Adaptive Source-Channel Coding

• Adaptive Source-Channel Coding (ASCC) is a communication strategy that adjusts encoding, channel protection, and decoding based on operating conditions such as SNR, bandwidth, and source-message probability.
• It employs methods like SNR-conditioned decoders, layered JSCC for progressive quality, and hypernetwork-based parameter modulation to optimize performance under varying channel and source conditions.
• ASCC bridges classical information-theoretic approaches with modern deep learning techniques, enabling dynamic rate adaptation, feedback-driven improvements, and cross-layer optimization for multi-user and multi-task scenarios.

Adaptive Source-Channel Coding (ASCC) denotes a class of communication strategies in which the source representation, channel protection, decoding rule, or all three are adjusted to operating conditions rather than kept fixed. In the literature considered here, the adaptation variable may be channel quality such as signal-to-noise ratio (SNR), available bandwidth or semantic rate, source-message probability, task type, interaction history on a two-way channel, or multi-user interference state. Accordingly, ASCC spans classical information-theoretic joint source-channel coding (JSCC) for correlated sources over two-way channels, layered and SNR-conditioned deep JSCC for images, digital semantic systems with adaptive power/modulation/rate control, and even source-model adaptations inside otherwise separated source-channel coding (SSCC) pipelines (Weng et al., 2019, Weng et al., 2020, Kurka et al., 2020, Ding et al., 2021, Xie et al., 2024, Oh et al., 4 Jan 2025, Yuan et al., 29 Sep 2025).

1. Scope and taxonomy

Within this body of work, ASCC is not a single mechanism but a family of adaptation patterns. Some papers adapt the coding rule to the channel state explicitly, for example by conditioning a neural encoder or decoder on SNR. Others adapt to bandwidth or rate by constructing layered representations whose quality improves as more channel uses become available. Still others adapt to the source itself, either by assigning different channel protection to more probable source messages or by changing contextual source modeling so that compressed descriptions become shorter and less fragile under residual channel errors (Bocharova et al., 2014, Kurka et al., 2020, Ding et al., 2021, Wang et al., 6 May 2026).

Adaptation axis	Representative mechanism	Representative papers
Channel quality	SNR-conditioned decoder, hypernetwork modulation, channel-adaptive attention, CQI-based re-encoding	(Ding et al., 2021, Xie et al., 2024, Zhang et al., 7 Jan 2025, Li et al., 2024)
Bandwidth or rate	Layered progressive JSCC, semantic-rate masking, information-bottleneck rate control	(Kurka et al., 2020, Sun et al., 2022, Zhang et al., 16 Jun 2026)
Source realization or source model	Probability-class-dependent channel code, contextual memory-enhanced arithmetic coding	(Bocharova et al., 2014, Wang et al., 6 May 2026)
Interaction history	Two-way encoders depending on past received outputs	(Weng et al., 2020, Weng et al., 2020)
Multi-user resource state	Joint source/channel rate, power, and beamforming optimization	(Yuan et al., 29 Sep 2025, Wang et al., 19 Jan 2026)

A recurrent distinction in this literature is between explicit adaptation and implicit robustness. Earlier deep JSCC systems were already known for graceful degradation, but several later works argue that graceful degradation alone is not equivalent to ASCC: explicit conditioning on SNR, rate, or other side information changes the learned source reconstruction rule itself rather than merely relying on passive robustness (Ding et al., 2021, Xie et al., 2024). Another important distinction is between JSCC and SSCC. Some adaptive schemes remain strictly separated, yet are still relevant to ASCC because they modify source coding behavior or channel protection according to task or operating point (Wang et al., 6 May 2026, Bocharova et al., 2014).

2. Information-theoretic foundations

In the two-way discrete-memoryless channel (DM-TWC) setting, adaptation is formalized by allowing each channel input to depend on the local source block and the terminal’s past received outputs,

$$X_{1,i}=f_{1,i}(S_1^k,Y_1^{i-1}),\qquad X_{2,i}=f_{2,i}(S_2^k,Y_2^{i-1}),$$

so the encoder is causal in the interaction history rather than being a fixed function of the source alone (Weng et al., 2020, Weng et al., 2020). This notion of adaptation is structurally different from modern CSI-driven link adaptation: it is interaction-aware JSCC for simultaneous bidirectional communication.

