Imagine you have this image
[[1 2 3] [4 5 6] [7 8 90]]
You flatten it into this format -
[1 2 3 4 5 6 7 8 90]
Now you are given the index of Pixel 90 to be 8.
How can you find that pixel 90 is in Row 3 and column 3?
OpenCL, similarly to other programming languages like C, C++, Java and so on, uses zero based indexing. So in this terms you are looking for Row 2 and Column 2.
Now to calculate which row that is we need to divide index position 8 by number of columns:
8 / 3 = 2
So in zero based indexing that is a second row.
Now to calculate which column that is we use modulo operator:
8 % 3 = 2
In the 2D case, a point (x,y) in a rectangle with the dimensions (sx,sy) can be represented in 1D space by a linear index n as follows:
n = x+y*sx
Converting the 1D index n back to (x,y) works as follows:
x = n%sx
y = n/sx
For the 3D case, a point (x,y,z) in the a box with dimensions (sx,sy,sz) can be represented in 1D as
n = x+(y+z*sy)*sx
and converted back to (x,y,z) like this:
z = n/(sx*sy);
temp = n%(sx*sy);
y = temp/sx;
x = temp%sx;
Note that "/" here means integer division (always rounds down the result) and "%" is the modulo operator.
I am searching for a method to create, in a fast way a random matrix A with the follwing properties:
A = transpose(A)
A(i,i) = 0 for all i
A(i,j) >= 0 for all i, j
sum(A) =~ degree; the sum of rows are randomly distributed by a distribution I want to specify (here =~ means approximate equality).
The distribution degree comes from a matrix orig, specifically degree=sum(orig), thus I know that matrices with this distribution exist.
For example: orig=[0 12 7 5; 12 0 1 9; 7 1 0 3; 5 9 3 0]
orig =
0 12 7 5
12 0 1 9
7 1 0 3
5 9 3 0
sum(orig)=[24 22 11 17];
Now one possible matrix A=[0 11 5 8, 11 0 4 7, 5 4 0 2, 8 7 2 0] is
A =
0 11 5 8
11 0 4 7
5 4 0 2
8 7 2 0
with sum(A)=[24 22 11 17].
I am trying this for quite some time, but unfortunatly my two ideas didn't work:
version 1:
I switch Nswitch times two random elements: A(k1,k3)--; A(k1,k4)++; A(k2,k3)++; A(k2,k4)--; (the transposed elements aswell).
Unfortunatly, Nswitch = log(E)*E (with E=sum(sum(nn))) in order that the Matrices are very uncorrelated. As my E > 5.000.000, this is not feasible (in particular, as I need at least 10 of such matrices).
version 2:
I create the matrix according to the distribution from scratch. The idea is, to fill every row i with degree(i) numbers, based on the distribution of degree:
nn=orig;
nnR=zeros(size(nn));
for i=1:length(nn)
degree=sum(nn);
howmany=degree(i);
degree(i)=0;
full=rld_cumsum(degree,1:length(degree));
rr=randi(length(full),[1,howmany]);
ff=full(rr);
xx=i*ones([1,length(ff)]);
nnR = nnR + accumarray([xx(:),ff(:)],1,size(nnR));
end
A=nnR;
However, while sum(A')=degree, sum(A) systematically deviates from degree, and I am not able to find the reason for that.
Small deviations from degree are fine of course, but there seem to be systmatical deviations in particulat of the matrices contain in some places large numbers.
I would be very happy if somebody could either show me a fast method for version1, or a reason for the systematic deviation of the distribution in version 2, or a method to create such matrices in a different way. Thank you!
Edit:
This is the problem in matsmath's proposed solution:
Imagine you have the matrix:
orig =
0 12 3 1
12 0 1 9
3 1 0 3
1 9 3 0
with r(i)=[16 22 7 13].
Step 1: r(1)=16, my random integer partition is p(i)=[0 7 3 6].
Step 2: Check that all p(i)<=r(i), which is the case.
Step 3:
My random matrix starts looks like
A =
0 7 3 6
7 0 . .
3 . 0 .
6 . . 0
with the new row sum vector rnew=[r(2)-p(2),...,r(n)-p(n)]=[15 4 7]
Second iteration (here the problem occures):
Step 1: rnew(1)=15, my random integer partition is p(i)=[0 A B]: rnew(1)=15=A+B.
