I need to create all possible permutation matrices for a matrix where every permutation matrix contains only one 1 in each column and each row, and 0 in all other places.
For example, below example in (1) is all possible permutation matrices for 2x2 matrix and in (2) is a all possible permutation matrices for 3x3 matrix and so on
So how can I get these matrices of a matrix NxN in MATLAB and store them into one three-dimensional matrix?
Here's my solution, using implicit expansion (tested with Octave 5.2.0 and MATLAB Online):
n = 3;
% Get all permutations of length n
p = perms(1:n);
% Number of permutations
n_p = size(p, 1);
% Set up indices, where to set elements to 1
p = p + (0:n:n^2-1) + (0:n^2:n^2*n_p-1).';
% Set up indices, where to set elements to 1 (for MATLAB R2016a and before)
%p = bsxfun(#plus, bsxfun(#plus, p, (0:n:n^2-1)), (0:n^2:n^2*n_p-1).');
% Initialize 3-dimensional matrix
a = zeros(n, n, n_p);
% Set proper elements to 1
a(p) = 1
The output for n = 3:
a =
ans(:,:,1) =
0 0 1
0 1 0
1 0 0
ans(:,:,2) =
0 1 0
0 0 1
1 0 0
ans(:,:,3) =
0 0 1
1 0 0
0 1 0
ans(:,:,4) =
0 1 0
1 0 0
0 0 1
ans(:,:,5) =
1 0 0
0 0 1
0 1 0
ans(:,:,6) =
1 0 0
0 1 0
0 0 1
Using repelem, perms and reshape:
n = 3; % matrix size
f = factorial(n); % number of permutation
rep = repelem(eye(n),1,1,f) % repeat n! time the diagonal matrix
res = reshape(rep(:,perms(1:n).'),n,n,f) % indexing and reshaping
Where res is:
res =
ans(:,:,1) =
0 0 1
0 1 0
1 0 0
ans(:,:,2) =
0 1 0
0 0 1
1 0 0
ans(:,:,3) =
0 0 1
1 0 0
0 1 0
ans(:,:,4) =
0 1 0
1 0 0
0 0 1
ans(:,:,5) =
1 0 0
0 0 1
0 1 0
ans(:,:,6) =
1 0 0
0 1 0
0 0 1
And according to your comment:
What I need to do is to multiply a matrix i.e Z with all possible
permutation matrices and choose that permutation matrix which
resulting a tr(Y) minimum; where Y is the results of multiplication of
Z with the permutation matrix. I Think I don't need to generate all
permutation matrices and store them in such variable, I can generate
them one by one and get the result of multiplication. Is that possible
?
You're trying to solve the assignment problem, you can use the well known hungarian algorithm to solve this task in polynomial time. No needs to generate a googleplex of permutation matrix.
Related
I am trying to create a neighbourhood graph from a given binary matrix B. Neighbourhood graph (A) is defined as an adjacency matrix such that
(A(i,j) = A(j,i) = 1)
if the original matrix B(i) = B(j) = 1 and i and j are adjacent to each (left, right, up, down or diagonal). Here I used the linear subscript to access the original matrix B. For example, consider the below matrix
B = [ 0 1 0;
0 1 1;
0 0 0 ];
My A will be a 9 * 9 graph as given below
A = [ 0 0 0 0 0 0 0 0 0;
0 0 0 0 0 0 0 0 0;
0 0 0 0 0 0 0 0 0;
0 0 0 0 1 0 0 1 0;
0 0 0 1 0 0 0 1 0;
0 0 0 0 0 0 0 0 0;
0 0 0 0 0 0 0 0 0;
0 0 0 1 1 0 0 0 0;
0 0 0 0 0 0 0 0 0 ];
Since in the original B matrix, B(4), B(5) and B(8) are adjacent with corresponding entries 1, the adjacency matrix A has 1 at A(4,5), A(5,4), A(4,8), A(8,4), A(5,8) and A(8,5).
How can I create such an adjacency matrix A given the matrix B in an efficient way?
This doesn't require any toolbox, and works for square or rectangular matrices. It uses array operations with complex numbers.
Consider a binary matrix B of size M×N.
Create an M×N matrix, t, that contains the complex coordinates of each nonzero entry of B. That is, entry t(r,c) contains r+1j*c if B(r,c) is nonzero, and NaN otherwise.
Compute an M*N×M*N matrix, d, containing the absolute difference for each pair of entries of B. Pairs of entries of B that are nonzero and adjacent will produce 1 or sqrt(2) in matrix d.
