Two-Basis Beamspace Denoising for Impulsive Noise in Massive MIMO

Large antenna arrays at millimeter-wave frequencies receive energy along only a few propagation paths. After a spatial Fourier transform into "beamspace" (the domain of arrival angles), the channel concentrates in a handful of angular bins, and the fastest good estimators simply zero out the weak bins. The best of these, BEACHES, picks its threshold without ground truth by using Stein's unbiased risk estimate (SURE), a formula that estimates mean squared error from the noisy data alone but is only valid for Gaussian noise. Real arrays also see impulsive disturbances: a hardware glitch on one element, a saturating analog-to-digital converter, or a burst of interference that hits a few antennas. Our observation is that such an impulse is sparse in the antenna domain but perfectly flat in beamspace. A spike on one antenna spreads with equal magnitude over every angular bin. Thresholding in beamspace alone is therefore the wrong tool. It either raises the threshold and erases weak paths, or it leaves the spike's floor in every bin.

We propose to estimate the channel and the impulses jointly, as a signal sparse in beamspace plus a signal sparse across antennas. The solver uses only FFTs and two soft thresholds per iteration. Both thresholds are tuned by a SURE-type risk estimate that remains valid under the model "Gaussian noise plus sparse, arbitrarily large impulses." Heavy-tailed modeling therefore enters as a physical structural prior on where the outliers live, rather than as a replacement likelihood such as Student-$t$. The goal is an estimator that keeps the $O(B \log B)$ cost of BEACHES, where $B$ is the number of antennas, and that does not collapse when a few antennas misbehave.

Core Claim

Model one pilot snapshot at a $B$-antenna array as

$$y = F^{H}x + s + w,$$

where $F$ is the unitary $B$-point spatial DFT, $x$ is the beamspace channel (few large entries), $s$ is the antenna-domain impulse vector (few nonzero entries of unbounded size), and $w$ is complex Gaussian noise with variance $\sigma^2$. We estimate both components by

$$(\hat x,\hat s)=\arg\min_{x,s}\ \tfrac12\|y-F^{H}x-s\|_2^2+\tau\|x\|_1+\lambda\|s\|_1 .$$

The antenna-domain identity basis and the DFT basis are maximally incoherent, with mutual coherence $1/\sqrt{B}$. Classical uncertainty-principle arguments therefore give a unique split roughly when the number of paths plus the number of impulses stays below order $\sqrt{B}$, which is about 16 for the 256-antenna array we test.

Minimizing over $s$ in closed form turns the objective into a beamspace lasso with a Huber loss on antenna residuals. The estimator itself is thus a known robust regression. The claim is not that the estimator is new. The new parts are three. First, the physical argument that its two-basis structure matches impulsive corruption in arrays, where beamspace-only denoising does not. Second, a closed-form, ground-truth-free rule for choosing $(\tau,\lambda)$. Third, evidence that the resulting method beats every fast alternative while staying within a small constant factor of BEACHES in runtime.

The tuning rule treats the impulse as part of the mean. It then applies SURE to the combined mean $F^{H}x+s$, using the known result that the degrees of freedom of a lasso equal the rank of its active columns. Here those columns come from the concatenated dictionary $[F^{H}, I]$:

$$\widehat{R}(\tau,\lambda)=\|y-F^{H}\hat x-\hat s\|_2^2+2\sigma^2\,\mathrm{rank}\big([F^{H}_{\mathcal A_x},\,I_{\mathcal A_s}]\big)-B\sigma^2 .$$

Here $\mathcal A_x$ and $\mathcal A_s$ are the active beams and antennas, and $\sigma^2$ is estimated robustly from the median of the residual. This risk is unbiased for the combined mean, not for the channel alone. Whether minimizing it also minimizes channel error is the main empirical question.

Relation to Prior Work

Two papers motivate the gap. BEACHES is FFT-plus-SURE soft thresholding in beamspace. We keep its speed and its unsupervised tuning philosophy, but change the signal model so that the risk estimate stays valid when a few antennas carry impulses. Robust Adaptive Sparse Channel Estimation in the Presence of Impulsive Noises combines sign-error LMS updates with sparsity penalties on a 128-tap time-domain FIR channel under $\alpha$-stable noise. It shows that impulsive noise costs real accuracy, but it uses a single domain, hand-tuned step sizes, and no array structure. We work in the spatial domain with two bases and tune from a risk estimate.

