Let $f:\mathbb R\to \mathbb R$ be such that for every sequence $(a_n)$ of real numbers,

$\sum a_n$ converges $\implies \sum f(a_n)$ converges

Prove there is some open neighborhood $V$, such that $0\in V$ and $f$ restricted to $V$ is a linear function.

This was given to a friend at his oral exam this week. He told me the problem is hard. I tried it myself, but I haven't made any significant progress.

  • $\begingroup$ f is continuous ? $\endgroup$ – Surb Jun 20 '15 at 9:52
  • 2
    $\begingroup$ For which sequence $\sum a_n$ converges but $\sum\sin(a_n)$ doesn't? $\endgroup$ – ajotatxe Jun 20 '15 at 10:11
  • 1
    $\begingroup$ @ajotatxe Consider the series $\sum a_n $ where $$a_{3n-2} = \frac{1}{n^{1/3}}, a_{3n-1} = -\frac{1}{2n^{1/3}}, a_{3n} = -\frac{1}{2n^{1/3}}.$$ Then $\sum a_n$ converges (to $0$). Using $\sin x = x - x^3/6 + O(x^5),$ you can see $\sum \sin a_n = \infty.$ $\endgroup$ – zhw. Jun 20 '15 at 19:58
  • 1
    $\begingroup$ @LeGranDODOM do you mind telling me at which university people ask these kind of questions in oral exams?- I would love to avoid this place in my life. $\endgroup$ – wewasss Jun 21 '15 at 13:25
  • 1
    $\begingroup$ @wewasss this one was given at Ecole Polytechnique this week. The examiner gave my friend a few hints after he got stuck though. $\endgroup$ – Gabriel Romon Jun 21 '15 at 13:27

I have added an answer to collect various resources where this fact is proven. This answer is community-wiki - feel free to add further references you are aware of.



  • Teodora-Liliana Radulescu, Vicentiu D. Radulescu and Titu Andreescu: Problems in Real Analysis: Advanced Calculus on the Real Axis, Problem 2.5.30, page 108

Online resources

  • $\begingroup$ Wow, I didn't know there was some lore about the problem. $\endgroup$ – Gabriel Romon Jun 21 '15 at 13:16
  • $\begingroup$ Neither did I. So I googled, found some papers, looked at references in them. $\endgroup$ – Martin Sleziak Jun 21 '15 at 13:19
  • $\begingroup$ Would it be appropriate for the site if I added some screenshots of the second article ? $\endgroup$ – Gabriel Romon Jun 21 '15 at 13:23
  • $\begingroup$ If you have a link to a copy of the paper, which is (legally) freely available, then adding a link would work just as fine. If it is not freely available, then posting it here would be questionable: To what extent should copyrighted material be made available on math.se?. $\endgroup$ – Martin Sleziak Jun 21 '15 at 13:28
  • $\begingroup$ Thanks for the correction @Did. I removed that claim. $\endgroup$ – Martin Sleziak Jun 21 '15 at 14:04

A start: There exists $\delta > 0$ such that $\frac{f(x)}{x}$ is bounded in $\{0<|x|<\delta\}.$ Proof: Otherwise there is a sequence of nonzero terms $x_n \to 0$ such that

$$\left|\frac{f(x_n)}{x_n}\right|> n$$

for all $n.$ We can pass to a subsequence $n_k$ such that $|x_{n_k}| < 1/k^2$ for all $k.$ We then have

$$(1)\,\,\,\,\left|\frac{f(x_{n_k})}{x_{n_k}}\right|> n_k \ge k.$$

Consider now the case where $x_{n_k}, f(x_{n_k})> 0$ for all $k.$ Then for each $k,$ there is a positive integer $m_k$ such that

$$(2)\,\,\,\,\frac{1}{k^2} < m_kx_{n_k} < \frac{2}{k^2}.$$

Think about the series

$$x_{n_1} + x_{n_1} + \cdots + x_{n_1} + x_{n_2} + \cdots + x_{n_2} + x_{n_3} + \cdots ,$$

where $x_{n_1}$ occurs $m_1$ times, $x_{n_2}$ occurs $m_2$ times, etc. By $(2)$ this series converges. But note $(1)$ shows the $k$th grouping in the "$f$-series" sums to

$$m_kf(x_{n_k}) > m_k\cdot k\cdot x_{n_k} > \frac{1}{k}.$$

Thus the $f$-series diverges, contradiction. So we are done in the case $x_{n_k}, f(x_{n_k})> 0.$ The other cases follow by similar arguments..


