Is there a way to calculate the following sum for large
$c > K$, closed-form or approximation ?:
$$
\sum_{n = 1}^{c}\left\{\sqrt{\,{K + n^{2}}\,}\right\}
$$
where $K$ and $n$ are positive integers and
$\left\{\cdots\right\}$ indicates the Fractional Part.
- 94,079
- 365
-
Hint: when $n$ is large compared to $\sqrt K$ then $\sqrt{K+n^2}-n \sim \frac{K}{2n}$. – Greg Martin Aug 22 '21 at 20:30
-
@RuanSunkel are you satisfied with my answer? – fwd Aug 24 '21 at 13:20
-
Hi @fwd, I am still busy digesting. So far I am happy with the answer, I learned a lot. Looking at the question and the answer, I think what I am really interested in is the $\mathcal{O}(1)$ term in $\sum_2$, calculating it given a specific $K$ and $c$, so basically calculating the actual value of $\sum_2$. Thinking of maybe asking a follow up question. – Ruan Sunkel Aug 24 '21 at 14:14
3 Answers
The given sum can be written as $$ \Sigma = \Sigma_1 - \Sigma_2 $$ where $$ \Sigma_1 = \sum_{n=1}^{c} \sqrt{n^2 + K} $$ and $$ \Sigma_2 = \sum_{n=1}^{c} \lfloor\sqrt{n^2 + K}\rfloor. $$ Now using $H_n = \log n + \gamma + \mathcal{O}(1/n)$ where $H_n$ is the $n$-th harmonic number, \begin{align*} \sum_{n=1}^{c} \sqrt{n^2 + K} - n &= \sum_{n=1}^c \frac{K}{\sqrt{n^2 + K} + n} \\ &\le \frac{K}{2} \sum_{n=1}^{c} \frac{1}{n} = \frac{K}{2}(\log c + \gamma + \mathcal{O}(1/c)). \end{align*} Hence $$ \Sigma_{1} = \sum_{n=1}^{c} \sqrt{n^2 + K} = \frac{1}{2}c(c+1) + \frac{K}{2}(\log c + \gamma) + \mathcal{O}(1/c) $$
Let $A = \lfloor\sqrt{K+1} - 1\rfloor$. Then, \begin{align*} \Sigma_{2} = \sum_{n=1}^{c} \lfloor\sqrt{n^2 + K}\rfloor &= \sum_{j=1}^{A}\sum_{\frac{K-(j+1)^2}{2(j+1)} < n \le \frac{K-j^2}{2j}}(n+j) + \sum_{\frac{K-1}{2} < n \le c}n \\ &= \sum_{n=1}^c n + \sum_{j=1}^{A}\sum_{\frac{K-(j+1)^2}{2(j+1)} < n \le \frac{K-j^2}{2j}}j\\ &= \frac{1}{2}c(c+1) + \sum_{j=1}^A j\left(\frac{3}{2} + \frac{K}{2j(j+1)}+ \mathcal{O}(1)\right)\\ &= \frac{1}{2}c(c+1) + \frac{3}{4}A(A+1) + \frac{K}{2}(H_A - 1) + \mathcal{O}\left(\sum_{j=1}^A j\right)\\ &= \frac{1}{2}c(c+1) + \frac{3}{4}A(A+1) + \frac{K}{2}(H_A - 1) + \mathcal{O}(1) \end{align*} since $\mathcal{O}\left(\sum_{j=1}^A j\right) = \mathcal{O}(K) = \mathcal{O}(1)$.
