When a radio receiver tries to recover data without a long training preamble, it faces a basic symmetry. The received signal $Y = HX + W$ is unchanged if the unknown channel $H$ is rotated by $90^\circ$ and the unknown symbols $X$ are rotated back by the same amount. For a square QAM constellation (a grid of amplitude-and-phase points), the receiver can recover the shape of the grid but not which way is "up." Current blind and semi-blind receivers handle this in one of three ways. They send one known pilot symbol and trust it absolutely, they encode data in phase differences between symbols and pay about 3 dB for it, or they converge first and then try every rotation against a checksum. All three treat orientation as a hard decision made outside the inference.
We propose to make orientation part of the inference. The finite group of rotations can be factored exactly out of expectation propagation (EP), a message-passing method that approximates a posterior by repeatedly matching means and variances. A Gaussian handles the part of the posterior that does not depend on rotation. A small discrete distribution, one four-way variable per user stream, carries the orientation. That distribution is updated softly by whatever rotation-sensitive information the packet already contains: CRC bits, device IDs, code constraints, or a single optional pilot. The regime where this matters is opportunistic and underlay spectrum sharing, where secondary users send short bursts at low power into idle or lightly loaded licensed bands. There, every pilot symbol and handshake erodes the gain from reusing idle spectrum, and a single anchor symbol received at low SNR is often wrong. Measurements summarized in Intelligent and efficient development of wireless networks: A review of cognitive radio networks report that over 60% of licensed spectrum below 6 GHz is idle at any time.
Core Claim
Plain Gaussian EP is a poor fit for this symmetry, and this is the point the obvious combination misses. The research file credits EP's forward-KL "mass-covering" behavior with preserving uncertainty across orientations. Applied to a posterior with four symmetric modes, however, mass-covering moment matching places the Gaussian mean at the centroid of the modes, which is zero. The receiver is then stuck at a saddle point, and in practice only random initialization or an anchor breaks it.
Our fix is to change the family EP projects onto. For stream $k$, with channel column $h_k$ and symbol row $x_k$, let $g_k \in \{0,1,2,3\}$ index the rotation $j^{g_k}$. The approximating family is
$$q(h_k, x_k, g_k) = \pi_k(g_k)\, \tilde q\big(j^{-g_k} h_k,\; j^{g_k} x_k\big),$$
where $\tilde q$ is an ordinary EP approximation (a Gaussian over the channel and factorized discrete-projected terms over symbols) expressed in a canonical orientation, and $\pi_k$ is a four-point distribution.
The likelihood and the symmetric constellation prior are invariant under the rotation. Their EP updates therefore act only on $\tilde q$, with the same cost as standard EP, because computing one component yields all four by rotation. Factors that are not invariant act mainly on $\pi_k$. These include CRC parity checks on demapped bits, known ID bits, and an optional pilot. Once $\pi_k$ becomes confident, it feeds back into the symbol beliefs and sharpens them.
A primary-user interferer is a second useful case. Its orientation is a nuisance variable. It never needs to be resolved, because cancelling it only requires the product $H_p X_p$, and that product is invariant under rotation. The orbit form therefore lets the receiver keep a primary interferer's orientation unresolved while subtracting its contribution.
The claim to test is this: at a fixed dedicated-pilot budget of zero or one symbol, orbit EP achieves lower block error rate than hard-anchored or post-hoc-resolved blind receivers in low-power underlay conditions. It also outputs a calibrated probability of wrong orientation, which a URLLC link can use to declare an erasure instead of delivering a rotated packet.
Relation to Prior Work
Several works motivate the problem, and each resolves orientation outside the inference loop.
- Hard single-symbol anchoring. Next-Generation URLLC with Massive Devices factorizes $Y$ with bilinear generalized approximate message passing (BiG-AMP). It uses a single scalar pilot plus IDs and CRC to initialize and to break phase and permutation ambiguity. The anchor is a hard seed; our pilot, CRC and ID bits are soft factors inside the iterations.
- Post-hoc resolution. Blind Data Detection in Massive MIMO via ℓ₃-Norm Maximization Over the Stiefel Manifold and Blind Joint MIMO Channel Estimation and Decoding recover the constellation up to ambiguity by geometric optimization, then fix orientation with a reference symbol after the fact.
- Differential modulation. Blind Grant-Free Random Access With Message-Passing-Based Matrix Factorization in mmWave MIMO mMTC avoids the question with differential modulation and accepts the SNR penalty. Our proposal is meant to recover that penalty.
Three papers supply method.
