State-Dependent DM Relay Channel

• State-dependent discrete memoryless relay channels are three-terminal networks where an i.i.d. state governs channel behavior and impacts capacity and coding schemes.
• These channels integrate techniques such as decode-and-forward, Gel’fand-Pinsker binning, and state description to address asymmetric channel-state information.
• Capacity analyses reveal that state asymmetry incurs specific mutual information penalties and necessitates tailored cooperative strategies for effective communication.

Searching arXiv for recent and foundational papers on the state-dependent discrete memoryless relay channel. The state-dependent discrete memoryless relay channel (SD-DMRC) is a three-terminal relay model in which the source-to-relay–destination communication law is controlled by an i.i.d. random state. In its basic form, the channel is specified by finite alphabets, an i.i.d. state sequence $S^n$, and a memoryless transition $W_{Y_2,Y_3|X_1,X_2,S}$. The core information-theoretic problem is capacity under asymmetric channel-state information (CSI): the state may be known noncausally only at the relay, noncausally only at the source, or strictly causally at the relay, and each CSI pattern induces a different balance among decode-and-forward, Gel'fand-Pinsker binning, compress-and-forward, state description, and cooperative compression (0911.4704).

1. Formal model and achievability notion

A standard SD-DMRC uses finite alphabets $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$, with states $S^n$ drawn i.i.d. according to $P_S$, and per-letter transition law

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$

The source transmits $X_1^n$, the relay transmits $X_2^n$, the relay observes $Y_2^n$, and the destination observes $Y_3^n$. Achievable rate is defined in the usual Shannon sense: a rate $W_{Y_2,Y_3|X_1,X_2,S}$0 is achievable if there exist $W_{Y_2,Y_3|X_1,X_2,S}$1 codes with average error probability tending to zero as $W_{Y_2,Y_3|X_1,X_2,S}$2. Depending on the model, encoder mappings may depend on the message, on the entire state sequence $W_{Y_2,Y_3|X_1,X_2,S}$3, on past relay outputs $W_{Y_2,Y_3|X_1,X_2,S}$4, or on strictly causal state $W_{Y_2,Y_3|X_1,X_2,S}$5. Cost constraints can also be imposed, for example $W_{Y_2,Y_3|X_1,X_2,S}$6, although many statements are presented without explicit costs (0911.4704).

The main asymmetry is the CSI pattern. In the noncausal relay-informed model, the relay knows $W_{Y_2,Y_3|X_1,X_2,S}$7 and transmits $W_{Y_2,Y_3|X_1,X_2,S}$8, while the source has no state knowledge and sends $W_{Y_2,Y_3|X_1,X_2,S}$9. In the source-informed model, the source knows $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$0 and the relay is uninformed. In strictly causal relay-state models, the relay knows only $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$1 when choosing its $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$2-th channel input, and some formulations add source–relay conferencing links of capacities $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$3 and $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$4 [(0911.4704); (Zaidi et al., 2010); (Li et al., 2011)].

2. Noncausal state available only at the relay

A central SD-DMRC instance is the channel with noncausal CSI only at the relay. An early formulation appears in "Cooperative Relaying with State Available at the Relay" (0711.4864), and a detailed treatment appears in "Cooperative Relaying with State Available Non-Causally at the Relay" (0911.4704). In this model, the lower bound is built from two auxiliaries $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$5 and $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$6 under the joint distribution

$$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$7

The achievable rate is

$$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$8

where the maximization is over $$\mathcal X_1,\mathcal X_2,\mathcal S,\mathcal Y_2,\mathcal Y_3$$9. The cardinality bounds are

$S^n$0

The coding architecture combines decode-and-forward relaying, codeword splitting, and Gel'fand-Pinsker binning at the relay (0911.4704).

Operationally, the source uses block-Markov transmission: in block $S^n$1, if the previous message index is $S^n$2 and the new message is $S^n$3, it sends $S^n$4. The relay, knowing $S^n$5 and $S^n$6, searches for a $S^n$7-sequence jointly typical with $S^n$8 and the current state block, and then generates $S^n$9 i.i.d. according to $P_S$0. The relay decodes the new source message by joint typicality conditioned on $P_S$1, while the destination uses backward decoding. The resulting two rate constraints are

$P_S$2

The proof uses standard union bounds, joint typicality, and a Gel'fand-Pinsker bin-covering argument for the existence of a suitable $P_S$3-bin index (0911.4704).

