Multi-User Adaptive Source-Channel Coding

• Multi-User Adaptive Source-Channel Coding (MU-ASCC) is a design principle that jointly optimizes source representation and channel protection by adapting to user conditions, channel states, and task objectives.
• It dynamically adjusts coding rates, power allocation, and beamforming to meet diverse performance metrics such as MSE, PSNR, and AoI across heterogeneous users.
• Research in MU-ASCC spans deep joint source-channel coding, layered JSCC, and semantic communication, highlighting trade-offs between adaptivity and system complexity.

to=arxiv_search 大发官网.search 天天彩票app{"query":"\"multi-user adaptive source-channel coding\" OR MU-ASCC", "max_results": 10} Multi-User Adaptive Source-Channel Coding (MU-ASCC) denotes a class of multiuser communication strategies in which source representation and channel protection are jointly designed and adapted to heterogeneous users, channel states, source statistics, side information, decoding deadlines, or task objectives. In the recent literature the term is used explicitly for a multi-user semantic and data communication framework that jointly optimizes source-channel coding rates, power allocation, and beamforming in a digital MU-MISO downlink, but the underlying idea spans earlier work on joint source-channel coding for broadcast, multiple-access, relay, erasure, and multicast settings, as well as deep joint source-channel coding with decoder-side SNR adaptation, unequal error protection, and feedback-driven scheduling (Yuan et al., 29 Sep 2025, 0807.2666).

1. Scope and defining characteristics

MU-ASCC is not a single canonical architecture. The literature instead presents a family of constructions whose common feature is adaptation across users rather than a fixed point-to-point code design. In different settings, the adaptive variable is the receiver SNR estimated from pilots, the source-message class induced by probability thresholds, the layer index in a degraded broadcast code, the decoding time of each user under variable-length coding, the joint fading state across users, the topology of a cooperative network graph, or the semantic importance of transmitted content (Ding et al., 2021, Rezazadeh et al., 2019, 0901.2396, Musy, 2011, Kazemi et al., 2017, Bao et al., 2010, Men et al., 2 Mar 2026).

The practical objectives also vary. Some schemes target distortion metrics such as MSE, PSNR, MS-SSIM, or classification accuracy; others optimize admissibility, random-coding exponents, source-channel rate, AoI, or NMSE. This suggests that MU-ASCC is best understood as a design principle: the source and channel interface is treated as user-dependent and state-dependent, rather than fixed once and for all (Yuan et al., 29 Sep 2025, Feng et al., 2019, 2505.19465).

2. Fundamental limits: joint design, separation, and modular interfaces

A central theme of the multiuser JSCC literature is that classical Shannon source-channel separation does not generally remain optimal once multiple correlated sources, multiuser channels, and receiver side information are introduced. For several multiuser channel models, necessary and sufficient conditions for optimal separation are obtained only under specific source and side-information structures; otherwise, joint or only operationally separated constructions are required (0807.2666). In particular, for a MAC with receiver side information, informational separation is optimal when $S_1 - W_1 - S_2$ forms a Markov chain, leading to conditions of the form

$$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$

The same body of work emphasizes that even when source-channel separation is optimal, the optimal source and channel codes are not necessarily the optimal codes for the underlying source coding or channel coding problems taken in isolation (0807.2666).

For relay networks, the separation issue becomes more nuanced. For MARCs and MABRCs with correlated sources and side information, sufficient conditions based on operational separation are derived, as well as necessary conditions on achievable source-channel rates; because operational separation is generally not optimal, joint source-channel coding schemes based on a combination of correlation preserving mapping and Slepian-Wolf coding are also derived (Murin et al., 2011). Yet for fading Gaussian MARCs and MABRCs, conditions are identified under which informational separation is optimal; this is described as the first proof of separation optimality for these relay settings (Murin et al., 2011, Murin et al., 2012).

