How does one compute $\mathbb{E}[\max_{1\le i < n} |\sigma(i) - \sigma(i+1)|]$ where the expectation is taken over a uniformly random permutation $\sigma \in \mathbb{P}_n$, the set of all permutations on $[n]$?

Is there a recursion for $c^n_t$ that counts the number of permutations $\sigma$ on $[n]$ such that $\max_{1\le i < n} |\sigma(i) - \sigma(i+1)| \le t$?

  • 2
    $\begingroup$ This seems very difficult. Even determining $c^n_t$ for small $t$ is not easy. Using the transfer-matrix method to insert new maximum elements gives $c^n_2=2n^2-10n+20\,$ for $n\geqslant 4\,$ (for smaller $n$, $c^n_2=n!$). Surprisingly, this is not in OEIS. $\endgroup$ – David Bevan Nov 7 '12 at 20:59
  • $\begingroup$ @quantumelixir, are you interested in an estimate for the expectation? I think the answer is $n - \Theta(\sqrt{n})$. $\endgroup$ – Yury Nov 19 '12 at 0:16
  • $\begingroup$ @Yury: I'm happy with a good estimate for the expectation. So please add a complete answer! Thanks :-) $\endgroup$ – quantumelixir Nov 19 '12 at 22:17
  • $\begingroup$ Is it known whether for a fixed t, $c^n_t$ is always eventually a polynomial in $n$ ? $\endgroup$ – Ewan Delanoy Nov 20 '12 at 5:10
  • $\begingroup$ @EwanDelanoy, no it is exponential in $n$. Very roughly, $c_t^n$ is $t^n$. There are exponentially many permutations even when $t=2$. Consider permutations of the form $[1\ 2]\ 3\ [4\ 5]\ 6\ [7\ 8]\ 9\ [10\ 11]\ \dots$ where you can choose any order for numbers in the brackets. For all such permutations $t\leq 2$. The number of such permutations is $2^{\lfloor (n+1)/3\rfloor}$. $\endgroup$ – Yury Nov 20 '12 at 5:56

The answer to your first question is $n - \sqrt{\pi n}/2 + o(\sqrt{n})$. The proof is quite involved. So I will only sketch the main idea. Let me know if you are interested in more details.

The answer to your second question is: $(\alpha t)^n \leq c_t^n \leq (\beta t)^n$ for some constants $\alpha, \beta > 0$.

Part I. An estimate for $\mathbb{E}[\max_{1\le i < n} |\sigma(i) - \sigma(i+1)|]$.

It will be convenient to work with the random variable $\Delta = n - \max_{1\le i < n} |\sigma(i) - \sigma(i+1)|$. Our goal is to prove that $\mathbb{E}[\Delta] = (\sqrt{\pi}/2 + o(1)) \sqrt{n}$. We will consider an algorithm that generates $\sigma$ uniformly at random and show that for this algorithm $\mathbb{E}[\Delta] = (\sqrt{\pi}/2 + o(1)) \sqrt{n}$ (of course, the answer doesn't depend on the way we generate $\sigma$, as long as $\sigma$ is distributed uniformly).

Denote $[n] = \{1,\dots,n\}$.

Let $k = n ^{1/2+ \varepsilon}$ (for some small $\varepsilon > 0$). Choose two disjoint subsets $S,T \subset [n]$ uniformly at random.


  • assign values from $\{1,\dots, k\}$ to $\sigma(i)$ for $i\in S$ (use each value once);
  • assign values from $\{n-k+1,\dots, n\}$ to $\sigma(i)$ for $i\in T$ (use each value once);
  • assign values from $\{k+1,\dots, n-k\}$ to $\sigma(i)$ for $i\in [n]\setminus (S\cup T)$ (use each value once).

We get a permutation $\sigma$ distributed uniformly at random.

Informal Observation. Observe that if $i\in S$ and $i+1\in T$ then $\sigma(i+1) - \sigma(i) \geq (n-k+1) - k = n - 2k +1$ and thus $\Delta \leq 2k - 1$. Similarly, if $i\in T$ and $i+1\in S$, then $\Delta \leq 2k - 1$. On the other hand, if there is no $i$ such that either $i\in S$ and $i+1\in T$ or $i\in T$ and $i+1\in S$ then for every $i$, $|\sigma(i) - \sigma(i+1)| \leq n - (k+1)$. So $\Delta \geq k + 1$.

Note that by the Birthday Paradox a pair $(i,i+1)$ with $i\in S$ and $i+1\in T$ should exist when $k\approx \sqrt{n}$. Therefore, we expect that $\Delta = \Theta(\sqrt n)$.

