```

```

# Hierarchical Correlation Plots

Correlation

Correlation is a method of characterizing the similarity of the contour of two sequences of numbers. In its most basic form, correlation is the sum of the multiplication between corresponding numbers in a pair of sequences:

To normalize the resulting correlation values into the range from -1.0 to +1.0, use Pearson's product moment correlation:

Correlation Examples

Here are some example correlation calculations to give an idea for how correlation values correspond to curves.

First consider the simplest case of comparing two sequences with two numbers. In the following figure the first pair of lines are slanted in the same direction. This causes the correlation value to be +1.0. The correlation value is often called the r-value in statistics, so in this case the text "r=1" means the correlation is 1.0.

In the second example of the following figure the lines have opposite slopes (the first is pointing up, and the second is pointing down). In this case the lines are negatively correlated.

The following set of examples demonstrates that normalized correlation (Pearson correlation) measures the gross contour of a pair of line. In this case the slope of the second line is varied; however, the correlation value remains constant at the maximum correlation.

Also notice that a Pearson correlation for a purely horizontal line (sequence with all numbers the same) cannot be caculated becase one or both of the terms inside the square root of the denominator is 0, making the entire value become 0/0 which is an undefined number. In the program which generates plots described below from the correlation values, the number 0/0 is assigned to be the correlation value 0.0.

The following figure demonstrates the correlation values for several sequences of length 3.

Finally, here is a demonstration of correlation for longer sequences behaves. In the following example, the red curve is a smooth arch consisting of 11 numbers. The blue curve represents a second sequence which consists of the numbers from the red curve plus a random offset which is gradually increased in the pairs of curves in the figure. Notice that for small random fluxuations, the correlation remains high, but as the randomness of the blue curve increases, the correlation between the two sequences drops.

```
```

Hierichical Correlation Plots

One problem that occurs when correlating two sequences is that a single number cannot describe in any detail the similarity of two sequences. Correlation can only describe the overall similarity of the two sequences when both are consider in their entirity. Consider the following two sequences (one in red and the other in blue):

The correlation between these two sequences is 0.235. This means that the two sequences are slightly similar to each other. This value is virtually meaningless, and two mostly random sequences could also generate a similar correlation value. The fact that there are two arches in the blue sequence and one arch in the red sequences cannot be described in the single correlation number. In order to do this sort of comparison, the two sequences have to be chopped up into smaller pieces. For example, the sequences could be divided into three sections:

In this case when dividing the sequences up into the first 25%, the middle 50% and the last 25%, three high correlation values pop out of the sequences rather than one slightly positive correlation values.

To examine the internal similarity of two sequences, cut the sequences up into smaller pieces. For example, if both sequences have six numbers {A,B,C,D,E,F}, then {A,B,C}, {D}, {C,D,E,F} and {A,B,C,D,E} are all sub-sequences which can be correlated in addition to the entire sequence.

How many sub-sequences should the two sequences be cut up into? In the above example, using three unequal segments best demonstrates the internal structure between the two sequences. However, in the general case where you do not know anything about the internal structures of the two sequences beforehand, it is best to segment the sequences into all possible sub-sequences.

The following plot schematic shows a two-dimensional plot which can display all of these sub-sequence correlations simultaneously. Note that the bottom row is not displayed in actual plots, and is just shown for reference to the original sequence (correlation with a single number isn't interesting).

Alternatively, each row can be stretched to create a rectangular plotting region:

```
```

Plot Colorization

Using Pearson's product moment correlation, the correlation values will be in the range from -1 to +1. This range can be colored with one range from -1 to +1, or it can be colored with two ranges: one from 0 to -1, and another from 0 to +1. Here are three possible coloration schemes for the hierarchical correlation plot (which could also be reversed):

Alternatively, in order to view more subtle changes in correlation, more colors can be fitted into the range between -1 and +1, such as using a hue value:

Where red/orange = high correlation, yellow = moderate correlation, green = slight correlations, light blue = slight negative correlation, dark blue = moderate negative correlations, and purple = strong negative correlation.

```

```

Example Hierarchical Correlation Plots

Now recall the double and single arch sequences from a previous section. Here is a hierarchical correlation plot of those sequences, with the two sequences underneath for reference.

