All possible encodings from a numerical mapping

March 2018 ยท 6 minute read

I haven’t done interview prep in a while, and I decided to get back into it after I saw a practice problem that caught my eye. I got it from this mailing list.

The problem

Suppose we have a mapping of letters to numbers, or an encoding, such that $a = 1, b = 2, \cdots z = 26$.

Given a string of digits, find the number of possible ways that this message could be decoded.

For example, 111 has three different mappings:

111
aaa

Table: Mapping 1

111
ka

Table: Mapping 2

111
ak

Table: Mapping 3

The approach

In an interview, you should never begin by coding. Think through the problem first, and identify a strategy. You want to have a high level overview of what you’re going to do before you do it, otherwise you run the risk of going down a tangent, and then realizing you got too focused on little details and ended up with a suboptimal, or even totally incorrect approach.

You should also ask your interviewer questions. For example, for the sake of our problem, are we guaranteed valid strings? Should we worry about the total length of strings? What do we do when passed an empty string? What do we do when we receive a string without numbers or that is a valid string but isn’t valid for our test cases? Can there be zeros before or after the “main” sequence?

For the purposes of this exercise, let’s assume we’re getting valid strings, and that all we have to worry about is the argument, not parsing input, or worrying about getting the result back to the user. Don’t worry about the empty string. Usually your interviewers aren’t out to trip you up on minutiae, and all of my interviews have generally been more focused on the actual algorithm and how well I think out my answer rather than the smaller details.

One of the first things that came to mind was that this problem can be broken down into subproblems, and that each letter is restricted to two digits at maximum. That is, because z = 26, and all letters are between [a, z] (inclusive). This means, that every single digit must correspond to a letter, and that every pair of digits can correspond to a letter.

The recursive approach

The first thing that came to mind for me is that this problem can be broken down into recursive subproblems.

Suppose $s$ is a string, and $s_i$ represents the string from elements 0 to the $i^{th}$ element.

The number of ways $s * i$ can be interpreted is equal to the number of ways $s * {i - 1}$ can be interpreted, unless the last two characters form a letter (that is, $0 < s * (i - 1) * 10 + s * (i - 2) \leq 26$). Then we have the number of ways that $s*(i - 1)$ can be interpreted and all the ways $s * (i - 2)$ can be interpreted.

In recursion, it’s really important to sort of just trust the recursive property. It can be hard to wrap your head around, and sometimes you might even overthink it and convince yourself you’re wrong, but as long as you have your subproblems defined correctly and good base cases, you will be fine.

Our base cases are $s_0$ and $s_1$. They can both be interpreted in only one way.

def recursive_count(s: str) -> int:
    """
    Determines the number of ways a string or array of digits can be decoded
    given the decoding rules provided by the problem.
    s: an array of single digits integers (every digit is between 1 and 26,
    inclusive)
    :return: the number of ways `s` can be decoded
    """
    if not s or len(s) == 1:
        return 1

    count = 0

    if int(s[-1]) > 0:
        count = recursive_count(s[:-1])

    prev_two = (int(s[-2]) * 10) + int(s[-1])

    if 0 < prev_two < 27:
        count += recursive_count(s[:-2])
    return count

This is not the most efficient way to go about this, however. We can optimize this further by memoizing. You may notice that this problem has a very familiar structure. We can break it down as such:

Suppose $f(x)$ is the function that returns the number of ways some string $x$ can be decoded. Then $f(x) = f(x*{n - 1}) + f(x*{n - 2})$. Of course, the second part of that statement is conditional, we have to check whether two letters can be decoded to a letter (e.g. are they less than or equal to 26), but this is very close to the recursive equation for the Fibonacci sequence.

Memoizing

We can improve the efficiency of this solution by caching the results of each subproblem in an area. This is called memoizing (not memorizing!).

In this situation, suppose we have some array A. We will enforce that A[i] returns the number of ways that the subset of the string from the first character up to the $i - 1^{th}$ character can be decoded. If we see that two characters can be decoded as a character, we will add the number of ways that the substring from $0$ to $i - 2$ can be interpreted (that’s just the string if we remove the pair of characters we just looked at).

In a way, we’re almost looking at paths. How many ways can we traverse the string when we can either traverse one letter at a time, or under certain conditions possibly skip a letter (since when you interpet a pair of letters, the next letter you decode is the one after that pair).

def dp_count(s :str) -> int:
    """
    Determines the number of ways a string or array of digits can be decoded
    given the decoding rules provided by the problem.
    s: an array of single digits integers (every digit is between 1 and 26,
    inclusive)
    :return: the number of ways `s` can be decoded
    """
    # add an extra element because cache[len(s)] should return our answer
    cache = [0] * (len(s) + 1)
    cache[0], cache[1] = 1, 1

    for i in range(2, len(s) + 1):
        if int(s[i-1]) > 0:
            cache[i] = cache[i-1]

        prev_two = (int(s[i-2]) * 10) + int(s[i-1])

        if 0 < prev_two < 27:
            cache[i] += cache[i-2]
    # this simply returns the last element in the array
    return cache[-1]

Now we have a solution with $O(n)$ space and time complexity, which is as efficient as you can get.

A nice little trick you can use to determine what the best possible efficiency for a particular problem is to think about how you would go about solving the problem. To me, it makes sense that this is an $O(n)$ problem because we have to look at all of the elements in the string/byte array at least once if we want to find out how many ways there are to decode it. We also know that we only have to look at up to two characters at any point because the upper bound for a decoded letter is two digits, but it has to be at least one digit, which is why we essentially have a sliding window of two characters.