Starting out with reading the referenced papers would be a good start. But in essence, they are using a specific way of sampling from a large array of data. T

Starting out with reading the referenced papers would be a good start. But in essence, they are using a specific way of sampling from a large array of data. They take certain elements from the large array Theta_{c} (the stroke representation), but only those that match the equation 1. The selection criterion (eq 1) is of the form z(i) = z(i-1) + min(1,f(dtheta)), with the elements being picked from the large data array (j) being the rounded down version of z, which is just an implementation detail really and not that important for the algorithm as the 4.3434th (for example) element of a discrete array does not make sense. What it does is varying the distance between samples (z(i) - z(i-1)) based on the difference in orientation of the samples. If there is a sharp corner you would need more samples to properly describe this, in contrast to a straight line. This is the curvature they are talking about (the second derivative of position, or the first derivative of orientation: delta orientation). Have fun!