A complementary modular view is provided by hybrid coding. There, the same codeword is used for both source coding and channel coding, yielding a source encoder and channel decoder that remain modular in analysis but avoid a strict digital interface. For communicating correlated sources over a DM-MAC, achievable distortions satisfy

$$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$

with symbol-by-symbol channel mappings $x_1(q,u_1,s_1)$ and $x_2(q,u_2,s_2)$ and reconstructions $\hat s_j(q,u_1,u_2,y)$ (Minero et al., 2013). In this sense, MU-ASCC includes both fully joint designs and modular interfaces that nevertheless preserve joint source-channel dependence.

3. Adaptive mechanisms and code construction primitives

The adaptive mechanisms appearing in MU-ASCC can be grouped by what is being conditioned upon. In message-dependent MAC coding, each user partitions source messages into two classes via a threshold $\gamma_\nu\in[0,1]$, assigning one of two input distributions to each class; the achievable random-coding exponent is then optimized over $(\gamma_1,\gamma_2)$, and the resulting exponent is larger than that achieved using only one input distribution for each user (Rezazadeh et al., 2019). In multi-class source-channel coding, source messages are assigned to classes $\mathcal{A}_i^k$ according to probability thresholds and each class uses a channel code of rate $R_i=\frac{1}{n}\log|\mathcal{A}_i^k|$; as the number of classes increases, performance improves on separate source-channel coding and approaches joint source-channel coding (Bocharova et al., 2014).

In receiver-centric variable-length MAC coding, adaptation occurs in time rather than only in rate. The decoding times $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$0 and $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$1 may differ, and the rate of user $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$2 is defined as $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$3. The achievable region depends on ratios $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$4 and $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$5 that quantify the overlap between average decoding times, and the framework permits users to be decoded at different instants (Musy, 2011). A related but broader form of adaptation appears in a broadcast approach for slowly fading MACs with receiver-only CSI, where each transmitter splits its information into $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$6 streams adapted to the combined states resulting from all users’ channels, rather than only to its own direct channel (Kazemi et al., 2017).

At the network level, GANCC adapts the code graph to the dynamic network graph on-the-fly and integrates channel coding with network coding through circulant LDPC codes. The resulting framework is distributed, real-time, and local-information-based, and it replaces fixed graph design with topology-matched sparse-graph construction in each cooperation round (Bao et al., 2010).

Adaptation axis	Representative mechanism	Source
Message statistics	Threshold-based classes and class-dependent input distributions	(Rezazadeh et al., 2019)
Message probability	Multi-class partition with class-dependent channel codes	(Bocharova et al., 2014)
Decoding time	Receiver-centric variable-length rates $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$7	(Musy, 2011)
Joint channel state	$$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$8 information streams per user	(Kazemi et al., 2017)
Network topology	Code graph matched to dynamic network graph	(Bao et al., 2010)

These constructions show that “adaptive” in MU-ASCC is broader than AMC-style rate switching. It can mean class-adaptive codebooks, state-adaptive superposition, topology-adaptive sparse graphs, or receiver-specific stopping times.

4. Deep MU-ASCC architectures and learned end-to-end optimization

A prominent learned instantiation is SNR-adaptive deep JSCC for wireless image transmission. The encoder maps an image $$\begin{aligned} H(S_1|W_1) &< b I(X_1;Y_1|X_2,Q),\ H(S_2|W_1) &< b I(X_2;Y_1|X_1,Q),\ H(S_1|W_1)+H(S_2|W_1) &< b I(X_1,X_2;Y_1|Q). \end{aligned}$$9 to a complex vector $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$0 under the average power constraint

$$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$1

and each user $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$2 receives $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$3 with $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$4. The decoder takes both $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$5 and an estimated SNR map derived from pilot signals, using separate convolutional layers whose outputs are added element-wise. The encoder is a CNN with 5 convolutional layers and PReLU activations; the decoder uses deconvolutional layers, PReLU, and Sigmoid activations; and an early Denoising Module contains two parallel residual branches, one with standard convolution and one with dilated convolution, whose outputs are averaged. Training minimizes

$$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$6

and performance is measured by PSNR (Ding et al., 2021). On CIFAR-10 with bandwidth ratios $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$7 and $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$8, training SNRs $$\begin{aligned} I(U_1;S_1|U_2,Q) &< I(U_1;Y|U_2,Q),\ I(U_2;S_2|U_1,Q) &< I(U_2;Y|U_1,Q),\ I(U_1,U_2;S_1,S_2|Q) &< I(U_1,U_2;Y|Q), \end{aligned}$$9 dB, and testing SNRs $x_1(q,u_1,s_1)$0 dB, the scheme uses a single model for all SNRs, consistently outperforms the baseline especially when the test SNR does not match the training SNR, and remains robust to noisy SNR estimation: with $x_1(q,u_1,s_1)$1 performance is almost unchanged, while with $x_1(q,u_1,s_1)$2 the PSNR drops only slightly under low-SNR conditions (Ding et al., 2021).

