4
$\begingroup$

The quartic form

$$(x^2 + y^2 + z^2)^2 - 3 ( x^3 y + y^3 z + z^3 x)$$

is non-negative for all real $x$, $y$, $z$, as one can check (with some effort). A theorem of Hilbert implies that there exist quadratic forms $Q_1$, $Q_2$, $Q_3$ so that

$$(x^2 + y^2 + z^2)^2 - 3( x^3 y + y^3 z + z^3 x) = Q_1^2 + Q_2^2 + Q_3^2$$

I would like to find an explicit writing of the quartic forms, with rational quadratic forms $Q_i$. Maybe more than $3$ terms are necessary.

|cite|improve this question
$\endgroup$
  • $\begingroup$ Why $3$ quadratic forms? What is special about $3$? $\endgroup$ – Rodrigo de Azevedo Sep 5 '17 at 15:55
  • $\begingroup$ In this case ( positive quartic forms) $3$ squares will do. I don't know how HIlbert proved his result. It is not valid for more variables or higher degree I suppose, there are counter examples. So $3$ is just the number that will do. $\endgroup$ – Orest Bucicovschi Sep 5 '17 at 15:59
  • $\begingroup$ I believe $2$ will do in this case. I will update my answer soon. $\endgroup$ – Rodrigo de Azevedo Sep 5 '17 at 16:14
  • $\begingroup$ @Rodrigo de Azevedo: That would be pretty nice then. So you can get a positive semidefinite matrix of rank $2$ in that subspace. Interesting. If you could just produce the matrix that would do it. $\endgroup$ – Orest Bucicovschi Sep 5 '17 at 16:18
12
$\begingroup$

$$(x^2 + y^2 + z^2)^2 - 3 ( x^3 y + y^3 z + z^3 x)=\frac{1}{2}\sum_{cyc}(x^2-y^2-xy-xz+2yz)^2=$$ $$=\frac{1}{6}\sum_{cyc}(x^2-2y^2+z^2-3xz+3yz)^2.$$

|cite|improve this answer
$\endgroup$
  • $\begingroup$ The first sum, is there some mistake since $-xy$ appears twice ? $\endgroup$ – Orest Bucicovschi Aug 30 '17 at 12:11
  • $\begingroup$ @orangeskid It was typo. See now please. By the way, if $Q_i$ are polynomials then your statement is wrong. $\endgroup$ – Michael Rozenberg Aug 30 '17 at 12:14
  • $\begingroup$ Excellent! Great answer! $\endgroup$ – Orest Bucicovschi Aug 30 '17 at 12:24
  • $\begingroup$ @orangeskid Thank you! Good luck! $\endgroup$ – Michael Rozenberg Aug 30 '17 at 12:25
  • 2
    $\begingroup$ @Jaideep Khare For the first solution we can write $\sum\limits_{cyc}(x^4+x^2y^2-3x^3y)=\sum\limits_{cyc}(ax^2+by^2+cz^2+dxy+exz+fyz)^2.$ Now, it's better to assume $a=-b$, $c=0$ and from here $d+e+f=0$ because the equality in $\sum\limits_{cyc}(x^4+x^2y^2-3x^3y)=0$ occurs for $x=y=z$. Now, we can write the system on $a$, $b$, $d$, $e$ and $f$ and playing with this system gives the answer. The second solution we can get by the similar way. $\endgroup$ – Michael Rozenberg Aug 30 '17 at 13:14
9
$\begingroup$

We are given a trivariate quartic form

$$ f (x, y, z) := (x^2 + y^2 + z^2)^2 - 3 ( x^3 y + y^3 z + z^3 x) $$

which can be rewritten as the sum of $9$ monomials

$$f (x, y, z) = x^4 + y^4 + z^4 + 2 x^2 y^2 + 2 x^2 z^2 + 2 y^2 z^2 - 3 x^3 y - 3 y^3 z - 3 z^3 x $$

and we would like to write it as the sum of squares (SOS) of quadratic forms. The fewer the better.

