# Tetration Forum

You're currently viewing a stripped down version of our content. View the full version with proper formatting.
Pages: 1 2

I suppose I ought to write up something more formal. My linear approximation is the same as the "standard" linear approximation for base e, but otherwise it's different. And it has a couple big advantages over the "standard" linear approximation.

First of all, it's $C^1$ continuous, meaning it's continuous and once differentiable for all real x. The "standard" linear approximation is only $C^0$ continuous for bases other than e.

More importantly, it helps highlight the fact that there would be an inflection point on the critical interval that I define, which helps expose more information about the slog (and tetration). I've always been partial to simple formulae that expose additional insights.

Anyway, I first discussed my linear approximation on Google groups, which is where I first met Gottfried Helms incidentally. (Which reminds me, Gottfried had put forth a guess as to where the inflection point would be located. I still haven't tried to determine the precise location of the infleciton point and whether it's consistent from base to base.)

So I went back and looked more closely at this post:

I had been focussing on tetration at the time, but we can easily make it work for the slog as well, at least for real bases greater than eta (greater than 1 with a caveat) and using a real domain.

For a given base $b$, let's find an integer constant $n$, defined by:

$\log_b^{\circ 2}(e) \lt \exp_b^{\circ n}(1) \le \log_b(e)$

Having found this constant, we can now define the $\mathrm{slog}_b(z)$ as:

$\mathrm{slog}_b(z) = \begin{cases}
\mathrm{slog}_b(b^z) - 1 & \text{if } z \le \log_b^{\circ 2}(e) \\
\vspace{2} \\
n + \frac{z-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)} & \text{if } \log_b^{\circ 2}(e) < z \le \log_b(e) \\
\vspace{5} \\
\mathrm{slog}_b(\log_b(z)) + 1 & \text{if } \log_b(e) < z
\end{cases}$

By the way, the caveat for bases between 1 and eta is that it gives us a linear approximation on the "corridor" between the upper and lower real fixed points. We would need to use complex numbers to generalize this outside this real interval, and my formula explicitly relies on use of real numbers. Therefore, rather than using 1 as a reference point for slog(z)=0, we would need to use a real number in the corridor. Using e is the simplest choice, though without looking at the complex slog, it's impossible to choose a "correct" reference point that corresponds with 1.
It's trivial to see that my linear approximation is continuous:

Let's start by examining the limit at $\log_b^{\circ 2}(e)$ from the right:

$
\begin{eqnarray}
{\Large \lim_{\normalsize z \to \log_b^{\circ 2}(e)^{+}}}\ \left[ \mathrm{slog}_b(z) \right]
& = & n + \frac{\log_b^{\circ 2}(e)-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)}
\end{eqnarray}
$

Then from the left:

$
\begin{eqnarray}
{\Large \lim_{\normalsize z \to \log_b^{\circ 2}(e)^{-}}}\ \left[ \mathrm{slog}_b(z) \right]
& = & \left. \mathrm{slog}_b\left(b^z\right) - 1\right|_{z=\log_b^{\circ 2}(e)} \\
& = & \mathrm{slog}_b\left(\log_b(e)\right) - 1 \\
\vspace{4} \\
& = & n + \frac{\log_b(e)-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)} - \frac{\log_b(e)-\log_b^{\circ 2}(e)}{\log_b(e)-\log_b^{\circ 2}(e)}\\
\vspace{4} \\
& = & n + \frac{\log_b^{\circ 2}(e)-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)} \\
\end{eqnarray}
$

The limits at the other end of the interval are similarly simple to demonstrate.
More importantly, we can demonstrate continuity of the first derivative:

As before, we can examine the limit at $\log_b^{\circ 2}(e)$ from the right:

$
\begin{eqnarray}
{\Large \lim_{\normalsize z \to \log_b^{\circ 2}(e)^{+}}}\ {\Large D_{\normalsize z}} \left[ \mathrm{slog}_b(z) \right]
& = & {\Large D_{\normalsize z}} \left[ n + \frac{z-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)} \right]_{z=\log_b^{\circ 2}(e)} \\
& = & \left. \frac{1}{\log_b(e)-\log_b^{\circ 2}(e)} \right|_{z=\log_b^{\circ 2}(e)} \\
& = & \frac{1}{\log_b(e)-\log_b^{\circ 2}(e)}
\end{eqnarray}
$