In the plot, high correlation is displayed as white, high negative correlation is displayed as black, and low correlation is displayed as gray. The plot now clearly demonstrates the three interesting sub regions of the pair of sequences. The far left and right sides of the plots are white which indicates that the two sequences are strongly correlated at their beginnings and ends. The dark central region indicates that they are strongly anti-correlated (doing opposite things) in the central region. This is where the red curve rises when the blue curve false, and vice-versa.

Example Temposcape Plots

A comparison of the beat-by-beat performance tempos for Chopin's Mazurka in F major, Op. 68, No. 3. Click on the thumnail images in order to view a large version of the plot.

 chiu1999 indjic2001 luisada1991 rubinstein1938 rubinstein1961 smith1975

Here are hue-colored plots of the same correlation comparisons. (The hue values need correcting).

 chiu1999 indjic2001 luisada1991 rubinstein1938 rubinstein1961 smith1975

Here are rectangular plots of the same hue plots above:

 chiu1999 indjic2001 luisada1991 rubinstein1938 rubinstein1961 smith1975
```

```

Polycorrelation plots

The basic method of plotting the correlation between two performances can also be extended to comparisons between multiple performances. In this case a color is displayed to indicate which of several sequences is closeset to the original sequence.

The following polycorrelation plots are shown for each performer. The colors in the plot represent the closest performance according to the correlation measurement at the given point in the plot. The colors represent the following performances:

 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975

The most interesting plots here are Rubinstein 1938 and Rubinstein 1961. In this case both plots show in the first half of the performance, that they are each the best fit to each other. In the Rubinstein 1938 plot, the dark blue represents the Rubinstein 1961 performance. The dark blue color progresses from the bottom of the plot to the top of the plot. Likewise with the light blue coloring of the Rubinstein 1961 plot.

If colors progress from the bottom to the top then that is a good case for influence between the performances. Colors found at the top of the picture which are not connected to the same color at the bottom of the plot shows an large-scale structural similarity, but not related to any surface similarities. In this case it is more likely that the similarity results from indirect influence on the performance, or there is no influence at all between the performances. Other items of interest:

• The bands of purple and red in the Indjic 2001 plot are on the phrase level which is perhaps demonstrating the change of (tempo) performance style between successive phrases.
• The Chiu plot shows potential influence from the Rubinstein 1961 recording with the moderately significat amounts of dark blue in the beginning and ending sections.

The absolute correlation values of the most similar performance can also be superimposed onto the polycorrelation plots to indicate the stregth of the correlation between the two performances. In the following plots, the brighther colors indicate a stronger correlation value, while a darker color indicates a weaker correlation.

 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975
In these example plots, the Smith 1975 performance shows the least comonality with the other performances. The Indjic 2001 performances does not show any strong correlation with other performances on the low level, but the large-scale tempo structure strongly resembles the Luisada 1991 performance.

Hierarchical Average Plots

These are arranged in a similar manner as the correlation plots, but the average value of a single sequence is displayed rather than the correlation between two sequences.

 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975

In these plots, the poco più vivo section is displayed in white since it is faster than the average tempo of the piece. Black regions at the bottom of the pictures indicate where the tempo slows down at phrase boundaries.

Of note:

• Smith 1975 has very arched phrasing. playing the middle of phrases faster, and the phrase endings slower. Smith 1975 also does a
• Indjic 2001 is also very regularly phrased, even more so than Smith 1975. In the return of the A section at the end of the piece (the last two phrases), the phrases are slightly less arched, and they show a tendancy to be double phrases (with a slow down in the middle of the 8 bar period) which is less prominent in the opening of the piece.
• Notice that the more recent performances smooth out the high-level tempo averages. Indjic and Luisada are smooth at the top of the plots, Smith 1975 is chunky, but fairly smooth, and both Rubinsteins are much more striated at the top levels.
• Chiu 1999 shows a tendance to break phrases into 4 bar fragments which is in particular contrast to Smith 1975 which generally shows 8-bar phrasing.

