Learned controllers are starting to decide who may transmit where and when in shared radio spectrum. Examples include models that compress channel measurements, predict interference, or flag idle resources that a secondary user could borrow. They are trained on channel measurements that are expensive to collect, so in practice the same small dataset is passed through the model dozens of times. Language-model research has recently made this repetition quantitative. The effective reuse rate is the factor by which fresh data would have to grow to match $K$ passes over $N$ samples, and the number of passes that remain useful grows with dataset size. Wireless learning has no such rule. Engineers therefore cannot say in advance whether the next hour of budget should go to a new measurement campaign or to more training epochs on data they already hold.
We propose that for data collected on large antenna arrays, the useful number of repetitions depends mainly on the spatial effective rank of the dataset, not on raw sample count. The effective rank is the number of independent spatial directions the channels actually occupy. Array physics fixes this quantity: it grows with antenna count and angular spread, and it shrinks when users' arrival angles overlap. It can be estimated from a small pilot sample before any model is trained. If this holds, a covariance measurement taken on day one tells an operator how far a finite channel dataset can be stretched.
Core Claim
Define the effective reuse rate as in Larger Datasets Can Be Repeated More. $E(K,N) = N'/N$, where $N'$ is the number of fresh samples a single-pass run needs to reach the test loss that $K$ epochs on $N$ samples reach. Let $K^*(N)$ be the saturation point, the largest $K$ with $E(K,N) \ge 0.8K$. That paper fits $K^* \approx 0.80 \log N + 5.21$ on 0.3B-parameter language models and supports it with theory for Zipf-distributed features.
Our claim is that for spatial channel data, $K^*$ collapses onto a single curve in $N/r_{\text{eff}}$. Here $r_{\text{eff}}$ is the effective rank of the sample spatial covariance $C$:
$$ r_{\text{eff}} = \frac{(\operatorname{tr} C)^2}{\operatorname{tr}(C^2)}. $$
The reasoning comes from two known results. First, the linear-regression analyses show that the gain from reuse is set by the shape of the data covariance spectrum. In Improved Scaling Laws in Linear Regression via Data Reuse, multi-pass SGD beats one-pass only up to an iteration budget that depends on the spectral exponents. In Scaling Laws in Linear Regression: Compute, Parameters, and Data, variance accumulated from repeated noisy samples eventually breaks the power law. Second, large arrays give channel covariances a very un-Zipf-like shape. As MIMO evolution summarizes, cross-user inner products vanish as antennas grow. A Coordinated Approach to Channel Estimation in Large-scale Multiple-antenna Systems shows that users with disjoint arrival-angle supports occupy asymptotically orthogonal subspaces. The combined covariance is then close to block-structured, with roughly one block of energy per angular cluster and a sharp drop after it, rather than a long heavy tail.
We therefore expect three things. Repetition should saturate earlier than the text law predicts when $N \gg r_{\text{eff}}$, because there is no rare tail left for extra passes to resolve. Repetition should stay useful longer when $N$ is comparable to $r_{\text{eff}}$. And adding antennas at fixed $N$ should push $K^*$ up, not leave it unchanged. The log-$N$ text law predicts no dependence on array size or angular spread, so this contrast is directly falsifiable.
The spectrum-sharing direction supplies the task and the stakes. More than 60% of licensed spectrum below 6 GHz sits idle at any given time and place (Intelligent and efficient development of wireless networks). Reclaiming it with learned controllers, such as the reinforcement-learning pricing in Iris or the diffusion-based interference coordination in xDiff, means training on buffered, finite channel history. Iris, for instance, reports roughly 30,000 epochs to re-adapt. Neither paper asks how much of that retraining is buying new information.
Research Program
The first component is a controlled measurement of $E(K,N)$ on spatial channel data. We use DeepMIMO scenario O1 at 140 GHz, the setup of Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback, which uses an 8×8 base-station array, a 4×4 user array, and 128 subcarriers. Its large grid of ray-traced user positions serves as a finite pool of unique samples. Measuring $E$ requires a fresh-data reference, so each training set of size $N$ is drawn as a small fraction of the pool, and single-pass runs use up to $K \cdot N$ unique samples.
Two physical factors are varied independently of $N$. Array size is varied by using sub-arrays of the 8×8 panel (2×2, 4×4, 8×8). Angular structure is varied by drawing users from narrow sectors, wide sectors, or deliberately overlapping sectors. This gives datasets with matched $N$ but effective ranks spanning more than an order of magnitude.
The second component uses two tasks. The first is CSI compression with the normalization-free distillation architecture, chosen because it has a published number on the same scenario. The second is a spatial-reuse feasibility task that represents spectrum sharing. Given the channels of a primary user and a candidate secondary user on a resource block, the model predicts whether the secondary user's beamformed interference at the primary user stays below a threshold. The label is computed exactly from the ray-traced channels, so ground truth is noise-free and only the learner's use of the data is under test.
On both tasks we fit $K^*$ as a function of $N$ alone and as a function of $N/r_{\text{eff}}$, and compare how well each collapses the curves across array sizes and sector configurations.
The third component turns the law into a decision. Suppose a pilot sample of a few hundred channels is available, along with a cost per new measurement and a cost per training epoch. The rule estimates $r_{\text{eff}}$ from the pilot sample, predicts $E(K,N)$, and recommends a split of the budget between collecting new data and running more epochs. We score each rule by regret: the loss it achieves compared with the best split found by exhaustive search on held-out configurations.
Baselines
The naive reference point is a fixed-epoch default: train for a fixed large number of epochs, as wireless papers typically do, and spend all remaining budget on data collection. Any useful rule must beat it on regret. If it does not, the question is not worth a scaling law.
