7

For which ${k, p \in \mathbb{R}}$ does $\sum_{m,n \in \mathbb{N}} \frac{1}{m^p + n^k}$ converge?

It is necessary but not sufficient that $p, k \gt 1$

I am looking for a simple solution, as the solutions I have seen so far have been rather perverted.

Mark
  • 5,696

4 Answers4

5

Breaking the sum into two parts we have $$ \sum_{m=1}^\infty\sum_{n=1}^\infty\frac1{m^p+n^k} =\color{#C00000}{\sum_{m=1}^\infty\sum_{n=1}^{\left\lfloor m^{p/k}\right\rfloor}\frac1{m^p+n^k}} +\color{#00A000}{\sum_{n=1}^\infty\sum_{m=1}^{\left\lceil n^{k/p}-1\right\rceil}\frac1{m^p+n^k}}\tag{1} $$ For the red (left hand) sum, where $m^p\ge n^k$, we have $$ \frac12\sum_{m=1}^\infty\sum_{n=1}^{\left\lfloor m^{p/k}\right\rfloor}\frac1{m^p} \le\sum_{m=1}^\infty\sum_{n=1}^{\left\lfloor m^{p/k}\right\rfloor}\frac1{m^p+n^k} \le\sum_{m=1}^\infty\sum_{n=1}^{\left\lfloor m^{p/k}\right\rfloor}\frac1{m^p}\tag{2} $$ and $$ \sum_{m=1}^\infty\sum_{n=1}^{\left\lfloor m^{p/k}\right\rfloor}\frac1{m^p} =\sum_{m=1}^\infty\frac{\left\lfloor m^{p/k}\right\rfloor}{m^p}\tag{3} $$ which converges iff $p\left(1-\frac1k\right)\gt1$, which is equivalent to $\frac1p+\frac1k\lt1$.

For the green (right hand) sum, where $n^k\gt m^p$, we have $$ \frac12\sum_{n=1}^\infty\sum_{m=1}^{\left\lceil n^{k/p}-1\right\rceil}\frac1{n^k} \le\sum_{n=1}^\infty\sum_{m=1}^{\left\lceil n^{k/p}-1\right\rceil}\frac1{m^p+n^k} \le\sum_{n=1}^\infty\sum_{m=1}^{\left\lceil n^{k/p}-1\right\rceil}\frac1{n^k}\tag{4} $$ and $$ \sum_{n=1}^\infty\sum_{m=1}^{\left\lceil n^{k/p}-1\right\rceil}\frac1{n^k} =\sum_{n=1}^\infty\frac{\left\lceil n^{k/p}-1\right\rceil}{n^k}\tag{5} $$ which converges iff $k\left(1-\frac1p\right)\gt1$, which is equivalent to $\frac1p+\frac1k\lt1$.

Thus, $(1)$ converges iff $\frac1p+\frac1k\lt1$.

robjohn
  • 345,667
3

Let us assume that $p,k>1$. Denote $$a_{m,n}=\frac{1}{m^p+n^k},$$

and for every $m\ge 1$, denote $$b_m=m^{-p(1-\frac{1}{k})}\quad\text{and}\quad S_m=\sum_{n=1}^\infty a_{m,n}<\infty.$$

Claim: There exist $0<C_1<C_2$, only dependent on $k$ and $p$, such that for every $m\ge 1$,
$$C_1\, b_m<S_m<C_2\, b_m.\tag{1}$$ As a corollary, $$\sum_{m,n=1}^\infty a_{m,n}=\sum_{m=1}^\infty S_m<\infty \iff \frac{1}{p}+\frac{1}{k}<1.\tag{2}$$

Proof: To prove $(1)$, we may suppose $m^\frac{p}{k}\ge 3$ for convenience. Write $S_m$ as $A_m+B_m$, where $$A_m=\sum_{1\le n\le m^\frac{p}{k}}a_{m,n}\quad\text{and}\quad\quad B_m=\sum_{n>m^\frac{p}{k}}a_{m,n}.$$

