Certified Interference Nulls: Paying for Spatial Protection with Pilot Effort

A base station with many antennas can steer a "null" toward a nearby receiver it does not want to disturb. A null is a direction in which its transmissions cancel out, so the receiver hears almost nothing. How deep that null is depends entirely on how well the station knows the receiver's channel, and it learns the channel only from pilot signals the receiver chooses to send. Those pilots cost the receiver battery and airtime, and only the receiver knows how much interference it can tolerate. This note proposes treating null depth as something the station can promise and the receiver can buy with pilot effort. The station publishes a short menu of the form "send this many pilots, and we guarantee your interference stays below this level with 95% probability." Receivers that need strong protection pick demanding items, and tolerant ones pick cheap items.

The obstacle is statistical. Each guarantee has to be honored with a known failure rate. Modern channel estimators exploit the fact that high-frequency channels arrive along a few angular paths, and they find those paths with sparsity-promoting regression such as the Lasso, an $\ell_1$-penalized least-squares fit. The penalty shrinks every recovered path gain toward zero. That shrinkage is an unknown, systematic error, and it is exactly what leaks through a null. A station that builds nulls on shrunken estimates will routinely break its promises, and receivers that pay for deep nulls are hurt most. Our central observation is that de-biased sparse estimation, a correction that restores Gaussian error with a computable covariance, turns the leakage through a null into a quantity with a closed-form tail bound. That bound is what makes the promise contractible.

Core Claim

If the station places its null orthogonal to a de-biased channel estimate, the residual interference at the protected receiver is the squared projection of an approximately Gaussian estimation error. The station can therefore compute an exact failure probability for any promised leakage level. Because the error covariance shrinks in proportion to pilot effort, the certified leakage becomes an explicit function of that effort. This yields a contract with a scalar private type, the receiver's sensitivity to interference. A scalar type satisfies the standard single-crossing condition, which is the requirement that higher types value extra effort-for-protection trades more, so sorting by self-selection works. Raw Lasso, post-Lasso refitting, and heavy-tailed Bayesian estimators can all produce good point estimates. None of them yields a calibrated leakage tail, so contracts built on them violate the receivers' participation guarantees after the fact, at a rate that grows with the depth of null promised.

Concretely, let $y = \Phi A x + n$ be the receiver's uplink pilots. Here $\Phi$ is the station's $L \times N$ measurement matrix over $N$ antennas using few radio chains, $A$ is an $N \times G$ dictionary of $G \gg L$ candidate angles, $x$ is the sparse vector of path gains, and $n$ is noise of variance $\sigma^2$ per pilot. The de-biased estimator of Confidence Intervals and Hypothesis Testing for High-Dimensional Regression gives $\hat{x}^u$ with $\hat{x}^u - x$ approximately complex Gaussian with covariance $\sigma^2 M \hat\Sigma M^H / q$. In this expression $M$ is the paper's decorrelating matrix, $\hat\Sigma$ is the sample Gram matrix of $\Phi A$, and $q$ is the receiver's pilot effort in pilot energy units. Mapping this to the antenna domain gives $C(q) = A\,\sigma^2 M\hat\Sigma M^H A^H / q$. For any transmit beam $w$ orthogonal to $A\hat{x}^u$, the leakage $|h^H w|^2$ is then approximately exponential, so

$$ \Pr\big(|h^H w|^2 > \varepsilon\big) \approx \exp\!\Big(-\frac{\varepsilon}{w^H C(q)\, w}\Big), $$

and the promise "leakage at most $\varepsilon$ with failure probability $\delta$" holds whenever $w^H C(q) w \,\ln(1/\delta) \le \varepsilon$. With a shrunken Lasso estimate the error has an unknown mean, and no such statement is available.

Research Program

The first component is certification-aware nulling. Plain zero-forcing places beams in the subspace orthogonal to estimated victim channels. We instead choose the station's beams to maximize its own users' rate subject to the quadratic constraints $w^H C_k(q_k) w \ln(K/\delta) \le \varepsilon_k$ for each of $K$ protected receivers, with a union bound over receivers. This is a second-order cone program. It shapes each null away from directions where the estimation error is large, rather than relying on a large regularization constant as regularized zero-forcing does. Regularized zero-forcing, the inverse $H^H(HH^H+\alpha I)^{-1}$, trades interference leakage against noise amplification, and the trade-off has been characterized asymptotically in Asymptotic analysis of downlink MISO systems over Rician fading channels. It is itself a shrinkage estimator of the inverse channel, so its residual leakage cannot be certified for the same reason the Lasso's cannot. We will check, and expect, that the station's rate cost of a guarantee falls roughly as $1/q$, since $C(q)$ scales as $1/q$. This is the "production function" the contract needs.

