Measure similar information using Kullback-Leibler (KL) distance matlab code - matlab

I am trying implement the distance measurement between two distributions. The detail is described in here and input image . Let short summarize the idea of the paper:
The input image is divided into inside region and outside region by using Heaviside function H
Calculate the distribution inside region and outside region, in which phi is boundary of inside and outside.
Calculate the Kullback-Leibler distance
I implement that scheme, but I have three problems:
Log function in paper is log or log2 in matlab?
Log(0) is infinite, But we know that distribution result will return many 0 values. How to ignore it? In my case, I plus with eps value, some people add h1(h1==0)=1, which is correct?
Could you see my code? Is it correct? I am not sure about my implementation.
This is my code to implement that scheme:
function main()
Img=imread('1.bmp');%please download at above link
Img=double(Img(:,:,1));
%% Initial boundary
c0=2; %const value
phi = ones(size(Img(:,:,1))).*c0;
phi(26:32,28:34) = -c0;
%% Heaviside function
epsilon=1
Hu=0.5*(1+(2/pi)*atan(phi./epsilon));
%% Inside and outside image
inImg=Img.*(1-Hu);
outImg=Img.*(Hu);
%% Let caclulate KL distance
h1 = histogram(inImg, 256, 0, 255); %Histogram of inside
h2 = histogram(outImg, 256, 0, 255);%Histogram of outside
lamda1=KLdist(h1,h2) % distance from h1 to h2
lamda2=KLdist(h2,h1) % distance from h2 to h1
end
%%%%%%%%%% function for KL distance%%%%%%%%%%%%%%%
function [d1,d2]=KLdist(h1,h2)
d1=sum(h1.*log2(h1+eps)-h1.*log2(h2+eps))
d2=sum(h2.*log2(h2+eps)-h2.*log2(h1+eps))
end
%%%%%%%%%%function for histogram calculation%%%%%%
function [h,bins] = histogram(I, n, min, max)
I = I(:);
range = max - min;
drdb = range / double(n); % dr/db - change in range per bin
h = zeros(n,1);
bins = zeros(n,1);
for i=1:n
% note: while the instructions say "within integer round off" I'm leaving
% this as float bin edges, to handle the potential float input
% ie - say the input was a probability image.
low = min + (i-1)*drdb;
high = min + i*drdb;
h(i) = sum( (I>=low) .* (I<high) );
bins(i) = low;
end
h(n) = h(n) + sum( (I>=(n*drdb)) .* (I<=max) ); % include anything we may have missed in the last bin.
h = h ./ sum(h); % "relative frequency"
end

Let answer one by one
Log function in paper is log or log2 in matlab?
Ans: This is natural log. In matlab, you just call log()
Log(0) is infinite, But we know that distribution result will return many 0 values. How to ignore it?
Ans: To ignore the log of zero. You need add some small value as log(x+eps) or log(x+(x==0)*eps), where x is your values
Could you see my code? Is it correct? I am not sure about my implementation.
Ans: Your code looks fine. You can base on my suggestion to improve your code. Good luck

Related

Where did i do wrong when i tried to approximatee this data using polynomial?