Here comes a proof in the general case. We first start with the following

Lemma: Let $-\varepsilon<y<0<x<\varepsilon$ and $N\in\mathbb{N}$. Then there is a some $M\in\mathbb{N}$ and $z_{1},\dots,z_{M}\in\left\{ x,y\right\} $ with $$ \left|\left\{ i\in\left\{ 1,\dots,M\right\} \,\mid\, z_{i}=x\right\} \right|=N $$ and $$ \left|\sum_{i=1}^{K}z_{i}\right|<\varepsilon\qquad\forall K\in\left\{ 1,\dots,M\right\} .\qquad\qquad\left(\ast\right) $$

Proof: We first inductively construct an infinite sequence $\left(w_{n}\right)_{n\in\mathbb{N}}\in\left\{ x,y\right\} ^{\mathbb{N}}$ with $\left|\sum_{i=1}^{n}w_{i}\right|<\varepsilon$ for all $n\in\mathbb{N}$. It is clear that the stated condition holds for $w_{1}:=x$, since $-\varepsilon<x<\varepsilon$. Now, let $w_{1},\dots,w_{n}$ be already constructed. Distinguish two cases:

  1. We have $\alpha:=\sum_{i=1}^{n}w_{i}\geq0$. Then choose $w_{n+1}:=y$. Then $$ \varepsilon>\alpha>\alpha+y=\sum_{i=1}^{n+1}w_{i}=\alpha+y\geq y>-\varepsilon, $$ i.e. $\left|\sum_{i=1}^{n+1}w_{i}\right|<\varepsilon$.

  2. We have $\alpha:=\sum_{i=1}^{n}w_{i}<0$. Then we choose $w_{n+1}:=x$. Similar to the above, it follows $$ -\varepsilon<\alpha<\alpha+x=\sum_{i=1}^{n+1}w_{i}=\alpha+x\leq x<\varepsilon $$ and hence $\left|\sum_{i=1}^{n+1}w_{i}\right|<\varepsilon$.

Now assume that $\left\{ i\in\mathbb{N}\,\mid\, w_{i}=x\right\} $ is finite. This implies $w_{i}=y$ for all $i\geq N_{0}$ with $N_{0}\in\mathbb{N}$ suitable. But this implies $$ \sum_{i=1}^{N}w_{i}=\sum_{i=1}^{N_{0}-1}w_{i}+\sum_{i=N_{0}}^{N}w_{i}=\sum_{i=1}^{N_{0}-1}w_{i}+\left(N-N_{0}\right)y\xrightarrow[N\to\infty]{}-\infty, $$ in contradiction to $\left|\sum_{i=1}^{N}w_{i}\right|<\varepsilon$ for all $N$.

Thus, $\left\{ i\in\mathbb{N}\,\mid\, w_{i}=x\right\} $ is infinite. It is easy to see that this implies existence of $M\in\mathbb{N}$ with $\left|\left\{ i\in\left\{ 1,\dots,M\right\} \,\mid\, w_{i}=x\right\} \right|=N$, so that we can set $z_{i}:=w_{i}$ for $i\in\left\{ 1,\dots,M\right\} $. $\square$

Remark: 1. A completely similar argument show that we can choose $z_{1},\dots,z_{M}\in\left\{ x,y\right\} $ with $$ \left|\left\{ i\in\left\{ 1,\dots,M\right\} \,\mid\, z_{i}=y\right\} \right|=N $$ and such that $\left(\ast\right)$ still holds.

