm,
% provided that the L-curve is convex in the given point. If this is never
% the case, then select the leftmost corner candidate in clist.

vz = find(diff(P(clist,2)) ...  % Points where the increase in solution
     >= abs(diff(P(clist,1)))); % (semi)norm is larger than or equal
                                % to the decrease in residual norm.
if length(vz)>1
	if(vz(1) == 1),  vz = vz(2:end);  end
elseif length(vz)==1
	if(vz(1) == 1),  vz = [];  end
end

if isempty(vz)		
        % No large increase in solution (semi)norm is found and the
        % leftmost corner candidate in clist is selected.
	index = clist(end);
else
        % The corner is selected as described above.
	vects = [P(clist(2:end),1)-P(clist(1:end-1),1) ...
			P(clist(2:end),2)-P(clist(1:end-1),2)];
	vects = sparse(diag(1./sqrt(sum(vects.^2,2)))) * vects;
	delta = vects(1:end-1,1).*vects(2:end,2) ...
			- vects(2:end,1).*vects(1:end-1,2);
	vv = find(delta(vz-1)<=0);
	if isempty(vv)
		index = clist(vz(end));
	else
		index = clist(vz(vv(1)));
	end
end

% Corner according to original vectors without Inf, NaN, and zeros removed.
k_corner = kept(index);

if fig  % Show log-log L-curve and indicate the found corner.
    figure(fig); clf
    diffrho2 = (max(P(:,1))-min(P(:,1)))/2;
    diffeta2 = (max(P(:,2))-min(P(:,2)))/2;
    loglog(rho, eta, 'k--o'); hold on; axis square;
    % Mark the corner.
    loglog([min(rho)/100,rho(index)],[eta(index),eta(index)],':r',... 
           [rho(index),rho(index)],[min(eta)/100,eta(index)],':r') 
    % Scale axes to same number of decades.
    if abs(diffrho2)>abs(diffeta2),
        ax(1) = min(P(:,1)); ax(2) = max(P(:,1));
        mid = min(P(:,2)) + (max(P(:,2))-min(P(:,2)))/2;
        ax(3) = mid-diffrho2; ax(4) = mid+diffrho2;
    else
        ax(3) = min(P(:,2)); ax(4) = max(P(:,2));
        mid = min(P(:,1)) + (max(P(:,1))-min(P(:,1)))/2;
        ax(1) = mid-diffeta2; ax(2) = mid+diffeta2;
    end
    ax = 10.^ax; ax(1) = ax(1)/2; axis(ax);
	xlabel('residual norm || A x - b ||_2')
	ylabel('solution (semi)norm || L x ||_2');
    title(sprintf('Discrete L-curve, corner at %d', k_corner));
end

% =========================================================================
% First corner finding routine -- based on angles

function index = Angles( W, kv)

% Wedge products
delta = W(1:end-1,1).*W(2:end,2) - W(2:end,1).*W(1:end-1,2);

[mm kk] = min(delta);
if mm < 0		% Is it really a corner?
	index = kv(kk) + 1;
else			% If there is no corner, return 0.
	index = 0;
end

% =========================================================================
% Second corner finding routine -- based on global behavior of the L-curve

function index = Global_Behavior(P, vects, elmts)

hwedge = abs(vects(:,2));  % Abs of wedge products between
                           % normalized vectors and horizontal,
                           % i.e., angle of vectors with horizontal.
[An, In] = sort(hwedge);   % Sort angles in increasing order.

% Locate vectors for describing horizontal and vertical part of L-curve.
count = 1;
ln = length(In);
mn = In(1);
mx = In(ln);
while(mn>=mx)
	mx = max([mx In(ln-count)]);
	count = count + 1;
	mn = min([mn In(count)]);
end
if count > 1
	I = 0; J = 0;
	for i=1:count
		for j=ln:-1:ln-count+1
			if(In(i) < In(j))
				I = In(i); J = In(j); break
			end
		end
		if I>0, break; end
	end
else
	I = In(1); J = In(ln);
end

% Find intersection that describes the "origin".
x3 = P(elmts(J)+1,1)+(P(elmts(I),2)-P(elmts(J)+1,2))/(P(elmts(J)+1,2) ...
			-P(elmts(J),2))*(P(elmts(J)+1,1)-P(elmts(J),1));
origin = [x3 P(elmts(I),2)];

% Find distances from the original L-curve to the "origin".  The corner
% is the point with the smallest Euclidian distance to the "origin".
dists = (origin(1)-P(:,1)).^2+(origin(2)-P(:,2)).^2;
[Y,index] = min(dists);