One paper supplies the likelihood idea. LL-BCS pairs a Laplacian likelihood with a Laplacian prior for radar time delays and solves it by EM (expectation–maximization). Our antenna-domain $\ell_1$ term is the maximum a posteriori counterpart of that likelihood. However, we replace the iterative EM, which reaches only a local optimum, with a convex problem solved by a fixed number of FFT passes. We also add array geometry.

Two robust-statistics papers constrain the design. Outlier-Robust Sparse Mean Estimation needs bounded fourth moments for its truncation step and a convex relaxation whose cost is far beyond per-slot budgets. Dictionary-Sparse Recovery From Heavy-Tailed Measurements handles a heavy-tailed sensing matrix, not heavy-tailed additive noise, and notes that its guarantees may not carry over to structured matrices. Our approach sidesteps moment assumptions on impulses entirely: impulses are estimated, not truncated. Our guarantees come from basis incoherence rather than random sensing.

Finally, VAE-leveraged MMSE and the CeBed deep estimators assume conditionally Gaussian noise. CeBed reports that deep models can fall below least squares out of distribution. That result motivates keeping a model-based estimator with an explicit outlier term.

Research Program

The first step is analysis. We will characterize how each noise model looks in beamspace and where BEACHES fails. We will quantify how a single antenna impulse of amplitude $A$ raises every bin by $A/\sqrt{B}$, and derive the path-power level below which Gaussian-SURE thresholding erases a path. We will then extend the identifiability argument to the realistic case where off-grid paths leak into neighboring bins. Leakage makes $x$ only approximately sparse and erodes the effective incoherence. We expect this to be the binding limit, and we will measure it rather than assume it away.

The second step is the solver. It is an accelerated proximal iteration on the objective, with a fixed iteration count $T$. Each iteration costs one FFT and one inverse FFT plus two soft thresholds, for a total of $O(TB\log B)$. We will also evaluate a warm-started path over a small grid of $(\tau,\lambda)$, so that the risk-based search reuses solutions. The target is a runtime within 10x of BEACHES. That is still two orders of magnitude inside the 1,500x gap BEACHES reports against NOMP (Newtonized orthogonal matching pursuit).

The third step is tuning. We implement $\widehat R$ with a robust noise-variance estimate. We then study its regret relative to oracle tuning, which chooses parameters using the true channel.

The fourth step is misspecification. Symmetric $\alpha$-stable noise is not "Gaussian plus sparse": every antenna is heavy-tailed. We will test whether the two-basis estimator degrades gracefully there, falling back toward a Huber-type beamspace estimator as $\lambda$ shrinks. We will also test whether $\widehat R$ still picks reasonable thresholds.

Baselines

The naive reference is least squares, meaning the raw antenna observation used directly as the channel estimate. It fixes the floor and matches the least-squares and maximum-likelihood references used in BEACHES.

The strongest fast competitor is BEACHES run unchanged. On clean QuaDRiGa mmMAGIC channels it reports a "2–3 dB signal-to-noise ratio (SNR) improvement over standard Maximum Likelihood (ML) estimation" and speedups "up to 1,500x faster than NOMP and up to 58,000x faster than ANM in nLoS scenarios." We found no reported result on Lacuna for BEACHES under impulsive noise, so we will produce that number ourselves on identical channel draws. We will use the same 256-antenna, 16-user, 60 GHz line-of-sight and non-line-of-sight setup.

The strongest accurate competitor is atomic norm minimization (ANM, a continuous-angle convex method in ANM for hybrid precoding) with an $\ell_1$ data-fit term replacing the $\ell_2$ one. This is the Laplacian-likelihood idea of LL-BCS carried to arrays. We give it oracle-tuned regularization. It serves as an accuracy ceiling at far higher cost.