I am starting to learn numerical analysis using MATLAB in my course. so far we have covered polynomial interpolation (spline, polyfit, constraint spline, etc.) I was doing this practice question and I can not get the correct answer. I have uploaded the code I used and the question, where did I do wrong? thanks in advance!
close all; clear all; clc;
format long e
x = linspace(0,1,8);
xplot = linspace(0,1);
f = #(x) atan(x.*(x+1));
y_val = f(xplot);
c = polyfit(x,f(x),7);
p = polyval(c,0.7);
err = abs(f(0.7)-p)/(f(0.7))
The question I encountered is seen in the picture
After some playing around, it seems to be a matter of computing the absolute error instead of the relative absolute error.
The code below yields the desired answer. And yes, it is pretty unclear from the question which error is intended.
% Definitions
format long e
x = linspace(0,1,8)';
xplot= linspace(0,1);
f = #(x) atan(x.*(x+1));
y_val = f(xplot);
% Degree of polynomial
n = 7;
% Points to evaluate function
point1 = 0.5;
point2 = 0.7;
% Fit
c= polyfit(x,f(x),n);
% Evaluate
approxPoint1 = polyval(c, point1);
approxPoint2 = polyval(c, point2);
% Absolute errors
errPoint1 = abs( f(point1) - approxPoint1)
errPoint2 = abs( f(point2) - approxPoint2)
What you did wrong was :
mixing absolute and relative values when calculating errors to feed resulting variable err.
incorrectly placing abs() parentheses when calculating err: Your abs() only fixes the numerator, but then the denominator. To obtain |f(0.7)| you also need another abs(f(0.7))
for point x=0.7 instead of
err = abs(f(0.7)-p)/(f(0.7))
could well simply be
err = abs(f(.7)-p));
you only calculate err for assessment point 0.5 . In order to choose among the possible candidates of what seems to be a MATLAB Associate test multichoice answer, one needs err on 0.5 and err on 0.7 and then match the pair in the correct order, among all offered possible answers.
Although it's common practice to approach polynomial approximation of N points with an N-1 degree polynomial, it's often possible to approximate below satisfactory error with lower degree polynomials than N-1.
Lower degree polynomials means less calculations, less time spent approximating points. If one obtains a fair enough approximation with for instance a degree 4 polynomial, why waste time calculating an approximation with a higher degree all the way up to N-1?
Since the question does not tell what degree should have the approximating polynomial, you have to find it sweeping all polynomial degrees from 1 to up to a reasonable order.
The last error I found is that you have used linspace without specifying amount of points, MATLAB then takes by default 100 points hoping it's going to be ok.
Well, in your question 100 points is way too low an amount of points as I am going to show after supplying the following lines, as mentioned in previous point, sweeping all possible approximating polynomials, NOT on default 100 points you tacitly chose.
np=8; % numel(x) amount supplied points
N=1e3; % amount points x grid we build to measure f(x)
x= linspace(0,1,np);
xplot = linspace(0,1,N);
f = #(x) atan(x.*(x+1)); % function to approximate
y_val = f(xplot); % f(x)
xm=[.5 .7]; % points where to asses error
% finding poly coeffs
err1=zeros(2,np+2); % 1st row are errors on x=.5, 2nd row errors on x=.7
figure(1);
ax1=gca
hp1=plot(xplot,y_val)
grid on;
hp1.LineWidth=3;
hp1.Color='r';
hold on
for k=1:1:np+2
c = polyfit(x,f(x),k);
p_01 = polyval(c,xm(1));
err1(1,k) = abs(f(xm(1))-p_01);
% err(1,k) = abs((f(0.5)-p_05)/(f(0.5)))
p_02 = polyval(c,xm(2));
err1(2,k) = abs(f(xm(2))-p_02);
% err(2,k) = abs((f(0.7)-p_07)/(f(0.7)))
plot(x,polyval(c,x),'LineWidth',1.5); %'Color','b');
end
err1
.
.
The only pair of errors matching in the correct order are indeed those of polynomial order 7, but the total smallest error corresponds to the approximating polynomial of order 6.
What happens when taking linspace without defining a large enough amount of points? let's have a look:
np=8; % numel(x) amount supplied points
% N=1e3; % amount points x grid we build to measure f(x)
x= linspace(0,1,np);
xplot = linspace(0,1); %,N);
f = #(x) atan(x.*(x+1)); % function to approximate
y_val = f(xplot); % f(x)
xm=[.5 .7]; % points where to asses error
% finding poly coeffs
err1=zeros(2,np+2); % 1st row are errors on x=.5, 2nd row errors on x=.7
figure(1);
ax1=gca
hp1=plot(xplot,y_val)
grid on;
hp1.LineWidth=3;
hp1.Color='r';
hold on
for k=1:1:np+2
c = polyfit(x,f(x),k);
p_01 = polyval(c,xm(1));
err1(1,k) = abs(f(xm(1))-p_01);
% err(1,k) = abs(f(0.5)-p_05)/abs(f(0.5))
p_02 = polyval(c,xm(2));
err1(2,k) = abs(f(xm(2))-p_02);
% err(2,k) = abs(f(0.7)-p_07)/abs(f(0.7))
plot(x,polyval(c,x),'LineWidth',1.5); %'Color','b');
end
err1
With only 100 points all errors come up way too large, not a single error anywhere near 1e-5 or 1e-6.
This is why one couldn't tell which pair to go for, because all obtained values where at least 5 orders of magnitude away from landing zone.
I was about to include a plot with legend, but the visualization of this particular approach is in this case and in my opinion at best misleading, as in both plots for 100 and 1000 points, at 1st glance, both look as if the errors should be similar regardless of the amount of grid points used.
But as shown above 1e2 points cannot approximate the function,
it's like pitch dark, looking for something and pointing torch 180 from where we should be aiming at, not a chance to spot it.
Yet 1e3 grid points produce a pair of errors matching one of the possible answers, this is option D.
I hope it helps, thanks for reading my answer.