  1. For the above sequence, we also have $$ \left|\sum_{i=K}^{M}z_{i}\right|=\left|\sum_{i=1}^{M}z_{i}-\sum_{i=1}^{K-1}z_{i}\right|\leq\left|\sum_{i=1}^{M}z_{i}\right|+\left|\sum_{i=1}^{K-1}z_{i}\right|<2\varepsilon\qquad\qquad\left(\square\right) $$ for all $K\in\left\{ 1,\dots,M\right\} $.

By the argument given by @zhw., there is some $\delta>0$ and $C>0$ with $\left|\frac{f\left(x\right)}{x}\right|\leq C$ for all $x\in\left(-\delta,\delta\right)\setminus\left\{ 0\right\} $. We will need this constant $C$ below.

First note that $f\left(0\right)=0$ is trivial (consider the sequence $a_{n}=0$). Now, let us assume towards a contradiction that $f$ is not linear on any neighborhood of $0$. This implies that for every $n\in\mathbb{N}$ there are $x_{n}\in\left(0,\frac{1}{n^{2}}\right)$ and $y_{n}\in\left(-\frac{1}{n^{2}},0\right)$ with $$ \alpha_{n}:=\frac{f\left(x_{n}\right)}{x_{n}}\neq\frac{f\left(y_{n}\right)}{y_{n}}=:\beta_{n}. $$ Indeed, if there where no such $x_{n},y_{n}$ for some $n\in\mathbb{N}$, we could fix some $y_{n}\in\left(-\frac{1}{n^{2}},0\right)$ and would get $$ \gamma:=\frac{f\left(y_{n}\right)}{y_{n}}=\frac{f\left(x\right)}{x} $$ for all $x\in\left(0,\frac{1}{n^{2}}\right)$. Likewise, fixing now an arbitrary $x_{n}\in\left(0,\frac{1}{n^{2}}\right)$ and get $$ \frac{f\left(y\right)}{y}=\frac{f\left(x_{n}\right)}{x_{n}}=\frac{f\left(y_{n}\right)}{y_{n}}=\gamma $$ for all $y\in\left(-\frac{1}{n^{2}},0\right)$. All in all, this implies $f\left(y\right)=\gamma\cdot y$ for all $y\in\left(-\frac{1}{n^{2}},\frac{1}{n^{2}}\right)$, since we have $f\left(0\right)=0$. Thus, $f$ is linear on a neighborhood of $0$ after all, a contradiction.

Note that (by choice of $C$ above), we have $\left|\alpha_{n}\right|\leq C$ and $\left|\beta_{n}\right|\leq C$. Let $\varepsilon_{n}:=\frac{1}{n^{2}}$ and choose $N_{n}\in\mathbb{N}$ with $N_{n}\geq\frac{1+C\varepsilon_{n}}{\left|\alpha_{n}-\beta_{n}\right|\left|x_{n}\right|}$. Note that this is possible, since $\left|\alpha_{n}-\beta_{n}\right|>0$ and $\left|x_{n}\right|>0$. By the Lemma above, we find some $M_{n}\in\mathbb{N}$ and $z_{1}^{\left(n\right)},\dots,z_{M_{n}}^{\left(n\right)}\in\left\{ x_{n},y_{n}\right\} $ with $$ N_{n}=\left|\left\{ i\in\left\{ 1,\dots,M_{n}\right\} \,\mid\, z_{i}^{\left(n\right)}=x\right\} \right| $$ and $$ \left|\sum_{i=1}^{K}z_{i}^{\left(n\right)}\right|<\varepsilon_{n}=\frac{1}{n^{2}}\qquad\forall K\in\left\{ 1,\dots,M_{n}\right\} .\qquad\qquad\left(\dagger\right) $$ In particular (for $K=M_{n}$), we get $$ \left|N_{n}x_{n}+\left(M_{n}-N_{n}\right)y_{n}\right|=\left|\sum_{i=1}^{M_{n}}z_{i}^{\left(n\right)}\right|<\frac{1}{n^{2}}.\qquad\qquad\left(\ddagger\right). $$

