Scanning the Array: Bidirectional State-Space Precoders for 1-Bit Massive MIMO

Massive MIMO base stations drive hundreds of antennas at once. MIMO (multiple-input multiple-output) means many antennas transmit together so that each user receives a clean signal. A cheap way to build such arrays is to give every antenna a 1-bit digital-to-analog converter, so each antenna can only emit a sign. The catch is that the precoder, the algorithm that decides what each antenna sends, must then choose a vector of signs. Good nonlinear precoders exist, but they are iterative optimizers solved afresh for every transmitted symbol vector. Cheap linear precoders lose badly: on a 128-antenna, 16-user benchmark, linear quantized precoding pays an 8 dB penalty against an ideal unquantized system at a bit error rate of $10^{-3}$, while the iterative nonlinear precoder SQUID sits only 3 dB away (Quantized Precoding for Massive MU-MIMO).

We propose to close that gap with a single learned pass whose cost grows linearly in the number of antennas. The pass is a bidirectional state-space model: a recurrent network with linear-time cost, like Mamba, that reads a sequence both left-to-right and right-to-left. Here it reads along the antenna array. The physical intuition is error feedback. When one antenna's sign is decided, the quantization error it creates is carried in the model's hidden state and compensated by its neighbours, in both directions along the aperture, so that the error lands where users do not hear it. We add two further pieces. Boundary-anchored position features, borrowed from position-specific n-grams in morphology, tell the model where the array ends. The Bussgang decomposition serves as both the starting point and the measuring instrument for the distortion the network produces.

Core Claim

A 1-bit precoder is at heart a spatial error-shaping problem. When the channel has structure along the array, which is the case for line-of-sight and ray-traced channels on uniform arrays, a bidirectional recurrent scan over antenna index should match SQUID's error rate at a fraction of its latency. A permutation-equivariant network of the same size should not, because it cannot exploit neighbour structure. The claim is conditional, and the condition can be tested. On spatially white (i.i.d. Rayleigh) channels, antenna order carries no information, and the scan should give no advantage over an order-agnostic model. That prediction is what separates a real mechanism from a lucky architecture.

The Bussgang decomposition writes a nonlinearity $y = g(x)$ as

$$y = \alpha x + e,$$

where $\alpha$ is a deterministic gain and $e$ is a distortion term uncorrelated with the input $x$. We use it in two ways. First, the network predicts a correction on top of the Bussgang-linearized zero-forcing precoder rather than learning from scratch. Second, we use it to measure the result: for the received signal $r = H\,\mathrm{sign}(x) + n$, with channel $H$ and noise $n$, we estimate the Bussgang gain and the distortion covariance $C_e$ empirically. We then report how much distortion energy the learned precoder leaves inside the users' channel subspace, $\mathrm{tr}(H C_e H^{H})$, compared with the energy it pushes outside it.

Relation to Prior Work

Two strands of communications work frame the problem. Quantized Precoding for Massive MU-MIMO uses Bussgang to analyse linear precoders under low-resolution converters and introduces SQUID, an iterative splitting method; it supplies our main baseline and benchmark. Asymptotic Analysis of One-Bit Quantized Box-Constrained Precoding in Large-Scale Multi-User Systems proves that Bussgang predictions diverge from reality once amplitude constraints make the pre-quantization signal non-Gaussian. This result constrains us. We do not use Bussgang to predict our precoder's rate. We use it only as an empirical decomposition, and we check directly how far its predicted rate departs from simulated rate for our outputs. That work also restricts itself to BPSK modulation with perfect channel knowledge, which motivates our QAM and noisy-channel-estimate conditions.

Learned beamformers so far rely on graph networks or autoencoders. GPASS uses an edge-update graph network for pinching-antenna placement and finds that it needs explicit antenna indices as inputs to respect element ordering. MIMO-PASS shows that its optimization objective is permutation-invariant to pinch order. These two findings pull in opposite directions, and our white-versus-structured-channel test is designed to resolve the tension for fixed arrays. Order should matter exactly when the channel carries spatial correlation. Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback finds that standard z-score normalization destroys physical scale in wireless learning. It constrains our design: we feed unnormalized channel magnitudes and tie the output scale to the Bussgang gain. Linear-complexity beamforming work such as Soft Max-Min for holographic surfaces (Holographic Multi-User Multi-Stream Beamforming Maintaining Rate-Fairness) motivates the linear-cost target. Those methods assume ideal converters; we assume 1-bit ones.

The architecture comes from sequence modeling. Caduceus introduces BiMamba, which shares weights between forward and backward scans, and builds reverse-complement equivariance into DNA models. We reuse the weight-shared bidirectional scan, and we impose the array's mirror symmetry the same way Caduceus imposes strand symmetry: reversing the antenna order and conjugating appropriately should map a solution to a solution. Bi-Mamba: Towards Accurate 1-Bit State Space Models binarizes weights with straight-through gradient estimators. We borrow the straight-through trick, but we apply it to the output signs that the hardware emits, not to the model's weights.

Morphology supplies the boundary features. Dealing with the Grey Sheep of the Romanian Gender System and Automatic Discrimination between Cognates and Borrowings show that marking word ends lets a model treat an identical substring differently as a stem or as a suffix. The array analogue is that edge antennas of a finite aperture see truncated neighbourhoods and must absorb error that cannot be passed further along. We therefore give each antenna two features: its distance to the first element and its distance to the last element, plus explicit edge tokens. This is a small, cheap ingredient, and an ablation isolates it.

