Sign-Flip Calibrated Activity Detection Across Decentralized Antenna Clusters

In grant-free massive access, thousands of sensors share a base station. Each sensor holds a non-orthogonal pilot code, only a small fraction wake up at any moment, and the receiver must decide which of them transmitted. Two features of the receiver are now hard to avoid. First, the antenna array is split into clusters, each with its own processor, because shipping raw antenna samples to one central chip exceeds interconnect limits. Second, the decision of whether a device is "active" is made by thresholding a score. Today that threshold is tuned by hand or justified by large-array, Gaussian-noise theory. When the noise is bursty or impulsive, or when a jammer is present, the false-alarm rate it actually delivers can be far from the one intended.

We propose a randomization test that gives each candidate device a p-value with finite-sample false-alarm control. The test relies only on the noise being symmetric in time: noise and interference on each pilot symbol are equally likely to appear with flipped sign. It runs at the central hub using one small matrix per cluster. We randomly flip the signs of pilot-symbol rows in the residual left after explaining the other active devices. Because the flip is applied identically across all antennas, spatial correlation, heteroskedastic bursts and heavy tails are preserved. The test statistic is a quadratic form, so every flip can be evaluated from the clusters' summed $L \times L$ pilot-domain covariance, where $L$ is the pilot length. The decentralized test is therefore numerically identical to the centralized one, and its interconnect cost does not grow with the number of antennas.

Core Claim

Three properties should hold together. The test stays calibrated where parametric thresholds fail. It loses nothing from array partitioning, unlike decentralized equalization. And binning pilot symbols makes it cheap enough for a baseband deadline. Concretely, let $Y_c \in \mathbb{C}^{L \times M_c}$ be the pilots received by cluster $c$, which has $M_c$ antennas. Let $P^\perp$ be the projector that removes the pilots of the currently estimated active set $S$. Let $a_n$ be the pilot of device $n$, and let $D$ be a diagonal matrix of random $\pm 1$ signs over pilot symbols. The statistic under a flip is

$$T_n(D) = \sum_{c=1}^{C} \lVert a_n^H D P^\perp Y_c \rVert^2 = a_n^H D P^\perp \Big(\sum_{c} Q_c\Big) P^\perp D\, a_n, \qquad Q_c = Y_c Y_c^H .$$

The observed statistic is $T_n(I)$, where $I$ is the identity (no flips). The p-value is the rank of $T_n(I)$ among $B$ flipped draws. Suppose the noise and interference rows are independent across pilot symbols and sign-symmetric, which allows them to be non-Gaussian, bursty and spatially correlated. Then with oracle residuals the test has exact size. The residual projection makes the actual test approximate. Following the attainable-versus-oracle argument of Wang and Lee, its size should approach nominal as $L-|S|$ grows. We will measure that gap rather than assume it away.

The flip has to act on pilot symbols. A per-antenna flip leaves each antenna's squared correlation unchanged and so carries no information. Flipping contiguous bins of symbols is still a valid symmetry. This is the point where binned "cheap" permutations transfer exactly.

Relation to Prior Work

The activity-detection literature motivates the gap. Liu and Yu apply approximate message passing to a single centralized receiver. Their asymptotic state evolution shows false alarms and misses vanishing as antennas grow, assuming i.i.d. Rayleigh channels and Gaussian noise. We target the finite array, non-Gaussian, partitioned regime that this analysis does not cover. Li et al. distribute asynchronous activity detection across access points with a Gaussian-likelihood, reliability-weighted fusion. They report ROC curves, but they set thresholds and penalties heuristically and do not examine impulsive noise or jammers. We keep their kind of detector as the score generator and change how its decision threshold is set. Evidence from other fields that asymptotic nulls break under extreme imbalance comes from exact permutational survival statistics. There, exact nulls remove the false positives that asymptotic tests produce when tiny groups are compared with large ones, which resembles testing one rare device against a sea of inactive ones.

Two sources supply the method. From Wang and Lee we take sign-symmetric randomization of regression residuals, which is valid under heteroskedastic and sub-Weibull errors. That work is real-valued and centralized. We move it to complex multi-antenna residuals, randomize along the pilot dimension while holding the antenna dimension fixed, and express it through cluster-summable covariances. From Cheap Permutation Testing we take a different idea: randomize bin assignments over precomputed bin-level sufficient statistics, keeping finite-sample validity while cutting per-draw cost. That work stores its bin tensor in one global memory and never meets a physical layer. In our setting the sufficient statistic is naturally a sum over hardware clusters.

The decentralized-baseband literature both constrains and enables the design. Partially decentralized receivers (Li et al., Jeon et al.) reach centralized performance by summing local Gram matrices. Fully decentralized fusion, in which each cluster decides alone, degrades monotonically with cluster count and is worst under uniform partitioning (Zhuang et al.). We deliberately put the test on the summed-statistic side. This sidesteps that degradation, and we test the fully decentralized p-value-combination alternative explicitly as a foil. The jammer model of Marti et al. supplies one of our stress conditions: a directional interferer 25 dB above users.

Research Program

The first piece is the test itself. The hub needs a screening estimate of the active set $S$, obtained from a standard detector run on the summed covariance. Every device outside $S$, and devices near the decision boundary, are then tested with $P^\perp$ built from $S$ minus the tested device. Sign flips can be replaced by uniform random phase rotations when the interference is circularly symmetric. Blocked flips over runs of adjacent pilot symbols handle short temporal correlation, the same restricted-permutation fix used for serially dependent data by PCID and empirical-null construction.

