previous overview next
Evidence once more
$
\def \EN {\quad \mbox{and} \quad}
\def \hieruit {\quad \Longrightarrow \quad}
$
In a previous section, called 'Quotient Function', evidence has been gathered
for the following theorem:
$$
\frac{b}{a}(dx) = e^{\, \left[ b/a \right]'(0) \: dx}
$$
I feel not entirely comfortable with the "proof" in this section, though. The
task could be re-formulated as follows: given $f(2.x) = f(x).f(x)$, continuous
in $x$, prove the theorem that: $f(x) = \exp(Px)$. It is clear that the theorem
is a sufficient condition for $f(2.x) = f(x).f(x)$ being true. But is it also a
necessary condition? Up to now, the proof has been accomplished by - sort of -
transfinite induction. Truth has been established for grid spacings approaching
zero - the domain of Calculus - and then, while working backwards, for finite
sized grid spacings, like in Numerical Analysis. In terms of Chaos Theory: the
theorem is proved for the Analytical Attractor in a/b(0) = 1 and then very much
extrapolated. Therefore trying to arrive at the same result via some other road
seems to be worthwile the effort.
The end-result from the section 'Direct Solver' is recalled in the first place:
$$
\left[ \begin{array}{ccccccc} . & . & . & & & & \\
& & . & . & . & & \\
& \displaystyle
- \frac{a_{54}}{a_{55}} \frac{a_{43}}{a_{44}}
& 0
& \displaystyle
1 - \frac{a_{54}}{a_{55}} \frac{a_{45}}{a_{44}}
- \frac{a_{56}}{a_{55}} \frac{a_{65}}{a_{66}}
& 0
& \displaystyle
- \frac{a_{56}}{a_{55}} \frac{a_{67}}{a_{66}}
& \\ & & & . & . & . &
\\ & & & & . & . & .
\end{array} \right]
\left[ \begin{array}{c} . \\ . \\ T_3 \\ T_4 \\ T_5 \\ T_6 \\ T_7 \\ .
\end{array} \right]
$$
Thus we see that the off-diagonal coefficients of the coarsened grid are given
by:
$$
a'_{53} = - \frac{a_{54}}{a_{55}} \frac{a_{43}}{a_{44}} \EN
a'_{57} = - \frac{a_{56}}{a_{55}} \frac{a_{67}}{a_{66}}
$$
Independent of the kind of mesh involved. Now imagine a piece of an 1-D mesh
again, but this time with a non-uniform grid spacing:
Where we have arranged it in such a way that the distance between the vertices
$(3)$ and $(4)$ is equal to the distance between the vertices $(6)$ and $(7)$;
and the distance between the vertices $(4)$ and $(5)$ is equal to the distance
between the vertices $(5)$ and $(6)$. Furthermore it is assumed that the matrix
coefficients are only dependent upon these two distances, called:
\begin{eqnarray*}
\mbox{distance }_{3-4} = \mbox{distance }_{6-7} = dx_1 \\
\mbox{distance }_{4-5} = \mbox{distance }_{5-6} = dx_2
\end{eqnarray*}
If matrix elements which are at the left of the main diagonal are denoted as
$a$ and matrix elements which are at the right of the main diagonal as $b$,
as has been done before, then it becomes readily evident that:
\begin{eqnarray*}
a_{43} = a(dx_1) \EN a_{54} = a(dx_2) \\
a_{67} = b(dx_1) \EN a_{56} = b(dx_2)
\end{eqnarray*}
Furthermore it follows that $a_{44} = a_{66}$ and:
$$
a'_{53} = a(dx_1 + dx_2) \EN a'_{57} = b(dx_1 + dx_2)
$$
Repeat:
$$
a'_{57} = - \frac{a_{56}}{a_{55}} \frac{a_{67}}{a_{66}} \EN
a'_{53} = - \frac{a_{54}}{a_{55}} \frac{a_{43}}{a_{44}}
$$
Hence the quotient of the off-diagonal coefficients of the coarsened grid can
also be written as follows:
$$
\frac{b(dx_1 + dx_2)}{a(dx_1 + dx_2)} =
\frac{b(dx_1)}{a(dx_1)} \frac{b(dx_2)}{a(dx_2)}
$$
Herewith it is established that the quotient function is characterized by:
$$
\frac{b}{a}(dx_1 + dx_2) = \frac{b}{a}(dx_1) \frac{b}{a}(dx_2)
$$
And we are triggered to pay some attention to functions which are characterized,
in general, by the following property:
$$
f(p+q) = f(p) . f(q)
$$
A couple of other properties can be derived herefrom:
$$
f(q) = f(q-p+p) = f(q-p) . f(p) \hieruit f(q-p) = f(q) / f(p)
$$ $$
\hieruit f(p-p) = f(p)/f(p) \hieruit f(0) = 1
$$
Employing the hypothesis that $f$ is a differentiable function and taking the
limit for $dx \rightarrow 0$, we see that:
$$
f'(x) = \frac{f(x+dx) - f(x)}{dx} = \frac{f(x).f(dx) - f(dx)}{dx} =
$$ $$
f(x) \frac{f(dx) - 1}{dx} = f(x) \frac{f(dx) - f(0)}{dx} = f'(0).f(x)
$$
Giving a differential equation for $f(x)$ which can be solved rather easily:
$$
\frac{df}{f} = f'(0).dx \hieruit ln(f) = f'(0).x + c
\hieruit f(x) = C e^{f'(0).x} = e^{f'(0).x}
$$
Where the boundary condition $f(0) = 1$ is employed in the last step. $f'(0)
= P$ is an arbitrary constant. The result is in agreement with our conjecture
that the only possible Quotient Functions are, indeed, given by: $a/b(dx) =
\exp(P.dx)$. If the above proof is valid, then this conjecture has been turned
into a theorem.