Step 2: Check that all p(i)<=rnew(i), which gives A<=4, B<=7. So A+B<=11, but A+B has to be 15. contradiction :-/
Edit2:
This is the code representing (to the best of my knowledge) the solution posted by David Eisenstat:
orig=[0 12 3 1; 12 0 1 9; 3 1 0 3; 1 9 3 0];
w=[2.2406 4.6334 0.8174 1.6902];
xfull=zeros(4);
for ii=1:1000
rndmat=[poissrnd(w(1),1,4); poissrnd(w(2),1,4); poissrnd(w(3),1,4); poissrnd(w(4),1,4)];
kkk=rndmat.*(ones(4)-eye(4)); % remove diagonal
hhh=sum(sum(orig))/sum(sum(kkk))*kkk; % normalisation
xfull=xfull+hhh;
end
xf=xfull/ii;
disp(sum(orig)); % gives [16 22 7 13]
disp(sum(xf)); % gives [14.8337 9.6171 18.0627 15.4865] (obvious systematic problem)
disp(sum(xf')) % gives [13.5230 28.8452 4.9635 10.6683] (which is also systematically different from [16, 22, 7, 13]
Since it's enough to approximately preserve the degree sequence, let me propose a random distribution where each entry above the diagonal is chosen according to a Poisson distribution. My intuition is that we want to find weights w_i such that the i,j entry for i != j has mean w_i*w_j (all of the diagonal entries are zero). This gives us a nonlinear system of equations:
for all i, (sum_{j != i} w_i*w_j) = d_i,
where d_i is the degree of i. Equivalently,
for all i, w_i * (sum_j w_j) - w_i^2 = d_i.
The latter can be solved by applying Newton's method as described below from a starting solution of w_i = d_i / sqrt(sum_j d_j).
Once we have the w_is, we can sample repeatedly using poissrnd to generate samples of multiple Poisson distributions at once.
(If I have time, I'll try implementing this in numpy.)
The Jacobian matrix of the equation system for a 4 by 4 problem is
(w_2 + w_3 + w_4) w_1 w_1 w_1
w_2 (w_1 + w_3 + w_4) w_2 w_2
w_3 w_3 (w_1 + w_2 + w_4) w_3
w_4 w_4 w_4 (w_1 + w_2 + w_3).
In general, let A be a diagonal matrix where A_{i,i} = sum_j w_j - 2*w_i. Let u = [w_1, ..., w_n]' and v = [1, ..., 1]'. The Jacobian can be written J = A + u*v'. The inverse is given by the Sherman--Morrison formula
A^-1*u*v'*A^-1
J^-1 = (A + u*v')^-1 = A^-1 - -------------- .
1 + v'*A^-1*u
For the Newton step, we need to compute J^-1*y for some given y. This can be done straightforwardly in time O(n) using the above equation. I'll add more detail when I get the chance.
First approach (based on version2)
Let your row sum vector given by the matrix orig [r(1),r(2),...,r(n)].
Step 1. Take a random integer partition of the integer r(1) into exactly n-1 parts, say p(2), p(3), ..., p(n)
Step 2. Check if p(i)<=r(i) for all i=2...n. If not, go to Step 1.
Step 3. Fill out your random matrix first row and colum by the entries 0, p(2), ... , p(n), and consider the new row sum vector [r(2)-p(2),...,r(n)-p(n)].
Repeat these steps with a matrix of order n-1.
The point is, that you randomize one row at a time, and reduce the problem to searching for a matrix of size one less.
As pointed out by OP in the comment, this naive algorithm fails. The reason is that the matrices in question have a further necessary condition on their entries as follows:
FACT:
If A is an orig matrix with row sums [r(1), r(2), ..., r(n)] then necessarily for every i=1..n it holds that r(i)<=-r(i)+sum(r(j),j=1..n).
That is, any row sum, say the ith, r(i), is necessarily at most as big as the sum of the other row sums (not including r(i)).
In light of this, a revised algorithm is possible. Note that in Step 2b. we check if the new row sum vector has the property discussed above.
Step 1. Take a random integer partition of the integer r(1) into exactly n-1 parts, say p(2), p(3), ..., p(n)
Step 2a. Check if p(i)<=r(i) for all i=2...n. If not, go to Step 1.
Step 2b. Check if r(i)-p(i)<=-r(i)+p(i)+sum(r(j)-p(j),j=2..n) for all i=2..n. If not, go to Step 1.
Step 3. Fill out your random matrix first row and colum by the entries 0, p(2), ... , p(n), and consider the new row sum vector [r(2)-p(2),...,r(n)-p(n)].
Second approach (based on version1)
I am not sure if this approach gives you random matrices, but it certainly gives you different matrices.
The idea here is to change some parts of your orig matrix locally, in a way which maintains all of its properties.
You should look for a random 2x2 submatrix below the main diagonal which contains strictly positive entries, like [[a,b],[c,d]] and perturbe its contents by a random value r to [[a+r,b-r],[c-r,d+r]]. You make the same change above the main diagonal too, to keep your new matrix symmetric. Here the point is that the changes within the entries "cancel" each other out.
Of course, r should be chosen in a way such that b-r>=0 and c-r>=0.
You can pursue this idea to modify larger submatrices too. For example, you might choose 3 random row coordinates r1, r2, r2 and 3 random column coordinates c1, c2, and c3 and then make changes in your orig matrix at the 9 positions (ri,cj) as follows: you change your 3x3 submatrix [[a b c],[d e f], [g h i]] to [[a-r b+r c] [d+r e f-r], [g h-r i+r]]. You do the same at the transposed places. Again, the random value r must be chosen in a way so that a-r>=0 and f-r>=0 and h-r>=0. Moreover, c1 and r1, and c3 and r3 must be distinct as you can't change the 0 entries in the main diagonal of the matrix orig.