SOS decomposition

A general trivariate quartic form is parametrized as follows

$$\begin{array}{c} \begin{bmatrix} x^2\\ y^2\\ z^2\\ x y\\ x z\\ y z\end{bmatrix}^\top \begin{bmatrix} q_{11} & q_{12} & q_{13} & q_{14} & q_{15} & q_{16}\\ q_{12} & q_{22} & q_{23} & q_{24} & q_{25} & q_{26}\\ q_{13} & q_{23} & q_{33} & q_{34} & q_{35} & q_{36}\\ q_{14} & q_{24} & q_{34} & q_{44} & q_{45} & q_{46}\\ q_{15} & q_{25} & q_{35} & q_{45} & q_{55} & q_{56}\\ q_{16} & q_{26} & q_{36} & q_{46} & q_{56} & q_{66}\end{bmatrix} \begin{bmatrix} x^2\\ y^2\\ z^2\\ x y\\ x z\\ y z\end{bmatrix} = \\\\ = q_{11} x^4 + q_{22} y^4 + q_{33} z^4 + 2 q_{14} x^3 y + 2 q_{15} x^3 z + 2 q_{24} x y^3 + 2 q_{26} y^3 z + 2 q_{35} x z^3 + 2 q_{36} y z^3 + (2 q_{12} + q_{44}) x^2 y^2 + (2 q_{13} + q_{55}) x^2 z^2 + (2 q_{23} + q_{66}) y^2 z^2 + (2 q_{16} + 2 q_{45}) x^2 y z + (2 q_{46} + 2 q_{25}) x y^2 z + (2 q_{34} + 2 q_{56}) x y z^2\end{array}$$

Note that the $6 \times 6$ matrix above, which we denote by $\mathrm Q$, is symmetric by construction. Hence, we have $1 + 2 + \cdots + 6 = 21$ unknowns (instead of $6^2 = 36$). We recover the particular quartic $f$ if the following $15$ affine equality constraints are satisfied

$$\begin{array}{rl} q_{11} &= 1\\ q_{22} &= 1\\ q_{33} &= 1\\ 2 q_{14} &= -3\\ q_{15} &= 0\\ q_{24} &= 0\\ 2 q_{26} &= -3\\ 2 q_{35} &= -3\\ q_{36} &= 0\\ 2 q_{12} + q_{44} &= 2\\ 2 q_{13} + q_{55} &= 2\\ 2 q_{23} + q_{66} &= 2\\ 2 q_{16} + 2 q_{45} & = 0\\ 2 q_{46} + 2 q_{25} &= 0\\ 2 q_{34} + 2 q_{56} &= 0\end{array}$$

Note that we have $21 - 15 = 6$ degrees of freedom. These $15$ equality constraints can be written much more economically in the following form

$$\mathcal A (\mathrm Q) = \mathrm b$$

where $\mathcal A : \mbox{Sym}_6 (\mathbb R) \to \mathbb R^{15}$ is a linear function and $\mathrm b \in \mathbb R^{15}$.

Our goal is to find a (symmetric) positive semidefinite $6 \times 6$ matrix $\rm Q$ such that the $15$ equality constraints above are satisfied. If such a matrix $\rm Q$ exists, then there exists a ("square root") matrix $\rm R$ such that $\rm Q = R R^T$ and, thus, $f$ can be written as a SOS of quadratic forms.

Approximate solution

We would like matrix $\rm Q$ to be low-rank, in order to have as few terms in the SOS decomposition as possible. Hence, we have the following rank-minimization problem

$$\begin{array}{ll} \text{minimize} & \mbox{rank} (\mathrm Q)\\ \text{subject to} & \mathcal A (\mathrm Q) = \mathrm b\\ & \mathrm Q \succeq \mathrm O_6\end{array}$$