The codeword-splitting interpretation is especially important. The relay allocates one layer to cooperation with the source and another layer to state pre-cancellation. The earlier treatment states this explicitly: $P_S$4 is a “cooperation layer” independent of the state, while $P_S$5 is a “GP-binning layer” correlated with the state. This decomposition isolates the two roles of the informed relay: coherent relaying and state mitigation (0711.4864).

3. Converse bounds and capacity for physically degraded channels

For the noncausal relay-informed SD-DMRC, converse arguments based on Fano’s inequality, the chain rule, data processing, and time-sharing yield the single-letter upper bound

$P_S$6

under distributions of the form

$P_S$7

This upper bound is stronger than the benchmark obtained by assuming that the channel state is available at the source, the relay, and the destination. The strengthening is carried by the penalty term $P_S$8, which reflects the fact that the source does not know the state and therefore cannot perform dirty-paper pre-cancellation on its own input [(0911.4704); (0711.4864)].

The physically degraded case admits an exact characterization. When the channel is physically degraded in the sense indicated in the literature as $P_S$9, the lower and upper bounds reduce to the same expression,

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$0

In this case, the capacity follows from coincidence of the inner and outer bounds. The same literature also notes that the penalty term quantifies the unavoidable loss caused by source ignorance of the state; symmetric-state cut-set bounds do not capture this loss [(0911.4704); (0711.4864)].

A recurring misconception is that a converse obtained by revealing the state to all nodes is already tight enough for asymmetric-CSI relay channels. The relay-informed SD-DMRC shows that this is generally false: the improved converse retains a state-asymmetry penalty that survives single-letterization. That penalty is not an artifact of the proof but a structural consequence of placing the state only at the relay (0911.4704).

4. Other CSI configurations of the SD-DMRC

The same state-dependent DM relay law supports several distinct CSI configurations, each with a different coding mechanism.

CSI configuration	Main mechanism	Representative papers
Noncausal state only at relay	Codeword splitting, Gel'fand-Pinsker binning, decode-and-forward	(0911.4704)
Noncausal state only at source	Future-state description or relay-input description	(Zaidi et al., 2010, Zaidi et al., 2011)
Strictly causal state at relay with conferencing	Message cooperation and state cooperation, joint decoding	(Li et al., 2011)

When the source knows the state noncausally and the relay does not, the main obstacle is the relay’s ignorance of the state. One lower bound introduces $W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$1 and $W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$2 as descriptions of future states for the relay and destination, together with auxiliaries $W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$3. The source sends compressed descriptions of the state two blocks ahead, uses Gel'fand-Pinsker binning for both message and description layers, and enables the relay to operate in a later block as if it knew the state. The destination uses a two-block window decoder. The achievable rate is

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$4

subject to explicit state-description constraints and the positivity condition

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$5

If $W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$6, the bound reduces to the classical decode-and-forward rate

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$7

If the source-to-relay link is very clean, the lower bound approaches

$W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$8

which coincides with the upper bound obtained by revealing the state to all nodes (Zaidi et al., 2010).

A second source-informed construction does not describe the state itself. Instead, the source computes the relay input that the relay would send if it knew the state, quantizes that input to $W_{Y_2,Y_3|X_1,X_2,S}(y_2,y_3|x_1,x_2,s).$9, and sends the description to the relay. The relay decodes $X_1^n$0 and transmits it in the next block. The corresponding lower bound has the form

$X_1^n$1

subject to

$X_1^n$2

The same work derives a nontrivial upper bound with auxiliaries $X_1^n$3 and $X_1^n$4,

$X_1^n$5

under the Markov structure

$X_1^n$6

This again replaces a naive cut-set perspective by a converse that explicitly tracks state ignorance at the uninformed nodes (Zaidi et al., 2011).

