SOURCE-CHANNEL CODING OF ANALOG DATA FOR DIGITAL MAGNETIC RECORDING

In the traditional approach of storing analog data (such as speech samples or image pixels) in digital magnetic recording, a binary representation of the analog data is stored with great reliability. Since such high reliability is unnecessary for many types of analog data, it is hypothesized that more data can be stored by increasing the density with which bits are placed on the media and using source-channel coding to both compress the data and protect it from the resulting increased bit error rate. This paper estimates the gains achievable with this approach. Central to this work is a model for the digital magnetic recording channel that contains a linear filter representing the Lorentzian step response and additive Gaussian processes representing electronic and media noises. The model differs from previous ones in that lineal bit density, as well as track density, are parameters that may be varied. Parameters for this model that typify disk storage systems are identified and the gains in data density for the proposed approach are estimated in two ways: by estimating the capacity of the channel model and by simulating a family of source-channel codes. From capacity bounds developed in a companion paper, it is estimated that data density may ultimately be increased by a factor of 20 or more. To assess the gains achievable with a practical system, a tree-structured family of source-channel codes is proposed and analyzed. It is found that these are much simpler than optimal unstructured source-channel codes, yet have comparable performance. By simulating them on a Gaussian source and the given channel model, it is estimated that gains on the order of 3 to 4 in data density are achievable with peak detection demodulation and gains of 2 to 3 are achievable with 1-D2 partial response maximum likelihood demodulation, each with moderate increases in bit density. Comparable gains are anticipated for speech and images.