The document discusses companding, a technique to reduce quantization errors in fixed point DSPs. It involves multiplying the input signal by a time-varying gain that is large when the input is small and small when the input is large, to keep the signal level constant. Simply applying this to the input/output of a DSP will distort its behavior. Instead, the states of the DSP must be transformed to also keep them constant, while preserving the original input/output relationship. The author proposes implementing this on a digital reverberator to compare the performance of companded and non-companded versions as the bit depth varies.