The explicit MU-ASCC framework for multi-user semantic and data communication extends adaptation to heterogeneous tasks. In a MU-SemDaCom MU-MISO downlink, a BS serves data users, who aim at accurate source data reconstruction, and semantic users, who focus on semantic task execution. Because the E2E distortion of DNN-based codecs under finite blocklength cannot be expressed in closed form, the framework approximates it by logistic regression, decomposing E2E distortion into source and channel terms. For a data user and a semantic user, respectively,

$x_1(q,u_1,s_1)$3

$x_1(q,u_1,s_1)$4

The BER is modeled via finite blocklength theory as

$x_1(q,u_1,s_1)$5

and the design solves a weighted-sum E2E distortion minimization over source coding rates, channel coding rates, power allocation, and beamforming vectors, using an AO framework in which rate adaptation is handled by subgradient descent and power/beamforming via UDD and SCA (Yuan et al., 29 Sep 2025). At $x_1(q,u_1,s_1)$6, the reported gains are $x_1(q,u_1,s_1)$7 classification accuracy and $x_1(q,u_1,s_1)$8 MS-SSIM versus ZF-WF-BPG, and $x_1(q,u_1,s_1)$9 classification accuracy and $x_2(q,u_2,s_2)$0 MS-SSIM versus ZF-WF-DJSCC; at $x_2(q,u_2,s_2)$1, performance improves further for both user types (Yuan et al., 29 Sep 2025).

Deep MU-ASCC also appears in multi-user CSI feedback. RCA-MUNet uses DJSCC so that each UE’s encoder outputs a real latent vector mapped to a complex vector for uplink transmission over AWGN, while the BS decoder fuses multi-user features through residual cross-attention blocks. The residual cross-attention mechanism uses the current UE embedding and the average embedding of the other UEs, and a two-stage training scheme first pre-trains across a wide range of uplink SNRs and then fine-tunes at a specific deployment SNR for a small number of epochs, e.g. 200–400. Training minimizes the sum of per-user CSI reconstruction MSE, and evaluation uses NMSE. The framework is reported to eliminate the cliff-effect of bit-level SSCC, provide graceful NMSE degradation with SNR, achieve 1–2 dB NMSE improvements over single-user and prior multi-user baselines, and keep the decoder parameter count flat as the number of UEs increases because encoder and decoder parameters are shared among UEs (2505.19465).

5. Representative operating regimes and applications

Application-layer layered JSCC for BEBC multicasting is an early and technically important multi-user adaptive regime. Independent parallel Gaussian sources are transmitted to multiple user classes with different erasure probabilities and capacities $x_2(q,u_2,s_2)$2. The broadcast code has $x_2(q,u_2,s_2)$3 layers, and the degraded capacity region is

$x_2(q,u_2,s_2)$4

Because parallel Gaussian sources are successively refinable under quadratic distortion, embedded scalar quantizers and layered rateless encoders can be optimized through convex programs such as min bandwidth, min weighted total distortion, and min-max distortion penalty. The reverse-waterfilling rate-distortion function is

$x_2(q,u_2,s_2)$5

and the minimum-bandwidth solution is

$x_2(q,u_2,s_2)$6

The resulting layered source-channel code concatenates embedded scalar quantization and systematic raptor coding, yielding user-adaptive reconstruction quality while avoiding catastrophic error propagation through direct bit-plane transmission and BP decoding (0901.2396).

