Scheduling Pilots by Their Information Value on Non-Gaussian Channels

A base station has to decide, slot by slot, which users get a pilot. A pilot is a known training symbol that lets the station re-estimate a user's channel, and there are fewer pilots than users. Users who go unprobed slowly drift out of date, so this is a sequential decision problem. The best existing treatment models it as a restless bandit, meaning a set of options whose states keep changing even when nobody looks at them. That treatment assumes each pilot reveals the channel exactly, and that the channel is a small Markov chain. We propose to drop both assumptions. Each user is instead scored by how much one noisy pilot would actually reduce the error of the best possible channel estimate. A learned denoiser is used to compute that score on channels that are not Gaussian.

The connecting idea is the information–estimation identity known as I-MMSE. It says that the information an observation carries equals an integral of the minimum mean squared error (MMSE) of estimation across signal-to-noise ratios. A corollary says that the error drop caused by one observation equals the expected squared change it causes in the conditional-mean estimate. On Gaussian channels, such as the 3GPP link-level models, all of this collapses to a closed-form Kalman-filter computation, and nothing learned can help. On ray-traced channels it does not collapse. Our central bet, which we will test directly, is that pilot scheduling gains from non-linear estimation exist only where the channel is non-Gaussian. This turns a known negative result about simplified channel models into a design rule for when learned scheduling is worth its cost.

Core Claim

For user $u$ at slot $t$, let $\mathcal{Y}_{t-1}$ be all of that user's past pilot observations. Let $y_t$ be the noisy pilot observation $y_t=\sqrt{\rho}\,h_u(t)+n$, at pilot SNR $\rho$, if a pilot were assigned now. Let $\hat h$ be the conditional-mean channel estimate. We define the pilot value as

$$\Delta_u(t)=\mathbb{E}\big\|\hat h_u(t\mid\mathcal{Y}_{t-1},y_t)-\hat h_u(t\mid\mathcal{Y}_{t-1})\big\|^2=\text{MMSE}_{\text{pred}}-\text{MMSE}_{\text{filt}},$$

where the equality holds for conditional-mean estimators by orthogonality. The information this pilot carries about the channel is

$$I(h_u(t);y_t\mid\mathcal{Y}_{t-1})=\tfrac12\int_0^{\rho}\text{mmse}_u(\gamma)\,d\gamma,$$

with $\text{mmse}_u(\gamma)$ the conditional error at pilot SNR $\gamma$. A single conditional denoiser trained across SNRs supplies both quantities.

The claim has three parts. First, a scheduler built on $\Delta_u$ inherits the structure of Asymptotically optimal pilot allocation over Markovian fading channels: a Lagrangian relaxation into per-user problems, followed by an index policy. The difference is that sensing is noisy and the belief is continuous. Second, on 3GPP channels this scheduler should match a Kalman-based scheduler. Third, on DeepMIMO ray-traced trajectories it should beat that scheduler in long-run spectral efficiency.

Relation to Prior Work

Each prior paper plays one of four roles here.

Two papers supply the method. MMG: Mutual Information Estimation via the MMSE Gap in Diffusion estimates mutual information as the area between conditional and unconditional denoising curves. Its "orthogonal principle" computes the gap as an expected squared difference of predictions rather than as a subtraction of two noisy errors. We reuse exactly that estimator to compute $\Delta_u$. MMG does this offline, on synthetic data, with an isotropic Gaussian corruption. We apply it online, inside a decision loop, with the past observation history as the conditioning input. Information-Theoretic Diffusion supplies the justification that MSE regression across noise levels suffices to learn the needed conditional expectations. It also reports a known weakness: at high SNR, neural estimates of the error gap are biased low. We treat that bias as a risk to measure.

One paper motivates the gap. An Information-Theoretic Analysis of Discrete-Time Control and Filtering Limitations by the I-MMSE Relationships writes directed information in feedback channels as a sum of prediction-minus-filtering MMSE terms. That is precisely our per-slot pilot value, summed over time. However, it only bounds given systems over AWGN channels and never chooses actions. It also notes that nonlinear filters make these MMSEs expensive to compute. We address that cost with a learned denoiser.

Another paper motivates the gap from the scheduling side. The restless-bandit work cited above gives the index-policy template and a reported gap below 0.1% to an upper bound on 20–40-user networks. It assumes noise-free sensing, and it assumes that probing resets the channel to its steady-state distribution. We remove both assumptions.

One paper constrains the design. Wireless Channel Modeling for Machine Learning -- A Critical View on Standardized Channel Models shows that 3GPP tapped/clustered delay line channels are near-Gaussian. On those channels, linear MMSE estimation is optimal and beats a diffusion denoiser. It reports nMSE 0.00915 against 0.01204 on CDL-A at 10 dB. This rules out claiming any learned gain on those models. It also predicts our null result there.

The evaluation setting comes from two papers. Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO supplies the DeepMIMO configuration. Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading supplies the mobility-transition protocol, with users accelerating from 3 to 120 to 500 km/h.

Research Program

The first component is a conditional channel denoiser. It is trained on channel trajectories and conditioned on a summary of the past pilot observations. It is trained by MSE regression across pilot SNRs, so that it outputs the conditional mean both with and without a hypothetical current pilot. At decision time, we draw a few samples of the hypothetical pilot $y_t$ from the model's own predictive distribution. $\Delta_u$ is then estimated with MMG's orthogonal form. The same network yields the information integral when we need an information-weighted variant.