A precursor to this adaptive viewpoint is the hybrid digital/analog JSCC construction for transmitting correlated sources over DM-TWCs. That scheme introduces auxiliaries $U_1,U_2$, forms channel inputs as $X_j=f_j(U_j,S_j)$, and derives the rate-one achievability conditions

$I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$

$I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$

thereby showing how source correlation can be preserved directly in the channel inputs rather than only after separate compression (Weng et al., 2019). The same framework subsumes uncoded transmission, rate-one SSCC, and correlation-preserving coding (Weng et al., 2019).

The later adaptive two-way results go further by embedding JSCC in a block-Markov or stationary-Markov construction. The central idea is to couple the terminals through previous source variables, previous channel interaction, and current source/channel descriptions, which enlarges the effective side information available at each decoder. The resulting achievable conditions contain mutual information terms such as

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$

with the symmetric condition for user 2, making explicit that decodability depends on current source variables, previous source variables, past interaction state, and current observations jointly (Weng et al., 2020, Weng et al., 2020). These papers also show strict inclusion relations: the adaptive scheme contains prior non-adaptive hybrid coding and induces an adaptive SSCC corollary based on Wyner-Ziv source coding and Han’s adaptive channel coding (Weng et al., 2020).

At the same time, this line of work is notable for delineating when adaptation is unnecessary. For certain symmetric DM-TWCs and source classes satisfying conditions such as $R_{\mathrm{WZ}}(D)=R_{S|S'}(D)$, complete JSCC theorems show that non-adaptive SSCC is optimal, so interactive dependence on past outputs does not enlarge the relevant distortion region (Weng et al., 2020, Weng et al., 2019). This suggests that ASCC is not synonymous with mandatory complexity: its value is contingent on channel interaction structure and on whether source correlation can be profitably exploited beyond separation.

3. SNR-adaptive deep JSCC

A major deep-learning development is the shift from “one model per operating SNR” to single-model multi-SNR systems. In SNR-adaptive deep JSCC for wireless image transmission, the encoder is a CNN

$$f_{\theta} : \mathbb{R}^{n} \rightarrow \mathbb{C}^{k}, \qquad y=f_\theta(x),$$

with average power constraint

$\frac{1}{k}E[y^{*}y] \leq P,$

and the receiver observes

$$z_i = y + n_i,\qquad n_i\sim \mathcal{CN}(0,\sigma_i^2).$$

The decoder then reconstructs

$U_1,U_2$0

where $U_1,U_2$1 is obtained from pilot-assisted SNR estimation. The practical mechanism is an SNR map, expanded to the same dimensionality as the received tensor and fused with channel features, plus a denoising module with convolutional and dilated-convolution residual branches (Ding et al., 2021). The paper frames this as decoder-side adaptation, explicitly contrasting the old decoder $U_1,U_2$2 with the new $U_1,U_2$3, and emphasizes applicability to multi-user scenarios where “each user has a different channel but the same decoder” (Ding et al., 2021).

Hypernetwork-based adaptation generalizes this idea by making encoder and decoder parameters explicit functions of the channel condition $U_1,U_2$4. In its general form,

$U_1,U_2$5

The practical Hyper-AJSCC realization avoids full weight generation and instead applies channel-dependent scaling to backbone layer parameters, for example

$U_1,U_2$6

with analogous modulation for biases and convolution kernels (Xie et al., 2024). This positions ASCC as structured parameter modulation rather than model switching, and the reported storage overhead is much smaller than attention-based adaptive baselines: 4,118 parameters and 16 KB for Hyper-AJSCC versus 67,840 parameters and 265 KB for ADJSCC (Xie et al., 2024).