And then from the left:

$
\begin{eqnarray}
{\Large \lim_{\normalsize z \to \log_b_{\circ 2}(e)^{-}}}\ {\Large D_{\normalsize z}} \left[ \mathrm{slog}_b(z) \right]
& = & {\Large \lim_{\normalsize z \to \log_b^{\circ 2}(e)^{-}}}\ {\Large D_{\normalsize z}} \left[ \mathrm{slog}_b(b^z) - 1\right] \\
& = & {\Large D_{\normalsize z}} \left[ n - 1 + \frac{b^z-\exp_b^{\circ n}(1)}{\log_b(e)-\log_b^{\circ 2}(e)} \right]_{z=\log_b^{\circ 2}(e)} \\
& = & \left. \frac{b^z \ln(b)}{\log_b(e)-\log_b^{\circ 2}(e)} \right|_{z=\log_b^{\circ 2}(e)} \\
& = & \frac{\frac{\log_b(e)}{\log_b(e)}}{\log_b(e)-\log_b^{\circ 2}(e)} \\
& = & \frac{1}{\log_b(e)-\log_b^{\circ 2}(e)}
\end{eqnarray}
$

As before, limits at the other end of the interval can also be demonstrated.
You can just solve for n explicitly as $n = \lfloor \text{slog}_b(e)\rfloor - 1$, since you can write your inequality that defines n as
$\exp_b^{-2}(e) < \exp_b^{n}(1) \le \exp_b^{-1}(e)$
$\exp_b^{-1}(e) < \exp_b^{(n+1)}(1) \le \exp_b^{0}(e)$
or ${}^{(n+1)}b = \exp_b^{(n+1)}(1) = \exp_b^a(e)$ where $-1 < a \le 0$. Taking the super-logarithm we get $n+1 = \text{slog}_b(\exp_b^a(e))$ which is a decreasing function of b, so if we take the floor: $n+1 = \lfloor \text{slog}_b(\exp_b^0(e)) \rfloor = = \lfloor \text{slog}_b(e) \rfloor$ so $n = \lfloor \text{slog}_b(e)\rfloor - 1$.

It is interesting to note that both the floor of slog and the ceiling of slog can be defined independently of the function itself. This means if you do use this to define n then there is no chance of circular definitions

Andrew Robbins

I suppose I ought to write up something more formal. My linear approximation is the same as the "standard" linear approximation for base e, but otherwise it's different. And it has a couple big advantages over the "standard" linear approximation.

First of all, it's $C^1$ continuous, meaning it's continuous and once differentiable for all real x. The "standard" linear approximation is only $C^0$ continuous for bases other than e.

More importantly, it helps highlight the fact that there would be an inflection point on the critical interval that I define, which helps expose more information about the slog (and tetration). I've always been partial to simple formulae that expose additional insights.

Anyway, I first discussed my linear approximation on Google groups, which is where I first met Gottfried Helms incidentally. (Which reminds me, Gottfried had put forth a guess as to where the inflection point would be located. I still haven't tried to determine the precise location of the infleciton point and whether it's consistent from base to base.)
....
I discovered the same critical section, which must contain the inflection point, for base n, which follows the general definition that t(x) = n^(t(x-1)). The critical section is between x-1 and x, with the following equations. t refers to the tetration estimation function for base n.

t(x-1) = -ln(ln(n))/ln(n)
t(x) = 1/ln(n)

Note, t(x+1) = e, for all base n values. Between x and x+1, the value of t(x) may be estimated by n^t(x-1), where t(x-1) is a line. For base e, the critical section is between t(-1) and t(0), where a linear estimate at x=0 produces

t(-1) = -ln(ln(e))/ln(e) = 0
t(0) = 1/ln(e) = 1

And t(1)=e. We know it contains the inflection point because one can show (with a bit of algebra), that the first derivative at t'(x) is equal to the first derivative at t'(x-1). Start from the definition of tetration, and take the derivative.

t(x) = n^t(x-1)
t'(x) = n^(t(x-1))*ln(n)*t'(x-1)