How to do circular convolution between 2 functions with cconv?

I was asked to do circular convolution between two functions by sampling them, using the functions cconv. A known result of this sort of convolution is: CCONV( sin(x), sin(x) ) == -pi*cos(x)
To test the above I did:
w = linspace(0,2*pi,1000);
l = linspace(0,2*pi,1999);
stem(l,cconv(sin(w),sin(w))
but the result I got was:
which is absolutely not -pi*cos(x).
Can anybody please explain what is wrong with my code and how to fix it?
In the documentation of cconv it says that:
c = cconv(a,b,n) circularly convolves vectors a and b. n is the length of the resulting vector. If you omit n, it defaults to length(a)+length(b)-1. When n = length(a)+length(b)-1, the circular convolution is equivalent to the linear convolution computed with conv.
I believe that the reason for your problem is that you do not specify the 3rd input to cconv, which then selects the default value, which is not the right one for you. I have made an animation showing what happens when different values of n are chosen.
If you compare my result for n=200 to your plot you will see that the amplitude of your data is 10 times larger whereas the length of your linspace is 10 times bigger. This means that some normalization is needed, likely a multiplication by the linspace step.
Indeed, after proper scaling and choice of n we get the right result:
res = 100; % resolution
w = linspace(0,2*pi,res);
dx = diff(w(1:2)); % grid step
stem( linspace(0,2*pi,res), dx * cconv(sin(w),sin(w),res) );
This is the code I used for the animation:
hF = figure();
subplot(1,2,1); hS(1) = stem(1,cconv(1,1,1)); title('Autoscaling');
subplot(1,2,2); hS(2) = stem(1,cconv(1,1,1)); xlim([0,7]); ylim(50*[-1,1]); title('Constant limits');
w = linspace(0,2*pi,100);
for ind1 = 1:200
set(hS,'XData',linspace(0,2*pi,ind1));
set(hS,'YData',cconv(sin(w),sin(w),ind1));
suptitle("n = " + ind1);
drawnow
% export_fig(char("D:\BLABLA\F" + ind1 + ".png"),'-nocrop');
end

Gaussian iterative curve fitting [duplicate]

