Spiking neural networks, which compute with binary pulses instead of continuous numbers, promise large energy savings for wireless receivers because each incoming pulse triggers an addition rather than a multiplication. The catch is that every spiking design choice is currently tuned by trial and error. Designers must pick how long to count spikes, how to scale inputs so neurons do not saturate, and how many settling steps to allow, and they redo this search for each antenna configuration. This proposal argues that the mathematics used to analyze massive multi-antenna (MIMO) systems can remove that guesswork. That mathematics, called large-system analysis, studies what happens as the number of antennas and users grows together at a fixed ratio, and finds that random quantities settle to predictable values.
The central observation is simple. A rate-coded spiking neuron can only represent values inside a bounded range, so a spiking linear solver does not compute the ordinary regularized least-squares answer. It computes a box-constrained version of it and then adds counting noise that shrinks with the observation window. Box-constrained estimators in high dimensions are exactly what the Convex Gaussian Min-Max Theorem (a tool that turns a hard high-dimensional optimization into a small scalar problem) has recently been shown to handle. Pairing the two gives closed-form, load-dependent predictions of error rate, spike energy, and settling latency for a spiking MIMO detector before anything is simulated.
Core Claim
Consider uplink detection with $y = Hx + n$, where $H$ is an $N \times K$ channel matrix ($N$ receive antennas, $K$ users, load $\gamma = N/K$), $x$ holds BPSK symbols, and $n$ is Gaussian noise. A recurrent spiking solver of the kind Computing Generalized Matrix Inverse on Spiking Neural Substrate deployed on IBM TrueNorth represents each unknown by a neuron firing rate in a bounded interval. After inputs are scaled by a factor $\eta$, it therefore settles to
$$ \hat{x}_{\text{box}} = \arg\min_{|x_i| \le 1/\eta} \; \|y - Hx\|^2 + \lambda \|x\|^2 , $$
where $\lambda$ is the regularization weight. The value read out after $T$ time ticks is $\hat{x}_{\text{box}} + e$, where the counting error $e$ has per-entry variance of order $1/(\eta^2 T)$, the variance of a Bernoulli rate estimate. We claim that as $N, K \to \infty$ at fixed $\gamma$, each user's output decouples into a scalar channel: a clipped Gaussian observation of $x_i$ with effective noise level $\sigma_\star(\gamma, \mathrm{SNR}, \lambda, \eta)$ given by Min-Max fixed-point equations, plus independent readout noise. From this we obtain three deterministic equivalents. The first is the bit error rate. The second is the mean firing rate, which fixes the energy per detected symbol:
$$ E_{\text{sym}} \approx E_{\text{AC}} \cdot K \cdot T \cdot \sum_{i=1}^{K} \bar r_i , $$
where $E_{\text{AC}}$ is the energy of one accumulate, $\bar r_i$ is user $i$'s asymptotic firing rate, and each spike fans out to $K$ recurrent synapses. The third is the settling time, which is governed by the condition number of $H^\top H + \lambda I$ and follows from the Marchenko–Pastur spectrum edges $(1 \pm \gamma^{-1/2})^2$.
The new element is that $\eta$ acts as a noise-shaping knob. A small $\eta$ widens the representable range and avoids clipping, but it shrinks rates and amplifies counting noise. A large $\eta$ does the reverse. This is the spiking counterpart of the amplitude limit $A$ that Asymptotic Analysis of One-Bit Quantized Box-Constrained Precoding showed can be optimized analytically for one-bit precoders. The optimal $\eta^\star(\gamma, T)$ becomes a formula rather than a per-matrix computation.
Relation to Prior Work
Two papers motivate the gap. Neuromorphic In-Context Learning for Energy-Efficient MIMO Symbol Detection builds a spiking transformer detector and reports 5.4x to 26.8x power reduction, but it evaluates only finite synthetic configurations and gives no theory for how its encoding noise scales with system size. We trade its learned, adaptive architecture for an analyzable linear solver. Shukla et al. supply the solver and the idea of the scaling factor. Their error bounds, however, come from deterministic matrix norms of a fixed matrix, so $\eta$ must be recomputed for every matrix. We replace those bounds with random-matrix limits.
