## Generating Mini-Crosswords

Introduction

The New York Times publishes “mini-crosswords,” which are crossword puzzles on a relatively small grid (usually 5×5), without any black squares, so that every row and column of the grid must spell a word. The figure below shows an example of a solution to such a puzzle.

How hard is it to create these puzzles? This post is motivated by a recent College Mathematics Journal article (see Reference (1) below) that considers this question, and describes an approach using the Metropolis-Hastings algorithm to randomly sample instances of puzzles.

But instead of just randomly sampling one puzzle at a time, can we actually enumerate all possible puzzles? In particular, my idea was to reduce the problem of finding a crossword puzzle solution to that of finding a (generalized) exact cover with appropriately crafted constraints. This would be handy, because we already have code for solving exact cover problems, using Knuth’s Dancing Links (DLX) algorithm (see here and here for similar past exercises).

Crossword as an exact cover

To state the problem more precisely: given a positive integer $n$ indicating the size of the grid ($n=5$ in the above example), and a dictionary $W$ of $m$ words each of length $n$ over alphabet $\Sigma$, we must construct an exact cover problem whose solution corresponds to a placement of letters in all $n^2$ grid positions such that each row and column spells a word in $W$.

To do this, we construct a zero-one matrix with $2mn$ rows, each corresponding to placing one of the $m$ dictionary words either “across” (in one of the $n$ rows of the puzzle) or “down” (in one of the $n$ columns). A solution will consist of a subset of $2n$ rows of the matrix: $n$ “across” words, one in each row of the puzzle, and $n$ “down” words, one in each column, each pair of which intersect in the appropriate common letter.

To represent the constraints, we initially need $2n^2\cdot|\Sigma|$ columns in our matrix (where $|\Sigma|=26$), each indexed by a tuple (puzzle row $i$, puzzle column $j$, alphabet letter $c$, across or down). For a given matrix row– representing placement of a word with letters $(w_1, w_2, ..., w_n)$ in a particular location and orientation in the puzzle grid– we set to one those $n\cdot|\Sigma|$ columns corresponding to the $n$ different $(i,j)$ grid locations where the word will be placed in the puzzle… where for each alphabet letter $c$, we set the “across” column to one if and only if either

• the word is “across” and $c$ matches the letter of the word in this location (i.e., $c=w_j$), or
• the word is “down” and $c$ does not match the letter of the word in this location (i.e., $c \neq w_i$).

If neither of these conditions is satisfied, we instead set the “down” column to one. The following Python code produces the number and list of (row, column) pairs of the resulting sparse matrix.

letters = 'abcdefghijklmnopqrstuvwxyz'
m = len(words)
n = len(words[0])
print(m * n * 2 * (n * 26), file=file)
b = [True, False]
for row, ((w, word), k, across) in enumerate(
product(enumerate(words), range(n), b)):
for col, (i, j, letter, horiz) in enumerate(
product(range(n), range(n), letters, b)):
if ((i if across else j) == k and
(word[j if across else i] == letter) == (across == horiz)):
print(row, col, file=file)


However, we’re not quite finished. Although the desired end result is a crossword with distinct words in each row and column, such as the 6×6 solution shown in (a) below, as Howard observes in the CMJ article, there are many valid solutions that use the same word more than once, including the extreme cases of symmetricword squares” such as the one shown in (b) below.

(a) 6×6 mini-crossword, where the word in each row and column is unique. (b) 6×6 “word square,” a symmetric grid with each row and corresonding column containing the same word.

We can eliminate this duplication by adding $m$ “optional” columns to the zero-one matrix, one for each word in the dictionary, and solve the resulting generalized exact cover problem, so that each word may be used at most once in a solution.

Results

All of the source code is available here, as well as on GitHub. My initial test used the 4×4 case discussed in Howard’s paper, with his dictionary of 1826 words. He describes a process for estimating the total number of possible puzzles by repeated sampling using the Markov chain Monte Carlo approach: “We estimate that there are approximately
73,000–74,000 distinct puzzles each with no repeated words.” This is pretty accurate; it turns out that there are exactly 74,339 (each contributing two symmetric pairs of solutions to the generalized exact cover problem, for a total of 148,678 solutions).

References:

1. Howard, C. Douglas, It’s Puzzling, College Mathematics Journal, 49(4) September 2018, p. 242-249 [DOI]
2. Knuth, D., Dancing Links, Millenial Perspectives in Computer Science, 2000, p. 187-214 (arXiv)

This entry was posted in Uncategorized. Bookmark the permalink.

### 3 Responses to Generating Mini-Crosswords

1. Marc says:

Hi! Can you provide the words.txt file used in the example?

• I left the word list out of the repository, since it is technically supplemental material to Howard’s paper, which unfortunately is behind Taylor&Francis’s paywall. So although it’s not my place to post it, it’s possible (likely?) that he would be happy to provide it if asked.

In the meantime, there are several other longer word lists here, sorted by frequency, used in various past analyses.

• Marc says:

Thanks! Yes – I could not access it because of the paywall. And thanks for the great content!

This site uses Akismet to reduce spam. Learn how your comment data is processed.