Systematic Error Bands for Moment-Matched Fading Proxies in Single-Antenna Ultra-Reliable Links

Wireless analysts routinely replace a channel distribution they cannot integrate with a simpler stand-in. The usual stand-in is a Gamma distribution whose mean and variance are set equal to those of the true channel, a procedure called moment matching. The resulting closed-form outage formulas then drive design choices such as how many reflecting elements a surface needs or how to split power between users. This proposal argues that moment matching carries a hidden, non-shrinking bias in exactly the place ultra-reliable systems care about. That place is the probability of deep fades at the $10^{-5}$ to $10^{-9}$ level. The bias should be treated the way precision physics treats a residual systematic uncertainty: as an error that more samples will not remove, which must be quantified and reported as an explicit error band next to the central estimate.

The idea in two sentences: a Gamma proxy's behavior near zero signal strength is fixed by the first two moments, which are dominated by the bulk of the distribution, so the proxy can predict the wrong rate at which outage falls with SNR. We pin the proxy's near-zero behavior to the exact one, and we compute the full range of outage probabilities consistent with everything we actually know about the channel. That range becomes a guaranteed band, and it is cheap to compute precisely because single-antenna links reduce the problem to one scalar random variable.

Core Claim

Our first claim is that, for single-antenna (SISO) links, moment-matched Gamma and Log-normal proxies produce a systematic, sample-size-independent error in rare-event outage. In SISO links the received signal is $y = hx + n$ with scalar gain $h$. This setting includes cascaded links through a reconfigurable intelligent surface (RIS, a panel of passive elements that re-radiates signals) and links with transceiver hardware flaws.

The error has two concrete sources. The first is a mismatch in diversity order, meaning the slope of outage versus SNR on a log-log plot. The proxy's slope is its shape parameter $k = \mathbb{E}[X]^2/\mathrm{Var}(X)$. The true slope is set by how much probability the channel places near zero. For a coherent sum of $N$ double-Rayleigh products, these need not coincide. The second source is error floors. Li et al. showed that I/Q imbalance in single-antenna backscatter NOMA drives diversity order to zero, yet no fixed-shape Gamma proxy can represent a floor.

Our second claim is that this error can be bounded rather than ignored. Let $X$ be the scalar SNR and let $\mu_j$ be its exactly computable moments. Let $T$ denote an exactly derived constraint on the small-$x$ behavior: the leading exponent and coefficient of $P(X\le x)$ as $x\to 0$, plus any floor mass from hardware impairments. The outage at threshold $\gamma$ is then bracketed by

$$ \underline{P}(\gamma)=\min_{F\in\mathcal{F}} F(\gamma),\qquad \overline{P}(\gamma)=\max_{F\in\mathcal{F}} F(\gamma), $$

where $\mathcal{F}$ is the set of all distributions on $[0,\infty)$ with moments $\mu_1,\dots,\mu_m$ that also satisfy $T$. Because $X$ is one-dimensional, this is a classical moment problem that can be solved as a small linear or semidefinite program. The SISO constraint is therefore what makes guaranteed bands tractable; with vector channels the same problem becomes high-dimensional.

The deliverable has two parts. The first is a tail-anchored proxy, which matches mean, variance, and the exact near-zero exponent. The second is its band, reported in the physics style as central value plus systematic interval.

Relation to Prior Work

Several papers motivate the gap by validating Gamma proxies only where they are easy to check. Xie et al. approximate cascaded STAR-RIS fading in multicell NOMA as Gamma via moments. Zhao et al. and Qian et al. do the same for ergodic rate and RIS SNR. All three treat $(k,\theta)$ as exact. Xie et al. validate with $10^6$ Monte Carlo trials, which cannot resolve outage below roughly $10^{-5}$. Their "almost perfect" agreement is therefore silent on the URLLC regime. We do not change their network models; we attach a tail-accurate correction and an error band to them.