Three papers supply the method. Ma et al. develop the Min-Max analysis for box constraints and, importantly, show that Bussgang linearization (treating distortion as uncorrelated Gaussian noise) fails there. This negative result is what forces our choice of tool. Detection of Spatially Modulated Signals via RLS gives large-system bounds for regularized least-squares detection, and Microscopic analysis for decoupling principle establishes per-user scalar decoupling via the replica method. We extend both to a readout that is bounded and noisy.
Linear One-Bit Precoding in Massive MIMO derives a closed-form optimal regularization for one-bit precoders. It treats baseband arithmetic as exact, whereas here the arithmetic itself is the source of distortion.
Two constraints limit the scope. The Min-Max tools in the file cover BPSK and i.i.d. Gaussian channels, so our theory starts there. Correlated channels are treated only empirically, as a stress test.
Research Program
The work has three parts.
The first is theory. We write the spiking solver's fixed point as the box-constrained problem above, apply the Min-Max theorem to obtain scalar fixed-point equations, and derive the three deterministic equivalents. From these we derive $\eta^\star$ and the minimal window $T$ that reaches a target error rate at each load $\gamma$. We also check whether the leaky integrate-and-fire reset changes the fixed point relative to ideal rate coding. If it does, the analysis must be restricted to integrate-and-fire neurons, and we will report that.
The second is a simulator. It is a bit-accurate integrate-and-fire implementation of the recurrent solver with low-bit integer weights, following TrueNorth-style constraints. It counts every synaptic event, so energy is measured by event counts rather than inferred from the formula.
The third is an energy comparison. We account digital baselines by multiply-accumulate counts and spiking designs by accumulate counts, both at the 45nm figures reported in SpikingBERT: 0.9 pJ per accumulate and 4.6 pJ per 32-bit multiply-accumulate. We also report the full operation counts so that readers can substitute other technology figures.
Baselines
The naive reference is ideal floating-point regularized least-squares (MMSE) detection with no hardware constraints. It fixes the best achievable error rate of the linear family and the multiply-accumulate energy floor. It is computed on the same channel draws, and its known deterministic equivalent doubles as a sanity check.
The strongest energy competitor is low-precision digital detection, specifically INT8 and INT4 fixed-point MMSE and truncated-polynomial-expansion detection following Power efficient low complexity precoding for massive MIMO systems. These avoid matrix inversion on conventional hardware and are what a practitioner would actually deploy. We give them the same channel realizations, and we tune the polynomial degree and bit width over a search budget matched to the budget spent tuning $\eta$ and $T$. The spiking design beats this baseline only if, at equal error rate (within 0.25 dB SNR), its counted energy per symbol is lower. The file contains no reported energy result for quantized digital MIMO detectors, so this baseline is run from scratch.
The strongest accuracy competitors are the learned detectors. MMNet (Adaptive Neural Signal Detection for Massive MIMO) reports being 4–8 dB better than MMSE and within 1.5 dB of maximum-likelihood detection on 64x16 QuaDRiGa channels. The spiking transformer of Song et al. reports 5.4x–26.8x power reduction versus its ANN counterpart, but the file contains no absolute error-rate number for it. We retrain both at our configurations with equal training budgets. We do not expect to beat MMNet in accuracy. The question is where on the energy–accuracy frontier each method sits.
The key ablation is a heuristic spiking solver. It uses the same solver, but $\eta$ is chosen by Shukla et al.'s per-matrix norm rule and $T$ by grid search. This isolates the value of the asymptotic design law. A second ablation keeps the spiking solver but replaces the Min-Max prediction with a Bussgang-style prediction, isolating the value of the analytical tool.