I have a set of frequency data with peaks to which I need to fit a Gaussian curve and then get the full width half maximum from. The FWHM part I can do, I already have a code for that but I'm having trouble writing code to fit the Gaussian.
Does anyone know of any functions that'll do this for me or would be able to point me in the right direction? (I can do least squares fitting for lines and polynomials but I can't get it to work for gaussians)
Also it would be helpful if it was compatible with both Octave and Matlab as I have Octave at the moment but don't get access to Matlab until next week.
Any help would be greatly appreciated!
Fitting a single 1D Gaussian directly is a non-linear fitting problem. You'll find ready-made implementations here, or here, or here for 2D, or here (if you have the statistics toolbox) (have you heard of Google? :)
Anyway, there might be a simpler solution. If you know for sure your data y will be well-described by a Gaussian, and is reasonably well-distributed over your entire x-range, you can linearize the problem (these are equations, not statements):
y = 1/(σ·√(2π)) · exp( -½ ( (x-μ)/σ )² )
ln y = ln( 1/(σ·√(2π)) ) - ½ ( (x-μ)/σ )²
= Px² + Qx + R
where the substitutions
P = -1/(2σ²)
Q = +2μ/(2σ²)
R = ln( 1/(σ·√(2π)) ) - ½(μ/σ)²
have been made. Now, solve for the linear system Ax=b with (these are Matlab statements):
% design matrix for least squares fit
xdata = xdata(:);
A = [xdata.^2, xdata, ones(size(xdata))];
% log of your data
b = log(y(:));
% least-squares solution for x
x = A\b;
The vector x you found this way will equal
x == [P Q R]
which you then have to reverse-engineer to find the mean μ and the standard-deviation σ:
mu = -x(2)/x(1)/2;
sigma = sqrt( -1/2/x(1) );
Which you can cross-check with x(3) == R (there should only be small differences).
Perhaps this has the thing you are looking for? Not sure about compatability:
http://www.mathworks.com/matlabcentral/fileexchange/11733-gaussian-curve-fit
From its documentation:
[sigma,mu,A]=mygaussfit(x,y)
[sigma,mu,A]=mygaussfit(x,y,h)
this function is doing fit to the function
y=A * exp( -(x-mu)^2 / (2*sigma^2) )
the fitting is been done by a polyfit
the lan of the data.
h is the threshold which is the fraction
from the maximum y height that the data
is been taken from.
h should be a number between 0-1.
if h have not been taken it is set to be 0.2
as default.
i had similar problem.
this was the first result on google, and some of the scripts linked here made my matlab crash.
finally i found here that matlab has built in fit function, that can fit Gaussians too.
it look like that:
>> v=-30:30;
>> fit(v', exp(-v.^2)', 'gauss1')
ans =
General model Gauss1:
ans(x) = a1*exp(-((x-b1)/c1)^2)
Coefficients (with 95% confidence bounds):
a1 = 1 (1, 1)
b1 = -8.489e-17 (-3.638e-12, 3.638e-12)
c1 = 1 (1, 1)
I found that the MATLAB "fit" function was slow, and used "lsqcurvefit" with an inline Gaussian function. This is for fitting a Gaussian FUNCTION, if you just want to fit data to a Normal distribution, use "normfit."
Check it
% % Generate synthetic data (for example) % % %
nPoints = 200; binSize = 1/nPoints ;
fauxMean = 47 ;fauxStd = 8;
faux = fauxStd.*randn(1,nPoints) + fauxMean; % REPLACE WITH YOUR ACTUAL DATA
xaxis = 1:length(faux) ;fauxData = histc(faux,xaxis);
yourData = fauxData; % replace with your actual distribution
xAxis = 1:length(yourData) ;
gausFun = #(hms,x) hms(1) .* exp (-(x-hms(2)).^2 ./ (2*hms(3)^2)) ; % Gaussian FUNCTION
% % Provide estimates for initial conditions (for lsqcurvefit) % %
height_est = max(fauxData)*rand ; mean_est = fauxMean*rand; std_est=fauxStd*rand;
x0 = [height_est;mean_est; std_est]; % parameters need to be in a single variable
options=optimset('Display','off'); % avoid pesky messages from lsqcurvefit (optional)
[params]=lsqcurvefit(gausFun,x0,xAxis,yourData,[],[],options); % meat and potatoes
lsq_mean = params(2); lsq_std = params(3) ; % what you want
% % % Plot data with fit % % %
myFit = gausFun(params,xAxis);
figure;hold on;plot(xAxis,yourData./sum(yourData),'k');
plot(xAxis,myFit./sum(myFit),'r','linewidth',3) % normalization optional
xlabel('Value');ylabel('Probability');legend('Data','Fit')

