Solving Challenging Inequality $x^{\frac{p^2+p+1}{p^2}}-x^{\frac{p+1}{p}} < \log(p)$

86 Views Asked by At

This is a continuation of the problem asked here. I made some edits to the original problem, but enough changes were made to warrant a new question.

The goal is, given a value of $p$, find the maximum value of $x$ that satisfies $$ x^{\frac{p^2+p+1}{p^2}}-x^{\frac{p+1}{p}} < \log(p) $$ given $p\in\mathbb{Z}\geq 3$.

For purposes of usage, a lower bound of $x<x_{max}$ would be signifigantly more useful, although an upper bound may potentially be interesting.

In their answer, user Claude Leibovici showed that for sufficiently large $p$, $x_{max}\approx \frac{1}{2}p^2$.

Below is my work that I appended to the bottom of the original question. It provides a lower bound estimate that turns out to be signfigantly lower than the true value (but not too bad) $$ e^{W(p^{k+1}-1} < x_{max} $$


We can also think of this as the inequality of two continuous functions $$ f(x;k,p) = x^{\frac{p^{k+1}-1}{p^k(p-1)}} $$ $$ g(;a,k,p) = a + k\log(p). $$

Note that $a$ is a constant derived from the problem that guarantees that $f(x; 1,p) < g(;a,1,p)$ for a given value of $x$. The initial question formulation was asking to find the largest value of $x$ that satisfied the following inequality $$ f(x;2,p) - f(x;1,p) < g(;a,2,p) - g(;a,1,p) $$ with the goal of if we know a priori that $f(x;1,p) < g(;a,1,p) $ we can then imply that $f(x;2,p) < g(;a,2,p).$

We can differentiate both sides of the second inequality with respect to $k$ to get $$ x^{\frac{p^{k+1}-1}{p^k(p-1)}} \log(x)\left(\frac{\log(p)}{p^k(p-1)}\right) < \log(p) $$ for which we want to find a value of $x$ max that makes this inequality true. Rearranging we get that $$ x^{\frac{p^{k+1}-1}{p^k(p-1)}} \log(x) < p^k(p-1) $$ setting $x = e^z$ and just using some temp placeholder variables we can say that $$ e^{az}z < b $$ which has solution $$ z < W(ab) \implies \log(x) < W(p^{k+1}-1) $$ so our max value is $x=e^{W(p^{k+1}-1)}$

Since the derivative of the left hand side seems to be decreasing on $[1,2]$, is it sufficient to say that if $x$ is less than our bound here it should also be less at the next point?

This may just be me rambling but maybe it is helpful to the task

1

There are 1 best solutions below

0
On

For this answer, I shall refer to what I wrote for this previous question.

Considering the problem of the zero of function $$f(x)=x^a-x^b -\log(p) \quad \text{where} \quad a=\frac{p^2+p+1}{p^2}\quad \text{and} \quad b=\frac{p+1} p$$ $$f'(x)=a x^{a-1}-b x^{b-1} \quad \text{and} \quad f''(x)=a(a-1) x^{a-2}-b(b-1) x^{b-2}$$ I suggested as an initial estimate $$x_0=\frac{p^2\log (p)}{W\left(p^2 \log (p)\right)}$$ and Newton iterates will be $$x_{n+1}=x_n-\frac{x_n^a-x_n^b -\log(p) }{a x_n^{a-1}-b x_n^{b-1} }$$ This converges very fast.

Using as before for a test, $p=123.456$ and working with unlimited precision we have $$\left( \begin{array}{cc} n & x_n \\ 0 & 8150.36957189589 \\ 1 & 7636.04252044307 \\ 2 & 7634.23710338033 \\ 3 & 7634.23708015417 \\ 4 & 7634.23708015417 \end{array} \right)$$ which shows that we know the analytical solution of the equation.

What has been checked is that, as a function of $p$, $f(x_n)$, $f'(x_n)$ and $f''(x_n)$ are always positive. So, by Darboux theorem, $x_n$ is an overestimate of the solution. So, to satisfy the inequality, $\color{red}{\text{a lower bound is } \lfloor x_n\rfloor ~~\forall n \ge 2}$

Now, let us consider the implicit function $$g(x,p)=x^a-x^b -\log(p)=0$$ $$\frac{\partial g(x,p)}{\partial p}=\frac{x^b \log (x)}{p^2}-\frac{(p+2) x^a \log (x)}{p^3}-\frac{1}{p}$$ $$\frac{\partial g(x,p)}{\partial x}=a x^{a-1}-b x^{b-1}$$ and then, at solution, $$\frac{dx}{dp}=-\frac{\frac{\partial g(x,p)}{\partial p} } {\frac{\partial g(x,p)}{\partial x} }$$

Numerical analysis shows that $$\frac{dx_n}{dp}\sim p-2$$

For $n=3,4$ a quick and dirty linear regression $\frac{dx_n}{dp}=a+b\,p$ gives with $R^2=0.999977$ $$\begin{array}{clclclclc} \text{} & \text{Estimate} & \text{Standard Error} & \text{Confidence Interval} \\ a & -1.86836 & 0.05913 & \{-1.98572,-1.75099\} \\ b & +1.03026 & 0.00101 & \{+1.02825,+1.03227\} \\ \end{array}$$

Edit

A much better estimate of the solution is given by the first iterate of Halley method $$x_1=x_0+\frac{2 f(x_0) f'(x_0)}{f(x_0) f''(x_0)-2 f'(x_0)^2}$$ with the same $x_0$ as before.

For $p=123.456$, $x_1=7634.28$.

More spectacular would be the first iterate of Householder method $$x_1=x_0+\frac{3 f(x_0) \left(f(x_0) f''(x_0)-2 f'(x_0)^2\right)}{f(x_0)^2 f'''(x_0)+6 f'(x_0)^3-6 f(x_0) f'(x_0) f''(x_0)}$$

Update

Just out of curiosity, I searched the maximum value of the ratio $$R_n=\frac{x^a} {x^b +\log(p)}$$ for Newton successive iterates $$\left( \begin{array}{cc} n & R_n \\ 0 & 1.11721588695 \\ 1 & 1.01728777375 \\ 2 & 1.00039248588 \\ 3 & 1.00000020331 \\ 4 & 1.00000000076 \end{array} \right)$$