A different branch studies strictly causal state at the relay together with orthogonal conferencing links of capacities $X_1^n$7 and $X_1^n$8. Two achievable schemes are described: a block-Markov plus Wyner-Ziv plus backward-decoding scheme, and a noisy-network-coding-style joint-decoding scheme. The second scheme is always at least as large as the first. In the one-letter characterization of the second scheme, any rate

$X_1^n$9

is achievable. This literature also identifies exact capacities in special cases, including no cooperation, message-only cooperation, state-only cooperation, and the limit $X_2^n$0 (Li et al., 2011).

5. Generalizations: private messages, partial CSI, and sensing

The SD-DMRC has been generalized to a state-dependent relay channel with private messages (SD-RCPM), in which the source sends a private message $X_2^n$1 to the relay, the relay sends a private message $X_2^n$2 to the destination, and the source sends a relayed message $X_2^n$3 to the destination. In this model, the source and relay may each observe a noisy version of the state, either causally or noncausally. For noncausal partial CSI, an achievable rate region is derived using Gel'fand-Pinsker coding at the informed nodes and compress-and-forward at the relay. For causal partial CSI, the analogous region is obtained using Shannon’s strategy and compress-and-forward. The noncausal region contains the rate constraints

$X_2^n$4

together with a Wyner-Ziv constraint for $X_2^n$5. In the causal case, the same inequalities appear without the $X_2^n$6 penalties. This extension shows that state asymmetry can be superposed with message asymmetry and with relay compression (Akhbari et al., 2010).

A recent reinterpretation embeds the state-dependent DM relay channel into bistatic integrated sensing and communications. In that model, the destination must both decode the message and estimate a state-parameter sequence $X_2^n$7 correlated with the channel state. The fundamental object is the capacity-distortion function

$X_2^n$8

where $X_2^n$9 is the average distortion. The upper bound extends the cut-set method by introducing an auxiliary sensing variable $Y_2^n$0: $Y_2^n$1 subject to a distortion constraint and

$Y_2^n$2

A lower bound is obtained by a hybrid partial-decode-and-compress-forward scheme with auxiliaries $Y_2^n$3, and the upper and lower bounds coincide for three channel classes: orthogonal sender components with $Y_2^n$4 independent of the relay’s channel, Cover-Kim relay channels, and a two-hop orthogonal sender/receiver class with $Y_2^n$5 on the first hop (Liu et al., 12 Jul 2025).

6. Structural themes and recurring technical principles

Across the literature, asymmetric CSI appears through explicit mutual-information penalties rather than through a mere change of optimization domain. In relay-informed models, the converse contains $Y_2^n$6; in relay-side Gel'fand-Pinsker schemes, the achievable rate loses $Y_2^n$7; in source-informed models, message and state-description layers lose terms such as $Y_2^n$8 or $Y_2^n$9. This suggests that the essential obstruction is not only the presence of the state but also which terminal must act without seeing it [(0911.4704); (Zaidi et al., 2011)].

A second recurring principle is that cooperation can target three different objects. One option is message cooperation, as in decode-and-forward layers or source-to-relay conferencing. A second is state cooperation, where an informed node describes the state itself or a compressed version of it. A third is relay-input cooperation, where the informed source describes the relay action that would have been chosen under full state knowledge. The source-informed literature makes this trichotomy especially explicit through the state-description scheme and the “analog-input” description scheme [(Zaidi et al., 2010); (Zaidi et al., 2011)].

A third theme is that strictly causal state can matter in multi-terminal settings. One work states this point explicitly: strictly causal state information at the relay, when compressed and forwarded, can boost rates even in memoryless i.i.d. state channels, in contrast with the point-to-point case where strictly causal state information is useless. The gain arises from cooperation and joint decoding, not from pointwise pre-cancellation (Li et al., 2011).

The SD-DMRC therefore occupies a distinct position within relay-channel theory. It retains the standard memoryless relay architecture, but the introduction of a random state and asymmetric CSI forces coding schemes to blend relaying, binning, state description, compression, and state-aware converse techniques. The exact form of the optimal tradeoff depends sharply on who knows the state, when it is known, and whether the informed node can convert that knowledge into cooperative side information for the other terminals.