Now define a sequence $\left(a_{k}\right)_{k\in\mathbb{N}}$ by $$ a_{k}=z_{\ell}^{\left(K\right)}\qquad\text{ for }k=\sum_{i=1}^{K-1}M_{i}+\ell\text{ with }K\in\mathbb{N}\text{ and }\ell\in\left\{ 1,\dots,M_{K}\right\} . $$ It is not too hard to see that this is well-defined. Furthermore, the series $\sum_{k}a_{k}$ is convergent, since it is Cauchy; indeed, for $\sum_{i=1}^{T-1}M_{i}+p=t\geq k=\sum_{i=1}^{K-1}M_{i}+\ell$, we have \begin{eqnarray*} \left|\sum_{n=k}^{t}a_{n}\right| & \leq & \left|\sum_{n=\ell}^{M_{K}}z_{n}^{\left(K\right)}\right|+\sum_{S=K+1}^{T-1}\left|\sum_{n=1}^{M_{S}}z_{n}^{\left(S\right)}\right|+\left|\sum_{n=p}^{M_{T}}z_{n}^{\left(T\right)}\right|\\ & < & 2\varepsilon_{K}+\sum_{S=K+1}^{T-1}\varepsilon_{S}+\varepsilon_{T}\\ & \leq & 2\sum_{n=K}^{T}\frac{1}{n^{2}}\leq2\sum_{n=K}^{\infty}\frac{1}{n^{2}}\xrightarrow[K\to\infty]{}0. \end{eqnarray*}

By assumption on $f$, this implies convergence of $\sum_{n}f\left(a_{n}\right)$. In particular, this series is Cauchy. But for $K\in\mathbb{N}$, we have \begin{eqnarray*} \left|\sum_{n=1+\sum_{i=1}^{K-1}M_{i}}^{\sum_{i=1}^{K}M_{i}}f\left(a_{n}\right)\right| & = & \left|\sum_{n=1}^{M_{K}}f\left(z_{n}^{\left(K\right)}\right)\right|\\ & = & \left|N_{n}f\left(x_{n}\right)+\left(M_{n}-N_{n}\right)f\left(y_{n}\right)\right|\\ & = & \left|N_{n}\alpha_{n}x_{n}+\left(M_{n}-N_{n}\right)\beta_{n}y_{n}\right|\\ & = & \left|N_{n}\left(\alpha_{n}-\beta_{n}\right)x_{n}+\beta_{n}\left(N_{n}x_{n}+\left(M_{n}-N_{n}\right)y_{n}\right)\right|\\ & \geq & N_{n}\left|\alpha_{n}-\beta_{n}\right|\left|x_{n}\right|-\left|\beta_{n}\right|\cdot\left|N_{n}x_{n}+\left(M_{n}-N_{n}\right)y_{n}\right|\\ & \overset{\text{see }\left(\ddagger\right)}{\geq} & N_{n}\left|\alpha_{n}-\beta_{n}\right|\left|x_{n}\right|-C\varepsilon_{n}\\ & \geq & 1 \end{eqnarray*} by choice of $N_{n}$, which shows that $\sum_{n}f\left(a_{n}\right)$ is not Cauchy and hence not convergent, a contradiction.


Here is a proof in the case that $f$ is differentiable at $0$.

Let $f(x) = rx + \epsilon(x)$, where $\lim_{x \to 0} {\epsilon(x)/x}=0$. I claim $\epsilon = 0$ in a a neighborhood of 0. Suppose not. Then, up to extracting subsequences and making a few sign choices that are without loss of generality, we may assume that there exists $a_n$ decreasing monotonically to $0$ with $\epsilon(a_n)>0$, and $\epsilon(a_n)/a_n$ decreasing monotonically to $0$. Up to extracting another subsequence, we are left with two cases: either $\epsilon(-a_n) \geq 0$ for all $n$, or $\epsilon(-a_n) \leq 0$ for all $n$. I claim that neither is actually possible.