Transformer-based semantic JSCC pushes adaptation further into the attention mechanism. SNR-EQ-JSCC injects a learned SNR embedding into the attention-block input,

$U_1,U_2$7

and modifies the query as

$U_1,U_2$8

with $U_1,U_2$9 and $X_j=f_j(U_j,S_j)$0 generated from SNR, plus correlation-based penalties

$X_j=f_j(U_j,S_j)$1

to stabilize training (Zhang et al., 7 Jan 2025). The same work also uses a two-level adaptation strategy over fast fading: average SNR for full-image semantic processing and instantaneous SNR for blockwise channel-side processing, with an average-SNR-only fallback that requires no retraining (Zhang et al., 7 Jan 2025).

The most explicit online deep-JSCC adaptation in this set is the coarse-to-fine framework for time-varying block fading. Its coarse encoder

$X_j=f_j(U_j,S_j)$2

adapts each latent layer to average SNR, while the fine encoder

$X_j=f_j(U_j,S_j)$3

re-encodes the remaining untransmitted blocks using instantaneous SNR as the channel changes (Li et al., 2024). This is significant because adaptation no longer occurs only through conditioning a fixed codeword; the future part of the source-channel mapping is recomputed during transmission. Limited-feedback CQI quantization and reinforcement-learning-based CQI selection are then added to reduce SNR feedback overhead (Li et al., 2024).

A related but architecturally distinct line criticizes layer-wise SNR injection and instead maps SNR into a unified semantic vector, then performs one-shot global reweighting of latent features. SA-RA-JSCC couples this with a semantic-rate-aware module and a top-$X_j=f_j(U_j,S_j)$4 code mask, so that adaptation responds jointly to channel quality $X_j=f_j(U_j,S_j)$5 and semantic-rate constraint $X_j=f_j(U_j,S_j)$6 (Zhang et al., 16 Jun 2026). This suggests a move from local channel-aware feature refinement toward globally coordinated latent-space control.

4. Bandwidth and rate adaptation

Bandwidth-agile JSCC addresses a different ASCC axis: adaptation to available channel uses rather than to channel quality alone. DeepJSCC-$X_j=f_j(U_j,S_j)$7 considers image transmission over $X_j=f_j(U_j,S_j)$8 parallel channels, with latent output partitioned as $X_j=f_j(U_j,S_j)$9 and reconstruction from any received subset $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$0. It studies successive refinement, where only nested subsets $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$1 are used, and multiple descriptions, where any nonempty subset is decodable (Kurka et al., 2020). The paper proposes three architectures—multiple-decoder, single-decoder with structured masking, and residual transmission—and shows that progressive layered transmission incurs negligible loss relative to a single transmission of equivalent total bandwidth, while preserving graceful degradation under SNR mismatch (Kurka et al., 2020). In ASCC terms, this is bandwidth adaptation through learned layered source-channel representations.

Rate adaptation also appears in a more representation-theoretic form. AIB-JSCC proposes the objective

$I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$2

where $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$3 captures relevance after channel corruption and $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$4 penalizes information retained in the transmitted representation (Sun et al., 2022). The key adaptive element is not online rate switching but dynamic training-time adjustment of $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$5 using a PID-style rule driven by validation MSE and bounded by

$I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$6

The reported effect is lower effective transmission rate together with improved reconstruction quality relative to fixed-$I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$7 baselines (Sun et al., 2022). This suggests that ASCC can be implemented as adaptive regularization of learned representations, even when the physical codeword length is fixed.

Short-blocklength PAC-based JSCC provides a coding-theoretic counterpart to this rate-aware view. In that framework, source PAC compression and channel PAC coding are concatenated, and practical joint decoding uses a combined metric

$I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$8

that adds channel path likelihood and an approximate source prior (Zheng et al., 2023). A central observation is that the optimal compression rate varies with SNR; after optimizing compression rate for each SNR, the reported gap to the JSCC finite-length bound is about $I(S_1;U_1\mid S_2,U_2)< I(U_1;Y_2\mid S_2,U_2),$9 dB at $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$0 (Zheng et al., 2023). This does not constitute online ASCC by itself, but it establishes source-rate adaptation as a necessary ingredient for finite-blocklength optimality.

