WorksEngineering
Learning Joint Synchronization, Channel Estimation, and Decoding: Comparison with a Finite-Blocklength Achievability Bound
Abstract
Learned schemes that jointly perform tasks such as synchronization and decoding have achieved promising results for short-packet communication. However, it remains unclear how close their performance is to information-theoretic bounds. To investigate this question, we extend the turbo autoencoder (TurboAE) proposed in Dörner et al. (2023) to a single-input single-output waveform channel with an unknown delay and memoryless block fading, for which an achievability bound from finite-blocklength information theory is available. Our extension incorporates a pilot-free, model-guided neural synchronizer and enables joint synchronization, channel estimation, and decoding. At a blocklength of complex symbols and a rate of one bit per channel use, the proposed TurboAE achieves a lower bit error rate (BER) than all evaluated pilot-assisted 5G baselines, with gains of dB at a BER of over a baseline involving a low-density parity-check (LDPC) code and dB over a baseline involving a polar code and successive cancellation list decoding (list size ). In terms of block error rate (BLER), the TurboAE outperforms the LDPC-coded baseline, but lags behind the polar-coded baseline by about dB at a BLER of . Moreover, all evaluated schemes remain at least dB away from a pilot-assisted achievability bound, indicating substantial room for improved designs.
Introduction
Modern coding schemes, such as low-density parity-check (LDPC) and polar codes [1], [2], achieve excellent error-correction performance at long blocklengths. However, their performance degrades significantly in the short-blocklength regime, for two main reasons. First, short blocklengths limit the ability of these codes to fully exploit their structural properties. Second, when the blocklength is short, the finite symbol budget must be shared among synchronization, channel estimation, and data transmission. This introduces a critical trade-off between pilot symbols and payload in conventional pilot-assisted communication systems.
Finite-blocklength information-theoretic analyses show that short-packet transmission can benefit from joint synchronization, packet detection, channel estimation, and decoding. Specifically, joint synchronization and decoding without pilots achieves nearly the same coding rate as decoding with perfect synchronization, which suggests that pilot-assisted schemes may incur a rate loss [3]. Similarly, joint packet detection and decoding achieves higher coding rates than pilot-based detection followed by decoding [4], and noncoherent transmission without explicit channel estimation is more energy efficient than pilot-assisted transmission [5]. However, how to design practical joint schemes that realize these gains remains an open question.
One promising approach to optimize the transmitter and receiver for short-packet communication is end-to-end learning using autoencoders [6]. However, the original autoencoder proposed in [6] is limited to relatively short messages of up to about 10 bits. To solve this problem, more sophisticated architectures have been proposed, such as the turbo autoencoder (TurboAE) [7] and its serial extension [8], inspired by the parallel/serial encoding and iterative decoding of turbo codes [9], [10]. In the analysis of both schemes, however, perfect frame synchronization was assumed [7], [8]. This assumption was later removed in [11], in which the serial TurboAE is extended to unsynchronized multipath channels, where it jointly performs packet detection, synchronization, equalization, and decoding. The TurboAE in [11] outperforms a pilot-assisted LDPC baseline for short blocklengths up to around 100 channel uses. However, it remains unclear how close the performance of such a learned scheme is to information-theoretic bounds.
In this paper, we address this question by considering a continuous-time waveform channel with an unknown delay and memoryless block fading, for which a finite-blocklength pilot-assisted achievability bound is available [12]. We adapt the TurboAE in [11] to this channel by extending the encoder, synchronizer, decoder, and training strategy. The resulting learned scheme achieves a lower bit error rate (BER) than all evaluated pilot-assisted 5G baselines. However, both the TurboAE and the 5G baselines remain at least 2.1 dB from the finite-blocklength bound for a block error rate (BLER) of , indicating substantial room for improved designs.
System Model
We use the same single-input single-output waveform model as in [12], which is illustrated in Figure 1. The goal is to transmit information bits within a packet comprising complex-valued channel uses. The packet is divided into blocks, each containing symbols, such that . Each block is transmitted over an independent fading channel, with fading coefficient remaining constant over the block. The vector of transmitted symbols corresponding to the th block is

Figure 1. Block diagram of the system model. Each symbol is pulse shaped. The resulting waveform is delayed by , , scaled by the fading coefficient , and corrupted by noise before being filtered and oversampled at the receiver.
