3
$\begingroup$

Continuing on the theme from my last question Variable Length Encoding of Integers, I have come up with a simple encoding scheme, but for which an efficient algorithm eludes me.

The constraints are simple enough: no (binary representation) number is allowed where it is divisible by 3, or a subset (prefix) of that representation is divisible by 3.

To terminate the number two bits are added so that the number is divisible by 3.

For example 1101 is allowed since neither 1101, 101, 01 nor 1 are divisible by 3.

However, 1011 is not allowed since 11 is divisible by 3.

The representation 1101 would then have 10 prepended to make it divisible by 3 (101101).

All this allows a stream of bits to be read, at each point testing to see if the number is divisible by three. If not, keep reading the next bit, until it is divisble by three. Hence allowing for a (unique) variable length encoding.

My question is about the mapping of integers on to this encoding scheme. However hard I try I can't seem to create a straightforward algorithm to do the mapping. Is there one?

Clarification

The examples above are to be read in the order right to left. So any prefix is on the right. So the stream of bits will be read right to left.

$\endgroup$
3
  • $\begingroup$ I had it my head the natural reading order of binary numbers, LSB on the right. I've clarified this above. $\endgroup$ Commented Jul 29, 2014 at 21:10
  • $\begingroup$ It sounds as if you're not going to encode those numbers that aren't allowable (i.e., have a prefix divisible by three). If that's the case, the allowable numbers are fairly rare. I counted only 6765 allowable numbers out of the first 1048577, which is just a hair more than 0.6%. $\endgroup$ Commented Jul 29, 2014 at 21:19
  • $\begingroup$ I realise a more efficient encoding is possible, say 'divisible by 7'. But the entire set of integers - without bound or gaps - is mapped onto this encoded/restricted set. So given a potentially infinite stream of bits to read, we can succesfully work out when a number terminates and convert back to conventional binary notation. $\endgroup$ Commented Jul 29, 2014 at 21:31

2 Answers 2

2
$\begingroup$

The first step is counting the number of strings of length $n$ in which no non-empty prefix is divisible by $3$. Surprisingly enough, these are given by Fibonacci numbers $F_n$ for $n > 0$ (exercise). I will show how to map the natural numbers to strings of these forms, from which your encoding will follow. Divide the natural numbers into parts of lengths $1,F_1,F_2,F_3,\ldots$. Using the formula $F_1+\cdots+F_n = F_{n+2}-1$, we see that numbers $m$ in the range $F_{n+1} \leq m < F_{n+2}$ will be encoded using strings of length $n$; additionally, $0$ will be encoded by the empty string. Given $m$, it should be straightforward to find $n$, for example using the Fibonacci recurrence or the approximate formula $F_n \approx \frac{1}{\sqrt{5}} \left(\frac{1+\sqrt{5}}{2}\right)^n$.

Suppose $m > 0$, and let $x = m - F_{n+1}$, so that $0 \leq x < F_n$. It remains to encode $x$ as a string of length $n$ in which no non-empty prefix is divisible by $3$. The idea is to interpret the recurrence $F_n = F_{n-1} + F_{n-2}$ accordingly. There are many ways of doing it, here is one. There are $F_{n-1}$ legal strings of the form $0s$, and $F_{n-2}$ legal strings of the form $1\alpha s$, where $\alpha$ is determined so neither $\alpha s$ nor $1\alpha s$ are divisible by $3$. We can encode the first $F_{n-1}$ values of $x$ using strings of the form $0s$, and the latter $F_{n-2}$ values of $x$ using strings of the form $1\alpha s$. To determine $s$ itself, we run this procedure recursively. The base case is the unique string of length $1$, namely $1$.

To illustrate the entire procedure, let us show how to encode $m = 7$. First we find that $F_5 = 5 \leq 7 < F_6 = 8$, and so $n = 4$ and $x = 7 - 5 = 2$. We write $F_4 = F_3 + F_2 = 2 + 1$, so that the encoded string is of the form $s_4 = 1\alpha_4 s_2$. The string $s_2$ itself is the encoding of $x - F_3 = 0$ of length $2$. Writing $F_3 = F_2 + F_1 = 1 + 1$, we see that $s_2$ has the form $s_2 = 0s_1$. The string $s_1$ itself is the encoding of $0$ of length $1$, which is $s_1 = 1$. Therefore $s_4 = 1\alpha_4 s_2 = 1\alpha_4 0s_1 = 1\alpha_4 01$. If $\alpha_4 = 0$ then $1001$ would be divisible by $3$, so $\alpha_4 = 1$ and the output is $1101$.

$\endgroup$
1
  • $\begingroup$ Thank you for your in depth answer. It is surprising that Fibonacci numbers are involved! $\endgroup$ Commented Jul 30, 2014 at 9:57
1
$\begingroup$

After a bit of head scratching, here's an alternative way of doing it:

Prepare a binary tree to represent all possible strings of binary digits. There is one root node which represents the empty string. At each node in the tree we can make a decision to either go left (0) or right (1). In this way any node is reachable, and represents a unique string of digits.

Now we prune the tree. We mark the remainder after dividing the binary number by three against each node. Discard any nodes which are reached by going right (1) and whose remainder is zero. Also discard all their children as these are unreachable.

This leaves a somewhat more sparse tree. The main property of this tree being, that whichever path is taken through it, the binary number is never divisible by three. Also, every node either has two children or one child.

For the mapping. Start from the root node. If there are two children go left (0) or right (1) as before, output 0 or 1 accordingly and move to the next digit. If there is one child, then there is no choice, take that child and output 0, stay on the same digit. Repeat until you reach the end of the binary number.

In this way, all nodes in the sparse tree are uniquely reachable.

The main thing to note, is that if we input successive binary numbers, the encoded output is not ordered strictly. Here's the output for the first few numbers (read right to left):

0 → 0
1 → 1
10 → 10
11 → 101
100 → 100
101 → 10001
110 → 1010
111 → 1101
1000 → 1000
1001 → 1000001
1010 → 100010
1011 → 100101
1100 → 10100
1101 → 110001
1110 → 11010
1111 → 11101
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.