I have indices
I = [nGrid x 9] matrix % mesh on fine grid (9 point rectangle)
J = [nGrid x 4] matrix % mesh on coarse grid (4 point rectangle)
Here, nGrid is large number depending on the mesh (e.g. 1.e05)
Then I want to do like
R_ref = [4 x 9] matrix % reference restriction matrix from fine to coarse
P_ref = [9 x 4] matrix % reference prolongation matrix from coarse to fine
R = sparse(size) % n_Coarse x n_Fine
P = sparse(size) % n_Fine x n_Coarse
for k = 1 : nGrid % number of elements on coarse grid
R(I(k,:),J(k,:)) = R_ref;
P(J(k,:),I(k,:)) = P_ref;
end
size is predetermined number.
Note that even if there is same index in (I,J), I do not want to accumulate. I just want to put stencils Rref and Pref at each indices respectively.
I know that this is very slow due to the data structure of sparse.
Usually, I use
sparse(row,col,entry,n_row,n_col)
I can use this by manipulate I, J, R_ref, P_ref by bsxfun and repmat.
However, this cannot be done because of the accumulation of sparse function
(if there exists (i,j) such that [row(i),col(i)]==[row(j),col(j)], then R(row(i),row(j)) = entry(i)+entry(j))
Is there any suggestion for this kind of assembly procedure?
Sample code
%% INPUTS
% N and M could be much larger
N = 2^5+1; % number of fine grid in x direction
M = 2^5+1; % number of fine grid in y direction
% [nOx * nOy] == nGrid
nOx = floor((M)/2)+1; % number of coarse grid on x direction
nOy = floor((N)/2)+1; % number of coarse grid on y direction
Rref = [4 4 -1 4 -2 0 -1 0 0
-1 -1 -2 4 4 4 -1 4 4
-1 -1 4 -2 4 -2 4 4 -1
0 4 4 0 0 4 -1 -2 -1]/8;
Pref = [2 1 0 1 0 0 0 0 0
0 0 0 1 1 1 0 1 2
0 0 1 0 1 0 2 1 0
0 2 1 0 0 1 0 0 0]'/2;
%% INDEX GENERATION
tri_ref = reshape(bsxfun(#plus,[0,1,2]',[0,N,2*N]),[],1)';
TRI_ref = reshape(bsxfun(#plus,[0,1]',[0,nOy]),[],1)';
I = reshape(bsxfun(#plus,(1:2:N-2)',0:2*N:(M-2)*N),[],1);
J = reshape(bsxfun(#plus,(1:nOy-1)',0:nOy:(nOx-2)*nOy),[],1);
I = bsxfun(#plus,I,tri_ref);
J = bsxfun(#plus,J,TRI_ref);
%% THIS PART IS WHAT I WANT TO CHANGE
R = sparse(nOx*nOy,N*M);
P = R';
for k = 1 : size(I,1)
R(J(k,:),I(k,:)) = Rref;
P(I(k,:),J(k,:)) = Pref;
end
Related
How to obtain the coordinates of the first and the last appearances (under column-major ordering) of each label present in a matrix?
Example of a label matrix (where labels are 1 to 4):
L = [
1 1 1 1 0 0 0 0
0 0 0 0 2 2 0 0
0 0 0 0 0 0 2 0
0 0 0 0 0 0 0 0
0 0 0 0 0 3 0 0
0 0 0 0 0 0 3 3
0 0 0 4 0 0 0 0
4 4 4 0 0 0 0 0
];
For the above example L, I would like to obtain a matrix of coordinates like:
M = [
1 1 1
1 4 1
2 5 2
3 7 2
5 6 3
6 8 3
8 1 4
7 4 4 ];
Where the 1st column of M contains horizontal coordinates, the 2nd contains vertical coordinates, and the 3rd column contains the label. There should be 2 rows for each label.
