A base station with many antennas has to recover several users' data symbols from a noisy mixture of them. Exact "optimal" detection costs exponentially many operations in the number of users. Cheap linear detectors lose 4 to 10 dB when the radio channel is strongly correlated, which is the normal case at millimeter-wave frequencies. The middle ground is iterative probabilistic detectors, which keep a Gaussian belief about the continuous quantities and a discrete belief about the symbols. These get close to optimal, but they pay a cubic-cost linear solve per iteration and are rerun from scratch for every received vector. At the same time, the channel itself changes slowly from one time slot to the next, even for users moving at walking or driving speed.
This note proposes to split the detector's internal state by how it changes over time. Everything that depends on the channel (the belief about the channel matrix, the linear-solver state, the preconditioner) is carried forward and warm-started from the previous slot. Everything that depends on the transmitted data (the beliefs about the symbols and their variance schedule) is restarted on every vector, because the data are independent from one vector to the next. The claim is that this split lets a detector treat the channel as uncertain and jointly infer channel and symbols, instead of pretending its channel estimate is exact, at roughly the per-vector cost of a detector that assumes perfect channel knowledge.
Core Claim
Most warm-starting in the literature reuses the previous solution because consecutive problems have nearly the same answer. Examples are shifting control sequences forward in receding-horizon control (Distributed Formation Control for Multi-Vehicle Systems With Splitting and Merging Capability, Real-Time Iteration Scheme for Diffusion Policy) and shifting branch-and-bound frontiers and dual bounds in hybrid control (Warm Start of Mixed-Integer Programs for Model Predictive Control of Hybrid Systems). In data detection this premise is half false. Consecutive symbol vectors are statistically independent, so seeding the symbol estimate with the previous vector's answer injects a wrong prior. The channel, however, is strongly correlated over time.
We therefore treat detection as a hybrid discrete–continuous inference problem over both the symbols $x$ (discrete, from a constellation such as 256-QAM) and the channel $H$ (continuous, drifting). The inference method is expectation consistency (EC), an approximate-inference scheme that alternates between a Gaussian estimator and a discrete-prior estimator. We then warm-start only the continuous, channel-side part of its fixed point across slots.
The new element is not warm-starting in general, and not EC detection in general. It is three things together. First, the time-based partition of a hybrid inference state. Second, a fixed per-slot iteration budget in the spirit of real-time iteration. Third, the complexity claim that joint channel–symbol inference then costs about the same as known-channel inference.
The closest prior works each supply one piece. Warm-Starting in Message Passing algorithms warm-starts inner conjugate-gradient solves across outer iterations of Vector Approximate Message Passing. It reaches near-optimal recovery in 20 outer iterations on an ill-conditioned problem where Memory AMP needs hundreds to thousands, but it operates within a single static problem, not across physical time. Probabilistic MIMO Symbol Detection with Expectation Consistency Approximate Inference supplies the EC detector, which comes within 0.4 dB of optimal in 5×5 systems. That paper solves each block cold with the channel known. Detect to Learn tracks the channel over time by feeding detected symbols back as pseudo-pilots, but it does so with a heuristic attention loss rather than a principled posterior over $H$.
Research Program
The model at slot $t$ is $y_t = H_t x_t + w_t$. The channel follows a first-order drift $H_t = \rho H_{t-1} + \sqrt{1-\rho^2}\, E_t$, where $\rho$ is set by the Doppler spread and $E_t$ is an innovation term. The detector maintains a Gaussian belief over $H_t$ that is refined within the slot using soft symbol estimates. In effect this is a soft, variance-aware version of decision feedback.
The EC iteration has three kinds of state. The channel factor and the LMMSE-type linear solve depend on $H_t$. The symbol factor and its variance limit depend on $x_t$. Our warm-start rule has three parts:
- Carried forward: the channel belief is propagated through the drift model, and the conjugate-gradient iterate and preconditioner for the Gram-type system are reused.
- Restarted: the symbol beliefs are reset to the constellation prior, and the "progressive variance limit" schedule is rerun from its start.
- Budgeted: each slot gets a fixed small number of inner and outer iterations.
Restarting the variance schedule matters. The EC paper uses that schedule specifically to avoid bad local minima, and carrying converged symbol variances forward would skip it.
