A Variational Bayes Approach to Decoding in a Phase-Uncertain Digital Receiver

This paper presents a Bayesian approach to symbol and phase inference in a phase-unsynchronized digital receiver. It primarily extends [Quinn 2011] to the multi-symbol case, using the variational Bayes (VB) approximation to deal with the combinatorial complexity of the phase inference in this case. The work provides a fully Bayesian extension of the EM-based framework underlying current turbo-synchronization methods, since it induces a von Mises prior on the time-invariant phase parmeter. As a result, we achieve tractable iterative algorithms with improved robustness in low SNR regimes, compared to the current EM-based approaches. As a corollary to our analysis we also discover the importance of prior regularization in elegantly tackling the significant problem of phase ambiguity.

💡 Research Summary

This paper addresses the problem of jointly estimating transmitted symbols and an unknown carrier‑phase offset in a digital receiver where the transmitter and receiver oscillators are not synchronized. While existing turbo‑synchronization techniques rely on an Expectation‑Maximization (EM) framework that iteratively refines soft symbol probabilities and a point estimate of the phase, they discard the full posterior uncertainty of the phase parameter. The authors propose a fully Bayesian treatment that introduces a von Mises (Tikhonov) prior on the phase, thereby preserving conjugacy and enabling analytic updates for the phase posterior.

For a single‑symbol transmission, the exact Bayesian posterior can be derived in closed form: the phase posterior remains von Mises with a concentration parameter κₗ = κ₀ + 2r (lᵀa)ᴴx, where κ₀ is the prior concentration, r is the noise variance, a is the constellation point, and l is a one‑hot indicator of the transmitted symbol. When κ₀ = 0 (uniform prior) the posterior concentration becomes independent of the transmitted symbol, leading to the well‑known phase‑ambiguity problem. A non‑zero κ₀ breaks this symmetry, yielding a unique MAP estimate for each symbol.

In the multi‑symbol case, the joint posterior involves a mixture of Mᴷ components (M = constellation size, K = number of symbols), which is computationally intractable. To overcome this, the paper applies the Variational Bayes (VB) approximation, factorizing the posterior into independent distributions over the phase φ and the label vectors {l₁,…,l_K}. Minimizing the Kullback‑Leibler divergence yields update equations that keep the phase posterior in the von Mises family and the label posteriors as multinomials. Two VB schemes are presented:

1. Offline VB – processes an entire batch of K symbols at once. The phase posterior is collapsed into a single von Mises distribution with concentration κ_{lK}=κ₀+2r∑_{i=1}^{K}(l_iᵀa_i)ᴴx_i, and each label posterior is a multinomial with parameters proportional to the prior symbol probabilities weighted by the exponential of the sufficient statistics.

2. Online VB – updates the phase posterior sequentially after each symbol. The mixture of M von Mises components generated by the current symbol is approximated by a single von Mises distribution, whose concentration is updated recursively: κ̂_{t}=κ̂_{t‑1}+2r(l_tᵀa_t)ᴴx_t. This single distribution then serves as the prior for the next symbol, enabling real‑time operation with O(M) complexity per step.

Simulation results use 20‑symbol batches transmitted over an AWGN channel. In the low‑SNR regime (‑15 dB to 5 dB), both offline and online VB algorithms significantly outperform the conventional EM‑based turbo‑synchronization and pilot‑assisted methods. The Bayesian approaches achieve higher symbol‑identification rates while requiring only a constant‑time update of the prior concentration, i.e., no extra computational burden beyond the standard VB updates.

A key insight is that the von Mises prior regularizes the rotational invariance inherent in QAM constellations. By selecting κ₀ > 0, the posterior phase distribution becomes unimodal, eliminating the phase‑ambiguity that plagues uniform‑prior methods. As the number of observations grows, the influence of the exact value of κ₀ diminishes, but a non‑zero prior is crucial for robust convergence in the early stages.

In summary, the paper delivers a principled Bayesian extension of turbo‑synchronization, replaces the point‑estimate EM step with a full posterior update, and demonstrates that variational approximations make the approach tractable for realistic multi‑symbol transmissions. The proposed framework offers improved low‑SNR robustness, eliminates the need for extensive pilot symbols, and incurs negligible computational overhead, making it attractive for low‑power, high‑throughput digital receivers operating under severe phase uncertainty.