With for-loop you can do it like that:
M=zeros(2*max(L(:)),3);
for k=1:max(L(:))
[r,c]=find(L==k);
s=sortrows([r c],2);
M(k*2-1:k*2,:)=[s(1,:) k; s(end,:) k];
end
M =
1 1 1
1 4 1
2 5 2
3 7 2
5 6 3
6 8 3
8 1 4
7 4 4
Maybe somehow with regionprops options you can do it without the loop...
I just had to try it with accumarray:
R = size(L, 1);
[rowIndex, colIndex, values] = find(L); % Find nonzero values
index = (colIndex-1).*R+rowIndex; % Create a linear index
labels = unique(values); % Find unique values
nLabels = numel(labels);
minmax = zeros(2, nLabels);
minmax(1, :) = accumarray(values, index, [nLabels 1], #min); % Collect minima
minmax(2, :) = accumarray(values, index, [nLabels 1], #max); % Collect maxima
temp = ceil(minmax(:)/R);
M = [minmax(:)-R.*(temp-1) temp repelem(labels, 2, 1)]; % Convert index to subscripts
M =
1 1 1
1 4 1
2 5 2
3 7 2
5 6 3
6 8 3
8 1 4
7 4 4
Here's what I got for timing with Dev-iL's script and Adiel's newest code (Note that the number of labels can't go above 127 due to how Adiel's code uses the uint8 values as indices):
| Adiel | Dev-iL | gnovice
-----------------------+---------+---------+---------
20 labels, 1000x1000 | 0.0753 | 0.0991 | 0.0889
20 labels, 10000x10000 | 12.0010 | 10.2207 | 8.7034
120 labels, 1000x1000 | 0.1924 | 0.3439 | 0.1387
So, for moderate numbers of labels and (relatively) smaller sizes, Adiel's looping solution looks like it does best, with my solution lying between his and Dev-iL's. For larger sizes or greater numbers of labels, my solution starts to take the lead.
If you're looking for a vectorized solution, you can do this:
nTags = max(L(:));
whois = bsxfun(#eq,L,reshape(1:nTags,1,1,[]));
% whois = L == reshape(1:nTags,1,1,[]); % >=R2016b syntax.
[X,Y,Z] = ind2sub(size(whois), find(whois));
tmp = find(diff([0; Z; nTags+1])); tmp = reshape([tmp(1:end-1) tmp(2:end)-1].',[],1);
M = [X(tmp), Y(tmp), repelem(1:nTags,2).'];
Or with extreme variable reuse:
nTags = max(L(:));
Z = bsxfun(#eq,L,reshape(1:nTags,1,1,[]));
[X,Y,Z] = ind2sub(size(Z), find(Z));
Z = find(diff([0; Z; nTags+1]));
Z = reshape([Z(1:end-1) Z(2:end)-1].',[],1);
M = [X(Z), Y(Z), repelem(1:nTags,2).'];
Here's my benchmarking code:
function varargout = b42973322(isGPU,nLabels,lMat)
if nargin < 3
lMat = 1000;
end
if nargin < 2
nLabels = 20; % if nLabels > intmax('uint8'), Change the type of L to some other uint.
end
if nargin < 1
isGPU = false;
end
%% Create L:
if isGPU
L = sort(gpuArray.randi(nLabels,lMat,lMat,'uint8'),2);
else
L = sort(randi(nLabels,lMat,lMat,'uint8'),2);
end
%% Equality test:
M{3} = DeviL2(L);
M{2} = DeviL1(L);
M{1} = Adiel(L);
assert(isequal(M{1},M{2},M{3}));
%% Timing:
% t(3) = timeit(#()DeviL2(L)); % This is always slower, so it's irrelevant.
t(2) = timeit(#()DeviL1(L));
t(1) = timeit(#()Adiel(L));
%% Output / Print
if nargout == 0
disp(t);
else
varargout{1} = t;
end
end
function M = Adiel(L)
M=[];
for k=1:max(L(:))
[r,c]=find(L==k);
s=sortrows([r c],2);
M=[M;s(1,:) k; s(end,:) k];
end
end
function M = DeviL1(L)
nTags = max(L(:));
whois = L == reshape(1:nTags,1,1,[]); % >=R2016b syntax.