The strongest external law is the text-derived relation $K^* \approx 0.80 \log N + 5.21$ from Larger Datasets Can Be Repeated More. It is the only published quantitative repetition threshold in the file. We use it both unchanged and with its two constants refit on our wireless data. The refit version is the real competitor, because it lets log-$N$ scaling adapt to the domain while still ignoring physics.
We also include an exponential-decay information model in the style of InfoLaw and the half-life model of Scaling Laws for Data Filtering. Each is fitted per configuration with a free decay half-life. This is the most flexible physics-free competitor. InfoLaw reports 0.15% mean and 0.96% maximum absolute error in predicting validation loss for language models up to 7B parameters. The Data Filtering paper reports no scalar error number in the file.
The key ablation replaces $r_{\text{eff}}$ with a physics-blind statistic of equal cost, such as raw input dimension $M$ or per-sample energy. This checks that covariance structure, not just dimension, carries the signal.
For task performance, the reference is the reported CSI result from the normalization-free distillation paper: −12.87 dB NMSE at 256 bits, against −11.98 dB for its baseline. Our training pipeline must reproduce roughly this number before any reuse measurement is trusted. For the spatial-reuse feasibility task, no reported result was found on Lacuna.
All laws receive the same grid of training runs for fitting. Evaluation is always on array sizes and sector configurations held out from fitting. Each model uses the same architecture, optimizer, and learning-rate sweep at every $(N,K)$.
Beating the baselines means two things. First, the $N/r_{\text{eff}}$ law must predict $K^*$ on held-out configurations with clearly lower error than the refit log-$N$ law. Second, the resulting allocation rule must achieve lower regret than every baseline across a range of measurement-to-compute cost ratios.
Why not just…?
Why not just use early stopping on a validation set? Early stopping picks the right number of epochs once the data already exists. It cannot tell an operator, before a measurement campaign, whether the next thousand samples are worth more than fifty more epochs. That planning question is what the allocation experiment tests. We include early stopping as an oracle-for-$K$ inside the regret evaluation. If covariance-based planning adds nothing beyond "collect as much as affordable, then early-stop," the regret numbers will show it, and we will report that as the finding.
Why not just generate unlimited synthetic channels and stay single-pass? A ray tracer fixes the geometry, so a given scenario is itself a finite set of user positions, and "infinite" simulation largely resamples the same spatial structure. Real captures are strictly finite as well. RadioLLM, which adapts a pretrained language model to radio tasks, needed few-shot adaptation to handle real ADS-B and Wi-Fi captures after synthetic training. The alternative is a partial threat for purely simulated work. Our law measures exactly when fresh simulated samples stop adding information beyond repetition, which addresses the threat directly.
Why not just fit an empirical $N$–$K$ power law per deployment? This is the most serious alternative, and it is why the refit baselines exist. The case for the physical law rests entirely on extrapolation: predicting a new array size or user layout from a pilot covariance, without a fresh grid of training runs. If the refit log-$N$ law or the per-configuration decay model extrapolates to held-out configurations as well as $N/r_{\text{eff}}$ does, the physical story adds nothing and we will say so.
Why not just rely on array orthogonality and skip learned spectrum controllers? The pilot-contamination analysis of Yin et al. shows that orthogonality fails when arrival-angle supports overlap. The wavenumber analysis of Sanguinetti et al. shows that mode leakage on finite apertures vanishes only when the receiver aperture approaches the link distance. Learned controllers therefore remain necessary. Our overlapping-sector configurations are designed to probe exactly the regime where orthogonality breaks.
Decisive Checks
The central test is the curve collapse. Plot $K^*$ for every combination of array size and sector configuration, first against $\log N$ and then against $\log(N/r_{\text{eff}})$. If the second plot does not visibly tighten the spread relative to the first on held-out configurations, the core claim is false.
The second test is directional. At fixed $N$, growing the array from 2×2 to 8×8 should raise $K^*$. Switching from disjoint to overlapping sectors should lower $r_{\text{eff}}$ and lower $K^*$. The text law predicts neither effect.
The third test is practical. The allocation rule driven by a few hundred pilot channels must achieve lower regret than both the fixed-epoch default and the refit log-$N$ law on held-out layouts, on both the CSI task and the spatial-reuse feasibility task. The CSI pipeline must first reproduce the −12.87 dB NMSE reference.
As a control, repeating the measurement on single-antenna RadioML signals, which have no spatial dimension, should show that $r_{\text{eff}}$ loses its predictive advantage there. That result would confirm the effect comes from array geometry rather than from wireless data in general.
References
- Larger Datasets Can Be Repeated More: A Theoretical Analysis of Multi-Epoch Scaling in Linear Regression
- Improved Scaling Laws in Linear Regression via Data Reuse
- Scaling Laws in Linear Regression: Compute, Parameters, and Data
- Evolution of MIMO Technology
- A Coordinated Approach to Channel Estimation in Large-Scale Multiple-Antenna Systems
- Intelligent and efficient development of wireless networks: A review of cognitive radio networks
- Iris: Deep Reinforcement Learning Driven Shared Spectrum Access Architecture for Indoor Neutral-Host Small Cells
- xDiff: Online Diffusion Model for Collaborative Inter-Cell Interference Management in 5G O-RAN
- Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO
- InfoLaw: Information Scaling Laws for Large Language Models with Quality-Weighted Mixture Data and Repetition
- Scaling Laws for Data Filtering - Data Curation cannot be Compute Agnostic
- RadioLLM: Introducing Large Language Model into Cognitive Radio via Hybrid Prompt and Token Reprogrammings
- Wavenumber-Division Multiplexing in Line-of-Sight Holographic MIMO Communications