%lmsp

% [pstar,iter]=lmsp(func,jac,x0,tol,maxiter)
%
% Use the Levenberg-Marquardt algorithm to minimize
%
% f(x)=sum(F_i(x)^2)
%
% func name of the function F(x)
% jac name of the Jacobian function J(x)
% x0 initial guess
% tol stopping tolerance
% maxiter maximum number of iterations allowed
%
% Returns
% pstar best solution found.
% iter Iteration count.
%
function [pstar,iter]=lmsp(func,jacsp,p0,tol,maxiter)
%
% Initialize p and oldp.
%
p0=[1.0; 0.5; 0.60; -300.0]; % initial guess for depth, shape, degree and dipole moment
p=p0;
fp=norm(feval(func,p),2)^2;
oldp=p0*2;
oldfp=fp*2;
%
% Initialize lambda.
%
lambda=0.0001;
%
% The main loop. While the current solution isn't good enough, keep
% trying... Stop after maxiter iterations in the worst case.
%
iter=0;
while (iter <= maxiter)
%
% Compute the Jacobian.
%

1
 J=feval(jacsp,p);
%
% Compute rhs=-J'*f
%
% rhs=-J'*feval(func,p);
%
% We use a clever trick here. The least squares problem
%
% min || [ J ] s - [ -F ] ||
% || [ sqrt(lambda)*I ] [ 0 ] ||
%
% Has the normal equations solution
%
% s=-inv(J'*J+lambda*I)*J'*F
%
% which is precisely the LM step. We can solve this least squares problem
% more accurately using the QR factorization then by computing
% inv(J'*J+lambda*I) explicitly.
%
rhs=-J'*feval(func,p);
myrhs=[-feval(func,p); zeros(length(p),1)];
s=[J; sqrt(lambda)*eye(length(p))]\myrhs;
%
% Check the termination criteria.
%
if ((norm(rhs,2)< sqrt(tol)*(1+abs(fp))) & ...
(abs(oldfp-fp)<tol*(1+abs(fp))) & ...
(norm(oldp-p,2)<sqrt(tol)*(1+norm(p,2))))
pstar=p;
return;
end
%
% See whether this improves chisq or not.
%
fpnew=norm(feval(func,p+s),2)^2;
%
% If this improves f, then make the step, and decrease lambda and make
% the step.
%
if (fpnew < fp)

2
 oldp=p;
oldfp=fp;
p=p+s;
fp=fpnew;
lambda=lambda/2;
if (lambda <10^(-12))
lambda=1.0e-12;
end
iter=iter+1;
else
%
% Didn't improve f, increase lambda, and try again.
%
lambda=lambda*2.5;
if (lambda >10^16)
lambda=10^16;
end
iter=iter+1;
end
%
% end of the loop.
%
end
%
% Return, max iters exceeded.
%
pstar=p;
----------------------------------------------------------------------------
% p=chi2cdf(x,m)
%
% Computes the Chi^2 CDF, using a transformation to N(0,1)
% Note that x and m must be scalars.
%
function p=chi2cdf(x,m)
if (x == (m-1))
p=0.5;
else
z=(x-m+2/3-0.08/m)*sqrt((m-1)*log((m-1)/x)+x-(m-1) )/abs(x-m+1);
p=phi(z);
end

3
function J=jacsp(p)
%
% global variables.
global SIGMA;
global V;
global x;
%
n=length(x);
J=zeros(n,4);
for i=1:n
J(i,1)=(p(4)*sin(p(3))/((x(i)^2+p(1)^2)^p(2))-
2*p(4)*(x(i)*cos(p(3))+p(1)*sin(p(3)))/((x(i)^2+p(1)^2)^p(2))*p(2)*p(1)/(x(i)^2+p(1)^2))/SIGMA(i);
J(i,2)=(-p(4)*(x(i)*cos(p(3))+p(1)*sin(p(3)))/((x(i)^2+p(1)^2)^p(2))*log(x(i)^2+p(1)^2))/SIGMA(i);
J(i,3)=(p(4)*(-x(i)*sin(p(3))+p(1)*cos(p(3)))/((x(i)^2+p(1)^2)^p(2)))/SIGMA(i);
J(i,4)=((x(i)*cos(p(3))+p(1)*sin(p(3)))/((x(i)^2+p(1)^2)^p(2)))/SIGMA(i);
end