Given $m^\frac{p}{k}\ge 3$, when $1\le n\le m^\frac{p}{k}$,

$$\frac{m^{-p}}{2} \le a_{m,n}\le m^{-p}\Longrightarrow \frac{b_m}{6}= \frac{m^\frac{p}{k}}{3}\cdot\frac{m^{-p}}{2} \le A_m\le m^\frac{p}{k}\cdot m^{-p}=b_m\,; \tag{3}$$

when $n>m^\frac{p}{k}$,

$$\frac{n^{-k}}{2} \le a_{m,n}\le n^{-k}\Longrightarrow \frac{b_m}{2^k(k-1)}=\int_{2m^\frac{p}{k}}^\infty\frac{dx}{2x^k}\le B_m\le\int_{\frac{1}{2}m^\frac{p}{k}}^\infty\frac{dx}{x^k}= \frac{2^{k-1} b_m}{k-1}. \tag{4}$$

$(1)$ follows from $(3)$ and $(4)$ immediately; $(2)$ follows from $(1)$ and the fact that $$\sum_{m=1}^\infty m^{-p(1-\frac{1}{k})}<\infty \iff p(1-\frac{1}{k})>1.$$

Hu Zhengtang
  • 3,757
  • 1
    @robjohn: Ah, see you again. Since you have posted your answer and its formulation is quite clear, I think you shouldn't delete it. However, I don't understand when you find an existing answer using the same argument as the answer you are typing, why you choose to complete your typing and post it, rather than abandon it.(In this case, judging from the time record, my answer was posted more than one hour earlier than yours, so I suppose your answer was far from being completed when you saw my answer.) – Hu Zhengtang Oct 19 '13 at 13:18
  • 1
    @robjohn: For me, when I am going to answer an question, before starting typing, I would make sure there is no existing answer using a similar argument as mine, and after the typing being completed, I will check it again before posting. If there is some new answer similar to mine, I will definitely abandon mine, because I try to avoid any suspicion of plagiarism. – Hu Zhengtang Oct 19 '13 at 13:31
  • Although we both break the sum into two parts at the point where $m^p=n^k$ (which seems to be the point of the second line of your proof), after I had time to decipher your proof, that is about the end of the similarity in our methods. Line $(3)$ looks somewhat similar, but there is the requirement that $m\ge3^{k/p}$ which is confusing. It seems that line $(4)$ is controlling the tail of the sum, which my answer avoids. A minor point would be to account for $S_m$ for $m\lt3^{k/p}$. – robjohn Oct 19 '13 at 14:05
  • 3
    In any case, just as it is with math books, I believe that two answers explaining the same ideas can be quite different depending on presentation. All else being equal, an answer which is easier to understand is better than an answer which is harder to follow. It would be sad if the latter were to preclude the former. This is starting to sound like a topic for a meta thread. – robjohn Oct 19 '13 at 14:05
  • 1
    @robjohn: Thank you for your reply. The artificial requirement $m^{\frac{k}{p}}\ge 3$ looks weird in the statement of the claim, so I removed it from the statement. I agree with your standpoint that "an answer which is easier to understand is better than an answer which is harder to follow", but I still do not feel like posting any answer sharing the same idea with the existing ones by myself just because I can provide a better formulation. – Hu Zhengtang Oct 19 '13 at 17:19
  • 1
    I don't think Rob plagiarized your answer, if that is your concern. His method seems slightly different. They are both fine answers and complement each other. – Potato Oct 20 '13 at 03:18
  • 1
    @Potato: No, I never thought robjohn plagiarized my answer. My first two comments were replying an earlier comment of robjohn, which has been deleted by him/her now. – Hu Zhengtang Oct 20 '13 at 04:04
3

Here is another approach:

Upper bound the sum by grouping by dyadic scales:

$\sum_{m,n} \frac{1}{m^k+n^p}\leq \sum_{l} \#\{m,n:\ 2^{l-1}\leq m^k+n^p \leq 2^l \} 2^{-(l-1)}$

We can upper bound the size of this set by the size of the set $\{m,n:\ m^k \leq 2^l, n^p \leq 2^l\}$, Thus the size is bounded above by $2^{l/k+l/p}$ and an upper bound on the sum is

$\sum_l 2^{1/k+1/p-1}$ which is finite if $1/k+1/p < 1$.

