Drawing uniform random numbers from within an ellipsoid

Here’s a Matlab function I wrote to draw a set of random numbers from a uniform distribution within an n-dimensional ellipsoid, where the size and orientation of the ellipsoid is defined by a covariance matrix.

function pnts = draw_from_ellipsoid( covmat, cent, npts )

% function pnts = draw_from_ellipsoid( covmat, cent, npts )
%
% This function draws points uniformly from an n-dimensional ellipsoid
% with edges and orientation defined by the the covariance matrix covmat. 
% The number of points produced in the n-dimensional space is given by 
% npts, with an output array of npts by ndims. The centre of each dimension
% is given in the vector cent.

% get number of dimensions of the covariance matrix
ndims = length(covmat);

% calculate eigenvalues and vectors of the covariance matrix
[v, e] = eig(covmat);

% check size of cent and transpose if necessary
sc = size(cent);
if sc(1) > 1
    cent = cent';
end 

% generate radii of hyperspheres
rs = rand(npts,1);

% generate points
pt = randn(npts,ndims);

% get scalings for each point onto the surface of a unit hypersphere
fac = sum(pt(:,:)'.^2);

% calculate scaling for each point to be within the unit hypersphere 
% with radii rs
fac = (rs.^(1/ndims)) ./ sqrt(fac');

pnts = zeros(npts,ndims);

% scale points to the ellipsoid using the eigenvalues and rotate with 
% the eigenvectors and add centroid
d = sqrt(diag(e));
for i=1:npts
    % scale points to a uniform distribution within unit hypersphere
    pnts(i,:) = fac(i)*pt(i,:);
    
    % scale and rotate to ellipsoid
    pnts(i,:) = (pnts(i,:) .* d' * v') + cent;
end

end

6 Replies to “Drawing uniform random numbers from within an ellipsoid”

  1. Thanks very much for this. In case anyone wants the code for Python, here you go:

    def draw_from_ellipsoid( covmat, cent, npts):
        # random uniform points within ellipsoid as per: https://ma.ttpitk.in/blog/?p=368
        ndims = covmat.shape[0]
        
        # calculate eigenvalues (e) and eigenvectors (v)
        eigenValues, eigenVectors = np.linalg.eig(covmat)
        idx = (-eigenValues).argsort()[::-1][:ndims]
        e = eigenValues[idx]
        v = eigenVectors[:,idx]
        e = np.diag(e)
        
        # generate radii of hyperspheres
        rs = np.random.uniform(0,1,npts)
        
        # generate points
        pt = np.random.normal(0,1,[npts,ndims]);
    
        # get scalings for each point onto the surface of a unit hypersphere
        fac = np.sum(pt**2,axis=1)
    
        # calculate scaling for each point to be within the unit hypersphere 
        # with radii rs
        fac = (rs**(1.0/ndims)) / np.sqrt(fac)
    
        pnts = np.zeros((npts,ndims));
    
        # scale points to the ellipsoid using the eigenvalues and rotate with 
        # the eigenvectors and add centroid
        d = np.sqrt(np.diag(e))
        d.shape = (ndims,1)
        
        for i in range(0,npts):
            # scale points to a uniform distribution within unit hypersphere
            pnts[i,:] = fac[i]*pt[i,:]
            pnts[i,:] = np.dot(np.multiply(pnts[i,:],np.transpose(d)),np.transpose(v)) + cent
            
        return pnts
    
  2. Hi, thanks a lot! Can you point me to the underlying theory?

    I get the intuition:

    You sample points from a normal distribution and normalize them to be on the unit sphere. That way, you get points with a uniform distribution in their orientation. Then you combine the uniform orientations with uniform radii (is that results still uniform?). Where is the exponent in r^*(1/ndim) coming from?
    Finally, you rotate and rescale.

    Looking forward to some hint,

    Felix

  3. Someone needs this in c++, here you go..

    Eigen::MatrixXf generate_samples_from_ellipsoid(Eigen::MatrixXf covmat, Eigen::VectorXf cent, int npts){
    int ndims = covmat.rows();
    Eigen::EigenSolver eigensolver;
    eigensolver.compute(covmat);
    Eigen::VectorXf eigen_values = eigensolver.eigenvalues().real();
    Eigen::MatrixXf eigen_vectors = eigensolver.eigenvectors().real();
    std::vector<std::tuple> eigen_vectors_and_values;

    for(int i=0; i<eigen_values.size(); i++){
    std::tuple vec_and_val(eigen_values[i], eigen_vectors.row(i));
    eigen_vectors_and_values.push_back(vec_and_val);
    }
    std::sort(eigen_vectors_and_values.begin(), eigen_vectors_and_values.end(),
    [&](const std::tuple& a, const std::tuple& b) -> bool{
    return std::get(a) <= std::get(b);
    });
    int index = 0;
    for(auto const vect : eigen_vectors_and_values){
    eigen_values(index) = std::get(vect);
    eigen_vectors.row(index) = std::get(vect);
    index++;
    }

    Eigen::MatrixXf eigen_values_as_matrix = eigen_values.asDiagonal();

    std::random_device rd{};
    std::mt19937 gen{rd()};
    std::uniform_real_distribution dis(0, 1);
    std::normal_distribution normal_dis{0.0f, 1.0f};

    Eigen::MatrixXf pt = Eigen::MatrixXf::Zero(npts, ndims).unaryExpr([&](float dummy){return (float)normal_dis(gen);});
    Eigen::VectorXf rs = Eigen::VectorXf::Zero(npts).unaryExpr([&](float dummy){return dis(gen);});
    Eigen::VectorXf fac = pt.array().pow(2).rowwise().sum();
    Eigen::VectorXf fac_sqrt = fac.array().sqrt();
    Eigen::VectorXf rs_pow = rs.array().pow(1.0/ndims);
    fac = rs_pow.array()/fac_sqrt.array();
    Eigen::MatrixXf points = Eigen::MatrixXf::Zero(npts, ndims);
    Eigen::VectorXf d = eigen_values_as_matrix.diagonal().array().sqrt();
    for(auto i(0); i<npts; i++){
    points.row(i) = fac(i)*pt.row(i).array();
    Eigen::MatrixXf fff = (points.row(i).array()*d.transpose().array());
    Eigen::VectorXf bn = eigen_vectors*fff.transpose();
    points.row(i) = bn.array() + cent.array();
    }
    std::cout << "points: " << points << std::endl;
    return points;
    }

Leave a Reply to Felix Cancel reply