Two cautions from prior work shape the design:
- Stability. The message-passing warm-start paper reports that textbook moment-based variance estimates become unstable once errors accumulate, and it needs robustified variants. We adopt the robust variant from the outset. We also measure drift of the channel belief over long runs.
- Fixed budgets. Closed-loop Analysis of ADMM-based Suboptimal Linear Model Predictive Control shows that warm-starting can keep stability guarantees even under a fixed, early-stopped iteration count. This motivates analyzing our fixed-budget detector as a tracking loop rather than as a converged solver.
The complexity analysis rests on a standard conjugate-gradient bound. Warm-starting the linear solve from the previous slot's solution leaves an initial error proportional to the channel innovation. The number of inner iterations needed to reach tolerance $\epsilon$ is therefore roughly
$$k_t \approx \tfrac{1}{2}\sqrt{\kappa_t}\,\log\!\big(\|\Delta_t\|/\epsilon\big),$$
where $\kappa_t$ is the condition number of the regularized Gram matrix and $\|\Delta_t\|$ is the slot-to-slot change in the solve's solution. The form comes from CG's geometric contraction rate $(\sqrt{\kappa}-1)/(\sqrt{\kappa}+1)$. The prediction follows directly: per-slot cost should fall logarithmically as mobility decreases. Poor conditioning from line-of-sight correlation, which is exactly where channel hardening fails (A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO), should raise $k_t$ only through $\sqrt{\kappa_t}$. It should not force a return to cubic-cost inversion.
The role of weighted model integration (WMI), the hybrid-inference formalism built on SMT solvers (Probabilistic Inference in Hybrid Domains, SMT-based Weighted Model Integration with Structure Awareness), is deliberately limited here. We borrow its framing of exact discrete–continuous posteriors. We also answer its stated open problem that queries are solved in isolation with no reuse across time. But we do not use polytope-integration solvers, for reasons given below.
Baselines
Naive reference. LMMSE detection using the latest pilot-based channel estimate, computed by exact inversion once per coherence block. This is the "do nothing clever" point that every receiver already has.
Strongest hardware-oriented method. The unfolded Gram-domain block coordinate-descent detector of A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO. It is the strongest competitor because it was designed and evaluated on exactly our target: 60 GHz QuaDRiGa urban microcell channels, 128 base-station antennas, 16 users, up to 256-QAM. It reports an SNR advantage of 4 dB to 10 dB over LMMSE in high-load 16×16 configurations. Beating it means lower SNR at a matched target coded error rate with equal or fewer multiply-accumulates per vector. We will reimplement it in floating point, with its unfolded parameters trained on the same channel draws and the same training budget we give our own tuning.
Strongest inference-quality method. The cold-started EC detector (Probabilistic MIMO Symbol Detection with Expectation Consistency Approximate Inference). Its reported numbers are within 0.4 dB of optimal with LDPC coding in 5×5 Rayleigh systems, and a 2.5 dB gain over Gaussian Tree Approximation in 32×32 with 16-QAM. We run it in two forms. One uses the plug-in channel estimate. The other uses perfect channel knowledge, which serves as an oracle ceiling.
Channel-tracking method. The decision-feedback tracker RC-AttStructNet-DF (Detect to Learn). The file reports only that it is roughly 100 times faster than MMNet and has lower error rates than LMMSE and MMNet. No reported BER value was found on Lacuna for this method under QuaDRiGa at 30 km/h, so we will rerun it ourselves.
Ablations. Three ablations isolate the key ingredient, all at an identical per-slot iteration budget:
- warm-start nothing;
- warm-start everything, including symbol beliefs and variance schedule;
- warm-start the channel side only (the proposal).
A fourth ablation drops the joint channel inference but keeps the warm-started linear solver. This separates the gain from treating the channel as uncertain from the gain from cheap amortized solves.
Fairness. All methods see the same channel realizations, pilots, and LDPC codes. Every method's hyperparameters get the same tuning budget on a held-out set of channel seeds.
Why not just…?
Why not just use LMMSE with a polynomial-expansion inverse and re-estimate the channel each block? Polynomial expansion relies on channel hardening, and the ASIC paper reports that hardening fails in correlated mmWave channels. There, LMMSE trails coordinate descent by 4 to 10 dB under high load. Our naive baseline measures this gap directly on the same QuaDRiGa scenarios.