[X,Y,Z] = ind2sub(size(whois), find(whois));
tmp = find(diff([0; Z; nTags+1])); tmp = reshape([tmp(1:end-1) tmp(2:end)-1].',[],1);
M = [X(tmp), Y(tmp), repelem(1:nTags,2).'];
end
function M = DeviL2(L)
nTags = max(L(:));
Z = L == reshape(1:nTags,1,1,[]);
[X,Y,Z] = ind2sub(size(Z), find(Z));
Z = find(diff([0; Z; nTags+1]));
Z = reshape([Z(1:end-1) Z(2:end)-1].',[],1);
M = [X(Z), Y(Z), repelem(1:nTags,2).'];
end
You can retrive the uniqe values (your labels) of the matrix with unique.
Having them retrived you can use find to get their indices.
Put together your matrix with it.
I would like to write a "syntactical sugar" Octave or Matlab zero-padding function, to which the user sends an n-dimensional object and a vector of <= n entries. The vector contains new, equal or larger dimensions for the object, and the object is zero-padded to match these dimensions. Any dimensions not specified are left alone. One expected use is, given for example a 5d block X of 3d medical image volumes, I can call
y = simplepad(X, [128 128 128]);
and thus pad the first three dimensions to a power of two for wavelet analysis (in fact I use a separate function nextpwr2 to find these dimensions) while leaving the others.
I have racked my brains on how to write this method avoiding the dreaded eval, but cannot thus far find a way. Can anyone suggest a solution? Here is more or less what I have:
function y = simplepad(x, pad)
szx = size(x);
n_pad = numel(pad);
szy = [pad szx(n_pad+1:end)];
y = zeros(szy);
indices_string = '(';
for n = 1:numel(szx)
indices_string = [indices_string, '1:', num2str(szx(n))];
if n < numel(szx)
indices_string = [indices_string, ','];
else
indices_string = [indices_string, ')'];
end
end
command = ['y',indices_string,'=x;'];
eval(command);
end
Here's a solution that should handle all the little corner cases:
function A = simplepad(A, pad)
% Add singleton dimensions (i.e. ones) to the ends of the old size of A
% or pad as needed so they can be compared directly to one another:
oldSize = size(A);
dimChange = numel(pad)-numel(oldSize);
oldSize = [oldSize ones(1, dimChange)];
pad = [pad ones(1, -dimChange)];
% If all of the sizes in pad are less than or equal to the sizes in
% oldSize, there is no padding done:
if all(pad <= oldSize)
return
end
% Use implicit zero expansion to pad:
pad = num2cell(pad);
A(pad{:}) = 0;
end
And a few test cases:
>> M = magic(3)
M =
8 1 6
3 5 7
4 9 2
>> simplepad(M, [1 1]) % No change, since the all values are smaller
ans =
8 1 6
3 5 7
4 9 2
>> simplepad(M, [1 4]) % Ignore the 1, pad the rows
ans =
8 1 6 0
3 5 7 0
4 9 2 0
>> simplepad(M, [4 4]) % Pad rows and columns
ans =
8 1 6 0
3 5 7 0
4 9 2 0
0 0 0 0
>> simplepad(M, [4 4 2]) % Pad rows and columns and add a third dimension
ans(:,:,1) =
8 1 6 0
3 5 7 0
4 9 2 0
0 0 0 0
ans(:,:,2) =
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
As I understand, you want just pass the some dynamic arguments to function.
You can do this by converting these arguments to cell and call your function with passing cell content. So, your function will look like:
function y = simplepad(x, pad)
szx = size(x);
n_pad = numel(pad);
szy = [pad szx(n_pad+1:end)];
y = x;
szyc = num2cell(szy);
y(szyc{:}) = 0; % warning: assume x array only grows
end
Consider a set of points arranged on a grid of size N-by-M.
I am trying to build the adjacency matrix such that
neighboring points are connected.