Here are the same performances plotted with a logarithmic vertical scale:

 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975
Here are the same plots displayed in color:
 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975
And the same plots displayed in color on a vertical log scale:
 Chiu 1999 Indjic 2001 Luisada 1991 Rubinstein 1938 Rubinstein 1961 Smith 1975

The hue colorizations don't show as much of the phrasing structure as the plain black and white pictures, but they do demonstrate the contrast between older and newer performances. The newer performances show in the top row have a greater contrast between slow and fast tempos. The bottom three older performances show less of a contrast between slow and fast sections (demonstrated by the smaller red section of the vivo section).

Below is a pair of plots which represent an average performance for the mazurka which was created by averaging the note-by-note tempos before generating the plots. Notice that the average performance closely matches the Chiu 1999 performance.

Average Difference Plots

The average tempos of mazurka performances consistently slow down over time. The following figure shows that for all mazurkas which have currently been analyzed, the average tempo drops about 3 beats per minute over each decade.

For example, here are the average tempos for Mauzurka in F major, Op. 68, No. 3:

 performer year tempo Indjic 2001 105 Chiu 1999 115 Luisada 1991 105 Smith 1975 129 Rubinstein 1961 129 Rubinstein 1938 134

But what does it mean for a performance to be slower or faster than average compared to another performance. How is it slower or faster? In other words, is the difference just due to a constant change such as slowing down the speed of a record, or does the difference in tempo change clump in certain regions of the performance.

The following hierarchical average tempo difference plots address this problem of characterizing the change in average tempos between performances.

The following plots compare the tempos of a pair of performances. If the first performance is faster than the second performance at a give time-scope in the piece, the plot is colored red (i.e., red = hotter; faster). If the first performance is slower, then the plot is colored blue (i.e., blue = cooler; slower). If the tempos are the same, the plot is colored white (the same within one beat per minute, but this is not important for this particular plotting scheme).

In these plots, the comparison of average tempos for the entire performance is displayed in the top-most corner of the triangle plots. For example when comparing Indjic 2001 with Chiu 1999, the top of the triangle is colored blue because Indjic's average tempo of 105 is slower than Chiu's average tempo of 115 MM. Likewise, when comparing Chiu 1999 to Indjic 2001, the same region is colored red because 115 is a faster tempo than 105. (notice that in the following grid of comparisons, the bottom left half is a color mirror of the top right half).

 indjic2001 chiu1999 luisada1991 smith1975 rubinstein1961 rubinstein1938 indjic2001 chiu1999 luisada1991 smith1975 rubinstein1961 rubinstein1938

Notice in the grid of comparisons above that the older performances have a predominantly red color for the plots, which indicates that they are mostly faster than more recent performances. However, their plots are not solid red, which would be indicative of a tempo which is continually faster for all beats and sections throughout the performance.

Note in particular that the poco più vivo sections of the mazurka which occurs just after the mid point in the composition is nearly always demarked between performance comparisons. So while the more recent performances have been slowing down, the tempo of the vivo section has been increasing. It is the non-vivo parts of the composition which are slowing down, and there is an increase in the tempo range throughout the piece which is facilitated by a decrease in overall tempo.

Here is a table showing the average tempo, as well as the tempo of the vivo and non-vivo parts of the composition which is a better characterization of the tempo changes:

 performer year avg tempo non-vivo avg. vivo avg Indjic 2001 105 68 171 Chiu 1999 115 73 198 Luisada 1991 105 66 187 Smith 1975 129 78 249 Rubinstein 1961 129 81 225 Rubinstein 1938 134 86 224
```

```

Contoured Average Difference Plots

The following set of plots compare the tempo differences in a more refined manner. The previous section only disinguished between two states of faster or slower. The following plots give a better quantative feel for the differences in tempo between two performances by coloring the plot according to percent change using the following mapping:

 indjic2001 chiu1999 luisada1991 smith1975 rubinstein1961 rubinstein1938 indjic2001 chiu1999 luisada1991 smith1975 rubinstein1961 rubinstein1938
```

```