previous overview next
Trigonometric Connection II
$
\def \half {\frac{1}{2}}
\def \kwart {\frac{1}{4}}
\def \komma {\quad , \quad}
\def \OF {\quad \mbox{or} \quad}
\def \hieruit {\quad \Longrightarrow \quad}
$
A number of conclusions can be drawn by properly employing the formula:
$$
a.b = \frac{1}{2 + 2.\cos(k.\pi/2^N)} \quad \mbox{with} \quad k = 0,1.2, ... , 2^N
$$
The function reaches a minimum $1/4$ for the maximum value of the denominator,
which is $4$. Because the minimum of the denominator is zero, it also reaches to
infinity, meaning that dangerous points can be found at arbitrary density,
increasing with $N$, and everywhere in the region $1/4 < a.b < \infty$ .
It is concluded therefore that a region which is safe everywhere can only exist
in the interval $0 \le a.b \le 4$, as we have seen (but not proved) before.
However, not all points in the region $1/4 < a.b < \infty$ share the same amount
of risk. It seems reasonable to conjecture that points which are very close to
$1/4$ are less "dangerous" that others. It will be shown now that this is indeed
true. Points close to $1/4$ correspond with small angles. Hence any coarsening
of the grid will correspond with
the doubling of a small angle. The closer the point is to $1/4$, the smaller
the angle will be. And hence it will require quite some doubling effort, before
this point comes in real danger. To be more specific. Suppose that $a.b$ almost
equals $1/4$. We want to use a Taylor expansion here. Some hand-held calculus
can be avoided by devising an input for our favourite Computer Algebra System
(MAPLE):
series(1/(2+2*cos(x)),x);
$$
\kwart + \frac{1}{16} x^2 + \frac{1}{96} x^4 + O(x^6)
$$
In our case, suppose that we have a point $1/4 + \delta$, corresponding with
an angle $\pi.1/2^N$. Then:
$$
\kwart + \delta \approx \kwart + \frac{1}{16} (\pi/2^N)^2
$$
Solving for the angle:
$$
\pi/2^N \approx \sqrt{16.\delta} \hieruit
\ln(\pi) - N . \ln(2) \approx 0.5 \: \ln(16) + 0.5 \: \ln(\delta) \hieruit
$$ $$
N \approx 0.5 \: ^2\log(1/\delta) \quad + \quad ^2\log(\pi) - 2
$$
If, for example, the deviation from $1/4$ is one part per million $= 10^{-6}$,
then $\delta \approx 2^{-20} \rightarrow 0.5 \: ^2\log(1/\delta) \approx 10$.
This means that a grid coarsening less than $10$ times offers no real danger.
Which in turn means that the finest mesh can be allowed to consist of
$2^{10}\approx 1000$ points, without possibly running into big trouble.
What more can be said about the stationary points we have found in the past:
$$
\frac{1}{4} \komma 1 \komma \frac{3 \pm \sqrt{5}}{2}
$$
Or, solve for the angle $\phi$ in:
$$
a.b = \frac{1}{2 + 2.\cos(\phi)} = \quad \kwart \komma 1 \komma
\frac{3}{2} \pm \frac{\sqrt{5}}{2} $$
A useful remark being:
$$
\left( \frac{3}{2} + \frac{\sqrt{5}}{2} \right)
\left( \frac{3}{2} - \frac{\sqrt{5}}{2} \right) = 1
$$
Solve for the cosines first:
$$
\cos(\phi) = \quad 1 \komma - \half \komma \frac{ \pm \sqrt{5} - 1}{4}
$$
The first angle is just zero: $\phi = 0$ and any doubling or halving of this
angle maps it upon itself. Hence $1/4$ is a single stationary point, as we have
seen.
The second angle must be $120^0$. A doubling of this angle
gives $\cos(240^0) = -1/2$, which is the same value as with $\phi = 120^0$.
Hence $1$ is also a single stationary point, as we have seen.