The full transmitted symbol vector is . We assume that the energy of each transmitted complex symbol is normalized to 1, i.e.,
The discrete symbols are converted into a continuous-time waveform via pulse shaping as follows:
Here, is the symbol duration, and is a unit-energy rectangular pulse
The transmitted waveform is then passed through a fading channel with unknown delay and fading gain. The corresponding waveform is
where denotes the fading coefficient for block and is the unknown uniformly distributed delay. Throughout, we consider the common delay case . Finally, in (5) is a complex white Gaussian noise process with power spectral density . Under the unit-symbol-energy normalization (2), the average signal-to-noise ratio (SNR) per complex symbol is .
Following [12], we assume that, to obtain an oversampled observation, the receiver filters with the unit-energy rectangular pulse , where and is the oversampling rate, and samples the filter output every seconds. The th received sample in block is
where . Note that is larger than to account for the unknown delay. We collect these samples in the vector
It turns out convenient to write as , where is the integer delay in samples, is the fractional delay, and is the normalized fractional delay. To capture oversampling, we next define
and use (8) to construct the following scaled, zero-padded vector shifted by samples:
Using (9), we can express in (7) as [12 Sec. II-C]
where . The noise samples are independent because the unit-energy receive filters span nonoverlapping sampling intervals. Stacking the , we obtain the received signal (see Figure 1)
Finite-Blocklength Bound & 5G Baselines
Both the finite-blocklength bound and the 5G baselines are obtained under the assumption that each fading block carries a common pilot vector of dimension , followed by a data vector of dimension , i.e.,
The receiver estimates the delay and fading gains from the pilot vectors. Then, it decodes the data by treating these estimates as perfect.
Synchronization and Channel Estimation
Since we assume that the pilots precede the data in (12), all pilot samples of block lie within the first entries of in (7). We denote these samples by . As in [12 Sec. II-C], we consider an estimator that neglects the interference between data symbols and pilots. For a candidate integer delay and fractional delay , we define the noiseless pilot vector per unit channel gain as
where is defined similarly as in (9), with and replaced by and , respectively. Using (10) and (13), we conclude that , where . We use a maximum likelihood (ML) estimator to jointly estimate the fading gains, integer delay, and fractional delay as
For fixed , the conditional gain estimate for block is
Substituting (15) into (14), we obtain the common-delay estimate
For fixed , the objective in (16) is a rational function of with quadratic numerator and denominator. Hence, the optimal is 0, 1, or a point in obtained by setting the derivative of (16) to zero. Evaluating these candidates for all , we obtain [12 Alg. 1]. Finally, and .
Finite-blocklength Achievability Bound
We consider a receiver that filters in (5) with and resamples at the symbol rate using the delay estimate , obtaining the data samples , . Then, the transmitted codeword is recovered by a mismatched scaled nearest-neighbor (SNN) decoder, which treats the estimated fading coefficients as correct and returns
where contains the codewords. We use the so-called random-coding union bound with parameter (RCUs) from finite-blocklength information theory. Specifically, we assume that the codeword symbols are generated as where . Define the event that transmitted symbols are synchronized to within one symbol period as , and let denote its complement. The bound is stated in the following theorem [12 Sec. III].
Theorem 1 (Pilot-assisted RCUs bound): For the pilot-assisted model above and every , there exists a codebook with codewords, each containing data symbols, whose average packet error probability , averaged over equiprobable messages, satisfies
where
Here, is an optimization parameter, and
where follows the same distribution as and is independent of all other quantities.
Pilot-Assisted 5G Baselines
Similar to the RCUs bound, the baseline receiver uses the delay and fading gain estimates from Section III-A, filters the waveform with , resamples at the symbol rate, and decodes by treating the estimated delay and channel coefficients as perfect. We compare three coding schemes: a 5G polar code with cyclic redundancy check (CRC)-aided successive-cancellation (SC) decoding (5G polar + CRC SC), a 5G polar code with CRC-aided successive cancellation list (SCL) decoding of list size eight (5G polar + CRC SCL-8), and a 5G LDPC code with 20 belief propagation iterations (5G LDPC BP-20). To satisfy the constraint (2), we assume that all coding schemes use phase-shift keying (PSK) modulation.
TurboAE Design
The proposed TurboAE, illustrated in Figure 2, consists of a learned encoder at the transmitter and a learned synchronizer and decoder at the receiver. While the encoder and decoder largely build on [11 Figs. 5 and 20], our main contributions, i.e., the synchronizer and the training strategy, differ from [11] in two respects. First, the synchronizer in [11] represents the integer delay as a one-hot probability vector and lets the fractional delay to be accounted for by the decoder. For the assumed rectangular pulse in this design, this would lead to significantly degraded performance. Our synchronizer instead estimates both integer and fractional delay and compensates for them before decoding (see Section IV-B). Second, in the training strategy used in [11], the synchronizer is trained jointly with the decoder using a categorical cross-entropy loss. To suit the proposed synchronizer design, we instead use a regression loss for synchronization and perform the training in two stages: first we train the encoder, decoder, and synchronizer separately; then jointly (see Section IV-D). We describe each component of the proposed architecture next.

