3
$\begingroup$

The standard Radial Basis Function:

$$K(x_1, x_2) = e^{-(x_1-x_2)^2/2}$$

is known to be positive definite (P.D.). This can be shown by, for example, the Bochner Theorem. By positive definite it is meant that

$$\sum_{i=1}^k \sum_{j=1}^k c_i c_j K(x_i, x_j) \geq 0$$

for any choice of $k \in \mathbb{N}$, $x_1, x_2, \ldots, x_k \in \mathbb{R}^n$ and $c_1, c_2, \ldots, c_k \in \mathbb{R}$.

From the above inequality, it seems obvious that the Hermitian part of the matrix with entries $$K_{ij}=K(x_i, x_j)$$ should have positive eigenvalues. Which brings me to the following problem:

G[x1_, x2_] := Exp[-1/2 (x1 - x2)^2];
list1 = RandomReal[{-10, 10}, 100];
(mat = Outer[G, list1, list1, 1]);
Eigenvalues[1/2 (mat + Transpose[mat])]

The resulting eigenvalues are not all positive. Why is that?

{14.4349, 12.5488, 11.8824, 10.0691, 9.54906, 8.65466, 7.27382, \
5.90905, 4.70692, 3.60972, 2.94933, 2.2864, 1.67143, 1.38664, \
0.926639, 0.694622, 0.479601, 0.346991, 0.218595, 0.144655, \
0.0947855, 0.0693128, 0.0390857, 0.0228503, 0.0137675, 0.0077189, \
0.00423406, 0.00209994, 0.00146088, 0.000639826, 0.000346808, \
0.000175915, 0.000081206, 0.0000427318, 0.0000218744, 0.000010557, 
5.11307*10^-6, 2.03378*10^-6, 1.00012*10^-6, 4.39773*10^-7, 
2.10151*10^-7, 7.53904*10^-8, 3.44437*10^-8, 1.22666*10^-8, 
4.87432*10^-9, 2.15116*10^-9, 7.33137*10^-10, 4.1995*10^-10, 
1.24422*10^-10, 4.71167*10^-11, 1.48485*10^-11, 6.61021*10^-12, 
2.78995*10^-12, 9.61269*10^-13, 2.65273*10^-13, 8.56566*10^-14, 
3.34234*10^-14, 1.14188*10^-14, 4.38365*10^-15, -2.17014*10^-15, 
2.05478*10^-15, 1.69751*10^-15, -1.34612*10^-15, 1.32415*10^-15, 
1.17034*10^-15, -1.09847*10^-15, 9.08927*10^-16, 
8.03778*10^-16, -7.81348*10^-16, -6.84237*10^-16, 
6.31285*10^-16, -5.8266*10^-16, 5.71942*10^-16, -4.67633*10^-16, 
4.53221*10^-16, 4.25156*10^-16, 
4.17792*10^-16, -3.92773*10^-16, -3.64044*10^-16, 3.39323*10^-16, 
3.18269*10^-16, -3.06691*10^-16, 2.9718*10^-16, -2.78096*10^-16, 
2.6519*10^-16, 2.12749*10^-16, -1.93748*10^-16, -1.81685*10^-16, 
1.74396*10^-16, 1.56957*10^-16, 
1.50008*10^-16, -1.42644*10^-16, -1.33609*10^-16, 
1.29917*10^-16, -1.0308*10^-16, 7.93279*10^-17, -7.7838*10^-17, 
4.323*10^-17, -1.54597*10^-17, 6.30094*10^-18}
$\endgroup$
  • $\begingroup$ The ones that are negative are numerically zero. See the docs of Chop. If you use numbers with more than machine precision, you will also see that the negative entries become as small as the precision can resolve. $\endgroup$ – Marius Ladegård Meyer Nov 24 '16 at 17:39
4
$\begingroup$

The worst ones are order 10^-15. This is effectively zero, so the negative values are only a result of mathematica working with finite numerical precision.

$\endgroup$
  • $\begingroup$ How can i control that precision. Suppose i want the results to be precise enough so that these small values are either zero or positive. $\endgroup$ – Iconoclast Nov 24 '16 at 17:59
  • 1
    $\begingroup$ reference.wolfram.com/language/howto/… $\endgroup$ – Wouter Nov 24 '16 at 20:30
  • $\begingroup$ Wouter, I would include your comment in the answer. It's really helpful. Thanks ;) $\endgroup$ – An old man in the sea. Aug 2 '17 at 20:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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