Next, we are interested specifically in t(x)=1/ln(n). we can substitute this in and get
t'(x) = t'(x-1), if t(x)=1/ln(n)

There are many equations that have the same first derivative at t'(x) and t'(x-1), but the simplest is a line, which has the same first derivative on either side. Other solutions for curves from t(x-1) to t(x) can be found that have smooth second order derivatives at t(x-1) and t(x), but this requires a 3rd order polynomial equation instead of a linear equation. I got as far as deriving the general solution for a fourth order equation that has smooth 1st derivatives, 2nd derivatives, and 3rd derivatives, for a base(n) estimate over the critical region. The coefficients require solving a quadratic equation. For a 5th order estimate that is 4th order smooth, the non-linear equations get hopelessly complex, though I may attempt a numerical estimate.

Andrew Robbins has done much better by solving the estimation for the hyper log function, which is linear, as opposed to the tetration function, which is non-linear.

This is not formally worded, but all sorts of questions arise. Where is the inflection point? The inflection point is the second derivative. In general, one can hypothesize that all even derivatives (2nd, 4th, 6th etc), will probably have an inflection point, and will go from -infinity at t(-2) to a critical section at t(-1) to t(0) containing the inflection point. to t(1), t(2), t(3).... then going off to +infinity.

Continuing on the hypothesis, the odd derivative ought to have minimum values in the -1 to 0 range, going off to infinity on either side.

My estimates give results close to Andrew's solutions. One reason why I solved the "3rd order" smooth derivative versions of these equations, is that the smooth 1st/2nd order derivative equations are mirror image symmatrical about the critical section, with the inflection point exactly in the middle of the critical section. A 4th order equation with a smooth 3rd derivative is no longer symmetrical, so it give more insight into the true nature of the curve. I developed the general base=n equations, as an attempt to understand the more general patterns of the tetration curve.

3rd order estimates for e^^0.5
1.646354229 from Andrew Robbin's tetration paper
1.644772458 -0.09% base n=e 3rd derivative smooth estimate using equations below
1.648721271 +0.14% baes n=e 1st/2nd derivative smooth estimate

Here are my base=e equations, for a tetration estimate, t(y) = a0 + a1*(x+0.5) + a2*(x+0.5)^2 + a3*(x+0.5)^3 + a4*(x+0.5)^4, where x is in the range (-1 to 0), where t(1)=e, t(0)=1, and t(-1)=0

a0=0.49760
a1=0.94949
a2=0.01918
a3=0.20204
a4=-0.03837

I can also post the exact equations used to generate these coefficients if anyone is interested.

For my 3rd derivative smooth estimate for base 10, I get 10^^1.38748975 ~= 100, versus Andrew's result of 1.39233. The error is 0.35% which seems a little higher than I would've expected.

For base n=10. t(y) = a0 + a1*(x+0.96997) + a2*(x+0.96997)^2 + a3*(x+0.96997)^3 + a4*(x+0.96997)^4. T(y) is a tetration estimate in the range (-1.46997 to -0.46997), that has smooth first, second, and third derivatives. If you plug -1 into this estimate, you get t(-1)=0, which is how I figured out the 0.96997 normalization value.

t(0.46997+1)=e,
t(-0.46997)=-1/ln(10) = 0.43429
t(-1.46997)=-ln(ln(10))/ln(10) = -0.362216

a0=0.02105
a1=0.70413
a2=0.11992
a3=0.36953
a4=-0.23985

A 3rd order equation with a smooth second derivative has a2=0 and a4=0. Interestingly, the a1 and a3 values are not changed between the 2nd derivative smooth results and the 3rd derivative smooth results. This is true for all values of n. A second order smooth equation, centered on the critical section is always of the form
t(n) = a0 +a1x +a3*x^3, with the same a1 and a3 values for the 3rd order smooth result. a0 does change, and the normalization range changes by a little as well.
Thats indeed an interesting approach to extend this linear approximation on a unit interval to a polynomial approximation there.
It would think that this approach is different from the approach of approximating the super exponential by polynomials in a point (instead of the unit interval) such that the first some derivations are smooth.
http://math.eretrandre.org/tetrationforu...php?tid=34 however didnt really come to conclusions.

