Introduction
Classical information theory treats decoding errors as a monolithic event. Chawla [1] obtained the tight lower bound \[ E_{ GaussPeak}({R}) > ({1}{C_2} + {1}{4C_1})^{-1}(1 - {{R}}{C_1}) \] via a two-phase scheme (orthogonal signaling + antipodal confirm/deny + sequential anytime state feedback).
Structured source coding [2] partitions the atypical set \(B_^{(n)}\) into \(k\) clusters, yielding dominant-cluster probability \(D (,\, 2^{2n}/k)\).
Structured Channel Coding fuses the two frameworks. The resulting reliability \(E_{ struc}({R},)\) is strictly larger than the unstructured bound for \( > 0\) under the refined dominant-cluster metric, but the paradigms coincide exactly as \( 0\).
Section [sec:prelim] recalls preliminaries. Section [sec:def] defines the framework. Section [sec:derive] derives \(()\) rigorously, discusses the fundamental difference from the unstructured setting, and provides a matching converse. Section [sec:phases] re-engineers the protocol. Conclusions follow in Section [sec:conc].
Preliminaries
Gaussian Feedback Reliability (Chawla 2006)
Forward and feedback channels are independent infinite-bandwidth continuous-time peak-power-constrained AWGN links (\(C_1 = P_1/N_1\), \(C_2 = P_2/N_2\)). The two-phase scheme achieves the quoted (tight) lower bound on reliability (error exponent per unit time).
Structured Source Coding
The atypical set \(B_^{(n)}\) is partitioned into \(k\) Hamming clusters. The dominant-cluster probability obeys \(D (,\, 2^{2n}/k)\), with matching worst-case lower bound \(D (1-o(1)) 2^{2n}/k\).
Definition of Structured Channel Coding
[Structured Atypical Error Set] After Phase 1 orthogonal signaling, let \({E}^{(T_1)}\) be the set of incorrectly decoded \(K\)-packet vectors. Partition \({E}^{(T_1)} = _{j=1}^k C_j\) by Hamming-distance \(k\)-means. The dominant-cluster error probability is \[ D ({m} C_{} {m} m) P_e^{(2)}. \]
[Structured Dominant-Cluster Bound] \[ D (P_e^{(2)},\, 2^{2n}/k), \] with \(n = T_1 W\) (effective discrete dimension). Strict improvement \(D < P_e^{(2)}\) holds when error mass concentrates near decision boundaries (Sanov's theorem).
Derivation Sketch of the Structured Reliability Function
Discretization and Derivation of \(()\)
The infinite-bandwidth AWGN channel with orthogonal signaling admits an exact equivalence to a discrete-time channel. Each orthogonal packet of duration \(T_1/K\) corresponds to one independent Gaussian observation (standard result: the continuous-time model reduces to \(n = T_1 W\) effective dimensions when signals are time-orthogonal and bandwidth is unlimited; see Gallager [3] or the orthogonal-signal analysis in Chawla [1], Chapter 2). The classical Phase-1 error exponent \(E_{ orth}(R)\) is already normalized per unit time, so the combinatorial bound \(2^{2n}/k\) from source coding transplants directly: the extra decay factor is \[ () = {2n - _2 k}{T_1}. \] Substituting \( = _2 k / [n(H_{ eff} + )]\) with \(H_{ eff} = {R}/C_1\) and \(n = T_1 W\) yields the normalized extra exponent \[ () = (0,\, 2 - (H_{ eff} + )) {W}{T_1} > 0 \] in the finite-blocklength regime \(n < (1/(2))_2(k/)\). Thus \[ E_{ orth}^{ struc}(R,) = E_{ orth}(R) + (). \]
Propagation to Phase-2 and Structured Reliability Lower Bound
The reduced input error probability \(D\) shortens the Phase-2 waiting time \(_{ wait}\) by \(()T_1 / \), where \( = 4C_1 C_2/(C_2 + 4C_1)\). Normalizing by expected total time gives
Fundamental Difference from the Unstructured Setting
The original unstructured bound is tight for the monolithic any-error probability. The structured bound [eq:Estruc] is larger because \(()>0\). However, the two cannot be compared directly except in the limit \( 0\): the structured reliability quantifies decay of the probability that the dominant error cluster is mishandled, not the probability of any error. This refined error criterion changes the performance metric itself. The \(()\) improvement is therefore genuine for the new metric and does not contradict the tightness of the unstructured bound.
Converse Upper Bound
We suggest that no scheme (arbitrary clustering + arbitrary Phase-2 code) can exceed [eq:Estruc].
[Sanov Lower Bound on \(D\) – Independent of Clustering] For any partition of \({E}^{(T_1)}\) into \(k\) clusters and any noise realization, Sanov's theorem on the convex rate function of the orthogonal channel implies that atypical error mass concentrates on a set of types whose size satisfies \[ D (1-o(1)) {2^{2n}}{k} \] in the worst case (uniform mass over error types near decision boundaries).
The conditional probability of each error vector is bounded by the large-deviation rate function. When mass is (nearly) uniform over the support of \({E}^{(T_1)}\) (possible under Sanov for convex rate functions), the largest cluster satisfies \(|C_{}| |{E}^{(T_1)}|/k\). Multiplying by the per-vector probability bound yields the stated lower bound on \(D\). This lower bound depends only on the channel law and \(k\), not on the specific clustering algorithm or Phase-2 code.
Because any Phase-2 anytime code must operate on an input error probability at least as large as this combinatorial lower bound on \(D\), the waiting-time reduction cannot exceed the value derived in achievability. Consequently the structured reliability satisfies the matching upper bound
Together with the lower bound [eq:Estruc], the structured reliability may be tight for the refined dominant-cluster metric. As \( 0\), both bounds recover the classical (tight) unstructured result.
[Consistency with Shannon Capacity] \(_{ 0} E_{ struc}({R},) = E_{ GaussPeak}({R})\) and effective capacity remains \(C_1\).
Re-engineered Four-Phase Protocol
Phase 1 uses structured orthogonal signaling with cluster assignment at the decoder. Phase 2 employs a structured ID code for the cluster label only (\(_2 k + 1\) bits). Phase 3 is hierarchical antipodal confirm/deny of the cluster index. Phase 4 uses a cluster-index-aware semi-orthogonal anytime code. The \(\)-overhead remains vanishing.
Conclusion
Structured Channel Coding provides a possibly tight reliability function for the refined dominant-cluster metric. The derivation of \(()\) and the Sanov-based converse resolve the technical issues of direct transplantation and tightness. Future work includes non-i.i.d.\ noise and adaptive clustering.
References
- [1]
A. Chawla, "Reliability of a Gaussian Channel in the Presence of Gaussian Feedback," Master's thesis, Massachusetts Institute of Technology, 2006.
- [2]
A. Chawla, "Structured Source Coding: Shannon Theory as the Vanishing-Clustering Limit," preprint, REAL Institute, 2026.
- [3]
R. G. Gallager, Information Theory and Reliable Communication. Wiley, 1968.