Unfortunately, the objective function, $\mbox{rank} (\mathrm Q)$, is non-convex. One popular heuristic is to minimize the nuclear norm of $\rm Q$, which is a convex proxy for $\mbox{rank} (\mathrm Q)$. Fortunately, the nuclear norm of a symmetric, positive semidefinite matrix is its trace, i.e., a linear function of the matrix's (diagonal) entries. Thus, we obtain the following semidefinite program (SDP)

$$\begin{array}{ll} \text{minimize} & \mbox{tr} (\mathrm Q)\\ \text{subject to} & \mathcal A (\mathrm Q) = \mathrm b\\ & \mathrm Q \succeq \mathrm O_6\end{array}$$

which is convex and, thus, computationally tractable. The following Python + NumPy + CVXPY script solves the SDP above:

from cvxpy import *
import numpy as np

np.set_printoptions(linewidth=125)

# matrix variable
Q = Semidef(6)

# objective function
objective = Minimize( trace(Q) )

# constraints
constraints = [   Q[0,0]            ==  1,
                  Q[1,1]            ==  1,
                  Q[2,2]            ==  1,
                2*Q[0,3]            == -3,
                  Q[0,4]            ==  0,
                  Q[1,3]            ==  0,
                2*Q[1,5]            == -3,
                2*Q[2,4]            == -3,
                  Q[2,5]            ==  0,
                2*Q[0,1] +   Q[3,3] ==  2,
                2*Q[0,2] +   Q[4,4] ==  2,
                2*Q[1,2] +   Q[5,5] ==  2,
                2*Q[0,5] + 2*Q[3,4] ==  0,
                2*Q[3,5] + 2*Q[1,4] ==  0,
                2*Q[2,3] + 2*Q[4,5] ==  0 ]

# create optimization problem
prob = Problem(objective, constraints)

# solve optimization problem
prob.solve()

print "Solution = \n", Q.value
print "Solution after rounding = \n", np.round(Q.value, 1)

# compute eigendecomposition of the solution
Lambda, V = np.linalg.eigh( Q.value )
print "Eigenvalues = \n", Lambda

This script produces the following output:

Solution = 
[[  1.00052427e+00  -4.93986993e-01  -4.93986993e-01  -1.50131075e+00   8.16263094e-04   1.48128400e+00]
 [ -4.93986993e-01   1.00052427e+00  -4.93986993e-01   8.16263094e-04   1.48128400e+00  -1.50131075e+00]
 [ -4.93986993e-01  -4.93986993e-01   1.00052427e+00   1.48128400e+00  -1.50131075e+00   8.16263095e-04]
 [ -1.50131075e+00   8.16263094e-04   1.48128400e+00   2.98994780e+00  -1.48087234e+00  -1.48087234e+00]
 [  8.16263094e-04   1.48128400e+00  -1.50131075e+00  -1.48087234e+00   2.98994780e+00  -1.48087234e+00]
 [  1.48128400e+00  -1.50131075e+00   8.16263095e-04  -1.48087234e+00  -1.48087234e+00   2.98994780e+00]]
Solution after rounding = 
[[ 1.  -0.5 -0.5 -1.5  0.   1.5]
 [-0.5  1.  -0.5  0.   1.5 -1.5]
 [-0.5 -0.5  1.   1.5 -1.5  0. ]
 [-1.5  0.   1.5  3.  -1.5 -1.5]
 [ 0.   1.5 -1.5 -1.5  3.  -1.5]
 [ 1.5 -1.5  0.  -1.5 -1.5  3. ]]
Eigenvalues = 
[ -3.66863824e-04   1.62150777e-03   1.62150779e-03   4.11202771e-02   5.96370990e+00   5.96370990e+00]

The SDP solver produces a matrix $\rm Q$ whose (numerical) rank is $2$. Note that one of the eigenvalues of $\rm Q$ is slightly negative due to numerical noise.

