problem description
I'm using Eigen for some matrix task. Say I have matrix A whose size is 4x3, then its transpose A^T is 3x4 size, then A^T * A is 3x3 size, thus the inverse, (A^T * A)^(-1), is also 3x3 size.
I would like to get (A^T * A)^(-1). By using the mentioned formula, and by manually defining A^T * A matrix then do inverse, I got different result for (A^T * A)^(-1) matrix. Curious why this two results mismatch and differs very much.
Reproduce
Eigen version: git commit 972cf0c28a8d2ee0808c1277dea2c5c206591ce6
System: Ubuntu 20.04
Compiler: Clang++, 10.0.0
Code:
#include <iostream>
#include <Eigen/Dense>
int main() {
{
printf("----- A^T * A by calculate\n");
Eigen::MatrixXd A(4, 3);
A << 1.70023e+06, 1303.93, 1,
1.99113e+06, 1411.07, 1,
1.97211e+06, 1404.32, 1,
1.69433e+06, 1301.67, 1;
std::cout << "A:" << std::endl << A << std::endl;
Eigen::MatrixXd AT = A.transpose();
std::cout << "A^T:" << std::endl << AT << std::endl;
Eigen::MatrixXd ATA = AT * A;
std::cout << "A^T * A:" << std::endl << ATA << std::endl;
std::cout << "(A^T * A)^(-1):" << std::endl << ATA.inverse() << std::endl;
}
{
printf("----- A^T * A by define\n");
Eigen::MatrixXd ATA(3, 3);
ATA << 1.36154e+13, 1.00015e+10, 7.3578e+06,
1.00015e+10, 7.3578e+06, 5420.99,
7.3578e+06, 5420.99, 4;
std::cout << "A^T * A:" << std::endl << ATA << std::endl;
std::cout << "(A^T * A)^(-1):" << std::endl << ATA.inverse() << std::endl;
}
return 0;
}
The actual outputs:
----- A^T * A by calculate
A:
1.70023e+06 1303.93 1
1.99113e+06 1411.07 1
1.97211e+06 1404.32 1
1.69433e+06 1301.67 1
A^T:
1.70023e+06 1.99113e+06 1.97211e+06 1.69433e+06
1303.93 1411.07 1404.32 1301.67
1 1 1 1
A^T * A:
1.36154e+13 1.00015e+10 7.3578e+06
1.00015e+10 7.35781e+06 5420.99
7.3578e+06 5420.99 4
(A^T * A)^(-1):
3.49282e-06 -0.00946871 6.40758
-0.00946871 25.6689 -17370.5
6.40758 -17370.5 1.17549e+07
----- A^T * A by define
A^T * A:
1.36154e+13 1.00015e+10 7.3578e+06
1.00015e+10 7.3578e+06 5420.99
7.3578e+06 5420.99 4
(A^T * A)^(-1):
6.1435e-09 -1.66513e-05 0.0112659
-1.66513e-05 0.0452221 -30.658
0.0112659 -30.658 20826.4
This is just a matter of numerical accuracy. As pointed out by #Damien in the comment, the matrix is ill-conditioned, and thus a small difference in the input can lead to a large change in the results. By copying from the output only the first five digits and using them to manually define the second matrix, a significant part is discarded that is handled internally but not displayed with the standard accuracy of std::cout.
Take for instance the entry ATA(0,0). By using is ATA << 1.36154e+13,..., any value of the order of 1.e7 or lower is not considered. This is, however, the order of the other entries in the matrix.
In short, both results are correct, but your manually defined matrix ATA is not the same as the one that is calculated in the first part. (You can take the difference between the two to verify this).
My today practice for inverse matrix 3x3 is using
template<typename Derived>
template<typename ResultType>
inline void MatrixBase<Derived>::computeInverseAndDetWithCheck(
ResultType& inverse,
typename ResultType::Scalar& determinant,
bool& invertible,
const RealScalar& absDeterminantThreshold
) const
Set manual absDeterminantThreshold is default:
1)for float 1e-5
2)for double 1e-12
3)for long double 1e-15
For example your CODE:
Eigen::Matrix3d ATA_inv;
bool invertible = false;
double determinant = 0.;
double Threshold = abs(ATA.diagonal().prod());//maybe any test
ATA.computeInverseAndDetWithCheck(ATA_inv,determinant,invertible,Threshold<1e-12?Threshold*0.1:1e-12);
Related
I want to normalize the vectors, yet I could not understand how to do such a basic operation. There does not exist a direct method like Vector_3::normalize().