Thus,
\begin{align*} \Sigma &= \frac{K}{2}(\log c + \gamma) - \frac{3}{4}A(A+1) - \frac{K}{2}(H_A - 1) + \mathcal{O}(1).\\ \end{align*}
- 3,351
-
Hi, thank you for the answer. I did a quick test and $\sum_2$ seems to be different than expected. For example, with $K=15$, and $c=16$, $\sum_2=146$, but $\sum_2$ provided in the answer gives $\sum_2=2822$. Please correct me if I am missing something. – Ruan Sunkel Aug 22 '21 at 18:14
-
-
-
Thank you @fwd. Can you please elaborate on the $\mathcal{O}(1)$ constant that depends on $K$? – Ruan Sunkel Aug 23 '21 at 17:57
-
@RuanSunkel I added more details to show where the $\mathcal{O}(1)$ term comes from. The larger $K$ is, the larger this term is. Numerical plots suggest that it grows linearly in $K$. But since $K$ is fixed, it's just $\mathcal{O}(1)$ – fwd Aug 23 '21 at 18:27
-
Partial solution: If we follow the example from Find a formula for $\sum\limits_{k=1}^n \lfloor \sqrt{k} \rfloor$. Then we have integer values for some integer $p$ where $p \le \sqrt{{k}^{2} + n} < p + 1 \Longleftrightarrow {p}^{2} \le {k}^{2} + n < \left({p + 1}\right)^{2} \Rightarrow p = \lfloor{\sqrt{{k}^{2} + n}}\rfloor$, or
$$ \sqrt{{p}^{2} - n} \le k < \sqrt{\left({p + 1}\right)^{2} - n} \Rightarrow 1 \le \lceil{\sqrt{{p}^{2} - n}}\rceil \le k \le \lceil{\sqrt{\left({p + 1}\right)^{2} - n}}\rceil - 1 \le c. $$
The maximum $p$ index is given by $\sqrt{{c}^{2} + n} < p + 1$ or ${p}_{max} = \lfloor{\sqrt{{c}^{2} + n}}\rfloor$ while the minimum $p$ index is ${p}_{min} = \lfloor{\sqrt{n +1}}\rfloor$. Also the lower $k$ index Iverson bracket constraint is needed to ensure that ${p}^{2} \ge n$, and the upper $k$ index constraint is to ensure that ${k}_{max} \le c$. So we can write
$$ \begin{aligned} {\Sigma}_{2} = \sum_{k = 1}^{c} \lfloor{\sqrt{{k}^{2} + n}}\rfloor {}={} & \sum_{p = \lfloor{\sqrt{n + 1}}\rfloor}^{\lfloor{\sqrt{{c}^{2} + n}}\rfloor} \sum_{k = \max \left({1, \lceil{\sqrt{{p}^{2} - n}}\rceil \left[{{p}^{2} \ge n}\right]}\right)}^{\min \left({c, \lceil{\sqrt{\left({p + 1}\right)^{2} - n}}\rceil - 1}\right)} \lfloor{\sqrt{{k}^{2} + n}}\rfloor \\ {}={} & \sum_{p = \lfloor{\sqrt{n + 1}}\rfloor}^{\lfloor{\sqrt{{c}^{2} + n}}\rfloor} p \sum_{k = \max \left({1, \lceil{\sqrt{{p}^{2} - n}\rceil} \left[{{p}^{2} \ge n}\right]}\right)}^{\min \left({c, \lceil{\sqrt{\left({p + 1}\right)^{2} - n}}\rceil - 1}\right)} 1 \\ {}={} & \sum_{p = \lfloor{\sqrt{n + 1}}\rfloor}^{\lfloor{\sqrt{{c}^{2} + n}}\rfloor} p \begin{bmatrix} 1 + \min \left({c, \lceil{\sqrt{\left({p + 1}\right)^{2} - n}}\rceil - 1}\right) \\ \mathop{-} \max \left({1, \lceil{\sqrt{{p}^{2} - n}}\rceil \left[{{p}^{2} \ge n}\right]}\right) \end{bmatrix}. \end{aligned} $$
At this point the remaining sums can be approximated using the Euler Maclaurin formula after dropping the floor and ceiling functions, and taking the dominant argument of the max and min argument lists. This will generate asymptotic expansions.
This solution holds for all values of $c$.
- 1,005
Upon further telescoping sum expansions and simplifications we can write for $c \ge \lfloor{(n-1)/2}\rfloor$ and with asymptotic expansions as $n \rightarrow \infty$
$${\Sigma}_{2} = \frac{1}{2}\, c \left({c + 1}\right) + \sum_{j = 1}^{A} \lfloor{\frac{K - {j}^{2}}{2\, j}}\rfloor \approx \frac{1}{2}\, {c}^{2} + \frac{1}{2}\, c + \frac{1}{4}\, K\, \log \left({K}\right) + \frac{1}{4} \left({2\, \gamma - 1}\right) K + \frac{11}{24} - \frac{49}{240}\, \frac{1}{K} + \frac{197}{1{,}260}\, \frac{1}{{K}^{2}} - \frac{421}{3{,}360}\, \frac{1}{{K}^{3}} + O \left({\frac{1}{{K}^{4}}}\right). $$
Which is exact. However, so far I can't find the correct summation for $c < \lfloor{(K-1)/2}\rfloor$.
See How to show that $\sum_{j = 1}^{A \left({n}\right)} \left\{{\frac{n - {j}^{2}}{2\, j}}\right\}$ satisfies Weyl's criterion. for a related question to solving the fractional part of the above floor sum. A change of variables, $K = n$.
- 1,005