For example, in a 3x3 grid with a graph:
1-2-3
| | |
4-5-6
| | |
7-8-9
We should have the corresponding adjacency matrix:
+---+------------------------------------------------------+
| | 1 2 3 4 5 6 7 8 9 |
+---+------------------------------------------------------+
| 1 | 0 1 0 1 0 0 0 0 0 |
| 2 | 1 0 1 0 1 0 0 0 0 |
| 3 | 0 1 0 0 0 1 0 0 0 |
| 4 | 1 0 0 0 1 0 1 0 0 |
| 5 | 0 1 0 1 0 1 0 1 0 |
| 6 | 0 0 1 0 1 0 0 0 1 |
| 7 | 0 0 0 1 0 0 0 1 0 |
| 8 | 0 0 0 0 1 0 1 0 1 |
| 9 | 0 0 0 0 0 1 0 1 0 |
+---+------------------------------------------------------+
As a bonus, the solution should work for both 4- and 8-connected neighboring points, that is:
o o o o
o X o vs. o X o
o o o o
This the code that I have so far:
N = 3; M = 3;
adj = zeros(N*M);
for i=1:N
for j=1:M
k = sub2ind([N M],i,j);
if i>1
ii=i-1; jj=j;
adj(k,sub2ind([N M],ii,jj)) = 1;
end
if i<N
ii=i+1; jj=j;
adj(k,sub2ind([N M],ii,jj)) = 1;
end
if j>1
ii=i; jj=j-1;
adj(k,sub2ind([N M],ii,jj)) = 1;
end
if j<M
ii=i; jj=j+1;
adj(k,sub2ind([N M],ii,jj)) = 1;
end
end
end
How can this improved to avoid all the looping?
If you notice, there is a distinct pattern to the adjacency matrices you are creating. Specifically, they are symmetric and banded. You can take advantage of this fact to easily create your matrices using the diag function (or the spdiags function if you want to make a sparse matrix). Here is how you can create the adjacency matrix for each case, using your sample matrix above as an example:
4-connected neighbors:
mat = [1 2 3; 4 5 6; 7 8 9]; % Sample matrix
[r, c] = size(mat); % Get the matrix size
diagVec1 = repmat([ones(c-1, 1); 0], r, 1); % Make the first diagonal vector
% (for horizontal connections)
diagVec1 = diagVec1(1:end-1); % Remove the last value
diagVec2 = ones(c*(r-1), 1); % Make the second diagonal vector
% (for vertical connections)
adj = diag(diagVec1, 1)+diag(diagVec2, c); % Add the diagonals to a zero matrix
adj = adj+adj.'; % Add the matrix to a transposed copy of
% itself to make it symmetric
And you'll get the following matrix:
adj =
0 1 0 1 0 0 0 0 0
1 0 1 0 1 0 0 0 0
0 1 0 0 0 1 0 0 0
1 0 0 0 1 0 1 0 0
0 1 0 1 0 1 0 1 0
0 0 1 0 1 0 0 0 1
0 0 0 1 0 0 0 1 0
0 0 0 0 1 0 1 0 1
0 0 0 0 0 1 0 1 0
8-connected neighbors:
mat = [1 2 3; 4 5 6; 7 8 9]; % Sample matrix
[r, c] = size(mat); % Get the matrix size
diagVec1 = repmat([ones(c-1, 1); 0], r, 1); % Make the first diagonal vector
% (for horizontal connections)
diagVec1 = diagVec1(1:end-1); % Remove the last value
diagVec2 = [0; diagVec1(1:(c*(r-1)))]; % Make the second diagonal vector
% (for anti-diagonal connections)
diagVec3 = ones(c*(r-1), 1); % Make the third diagonal vector
% (for vertical connections)
diagVec4 = diagVec2(2:end-1); % Make the fourth diagonal vector
% (for diagonal connections)
adj = diag(diagVec1, 1)+... % Add the diagonals to a zero matrix
diag(diagVec2, c-1)+...
diag(diagVec3, c)+...