The second component is the contract. Receiver $k$ has a private sensitivity $\theta_k$ and utility $\theta_k v(\varepsilon) - c\,q$, where $v$ is decreasing in leakage. The station's utility is its own sum rate under the certified constraints. The station publishes a menu of $(q_j, \varepsilon_j)$ pairs that satisfy individual rationality, meaning every receiver is no worse off than refusing, and incentive compatibility, meaning each receiver prefers its own item. The structure follows standard screening, and the template of paying with a non-monetary good comes from Paid with Models: Optimal Contract Design for Collaborative Machine Learning, which rewards data contributors with model accuracy. We reward with certified protection instead. Protection, unlike model accuracy, is rivalrous, because every null consumes the station's spatial degrees of freedom. This coupling across receivers through the cone program is the main new technical difficulty in solving the menu.

We deliberately make the channel something the station measures rather than something the receiver reports. This sidesteps the multi-dimensional screening problem the file flags as intractable. A receiver can lie about its sensitivity, which the menu handles, but it cannot fabricate a channel vector. A receiver might also shirk by sending weaker pilots than the item requires. This is detectable because the de-biased residuals give a consistent estimate of effective pilot SNR, so effort is statistically verifiable. We will bound the shirking a receiver could get away with.

The closest prior design is Contract-Based Interference Coordination in Heterogeneous Cloud Radio Access Networks. It screens a macro station's scalar channel-quality type and coordinates interference by scheduled silence periods, assuming a known type prior and no channel estimation. Our design replaces silence with concurrent spatial nulls, replaces a reported type with measured channels plus a private sensitivity, and makes the estimation error the object being contracted on. Incentive Design for Efficient Federated Learning in Mobile Networks: A Contract Theory Approach shows that menus beat uniform Stackelberg pricing under asymmetric information. We take that as motivation for using menus at all rather than as a competitor on certification.

The third component removes the known-prior assumption shared by all these contract papers. Under an incentive-compatible menu, a receiver's choice reveals only which interval of sensitivities it belongs to. Estimating the population distribution of $\theta$ from many such interval-censored observations is a missing-data problem. Expectation-maximization (EM) handles it directly: the E-step computes each receiver's expected type given its interval, and the M-step refits a parametric or mixture prior. The station alternates EM updates with menu redesign across rounds. This is the alternating hidden-state and parameter logic of Learning linear state-space models with sparse system matrices, applied to types instead of physical states.

Unlike that paper's passive sensors, receivers here are strategic. A long-lived receiver might under-claim sensitivity today to make tomorrow's menu cheaper, which is the ratchet effect. We do not claim to solve this. We treat receivers as short-lived or myopic, then measure in simulation how much forward-looking receivers degrade the learned menu. A Renegotiable contract-theoretic incentive mechanism for Federated learning addresses changing types through renegotiation but still assumes the prior is known.

Baselines

The do-nothing reference is time-division silence: the station stops transmitting while the protected receiver is active, as in Phase II of the coordination scheme in Contract-Based Interference Coordination in Heterogeneous Cloud Radio Access Networks. This gives perfect protection at a large rate cost. That paper reports qualitatively that its coordinated scheme's sum rate grows "almost linearly" with SNR while orthogonal slicing plateaus, and it reports no numeric value we can reuse. We run silence in the same simulated network at the same transmit power. Beating it means higher station sum rate at matched protection-violation rates.

The naive spatial reference builds zero-forcing nulls on raw Lasso estimates and promises leakage as if the estimate were exact, with the Lasso penalty tuned by cross-validation. This is what a practitioner would deploy first. A second strong estimation baseline refits least squares on the Lasso-selected support (post-Lasso).

The strongest estimation competitor is EM-based sparse estimation with a heavy-tailed Student-t prior on path gains, the mechanism of Learning linear state-space models with sparse system matrices. It reduces shrinkage relative to the Lasso and produces a posterior covariance that can be plugged into the same tail bound. On that paper's synthetic 10-dimensional sparse system it reports 5.70% mean relative error versus 5.86% for a maximum-likelihood baseline, with 95% topology recovery. No reported result was found on Lacuna for any of these estimators on leakage certification.

The strongest nulling competitor is worst-case robust nulling: the station assumes the error lies in a norm ball whose radius is tuned to achieve the target failure rate on held-out channel draws. We also include regularized zero-forcing with $\alpha$ tuned on the same held-out draws. No reported result was found on Lacuna for robust or regularized nulling under certified-leakage contracts.

All estimators see identical pilots, dictionaries, and SNRs. Every tunable constant gets the same validation budget of channel realizations, and all solve the same cone program. The only difference is the error model fed into the constraint. Beating these baselines means achieving the target 5% violation rate while losing less station rate than robust nulling, which meets the target only by over-protecting.