I have also another question: Do your computations lead to the same interval as for the linear approximation, or is it different? And yes I would like to see some more equations
bo198214 Wrote:Thats indeed an interesting approach to extend this linear approximation on a unit interval to a polynomial approximation there.
It would think that this approach is different from the approach of approximating the super exponential by polynomials in a point (instead of the unit interval) such that the first some derivations are smooth.
http://math.eretrandre.org/tetrationforu...php?tid=34 however didnt really come to conclusions.

I have also another question: Do your computations lead to the same interval as for the linear approximation, or is it different? And yes I would like to see some more equations
Thanks for your reply. I've been browsing the forum's more here, and there's a lot of interesting posts! This is my first attempt at using tex.

My approximation equations use sexp instead of slog.

The resulting equations are non-linear, instead of linear equations. Therefore, the approximation is limited to a fourth order equation, instead of an "nth" order equation.

The approximation equation is centered on the critical section. For base(e) the approximation is centered at x= -0.5 instead of at x=0 or at x=-1. For the general base(n), the approximation is centered about the critical unit section where first derivatives are equal on either side of the unit section.

Equations to solve for a0..a4, so that the 1st, 2nd, and 3rd derivatives are smooth. First, solve for the "odd" terms, a1 and a3

$f(x) = a0 + a1*x + a2*x^2 + a3*x^3 + a4*x^4$

Starting with the definition of sexp for base n
$f(x)= n^{f(x-1)}$
$f'(x)= n^{f(x-1)}*\ln(n)*f'(x-1)$

The approximation is centered around x=0, for the range x[-0.5 to +0.5]. Later, the normalization coefficent will be estimated so that in the critical section, the following approximation holds.
$\text{sexp}(x+\text{norm})$ ~= $f(x)$

Next, continue by taking the first derivative.
$f'(0.5) = n^{f(-0.5)}*\ln(n)*f'(-0.5)$
if we set $n^{f(-0.5)}$ equal to $1/\ln(n)$, then we have $f'(0.5)=f'(-0.5)$
$f(-0.5) = -\ln(\ln(n))/\ln(n)$
$f(+0.5) = 1/\ln(n)$
$f(+1.5) = e$

Continue by taking the 2nd and 3rd derivatives. The f''' third derivative equations are used to calculate a2 and a4.
$f''(x) = n^{f(x-1)}*\ln(n)*f''(x-1) + n^{f(x-1)}*\ln(n)*\ln(n)*(f'(x-1))^{2}$
$f''(0.5) = f''(-0.5) + \ln(n)*(f'(-0.5))^2$
$f'''(x) = n^{f(x-1)}*\ln(n)*f'''(x-1) + n^{f(x-1)}*\ln(n)*\ln(n)*\ln(n)*(f'(x-1))^{3} + n^{f(x-1)}*\ln(n)*\ln(n)*3*f'(x-1)*f''(x-1)$
$f'''(0.5) = f'''(-0.5) + \ln(n)*3*f'(-0.5)*f''(-0.5) + \ln(n)*\ln(n)*(f'(-0.5))^{3}$

Now calculate the difference between f(+0.5) and f(-0.5). This difference is assigned to the shorthand constant "m", which is equal to the difference between f(0.5) and f(-0.5).
$m = (1 + \ln(\ln(n))) / \ln(n)$
$f(+0.5) = a0 + a1/2 + a2/4 + a3/8 + a4/16$
$f(-0.5) = a0 - a1/2 + a2/4 - a3/8 + a4/16$
$f(+0.5) - f(-0.5) = m = a1 + a3/4$
$a3 = 4*(m-a1)$

Next, average f'(+0.5) and f'(-0.5), noting that the two are equal, and using the shorthand value "p" for the first derivative of f(x) at +/- 1/2.
$f'(-0.5) = a1 - a2 + 3a3/4 -a4/2$
$f'(+0.5) = a1 + a2 + 3a3/4 + a4/2$
$(f'(+0.5) + f'(-0.5))/2 = f'(+0.5)$
$(f'(+0.5) + f'(-0.5))/2 = a1 + (3*a3)/4$
$p = f'(-0.5) = f'(+0.5) = a1 + (3*a3)/4$