Let $A$ be the set of pairs $(i,i+1)$ such that $i\in S$ and $i+1\in T$ or $i\in T$ and $i+1\in S$. We estimate the size of $A$. The probability that $i\in S$ and $i+1\in T$ is $\frac{|S| \cdot |T|}{n(n-1)} = \frac{k^2}{n(n-1)}$; the probability that $i\in T$ and $i+1\in S$ is also $\frac{k^2}{n(n-1)}$. Thus the expected number of pairs is $2k^2/n$. Moreover, the number of such pairs is concentrated near $2k^2/n$.

It can happen that $(i,i+1)\in A$ and $(i+1,i+2)\in A$. But the expected number of such $i$ is approximately $k^3 / n^2$. So by the Markov inequality there is at least one such $i$ with probability at most $k^3/n^2 = n^{3\varepsilon - 1/2}$. We can ignore this event.

Let \begin{align} S' &= \{i\in S: (i,i+1) \in A \text{ or } (i-1,i)\in A\},\\ T' &= \{i\in T: (i,i+1) \in A \text{ or } (i-1,i)\in A\}. \end{align} Recall that we assigned values from $\{1,\dots,k\}$ to $\sigma(i)$ for $i\in S$. Let us a bit modify the assignment procedure:

  1. For each $i\in S'$ assign $\sigma(i)$ a value from $\{1,\dots, k\}$ independently uniformly at random.
  2. If we used each value from $\{1,\dots, k\}$ at most once, assign unused values from $\{1,\dots, k\}$ randomly to unassigned $\sigma(i)$, $i\in S\setminus S'$ using each value only once.
  3. If we used some value more than once, assign values from $\{1,\dots, k\}$ to all $\sigma(i)$, $i\in S$ using each value only once (reassign some $\sigma(i)$).

Note that the expected number of pairs $(i,j) \in S'\times S'$ such that we assigned $\sigma(i) = \sigma(j)$ at the first step is approximately $|S'|^2 / k = 4k^3/n^2 = 4n^{ - 1/2 + 3\varepsilon}$. The probability that there is at least one such pair is at most $4n^{ - 1/2 + 3\varepsilon}$. We can ignore this event. So we can assume that the assignment algorithm always performs steps 1 and 2 (and never performs step 3). We similarly modify the procedure for assigning values to $\sigma(i)$ for $i\in T'$.

Then for every pair $(i,i+1) \in A$, $\sigma(i)$ is distributed uniformly at random in $\{1,\dots, k\}$ and $\sigma(i+1)$ is distributed uniformly at random in $\{n-k+1,\dots, n\}$ or the other way around. Values of $\sigma(i), \sigma(i+1)$ for different pairs $(i,i+1) \in A$ are independent.

We have that for every $(i,i+1) \in A$, the random variable $n+1 - |\sigma(i) - \sigma(i+1)|$ is distributed as a sum of two random variables uniformly distributed in $\{1,\dots,k\}$. Our goal now is to estimate the minimum of $2k^2/n$ independent copies of such random variable.

Consider two i.i.d. random variables $x \in_U (0,1)$ and $y \in_U (0,1)$. Then $\lceil k x\rceil$ and $\lceil k y\rceil$ are uniformly distributed in $\{1,\dots,k\}$. So $n+1 - |\sigma(i) - \sigma(i+1)|$ is distributed as $\lceil k x\rceil + \lceil k y\rceil \approx k(x+y)$ (up to an additive constant 2).

Lemma. Let $x_1,\dots,x_N$ and $y_1,\dots,y_N$ be independent random variables uniformly distributed on $(0,1)$. Let $z_i = x_i + y_i$. Then $$\mathbb{E}[\min_i\{z_i\}] \approx \sqrt{\frac{\pi}{2N}}.$$

Proof: The distribution function $F(t)$ (c.d.f.) of $z_i$ is given by $$F(t) = \begin{cases} 0, & \text{ if } t \leq 0;\\ t^2/2, & \text{ if } 0\leq t \leq 1;\\ 2t-1 - t^2/2, & \text{ if } 1\leq t \leq 2;\\ 1, &\text{ if } t > 1. \end{cases} $$

Note that if $z_i < z_j$ then $F(z_i) < F(z_j)$. Thus $\arg\min_i F(z_i) = \arg\min_i z_i$. Each $F(z_i)$ is uniformly distributed on $[0,1]$. Thus $\min_i\{ F(z_i)\}$ has probability density $N\cdot (1-s)^{N-1}$. We have $$\mathbb{E}[\min_i\{z_i\}]=\int_0^1 N\cdot s^{N-1} F^{-1}(s) ds \approx \int_0^1 N\cdot s^{N-1}\sqrt{2s} ds = \sqrt{2}\cdot N \cdot B(N,3/2)\approx \sqrt{2}\cdot\Gamma(3/2) \cdot N \cdot N^{-3/2}=\sqrt{\pi/(2N)},$$ where $B$ is the beta function and $\Gamma$ is the gamma function.


