Introduction
Dirty-paper coding (DPC) is a cornerstone of information theory, famously described by Costa as the problem of "Writing on Dirty Paper." It demonstrates that if a transmitter knows the interference (the "dirt") beforehand, it can adapt its signal to completely cancel the interference's effect, achieving the same capacity as an interference-free channel. However, traditional DPC assumes perfect Channel State Information (CSI) at both the transmitter and receiver—a condition rarely met in practical wireless systems or digital watermarking.
This research investigates the performance of DPC under more realistic constraints: a receiver that must estimate the channel from noisy training sequences and a transmitter that has no channel information at all.
Methodology
The study models the communication link as a "composite channel," evaluating performance based on the average error probability over channel estimation errors. The receiver uses Maximum-Likelihood (ML) estimation to determine the channel state. Two scenarios are considered: standard wireless broadcast channels, where interference is independent of training, and robust watermarking, where interference multiplicatively affects the training sequence.
Using Gaussian inputs, the authors derived a closed-form lower bound for the capacity of the fading Costa channel under imperfect receiver estimation and no transmitter CSI (Equation 22). This involves determining the optimal scaling parameter $\alpha$, which dictates how much of the interference to pre-cancel.
Practical Parameter Tuning
A significant challenge in implementing DPC under uncertainty is the numerical complexity of finding the optimal scaling parameter $\alpha^*$. However, the authors find that the system is remarkably robust.
As shown in the figure above, simply using the statistical mean of the parameter ($\bar{\alpha}$) results in performance nearly identical to the mathematically optimal value. This finding significantly simplifies real-world implementation by removing the need for complex real-time optimization.
Key Insights and Results
Contrary to classic DPC theory, where interference power ($Q$) has no impact on capacity, this study reveals that with imperfect CSI, the achievable rate becomes highly sensitive to $Q$. Heavy interference cannot be perfectly canceled if the channel estimate is noisy.
Furthermore, the research highlights the critical role of the training sequence length ($N$).
As illustrated, a training sequence of $N=10$ symbols requires an SNR of 18 dB to achieve a rate of 2 bits per channel use—an 11 dB penalty compared to perfect CSI. If the training is reduced to $N=1$, the penalty widens drastically to 27 dB. Interestingly, feeding the receiver's noisy estimate back to the transmitter provides only a marginal gain (approximately 1 dB), suggesting that transmitter CSI feedback is often unnecessary if the receiver's estimate is sufficiently accurate.
Conclusion
While imperfect channel estimation introduces performance penalties, DPC remains a viable and powerful technique. The discovery that transmitter ignorance and the use of mean parameters result in minimal losses paves the way for more practical DPC implementations in fading environments. Future research may explore non-Gaussian input distributions to further refine these capacity bounds.
Summary updated