The Kolmogorov Complexity of Irish traditional dance music

The Kolmororov Complexity [2, 5] of s sequence of letters is defined to be the length of the shortest program which will output that sequence of letters. As such, it is known to be uncomputable. Nonetheless, it can be estimated in a reliable way using lossless compression techniques such as Lempel-Ziv compression [8]. Li and Sleep [3, 4] show how these ideas can be used to estimate the similarity between different melodies.

The Lempel-Ziv (family) of lossless compression algorithms is known to be asymptotically optimal [6]. In the following sections, we examine the compression of a common reel Sally Gardens using two Lempel-Ziv variants: LZ77 and LZ78.

We select a (musically) representative sample of 60 reels from the online database [1]. The most common key for reels is D (about 45%), the second most common G (about 30%), with the remaining 25% in a variety of keys. We examined tunes from across the key spectrum. To enable fair comparison, we only examined “standard” length reels, i.e. ones with exactly two parts: In our representation this corresponds to a sequence of 128 letters. In summary,

• •

  Each reel has two parts

• •

  Each part has exactly 64 (quaver) notes

• •

  Each bar has exactly 8 (quaver) notes

• •

  Any longer notes are written as sequences of quavers adding up to the same note length. So, any crotchet in the tune is written as two identical quavers, any dotted crotchet as three quavers, etc.

• •

  Each part of the tune has 8 bars, and overall the tune has 16 bars

Our analysis of the representative sample of 60 reels gives an average compression ratio of 2.79 (to two decimal places) with standard deviation 0.46.

From the same database [1] we select a (musically) representative sample of 60 jigs. We examined tunes from across the key spectrum. To enable fair comparison, we only examined “standard” length jigs, i.e. ones with exactly two parts: In our representation this corresponds to a sequence of 96 letters. In summary,

• •

  Each jig has two parts

• •

  Each part has exactly 48 (quaver) notes

• •

  Each bar has exactly 6 (quaver) notes

• •

  Any longer notes are written as sequences of quavers adding up to the same note length. So, any crotchet in the tune is written as two identical quavers, any dotted crotchet as three quavers, etc.

• •

  Each part of the tune has 8 bars, and overall the tune has 16 bars

Our analysis of the representative sample of 60 jigs gives an average compression ratio of 2.61 (to two decimal places) with standard deviation 0.40.

In Figure 2 we display our results, where we use 20 bins for the different compression ratios.

In order to be able to account for differences in the (approximation to the) Kolmogorov complexity of reels versus jigs, we need to take in to account the different lengths of these melodies / alphabet strings. It should be clear that, for finite fixed alphabet size, the compression ratio will in general be larger for longer strings (for longer strings, letters will repeat more often). In our case, the comparison is between strings of length 96 (jigs) and length 124 (reels). We would expect, all other things being equal, a slightly higher compression ration for the longer string (reels). To account for this effect (which is not intrinsic to the music), we generate sequences of random letters from the 13 letter alphabet $A=\{a,b,c,d,e,f,g,h,i,j,k,l,m\}$. This corresponds to the lower octave, and most of the higher octave in the tunes considered (the highest note in the high octave is not used).

For strings of length $[50,100,150,200]$ we get average compression ratios of $[1.13,1.24,1.33,1.40]$ respectively. In the region of interest (lengths between 96 and 128), the dependence of compression ratio on string length is approximately linear. In particular, for random strings of length 96/128 chosen from the alphabet $A$, we get average compression ratios of 1.23/1.29 respectively. We use this to normalize the results from Sections 4 and 3 to calculate what we would expect the average compression ratio for jigs to be, were they of the same length as reels: $\frac{1.29}{1.23}(2.61)\approx 2.73$.

Our result therefore is to conclude reels have slightly less intrinsic (Kolmogorov) complexity than jigs, given the two averaged, normalized compression ratios of approximately 2.79 and 2.73 respectively. Separately from comparing categories such as reels and jigs, our estimations of Kolmogorov complexity can be used to distinguish more repetitive tunes from less repetitive ones. This should be of use to those learning Irish traditional dance music, who may want to start with tunes of low Kolmogorov complexity. We emphasize that we are not concerned with efficiency issues in actually compressing the sequences considered, and in fact the compression algorithms described may be very inefficient. We are only concerned with the size (length) of the resulting compression, how we can minimize that to get the best estimate of Kolmogorov complexity, and what that size can tell us about the tune being compressed.

In future work, we plan to apply this analysis to other (large) families of tunes, such as hornpipes and polkas. We are also exploring how Kolmogorov complexity, and other techniques from information theory can be applied to harmonic music (for which there is a much larger corpus).