Build the result matrix, A, such that it contains 1 iff the corresponding entry in d equals 1 or sqrt(2). Equivalently, and more robust to numerical errors, iff the corresponding entry in d is between 0 and 1.5.
Code:
B = [0 1 0; 0 1 1; 0 0 0]; % input
t = bsxfun(#times, B, (1:size(B,1)).') + bsxfun(#times, B, 1j*(1:size(B,2)));
t(t==0) = NaN; % step 1
d = abs(bsxfun(#minus, t(:), t(:).')); % step 2
A = d>0 & d<1.5; % step 3
To get B back from A:
B2 = zeros(sqrt(size(A,1)));
B2(any(A,1)) = 1;
Here is a solution using image processing toolbox* that creates sparse matrix representation of the adjacency matrix:
B = [ 0 1 0;
0 1 1;
0 0 0 ]
n = numel(B);
C = zeros(size(B));
f = find(B);
C(f) = f;
D = padarray(C,[1 1]);
%If you don't have image processing toolbox
%D = zeros(size(C)+2);
%D(2:end-1,2:end-1)=C;
E = bsxfun(#times, im2col(D,[3 3]) , reshape(B, 1,[]));
[~ ,y] = find(E);
result = sparse(nonzeros(E),y,1,n,n);
result(1:n+1:end) = 0;
*More efficient implementation of im2col can be found here.
I want to obtain all the possible permutations of one vector elements by another vector elements. For example one vector is A=[0 0 0 0] and another is B=[1 1]. I want to replace the elements of A by B to obtain all the permutations in a matrix like this [1 1 0 0; 1 0 1 0; 1 0 0 1; 0 1 1 0; 0 1 0 1; 0 0 1 1]. The length of real A is big and I should be able to choose the length of B_max and to obtain all the permutations of A with B=[1], [1 1], [1 1 1],..., B_max.
Thanks a lot
Actually, since A and B are always defined, respectively, as a vector of zeros and a vector of ones, this computation is much easier than you may think. The only constraints you should respect concerns B, which shoud not be empty and it's elements cannot be greater than or equal to the number of elements in A... because after that threshold A will become a vector of ones and calculating its permutations will be just a waste of CPU cycles.
Here is the core function of the script, which undertakes the creation of the unique permutations of 0 and 1 given the target vector X:
function p = uperms(X)
n = numel(X);
k = sum(X);
c = nchoosek(1:n,k);
m = size(c,1);
p = zeros(m,n);
p(repmat((1-m:0)',1,k) + m*c) = 1;
end
And here is the full code:
clear();
clc();
% Define the main parameter: the number of elements in A...
A_len = 4;
% Compute the elements of B accordingly...
B_len = A_len - 1;
B_seq = 1:B_len;
% Compute the possible mixtures of A and B...
X = tril(ones(A_len));
X = X(B_seq,:);
% Compute the unique permutations...
p = [];
for i = B_seq
p = [p; uperms(X(i,:).')];
end
Output for A_len = 4:
p =
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
1 1 0 0
1 0 1 0
1 0 0 1
0 1 1 0
0 1 0 1
0 0 1 1
1 1 1 0
1 1 0 1
1 0 1 1
0 1 1 1
I have a 3D matrix A (size m*n*k) where m=latitude, n*longitude and k=time.
I want only specific values from first and second dimension, specified by a logical matrix B (size m*n), and I want only the timesteps specified by vector C (size k).
In the end this should become a 2D matrix D, since the first two dimesions will colapse to one.
What is the most easy approach to do this?
And also is it possible to combine logical with linear indizes here? For example B is logical and C is linear?
Sample code with rand:
A=rand(10,10,10);
B=randi([0 1], 10,10);
C=randi([0 1], 10,1);
D=A(B,C) %This would be my approach which doesnt work. The size of D should be sum(B)*sum(c)
Another example without rand:
A=reshape([1:27],3,3,3);
B=logical([1,0,0;1,0,0;0,0,0]);
C=(1,3); %get data from timestep 1 and 5
D=A(B,C);%What I want to do, but doesnÄt work that way
D=[1,19;2,20];%Result should look like this! First dimension is now all data from dimesion 1 and 2. New dimesion 2 is now the time.
A = rand(4,4,4);
B = randi([0 1], 4,4)
B =
1 1 0 1
1 0 1 1
0 0 1 0
1 0 1 1
>> C = randi([0 1],1,1,4);
>> C(:)
ans =
0
1
1
0
Then use bsxfun or implicit expansion expansion whith .* if newer Matlab version to generate a matrix of logical for you given coordinates.
