SymEigen.h

// -*- c++ -*-

// Copyright (C) 2005,2009 Tom Drummond (twd20@cam.ac.uk)

//All rights reserved.
//
//Redistribution and use in source and binary forms, with or without
//modification, are permitted provided that the following conditions
//are met:
//1. Redistributions of source code must retain the above copyright
//    notice, this list of conditions and the following disclaimer.
//2. Redistributions in binary form must reproduce the above copyright
//   notice, this list of conditions and the following disclaimer in the
//   documentation and/or other materials provided with the distribution.
//
//THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND OTHER CONTRIBUTORS ``AS IS''
//AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
//IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
//ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR OTHER CONTRIBUTORS BE
//LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
//CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
//SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
//INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
//CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
//ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
//POSSIBILITY OF SUCH DAMAGE.

#ifndef __SYMEIGEN_H
#define __SYMEIGEN_H

#include <algorithm>
#include <iostream>
#include <cassert>
#include <cmath>
#include <utility>
#include <complex>
#include <TooN/lapack.h>

#include <TooN/TooN.h>

namespace TooN {
static const double root3=1.73205080756887729352744634150587236694280525381038062805580;

namespace Internal{

        using std::swap;

    static const double symeigen_condition_no=1e9;

    template <int Size> struct ComputeSymEigen {

        template<int Rows, int Cols, typename P, typename B>
        static inline void compute(const Matrix<Rows,Cols,P, B>& m, Matrix<Size,Size,P> & evectors, Vector<Size, P>& evalues) {

            SizeMismatch<Rows, Cols>::test(m.num_rows(), m.num_cols());  //m must be square
            SizeMismatch<Size, Rows>::test(m.num_rows(), evalues.size()); //m must be the size of the system
            

            evectors = m;
            FortranInteger N = evalues.size();
            FortranInteger lda = evalues.size();
            FortranInteger info;
            FortranInteger lwork=-1;
            P size;

            // find out how much space fortran needs
            syev_((char*)"V",(char*)"U",&N,&evectors[0][0],&lda,&evalues[0], &size,&lwork,&info);
            lwork = int(size);
            Vector<Dynamic, P> WORK(lwork);

            // now compute the decomposition
            syev_((char*)"V",(char*)"U",&N,&evectors[0][0],&lda,&evalues[0], &WORK[0],&lwork,&info);

            if(info!=0){
                std::cerr << "In SymEigen<"<<Size<<">: " << info
                        << " off-diagonal elements of an intermediate tridiagonal form did not converge to zero." << std::endl
                        << "M = " << m << std::endl;
            }
        }
    };

    template <> struct ComputeSymEigen<2> {

        template<typename P, typename B>
        static inline void compute(const Matrix<2,2,P,B>& m, Matrix<2,2,P>& eig, Vector<2, P>& ev) {
            double trace = m[0][0] + m[1][1];
            //Only use the upper triangular elements.
            double det = m[0][0]*m[1][1] - m[0][1]*m[0][1]; 
            double disc = trace*trace - 4 * det;
            
            //Mathematically, disc >= 0 always.
            //Numerically, it can drift very slightly below zero.
            if(disc < 0)
                disc = 0;

            using std::sqrt;
            double root_disc = sqrt(disc);
            ev[0] = 0.5 * (trace - root_disc);
            ev[1] = 0.5 * (trace + root_disc);
            double a = m[0][0] - ev[0];
            double b = m[0][1];
            double magsq = a*a + b*b;
            if (magsq == 0) {
                eig[0][0] = 1.0;
                eig[0][1] = 0;
            } else {
                eig[0][0] = -b;
                eig[0][1] = a;
                eig[0] *= 1.0/sqrt(magsq);
            }
            eig[1][0] = -eig[0][1];
            eig[1][1] = eig[0][0];
        }
    };

    template <> struct ComputeSymEigen<3> {

        template<typename P, typename B>
        static inline void compute(const Matrix<3,3,P,B>& m, Matrix<3,3,P>& eig, Vector<3, P>& ev) {
            //method uses closed form solution of cubic equation to obtain roots of characteristic equation.
            using std::sqrt;
            using std::min;
            using std::swap;
            
            double a_norm = norm_1(m);
            double eps = 1e-7 * a_norm;

            if(a_norm == 0)
            {
                eig = TooN::Identity;
                return;
            }

