Approximating Definite Integrals as Sums
11 views (last 30 days)
Show older comments
I have a function
that I want to integrate over
and
. I know the answer of this integration is
. When I use
function, I get the correct result but when I try to evaluate this integral using
loops, my result is always
. I increased points to
but still get the same result. By using
loops for integration. The sum approximation is
. I am following a research paper which claims that they used
loops with gaussian meshes and get
as answer.
My question is why exactly I am not getting the correct result and what exactly are gaussian meshes? The function
and the code runner.m that I am using are given below:
runner.m
%runner.m
clear; clc;
NBZ = 500; %number of x and y points. total points = NBZ^2
JNN = 0; %a parameter
a = 1;
% x and y limits... and x and y points.
xmin = -2*pi/(3*a);
xmax = 4*pi/(3*a);
ymin = -2*pi/(sqrt(3)*a);
ymax = 2*pi/(sqrt(3)*a);
dx = (xmax-xmin)/(NBZ-1); %Delta x
dy = dx; %Delta y
xs = xmin:dx:xmax; %array of x points
ys = ymin:dy:ymax; %array of y points
dsum= 0;
for ny = 1:NBZ
y = ys(ny);
for nx = 1:NBZ
x = xs(nx);
out = F(x,y,JNN);
dsum = dsum + out*dx*dy;
end
end
answer = dsum
%it gives: answer = -0.6778
F(x,y)
function out = F(x,y,JNN)
JN = 4;
D = 1;
s = 1;
h11 = 0;
h12 = -(JN+1i*D)*s*(1+exp(1i*(-x-sqrt(3)*y)/2))...
-JNN*s*(exp(1i*x)+exp(1i*(+x-sqrt(3)*y)/2));
h13 = -(JN-1i*D)*s*(1+exp(1i*(+x-sqrt(3)*y)/2))...
-JNN*s*(exp(1i*x)+exp(1i*(-x-sqrt(3)*y)/2));
h23 = -(JN+1i*D)*s*(1+exp(1i*x))-2*JNN*s*exp(1i*x/2)*cos(sqrt(3)/2*y);
h = [h11 h12 h13;
conj(h12) h11 h23;
conj(h13) conj(h23) h11];
[evecs, evals] = eig(h);
v1 = evecs(:,1);
v2 = evecs(:,2);
v3 = evecs(:,3);
e1 = evals(1,1);
e2 = evals(2,2);
e3 = evals(3,3);
X = [ 0, - exp(- (x*1i)/2 - (3^(1/2)*y*1i)/2)*(1/2 - 2i) - JNN*(exp(x*1i)*1i + (exp((x*1i)/2 - (3^(1/2)*y*1i)/2)*1i)/2), - exp((x*1i)/2 - (3^(1/2)*y*1i)/2)*(1/2 + 2i) - JNN*(exp(x*1i)*1i - (exp(- (x*1i)/2 - (3^(1/2)*y*1i)/2)*1i)/2)
- exp((conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*(1/2 + 2i) + JNN*((exp(- (conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*1i)/2 + exp(-conj(x)*1i)*1i), 0, exp(x*1i)*(1 - 4i) - JNN*exp((x*1i)/2)*cos((3^(1/2)*y)/2)*1i
- exp(- (conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*(1/2 - 2i) - JNN*((exp((conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*1i)/2 - exp(-conj(x)*1i)*1i), exp(-conj(x)*1i)*(1 + 4i) + JNN*exp(-(conj(x)*1i)/2)*cos((3^(1/2)*conj(y))/2)*1i, 0];
Y = [ 0, - 3^(1/2)*exp(- (x*1i)/2 - (3^(1/2)*y*1i)/2)*(1/2 - 2i) + (3^(1/2)*JNN*exp((x*1i)/2 - (3^(1/2)*y*1i)/2)*1i)/2, 3^(1/2)*exp((x*1i)/2 - (3^(1/2)*y*1i)/2)*(1/2 + 2i) + (3^(1/2)*JNN*exp(- (x*1i)/2 - (3^(1/2)*y*1i)/2)*1i)/2
- 3^(1/2)*exp((conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*(1/2 + 2i) - (3^(1/2)*JNN*exp(- (conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*1i)/2, 0, 3^(1/2)*JNN*exp((x*1i)/2)*sin((3^(1/2)*y)/2)
3^(1/2)*exp(- (conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*(1/2 - 2i) - (3^(1/2)*JNN*exp((conj(x)*1i)/2 + (3^(1/2)*conj(y)*1i)/2)*1i)/2, 3^(1/2)*JNN*sin((3^(1/2)*conj(y))/2)*exp(-(conj(x)*1i)/2), 0];
o1a = ( ((v2'*X*v1)*(v1'*Y*v2)) - ((v2'*Y*v1)*(v1'*X*v2)))/((e2-e1)^2);
o1b = ( ((v3'*X*v1)*(v1'*Y*v3)) - ((v3'*Y*v1)*(v1'*X*v3)))/((e3-e1)^2);
o1 = 1i*(o1a+o1b);
out = real(o1/(2*pi));
end
0 Comments
Accepted Answer
Matt J
on 2 Nov 2020
Edited: Matt J
on 2 Nov 2020
I don't know what gaussian meshes are, but the Riemann sum works with a just few fixes to the way the sampling is set up:
NBZ = 600; %number of x and y points. total points = NBZ^2
JNN = 0; %a parameter
a = 1;
% x and y limits... and x and y points.
xmin = -2*pi/(3*a);
xmax = 4*pi/(3*a);
ymin = -2*pi/(sqrt(3)*a);
ymax = 2*pi/(sqrt(3)*a);
xs = linspace(xmin,xmax,NBZ+1); %array of x points
ys = linspace(ymin,ymax,NBZ+1); %array of y points
xs(end)=[]; ys(end)=[];
dx=xs(2)-xs(1);
dy=ys(2)-ys(1);
dsum= 0;
for ny = 1:NBZ
y = ys(ny);
for nx = 1:NBZ
x = xs(nx);
out = F(x,y,JNN);
dsum = dsum + out;
end
end
answer = dsum*dx*dy
More Answers (0)
See Also
Categories
Find more on Linear Algebra in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!