Hello all, this question is a variant (and probably a more difficult one) of a (promptly answered ) question that I asked here, at Is it true that all the "irrational power" functions are almost polynomial ?.

For $n\geq 1$, let $f(n)$ denote the "integer part" (largest integer below) $n^{\frac{3}{2}}$, and let $g(n)=f(n+2)-2f(n+1)+f(n)$. Question : Is it true that $g(n)$ is always in $\lbrace -1,0,1\rbrace$ (excepting the initial value $g(1)=2$) ? I checked this up to $n=100000$.

It is not too hard to check that if $t$ is large enough compared to $r$ (say $t\geq \frac{3r^2+1}{4}$), $f(t^2+r)$ is exactly $t^3+\frac{3rt}{2}$ (or $t^3+\frac{3rt-1}{2}$ if $t$ and $r$ are both odd ) and similarly $f(t^2-r)$ is exactly $t^3-\frac{3rt}{2}$ (or $t^3-\frac{3rt+1}{2}$ if $t$ and $r$ are both odd ). So we already know that the answer is "yes" for most of the numbers.

  • $\begingroup$ What have you worked out about the analogous non-discretized question? That is, if you put $h(x) = (x+2)^{3/2} - 2(x+1)^{3/2} + x^{3/2}$, can you show that $h$ is bounded by small constants? This would be my first attempt to get a handle on the problem; if it works, then you can examine more closely what happens upon taking "floors". (I don't think the problem has anything really to do with number theory per se.) $\endgroup$
    – Yemon Choi
    Mar 15, 2010 at 6:37
  • $\begingroup$ @ Yemon : Bjorn Ponnen's method (explained in his answer to my previous question) yields the bound $|h(x)|\leq \frac{c}{\sqrt{x}}$, but it is not clear at all how to deduce a result on the floors from this, apart from the easy (and insufficient) bound $|g(n)| \leq 4+ \frac{c}{\sqrt{n}}$. I added the "number theory" tag beacuse I encountered this problem when working on Hall's conjecture on Mordell's equation $\endgroup$ Mar 15, 2010 at 7:36
  • 1
    $\begingroup$ It's still insufficient, but I think you can easily do better than 4. If [x] denotes the floor of x, then (x-1)-2y+(z-1) < [x]-2[y]+[z] < x -2(y-1) + z, which shows that |g(n)-h(n)|<2. $\endgroup$ Mar 15, 2010 at 7:45
  • 5
    $\begingroup$ One difficulty is that it seems that this would be false over the reals, e.g., for n=20.54 or n=112.675. $\endgroup$ Mar 15, 2010 at 8:01
  • 2
    $\begingroup$ I don't think the range of t and r you gave in the question cover most of the cases. In fact, the asymptotic density of that set appears to be zero. $\endgroup$
    – S. Carnahan
    Mar 15, 2010 at 9:01

2 Answers 2


Here is a proof that $|g(n)|\le 1$ for all but finitely many $n$. You can extract an explicit bound for $n$ from the argument and check the smaller values by hand.

If $f(n)=n^{3/2}$ without the floor, then $g(n)\sim \frac{3}{4\sqrt n}$, so it is positive and tends to 0. When you replace $n^{3/2}$ by its floor, $g(n)$ changes by at most 2, hence the only chance for failure is to have $g(n)=2$ when the fractional parts of $n^{3/2}$ and $(n+2)^{3/2}$ are very small and the fractional part of $(n+1)^{3/2}$ is very close to 1 (the difference is less than $\frac{const}{\sqrt{n}}$).