            //Polynomial terms of |a - l * Identity|
            //l^3 + a*l^2 + b*l + c

            const double& a11 = m[0][0];
            const double& a12 = m[0][1];
            const double& a13 = m[0][2];

            const double& a22 = m[1][1];
            const double& a23 = m[1][2];

            const double& a33 = m[2][2];

            //From matlab:
            double a = -a11-a22-a33;
            double b = a11*a22+a11*a33+a22*a33-a12*a12-a13*a13-a23*a23;
            double c = a11*(a23*a23)+(a13*a13)*a22+(a12*a12)*a33-a12*a13*a23*2.0-a11*a22*a33;

            //Using Cardano's method:
            double p = b - a*a/3;
            double q = c + (2*a*a*a - 9*a*b)/27;

            double alpha = -q/2;

            //beta_descriminant <= 0 for real roots!
            //force to zero to avoid nasty rounding issues.
            double beta_descriminant = std::min(0.0, q*q/4 + p*p*p/27);

            double beta = sqrt(-beta_descriminant);
            double r2 = alpha*alpha  - beta_descriminant;

            double cuberoot_r = pow(r2, 1./6);

            double theta3 = atan2(beta, alpha)/3;

            double A_plus_B = 2*cuberoot_r*cos(theta3);
            double A_minus_B = -2*cuberoot_r*sin(theta3);

            //calculate eigenvalues
            ev =  makeVector(A_plus_B, -A_plus_B/2 + A_minus_B * sqrt(3)/2, -A_plus_B/2 - A_minus_B * sqrt(3)/2) - Ones * a/3;

            if(ev[0] > ev[1])
                swap(ev[0], ev[1]);
            if(ev[1] > ev[2])
                swap(ev[1], ev[2]);
            if(ev[0] > ev[1])
                swap(ev[0], ev[1]);

            // for the vector [x y z]
            // eliminate to compute the ratios x/z and y/z
            // in both cases, the denominator is the same, so in the absence of
            // any other scaling, choose the denominator to be z and 
            // tne numerators to be x and y.
            //
            // x/z and y/z,  implies ax, ay, az which vanishes
            // if a=0
            //calculate the eigenvectors
            eig[0][0]=a12 * a23 - a13 * (a22 - ev[0]);
            eig[0][1]=a12 * a13 - a23 * (a11 - ev[0]);
            eig[0][2]=(a11-ev[0])*(a22-ev[0]) - a12*a12;
            eig[1][0]=a12 * a23 - a13 * (a22 - ev[1]);
            eig[1][1]=a12 * a13 - a23 * (a11 - ev[1]);
            eig[1][2]=(a11-ev[1])*(a22-ev[1]) - a12*a12;
            eig[2][0]=a12 * a23 - a13 * (a22 - ev[2]);
            eig[2][1]=a12 * a13 - a23 * (a11 - ev[2]);
            eig[2][2]=(a11-ev[2])*(a22-ev[2]) - a12*a12;
            
            //Check to see if we have any zero vectors
            double e0norm = norm_1(eig[0]);
            double e1norm = norm_1(eig[1]);
            double e2norm = norm_1(eig[2]);