I call squared_length() and then CGAL::sqr_root() to find the vector length. Then when I want to divide my vector to its length, compiler does not permit since CGAL types are not compatible. Where am I wrong?
I'm expanding here the comment by #MarcGlisse.
There are two scenarios - you want exact calculations or you don't. In the first case you can use the Exact_predicates_exact_constructions_kernel_with_sqrt and the normalized vector will be exact.
In the second case you can use many kernels, but the normalized vector will probably have the length, slightly different from 1. Also some kernels don't have the function CGAL::sqrt at all - however you can always convert squared distances to the double type, using the function CGAL::to_double. Another way is to use a function CGAL::approximate_sqrt, which automatically does this conversion if it's necessary. These conversions make calculations not exact. Please see the example below:
#include <iostream>
#include <CGAL/Exact_predicates_exact_constructions_kernel_with_sqrt.h>
#include <CGAL/Exact_predicates_exact_constructions_kernel.h>
#include <CGAL/Simple_cartesian.h>
using KernelExactWithSqrt = CGAL::Exact_predicates_exact_constructions_kernel_with_sqrt;
using KernelExact = CGAL::Exact_predicates_exact_constructions_kernel;
using KernelInexact = CGAL::Simple_cartesian<double>;
template <typename T>
auto normalize(T const& V)
{
auto const slen = V.squared_length();
auto const d = CGAL::approximate_sqrt(slen);
return V / d;
}
template <typename T>
auto check()
{
typename T::Point_2 const p{0, 0}, q{1, 1};
auto const n = normalize(typename T::Vector_2{p, q});
return n.squared_length() == 1;
}
int main()
{
std::cout << "====== Exact kernel with square root ======" << std::endl;
std::cout << check<KernelExactWithSqrt>() << std::endl;
std::cout << "====== Inexact kernel ======" << std::endl;
std::cout << check<KernelInexact>() << std::endl;
std::cout << "====== Exact kernel with automatic conversion to double ======" << std::endl;
std::cout << check<KernelExact>() << std::endl;
}
The output:
====== Exact kernel with square root ======
1
====== Inexact kernel ======
0
====== Exact kernel with automatic conversion to double ======
0
So, this example shows that the Exact_predicates_exact_constructions_kernel_with_sqrt guarantees you that the normalized vector will be what you expect.
Does armadillo offer functions that apply element-wise operations only on the diagonal elements of a matrix? Say I want to run an inverse on it then I get infs on off-diagonals. I know that I can get it with arma::inv(m), but for diagonal matrices it seems faster to run it element-wise.
Minimal example:
#include <iostream>
#include <armadillo>
int main() {
arma::Col<double> v = {1.2, 3.14, 2.71, 0.1};
arma::Mat<double> m = arma::diagmat(v);
std::cout << " A diagonal matrix:\n" << m << "\n\n";
std::cout << " It's elemenet-wise inverse:\n" << 1.0/m << std::endl;
return 0;
}
I have to take the coordinates of the vertices of a triangle from the user and tell if it is a right-angled triangle or not. I'm using Pythagoras Theorem to Find out i.e. h * h = b * b + p * p
But surprisingly this doesn't work for some specific right-angled triangles.
Here is one such Triangle:
Vertex A: (x, y) = (1, 3)
Vertex B: (x, y) = (1, 1)
Vertex C: (x, y) = (5, 1)
It calculates perfectly, which I figured out by printing the calculation, but still doesn't work.
Then I tried by using sqrt() function from the cmath library this way:
h = sqrt(b * b + p * p)
Logically it is the same, but it worked.
I want to understand, why the earlier method is not working?