You can repeat such things over and over again, say 100 times, until you find something which looks random. Note that this idea uses the fact that you have existing knowledge of a solution, this is the matrix orig, while the first approach does not use such knowledge at all.
I need this for Lagrange polynomials. I'm curious how one would do this without a for loop. The code currently looks like this:
tj = 1:n;
ti = zeros(n,n-1);
for i = 1:n
ti(i,:) = tj([1:i-1, i+1:end]);
end
My tj is not really just a 1:n vector but that's not important. While this for loop gets the job done, I'd rather use some matrix operation. I tried looking for some appropriate matrices to multiply it with, but no luck so far.
Here's a way:
v = [10 20 30 40]; %// example vector
n = numel(v);
M = repmat(v(:), 1, n);
M = M(~eye(n));
M = reshape(M,n-1,n).';
gives
M =
20 30 40
10 30 40
10 20 40
10 20 30
This should generalize to any n
ti = flipud(reshape(repmat(1:n, [n-1 1]), [n n-1]));
Taking a closer look at what's going on. If you look at the resulting matrix closely, you'll see that it's n-1 1's, n-1 2's, etc. from the bottom up.
For the case where n is 3.
ti =
2 3
1 3
1 2
So we can flip this vertically and get
f = flipud(ti);
1 2
1 3
2 3
Really this is [1, 2, 3; 1, 2, 3] reshaped to be 3 x 2 rather than 2 x 3.
In that line of thinking
a = repmat(1:3, [2 1])
1 2 3
1 2 3
b = reshape(a, [3 2]);
1 2
1 3
2 3
c = flipud(b);
2 3
1 3
1 2
We are now back to where you started when we bring it all together and replace 3's with n and 2's with n-1.
Here's another way. First create a matrix where each row is the vector tj but are stacked on top of each other. Next, extract the lower and upper triangular parts of the matrix without the diagonal, then add the results together ensuring that you remove the last column of the lower triangular matrix and the first column of the upper triangular matrix.
n = numel(tj);
V = repmat(tj, n, 1);
L = tril(V,-1);
U = triu(V,1);
ti = L(:,1:end-1) + U(:,2:end);
numel finds the total number of values in tj which we store in n. repmat facilitates the stacking of the vector tj to create a matrix that is n x n large. After, we use tril and triu so that we extract the lower and upper triangular parts of the matrices without the diagonal. In addition, the rest of the matrix is all zero except for the relevant triangular parts. The -1 and 1 flags for tril and triu respectively extract this out successfully while ensuring that the diagonal is all zero. This creates a column of extra zeroes appearing at the last column when calling tril and the first column when calling triu. The last part is to simply add these two matrices together ignoring the last column of the tril result and the first column of the triu result.
Given that tj = [10 20 30 40]; (borrowed from Luis Mendo's example), we get:
ti =
20 30 40
10 30 40
10 20 40
10 20 30
There are 3 points in 3D space. There are 2 orthogonal coordinate systems with the same origin. I know coordinates of those 3 points in both coordinate systems. Given a new point with its coordinates in the first coordinate system, how can I find its coordinates in the second coordinate system?
I think it's possible to get a rotation matrix using given points which does this, but I did not succeed doing this.
You can do it using matrix inverses. Three matrix-vector multiplications (e.g. transforming three 3D vectors by a 3x3 matrix) is equivalent to multiplying two 3x3 matrices together.
So, you can put your first set of points in one matrix, call it A:
0 0 1 < vector 1
0 1 0 < vector 2
2 0 0 < vector 3
Then put your second set of points in a second matrix, call it C. As an example, imagine a transform that scales by a factor of 2 around the origin and flips the Y and Z axes:
0 2 0 < vector 1
0 0 2 < vector 2
4 0 0 < vector 3
So, if A x B = C, we need to find the matrix B, which we can find by finding the A-1:
Inverse of A:
0 0 0.5
0 1 0
1 0 0
The multiply A-1 x C (in that order):
2 0 0
0 0 2
0 2 0
This is a transform matrix B that you can apply to new points. Dot-product multiply the vector by the first column to get the transformed X, second column to get the transformed Y, etc.
Im trying to learn chain matrix multiplication.
Suppose A is a 10 × 30 matrix, B is a 30 × 5 matrix, and C is a 5 × 60 matrix. Then,
How do we get the following number of operations? (Is it number of rows into columns ???)
(AB)C = (10×30×5) + (10×5×60) = 1500 + 3000 = 4500 operations
A(BC) = (30×5×60) + (10×30×60) = 9000 + 18000 = 27000 operations.
http://www.geeksforgeeks.org/dynamic-programming-set-8-matrix-chain-multiplication/
The number of operations is the number of multiplications required to calculate the result. A * B will result in a 10 x 5 matrix. Each entry in this matrix is the dotproduct of the respective row of A with the column of B with the same index. Thus: A * B requires calculation of 10 x 5 cells, where each cell is the sum of 30 multiplication, so 10 x 5 x 30. Though this is a rather strange representation.