- EP with iterative channel refinement. Cell-Free Massive MIMO Detection: A Distributed Expectation Propagation Approach shows that EP with iterative channel refinement from detected symbols gains up to 4.5 dB over non-orthogonal-pilot estimates after four iterations. It still needs pilots to set orientation. We reuse its EP detector structure and remove that dependence.
- Spectral initialization. Spectral Method for Phase Retrieval: An Expectation Propagation Perspective gives an EP-derived spectral initializer and weak-recovery thresholds for problems with a global phase ambiguity. A spectral estimate recovers a subspace, which is exactly the rotation-invariant quantity. We use it to initialize $\tilde q$, though its guarantees are for Haar matrices and intensity measurements, not bilinear constellation channels.
- Blind EM with belief propagation. Blind Channel Estimation and Joint Symbol Detection With Data-Driven Factor Graphs interleaves EM channel estimation with belief propagation and finds that EM converges to a surrogate channel matched to the approximate detector. This warns us to judge orbit EP by bit and block errors, not channel mean squared error alone.
Two papers constrain the design.
- When blind processing pays. Fundamental limits via CRB of semi-blind channel estimation in Massive MIMO systems shows that semi-blind processing helps mainly when pilot power is low, and that a pilot count fixed in proportion to users leaves a non-vanishing error floor. That is a statement about channel estimation error, not about orientation. We therefore expect our gains in the low-power secondary regime and not in high-power licensed links. We will measure the error floor rather than claim to remove it.
- EP instability. Bounding errors of Expectation-Propagation is a reminder that EP cavity steps can produce negative variances. We will use damping and report non-convergence rates.
No work in the file places EP-based blind detection in cognitive or shared-spectrum links. The cognitive-access papers, for example Interference-based Optimal Power-efficient Access Scheme for Cognitive Radio Networks, work at the queue and power level with abstracted physical layers.
Research Program
The receiver. We will build orbit EP on top of an EP bilinear receiver in the style of the cell-free ICD-EP detector, extended to fully blind operation. Initialization comes from a PCA-EP-style spectral estimate of the channel subspace. Two message schedules will be derived.
- Symbol-level soft messages. For each candidate rotation $g_k$, demapped bit log-likelihood ratios are scored against CRC parity and known ID bits. This gives a message on $g_k$.
- Feedback of orientation. $\pi_k$ reweights the symbol-level discrete prior before the next moment-matching step.
Permutation ambiguity. User permutation is handled by soft assignment of recovered streams to device IDs, using the same ID likelihoods. Permutation is kept outside the orbit variable, because the full permutation group does not factorize across streams the way per-stream rotations do.
The primary interferer. In the underlay setting we add the primary interferer as an extra stream whose orbit variable is marginalized rather than resolved.
Analysis. The analytic part asks when the soft orientation posterior becomes confident. We expect the log-odds of the correct rotation to grow roughly linearly in the number of informative parity and ID bits, weighted by their reliability. The reason is that each is a conditionally near-independent vote, which makes a sum of log-likelihood ratios. This yields a "pilot-equivalent" count for a given CRC length and SNR. We will check this prediction empirically rather than assume it.
Baselines
Pilot-aided detection. The naive reference is conventional pilot-aided linear MMSE channel estimation followed by EP detection, with pilot length equal to the number of active users. It shows what the overhead buys. We will also run it with pilots shortened to match our budget, so that the cost of shortened pilots is visible.
Hard-anchor BiG-AMP (strongest competitor). This is the reference-information receiver of Next-Generation URLLC with Massive Devices: one scalar pilot, IDs, CRC and SIC. It targets exactly our pilot budget. On the one-ring model with $M=512$, it reports maintaining 99.99999% reliability with up to 65 simultaneous devices, and reducing latency from 18.3 ms to 0.6 ms. It will be run on the same one-ring channels, packet lengths, CRC lengths, SNR grid and iteration budget as orbit EP. Its damping and initialization will be tuned with the same number of trials. Beating it means supporting more devices at that reliability target, or reaching it at lower SNR, with the same reference information.
Differential blind receiver. This is the UAMP matrix-factorization receiver of Blind Grant-Free Random Access With Message-Passing-Based Matrix Factorization in mmWave MIMO mMTC. On its synthetic benchmark ($M=100$, $U=300$, 10 paths, 10% activity, $L=200$) it reports detection about 1 dB from pilot-assisted bounds and zero active-user error above 3 dB. We will rerun it on that exact setup, where beating it means a coherent gain at matched activity-detection error.