Figure 2. End-to-end TurboAE architecture.
Encoder
The encoder contains two 1D convolutional blocks, separated by a straight-through binarizer [13] and a fixed interleaver. The outputs of the encoder are normalized to satisfy (2).
Synchronizer
The learned synchronizer, illustrated in Figure 3, maps the oversampled observations to one continuous common-delay estimate and uses it to obtain the delay-corrected observations . It builds on two key ideas. First, instead of predicting directly, it predicts the integer and fractional delays ( and ) separately and combines them into . Second, rather than predicting in a single step, it first forms a model-based prior estimate by conventional feature extraction and then corrects it with neural networks. We next describe each block in Figure 3.

Figure 3. Learned model-guided neural synchronizer. Candidate features provide integer-delay probabilities and fractional-delay priors, while global temporal features refine the fractional-delay estimates before soft combination.
a) Candidate features: For each possible integer delay , let index the samples the packet would occupy if the integer delay were . We consider the following nine candidate features : the normalized assumed delay ; three power features, i.e., the empirical mean and standard deviation of the power of the samples inside and the alignment score , defined as the average power inside minus the power outside ; two edge features, i.e., and , where is the difference between the power of the first/last sample inside and the average power outside ; and
is an estimate of the fractional-delay prior (see (10)); and three correlation features, i.e., the real part, imaginary part, and magnitude of the mean of inside .
b) Embedding & heads: We use a two-layer fully connected network to embed into . Furthermore, we use two separate linear heads ( and ) combined with tanh activation to produce the integer-delay probabilities and initial fractional delay , where is a trainable scalar and the tanh terms determine the influence of the correction according to
c) Global temporal features: In parallel, we use another neural network , consisting of residual 1D convolutional blocks and an attention-pooling layer, to learn temporal features directly from as complementary information for the fractional-delay correction according to
d) Soft combination: Combining (22) with the refined fractional-delay estimates , we obtain the final common delay estimate
e) Delay correction: Let and . In block , we remove the first samples that carry only noise and retain . Then, we undo the interference between neighboring symbols caused by the fractional delay. By (10), this interference is described by
where repeats each symbol times, is the identity matrix, is a column vector of ones, and denotes the Kronecker product. The zero row has entries. Treating as exact, we recover the faded symbols by regularized least squares with regularization parameter and repeat them according to the oversampling factor :
Decoder
The decoder uses two groups of 1D convolutional networks, denoted by Dec 1 and Dec 2 in Figure 2, which iteratively exchange information through the fixed interleaver . The input is first normalized and embedded along the symbol dimension, and then fed into the decoder for multiple iterations, each with its own trainable Dec 1 and Dec 2. Starting from a zero prior , in each iteration, Dec 1 combines the received features with , Dec 2 refines the deinterleaved output of Dec 1, and the interleaved output of Dec 2 becomes the next . Each network subtracts its input information from its output, so only new (extrinsic) information is passed on. In the last iteration, the decoder returns the bit probabilities.
Algorithm 1 Two-stage training of the TurboAE
Input: Total steps $N_{\mathrm{step}}$, first-stage steps $N_{\mathrm{pre}}$, and pass counts $K_{\mathrm{Enc}}$, $K_{\mathrm{Dec}}$, $K_{\mathrm{SyncEnc}}$, $K_{\mathrm{Sync}}$, $K_{\mathrm{JointEnc}}$, $K_{\mathrm{JointDec}}$
Output: $\theta_{\mathrm{Enc}}$, $\theta_{\mathrm{Dec}}$, and $\theta_{\mathrm{Sync}}$
1: Initialize $\theta_{\mathrm{Enc}}$, $\theta_{\mathrm{Dec}}$, $\theta_{\mathrm{Sync}}$, and their Adam optimizers
2: for $i = 1, \ldots, N_{\mathrm{step}}$ do
3: if $i \leq N_{\mathrm{pre}}$ then ▷ stage 1: separate training
4: PASS(ENC, $K_{\mathrm{Enc}}$, $\mathcal{L}_b$, $\{\theta_{\mathrm{Enc}}\}$) ▷ perfect sync.
5: PASS(DEC, $K_{\mathrm{Dec}}$, $\mathcal{L}_b$, $\{\theta_{\mathrm{Dec}}\}$) ▷ perfect sync.