Exact solution

The approximate solution produced by the SDP solver suggests that the exact solution is

$$\mathrm Q = \begin{bmatrix} 1 & - \frac{1}{2} & - \frac{1}{2} & - \frac{3}{2} & 0 & \frac{3}{2}\\- \frac{1}{2} & 1 & - \frac{1}{2} & 0 & \frac{3}{2} & - \frac{3}{2}\\- \frac{1}{2} & - \frac{1}{2} & 1 & \frac{3}{2} & - \frac{3}{2} & 0\\- \frac{3}{2} & 0 & \frac{3}{2} & 3 & - \frac{3}{2} & - \frac{3}{2}\\0 & \frac{3}{2} & - \frac{3}{2} & - \frac{3}{2} & 3 & - \frac{3}{2}\\ \frac{3}{2} & - \frac{3}{2} & 0 & - \frac{3}{2} & - \frac{3}{2} & 3\end{bmatrix}$$

It is easy to verify that, indeed, this is the exact solution. Using Python + SymPy:

from sympy import *

# build Q matrix
U = Matrix([[    1,   -1,   -1,   -3,    0,    3],
            [    0,    1,   -1,    0,    3,   -3],
            [    0,    0,    1,    3,   -3,    0],
            [    0,    0,    0,    3,   -3,   -3],
            [    0,    0,    0,    0,    3,   -3],
            [    0,    0,    0,    0,    0,    3]])
Q = (U + U.T) / 2

print "Determinant = ", Q.det()
print "Rank        = ", Q.rank()
print "Eigenvalues = ", Q.eigenvals()

# compute Cholesky decomposition
L = Q.cholesky()

print "6x2 square root = ", L[:,[0,1]]


#####################################
# compute the exact SOS decomposition
#####################################

x, y, z = symbols('x y z')

# quartic form
f = (x**2 + y**2 + z**2)**2 - 3 * ( x**3 * y + y**3 * z + z**3 * x)

# quadratic forms
q1 = (L[:,0].T * Matrix([x**2, y**2, z**2, x*y, x*z, y*z]))[0,0]
q2 = (L[:,1].T * Matrix([x**2, y**2, z**2, x*y, x*z, y*z]))[0,0]

print "Error polynomial = ", simplify(q1**2 + q2**2 - f)

This script produces the following output:

Determinant =  0
Rank        =  2
Eigenvalues =  {0: 4, 6: 2}
6x2 square root =  Matrix([[   1,          0],
                           [-1/2,  sqrt(3)/2],
                           [-1/2, -sqrt(3)/2],
                           [-3/2, -sqrt(3)/2],
                           [   0,    sqrt(3)],
                           [ 3/2, -sqrt(3)/2]])
Error polynomial =  0

Thus, a SOS decomposition of $f$ is

$$\boxed{\begin{array}{rl} f (x,y,z) &= \left( x^{2} - \frac{y^{2}}{2} - \frac{z^{2}}{2} - \frac{3}{2} x y + \frac{3}{2} y z \right)^2 + \left( \frac{\sqrt{3}}{2} \left( y^{2} - z^{2} - x y + 2 x z - y z \right) \right)^2\\ &= \frac{1}{4} \left( 2 x^{2} - y^{2} - z^{2} - 3 x y + 3 y z \right)^2 + \frac{3}{4} \left( y^{2} - z^{2} - x y + 2 x z - y z \right)^2\end{array}}$$

References

|cite|improve this answer
$\endgroup$
  • $\begingroup$ Oh, great, so this is a SOS method, right? $\endgroup$ – Orest Bucicovschi Sep 3 '17 at 0:37
  • $\begingroup$ Yes, it is Parrilo's method. $\endgroup$ – Rodrigo de Azevedo Sep 3 '17 at 0:39
  • $\begingroup$ That is indeed impressive! Did not expect to have a solution in $2$ terms. Do you think that there are just finitely many solutions of rank $2$? $\endgroup$ – Orest Bucicovschi Sep 8 '17 at 16:41
  • $\begingroup$ The column space of $\rm Q$ is $2$-dimensional. Every orthonormal basis for this $2$-dimensional column space leads to a SOS decomposition with 2 terms. $\endgroup$ – Rodrigo de Azevedo Sep 8 '17 at 23:10

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.