>> idx = logical(bsxfun(#times,B,C))
idx(:,:,1) =
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
idx(:,:,2) =
1 1 0 1
1 0 1 1
0 0 1 0
1 0 1 1
idx(:,:,3) =
1 1 0 1
1 0 1 1
0 0 1 0
1 0 1 1
idx(:,:,4) =
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
Then your output is D = A(idx). However, note that this D is now an Nx1 array. Where N is number of true elements is B times number of true elements in C. 10x True in B and 2x True in C:
>> size(D)
ans =
20 1
An easy way to do it is to first reshape A into an m*n-by-k matrix, then do your indexing:
result = reshape(A, [], size(A, 3));
result = result(B, C);
In this case C can be either a logical vector or vector of indices.
In each iteration I want to add 1 randomly to binary vector,
Let say
iteration = 1,
k = [0 0 0 0 0 0 0 0 0 0]
iteration = 2,
k = [0 0 0 0 1 0 0 0 0 0]
iteration = 3,
k = [0 0 1 0 0 0 0 1 0 0]
, that goes up to length(find(k)) = 5;
Am thinking of for loop but I don't have an idea how to start.
If it's important to have the intermediate vectors (those with 1, 2, ... 4 ones) as well as the final one, you can generate a random permutation and, in your example, use the first 5 indices one at a time:
n = 9; %// number of elements in vector
m = 5; %// max number of 1's in vector
k = zeros(1, n);
disp(k); %// output vector of all 0's
idx = randperm(n);
for p = 1:m
k(idx(p)) = 1;
disp(k);
end
Here's a sample run:
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1
1 0 0 0 0 0 0 0 1
1 0 0 1 0 0 0 0 1
1 0 0 1 1 0 0 0 1
1 1 0 1 1 0 0 0 1
I wouldn't even use a loop. I would generate a random permutation of indices that sample from a vector going from 1 up to the length of k without replacement then just set these locations to 1. randperm suits the task well:
N = 10; %// Length 10 vector
num_vals = 5; %// 5 values to populate
ind = randperm(N, num_vals); %// Generate a vector from 1 to N and sample num_vals values from this vector
k = zeros(1, N); %// Initialize output vector k to zero
k(ind) = 1; %// Set the right values to 1
Here are some sample runs when I run this code a few times:
k =
0 0 1 1 0 1 1 0 0 1
k =
1 0 0 0 1 0 1 1 0 1
k =
1 0 0 0 1 0 1 1 0 1
k =
0 1 1 1 0 0 1 0 0 1
However, if you insist on using a loop, you can generate a vector from 1 up to the desired length, randomly choose an index in this vector then remove this value from the vector. You'd then use this index to set the location of the output:
N = 10; %// Length 10 vector
num_vals = 5; %// 5 values to populate
vec = 1 : N; %// Generate vector from 1 up to N
k = zeros(1, N); %// Initialize output k
%// Repeat the following for as many times as num_vals
for idx = 1 : num_vals
%// Obtain a value from the vector
ind = vec(randi(numel(vec), 1));
%// Remove from the vector
vec(ind) = [];
%// Set location in output to 1
k(ind) = 1;
end
The above code should still give you the desired effect, but I would argue that it's less efficient.
I have a m-dimensional vector of integers ranging from 1 to n. These integers are column indexes for m × n matrix.
I want to create a m × n matrix of 0s and 1s, where in m-th row there's a 1 in the column that is specified by m-th value in my vector.
Example:
% my vector (3-dimensional, values from 1 to 4):
v = [4;
1;
2];
% corresponding 3 × 4 matrix
M = [0 0 0 1;
1 0 0 0;
0 1 0 0];
Is this possible without a for-loop?
Of course, that's why they invented sparse matrices:
>> M = sparse(1:length(v),v,ones(length(v),1))
M =
(2,1) 1
(3,2) 1
(1,4) 1
which you can convert to a full matrix if you want with full:
>> full(M)
ans =
0 0 0 1
1 0 0 0
0 1 0 0
Or without sparse matrix:
>> M = zeros(max(v),length(v));
>> M(v'+[0:size(M,2)-1]*size(M,1)) = 1;
>> M = M'
M =
0 0 0 1
1 0 0 0
0 1 0 0
Transposition is used because in matlab arrays are addressed by columns
In Octave, at least as of 3.6.3, you can do this easily using broadcasting:
M = v==1:4