diag(diagVec4, c+1);
adj = adj+adj.'; % Add the matrix to a transposed copy of
% itself to make it symmetric
And you'll get the following matrix:
adj =
0 1 0 1 1 0 0 0 0
1 0 1 1 1 1 0 0 0
0 1 0 0 1 1 0 0 0
1 1 0 0 1 0 1 1 0
1 1 1 1 0 1 1 1 1
0 1 1 0 1 0 0 1 1
0 0 0 1 1 0 0 1 0
0 0 0 1 1 1 1 0 1
0 0 0 0 1 1 0 1 0
Just for fun, here's a solution to construct the adjacency matrix by computing the distance between all pairs of points on the grid (not the most efficient way obviously)
N = 3; M = 3; %# grid size
CONNECTED = 8; %# 4-/8- connected points
%# which distance function
if CONNECTED == 4, distFunc = 'cityblock';
elseif CONNECTED == 8, distFunc = 'chebychev'; end
%# compute adjacency matrix
[X Y] = meshgrid(1:N,1:M);
X = X(:); Y = Y(:);
adj = squareform( pdist([X Y], distFunc) == 1 );
And here's some code to visualize the adjacency matrix and the graph of connected points:
%# plot adjacency matrix
subplot(121), spy(adj)
%# plot connected points on grid
[xx yy] = gplot(adj, [X Y]);
subplot(122), plot(xx, yy, 'ks-', 'MarkerFaceColor','r')
axis([0 N+1 0 M+1])
%# add labels
[X Y] = meshgrid(1:N,1:M);
X = reshape(X',[],1) + 0.1; Y = reshape(Y',[],1) + 0.1;
text(X, Y(end:-1:1), cellstr(num2str((1:N*M)')) )
I just found this question when searching for the same problem. However, none of the provided solutions worked for me because of the problem size which required the use of sparse matrix types. Here is my solution which works on large scale instances:
function W = getAdjacencyMatrix(I)
[m, n] = size(I);
I_size = m*n;
% 1-off diagonal elements
V = repmat([ones(m-1,1); 0],n, 1);
V = V(1:end-1); % remove last zero
% n-off diagonal elements
U = ones(m*(n-1), 1);
% get the upper triangular part of the matrix
W = sparse(1:(I_size-1), 2:I_size, V, I_size, I_size)...
+ sparse(1:(I_size-m),(m+1):I_size, U, I_size, I_size);
% finally make W symmetric
W = W + W';
Just came across this question. I have a nice working m-function (link: sparse_adj_matrix.m) that is quite general.
It can handle 4-connect grid (radius 1 according to L1 norm), 8-connect grid (radius 1 according to L_infty norm).
It can also support 3D (and arbitrarily higher domensional grids).
The function can also connect nodes further than radius = 1.
Here's the signiture of the function:
% Construct sparse adjacency matrix (provides ii and jj indices into the
% matrix)
%
% Usage:
% [ii jj] = sparse_adj_matrix(sz, r, p)
%
% inputs:
% sz - grid size (determine the number of variables n=prod(sz), and the
% geometry/dimensionality)
% r - the radius around each point for which edges are formed
% p - in what p-norm to measure the r-ball, can be 1,2 or 'inf'
%
% outputs
% ii, jj - linear indices into adjacency matrix (for each pair (m,n)
% there is also the pair (n,m))
%
% How to construct the adjacency matrix?
% >> A = sparse(ii, jj, ones(1,numel(ii)), prod(sz), prod(sz));
%
%
% Example:
% >> [ii jj] = sparse_adj_matrix([10 20], 1, inf);
% construct indices for 200x200 adjacency matrix for 8-connect graph over a
% grid of 10x20 nodes.
% To visualize the graph:
% >> [r c]=ndgrid(1:10,1:20);
% >> A = sparse(ii, jj, 1, 200, 200);;
% >> gplot(A, [r(:) c(:)]);
Your current code doesn't seem so bad. One way or another you need to iterate over all neighbor pairs. If you really need to optimize the code, I would suggest:
loop over node indices i, where 1 <= i <= (N*M)
don't use sub2ind() for efficiency, the neighbors of node i are simpy [i-M, i+1, i+M, i-1] in clockwise order
Notice that to get all neighbor pairs of nodes:
you only have to compute the "right" neighbors (i.e. horizontal edges) for nodes i % M != 0 (since Matlab isn't 0-based but 1-based)
you only have to compute "above" neighbors (i.e. vertical edges) for nodes i > M
there is a similar rule for diagonal edges
This would leed to a single loop (but same number of N*M iterations), doesn't call sub2ind(), and has only two if statements in the loop.
