Where Does the Half-Duplex Cognitive Radio Gap Come From?

A cognitive radio is a secondary transmitter that shares spectrum with a licensed (primary) user and tries to help it, or at least not hurt it, while sending its own data. Most clean capacity results assume the cognitive radio knows the primary's message in advance, which is physically impossible. A realistic cognitive radio must first overhear the primary. If it cannot transmit and receive at the same time, it has to split time between a listening phase and a talking phase. For this realistic half-duplex model, the best known schemes come within a constant number of bits of the best known upper bound on what any scheme could achieve, but that constant runs from about 1 to 8.5 bits per channel use depending on the regime (On the Gaussian Interference Channel with Half-Duplex Causal Cognition). At practical signal-to-noise ratios, 8.5 bits is a large amount of unexplained loss.

This note proposes to find out how much of that gap comes from the decoder. The schemes in question decode interfering messages one at a time and uniquely. A different rule, simultaneous non-unique decoding, lets a receiver jointly test candidate message pairs and insist only that its own message is unambiguous; it never commits to which interfering message was sent. We will build a half-duplex causal-cognition scheme whose destinations decode non-uniquely across both the listening and talking phases. We will then compute exact numerical achievable regions and outer bounds so that the gap can be split into three parts: a decoder loss, a Gaussian-translation loss, and outer-bound slack. The deliverable is an accounting, and a possibly smaller gap, rather than only a new achievable region.

Core Claim

In the half-duplex Gaussian causal cognitive interference channel, successive decoding forces each destination to fully resolve the cognitive radio's relayed copy of the primary message and the other user's shared ("public") message part. These rate constraints are artificial. Non-unique decoding over both phases removes them, and we expect this to shrink the gap measurably in moderate and mixed interference, where the reported gap is largest. We expect little or no change in very weak or very strong interference, where treating interference as noise or decoding it outright is already near-optimal.

The claim is falsifiable, and a null result is informative. If the non-unique region turns out to equal the region obtained by time-sharing over successive-decoding orders and rate splits, then the 8.5-bit figure is mostly outer-bound slack, and the field should spend its effort on converses rather than on schemes.

Positioning

Three lines of prior work motivate this project. Cardone, Tuninetti, Knopp and Salim characterize the full-duplex causal channel, where the cognitive radio can overhear while it transmits, to within 1 bit in the symmetric case and within 2 bits in several others (On the Capacity of the Two-User Gaussian Causal Cognitive Interference Channel; On the Gaussian Interference Channel with Causal Cognition, or with Unilateral Source Cooperation). Their half-duplex follow-up derives exact generalized degrees of freedom but leaves 1 to 8.5 bits of constant gap. According to our research file, the authors point to joint decoding as a way to reduce it (On the Gaussian Interference Channel with Half-Duplex Causal Cognition). Generalized degrees of freedom are a high-SNR measure of how rate grows with SNR. Because they are already exact, any gain we find must be a bounded number of bits, and we frame the question at that resolution.

The second line supplies the decoding method. Benammar, Piantanida and Shamai use non-unique interference decoding with prior (non-causal) message knowledge and prove it is strictly capacity-achieving in a mixed regime of the multicast cognitive channel (Capacity Results for the Multicast Cognitive Interference Channel). The same authors show it beats ignoring interference on compound broadcast channels (On the Compound Broadcast Channel: Multiple Description Coding and Interference Decoding). Kang, Lee, Chung and Suh combine rate splitting with non-unique decoding of relay compression indices in interference relay networks (A New Achievable Scheme for Interference Relay Channels). None of these handles a cognitive radio that learns the primary message causally during a listening phase. That combination is ours. We reuse their decoding analysis, and we use the Fourier-Motzkin bookkeeping of Kang et al. to eliminate auxiliary rates.

The third line constrains us. Zaidi, Piantanida and Shamai prove that in additive Gaussian channels, strictly causal state knowledge adds zero rate once one encoder does dirty-paper coding (pre-cancelling interference it knows in advance) (Capacity Region of Cooperative Multiple-Access Channel With States). Their setting involves channel states, not overheard messages. Still, it warns that Gaussian structure can erase gains that exist for discrete channels, so we will not assume a gain exists. Baccelli, El Gamal and Tse show that, with point-to-point codes, the optimal decoding region is a union of simpler "decode or treat as noise" regions (Interference Networks With Point-to-Point Codes). This is the strongest reason to suspect that careful successive decoding might match non-unique decoding, and it drives our main ablation.

Research Program

The scheme follows the half-duplex structure of Cardone et al. A fraction $\gamma$ of the block is spent listening and $1-\gamma$ talking. The primary splits its message into a part the cognitive radio decodes during listening and forwards, plus public and private parts in the style of Han-Kobayashi rate splitting. Han-Kobayashi splitting divides a message into a piece the unintended receiver may decode and a piece it treats as noise. The cognitive radio superimposes its own message on the forwarded primary bits and dirty-paper codes against the part it knows.