Shekhar et al. supply the strongest existing method. They use two-step Log-normal moment matching for cell-free massive MIMO under pilot contamination and report a close fit from −10 to 30 dB that is far better than the use-and-then-forget bound. Their result relies on averaging across many access points, which thins the tails. SISO links lack this averaging, and Shekhar et al. themselves leave a general tail guarantee open. We supply that guarantee for the scalar case.

Li et al. constrain the design. Their zero-diversity floor under I/Q imbalance means any proxy family without floor mass is structurally wrong at high SNR, so our constraint set $T$ must admit a floor.

The methodological lens comes from precision physics. In those analyses, systematics are separated from statistics and budgeted explicitly (Aleksić et al. on systematic-limited Crab Nebula measurements). Agreement between independent channels with uncorrelated systematics serves as a cross-check: Daya Bay's hydrogen-capture measurement agreed with the gadolinium channel's detector ratios within 0.6% (An et al.). We borrow the idea of reporting a systematic band. We also borrow the habit of checking a prediction against two independent "truths" whose errors differ: exact simulation and ray-traced data. We do not borrow ratio cancellation. Proxy error is not a common multiplicative factor, so ratios would not remove it.

Research Program

The first step is analytic: deriving the tail anchor $T$ for the SISO channels used in the cited literature. These are optimally phased RIS links with Rayleigh, Rician, and Nakagami hops, and the AmBC-NOMA link of Li et al. with I/Q imbalance. For each, we derive the leading small-$x$ behavior of the SNR distribution, including logarithmic corrections from products of Rayleigh variables and the floor mass from impairments. We then compare the true exponent against the moment-matched $k$ as a function of $N$ and Rician factor. This comparison alone tells designers when the standard Gamma proxy is optimistic or pessimistic.

The second step builds the tail-anchored proxy. One candidate is a two-component mixture: a small-$x$ component carrying the exact exponent and coefficient, plus a Gamma bulk, with weights set so that mean and variance still match. The other is a generalized Gamma whose extra power parameter is pinned by $T$ rather than by a third moment. Both keep closed-form CDFs, so they can be dropped into existing stochastic-geometry pipelines such as that of Xie et al.

The third step computes the moment-problem band. We solve the band as a function of $\gamma$, which yields upper and lower outage curves. We then study three effects. The first is how band width shrinks as more moments are added. The second is how it shrinks when the tail constraint is added. Our expectation is that moments alone leave bands spanning orders of magnitude at $10^{-6}$, because moments weight large $x$, while the tail anchor collapses them. The third is how the band propagates into design quantities, such as the minimum $N$ or SNR margin needed to meet a target outage.

Baselines

The naive reference point is the standard two-moment Gamma proxy exactly as used by Xie et al., with no band. Every method sees the same channel models and parameters. A method beats this baseline if its predicted outage at targets from $10^{-3}$ to $10^{-9}$ stays within a factor of two of the ground truth where the naive proxy is off by more. A second beating condition concerns design decisions: the SNR margin our method recommends must meet the target when the naive one fails it.

The strongest existing competitor is two-step Log-normal moment matching from Shekhar et al.. It is the only method in the file explicitly credited with good tail behavior, and it handles estimation error. We reimplement it on our SISO channels using the same moments we give our method. That paper reports only qualitative agreement with Monte Carlo over −10 to 30 dB and outage-reduction percentages tied to antenna scaling (69.45% for cell-free versus 16.61% collocated). No reported result was found on Lacuna for its outage error at probabilities below $10^{-5}$ or in single-antenna settings, so the comparison number must be produced here.

A third baseline is a higher-moment proxy: a generalized Gamma or Gamma mixture fit to three or four exact moments, with no tail anchor.

Ground truth is importance-sampled Monte Carlo, which samples deep fades more often and reweights them, accurate to about 10% relative error at $10^{-9}$. Every method receives the same compute budget for its closed form. Ground truth gets unlimited budget, and we report its wall-clock cost separately.

