In linear-real coding, the transmitted signals are (possibly redundant) linear combinations of the data signals. The linear combination of data signals can have a block pattern, resulting in linear-real block coders, or a stationary pattern, resulting in linear-real stationary (shift-register) coders. Stationary coding is shown to be a limiting case of block coding. Both methods appear to be practical for the control of burst and impulse noise. However, stationary coding appears to have some advantages and is the only one we study here. We propose shift register implementations which promise the required precision and dispersion at less cost than tuned RLC circuits. Error properties of both block and stationary coders are similar, but it is easier to learn concepts by analyzing the block coders. When the receiver is able, by using some of the techniques we discuss, to estimate the noise covariance matrix for each codeblock, the resulting noise power is less than that for receivers not using the statistics for each codeblock. Nonlinear memoryless filters, such as clippers, are especially effective when used with linear-real coders. We propose a memoryless filter which attenuates the input signal more severely when a second input to the filter indicates the channel is having a noise burst. If the memoryless filter is designed for the worst case noise, then performance will not degrade with decreased noise when the nonlinearity is odd and monotonic.