Let us call a set $S$ of positive integers "good" if there does not exist a pair of distinct subsets $A,B\subseteq S$ who have equal size and an equal sum. Equivalently, a set $S$ is good if the function $f:P(S)\rightarrow \mathbb N^2$ defined as $$f(A)=\left(|A|,\sum_{a\in A}a\right)$$ is injective.

Given any $d$, what is the smallest $n$ such that there exists a subset $S$ of $\{1,\ldots,n\}$ which is good and has $d$ elements?

I thought of this problem as a variation on a related one where we drop the requirement that $A$ and $B$ have the same size (i.e. ask that $g(A)=\sum_{a\in A}a$ be injective on the power set of $S$). (This post previously claimed this variant was easy; it has been pointed out in the comments by @Shagnik that it is open).

A potentially useful reformulation of the new problem is that one may also define this as looking for subsets of $\{(1,1),(1,2),(1,3),\ldots,(1,n)\}$ such that no two subsets have equal sum.

I can get lower bounds on $n$. In particular, note that a $k$ element subset of $\{1,\ldots,n\}$ must have its sum lie between $1+2+\ldots+k$ and $(n-k+1)+\ldots+n$. In particular, it can take on only $k(n-k)+1$ possible sums. However, there are ${d \choose k}$ subsets of $S$ needing distinct sums. Thus, we get the bound: $${d\choose k}\leq k(n-k)+1.$$ which must hold for all $0\leq k \leq d$. If one sets $k=d/2$ and works through using Stirling's approximation, dropping some lower order terms that crop up, one gets that $n$ should be at least asymptotic to $\frac{2\sqrt{2}\cdot 2^d}{\sqrt{\pi}\cdot d^{3/2}}$. That is, the ratio of $n$ to that expression tends to at least $1$ as $d$ goes to $\infty$. For upper bounds, we can obviously bound that $n\leq 2^{d-1}$ since $\{1,2,\ldots,2^{d-1}\}$ is a good set. However, this still leaves a sizeable gap between upper and lower bounds.

I do not know whether it is optimal, but a greedy algorithm where we define a sequence $a_n$ by setting $$a_n=\min\{x:\{a_1,a_2,\ldots,a_{n-1},x\}\text{ is good}\}$$ yields a sequence starting as $1,\,2,\,3,\,5,\,8,\,14,\,25,\,45,\,85,\,162,\,310,\,595,\,1165,\,2285,\,4485,\,8808$.

I suspect this is a difficult problem to answer in full, but I would be interested in any refinements of either upper or lower bounds or any literature discussing the problem.

  • $\begingroup$ If you remove the requirement the subsets have the same size it is trivially $2^{n-1}$ right? $\endgroup$ – Jorge Fernández Hidalgo Jul 15 '16 at 23:00
  • 1
    $\begingroup$ $1,2,3,5,8,14$ are correct for the first $6$ terms, by brute-force computation. But I get $a_7 = 23$. $\endgroup$ – Robert Israel Jul 15 '16 at 23:12
  • 2
    $\begingroup$ $\{1,2,3,11,17,20,23\}$. $\endgroup$ – Robert Israel Jul 15 '16 at 23:15
  • 1
    $\begingroup$ $3,5,6,7$ is the only counter of size $4$. $\endgroup$ – Jorge Fernández Hidalgo Jul 16 '16 at 19:21
  • 1
    $\begingroup$ From the Bohman paper about this: once you have a counterexample, you can get a larger one by doubling all the numbers and adding $1$. The odd sums all use $1$, and the even ones don't, so it reduces to the smaller counterexample. Your example of size $5$ arises the from one of size $4$ in this way. $\endgroup$ – Shagnik Jul 16 '16 at 19:28

You can improve the lower bound a little bit. Suppose we have $\{ a_1, a_2, ..., a_d\} \subseteq [n]$ such that no two distinct subsets of equal size have the same sum. As you noted, this is equivalent to no two distinct subsets (of any size) of $\left\{ \begin{pmatrix} 1 \\ a_1 \end{pmatrix} , \begin{pmatrix} 1 \\ a_2 \end{pmatrix} , ..., \begin{pmatrix} 1 \\ a_d \end{pmatrix} \right\}$ having the same sum.

For $1 \le i \le d$, let $\varepsilon_i$ be independent uniform $\{0,1\}$ random variables, and consider the random variables $X$ and $Y$ given by $$ \begin{pmatrix} X \\ Y \end{pmatrix} = \sum_{i=1}^d \varepsilon_i \begin{pmatrix} 1 \\ a_i \end{pmatrix}. $$

We have $X \sim \text{Bin} \left( d,\frac12 \right)$, and so it has mean $\mu_Y = \frac{d}{2}$ and variance $\sigma_X^2 = \frac{d}{4}$. By Chebyshev's inequality, $\mathbb{P}\left( |X - \mu_X| > \sqrt{\frac{3d}{4}} \right) \le \frac{\sigma_X^2}{(3d /4)} = \frac13$.

Since the $\varepsilon_i$ are independent, the variance of $Y$ is given by $\sigma_Y^2 = \sum_{i=1}^d \text{Var}(\varepsilon_i a_i) = \frac14 \sum_{i=1}^d a_i^2 \le \frac{dn^2}{4}$. Hence, if $\mu_Y$ is the expectation of $Y$, Chebyshev gives $\mathbb{P} \left( | Y - \mu_Y | > \sqrt{\frac{3dn^2}{4}} \right) \le \frac13$. (We have $\mu_Y = \frac12 \sum_{i=1}^d a_i$, but we won't be using this fact.)

Hence with probability at least $\frac13$, we have $| X - \mu_X | \le \frac{\sqrt{3d}}{2}$ and $| Y - \mu_Y | \le \frac{n \sqrt{3d}}{2}$. That is to say, for at least $\frac13 2^d$ choices of the coefficients $\varepsilon_i$, $\begin{pmatrix} X \\ Y \end{pmatrix}$ lies within some $\left( \sqrt{3d} + 1 \right)$-by-$\left(n \sqrt{3d} + 1\right)$ grid.

Since these sums must all be distinct, it follows that $$ \frac13 2^d \le \left( \sqrt{3d} + 1 \right) \left( n \sqrt{3d} + 1 \right), $$ from which we can deduce $n \ge \left( 1 - o(1) \right) \frac{2^d}{9 d}$.

Note that for the version where you don't require that the subsets have the same size, we don't need the variable $X$ and its $\sqrt{d}$ factor, so we get a lower bound of the form $\frac{c2^d}{\sqrt{d}}$ for some constant $c > 0$.

I am not sure if restricting the size of the subsets really makes the problem that much easier, and there is no known upper bound better than $c' 2^d$ for some other constant $c' > 0$, so I suspect it might be difficult to close the gap further.


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.