numerical integration for Gaussian function - indefinite integral

My approach
fun = #(y) (1/sqrt(pi))*exp(-(y-1).^2).*log(1 + exp(-4*y))
integral(fun,-Inf,Inf)
This gives NaN.
So I tried plotting it.
y= -10:0.1:10;
plot(y,exp(-(y-1).^2).*log(1 + exp(-4*y)))
Then understood that domain (siginificant part) is from -4 to +4.
So changed the limits to
integral(fun,-10,10)
However I do not want to always plot the graph and then know its limits. So is there any way to know the integral directly from -Inf to Inf.
Discussion
If your integrals are always of the form
I would use a high-order Gauss–Hermite quadrature rule.
It's similar to the Gauss-Legendre-Kronrod rule that forms the basis for quadgk but is specifically tailored for integrals over the real line with a standard Gaussian multiplier.
Rewriting your equation with the substitution x = y-1, we get
.
The integral can then be computed using the Gauss-Hermite rule of arbitrary order (within reason):
>> order = 10;
>> [nodes,weights] = GaussHermiteRule(order);
>> f = #(x) log(1 + exp(-4*(x+1)))/sqrt(pi);
>> sum(f(nodes).*weights)
ans =
0.1933
I'd note that the function below builds a full order x order matrix to compute nodes, so it shouldn't be made too large.
There is a way to avoid this by explicitly computing the weights, but I decided to be lazy.
Besides, event at order 100, the Gaussian multiplier is about 2E-98, so the integrand's contribution is extremely minimal.
And while this isn't inherently adaptive, a high-order rule should be sufficient in most cases ... I hope.
Code
function [nodes,weights] = GaussHermiteRule(n)
% ------------------------------------------------------------------------------
% Find the nodes and weights for a Gauss-Hermite Quadrature integration.
%
if (n < 1)
error('There is no Gauss-Hermite rule of order 0.');
elseif (n < 0) || (abs(n - round(n)) > eps())
error('Given order ''n'' must be a strictly positive integer.');
else
n = round(n);
end
% Get the nodes and weights from the Golub-Welsch function
n = (0:n)' ;
b = n*0 ;
a = b + 0.5 ;
c = n ;
[nodes,weights] = GolubWelsch(a,b,c,sqrt(pi));
end
function [xk,wk] = GolubWelsch(ak,bk,ck,mu0)
%GolubWelsch
% Calculate the approximate* nodes and weights (normalized to 1) of an orthogonal
% polynomial family defined by a three-term reccurence relation of the form
% x pk(x) = ak pkp1(x) + bk pk(x) + ck pkm1(x)
%
% The weight scale factor mu0 is the integral of the weight function over the
% orthogonal domain.
%
% Calculate the terms for the orthonormal version of the polynomials
alpha = sqrt(ak(1:end-1) .* ck(2:end));
% Build the symmetric tridiagonal matrix
T = full(spdiags([[alpha;0],bk,[0;alpha]],[-1,0,+1],length(alpha),length(alpha)));
% Calculate the eigenvectors and values of the matrix
[V,xk] = eig(T,'vector');
% Calculate the weights from the eigenvectors - technically, Golub-Welsch requires
% a normalization, but since MATLAB returns unit eigenvectors, it is omitted.
wk = mu0*(V(1,:).^2)';
end
I've had success with transforming such infinite-bounded integrals using a numerical variable transformation, as explained in Numerical Recipes 3e, section 4.5.3. Basically, you substitute in y=c*tan(t)+b and then numerically integrate over t in (-pi/2,pi/2), which sweeps y from -infinity to infinity. You can tune the values of c and b to optimize the process. This approach largely dodges the question of trying to determine cutoffs in the domain, but for this to work reliably using quadrature you have to know that the integrand does not have features far from y=b.
A quick and dirty solution would be to look for a position, where your function is sufficiently small enough and then taking it as limits. This assumes that for x>0 the function fun decreases montonically and fun(x) is roughly the same size as fun(-x) for all x.
%// A small number
epsilon = eps;
%// Stepsize for searching bound
stepTest = 1;
%// Starting position for searching bound
position = 0;
%// Not yet small enough
smallEnough = false;
%// Search bound
while ~smallEnough
smallEnough = (fun(position) < eps);
position = position + stepTest;
end
%// Calculate integral
integral(fun, -position, position)
If your were happy with plotting the function, deciding by eye where you can cut, then this code will suffice, I guess.