Now we calculate the difference between f''(+0.5) and f''(-0.5), and then substitute in for p from the equation just derived.
$f''(+0.5) = 2*a2 + 3*a3 + 3*a4$
$f''(-0.5) = 2*a2 - 3*a3 + 3*a4$
$f''(+0.5) - f''(-0.5) = 6*a3$
$f''(+0.5) - f''(-0.5) = 6*a3 = \ln(n)*p*p$
$6*a3 = \ln(n)*( a1 + 3a3/4)^{2}$

Now, substitute in from a3 = 4*(m-a1), and then with a little algebra, generate the quadratic equation for a1. Solve the quadratic equation for a1 in terms of the coefficients, a, b, and c. Then solve for a3 in terms of a1.
$24(m-a1) = \ln(n)*(a1 + 3(m-a1))^{2}$
$a = 4*\ln(n)$
$b = (24 - 12*\ln(n)*m)$
$c = 9*m*m*\ln(n)-24m$
$a1 = -(b + sqrt{b^{2} - 4ac})/2a$
$a3 = 4*(m-a1)$

The interesting thing about this result is that the value for a1 and a3 are the same for a 3rd order approximation, as they are for a 4th order approximation. I will post the equations for a0, a2, and a4 next.

Moderators Note: Changed ln to \ln in tex code, this lets ln appear non-italic as it should be. Looks much nicer
bo198214 Wrote:.... And yes I would like to see some more equations

There are three more unknown coefficients to calculate, based on three equations. First, the average of f(+0.5,-0.5) is calculated in terms of the coefficients, and in terms of the boundary definition. Second, the difference between f'(+0.5,-0.5) is calculated in terms of the coefficients, and is also know to be zero due to the boundary definition. Third, the difference between f'''(+0.5,-0.5) is calculated in terms of the coefficients. That difference can also be calculated in terms of the other derivatives and boundary conditions.

There are also important values used in this section that have already been defined, such as f(0.5), f(-0.5), a1, a3, and p from my previous post.
$f(+0.5) = 1/\ln(n)$
$f(-0.5) = -\ln(\ln(n))/\ln(n)$
$m = (1 + \ln(\ln(n))) / \ln(n)$
$ave = (f(-0.5) + f(0.5))/2 = 1/2(-\ln(\ln(n)) / \ln(n)) + 1/2(1/\ln(n))$
$ave = (1-\ln(\ln(n))/(2*\ln(n))$
$ave = a0 + a2/4 + a4/16$
$f'(+0.5) - f'(-0.5) = 0 = 2a2 + a4$
$a4 = -2*a2$
$ave = a0 - a2/8$
$f''(-0.5) = (-3*a3 + 2*a4)$

At this point, it is worth noting that if a third order estimate were used, there would not be an a4 term, and therefore a4=0 and a2=0. For that third order approximation, it is then trivial to determine that a0 is equal to "ave", which was just calculated. Continuing on, calculate the difference between f'''(0.5) and f'''(-0.5), in terms of the a0..a4 coefficients, and then substitute in the known values for f'(x)=p, and substitute in the equation for a4 in terms of a2, and substitute in the equation for f''(-0.5) in terms of a3, a4.
$f'''(+0.5) - f'''(-0.5) = 24*a4 = \ln(n)*3*p*f''(-0.5) + \ln(n)*\ln(n)*p^{3}$
$24*a4 = \ln(n)*3*p*f''(-0.5) + \ln(n)*\ln(n)*p^3$
$24*a4 = \ln(n)*3*p*(-3*a3 + 2*a4) + \ln(n)*\ln(n)*p^{3}$
$24*a4 - \ln(n)*3*p*2*a4 = \ln(n)*3*p*(-3*a3) + \ln(n)*\ln(n)*p^{3}$
$24*a4 - \ln(n)*p*6*a4 = -\ln(n)*9*p*a3 + \ln(n)*\ln(n)*p^{3}$
$a4*(24 - \ln(n)*p*6) = -\ln(n)*9*p*a3 + \ln(n)*\ln(n)*p^{3}$
$a4 = (-\ln(n)*9*p*a3 + \ln(n)*\ln(n)*p^{3}) / (24 - \ln(n)*6*p$

Next, use the value calculated for a4 to generate a2, and a0, using the value of ave in terms of the ln(n), from above.
$a2 = -a4/2$
$a0 = ave - (a2/" align="middle" />

Finally, the normalization value must be approximated by iteration to determine sexp(x+norm) ~= f(x). Find the value of x where f(x) has a known defined value of sexp, either a value 0, or a value of 1, corresponding to sexp(-1) or sexp(0).