For the time-domain setting, the comparison is the reweighted sparse sign-LMS variants from the impulsive-noise FIR paper. They report roughly "3dB in Mean Square Error (MSE) over the standard SLMS" at 10 dB SNR, $\alpha=1.2$, and 8 active taps out of 128. We run a batch version of our estimator on the same number of training samples.

The ablation that isolates the key ingredient is the same solver with $\lambda=\infty$, meaning no impulse component. This is plain beamspace lasso with our SURE and should reproduce BEACHES. A second ablation keeps the two-basis model but tunes $(\tau,\lambda)$ with a fixed grid chosen on a validation set, which isolates the value of the risk estimate.

Beating the fast baselines means lower channel NMSE (normalized mean squared error) and lower uncoded bit error rate at matched runtime whenever impulses are present. It also means staying within 0.3 dB of BEACHES when they are absent.

Why not just…?

Why not just estimate the noise level robustly and run BEACHES? A median-based noise estimate removes the grossest inflation. It does not remove the impulse itself, which still lifts every beamspace bin by $A/\sqrt{B}$. The threshold must then either sit above that floor and delete weak paths, or sit below it and admit spurious ones. This is a real competitor when impulses are small relative to $\sigma$. We include "BEACHES with median noise estimate" as a baseline and plot the regime where it suffices.

Why not just clip or median-filter the antennas first, then run BEACHES? This is the most serious threat, and it is the natural engineering answer. Clipping by amplitude cannot distinguish an impulse from a strong line-of-sight path, which also produces large antenna samples. Outlier-Robust Sparse Mean Estimation also shows that ad hoc truncation of heavy-tailed data distorts the remaining distribution unless moment conditions hold. Still, a well-chosen detect-and-erase step, using the residual after one BEACHES pass, may capture most of our gain. We implement it with an oracle-tuned clip level. If it matches the joint estimator, we will report that the contribution reduces to the tuning rule.

Why not just use a Laplacian-likelihood Bayesian estimator or robust ANM? These are accurate. However, LL-BCS relies on EM iterations that reach only local optima, and ANM solves semidefinite programs that BEACHES measured at up to 58,000x slower. We report the accuracy gap to robust ANM honestly. The case for our method rests on closing most of that gap at FFT cost.

Why not just train a deep estimator on impulsive data? CeBed shows that deep estimators "saturate and can even degrade below LS at high SNRs when tested on unseen SNR ranges." We test a residual-CNN estimator trained on one impulse rate and evaluated at others. The prediction is that our model-based method holds up better under unseen impulse rates.

Proposed Experiments

The main benchmark is the QuaDRiGa mmMAGIC urban micro-cell configuration from BEACHES: 256 antennas, 16 users, 60 GHz, with line-of-sight and non-line-of-sight propagation. We add three noise families on top of Gaussian noise across an SNR sweep. The first is Bernoulli–Gaussian impulses with per-antenna probability $p$ from 0 to 5% and impulse-to-noise ratio from 10 to 40 dB. The second is persistent faulty elements repeated across snapshots. The third is symmetric $\alpha$-stable noise with $\alpha\in\{1.2,1.5,1.8,2\}$. Metrics are channel NMSE, uncoded bit error rate with the equalizer used in BEACHES, measured runtime, and the NMSE regret of $\widehat R$ relative to oracle tuning. A secondary check reruns the 128-tap, $\alpha=1.2$ FIR setup of the sign-LMS paper.

Four outcomes are decisive, and we would treat each as binding. First, with $p=0$ our method must match BEACHES; otherwise the model costs too much on clean data. Second, as impulses are added, the gap over BEACHES should grow with impulse-to-noise ratio and collapse once paths plus impulses exceed roughly $\sqrt{B}$. This would confirm the incoherence mechanism rather than generic robustness. Third, the gap over oracle-tuned clip-then-BEACHES must be positive, especially in line-of-sight scenarios where strong paths confuse clipping; if it is not, we narrow the claim to tuning alone. Fourth, oracle regret should stay small even though $\widehat R$ targets the combined mean rather than the channel. If regret is large under $\alpha$-stable noise, that marks where the sparse-impulse model stops being the right description of heavy tails.