The second piece is the decentralized and binned execution. Each cluster sends $Q_c$ once per coherence block. If pilot symbols are grouped into $s$ bins, the hub forms a device-specific $s \times s$ matrix from the summed covariance at a cost of $O(L^2)$ per device. Each flip then costs $O(s^2)$, and $2^s$ distinct patterns are available. That pattern count sets the smallest attainable p-value, so $s$ trades resolution at very small false-alarm targets against power.

The third piece is analysis. We will bound the size distortion that the residual projection and active-set selection introduce, in terms of $L-|S|$ and the pilots' correlation with $S$. We will also characterize where the symmetry assumption breaks. A narrowband tone jammer that is correlated across pilot symbols violates it, and blocked flips only partly repair this.

Baselines

The naive reference is the parametric threshold. Estimate the noise power from the residual and threshold $T_n(I)$ at the chi-square quantile implied by Gaussian noise at the target false-alarm level $\alpha$. It gets the same summed covariance and the same screening set, so the only difference is the null distribution. If our test does not beat it under Gaussian noise, that is expected. The comparison that matters is calibration under impulsive, bursty and jammed noise.

The strongest existing detector is the reliability-aware likelihood fusion of Li et al.. It reports a missed-detection rate more than 10 times lower than block coordinate descent at the same false-alarm rate, and near-ideal performance with 4 bits per transmitted scalar. In the centralized large-system regime, Liu and Yu report that for 2,000 users with 100 active and pilot length 110, 8 antennas reach a $10^{-5}$ detection error rate. We will rerun both in their reported settings to reproduce these points before any comparison. Their thresholds will be tuned on Gaussian-noise validation data with the same tuning budget we spend choosing $s$ and $B$. Our test will use their scores as screening, so any gain is attributable to calibration. No reported result was found on Lacuna for calibrated false-alarm control in decentralized activity detection under non-Gaussian noise.

"Beating" these baselines means two things. Our empirical false-alarm rate stays within Monte Carlo error of nominal $\alpha$ across all noise conditions while theirs drifts. And at matched realized false alarms, our missed-detection rate is no worse than theirs by more than a small, pre-registered margin.

Three ablations isolate the key ingredients. Oracle residuals versus estimated residuals measures the cost of selection and projection. Full flips versus binned flips measures the cost of binning. Summed-covariance testing versus per-cluster tests combined by Fisher's method measures the value of keeping partial decentralization.

Why not just…?

Why not just calibrate the threshold offline by simulation? If the noise distribution at deployment matches the one simulated, this works and costs nothing at run time. That is a real threat in stationary conditions. It fails when interference changes, for example when a jammer appears or impulsive bursts begin. We test this directly: calibrate under Gaussian noise, deploy under bursty and jammed noise, and report false-alarm drift against our test, which recalibrates on every coherence block.

Why not just use decoy pilots, unassigned random codes tested against the same residual, as an empirical null? This is the most serious competitor. When pilots are i.i.d. random, decoys are close to exchangeable with inactive devices whatever the noise distribution. It may match our test there, apart from a subtle selection effect: real inactive pilots took part in screening, while decoys did not. Its validity rests on symmetry of the codebook, ours on symmetry of the noise. We will run both on i.i.d. Gaussian pilots and on structured codebooks such as cyclically shifted sequences. With structured codes, decoys are not exchangeable with assigned codes, while sign flips are unaffected. If decoys match us on both codebooks, we will report that and narrow the claim accordingly.

Why not just rely on large-array theory, since massive arrays make everything Gaussian? The averaging argument applies to the whole array. Even when the array is large, impulsive bursts hit all antennas on the same pilot symbol together, so the statistic is not Gaussian. Wang and Lee found asymptotic interval coverage collapsing to 60–70% of a nominal 95% under heavy tails and correlation. Our parametric-threshold baseline measures the analogous failure here.

Why not just let each cluster test locally and combine p-values? This minimizes the hub's work, but each cluster sees few antennas. It is the detection analogue of fully decentralized fusion, which Zhuang et al. show degrades as the number of clusters grows. The Fisher-combination ablation measures its power loss against the summed-covariance test at equal fronthaul.

Proposed Experiments

The simulations reuse settings from the file. One is the 2,000-user, 100-active configuration of Liu and Yu at pilot lengths 90 and 110, with arrays of 64 to 256 antennas split into 1 to 16 clusters. The other is the 100-device, roughly 10%-active cell-free layout of Li et al.. Spatially correlated channels come from the QuaDRiGa urban microcell model used by Marti et al.. The noise conditions are Gaussian, heavy-tailed complex Student-t, Gaussian with random high-power bursts on a subset of pilot symbols, a temporally i.i.d. jammer 25 dB above users, and a narrowband tone jammer as a deliberate violation of the symmetry assumption.

Metrics are the empirical false-alarm rate against nominal $\alpha$ from $10^{-1}$ to $10^{-3}$, missed detection at matched realized false alarms, fronthaul volume per coherence block, and hub operation counts for $B=1000$ across bin counts $s$.

Four checks decide the outcome. First, under bursty and jammed noise the parametric and tuned-threshold baselines must visibly exceed $\alpha$ while our test stays nominal. If they do not, the method is unnecessary. Second, the decentralized test must equal the centralized one to numerical precision at every cluster count, while the Fisher-combination variant loses power as clusters multiply. Third, binned flips with moderate $s$ must preserve power at substantially lower cost, mirroring what Cheap Permutation Testing found for kernel tests. Fourth, the comparison with decoys on structured codebooks determines whether noise symmetry or codebook symmetry is the better thing to randomize over.