I calculated the normalization values for 2,e,3, and 10 in the following spread sheet for 3rd and 4th order approximations.
Let me summarize, we have a polynomial $p$ in some interval $[x_0-1,x_0)$. (For for being unambiguous I will use the variable $b$ for the base instead of $n$ which is usually reserved for natural numbers.)
$p(x) = a_0 + a_1*x + a_2*x^2 + a_3*x^3 + a_4*x^4$

There are 5 constants and $x_0$ to be determined. To do that we consider 4 equations $f^{(n)}(x_0)=(b^{f(x)})^{(n)}(x_0-1)$, $n=0,\dots,3$ plus $f'(x_0)=f'(x_0-1)$. These are (for simplicity and that I can reuse your text I set $x=x_0$ and $f=p$):

(0) $f(x)= b^{f(x-1)}$
(1) $f'(x)= b^{f(x-1)}*\ln(b)*f'(x-1)=f(x)\ln(b)f'(x-1)$
(*) $f'(x)=f'(x-1)$
(2) $f''(x) = b^{f(x-1)}*\ln(b)*f''(x-1) + b^{f(x-1)}*\ln(b)*\ln(b)*(f'(x-1))^{2}$
(3) $f'''(x) = b^{f(x-1)}*\ln(b)*f'''(x-1) + b^{f(x-1)}*\ln(b)*\ln(b)*\ln(b)*(f'(x-1))^{3} + b^{f(x-1)}*\ln(b)*\ln(b)*3*f'(x-1)*f''(x-1)$

From (0), (1) and (*) we get
$f(x) = 1/\ln(b) = \log_b(e)$

As we have 6 values to be determined (5 constants and $x_0$) but only have 5 equations, we need another equation.

But as we know the values $p(x_0-1)=\log_b(\log_b(e))$ and $p(x_0)=\log_b(e)$ independently from $x_0$ we just search for the $n$ such that $f(n)=\exp_b^n(1)$ is inside the interval $[p(x_0-1),p(x_0))$, there can only be on such $n$. And then the additional equation
$p(n)=\exp_b^n(1)$
now completely determines $a_0,\dots,a_4,x_0$.

This is the approach that I would generalize from JayD's linear approximation considering your explanations. Is it how you calculate your values? Particularly I dont see why you need the special values -0.5 and 0.5.

And writing this, I wonder whether it would be more straight forward, to take the first 4 derivations (instead of now 3) into account, instead of introducing equation (*)? But then perhaps $p(x_0)$ would depend on $x_0$ and we could not find an appropriate $n$, who knows.
Thank you for your comments. I generated my approximations before seeing "http://math.eretrandre.org/tetrationforum/archive/index.php/thread-33.html" or anything else on this forum. Given that these are the very first posts I have ever made to a math forum, and given that its 30 years since I took calculus, I think I did ok!

The critical section is interesting because it is where the point of inflection in the tetration curve is. The critical section is almost linear, and the inflection points position is the most visible feature of this section of curve. That's what motivated the boundary condition for f'(x) = f'(x-1); also it makes the equations simpler. Secondly, the inflection point seems like a universal section of the curve, that applies to any tetration base larger than e^(1/e). The approximation was verified for base 2, e, 3, and base 10.

Centering the approximation around +/- 0.5 was to center the taylor series radius of convergence on the unit length being approximated. A finite taylor series doesn't have a radius of convergence, so it may not matter. Centering the series on a linear section may help with numerical approximations for a higher order derivatives.

The inflection point still seems like an important feature of the tetration curve. Jay's graphs (in the link above), show the inflection points, which is the minimum of the odd derivatives, gradually moving from the ~= -0.5 range to the +0.35 range for higher even derivatives. It would be interesting to see a table of the "x" coordinate of the inflection points, along with the slope of the tetration curve at the point of inflection for various even derivatives.
Pages: 1 2