simulating a random walk in matlab

i have variable x that undergoes a random walk according to the following rules:
x(t+1)=x(t)-1; probability p=0.3
x(t+1)=x(t)-2; probability q=0.2
x(t+1)=x(t)+1; probability p=0.5
a) i have to create this variable initialized at zero and write a for loop for 100 steps and that runs 10000 times storing each final value in xfinal
b) i have to plot a probability distribution of xfinal (a histogram) choosing a bin size and normalization!!* i have to report the mean and variance of xfinal
c) i have to recreate the distribution by application of the central limit theorem and plot the probability distribution on the same plot!
help would be appreciated in telling me how to choose the bin size and normalize the histogram and how to attempt part c)
your help is much appreciated!!
p=0.3;
q=0.2;
s=0.5;
numberOfSteps = 100;
maxCount = 10000;
for count=1:maxCount
x=0;
for i = 1:numberOfSteps
random = rand(1, 1);
if random <=p
x=x-1;
elseif random<=(p+q)
x=x-2;
else
x=x+1;
end
end
xfinal(count) = x;
end
[f,x]=hist(xfinal,30);
figure(1)
bar(x,f/sum(f));
xlabel('xfinal')
ylabel('frequency')
mean = mean(xfinal)
variance = var(xfinal)
For the first question, check the help for hist on mathworks homepage
[nelements,centers] = hist(data,nbins);
You do not select the bin size, but the number of bins. nelements gives the elements per bin and center is all the bin centers. So to say, it would be the same to call
hist(data,nbins);
as
[nelements,centers] = hist(data,nbins);
plot(centers,nelements);
except that the representation is different (line or pile). To normalize, simply divide nelements with sum(nelements)
For c, here i.i.d. variables it actually is a difference if the variables are real or complex. However for real variables the central limit theorem in short tells you that for a large number of samples the distribution will limit the normal distribution. So if the samples are real, you simply asssumes a normal distribution, calculates the mean and variance and plots this as a normal distribution. If the variables are complex, then each of the variables will be normally distributed which means that you will have a rayleigh distribution instead.
Mathworks is deprecating hist that is being replaced with histogram.
more details in this link
You are not applying the PDF function as expected, the expression Y doesn't work
For instance Y does not have the right X-axis start stop points. And you are using x as input to Y while x already used as pivot inside the double for loop.
When I ran your code Y generates a single value, it is not a vector but just a scalar.
This
bar(x,f/sum(f));
bringing down all input values with sum(f) division? no need.
On attempting to overlap the ideal probability density function, often one has to do additional scaling, to have both real and ideal visually overlapped.
MATLAB can do the scaling for us, and no need to modify input data /sum(f).
With a dual plot using yyaxis
You also mixed variance and standard deviation.
Instead try something like this
y2=1 / sqrt(2*pi*var1)*exp(-(x2-m1).^2 / (2*var1))
ok, the following solves your question(s)
codehere
clear all;
close all;
clc
p=0.3; % thresholds
q=0.2;
s=0.5;
n_step=100;
max_cnt=10000;
n_bin=30; % histogram amount bins
xf=zeros(1,max_cnt);
for cnt=1:max_cnt % runs loop
x=0;
for i = 1:n_step % steps loop
t_rand1 = rand(1, 1);
if t_rand1 <=p
x=x-1;
elseif t_rand1<=(p+q)
x=x-2;
else
x=x+1;
end
end
xf(cnt) = x;
end
% [f,x]=hist(xf,n_bin);
hf1=figure(1)
ax1=gca
yyaxis left
hp1=histogram(xf,n_bin);
% bar(x,f/sum(f));
grid on
xlabel('xf')
ylabel('frequency')
m1 = mean(xf)
var1 = var(xf)
s1=var1^.5 % sigma
%applying central limit theorem %finding the mean
n_x2=1e3 % just enough points
min_x2=min(hp1.BinEdges)
max_x2=max(hp1.BinEdges)
% quite same as
min_x2=hp1.BinLimits(1)
max_x2=hp1.BinLimits(2)
x2=linspace(min_x2,max_x2,n_x2)
y2=1/sqrt(2*pi*var1)*exp(-(x2-m1).^2/(2*var1));
% hold(ax1,'on')
yyaxis right
plot(ax1,x2,y2,'r','LineWidth',2)
.
.
.
note I have not used these lines
% Xp=-1; Xq=-2; Xs=1; mu=Xp.*p+Xq.*q+Xs.*s;
% muN=n_step.*mu;
%
% sigma=(Xp).^2.*p+(Xq).^2.*q+(Xs).^2.s; % variance
% sigmaN=n_step.(sigma-(mu).^2);
People ususally call sigma to variance^.5
This supplied script is a good start point to now take it to wherever you need it to go.