6: PASS(ENC, $K_{\mathrm{SyncEnc}}$, $\mathcal{L}_s$, $\{\theta_{\mathrm{Enc}}\}$)
7: PASS(SYNC, $K_{\mathrm{Sync}}$, $\mathcal{L}_s$, $\{\theta_{\mathrm{Sync}}\}$)
8: else ▷ stage 2: joint training
9: PASS(ENC+SYNC, $K_{\mathrm{JointEnc}}$, $\mathcal{L}_J$, $\{\theta_{\mathrm{Enc}}, \theta_{\mathrm{Sync}}\}$)
10: PASS(DEC+SYNC, $K_{\mathrm{JointDec}}$, $\mathcal{L}_J$, $\{\theta_{\mathrm{Dec}}, \theta_{\mathrm{Sync}}\}$)
11: end if
12: end for
Training
Let , , and denote the encoder, decoder, and synchronizer parameters. To train the encoder and decoder pair, we use the mean binary cross-entropy (BCE) loss
where and are the transmitted bits and corresponding predicted probabilities. To train the encoder and synchronizer pair, we use the mean smooth- loss
where if and otherwise, with threshold . For joint training, we use , where the parameter balances the two losses.
In Algorithm 1, we extend the alternating training of [7], in which the encoder is trained with the decoder frozen and vice versa to avoid poor local optima. In each PASS, we perform gradient-based updates of to minimize , each time on a new batch of random messages and channel realizations, with all other parameters frozen. In stage one, the encoder and decoder are first trained with perfect timing to minimize the BCE loss (27); then the encoder and synchronizer are trained to minimize the synchronization loss (28), so that the encoder learns the two objectives alternately. In stage two, the synchronizer is updated with either the encoder or the decoder under , with the remaining component frozen.
Results
For the numerical results,2 we consider complex symbols and a rate of bit per channel use, as in [11]. We normalize the symbol period to and set . We consider fading blocks. Oversampling by gives ; so each block contains samples. As in [11], we use Zadoff–Chu pilots [15] for the 5G baselines and the RCUs bound, which is optimized over and . The 5G baselines are simulated with Sionna [16]. Under block fading, the optimized 5G baselines (Section III-C) involve the transmission of pilots per block, 16-PSK for 5G polar + CRC SC, and 32-PSK for both 5G polar + CRC SCL-8 and 5G LDPC BP-20. The TurboAE is trained with Adam for 700 steps with an initial batch size of 500 and a learning rate of .
To single out the benefit of joint synchronization, channel estimation, and decoding, we first consider an additive white Gaussian noise (AWGN) channel with perfect synchronization, where the baselines use QPSK. In this case (Figure 4a), TurboAE has the lowest BER up to about , but its BER decays more slowly than that of 5G polar + CRC SCL-8 and 5G LDPC BP-20. Hence, it offers no coding gain over them at moderate-to-high SNR.

Figure 4. BER (left) and BLER (right) versus SNR for and . Top: AWGN; bottom: Rayleigh block fading (, , , common delay ). The 5G baselines use QPSK under AWGN; under block fading, 5G polar + CRC SC uses 16-PSK, while 5G polar + CRC SCL-8 and 5G LDPC BP-20 use 32-PSK. Two pilots per block are transmitted in these schemes. The TurboAE uses learned symbols without dedicated pilots.
For the considered block-fading channel (Figure 4c), instead, TurboAE achieves the lowest BER over the entire considered SNR range and requires 1.9 dB less SNR than 5G LDPC BP-20 to achieve a BER of . We attribute this advantage to joint synchronization and decoding, which avoids the pilot overhead. This is consistent with [11 Fig. 13], where, for the same packet size and rate, the TurboAE-based scheme outperforms a LDPC baseline with a 20-symbol preamble by 2.7 dB in BER over a multipath channel.
In terms of BLER under AWGN (Figure 4b), TurboAE requires a higher SNR than 5G polar + CRC SCL-8, 5G LDPC BP-20, and the RCUs bound. Instead, for the considered block-fading channel (Figure 4d), TurboAE slightly outperforms 5G polar + CRC SC but, at a BLER of , requires 1.4, 2.4, and 4.5 dB more SNR than 5G LDPC BP-20, 5G polar + CRC SCL-8, and the RCUs bound, respectively. Hence, the 1.9 dB BER gain over 5G LDPC BP-20 turns into a 1.4 dB BLER loss. As also noted in [11], this discrepancy between BER and BLER performance may be partly due to the use of the BCE loss (27), which targets individual bits rather than whole packets. Inspired by [11], we consider a concatenated scheme, in which we (i) append a 3-bit CRC with generator polynomial , (ii) retain at the decoder the eight most likely candidates under the predicted bit probabilities, and (iii) select the most likely CRC-valid candidate. This improves the TurboAE BLER by about 1.6 dB. On the contrary, finetuning with a BLER-targeted loss [14] (see Figure 4d) or adding a block-reconstruction penalty to the loss (27) improves the BLER only slightly.