Classical source-aware rate adaptation predates these neural systems. Multi-class source-channel coding partitions source messages into probability-based classes $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$1, assigns a distinct channel code to each class, and performs parallel per-class decoding followed by MAP selection (Bocharova et al., 2014). The achievable exponent

$I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$2

improves on separate source-channel coding, and the performance approaches JSCC as the number of classes increases (Bocharova et al., 2014). This is ASCC in a source-driven rather than CSI-driven sense: the protection level adapts to source-message probability.

5. Digital, semantic, multi-user, and cross-task ASCC

A substantial recent theme is digital compatibility. Blind training for channel-adaptive digital semantic communication introduces $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$3 trainable parallel BSCs between the semantic encoder and decoder, with bit-flip probabilities

$I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$4

and a relaxed binary interface used for end-to-end training (Oh et al., 4 Jan 2025). At inference time, the system selects a pretrained model and adapts symbol powers and modulation orders so that the realized BER profile matches the learned reliability profile $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$5, through adaptive power control (APC) or joint adaptive modulation and power control (AMPC) (Oh et al., 4 Jan 2025). In ASCC terms, the source representation learns its desired unequal error protection structure offline, and the physical layer realizes it online.

Multi-user adaptive source-channel coding extends this digital perspective to heterogeneous tasks. In MU-ASCC for downlink MU-MISO semantic and data communication, a base station jointly optimizes DNN-based source coding rates $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$6, finite-blocklength channel coding rates $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$7, transmit powers $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$8, and beamformers $I(S_2;U_2\mid S_1,U_1)< I(U_2;Y_1\mid S_1,U_1),$9 to minimize a weighted sum of end-to-end distortions across data users and semantic users (Yuan et al., 29 Sep 2025). The key modeling device is a logistic decomposition of end-to-end distortion into source and channel terms, for example

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$0

with an analogous semantic-task expression (Yuan et al., 29 Sep 2025). The resulting weighted-sum distortion problem is then attacked by alternating optimization, subgradient-based adaptive rate updates, and uplink-downlink duality for power/beamforming (Yuan et al., 29 Sep 2025). Relative to earlier “multi-user” deep JSCC claims that mainly meant common architectures over separate AWGN links, this is a more explicit multi-user ASCC formulation with shared wireless resource optimization (Ding et al., 2021, Yuan et al., 29 Sep 2025).

Integrated sensing and semantic communication adds another coupling dimension. In that setting, the semantic branch has source rate

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$1

channel coding rate

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$2

and average channel uses

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$3

while the transmitter simultaneously beams a sensing waveform and a semantic waveform under a total power constraint (Wang et al., 19 Jan 2026). The end-to-end semantic distortion is approximated as

$$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$4

and a hybrid Cramér-Rao bound is imposed for localization quality under imperfect time synchronization (Wang et al., 19 Jan 2026). The problem is then to co-optimize coding rate and beamforming under channel-use, power, and sensing constraints. This broadens ASCC from communication-only design to cross-task adaptation between semantic fidelity and sensing accuracy (Wang et al., 19 Jan 2026).

These digital and multi-task systems suggest that contemporary ASCC increasingly operates as a cross-layer optimization problem. Adaptation may involve not only the encoder and decoder but also modulation, power, beamforming, sensing covariance, or user scheduling. A plausible implication is that the classical boundary between source-channel coding and wireless resource allocation is becoming less tenable in task-oriented systems.

6. Misconceptions, limitations, and open directions

One common misconception is that ASCC refers only to transmitter-side link adaptation based on CSI. The cited literature is broader. ASCC includes receiver-side conditioning on channel state without transmitter-side rate change (Ding et al., 2021), source-driven unequal protection based on message probability (Bocharova et al., 2014), layered bandwidth agility (Kurka et al., 2020), and interaction-aware two-way coding based on past received outputs (Weng et al., 2020, Weng et al., 2020). It also includes schemes that remain separated at the protocol level but still adapt the source model in ways that materially change end-to-end robustness (Wang et al., 6 May 2026).