Thus the minimum of $2k^2/n$ independent copies of $x+y$ approximately equals $\sqrt{\pi n/(4k^2)}$ in expectation. The minimum of $2k^2/n$ independent copies of $k(x+y)$ approximately equals $\sqrt{\pi n}/2$ in expectation.

We get that the minimum over all pairs $(i,i+1)\in A$ of $n+1 - |\sigma(i) - \sigma(i+1)|$ is approximately $\sqrt{\pi n}/2$. If $(i,i+1)\notin A$ then $n+1 - |\sigma(i) - \sigma(i+1)| \geq k +1 \gg \sqrt{n}$. We showed that $\Delta$ is approximately equal to $\sqrt{\pi n}/2$ in expectation.

Part II. Simulation.

Our estimate agrees with data from a computational simulation, which I did in MATLAB. For $n$ between $100$ and $10000$, $$\sqrt{\pi n}/2 -0.75 < \Delta_{\text{simulation}}(n) < \sqrt{\pi n}/2 + 0.1,$$ where $\Delta_{\text{simulation}}(n)$ is an estimate for $\Delta$ obtained in the simulation. Simulation results

Part III. An estimate for $c_t^n$.

We prove first that $c_t^n \leq n (2t)^{n-1}$. There are $n$ ways to choose $\sigma(1)$. Given $\sigma(1)$, there are at most $2t$ ways to choose $\sigma(2)$: $$\sigma(2)\in \{\sigma(1) -2t,\dots,\sigma(1)-1,\sigma(1)+1,\dots,\sigma(1)+t\}.$$ Similarly, there are at most $2t$ ways to choose each $\sigma(i)$ for $i\in\{2,\dots, n\}$. So the total number of permutations $\sigma$ (for a given value of $t$) is at most $n (2t)^{n-1}<(\beta t)^n$ (for some $\beta>0$).

Now we prove that there are at least $(t!)^{\lfloor n/(t+1)\rfloor} > (\alpha t)^n$ permutations (for some $\alpha > 0$). Consider the set of permutations of the form $$[1, 2, \dots t], (t+1), [(t+1) + 1, (t+1) + 2 \dots (t + 1) + t], 2(t+1), [2(t+1) + 1 \dots 2(t+1) +t], 3(t+1) \dots$$ where we can arbitrarily permute numbers inside each bracket. Note that $|\sigma(i) -\sigma(i+1)| \leq t$ for all these permutations.

Let's count the number of such permutations. There are $t!$ ways to arrange numbers inside each bracket. There are ${\lfloor n/(t+1)\rfloor} $ brackets. So the total number of permutations is (at least) $(t!)^{\lfloor n/(t+1)\rfloor}$. By Stirling's formula, it is at least $(\alpha t)^n$ for some $\alpha > 0$.

| cite | improve this answer | |
  • $\begingroup$ Very nice proof. Do you happen to know a reference where this subject is treated more extensively ? $\endgroup$ – Ewan Delanoy Nov 20 '12 at 11:12
  • $\begingroup$ @EwanDelanoy, thanks! No, unfortunately, I don't know any references. $\endgroup$ – Yury Nov 20 '12 at 13:53
  • 2
    $\begingroup$ Wow, it took quite some time for me to go through the proof, but really really comprehensive answer Yury! Great answer, and many thanks :-) $\endgroup$ – quantumelixir Nov 21 '12 at 10:30

A naive approximation for the second question. The restriction of having a permutation of length $n$ with consecutive distances not greater than $t$ prohibits a total number of pairs given by $ 2 \times (1 + 2 + \cdots n - t-1) = (n-t)(n-t-1)$

The total number of a priori pairs is $n(n-1)$. The probability that a given pair violates the restriction is then $$\frac{(n-t)(n-t-1)}{n (n-1)}$$

If we assume (very rough approximation!) that the probability that all the pairs violate/fit the restriction are independent, the we have a probability of "success" given by

$$p_{n,t}=\left(1-\frac{(n-t-1)(n-t)}{n (n-1)}\right)^{n-1} = \left( \frac{t \, (2 n - t -1)}{n \, (n-1)} \right)^{n-1}$$

If we further assume that $n \gg 1$:

$$p_{n,t} \approx \left(\frac{2 t }{n}\right)^{n-1} e^{-(t-1)/2} $$

The number of allowed permutations then can be approximated as

$$c_{n,t} = n! \; p_{n,t} \approx \left(\frac{2 t }{e}\right)^{n-1} e^{-(t-1)/2} \sqrt{2 \pi \, n}$$

I've made a few simulations and the approximation does not seem too bad.

| cite | improve this answer | |

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.