For the other direction, lower bound the sum by

$\sum_{m,n} \frac{1}{m^k+n^p}\geq \sum_{l} \#\{m,n:\ 2^{l-1}\leq m^k+n^p \leq 2^l \} 2^{-l}$

A lower bound on this count is to bound the set $\{m,n:\ 2^{l-2}\leq m^k \leq 2^{l-1}, 2^{l-2} \leq n^p \leq 2^{l-1}\}$ which has size of order $(2^{-1/k}-2^{-2/k})(2^{-1/p}-2^{-2/p})2^{l/k+l/p}$, and thus if $1/k+1/p > 1$ the sum diverges.

Evan
  • 3,861
  • This answer looks pretty elegant to me, just hope that a dyadic proof does not fall into the "rather perverted" category. – Luke Skywalker Oct 20 '13 at 06:03
  • Very nice! I wonder if there is some heuristic that can let you know beforehand if this kind of technique will work out. – Mark Oct 20 '13 at 07:13
  • seems like anything that let's you rearrange or group terms in the sum can be attacked this way, it's a fairly common technique in harmonic analysis. very powerful – Evan Oct 20 '13 at 13:07
1

We split the sum under consideration into the three sums of positive terms: $$\sum_{m=1}^\infty \sum_{n=1}^\infty \frac 1 {m^p+n^k}=\sum_{m=2}^\infty \sum_{n=2}^\infty \frac 1 {m^p+n^k}+\sum_{m=1}^\infty \frac 1 {m^p+1}+\sum_{n=2}^\infty \frac 1 {n^k+1}.$$ If $p>1$ and $k>1$, then the two last series converge. Next, $$ \sum_{m=2}^\infty \sum_{n=2}^\infty \frac 1 {m^p+n^k}\le \sum_{m=2}^\infty \sum_{n=2}^\infty \int_{m-1}^{m} \int_{n-1}^{n} \frac {dx\,dy} {x^p+y^k}=\int_1^\infty \int_1^\infty \frac {dx\, dy} {x^p+y^k}.$$ Now we change $$ \left\{ x={r}^{2/p} \left( \cos \left( \phi \right) \right) ^{2/p},y={r}^{2/k} \left( \sin \left( \phi \right) \right) ^{2/k} \right\}. $$ The Maple code $$ with(VectorCalculus):$$ $$simplify(Jacobian([r^{2/p}*cos(phi)^{2/p}, r^{2/k}*sin(phi)^{2/k}], [r, phi], 'determinant')[2], trig, power);$$ calculates the Jacobian $$4\,{r}^{-{\frac {pk-2\,k-2\,p}{pk}}} \left( \cos \left( \phi \right) \right) ^{-{\frac {p-2}{p}}} \left( \sin \left( \phi \right) \right) ^{-{\frac {-2+k}{k}}}{p}^{-1}{k}^{-1}. $$ Thus, $$ \int_1^\infty \int_1^\infty \frac {dx\, dy} {x^p+y^k}\le$$ $$\int_1^\infty r^{-{\frac {pk-2\,k-2\,p}{pk}}-2} \,dr\int_{\phi_1(r)}^{\phi_2(r)}\left( \cos \left( \phi \right) \right) ^{-{\frac {p-2}{p}}} \left( \sin \left( \phi \right) \right) ^{-{\frac {-2+k}{k}}}{p}^{-1}{k}^{-1}\,d\phi \le $$ $$ \int_1^\infty r^{-{\frac {pk-2\,k-2\,p}{pk}}-2} \,dr\int_{0}^{\pi/2}\left( \cos \left( \phi \right) \right) ^{-{\frac {p-2}{p}}} \left( \sin \left( \phi \right) \right) ^{-{\frac {-2+k}{k}}}{p}^{-1}{k}^{-1}\,d\phi. $$ The inequalities appear because we extend the region of the integration from the quadrant $\{(x,y):x \ge 1, y \ge 1 \} $ to $\{(r,\phi): r\ge 1, \phi \ge \phi_1(r)=\arcsin(1/r), \phi \le \phi_2(r)=\pi/2-\arcsin(1/r)\} \subset$ $ \{(r,\phi):r\ge 1,\, \phi \ge 0, \,\phi \le \pi/2\}.$ The integral $$ \int_0^{\pi/2}\left( \cos \left( \phi \right) \right) ^{-{\frac {p-2}{p}}} \left( \sin \left( \phi \right) \right) ^{-{\frac {-2+k}{k}}}{p}^{-1}{k}^{-1}\,d\phi$$ converges because $p>1,\,k>1.$ The integral $$ \int_1^\infty r^{-{\frac {pk-2\,k-2\,p}{pk}}-2} \,dr$$ converges iff $$\frac 1 p +\frac 1 k <1. $$