Research Program

The model takes the channel $H$ and the symbol vector $s$ for one channel use. It forms per-antenna tokens from the column $H_{:,i}$, the Bussgang-ZF pre-quantization value at antenna $i$, and the boundary features. A stack of weight-shared bidirectional selective state-space layers then outputs a soft value at each antenna. For two-dimensional arrays, the scan follows a raster ordering with multiple orientations. Cost is $O(Nd)$ per layer for $N$ antennas and hidden width $d$, after the $O(NK)$ cost of forming features for $K$ users. The output is passed through the sign function with a straight-through gradient. Training minimizes the mean squared error between $\beta H\,\mathrm{sign}(x)$ and $s$, where $\beta$ is the Bussgang-style receive scaling. It also includes a symbol-error surrogate for QPSK and 16-QAM.

We train across antenna counts and user loads, and we test generalization to unseen $N$. Analysis has three parts. The first is the in-subspace versus out-of-subspace distortion split described above. The second is the gap between Bussgang-predicted and simulated rate. The third probes whether hidden states track accumulated local error, the behaviour expected of a learned spatial error-feedback loop.

Baselines

The naive reference is quantized zero-forcing analysed with Bussgang: compute the unquantized zero-forcing precoder and take signs. Quantized Precoding for Massive MU-MIMO reports this at an 8 dB penalty from ideal at BER $10^{-3}$ with 128 antennas and 16 users. Beating it is necessary but proves little.

The strongest existing method is SQUID from the same paper, reported "only 3 dB away" from the infinite-resolution baseline at BER $10^{-3}$ in that setting. We will run the authors' algorithm on identical channel draws. Its iteration count will be swept, so that we can plot a full error-rate-versus-wall-clock curve on the same GPU and CPU. We will not compare against a single operating point. Beating SQUID means reaching its BER at $10^{-3}$ within 0.5 dB at substantially lower measured latency, or doing better than it at equal latency. Matching it slowly counts as a loss.

For structured channels we also report SQUID on DeepMIMO data. The file gives no reported result on Lacuna for SQUID on DeepMIMO channels, so we generate our own.

The architectural competitors receive the same training data, parameter budget, and hyperparameter search trials. The first is a permutation-equivariant set or graph network over antennas in the spirit of GPASS. The second is a bidirectional transformer over antennas, which is quadratic in cost. The third is a causal single-direction state-space model. We also run a hand-designed, non-learned spatial error-feedback quantizer that passes each antenna's quantization error to the next with fixed weights. It asks whether learning adds anything beyond a classical recursion.

Three ablations isolate our ingredients. The first removes the boundary features. The second removes the Bussgang-ZF input, so the model learns from scratch. The third shuffles antenna order at train and test time. That last one is the key isolating ablation: if performance survives shuffling on structured channels, the scan is not doing what we claim.

Why not just…?

Why not just run SQUID with fewer iterations? This is the real threat, since SQUID is already designed to be cheap. The error-rate-versus-latency sweep settles it directly. If truncated SQUID dominates our curve everywhere, the proposal fails on its main claim, and we will report that. Our bet is that a learned pass amortizes the per-symbol optimization, and amortization is where learned solvers usually win.

Why not just use a graph network or transformer over antennas? The graph network ignores order unless indices are injected, as GPASS needed. The transformer pays quadratic cost in $N$, which grows at exactly the array sizes where 1-bit hardware pays off. Both are run at matched size. The decisive evidence is the pattern across channels: our model should beat the equivariant one on structured channels and tie it on i.i.d. channels.

Why not just buy more converter bits? Quantized Precoding for Massive MU-MIMO identifies 3–4 bits as a practical sweet spot. That is a genuine alternative for many deployments, and we do not dispute it. We will run our method at 2 and 3 bits too, where the straight-through quantizer generalizes. If the gain over quantized zero-forcing vanishes by 2 bits, the contribution is confined to the extreme low-power regime, and we will scope the claim accordingly.

Why not just use classical spatial error feedback? If a fixed error-feedback recursion captures most of the gain, the learned model is unnecessary. It is included as a baseline precisely so that this outcome is visible.

Proposed Experiments

The first benchmark is the narrowband 128-antenna, 16-user setting of Quantized Precoding for Massive MU-MIMO. We run it under i.i.d. Rayleigh fading, correlated channels, and line-of-sight uniform-linear-array channels, using QPSK and 16-QAM. The second uses ray-traced channels from DeepMIMO scenario O1, as in Normalization-Free Knowledge Distillation, which reports an 8×8 planar array configuration. We use that configuration and also a linear-array variant built from the same scenario.

We further test with noisy least-squares channel estimates, because both the box-constrained analysis and the beamforming literature flag perfect channel knowledge as unrealistic. Metrics are BER against SNR, the SNR gap to ideal at BER $10^{-3}$, and measured latency per symbol vector. We also report the Bussgang in-subspace distortion fraction and the gap between Bussgang-predicted and simulated rate.

Two decisive checks frame the result. The first is whether the bidirectional scan beats the equal-size permutation-equivariant model on structured channels while tying it on i.i.d. channels, and loses its advantage when antenna order is shuffled. The second is whether our curve sits below SQUID's on the latency-versus-BER plane. The first check establishes that spatial error shaping along the array is the mechanism. The second establishes that it is worth deploying.