Conclusion
We developed an end-to-end learned scheme for short-packet transmission over a continuous-time waveform channel with an unknown delay and memoryless block fading. In the proposed scheme, the delay is estimated without dedicated pilots. The scheme contains an encoder, a synchronizer, and a decoder, which are trained alternately. Under the same blocklength and rate constraints, TurboAE achieves a lower BER than all evaluated pilot-assisted 5G baselines but a higher BLER than 5G polar + CRC SCL-8; with a 3-bit CRC, its BLER is comparable to that of 5G LDPC BP-20. Finally, our work uncovered a large BLER gap between all evaluated schemes and the RCUs bound, motivating future work on new architectures and training methods for short-packet communication.
References
- [1]R. G. Gallager, “Low-density parity-check codes,” IRE Trans. Inf. Theory, vol. 8, no. 1, pp. 21–28, Jan. 1962.
- [2]E. Arıkan, “Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels,” IEEE Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009.DOI
- [3]Y. Polyanskiy, “Asynchronous communication: Exact synchronization, universality, and dispersion,” IEEE Trans. Inf. Theory, vol. 59, no. 3, pp. 1256–1270, Mar. 2013.
- [4]A. Lancho, J. Östman, and G. Durisi, “On joint detection and decoding in short-packet communications,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Madrid, Spain, Dec. 2021.arxiv.org/abs/2109.13669
- [5]J. Östman, G. Durisi, E. G. Ström, M. C. Coşkun, and G. Liva, “Short packets over block-memoryless fading channels: Pilot-assisted or noncoherent transmission?” IEEE Trans. Commun., vol. 67, no. 2, pp. 1521–1536, Feb. 2019.
- [6]T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,” IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, Dec. 2017.arxiv.org/abs/1702.00832
- [7]Y. Jiang, H. Kim, H. Asnani, S. Kannan, S. Oh, and P. Viswanath, “Turbo autoencoder: Deep learning based channel codes for point-to-point communication channels,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 32, Vancouver, BC, Canada, Dec. 2019.arxiv.org/abs/1911.03038
- [8]J. Clausius, S. Dörner, S. Cammerer, and S. ten Brink, “Serial vs. parallel turbo-autoencoders and accelerated training for learned channel codes,” in Proc. Int. Symp. Topics Coding (ISTC), Montreal, QC, Canada, Aug. 2021, pp. 1–5.
- [9]C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon limit error-correcting coding and decoding: Turbo-codes,” in Proc. IEEE Int. Conf. Commun. (ICC), vol. 2, Geneva, Switzerland, May 1993, pp. 1064–1070.
- [10]S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara, “Serial concatenation of interleaved codes: performance analysis, design, and iterative decoding,” IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 909–926, 1998.
- [11]S. Dörner, J. Clausius, S. Cammerer, and S. ten Brink, “Learning joint detection, equalization and decoding for short-packet communications,” IEEE Trans. Commun., vol. 71, no. 2, pp. 837–850, Feb. 2023.arxiv.org/abs/2207.05699
- [12]A. O. Kislal, M. Rajiv, G. Durisi, E. G. Ström, and U. Mitra, “Is synchronization a bottleneck for pilot-assisted URLLC links?” IEEE Trans. Wireless Commun., vol. 23, no. 12, pp. 17945–17958, Dec. 2024.arxiv.org/abs/2401.09089
- [13]Y. Bengio, N. Léonard, and A. Courville, “Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation,” in arXiv:1308.3432, Aug. 2013.arxiv.org/abs/1308.3432
- [14]R. Weis mayr, G. Marti, C. Dick, H. Song, and C. Studer, “Bit error and block error rate training for ML-assisted communication,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Jun. 2023.
- [15]5G; NR; Physical channels and modulation, 3GPP Technical Specification 38.211, Rev. 18.6.0, Apr. 2025.
- [16]J. Hoydis, S. Cammerer, F. Ait Aoudia, M. Nimier-David, L. Maggi, G. Marcus, A. Vem, and A. Keller, “Sionna,” 2022, https://nvlabs.github.io/sionna/.arxiv.org/abs/2203.11854