Why not just…?
Why not just run truncated polynomial expansion or INT8 MMSE on ordinary hardware? This is the real threat. Rate coding needs $T$ ticks, and each spike fans out to $K$ accumulates. If high precision forces a large $T$, the 5x per-operation advantage of accumulates can vanish entirely. The file offers no evidence that settles this for MIMO detection. Shukla et al.'s 10x gain over FPGAs came on small static problems. The energy deterministic equivalent predicts the crossover load and target error rate at which the spiking solver wins or loses, and the counted-energy experiment verifies it. If spiking never wins at a useful error rate, that negative result is itself the deliverable.
Why not just grid-search $\eta$ and $T$ in simulation? For a single configuration, a grid search will likely match the theory's choice. The theory earns its keep by predicting $\eta^\star$ and $T$ for unseen loads and system sizes, by explaining how settling time scales, and by avoiding per-matrix norm computation at runtime. We test exactly this. We tune by grid at small $N$, then compare grid-transferred versus theory-predicted settings at $N = 512$ and $1024$.
Why not just treat spike quantization with Bussgang linearization? This approach is valid when neurons rarely saturate, and we expect it to agree with our predictions when $\eta$ is small. Ma et al. show that it breaks down under box constraints, which is precisely the saturated regime where $\eta^\star$ lives. The Bussgang ablation will show whether its error-rate predictions diverge near the optimum. If they do not diverge, we will say that the simpler tool suffices.
Why not just use the spiking transformer, which already adapts to channels? It adapts, but it has no predictive theory, and its cost grows with model size. Our design gives up adaptivity in exchange for guarantees. The correlated-channel experiment shows how much accuracy that trade costs.
Proposed Experiments
Theory validation uses i.i.d. Rayleigh BPSK channels with $K$ from 16 to 512, $\gamma \in \{1.5, 2, 4, 8\}$, SNR from 0 to 20 dB, and $T$ from 4 to 256. The validation scale matches the up-to-1000-antenna Monte Carlo checks of Ma et al. The metrics are the gap between predicted and simulated bit error rate, the mean firing rate, and settling ticks. The decisive check is agreement within 0.2 dB at $N \ge 256$. Agreement should tighten with $N$ at the usual fluctuation rate of order $1/\sqrt{N}$.
The second experiment compares the theory-predicted $\eta^\star$ with the grid optimum across loads. The ablation's error-rate penalty from using the per-matrix norm rule is the measured value of the design law.
The third experiment traces energy–accuracy frontiers against all baselines. It reports the load and SNR regions where counted spiking energy falls below the INT8 and polynomial-expansion baselines at equal error rate.
The fourth experiment is a stress test on QuaDRiGa urban macrocell channels in the 64x16 setting used by MMNet. Here the i.i.d. theory is expected to lose accuracy. We measure how far, and whether substituting the empirical Gram spectrum for Marchenko–Pastur restores the prediction of $\eta^\star$.
The project fails cleanly on either of two outcomes. The first is that the theory misses simulation by more than 1 dB at large $N$ on i.i.d. channels. The second is that no operating point exists where spiking energy beats INT8 digital detection at an error rate within 1 dB of MMSE.
References
- Computing Generalized Matrix Inverse on Spiking Neural Substrate
- Asymptotic Analysis of One-bit Quantized Box-Constrained Precoding in Large-Scale Multi-User Systems
- Neuromorphic In-Context Learning for Energy-Efficient MIMO Symbol Detection
- Detection of Spatially Modulated Signals via RLS: Theoretical Bounds and Applications
- Microscopic analysis for decoupling principle of linear vector channel
- Linear One-Bit Precoding in Massive MIMO: Asymptotic SEP Analysis and Optimization
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- Power efficient low complexity precoding for massive MIMO systems
- Adaptive Neural Signal Detection for Massive MIMO