Another misconception is that every adaptive method in this area is JSCC. “Contextual Memory-Enhanced Source Coding for Low-SNR Communications” is explicitly an SSCC system: arithmetic source coding driven by a memory-augmented model, followed by LDPC and BPSK over AWGN or Rayleigh fading (Wang et al., 6 May 2026). Its Parameterized Contextual Memory and Mixture-of-Memory-Experts Router adapt token prediction to local context, which shortens the average codelength and reduces sensitivity to residual channel errors, but the training objective is still source-model cross-entropy rather than end-to-end noisy-channel distortion (Wang et al., 6 May 2026). This makes it ASCC-adjacent rather than full JSCC.

A third recurring issue is the strength of the channel-state assumptions. Some systems rely on decoder-side pilot-assisted SNR estimation only (Ding et al., 2021). Others assume that the channel condition is perfectly estimated at the receiver and fed back to the transmitter, so both sides know $$I(\tilde S_1;\tilde U_1) < I(\tilde U_1; S_2,U_2,\tilde S_2,\tilde U_2,\tilde W_2,X_2,Y_2),$$5 (Xie et al., 2024). Transformer-based semantic systems may require average SNR, fading coefficients, or both (Zhang et al., 7 Jan 2025), and coarse-to-fine block-fading schemes depend on timely per-block feedback or CQI decisions (Li et al., 2024). Several papers explicitly neglect or only lightly analyze feedback overhead, estimator mismatch, or latency effects (Zhang et al., 7 Jan 2025, Li et al., 2024).

Dataset and channel coverage remain narrow in much of the literature. SNR-adaptive deep JSCC and Hyper-AJSCC are evaluated on CIFAR-10 over AWGN (Ding et al., 2021, Xie et al., 2024). SNR-EQ-JSCC and SA-RA-JSCC focus on image transmission with DIV2K and related image benchmarks (Zhang et al., 7 Jan 2025, Zhang et al., 16 Jun 2026). Some works move to OFDM CDL-A block fading or Rayleigh fading, but many adaptive claims are still validated mainly on AWGN or simplified fading settings (Li et al., 2024, Wang et al., 6 May 2026). This suggests that robustness to richer nonstationary channels, MIMO effects, burst errors, or cross-dataset shifts remains incompletely characterized.

Open directions are repeatedly identified across the papers. These include extending adaptive deep JSCC to fading or MIMO channels, integrating a proper learned CSI/SNR estimator instead of injecting synthetic SNR noise, moving beyond images toward text, video, or multimodal data, and expanding from scenario-level multi-user adaptation to richer shared-resource formulations (Ding et al., 2021, Xie et al., 2024, Li et al., 2024). On the digital side, variable-rate learned compression and practical code-specific finite-blocklength models are identified as next steps beyond lookup-table source models and idealized BER approximations (Yuan et al., 29 Sep 2025, Wang et al., 19 Jan 2026). On the source side, MASC suggests that context-adaptive entropy models could be combined with explicit SNR-aware channel adaptation, HARQ, unequal error protection, or adaptive source truncation, though that combination is not yet realized in the cited work (Wang et al., 6 May 2026).

Taken together, these works depict ASCC as an expanding design space rather than a settled methodology. The central unifying principle is not a particular architecture but the deliberate matching of source representation, channel protection, and operating conditions. Whether implemented by block-Markov interaction, layered latent transmission, SNR-conditioned hypernetworks, attention reweighting, trainable BER profiles, weighted-sum distortion optimization, or contextual source modeling, ASCC is fundamentally about replacing fixed operating points with condition-aware source-channel decisions (Weng et al., 2019, Kurka et al., 2020, Xie et al., 2024, Oh et al., 4 Jan 2025, Yuan et al., 29 Sep 2025).