Here is a simplified version of My Code:
#include <iostream>
#include <cmath>
using namespace std;
class Vertex {
double x, y;
public:
void take_input(char obj) {
cout << endl << " Taking Coordinates of Vertex " << obj << ": " << endl;
cout << " Enter the x component: ";
cin >> x;
cout << " Enter the y component: ";
cin >> y;
}
double distance(Vertex p) {
double dist = sqrt((x-p.x)*(x-p.x) + (y-p.y)*(y-p.y));
return dist;
}
};
class Triangle {
Vertex a, b, c;
public:
void take_inp(string obj) {
cout << endl << "Taking Vertices of the Triangle " << obj << ": " << endl;
cout << " Verteces should be in a counter clockwise order (as per convention)." << endl;
a.take_input('A');
b.take_input('B');
c.take_input('C');
}
void is_rt_ang() {
double h = a.distance(c)*a.distance(c);
double bp = a.distance(b)*a.distance(b) + b.distance(c)*b.distance(c);
/*
// Strangely this attempt works which is logically the same:
double h = a.distance(c);
double bp = sqrt(a.distance(b)*a.distance(b) + b.distance(c)*b.distance(c));
*/
if (h == bp) {
cout << "Angle is 90" << endl;
cout << h << " = " << bp << endl;
cout << "It is Right-Angled" << endl;
}
else {
cout << "Angle is not 90!" << endl;
cout << h << " != " << bp << endl;
cout << "It is Not a Right-Angled" << endl;
}
}
};
int main()
{
Triangle tri1, tri2;
tri1.take_inp("tri1");
tri1.is_rt_ang();
return 0;
}
The line
double dist = sqrt((x-p.x)*(x-p.x) + (y-p.y)*(y-p.y));
in the Vertex::distance method gives you an approximation of a square root which is rarely going to coincide with an exact answer. This is because most real numbers can't be represented in floating point arithmetic.
But in given code sample you can make do without sqrt. Replace Vertex::distance method with a method
double distance_square(Vertex p) {
double dist_square = (x-p.x)*(x-p.x) + (y-p.y)*(y-p.y);
return dist_square;
}
and call it like this in Triangle::is_rt_ang:
double h = a.distance_square(c);
double bp = a.distance_square(b) + b.distance_square(c);
This solution is still flawed because floating-point multiplication is also a subject to rounding errors. But if it is guaranteed that you are going to work only with integer coordinates, you can replace all doubles in your code with ints and for them there is no problem with multiplication (besides possibly going out of bounds for large numbers).
EDIT: Also a comment on printing
It calculates perfectly, which I figured out by printing the
calculation, but still doesn't work.
When you print doubles you need to set precision manually in order to avoid rounding. If in your code I replace a line
cout << h << " != " << bp << endl;
with
cout << std::setprecision(std::numeric_limits<double>::digits10) << std::fixed << h << " != " << bp << endl;
then for example triangle from the question I get the output
Angle is not 90!
20.000000000000004 != 20.000000000000000
It is Not a Right-Angled
For this to compile you will need to add #include <limits> and #include <iomanip>.
In your is_rt_ang function you're assuming that your hypotenuse is always going to be the edge AC, but it doesn't seem like you're doing anything to verify this.
double h = a.distance(c)*a.distance(c);
double bp = a.distance(b)*a.distance(b) + b.distance(c)*b.distance(c);
You could try getting the squares of all your distances first, (AC)^2, (AB)^2, and (BC)^2, then finding the candidate for hypotenuse by taking the max value out of the three, then do something like:
bool isRightTriangle = max == (min1 + min2)
You may also be running into some kind of round-off error with floating point numbers. It is common to use a an epsilon value when comparing floating point numbers because of the inherent round-off errors with them. If you don't need floating point values maybe use an integer, or if you do need floating point values try using an epsilon value in your equalities like:
abs(h - bp) <= epsilon
You should be able to find more information about floating point values, round-off errors, and machine epsilons on the web.
Here is a link to a SO Q/A that talks about floating point math that may be a good resource for you: Is floating point math broken?
I got two different .cpp files with two equal calculations.. But they don't return the same results?
A code
double theta = (double)maxLoc.y/angleBins*CV_PI;
std::cout << theta << " " << abs(sin(theta)) << std::endl;
A result
1.53589 0.999391
B code
double theta = (double)maxLoc.y / angleBins * CV_PI;
std::cout << theta << " " << abs(sin(theta)) << std::endl;
B result
1.53589 0
You're probably calling the abs function from C in the second code snippet which takes an int as parameter. Using std::abs in both code snippets (and the correct header #include<cmath>) will fix the problem.