            //Some of the vectors vanish: we're computing
            // x/z and y/z, which implies ax, ay, az which vanishes
            // if a=0;
            //
            // So compute the other two choices.
            if(e0norm < eps)
            {
                eig[0][0] += a12 * (ev[0] - a33) + a23 * a13;
                eig[0][1] += (ev[0]-a11)*(ev[0]-a33) - a13*a13;
                eig[0][2] += a23 * (ev[0] - a11) + a12 * a13;
                e0norm = norm_1(eig[0]);
            }

            if(e1norm < eps)
            {
                eig[1][0] += a12 * (ev[1] - a33) + a23 * a13;
                eig[1][1] += (ev[1]-a11)*(ev[1]-a33) - a13*a13;
                eig[1][2] += a23 * (ev[1] - a11) + a12 * a13;
                e1norm = norm_1(eig[1]);
            }

            if(e2norm < eps)
            {
                eig[2][0] += a12 * (ev[2] - a33) + a23 * a13;
                eig[2][1] += (ev[2]-a11)*(ev[2]-a33) - a13*a13;
                eig[2][2] += a23 * (ev[2] - a11) + a12 * a13;
                e2norm = norm_1(eig[2]);
            }


            //OK, a AND b might be 0
            //Check for vanishing and add in y/x, z/x which implies cx, cy, cz
            if(e0norm < eps)
            {
                eig[0][0] +=(ev[0]-a22)*(ev[0]-a33) - a23*a23;
                eig[0][1] +=a12 * (ev[0] - a33) + a23 * a13;
                eig[0][2] +=a13 * (ev[0] - a22) + a23 * a12;
                e0norm = norm_1(eig[0]);
            }

            if(e1norm < eps)
            {
                eig[1][0] +=(ev[1]-a22)*(ev[1]-a33) - a23*a23;
                eig[1][1] +=a12 * (ev[1] - a33) + a23 * a13;
                eig[1][2] +=a13 * (ev[1] - a22) + a23 * a12;
                e1norm = norm_1(eig[1]);
            }

            if(e2norm < eps)
            {
                eig[2][0] +=(ev[2]-a22)*(ev[2]-a33) - a23*a23;
                eig[2][1] +=a12 * (ev[2] - a33) + a23 * a13;
                eig[2][2] +=a13 * (ev[2] - a22) + a23 * a12;
                e2norm = norm_1(eig[2]);
            }

            //Oh, COME ON!!!
            if(e0norm < eps || e1norm < eps || e2norm < eps)
            {
                //This is blessedly rare
        
                double ns[] = {e0norm, e1norm, e2norm};
                double is[] = {0, 1, 2};
                
                //Sort them
                if(ns[0] > ns[1])
                {
                    swap(ns[0], ns[1]);
                    swap(is[0], is[1]);
                }
                if(ns[1] > ns[2])
                {
                    swap(ns[1], ns[2]);
                    swap(is[1], is[2]);
                }
                if(ns[0] > ns[1])
                {
                    swap(ns[0], ns[1]);
                    swap(is[0], is[1]);
                }


                if(ns[1] >= eps)
                {
                    //one small one. Use the cross product of the other two
                    normalize(eig[1]);
                    normalize(eig[2]);
                    eig[is[0]] = eig[is[1]]^eig[is[2]];
                }
                else if(ns[2] >= eps)
                {
                    normalize(eig[is[2]]);
                    
                    //Permute vector to get a new vector with some orthogonal components.
                    Vector<3> p = makeVector(eig[is[2]][1], eig[is[2]][2], eig[is[2]][0]);

                    //Gram-Schmit
                    Vector<3> v1 = unit(p - eig[is[2]] * (p * eig[is[2]]));
                    Vector<3> v2 = v1^eig[is[2]];

                    eig[is[0]] = v1;
                    eig[is[1]] = v2;
                }
                else
                    eig = TooN::Identity;
            }
            else
            {
                normalize(eig[0]);
                normalize(eig[1]);
                normalize(eig[2]);
            }
        }
    };