Symmetric EP with post-hoc CRC search (key ablation). This is identical to orbit EP except that $\pi_k$ is frozen uniform during iteration and orientation is chosen after convergence by testing all four rotations against the CRC. It isolates the proposal's key ingredient, which is soft feedback of orientation during inference.
Mean-field variational inference. This is a mean-field variational version with the same factor graph and initialization. It tests whether EP's moment matching matters once symmetry is handled.
Why not just…?
Why not just keep one pilot symbol and anchor to it? This works at moderate SNR, as the BiG-AMP results show. A hard anchor, however, rests on the correct detection of one symbol. In underlay links, the CRB analysis of Fundamental limits via CRB of semi-blind channel estimation points to low pilot power as the regime where semi-blind methods pay, and there a single symbol is misread often. The comparison against the hard-anchor baseline across SNR and interference-to-noise ratio settles whether soft pooling of pilot, ID and CRC evidence buys anything.
Why not just run a blind solver and try all four rotations against the CRC? This is the real threat. If the ablation with uniform $\pi_k$ matches orbit EP in block error rate, the contribution shrinks to calibration and erasure signaling, and we will report it that way. We expect a gap only at low SNR. There, every rotation fails the CRC because of residual bit errors, so post-hoc search returns nothing. Soft parity votes can still identify the right orientation, and feeding it back can reduce those bit errors. The ablation is designed to show whether this happens.
Why not just use differential modulation? It eliminates the ambiguity at a cost of about 3 dB, per the characterization of Blind Grant-Free Random Access. It also degrades with more active users and does not extend cleanly to dense QAM. Power-limited secondary users can least afford the loss: Interference-based Optimal Power-efficient Access Scheme shows secondary throughput falling sharply as energy budgets tighten. Head-to-head block error rate at matched spectral efficiency on the Wang et al. benchmark measures how much of the 3 dB orbit EP recovers.
Why not just use variational inference, which converges more reliably? Mode-seeking variational inference picks one rotation, which is harmless only if a later step fixes orientation. It gives no calibrated probability of being wrong. The mean-field baseline tests both error rate and calibration directly. If it matches orbit EP on both, the EP-specific part of the claim fails.
Proposed Experiments
Benchmarks. We use three settings.
- Synthetic mmWave mMTC. The benchmark from Wang et al., for comparison with differential detection.
- One-ring massive MIMO. The setup with $M=512$ from Ke et al., for comparison with hard anchoring.
- Underlay variant. The one-ring setup with an added QPSK primary interferer at controlled interference-to-noise ratio, and secondary transmit power capped. We found no reported result on Lacuna for blind or semi-blind detection under primary-user interference, so all numbers in this setting are new and every baseline is rerun there.
Metrics. We report block error rate versus SNR at dedicated pilot budgets of 0, 1 and 2 symbols. We report orientation error rate and its calibration, measured as the reliability of $\pi_k$. We report the number of pilot symbols saved at a fixed block error rate target, EP non-convergence rate, and runtime per iteration.
Decisive checks. Three checks decide the outcome.
- Soft feedback versus post-hoc search. Orbit EP must beat the post-hoc CRC ablation at low SNR, or the core mechanism is not needed.
- Pilot-free operation. At zero pilots, orbit EP should approach the one-pilot hard-anchor receiver. This would show that CRC and ID bits alone can replace the anchor.
- Nuisance interferer. In the underlay setting, marginalizing the primary stream's orientation should cancel interference as well as an oracle that knows that orientation. If it does not, the nuisance-orbit argument is wrong and must be revised.
References
- Intelligent and efficient development of wireless networks: A review of cognitive radio networks
- Next-Generation URLLC With Massive Devices: A Unified Semi-Blind Detection Framework for Sourced and Unsourced Random Access
- Blind Data Detection in Massive MIMO via ℓ3-norm Maximization over the Stiefel Manifold
- Blind Joint MIMO Channel Estimation and Decoding
- Blind Grant-Free Random Access With Message-Passing-Based Matrix Factorization in mmWave MIMO mMTC
- Cell-Free Massive MIMO Detection: A Distributed Expectation Propagation Approach
- Spectral Method for Phase Retrieval: An Expectation Propagation Perspective
- Blind Channel Estimation and Joint Symbol Detection With Data-Driven Factor Graphs
- Fundamental Limits via CRB of Semi-Blind Channel Estimation in Massive MIMO Systems
- Bounding errors of Expectation-Propagation
- Interference-based optimal power-efficient access scheme for cognitive radio networks