Variable-Length Feedback Coding

Updated 3 April 2026

Variable-length feedback coding is a paradigm where the blocklength adapts in real-time via receiver feedback to optimize capacity, error exponents, and delay performance.
It leverages sequential information thresholding and strategies like random coding and ROVA to enhance reliability and reduce transmission latency.
Practical designs address constraints such as power, delay, and limited feedback opportunities, with recent deep learning methods further boosting performance.

Variable-length feedback coding refers to communication schemes in which the blocklength (number of channel uses per message) is not predetermined but instead is adaptively determined by real-time feedback from the receiver, subject to an average blocklength or delay constraint. These schemes fundamentally exploit the sequential nature of information accumulation and availability of feedback to achieve faster convergence to capacity, improved error exponents, and superior delay performance compared to fixed-length or non-feedback protocols. This entry surveys the theoretical underpinnings, main coding strategies, mathematical characterizations, and prominent research directions in the design and analysis of variable-length feedback codes across channel types, practical constraints, and code constructions.

1. Mathematical Foundations and Classical Results

The canonical setting considers a discrete memoryless channel (DMC) with input $X_i$ , output $Y_i$ , and transition law $P_{Y|X}$ , where a message $W \in \{1,\dots,M\}$ is transmitted over possibly infinite channel uses, with per-use noiseless instantaneous (often 1-bit) feedback. The central performance metric is the achievable rate $R = (\log M)/\mathbb{E}[\tau]$ , where $\tau$ is the (random) stopping time—i.e., when the decoder reliably decides on the transmitted message.

The early information-theoretic analysis established that:

Capacity is unchanged by feedback for memoryless channels: $C = \max_{P_X} I(X;Y)$ .
Error exponent (reliability function) is dramatically improved with feedback, notably via Burnashev's bound:

$E(R) = C_1 \left(1 - \frac{R}{C}\right), \quad C_1 = \max_{x,x'} D(P_{Y|X=x}\|P_{Y|X=x'}),$

where $D(\cdot\|\cdot)$ denotes Kullback–Leibler divergence (Naghshvar et al., 2013, Yavas et al., 2024).

Zero dispersion regime arises: For point-to-point channels with stop-feedback, the second-order term in the normal approximation vanishes, so rate converges to capacity with $O(\log N/N)$ penalty rather than $Y_i$ 0 (Lai et al., 2024, Papoutsidakis et al., 2024).

The fundamental mechanism underlying these improvements is the sequential information thresholding: as soon as the statistical confidence (e.g., posterior probability, information density) in the true message crosses a design threshold, the communication ceases.

2. Achievability Bounds and Code Design Strategies

Random Coding and Stopping Rules

The general achievability framework is:

Codebook generation: Create $Y_i$ 1 infinite-length codewords, each drawn iid from $Y_i$ 2.
Encoding: At time $Y_i$ 3, transmit the $Y_i$ 4-th symbol of $Y_i$ 5's codeword.
Decoding: The receiver computes a reliability metric (e.g., information density, posterior, or minimum distance) and applies a stopping rule $Y_i$ 6.
Feedback: A single-bit feedback (ACK/NACK or stop/go) is sent when the stopping condition is met.

This template admits multiple concrete instances:

Information density thresholding: $Y_i$ 7 (Chen et al., 2013, Yavas et al., 2024).
Posterior matching/partition: Inputs selected to shape posteriors toward uniformity or maximal JS-divergence (Naghshvar et al., 2013, Antonini et al., 2024).

For variable-length codes with periodic or sparse decoding (decoding attempted only every $Y_i$ 8 symbols), the capacity penalty is additive and linear in $Y_i$ 9; as long as $P_{Y|X}$ 0, capacity is maintained in the limit (Chen et al., 2013, Kim et al., 2015, Yavas et al., 2021).

Practical Algorithms

Reliability Output Viterbi Algorithm (ROVA): For convolutional codes, ROVA computes the exact codeword posterior, enabling rigorous stopping at a prescribed undetected-error probability (Williamson et al., 2013, Williamson et al., 2014).
Tail-biting convolutional codes with ROVA: Combine variable-length decision feedback with tail-biting structure for rate and latency gains over both random coding and CRC-based ARQ at short blocklengths (Williamson et al., 2014).
Deep learning-aided codes: Transformer-based neural architectures that dynamically allocate transmission effort across grouped bits, learning optimal feedback-based stopping and noise-refinement phases (Lai et al., 2024, Ding et al., 8 Feb 2026).

Numerical and Theoretic Achievability

For DMCs, the random coding analysis (using e.g., the random coding union (RCU) bound) delivers non-asymptotic bounds:

$P_{Y|X}$ 1

where $P_{Y|X}$ 2 depends on type-II error probabilities for $P_{Y|X}$ 3-long codewords (Chen et al., 2013). Choosing $P_{Y|X}$ 4 suffices to approach capacity up to $P_{Y|X}$ 5.

For the Gaussian channel, minimum-distance VLSF codes achieve:

$P_{Y|X}$ 6

with no $P_{Y|X}$ 7 penalty (Papoutsidakis et al., 2024). DeepVLF codes can further approach optimality with fewer channel uses (Lai et al., 2024, Ding et al., 8 Feb 2026).