First, the easy case: assume $\epsilon(-a_n) \geq 0$ for all $n$. Let $k_n$ be the least integer greater than $1 \over \epsilon(a_n)$. Define $b_i$ to be alternatingly $a_1$ and $-a_1$ for the first $2k_1$ entries. Then define $b_i$ to be alternatingly $a_2$ and $-a_2$ for the subsequent $2k_2$ entries, and so forth. It is easy to check that $\sum b_i$ converges (not absolutely), but $\sum f(b_i) \geq \sum k_n f(a_n) = \infty$.

Now, the hard case: assume $\epsilon(-a_n) \leq 0$ for all $n$. The index manipulation here is pretty awful, but here goes. Let $l_1 = 1$, let $m_{1,1}=k_1$, and let $b_i = a_{l_1}$ for $i \in \{1,...,m_{1,1}\}$. Now, given $l_\alpha$, $m_{\alpha,1}$, and $b_i$ for $i \in \{1,...,m_{\alpha,1}\}$, choose $j_\alpha > l_\alpha$ such that $a_{j_\alpha} < 1/2^\alpha$ and $${\epsilon(-a_{j_\alpha}) \over - a_{j_\alpha}} < {1 \over 4^\alpha} \epsilon(a_{l_\alpha}).$$ Now, choose $p_\alpha > 0$ such that $0 < p_\alpha a_{j_\alpha} - k_{l_\alpha} a_{l_\alpha} < a_{j_\alpha}$. Let $m_{\alpha,2} = m_{\alpha,1} + p_\alpha$, and let $b_i = -a_{j_\alpha}$ for $i \in \{m_{\alpha,1} + 1,...,m_{\alpha,2}\}$. Now pick $l_{\alpha + 1} > j_\alpha$, define $m_{\alpha + 1,1} = m_{\alpha,2} + k_{l_{\alpha + 1}}$, and define $b_i = a_{l_{\alpha + 1}}$ when $m_{\alpha,2} < i \leq m_{\alpha + 1,1}$.

To check that indeed $\sum b_i < \infty$ and $\sum f(b_i) = \infty$, note that $$|\sum\limits_{i=m_{\alpha - 1, 2} + 1}^{m_{\alpha,2}} b_i| = |p_\alpha a_{j_\alpha} - k_{l_\alpha} a_{l_\alpha}| < a_{j_\alpha} < {1 \over 2^\alpha},$$, whence $\sum b_i$ converges. Edit: well, we need to bound the partial sums within each $\sum\limits_{i=m_{\alpha - 1, 2} + 1}^{m_{\alpha,2}} b_i$, but this can be done by rearranging terms since the terms tend to zero as $\alpha \to \infty$. On the other hand, $$\sum\limits_{i=m_{\alpha - 1, 2} + 1}^{m_{\alpha,2}} f(b_i) \geq 1 - r a_{j_\alpha} + p_\alpha \epsilon(-a_{j_\alpha}).$$ Since $a_n \to 0$, we only need to check that $p_\alpha \epsilon(-a_{j_\alpha})$ is not too negative. By our choice of $j_\alpha$, $\epsilon(-a_{j_\alpha}) \geq {-1 \over 4^\alpha} a_{j_\alpha} \epsilon(a_{l_\alpha})$. By our choice of $p_\alpha$, $p_\alpha a_{j_\alpha} < 2 k_{l_\alpha} a_{l_\alpha}$, whence $p_\alpha \epsilon(-a_{j_\alpha}) \geq -2 a_{l_\alpha} k_{l_\alpha} \epsilon(a_{l_\alpha}) \geq -4 a_{l_\alpha}$. For $\alpha$ large, $a_{l_\alpha}$ is very small, so $$\sum\limits_{i=m_{\alpha - 1, 2} + 1}^{m_{\alpha,2}} f(b_i) \geq 1/2.$$ This proves the claim for $f$ differentiable.

By zhw.'s answer, the difference quotients near $0$ are bounded. By Heine-Borel, for each $a_n \to 0$, there is a subsequence $a_{n_k}$ such that $f(a_{n_k})/a_{n_k}$ converges. I'm sure that using similar arguments to the ones so far, you could show that all the subsequential limits are the same. This would show that, in fact, $f(x)/x$ converges as $x \to 0$, which would complete the proof.


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.