The second component is the scheduler. Following the Lagrangian relaxation of the restless-bandit paper, each user faces a single-user problem: probe or not, with a price on probing. We do not assume closed-form indexability, which that paper notes is hard to verify beyond two states. Instead, we solve each single-user problem by fitted value iteration over a compact belief summary: predicted error, estimate magnitude, time since last probe, and Doppler estimate. The per-slot reward is the achievable-rate lower bound under imperfect channel knowledge, in which estimation error counts as extra noise. This makes $\Delta_u$ enter the reward directly. The probing price at which a user becomes indifferent serves as its index, and the $M$ highest-index users are probed.

The third component is analysis. We will check whether indifference prices are monotone in the probing price, which is an empirical indexability check. We will measure how far the denoiser's $\Delta_u$ departs from the realized MSE drop as a function of pilot SNR, which quantifies the known high-SNR bias.

Baselines

The naive reference point is round-robin pilot assignment by staleness: probe whoever was probed longest ago. Alongside it we run a myopic scheduler that probes the users with the highest predicted immediate rate. Both use the same estimator as our method, so differences isolate the scheduling rule.

The prior structured method is the Whittle index policy of Asymptotically optimal pilot allocation over Markovian fading channels. We fit a $K$-state Markov chain to quantized channel gains from the same trajectories and apply that paper's closed-form index. Its estimation stage is then evaluated honestly, with noisy pilots. Its reported figure is a suboptimality gap below 0.1% under its own perfect-sensing model on 20–40 users. Beating it means higher sum spectral efficiency once sensing is noisy.

The strongest competitor is a Gaussian scheduler. It is the same relaxation and value iteration as our method, but the belief is a Kalman filter fitted to the second-order statistics of the training trajectories, and $\Delta_u$ comes from closed-form covariance updates. On Gaussian channels this is the provably right answer per the critical view on standardized models. It gets the same training trajectories and the same per-user value-iteration budget. No reported result was found on Lacuna for pilot scheduling of this kind on DeepMIMO. The closest anchor is that paper's static estimation comparison on CDL-A, where linear MMSE reaches nMSE 0.00915 against 0.01204 for diffusion. We count our method as beating the Gaussian scheduler only if it improves long-run sum spectral efficiency on DeepMIMO by more than three seed-level standard errors and by at least 2% relative.

We run two ablations to isolate the key ingredient. In the first, the learned denoiser does estimation but scheduling uses the Kalman $\Delta_u$. This separates better estimation from better pilot choice. In the second, $\Delta_u$ is computed by subtracting two empirical MSEs instead of the orthogonal form. This tests whether MMG's variance reduction matters inside a decision loop.

Why not just…?

Why not just use a Kalman filter and its closed-form error drop? This is the real threat. On 3GPP channels we expect it to tie our method, and we will report that tie as a result rather than hide it. Whether it loses on ray-traced trajectories is exactly what the DeepMIMO comparison decides. The static evidence for non-Gaussian structure there, an NMSE of −12.87 dB versus −11.98 dB in the CSI feedback paper, suggests room for a gain but does not guarantee that it survives in scheduling. If the gap falls under our 2% threshold, the conclusion is that Gaussian scheduling suffices, and the proposal's contribution reduces to that negative finding.

Why not just reward the rate bound and skip I-MMSE entirely? For the per-slot reward this is essentially what we do. The I-MMSE corollary earns its place in the computation: it provides an estimate of the error drop as one squared prediction difference, which the subtraction ablation tests against the naive alternative. The information integral is a secondary variant. If it adds nothing beyond the error drop, we will say so.

Why not just train a model-free reinforcement learning agent on raw pilot histories? With $N$ coupled users the joint action space is combinatorial, and the restless-bandit literature exists because direct solution is intractable. We include a per-user deep Q-learning agent under the same relaxation and training budget. This shows whether the explicit error-drop belief summary beats a learned one.

Why not just use outer-loop link adaptation or conformal link adaptation to absorb estimation error? Those methods choose a modulation given whatever estimate exists. They do not decide who gets probed. Conformal-AMC reports an 82.4% singleton selection rate against 75.3% for outer-loop adaptation, but it treats estimation as a black box. These methods are complementary to our scheduler rather than competitors, so we keep link adaptation fixed across all compared methods.

Proposed Experiments

Channel suites. We use two suites. The first is 3GPP TDL-A/C and CDL-A, with Jakes Doppler and the 3→120→500 km/h transitions from the conformal link-adaptation paper. The second is DeepMIMO O1 at 140 GHz. It uses an 8×8 base-station array, a 4×4 user array, 128 subcarriers, and a 6:2:2 split. Temporal trajectories are built by moving users along the ray-traced position grid at the same speeds and interpolating between grid points. This construction is synthetic in its timing, and we flag it as such.

Scale and conditions. Networks have $N\in\{20,40,100\}$ users and $M/N\in\{0.1,0.25\}$. Pilot SNR ranges from 0 to 20 dB.

Metrics. The metrics are long-run sum spectral efficiency under the imperfect-knowledge rate bound, time-averaged channel nMSE, 5th-percentile user rate, and scheduler compute per slot.

Decisive checks. The first decisive check is the Gaussian-trap prediction. On 3GPP suites, our scheduler and the Gaussian scheduler should be statistically indistinguishable. A learned gain there would indicate leakage or a bug rather than a discovery. The second decisive check is the DeepMIMO gap against the 2% threshold. This gap should widen at higher speeds and in shadowed regions, where non-Gaussianity is strongest.

Calibration and indexability. A calibration check plots the predicted $\Delta_u$ against the realized MSE drop across pilot SNR. Low-biased estimates at high SNR would lead the scheduler to under-probe high-SNR users, and we will test whether that effect explains any loss. The indexability check reports how often indifference prices violate monotonicity. Because the index policy is a heuristic whenever indexability fails, frequent violations would weaken any claim of near-optimality.