For each node in the graph add a connection to the right and one downwards. Check that you don't overreach your grid. Consider the following function that builds the adjacency matrix.
function adj = AdjMatrixLattice4( N, M )
% Size of adjacency matrix
MN = M*N;
adj = zeros(MN,MN);
% number nodes as such
% [1]---[2]-- .. --[M]
% | | |
% [M+1]-[M+2]- .. -[2*M]
% : : :
% [] [] .. [M*N]
for i=1:N
for j=1:N
A = M*(i-1)+j; %Node # for (i,j) node
if(j<N)
B = M*(i-1)+j+1; %Node # for node to the right
adj(A,B) = 1;
adj(B,A) = 1;
end
if(i<M)
B = M*i+j; %Node # for node below
adj(A,B) = 1;
adj(B,A) = 1;
end
end
end
end
Example as above AdjMatrixLattice4(3,3)=
0 1 0 1 0 0 0 0 0
1 0 1 0 1 0 0 0 0
0 1 0 0 0 1 0 0 0
1 0 0 0 1 0 1 0 0
0 1 0 1 0 1 0 1 0
0 0 1 0 1 0 0 0 1
0 0 0 1 0 0 0 1 0
0 0 0 0 1 0 1 0 1
0 0 0 0 0 1 0 1 0
I am trying to use the convhull function in a loop and for that I need to split matrices into submatrices of different sizes. Here is the code I am using:
x1=data(:,5); % x centre location
y1=data(:,16); % y centre location
z1=phi*90; % phi angle value
n=300;
%Create regular grid across data space
[X,Y] = meshgrid(linspace(min(x1),max(x1),n), linspace(min(y1),max(y1),n));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% PLOT USING SCATTER - TRYING TO ISOLATE SOME REGIONS %%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
c=z1>10 & z1 < 20;
c=c.*1;
j=1;
for i=1:length(z1)
if z1(i)< 20 && z1(i)> 10
c(i) = 1;
else
c(i)= 0;
end
end
C=[c c c];
C = ~C;
elementalLengthA = cellfun('length',regexp(sprintf('%i',all(C,2)),'1+','match'));
elementalStartA = regexp(sprintf('%i',all(C,2)),'1+','start');
result = cell(length(elementalLengthA),1);
for i = 1:length(elementalLengthA)
result(i) = {C(elementalStartA(i):elementalStartA(i)+elementalLengthA(i)-1,:)};
length(x1(i))=length(cell2mat(result(i)));
length(y1(i))=length(cell2mat(result(i)));
end
My for loop doens't work properly and I get this error: ??? Subscript indices must either be real positive integers or
logicals.
My matrix C is an nx3 matrix made of lines of 1 and 0. With the result(i) line I am splitting the C matrix into submatrices of 1. Let's say
c = [1 1 1;
0 0 0;
0 0 0;
1 1 1;
1 1 1;
1 1 1;
0 0 0;
1 1 1;
1 1 1;]
Then
>> cell2mat(result(1))
ans =
1 1 1
>> cell2mat(result(2))
ans =
1 1 1
1 1 1
1 1 1
>> cell2mat(result(3))
ans =
1 1 1
1 1 1
Now x1 and y1 are two vector column nx1. And I want to split them according to the length of C submatrices. so length(x1(1)) should be 1, length(x1(2))=3, length(x1(3))=2 and same for the y vector.
Is it possible to do that?
