Linear Universal Decoding for Compound Channels: a Local to Global Geometric Approach

Over discrete memoryless channels (DMC), linear decoders (maximizing additive metrics) afford several nice properties. In particular, if suitable encoders are employed, the use of decoding algorithm with manageable complexities is permitted. Maximum likelihood is an example of linear decoder. For a compound DMC, decoders that perform well without the channel’s knowledge are required in order to achieve capacity. Several such decoders have been studied in the literature. However, there is no such known decoder which is linear. Hence, the problem of finding linear decoders achieving capacity for compound DMC is addressed, and it is shown that under minor concessions, such decoders exist and can be constructed. This paper also develops a “local geometric analysis”, which allows in particular, to solve the above problem. By considering very noisy channels, the original problem is reduced, in the limit, to an inner product space problem, for which insightful solutions can be found. The local setting can then provide counterexamples to disproof claims, but also, it is shown how in this problem, results proven locally can be “lifted” to results proven globally.

💡 Research Summary

The paper tackles the long‑standing open problem of whether a linear decoder—one that selects the transmitted codeword by maximizing an additive metric—can achieve the capacity of a compound discrete memoryless channel (DMC) without prior knowledge of the actual channel law. A compound DMC is defined by a family of possible transition matrices ({W_\theta}_{\theta\in\Theta}); the encoder and decoder must be fixed for all (\theta). While universal decoders based on maximum‑likelihood, mismatched metrics, or list decoding have been studied, none of them are linear in the sense of additive per‑symbol scores, and linear decoders are attractive because they admit low‑complexity implementations (O(n) operations for blocklength n).

The authors introduce a “local geometric approach” that starts from the extreme case of a very noisy (VN) channel. In a VN channel the output distribution is almost uniform, and the true transition probabilities differ from uniform by a small perturbation (\epsilon). Expanding the log‑likelihood to first order yields a linear function of the input–output pair, i.e., (\log W(y|x) \approx \log \frac{1}{|\mathcal{Y}|} + \epsilon \phi(x,y)). Consequently, the log‑likelihood ratio between any two VN channels reduces to an inner product in a Euclidean space whose coordinates correspond to the (\phi) values for each ((x,y)) pair.

By mapping every possible channel (W_\theta) to a vector (\mathbf{w}\theta) in (\mathbb{R}^{|\mathcal{X}||\mathcal{Y}|}), the compound decoding problem becomes: find a linear functional (q) (i.e., a vector (\mathbf{q})) that maximizes the worst‑case inner product (\langle \mathbf{q}, \mathbf{w}\theta\rangle) over (\theta). This is a classic min‑max (saddle‑point) problem. The optimal linear metric is shown to be the support function of the convex hull of the channel vectors: \