Two ablations isolate the key ingredient. The first removes the tail anchor from our proxy, leaving mean and variance only, and checks whether accuracy collapses back toward the naive proxy. The second removes the moments from the band, leaving the tail constraint only. Together these show which piece of information carries the tail.

Why not just…?

Why not just match three or four moments instead of two? This is the most serious threat. If a three-moment generalized Gamma already lands within a factor of two at $10^{-7}$ across our channels, the tail-anchored point proxy adds little. In that case the contribution narrows to the guaranteed band, and we will say so. Our reason to expect higher moments to fail is that $\mathbb{E}[X^j]$ for $j\ge 3$ is increasingly dominated by large SNR values, while outage lives near zero. The higher-moment baseline and the band-width-versus-number-of-moments study test this directly.

Why not just run Monte Carlo? Monte Carlo is the ground truth here, and importance sampling makes rare events reachable. The file stresses, via Xie et al. and Zheng et al., that design loops evaluate outage thousands of times. We will report time per outage evaluation and total time for a full power-split optimization with each method. If importance sampling turns out fast enough for those loops, the case for closed forms weakens to interpretability, and we will report that.

Why not just use the known high-SNR asymptotic outage expression? Asymptotics such as the diversity-order analysis of Li et al. are exact as SNR grows but can be badly off at moderate SNR, where the Gamma bulk is accurate. Our proxy is designed to agree with the asymptote in one regime and with the Gamma in the other. We include the pure asymptote as a baseline and measure the SNR range over which each method stays within a factor of two.

Why not just add hardware distortion as extra Gaussian noise? Treating impairments as noise, as the MISO secrecy analyses of Shi et al. do for channel uncertainty, shifts the mean but keeps a fixed-shape proxy. It therefore cannot reproduce a floor. The I/Q imbalance experiments test whether this matters numerically.

Proposed Experiments

The first experiment uses synthetic SISO RIS links with $N$ from 4 to 256 under Rayleigh, Rician, and Nakagami hops, reproducing the cascaded model of Xie et al.. We report the log of predicted-over-true outage for each method at target probabilities from $10^{-3}$ to $10^{-9}$. We also report band coverage, meaning the fraction of configurations where the importance-sampled truth falls inside the band, which should be 100% by construction, so any miss signals a derivation error. Finally we report band width in decades.

The decisive check is whether the naive Gamma error grows as the target probability shrinks while ours stays flat. A second check is whether the sign of the naive error flips predictably with the gap between the moment-matched $k$ and the true exponent. If the naive Gamma is already within a factor of two everywhere, the premise fails.

The second experiment applies the method to the AmBC-NOMA I/Q-imbalance setup of Li et al. combined with an RIS hop. It tests whether the floor-aware band contains the true high-SNR plateau while all fixed-shape proxies predict continued decay.

The third experiment is the independent cross-check, analogous to the second detection channel at Daya Bay. We extract single-antenna sub-channels, one entry of the $4\times4$ complex matrix, from UrbanMIMOMap. Its 0.5 m grid lets us form empirical local SNR distributions over small neighborhoods that include blockage and multipath not present in textbook fading. The dataset's reported benchmark is RadioUNet map reconstruction (NMSE 0.0793, SSIM 0.7989), which measures map quality rather than tail accuracy. No reported result was found on Lacuna for outage-tail accuracy on it.

Here exact moments are replaced by empirical ones. The test is whether the bands, inflated for moment-estimation statistics, still cover the empirical lower tails. Neighborhood sample sizes limit this check to roughly the $10^{-3}$ to $10^{-4}$ range. It tests model realism rather than URLLC depth, and we will state that limit plainly.

The fourth experiment measures downstream design error. For a target outage of $10^{-6}$, we compare the RIS size and NOMA power split each method recommends against the choice that truly meets the target. The headline number is the fraction of designs the naive proxy certifies that actually violate the target.