};

template <int Size=Dynamic, typename Precision = double>
class SymEigen {
public:
    inline SymEigen(){}

        inline SymEigen(int m) : my_evectors(m,m), my_evalues(m) {}

    template<int R, int C, typename B>
    inline SymEigen(const Matrix<R, C, Precision, B>& m) : my_evectors(m.num_rows(), m.num_cols()), my_evalues(m.num_rows()) {
        compute(m);
    }

    template<int R, int C, typename B>
    inline void compute(const Matrix<R,C,Precision,B>& m){
        SizeMismatch<R, C>::test(m.num_rows(), m.num_cols());
        SizeMismatch<R, Size>::test(m.num_rows(), my_evectors.num_rows());
        Internal::ComputeSymEigen<Size>::compute(m, my_evectors, my_evalues);
    }

    template <int S, typename P, typename B>
    Vector<Size, Precision> backsub(const Vector<S,P,B>& rhs) const {
        return (my_evectors.T() * diagmult(get_inv_diag(Internal::symeigen_condition_no),(my_evectors * rhs)));
    }

    template <int R, int C, typename P, typename B>
    Matrix<Size,C, Precision> backsub(const Matrix<R,C,P,B>& rhs) const {
        return (my_evectors.T() * diagmult(get_inv_diag(Internal::symeigen_condition_no),(my_evectors * rhs)));
    }

    Matrix<Size, Size, Precision> get_pinv(const double condition=Internal::symeigen_condition_no) const {
        return my_evectors.T() * diagmult(get_inv_diag(condition),my_evectors);
    }

    Vector<Size, Precision> get_inv_diag(const double condition) const {
        Precision max_diag = -my_evalues[0] > my_evalues[my_evalues.size()-1] ? -my_evalues[0]:my_evalues[my_evalues.size()-1];
        Vector<Size, Precision> invdiag(my_evalues.size());
        for(int i=0; i<my_evalues.size(); i++){
            if(fabs(my_evalues[i]) * condition > max_diag) {
                invdiag[i] = 1/my_evalues[i];
            } else {
                invdiag[i]=0;
            }
        }
        return invdiag;
    }

    Matrix<Size,Size,Precision>& get_evectors() {return my_evectors;}

    const Matrix<Size,Size,Precision>& get_evectors() const {return my_evectors;}


    Vector<Size, Precision>& get_evalues() {return my_evalues;}
    const Vector<Size, Precision>& get_evalues() const {return my_evalues;}

    bool is_posdef() const {
        for (int i = 0; i < my_evalues.size(); ++i) {
            if (my_evalues[i] <= 0.0)
                return false;
        }
        return true;
    }

    bool is_negdef() const {
        for (int i = 0; i < my_evalues.size(); ++i) {
            if (my_evalues[i] >= 0.0)
                return false;
        }
        return true;
    }

    Precision get_determinant () const {
        Precision det = 1.0;
        for (int i = 0; i < my_evalues.size(); ++i) {
            det *= my_evalues[i];
        }
        return det;
    }

    Matrix<Size, Size, Precision> get_sqrtm () const {
        Vector<Size, Precision> diag_sqrt(my_evalues.size());
        // In the future, maybe throw an exception if an
        // eigenvalue is negative?
        for (int i = 0; i < my_evalues.size(); ++i) {
            diag_sqrt[i] = std::sqrt(my_evalues[i]);
        }
        return my_evectors.T() * diagmult(diag_sqrt, my_evectors);
    }

    Matrix<Size, Size, Precision> get_isqrtm (const double condition=Internal::symeigen_condition_no) const {
        Vector<Size, Precision> diag_isqrt(my_evalues.size());

        // Because sqrt is a monotonic-preserving transformation,
        Precision max_diag = -my_evalues[0] > my_evalues[my_evalues.size()-1] ? (-std::sqrt(my_evalues[0])):(std::sqrt(my_evalues[my_evalues.size()-1]));
        // In the future, maybe throw an exception if an
        // eigenvalue is negative?
        for (int i = 0; i < my_evalues.size(); ++i) {
            diag_isqrt[i] = std::sqrt(my_evalues[i]);
            if(fabs(diag_isqrt[i]) * condition > max_diag) {
                diag_isqrt[i] = 1/diag_isqrt[i];
            } else {
                diag_isqrt[i] = 0;
            }
        }
        return my_evectors.T() * diagmult(diag_isqrt, my_evectors);
    }

private:
    // eigen vectors laid out row-wise so evectors[i] is the ith evector
    Matrix<Size,Size,Precision> my_evectors;

    Vector<Size, Precision> my_evalues;
};

}

#endif