Why not just warm-start the whole detector from the previous vector, the way receding-horizon controllers shift their plans? Consecutive symbol vectors are independent, so a shifted symbol estimate is a biased prior, not a nearby solution. Carrying converged variances forward would also bypass the variance annealing that EC needs to avoid bad minima. The "warm-start everything" ablation tests this. We predict it will do worse than cold-starting at moderate-to-high SNR. If it does not, the partition principle is wrong and we will say so.
Why not just warm-start the linear solver across slots and keep treating the channel estimate as exact? This is the most serious threat. It is the obvious cross-slot extension of warm-started message passing, and it may capture most of the complexity savings. The "warm-started solver, no joint channel inference" ablation settles it. If that variant matches the full proposal in coded error rate at 30 km/h, the joint hybrid inference adds nothing, and the contribution shrinks to an engineering note. The proposal is worth pursuing only if channel-uncertainty-aware inference closes a measurable part of the gap to the perfect-channel oracle under mobility.
Why not just run exact hybrid inference with WMI solvers? Given the symbols, the integral over a Gaussian channel and noise has a closed form: the observation is Gaussian with a symbol-dependent covariance. So there is no polytope integral that needs an SMT solver, and the hardness is purely the discrete sum over $|\mathcal{X}|^K$ symbol vectors. We use exact enumeration of this closed form as the ground-truth posterior for small systems, not SMT-based WMI.
Why not just train an end-to-end neural receiver? The file's evidence points against this for our target. Detect to Learn reports being roughly 100 times faster than MMNet, and the ASIC paper argues that unstructured networks do not fit base-station area and power budgets. We include the decision-feedback tracker as the learned competitor rather than a larger unstructured network.
Proposed Experiments
Main benchmark. QuaDRiGa mmMAGIC urban microcell at 60 GHz with 128 receive antennas and 16 users, matching the ASIC paper. User speeds are swept from static to 30 km/h, the latter matching Detect to Learn, with 16-QAM to 256-QAM and LDPC coding. The primary metric is the SNR needed to reach a target coded block error rate, reported at matched per-vector multiply-accumulate counts. We also report mutual information per dimension. Complexity metrics are inner and outer iterations per slot and their dependence on Doppler and on the measured condition number.
Exact-reference benchmark. Small i.i.d. Rayleigh and correlated systems (4×4 with 16-QAM, 5×5 with QPSK, matching the EC paper's small-system regime). Here the exact joint posterior under channel uncertainty is computed by closed-form enumeration. This lets us measure each detector's gap to the true posterior's bit log-likelihood ratios, not just its error rate.
Decisive checks. There are four, each able to falsify part of the claim:
- Cost scaling. Warm-started per-slot inner iterations should fall roughly logarithmically as speed decreases, consistent with the conjugate-gradient bound. They should grow with line-of-sight correlation only through $\sqrt{\kappa}$.
- Oracle gap. Channel-side-only warm-starting should recover a substantial share of the gap between plug-in EC and perfect-channel EC at 30 km/h, at a cost within a small constant factor of plug-in EC.
- Partition test. The warm-start-everything ablation should underperform the channel-only rule.
- Long-run stability. Over tens of thousands of slots, the channel-belief error should stay bounded under the robust variance estimator and diverge under the textbook one, reproducing the instability reported for warm-started message passing.
If the first two checks fail, the idea does not survive.
References
- Distributed Formation Control for Multi-Vehicle Systems With Splitting and Merging Capability
- Real-time Iteration Scheme for Diffusion Policy
- Warm Start of Mixed-Integer Programs for Model Predictive Control of Hybrid Systems
- Warm-Starting in Message Passing algorithms
- Probabilistic MIMO Symbol Detection With Expectation Consistency Approximate Inference
- Detect to Learn: Structure Learning With Attention and Decision Feedback for MIMO-OFDM Receive Processing
- Closed-Loop Analysis of ADMM-Based Suboptimal Linear Model Predictive Control
- A Deep-Unfolding-Optimized Coordinate-Descent Data-Detector ASIC for mmWave Massive MIMO
- Probabilistic Inference in Hybrid Domains
- SMT-based Weighted Model Integration with Structure Awareness