% Computes the differences between forward model prediction and data,

% normalized by the standard deviation for the slug test example.
function fvec=funsp(p)
%
% global variables.
global SIGMA;
global V;
global x;
%global Vnol;
%global xnol;
%global K;
%global teta;
%global Vmax;
%global Vmin;
%
% Compute the function values.
%
fvec=zeros(length(x),1);
for i=1:length(x)
fvec(i)=(p(4)*((x(i)*cos(p(3))+p(1)*sin(p(3)))/(((x(i))^2+p(1)^2)^p(2))) - V(i))/SIGMA(i);
end

4
%set figure fontsizes
function void=bookfonts
set(gca,'FontSize',18,'LineWidth',1.0);

%
% Calculates the normal distribution.
%
% z=phi(x)
%
% gives z=int((1/sqrt(2*pi))*exp(-t^2/2),t=-infinity..x)
%
%%%%%%%%%%%%% Forget the previous lines, now the real data!!!!

function z=phi(x)
if (x>=0);
z=.5+.5*erf(x/sqrt(2));
else
z=1-phi(-x);
end

% Nama file :LMinv

% Reset to a known state.
%
clear
rand('state',0);
randn('state',0);
%
% Global variables for observed V (SP anomaly) and x (measurement points)
global V;
global x;
global SIGMA;
%
% The unknown/estimated parameters are z=p(1), q=p(2), tetaak=p(3) and Kak=p(4):
% z is the depth of burial center and q is the shape factor
%
% The data. V is measured to the nearest mV
%%load V.dat;
%%load x.dat;
%load data2c.dat %%%%%%%%%%%%% Forget the previous lines, now the real data!!!!

5
%x=data2c(:,1);
%V=data2c(:,2);
load d1.dat %%%%%%%%%%%%% Forget the previous lines, now the real data!!!!
x=d1(:,1);
V=d1(:,2);
%%c% Fixed parameter values.
Data=load('d1.dat');
%
% Fixed parameter values
%
%
%sigma= 1mV (assumption); standard deviation is 1mV for all data
for i=1:length(x)
SIGMA(i)=1;
end
% Solve the least squares problem with LM.
%
%p0=[0.001; 1.0];
tic
p0=[1.0; 0.5; 0.60; -300.0]; % initial guess for depth, shape, degree and dipole moment
[pest,itercnt]=lmsp('funsp','jacsp',p0,1.0e-12,100); %calling lmsp.m, executing funsp.m and jacsp.m
%pest=estimated parameter; itercnt=number of iteration
%
toc
time = toc;
%% Inversi sebetulnya telah selesai

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Check the chi^2 value.
%
chi2=norm(funsp(pest))^2
pvalue=1-chi2cdf(chi2,length(V)-2)
%
% Compute the covariance matrix, correlation matrix, and confidence intervals.
%
Jest=jacsp(pest); % Jacobian of the estimated parameter

C=inv(Jest'*Jest);
m=max(size(C));
for i=1:m

6
 for j=1:m
Corm(i,j)=C(i,j)/(sqrt(C(i,i))*sqrt(C(j,j)));
end
end
fprintf('Iteration number = %.i\n',itercnt);
fprintf('z=%.5f +- %.5f\n',[pest(1) 1.96*sqrt(C(1,1))]);
fprintf('q=%.5f +- %.5f\n',[pest(2) 1.96*sqrt(C(2,2))]);
fprintf('teta=%.5f +- %.5f\n',[pest(3)*(180/pi) 1.96*sqrt(C(3,3))*(180/pi)]);
fprintf('K=%.5f +- %.5f\n',[pest(4) 1.96*sqrt(C(4,4))]);
fprintf('Covariance matrix is \n');
C
fprintf('Correlation matrix is \n');
Corm
%
% Now, plot the data and fitted curve.
%
figure(1);
clf;
bookfonts;
%tfit=0.001:1.0:51;
xfit=x;
for i=1:length(xfit)
Vfit(i)=pest(4)*((xfit(i)*cos(pest(3))+pest(1)*sin(pest(3)))/(((xfit(i))^2+pest(1)^2)^pest(2)));
end
plot(xfit,Vfit,'r-');

hold on
errorbar(x,V,SIGMA,'ko');
xlabel('distance (m)');
ylabel('SP anomaly (mV)');
legend 'calculated data'
grid on

%save padamelan2c.mat
save hasild1xxx.mat