The key ablation swaps the de-biased estimate for the raw Lasso estimate inside the otherwise identical certified pipeline. This isolates the contribution of removing shrinkage bias. A second ablation replaces the estimated covariance with the oracle covariance, which separates estimator bias from covariance-estimation error.

On the economic side, we compare against a flat price, a single effort-for-protection rate offered to everyone, the uniform-pricing alternative that Incentive Design for Efficient Federated Learning in Mobile Networks: A Contract Theory Approach shows loses principal profit. That paper gives no numeric gap in the file, so we measure our own. We also compare against an oracle menu designed with the true type distribution. The EM-learned menu succeeds if its station utility converges to within a few percent of the oracle's.

Why not just…?

Why not just refit least squares on the Lasso support? Refitting removes shrinkage on the paths the Lasso found, so it is a serious competitor whenever support recovery is perfect. The failure is weak paths the Lasso drops entirely. Their energy leaks straight through the null, and the refitted error covariance knows nothing about them, so the promised bound is too tight. The de-biased estimator stays dense on purpose and assigns variance to every angular bin. The sparsity-loss concern raised in Confidence Intervals and Hypothesis Testing for High-Dimensional Statistical Models does not bite here, because a null only needs a vector, not a sparse one. The experiment sweeps the power of the weakest path. We expect post-Lasso violation rates to rise sharply once that power approaches the noise floor, while de-biased rates stay near the target. If they do not diverge, post-Lasso is good enough and we will say so.

Why not just use a heavy-tailed Bayesian prior fitted by EM? This is the most credible threat. The Student-t approach of Learning linear state-space models with sparse system matrices already mitigates shrinkage, and its posterior covariance can enter our constraint directly. That paper gives no coverage guarantees, however, and Bayesian posteriors under a misspecified prior are often overconfident. Calibration is exactly what a contract needs. We run it as a full drop-in alternative. If its empirical violation rate matches ours across path counts and SNRs, the contribution narrows to the contract and prior-learning layers, and we would recommend de-biasing the Bayesian estimate rather than replacing it.

Why not just add a fixed safety margin, through robust nulling or larger regularization? A margin tuned on validation data can hit any average violation target. It does so by over-protecting receivers with well-estimated channels and under-protecting those with poor ones, because a single margin ignores how the error covariance $C(q)$ varies across receivers and effort levels. That variation is precisely what lets the menu price effort, so a fixed margin also erases the incentive structure. We will report per-receiver violation rates, not just the average, and the station's rate cost at matched average violation.

Why not just schedule silence or orthogonal slots? The H-CRAN contract work already shows qualitatively that orthogonal slicing plateaus in sum rate while concurrent coordinated transmission keeps growing. The over-the-air work in Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air shows that spatial suppression drives inter-task interference down as $O(1/N_r)$ in the number of receive antennas while orthogonal schemes pay latency linear in the number of users. Silence remains our naive reference so the gain is measured rather than assumed.

Proposed Experiments

The first check validates the statistics before any wireless model enters. We reproduce de-biased coverage on synthetic Gaussian designs and on the riboflavin genomic data ($n=71$, $p=4088$), where Confidence Intervals and Hypothesis Testing for High-Dimensional Regression reports 95% intervals with actual coverage between 96% and 98%. Matching that range confirms our implementation of $M$.

The main testbed is a Monte Carlo simulation in the style of the H-CRAN environment: a 64- to 256-antenna station with few radio chains, multipath channels with 2 to 20 paths on and off an angular grid, several protected receivers, and its own served users. Its metrics are the per-receiver violation rate against the 5% target, the station's sum rate, the rate of ex-post participation violations, and the station's rate cost per unit of certified protection.

Two sweeps decide the claim. The first varies the number of paths. Validity of de-biasing requires sparsity of order $o(\sqrt{n}/\log p)$, so certification should hold for a few paths and break as scattering becomes rich. Locating that break point tells practitioners when the method applies, roughly millimeter-wave rather than rich sub-6 GHz scattering. The second varies off-grid angle mismatch. Dictionary model error is not covered by the de-biasing theory, and we expect a leakage floor; we will quantify it and test whether a finer grid restores calibration.

For the contract layer, we sample sensitivity types from shifting mixture distributions. We track the EM estimate of the prior in total variation, the station's utility regret against the oracle menu over rounds, and the degradation when a fraction of receivers are forward-looking and shade their claims.

The proposal fails if the Lasso-in-pipeline ablation already meets the violation target. It is weakened to an estimation note if the Bayesian estimator matches de-biased calibration. It succeeds if de-biased certification is the only estimator that holds the 5% target across receivers, while costing the station less rate than robust margins and silence.