EDIT:
Just to make it more clear
For instance
x1 =
1
2
3
4
5
6
7
8
9
and
y1 =
2
4
6
8
10
12
14
16
18
I want to get this as an output:
x1(1)=[1], x1(2)=[4 5 6]' and x1(3)=[8 9]'
y1(1)=[2], y1(2)[8 10 12]' and y1(3)=[16 18]'
Thanks
Dorian
I am new to MATLAB and I want to formulate the following lease square expression in Matlab. I have some codes that I am typing here. But the optimization problem solution seems not to be correct. Does anyone has an idea why?
First, I want to solve the heat equation
$$T_t(x,t) = - L_x . T(x,t) + F(x,t)$$
where L_x is Laplacian matrix of the graph.
then find y from the following least square.
$$ \min_y \sum_{j} \sum_{i} (\hat{T}_j(t_i) - T_j(t_i, y))^2$$
Thanks in advance!!
Here is my code:
%++++++++++++++++ main ++++++++++++++++++++
% incidence matrix for original graph
C_hat = [ 1 -1 0 0 0 0;...
0 1 -1 0 0 -1;...
0 0 0 0 -1 1;...
0 0 0 1 1 0;...
-1 0 1 -1 0 0];
% initial temperature for each vertex in original graph
T_hat_0 = [0 7 1 9 4];
[M_bar,n,m_bar,T_hat_heat,T_hat_temp] = simulate_temp(T_hat_0,C_hat);
C = [ 1 1 -1 -1 0 0 0 0 0 0;...
0 -1 0 0 1 -1 1 0 0 0;...
0 0 1 0 0 1 0 -1 -1 0;...
0 0 0 1 0 0 -1 0 1 -1;...
-1 0 0 0 -1 0 0 1 0 1];
%
% initial temperature for each vertex in original graph
T_0 = [0 7 1 9 4];
%
% initial temperature simulation
[l,n,m,T_heat,T_temp] = simulate_temp(T_0,C);
%
% bounds for variables
lb = zeros(m,1);
ub = ones(m,1);
%
% initial edge weights
w0 = ones(m,1);
% optimization problem
% w = fmincon(#fun, w0, [], [], [], [], lb, ub);
%++++++++++++++++++++ function++++++++++++++++++++++++++++
function [i,n,m,T_heat,T_temp] = simulate_temp(T,C)
%
% initial conditions
delta_t = 0.1;
M = 20; %% number of time steps
t = 1;
[n,m] = size(C);
I = eye(n);
L_w = C * C';
T_ini = T';
Temp = zeros(n,1);
% Computing Temperature
%
for i=1:M
K = 2*I + L_w * delta_t;
H = 2*I - L_w * delta_t;
%
if i == 1
T_heat = (K \ H) * T_ini;
%
t = t + delta_t;
else
T_heat = (K \ H) * Temp;
%
t = t + delta_t;
end
% replacing column of T_final with each node temperature in each
% iteration. It adds one column to the matrix in each step
T_temp(:,i) = T_heat;
%
Temp = T_heat;
end
end
%++++++++++++++++++ function+++++++++++++++++++++++++++++++++++++++++
function w_i = fun(w);
%
for r=1:n
for s=1:M_bar
w_i = (T_hat_temp(r,s) - T_temp(r,s)).^2;
end
end
To give a more clear answer, I need more information about what form you have the functions F_j and E_j in.
I've assumed that you feed each F_j a value, x_i, and get back a number. I've also assumed that you feed E_j a value x_i, and another value (or vector) y, and get back a value.
I've also assumed that by 'i' and 'j' you mean the indices of the columns and rows respectively, and that they're finite.
All I can suggest without knowing more info is to do this:
Pre-calculate the values of the functions F_j for each x_i, to give a matrix F - where element F(i,j) gives you the value F_j(x_i).
Do the same thing for E_j, giving a matrix E - where E(i,j) corresponds to E_j(x_i,y).
Perform (F-E).^2 to subtract each element of F and E, then square them element-wise.
Take sum( (F-E).^2**, 2)**. sum(M,2) will sum across index i of matrix M, returning a column vector.
Finally, take sum( sum( (F-E).^2, 2), 1) to sum across index j, the columns, this will finally give you a scalar.