`l2Norm` and derivatives incorrect for complex valued matrices

I think the implementations for the L2 norm and variants are incorrect for complex valued matrices. This test code demonstrates the issues:

#include <iostream>
#include <fstream>

#include <complex>
#include <blaze/Blaze.h>

using namespace std::complex_literals;
using cfloat = std::complex<float>;

int main(int argc, char** argv)
{
    size_t n = 2;

    blaze::DynamicMatrix<cfloat, blaze::columnMajor> m(n, n);

    m = 1.0if;

    std::cout << blaze::sqrNorm(m) << std::endl; // Prints: (-4,0)
    std::cout << blaze::l2Norm(m) << std::endl; // Prints (1.50996e-07,2)
}

In this case the outputs should be 4 and 2 respectively. It is clear that a normal square, i.e. z*z is being used instead of |z|^2.

I think the problem is the specialisations found in blaze/math/expressions/DMatNormExpr.h (for example here) that replace the Abs with a Noop which is only a correct optimisation for real types.

One solution might be to replace the Pow2 call with a replacement Abs2 that is specialised appropriately for complex numbers. That would retain the optimisation for both real and complex numbers, but would have the appropriate behaviour.

I'd be happy to put in a PR if needed but I could do with some guidance as to where are the places I'd need to add something like Abs2 particularly in terms of the SIMD versions.

Comments (7)