3. Variable-Length Feedback under Practical Constraints

Sparse and Bursty Feedback

Sparse stopping times: Limiting feedback to $P_{Y|X}$ 8 decoding times (e.g., packet boundaries, half-duplex constraints) yields a second-order rate penalty of $P_{Y|X}$ 9, still much better than without feedback (Yavas et al., 2021).
Bursty feedback of high rate: Allowing finite ( $W \in \{1,\dots,M\}$ 0) bursts of unlimited-rate feedback, where each burst can carry arbitrary decoder estimates, enables multi-phase protocols and can outperform one-bit stop-feedback codes—especially at very low error rates and short blocklengths (Chen et al., 2023).

Delay and Power Constraints

With strict average or maximal delay constraints (fixed total channel uses $W \in \{1,\dots,M\}$ 1), the rate penalty under feedback is $W \in \{1,\dots,M\}$ 2 as opposed to $W \in \{1,\dots,M\}$ 3 for non-feedback (Kim et al., 2015). For energy-limited AWGN channels, variable-length feedback codes attain the minimum-energy limit for lossless (and lossy) transmission with $W \in \{1,\dots,M\}$ 4 overhead (Kostina et al., 2015).

Periodic Decoding and Finite Block-Lengths

When only a finite set of possible decoding times is allowed, the optimal design involves solving an integer program or performing gap-constrained sequential differential optimization (SDO) to select the set $W \in \{1,\dots,M\}$ 5 that minimizes average blocklength subject to reliability (Yang et al., 2022, Yavas et al., 2021). Empirically, a small number of decoding times suffices to attain the full-infinite limit within fractional slack.

4. Extensions: Lossy Coding, Sensing, and Universal Schemes

Lossy Source-Channel Coding

For transmission under a distortion constraint $W \in \{1,\dots,M\}$ 6 (rate-distortion $W \in \{1,\dots,M\}$ 7), variable-length feedback enables strong separation optimality:

$W \in \{1,\dots,M\}$ 8

where $W \in \{1,\dots,M\}$ 9 is the optimal error exponent of excess distortion probability, $R = (\log M)/\mathbb{E}[\tau]$ 0 is Burnashev's constant. No joint design or "hybrid" codes are needed—the optimal exponent is achieved by separation, with a two-phase scheme (message + confirmation) (Truong et al., 2017, Kostina et al., 2015).

Sensing–Communication Tradeoffs

In integrated sensing and communication (ISAC), VLF codes create structural tradeoffs between channel coding rate $R = (\log M)/\mathbb{E}[\tau]$ 1 and sensing error exponent $R = (\log M)/\mathbb{E}[\tau]$ 2, which cannot be decomposed as in fixed-length codes. Optimal $R = (\log M)/\mathbb{E}[\tau]$ 3 pairs can be achieved by Gaussian-mixture or chi-distributed signaling; the region is strictly convex (Papoutsidakis et al., 24 Jan 2025).

Universal VLF Schemes

When the channel law is unknown, variable-length feedback codes using type-based empirical mutual information as the decoding metric achieve the same first-order $R = (\log M)/\mathbb{E}[\tau]$ 4 rate with an increased second-order penalty, characterized in terms of the alphabet size (Yavas et al., 2024).

5. Spectrum and Autocorrelation of Variable-Length Signals

Variable-length signaling modifies the power spectral density (PSD) of the transmitted signal. Under constant average symbol duration, the autocorrelation function can be expressed as:

$R = (\log M)/\mathbb{E}[\tau]$ 5

where $R = (\log M)/\mathbb{E}[\tau]$ 6 is the pulse duration density. At high SNR, the spectrum converges to that of fixed-length symbol transmission (Hsu et al., 2022).

6. Deep Learning–Aided Variable-Length Feedback Codes

Recent research leverages deep neural networks, in particular transformer architectures and group-wise processing, to model and optimize multi-round interaction protocols. Learned codes, such as DeepVLF and its variants, outperform previous state-of-the-art learned feedback codes:

Bit-grouped dynamic stopping: Encodes stopping rules at the group level, enabling fine-grained rate adaptation.
Emergent two-phase structure: Learned encoders display message and refinement phases (akin to Schalkwijk–Kailath).
Performance: Achieve 20%–55% reduction in channel uses and significant gains in block error rates over previous fixed-length codes (Lai et al., 2024, Ding et al., 8 Feb 2026).

This suggests that the information-theoretically optimal structure of variable-length feedback is accessible to data-driven, end-to-end training within power and latency constraints.

7. Open Problems and Research Directions

Tight converse bounds for stop-feedback and bursty-feedback codes on general channels: For example, in BECs, a gap remains between one-shot VLF converse and the best known VLSF achievability, especially at moderate $R = (\log M)/\mathbb{E}[\tau]$ 7 (Devassy et al., 2016).
Feedback sparsity–performance tradeoff: Quantifying how rapidly performance saturates as the number of feedback opportunities increases, including device- and protocol-level constraints (Antonini et al., 2024, Yavas et al., 2021).
Multiuser and broadcast settings: While point-to-point variable-length feedback eliminates dispersion, for multiuser (e.g., broadcast-common-message) scenarios, positive dispersion reappears due to required simultaneous threshold-crossing in all receivers (Trillingsgaard et al., 2015).
Analytic characterizations of code spectra under variable-length signaling: Including spectral efficiency and occupied bandwidth for uncoded and coded signaling (Hsu et al., 2022).

Variable-length feedback coding unifies and generalizes several operational regimes—zero-error, fixed or lossy source transmission, various feedback and complexity constraints—providing a comprehensive framework for ultra-reliable and low-latency communication, as well as a fertile ground for new algorithmic and deep-learning-based approaches.