The new element is at the destinations. Each destination jointly processes its outputs from both phases. It declares its own message indices if there is some set of interfering and forwarded indices that makes all observations jointly typical, meaning statistically consistent with the transmitted codewords. It never has to decode the forwarded or public indices uniquely. We derive the resulting single-letter region for Gaussian inputs and eliminate auxiliary rates. The listening fraction and power splits are then free parameters.

The analytical step is a gap decomposition. Write $\bar R(\theta)$ for the best known outer bound on sum rate at channel parameters $\theta$, and $R_{\text{SND}}(\theta)$, $R_{\text{SD}}(\theta)$ for the numerically optimized sum rates with non-unique and with successive decoding. The subscripts stand for simultaneous non-unique decoding and successive decoding. Then

$$ \bar R - R_{\text{SD}} = \underbrace{(R_{\text{SND}} - R_{\text{SD}})}_{\text{decoder loss}} + \underbrace{(\bar R - R_{\text{SND}})}_{\text{remaining gap}}. $$

We will also report the analytical constant-gap bound for our scheme and check that it improves on the 8.5-bit worst case in the regimes where the decoder loss is positive. If time permits, we will tighten the remaining term with a numerically computed cut-set-style bound that includes the half-duplex listen/talk schedule. This separates achievability slack from converse slack.

Baselines

The naive reference point is treating interference as noise at both destinations, with the cognitive radio only dirty-paper coding against whatever it learned while listening. It shows how much any interference decoding is worth. Mirmohseni et al. indicate this is optimal only in weak interference (Capacity bounds for multiple access-cognitive interference channel).

The strongest existing method is the successive-decoding scheme of On the Gaussian Interference Channel with Half-Duplex Causal Cognition, with its reported constant gap of approximately 1 to 8.5 bits to the upper bound. The file gives no per-SNR sum-rate values for it; no reported result was found on Lacuna for pointwise half-duplex sum rates, so we will re-evaluate that scheme ourselves from its rate expressions. For fairness, it will use the same Gaussian inputs, the same channel grid, and the same optimizer and search budget over $\gamma$ and power splits as our scheme. Beating it means a strictly larger optimized sum rate on a set of channel parameters of non-negligible measure, together with a smaller worst-case analytical gap in at least one regime they identify.

The key ablation isolates the decoder. We keep our codebook structure fixed and replace non-unique decoding with the best successive decoder: the maximum over all decoding orders and rate splits, including time-sharing across them. If this ablation matches our scheme everywhere, the decoding rule contributes nothing. As an upper reference, we plot the full-duplex results within 1 to 2 bits from On the Capacity of the Two-User Gaussian Causal Cognitive Interference Channel, which bound how much half-duplex operation itself costs.

Why not just…?

Why not just use successive decoding with smarter rate splitting and time-sharing? This is the real threat. Baccelli et al. show that for point-to-point codes the optimal region decomposes into simple decode-or-ignore pieces (Interference Networks With Point-to-Point Codes), and a similar equivalence could hold here. Our setting differs in one way that may matter. The forwarded primary codeword appears in both phases, and a single successive order cannot exploit the joint observation. We do not argue this away. The decoder ablation above settles it, and we will report the regimes where the two coincide.

Why not just tighten the outer bound instead? It is possible that most of the 8.5 bits is converse slack; the open outer-bound problems noted in On the Capacity of the Two-User Gaussian Causal Cognitive Interference Channel make this plausible. The decomposition is built to detect exactly this case. A small decoder-loss term plus a large remaining gap would redirect effort to converses, and our numerical half-duplex bound gives a first attempt at that.

Why not just treat interference as noise and let dirty-paper coding do the work? Dirty-paper coding cancels only what the cognitive radio already knows. In half-duplex operation, it knows nothing about the primary's talking-phase content until listening ends. The multicast cognitive channel shows that interference decoding is strictly required in mixed regimes even with full prior knowledge (Capacity Results for the Multicast Cognitive Interference Channel). The naive baseline will measure this loss directly.

Proposed Experiments

No empirical wireless datasets exist in the reviewed corpus. As is standard in this field, the benchmarks are canonical channel models. The primary benchmark is the two-user Gaussian half-duplex causal cognitive interference channel. We sweep direct, cross and overhearing link strengths on a logarithmic grid spanning weak, moderate, strong and mixed interference, at SNRs from 0 to 60 dB. Metrics are the optimized sum rate, the achievable rate region at selected points, and the gap in bits to the outer bound. We also report the decoder-loss term from the decomposition and, at the high-SNR end, a check that our scheme recovers the exact generalized degrees of freedom of Cardone et al.

As a secondary sanity benchmark, we set the listening fraction to zero with full prior message knowledge. Our scheme should then reproduce the known capacity of the non-causal cognitive channel in its strong-interference regime (Three-User Cognitive Interference Channel: Capacity Region with Strong Interference).

Three checks decide the question. First, is $R_{\text{SND}} - R_{\text{SD}}$ strictly positive anywhere once successive decoding is given time-sharing? Second, does the positive region coincide with the moderate and mixed regimes where the reported 8.5-bit worst case occurs? Third, does the analytical worst-case gap shrink there? A positive answer to all three establishes the decoding rule as the missing ingredient. A negative answer to the first attributes the published gap to outer bounds and half-duplex scheduling.