A different regime is freshness-oriented adaptation. In a two-user broadcast symbol erasure channel with perfect instantaneous feedback, each update contains $x_2(q,u_2,s_2)$7 symbols and the source can send one encoded symbol per slot. The adaptive scheme operates in two phases: when both users are synchronized it sends rateless coded symbols of the current update; once the strong user has moved on and the weak user is still decoding the previous update, it sends either a new raw symbol or a mixture of a new symbol and symbols from the pending update, depending on feedback. The strong user’s AoI remains

$x_2(q,u_2,s_2)$8

identical to the greedy benchmark, while the weak user’s expected AoI scales as $x_2(q,u_2,s_2)$9, compared with $\hat s_j(q,u_1,u_2,y)$0 under greedy switching (Feng et al., 2019). Although the objective here is AoI rather than distortion, the coding rule is still multi-user adaptive in the sense that each transmission is selected according to the users’ decoding states.

Intent-guided video communication introduces yet another dimension: semantic importance. Video TokenCom uses a pretrained video tokenizer to obtain discrete video tokens, then leverages a user text prompt, a CLIP heatmap, and optical-flow propagation to identify token regions corresponding to user-intended semantics across time. Intended tokens are encoded with full codebook precision, while non-intended tokens use reduced codebook precision differential encoding. The resulting bits per pixel are

$\hat s_j(q,u_1,u_2,y)$1

and class-wise UEP is selected through a MILP that balances distortion and delay under bandwidth and BLER constraints (Men et al., 2 Mar 2026). On UVG at BPP $\hat s_j(q,u_1,u_2,y)$2, the reported average PSNR is 26.36 for TokenCom versus 24.47 for VC-DM and 23.28 for H.265, with LPIPS 0.095 versus 0.104 and 0.184, and FVD 1289 versus 2087 and 4010; at SNR $\hat s_j(q,u_1,u_2,y)$3 dB, TokenCom reduces FVD by approximately 1500 versus H.265, and side-information overhead is approximately $\hat s_j(q,u_1,u_2,y)$4 of payload (Men et al., 2 Mar 2026).

6. Trade-offs, misconceptions, and research directions

A common misconception is that MU-ASCC is simply “source coding plus adaptive channel coding.” The surveyed literature shows a much broader design space. Adaptation may enter through source correlations and side information, joint state-dependent stream splitting, pilot-aided decoder conditioning, class-dependent input distributions, irregular encoding for relay and destination, or semantic masks and token classes (Bahrami et al., 2015, Kazemi et al., 2017, Ding et al., 2021, Rezazadeh et al., 2019, Murin et al., 2012, Men et al., 2 Mar 2026). Equally important, the separation question is genuinely regime-dependent: separation is not optimal in general for sending correlated sources over multiuser channels, yet it can be optimal under specific Markov, side-information, or fading conditions (Bahrami et al., 2015, 0807.2666, Murin et al., 2011).

Another recurring trade-off is between adaptivity and structural complexity. As the number of classes or layers grows, performance can approach the joint source-channel limit, but the number of codebooks, thresholds, or power-allocation variables also increases; in the broadcast strategy adapted to the multiuser channel, each user generates $\hat s_j(q,u_1,u_2,y)$5 streams, while in multi-class coding the exponent improves as the number of classes increases (Bocharova et al., 2014, Kazemi et al., 2017). Deep architectures address some of this complexity by amortizing adaptation into training, yet they introduce model-selection and deployment issues such as SNR-specific fine-tuning, discrete source-rate quantization to the nearest available DNN model, and logistic-model fitting of E2E distortion (2505.19465, Yuan et al., 29 Sep 2025).

The literature also highlights a tension between robustness and digital compatibility. Deep JSCC-based CSI feedback and SNR-adaptive image transmission explicitly target the cliff-effect or SNR mismatch, while the MU-SemDaCom framework is explicitly digital-compliant and the TokenCom framework combines pretrained tokenization with semantic-aware UEP and class-wise MCS selection (2505.19465, Ding et al., 2021, Yuan et al., 29 Sep 2025, Men et al., 2 Mar 2026). A plausible implication is that future MU-ASCC systems will continue to combine modular digital interfaces with increasingly task-specific and user-specific adaptation mechanisms, rather than converging on a single universal architecture.