Let $a,b,c$ denote the nearest integers to $n^{3/2}$, $(n+1)^{3/2}$ and $(n+2)^{3/2}$. Then $c-2b+a=0$ because it is an integer very close to $(n+2)^{3/2}-2(n+1)^{3/2}+n^{3/2}$. Denote $m=c-b=b-a$. Then $(n+1)^{3/2}-n^{3/2}<m$ and $(n+2)^{3/2}-(n+1)^{3/2}>m$. Observe that $$ \frac{3}{2}\sqrt{n}<(n+1)^{3/2} - n^{3/2} < \frac{3}{2}\sqrt{n+1} $$ (the bounds are just the bounds for the derivative of $x^{3/2}$ on $[n,n+1]$. Therefore $$ \frac{3}{2}\sqrt{n} < m < \frac{3}{2}\sqrt{n+2} $$ or, equivalently, $$ n < \frac49 m^2 < n+2. $$ If $m$ is a multiple of 3, this inequality implies that $n+1=\frac49 m^2=(\frac23m)^2$, then $(n+1)^{3/2}=(\frac23m)^3$ is an integer and not slightly smaller than an integer as it should be. If $m$ is not divisible by 3, then $$ n+1 = \frac49 m^2 + r $$ where $r$ is a fraction with denominator 9 and $|r|<1$. From Taylor expansion $$ f(x+r) = f(x) +r f'(x) +\frac12 r^2 f''(x+r_1), \ \ 0<r_1<r, $$ for $f(x)=x^{3/2}$, we have $$ (n+1)^{3/2} = (\frac49 m^2 + r)^{3/2} = \frac8{27}m^3 + mr + \delta $$ where $0<\delta<\frac1{4m}$. This cannot be close to an integer because it is close (from above) to a fraction with denominator 27.

  • $\begingroup$ Initially I mistyped 3/2 as 2/3 several times, fixed now. $\endgroup$ Mar 15, 2010 at 12:47

It's not too hard to put a bound on the size of second differences (since without the truncation, they are bounded above by a constant times $n^{-1/2}$), but getting the bound down to one seems delicate. It looks like it can be done with mindless brute force, though. I won't write all of the cases, but here is a start. Write $n = t^2 + r$, for integers $t,r$ satisfying $|r| \leq t$. The binomial theorem implies $n^{3/2} = t^3 + (3/2)tr + (3/8)r^2/t - (1/16)r^3/t^3 + (3/128)r^4/t^5 - \dots$. I'll look at the case where $t$ is even. Then

  • $f(n) = t^3 + (3/2)tr + \lfloor (3/8)r^2/t - (1/16)r^3/t^3 + (3/128)r^4/t^5 - \dots \rfloor$
  • $f(n+1) = t^3 + (3/2)t(r+1) + \lfloor (3/8)(r+1)^2/t - (1/16)(r+1)^3/t^3 + (3/128)(r+1)^4/t^5 - \dots \rfloor$
  • $f(n+2) = t^3 + (3/2)t(r+2) + \lfloor (3/8)(r+2)^2/t - (1/16)(r+2)^3/t^3 + (3/128)(r+2)^4/t^5 - \dots \rfloor$.

$g(n)$ then has no contributions from the first two terms of each series, since they cancel. Therefore:

$g(n) = \lfloor (3/8)r^2/t - (1/16)r^3/t^3 + (3/128)(r+1)^4/t^5 - \dots \rfloor +$ $\qquad + 2\lfloor (3/8)r^2/t + (3/4)r/t - (1/16)(r+1)^3/t^3 + (3/8t) + (3/128)(r+2)^4/t^5 - \dots \rfloor +$ $\qquad + \lfloor (3/8)r^2/t + (3/2)r/t - (1/16)(r+2)^3/t^3 + (3/2t) + (3/128)(r+2)^4/t^5 - \dots \rfloor$.

At this point, you can break the analysis into more cases involving the fractional part of $(3/8)r^2/t$ and the size of $r/t$, and then invoke some estimates about the remaining parts of the sum.

  • $\begingroup$ I've already looked into this way, and it seems that the number of cases and subcases grows hopelessly ; also, the initial "modulo 2" phenomenon (difference between t odd and t even) branches into more complicated phenomena with larger moduli. $\endgroup$ Mar 15, 2010 at 8:05

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.