The
3rd/4th angles are somewhat more complicated. The smallest of the two will
correspond with the plus sign. Doubling this one gives:
$$
\cos(2.\phi) = 2.\cos^2(\phi) - 1 =
2 \left( \frac{ + \sqrt{5} - 1}{4} \right)^2 - 1 =
\frac{5 - 2\sqrt{5} + 1}{8} - 1 = \frac{ - \sqrt{5} - 1}{4}
$$
Which is precisely the second one. Let's double it again:
$$
\cos(4.\phi) = 2.\cos^2(2.\phi) - 1 =
2 \left( \frac{ - \sqrt{5} - 1}{4} \right)^2 - 1 =
\frac{5 + 2\sqrt{5} + 1}{8} - 1 = \frac{ + \sqrt{5} - 1}{4}
$$
Giving back the first of the two. We find that the angles are determined by:
$$
\cos(4.\phi) = \cos(\phi) \hieruit 4.\phi = k.2.\pi \pm \phi \hieruit
$$ $$
\phi = k \frac{2.\pi}{5}
\quad \left( \OF \phi = k \frac{2.\pi}{3} \right) \quad
\qquad \mbox{where} \quad k = 1,2,4,8, \: ...
$$
The product-function values corresponding with the denominator $5$ are
stationary with a multiplicity of two: the first point gives the second point,
the second point gives back the first point, and so on and so forth.
It is questioned now whether there exist also points which have a multiplicity
of three or higher. The answer is affirmative. Take for example an angle which
gives the same function value again, but only after three times doubling itself:
$$
\cos(8.\phi) = \cos(\phi) \hieruit 8.\phi = k.2.\pi \pm \phi \hieruit
$$ $$
\phi = k \frac{2.\pi}{7} \OF \phi = k \frac{2.\pi}{9}
\qquad \mbox{where} \quad k = 1,2,4,8, \: ...
$$
We have found some clues for the behaviour of points in the area $(a.b) > 1/4$,
corresponding with angles $k.2.\pi/D$ and denominators $D=2,3,4,5,7,8,9, ...$.
A denominator like $6$ can be produced by $1/3=2/6$ and halving the latter. The
same story can be told for $10$ : $1/5=2/10$. Denominators $31$ and $33$ can
be produced by $2^5.\phi=k.2.\pi \pm \phi$, corresponding with multiplicities
$\le 5$. Then $3/33 = 1/11$ , providing material for angles $k.2.\pi/11$.
It is thus observed that, in the "dangerous" area, there also exist infinitely
many points which are not "dangerous" at all, since they will always be mapped
upon each other with any further grid coarsening. But, as has been suggested
earlier, this situation is similar to unstable equilibrium in physics,
like a pencil balancing on its tip.
Fractions $k/2^N$ are expanded in a computer as binary numbers. For example:
$$
0.00101001010011101010110101110110
$$
But also numbers like $1/3$ are represented, in a computer, as a finite binary
fraction:
$$
0.01010101010101010101010101010101
$$
Given a finite (though maybe very large) binary representation of any number in
the area $1/4 < a.b < \infty$, there is no sensible way to tell if this number
is predestined to be dangerous or not. This remains true, even if we ever would
have a computer at our disposal with immensely large words.
This story vaguely
reminds to the (in)famous controversy between Formalists (Hilbert) and
Intuitionists
(Brouwer). The formalist mathematicians would say that any point in our product
space is either dangerous or not dangerous. But the intuitionist mathematicians
would argue that the Law of the Excluded Middle is not universally valid.
It cannot be decided whether a point is dangerous or not. From a realistic point
of view, one might find that the intuitionists are merely right. Our conclusion
would be, then, that the behaviour of the products of real world off-diagonal
coefficients $(a.b)$ is, actually, unpredictable, if their values are to
be chosen in the domain $1/4 < (a.b) < \infty$.
We have also seen that there exists another, "safe" domain, which fortunately
is quite distinct from the dangerous area, namely $0 \le (a.b) \le 1/4$. The
boundary $1/4$ of this safe domain may even be defined with a sensible pinch
of salt. And it is expected that no weird mathematical problems will show up
for this area of interest.