In matlab/octave pairwise distances between matrices as required for e.g. k-means are calculated by one function call (see cvKmeans.m), to distFunc(Codebook, X) with as arguments two matrices of dimensions KxD.
In Eigen this can be done for a matrix and one vector by using broadcasting, as explained on eigen.tuxfamily.org:
(m.colwise() - v).colwise().squaredNorm().minCoeff(&index);
However, in this case v is not just a vector, but a matrix. What's the equivalent oneliner in Eigen to calculate such pairwise (Euclidean) distances across all entries between two matrices?
I think the appropriate solution is to abstract this functionality into a function. That function may well be templated; and it may well use a loop - the loop will be really short, after all. Many matrix operations are implemented using loops - that's not a problem.
For example, given your example of...
MatrixXd p0(2, 4);
p0 <<
1, 23, 6, 9,
3, 11, 7, 2;
MatrixXd p1(2, 2);
p1 <<
2, 20,
3, 10;
then we can construct a matrix D such that D(i,j) = |p0(i) - p1(j)|2
MatrixXd D(p0.cols(), p0.rows());
for (int i = 0; i < p1.cols(); i++)
D.col(i) = (p0.colwise() - p1.col(i)).colwise().squaredNorm().transpose();
I think this is fine - we can use some broadcasting to avoid 2 levels of nesting: we iterate over p1's points, but not over p0's points, nor over their dimensions.
However, you can make a oneliner if you observe that |p0(i) - p1(j)|2 = |p0(i)|2 + |p1(j)|2 - 2 p0(i)T p1(j). In particular, the last component is just matrix multiplication, so D = -2 p0T p1 + ...
The blank left to be filled is composed of a component that only depends on the row; and a component that only depends on the column: these can be expressed using rowwise and columnwise operations.
The final "oneliner" is then:
D = ( (p0.transpose() * p1 * -2
).colwise() + p0.colwise().squaredNorm().transpose()
).rowwise() + p1.colwise().squaredNorm();
You could also replace the rowwise/colwise trickery with an (outer) product with a 1 vector.
Both methods result in the following (squared) distances:
1 410
505 10
32 205
50 185
You'd have to benchmark which is fastest, but I wouldn't be surprised to see the loop win, and I expect that's more readable too.
Eigen is more of a headache than I thought on first sight.
There is no reshape() functionality for example (and conservativeResize is something else).
It also seems (I'd like to be corrected) to be the case that Map does not just offer a view on the data, but assignments to temporary variables seem to be required.
The minCoeff function after the colwise operator cannot return a minimum element and an index to that element.
It is unclear to me if replicate is actually allocating duplicates of the data. The reason behind broadcasting is that this is not required.
matrix_t data(2,4);
matrix_t means(2,2);
// data points
data << 1, 23, 6, 9,
3, 11, 7, 2;
// means
means << 2, 20,
3, 10;
std::cout << "Data: " << std::endl;
std::cout << data.replicate(2,1) << std::endl;
column_vector_t temp1(4);
temp1 = Eigen::Map<column_vector_t>(means.data(),4);
std::cout << "Means: " << std::endl;
std::cout << temp1.replicate(1,4) << std::endl;
matrix_t temp2(4,4);
temp2 = (data.replicate(2,1) - temp1.replicate(1,4));
std::cout << "Differences: " << std::endl;
std::cout << temp2 << std::endl;
matrix_t temp3(2,8);
temp3 = Eigen::Map<matrix_t>(temp2.data(),2,8);
std::cout << "Remap to 2xF: " << std::endl;
std::cout << temp3 << std::endl;
matrix_t temp4(1,8);
temp4 = temp3.colwise().squaredNorm();
std::cout << "Squared norm: " << std::endl;
std::cout << temp4 << std::endl;//.minCoeff(&index);
matrix_t temp5(2,4);
temp5 = Eigen::Map<matrix_t>(temp4.data(),2,4);
std::cout << "Squared norm result, the distances: " << std::endl;
std::cout << temp5.transpose() << std::endl;
//matrix_t::Index x, y;
std::cout << "Cannot get the indices: " << std::endl;
std::cout << temp5.transpose().colwise().minCoeff() << std::endl; // .minCoeff(&x,&y);
This is not a nice oneliner and seems overkill just to compare every column in data with every column in means and return a matrix with their differences. However, the versatility of Eigen does not seem to be such that this can be written down much shorter.