user64494
  • 5,811
  • For the completeness, the inequality $$\sum_{m=1}^\infty \sum_{n=1}^\infty \frac 1 {m^p+n^k} \ge \int_1^\infty \int_1^\infty \frac {dx, dy} {x^p+y^k} $$ should be regarded. – user64494 Oct 11 '13 at 04:56
  • I'm not seeing why the trigonometric integral converges. What if the power on the sine term is negative and the function becomes unbounded near 0? Also if it converges I see that your condition is sufficient but I don't see that it's necessary. Thanks – Mark Oct 11 '13 at 20:08
  • The integrand is equivalent to $\phi^{2/k-1}p^{-1}k^{-1} $ as $\phi \to 0.$ The exponent $2/k-1 >-1$ if $k>1$. This implies the converges at $\phi=0$. The consideration of the integrand at $\phi=\pi/2$ is similar. What do you mean by "your condition"? – user64494 Oct 11 '13 at 20:19
  • Under the condition $\frac{1}{p} + \frac{1}{k} \lt 1$ the integral will converge. But what if the condition is not true? Maybe the integral can still converge? – Mark Oct 11 '13 at 20:30
  • Which integral do you mean? – user64494 Oct 11 '13 at 20:34
  • This one: $\int \frac{1}{x^p + y^k} dx dy$ You showed it converges (and therefore the sum converges) by upper bounding it with another integral. – Mark Oct 11 '13 at 20:50
  • The question is about the convergence of the double sum. In my answer the necessary and sufficient condition $\frac 1 p+\frac 1 k < 1$ of the convergence under consideration is proved. This is also the necessary and sufficient condition convergence of the integral $$\int_0^\infty \int_0^\infty \frac {dx,dy} {x^p+y^k}$$ because of the inequality $$\sum_{m=2}^\infty \sum_{n=2}^\infty \frac 1 {m^p+n^k}\le \sum_{m=2}^\infty \sum_{n=2}^\infty \int_{m-1}^{m} \int_{n-1}^{n} \frac {dx,dy} {x^p+y^k}=\int_1^\infty \int_1^\infty \frac {dx, dy} {x^p+y^k}.$$ – user64494 Oct 11 '13 at 21:32
  • 2
    @user64494: I think Mark is referring to the inequality that arises when you extend the domain. If the double sum converges, then (by the inequality in your comment) the double integral in the left side also converges, but the polar extended integral might in principle diverge (and thus $1/p+1/k>1$). – Luke Skywalker Oct 12 '13 at 23:12