Testing Angle Reciprocity Before Trusting It

In frequency-division duplex (FDD) wireless systems, the uplink and downlink use different frequency bands. Fast channel coefficients therefore differ between the two directions, but the physical geometry is shared: the buildings and reflectors are the same. The directions from which signal paths arrive and depart are therefore roughly the same on both links. This is called angle reciprocity. Several recent schemes rely on it to cut downlink training and feedback cost sharply. The base station estimates path angles from uplink pilots, sends downlink pilots aimed only along a few dominant angles, and the handset feeds back just the gains of those few paths. Every such scheme assumes the uplink angles are correct for the downlink. Hardware distortion, uncalibrated front ends, wide frequency gaps, and user motion can all break that assumption, and when they do, nothing in the protocol notices.

We propose adding a cheap statistical check to the downlink pilot. For each dominant path, the base station sends two extra pilots aimed slightly to either side of the uplink angle. The handset reports how the received energy splits between them. Under correct reciprocity the split is balanced, whatever the unknown path gain, so the imbalance can be turned into a calibrated p-value per path per frame. Across all paths and users, these p-values feed a multiple-testing rule that controls how often the system raises false alarms. Only flagged paths get angle correction or extra training. The work contributes three things: a hypothesis test that is valid regardless of the unknown fast-fading gain; error control across many simultaneous decisions; and a measured answer to when angle reciprocity actually fails.

Core Claim

A per-path, gain-invariant test statistic built from two offset pilots can detect angle-reciprocity violations at a controlled false-alarm rate. This holds across signal-to-noise ratios and channel scenarios without retuning. Using the test to decide when to correct angles recovers most of the spectral efficiency lost to mismatch. The overhead is a few extra pilots per dominant path, far below the full-channel training it replaces.

What is new is this online, per-path, calibrated decision. The closest existing statistical treatment is Learning the CSI Recovery in FDD Systems. It uses maximum mean discrepancy, a two-sample distance between distributions, offline and globally to argue that uplink and downlink channels share a "statistical fingerprint." That result motivates the premise but cannot flag a specific path in a specific frame. Statistical Beamforming for FDD Downlink Massive MIMO via Spatial Information Extraction and Beam Selection uses "statistical" to mean long-term angular power spectra and treats angle mismatch as a perturbation. It never tests mismatch as a hypothesis.

The Test

Let $\hat\theta_l$ be the uplink-estimated angle of dominant path $l$ and $\delta$ a small angular offset, on the order of half a beamwidth. The base station sends pilots beamformed toward $\hat\theta_l+\delta$ and $\hat\theta_l-\delta$. The handset measures received energies $P_l^{+}$ and $P_l^{-}$ and reports the normalized imbalance

$$ T_l = \frac{P_l^{+} - P_l^{-}}{P_l^{+} + P_l^{-}} . $$

If the true downlink angle equals $\hat\theta_l$, the two offset beams see the same array gain. The unknown complex path gain, which is not reciprocal in FDD, multiplies both energies equally and cancels in the ratio. The idea is borrowed from monopulse direction finding in radar.

Under the null hypothesis, $T_l$ is zero-mean. Its spread depends only on the per-pilot signal-to-noise ratio, which the handset already estimates. The null distribution can therefore be computed in closed form, as a ratio of noncentral chi-square variables, or tabulated by simulation. Under the alternative, the sign of $T_l$ indicates the direction of the angle error and its size grows with the error. A rejected test thus also yields a first-order correction.

Many paths and users are tested each frame, so testing each at level 0.05 would produce frequent false alarms. We control the false discovery rate across the per-frame family with Benjamini–Hochberg. As a stricter comparison we also use Bonferroni-type family-wise control, the correction Turning Analysis during Standardized Test Using On-Shoe Wearable Sensors in Parkinson's Disease applies for the same multiplicity problem in a clinical setting.

Mismatch from hardware or calibration persists over time, so we also run a sequential version. It is a cumulative-sum test on per-frame statistics, which lets offset pilots be sent only in a fraction of frames once reciprocity has been confirmed.

Research Program

The first strand derives and validates the null distribution of $T_l$. This includes finite-resolution uplink angle estimates, where $\hat\theta_l$ itself is noisy, and leakage from neighboring paths. Leakage is the main threat to validity in rich scattering, the regime where Downlink Pilot Precoding and Compressed Channel Feedback for FDD-Based Cell-Free Systems already warns that selecting a few dominant paths degrades.

The second strand embeds the test in the dominant-path protocol of Beamforming Design with Partial Channel Estimation and Feedback for FDD RIS-Assisted Systems. That protocol splits channel parameters into slowly varying, reciprocal path angles and fast, non-reciprocal path gains, and it cuts feedback from all 256 surface elements down to the gains of 4 dominant paths. Our addition is small: two offset pilots per path and a few feedback bits for $T_l$. On rejection, a path's angle is shifted using the sign and size of $T_l$ and re-probed. If rejections are widespread, the system falls back to broader downlink training.

The third strand is a characterization study. We use the test itself to map where angle reciprocity fails as a function of carrier gap, signal-to-noise ratio, hardware distortion level, user speed, and angular spread. Rejection rates and effect sizes are reported across these factors with analysis of variance and post-hoc correction. This turns the qualitative warnings in the literature into measured boundaries.

As a secondary evaluation lens, we report Bland–Altman limits of agreement between predicted and true downlink beam gains alongside normalized mean squared error. Comparative analysis of two mathematical algorithms for the calculation of computed tomography perfusion parameters in the healthy and diseased pancreas shows that correlation alone hides systematic bias between two estimators. Normalized mean squared error has a similar blind spot here: it averages away a consistent angle offset that matters for beamforming.

Baselines

The naive reference point is the dominant-path scheme of Beamforming Design with Partial Channel Estimation and Feedback for FDD RIS-Assisted Systems run as published. It always trusts the uplink angles and has no detection. It is the right reference because our test adds overhead only to fix its failure mode. The paper reports that feedback drops from 256 elements to 4 dominant paths and that spectral efficiency stays high with 8 feedback bits. However, no reported result was found on Lacuna for its spectral efficiency under injected angle mismatch or hardware distortion. We will measure that ourselves on its 28 GHz setup: 4×4 base-station array, 16×16 surface, $L=6$ paths, $L_s=4$ selected.

The strongest existing reciprocity-based method is the cell-free scheme of Downlink Pilot Precoding and Compressed Channel Feedback for FDD-Based Cell-Free Systems. It estimates angles with subspace methods and selects dominant paths by signal-to-leakage-and-noise ratio. It reports over 80% feedback reduction and about 10 dB gain over conventional quantized feedback at 15 bps/Hz spectral efficiency. We run it with and without our test on the same channel realizations and the same pilot and feedback budget. The budget is counted in total pilot symbols and feedback bits, so the offset pilots are charged against us.

The full-overhead reference is conventional downlink training with Type II-style high-resolution feedback, which does not rely on reciprocity. As Pushing The Limit of Type I Codebook For FDD Massive MIMO Beamforming: A Channel Covariance Reconstruction Approach notes, it gains up to 30% over Type I codebooks at heavy uplink cost. It bounds what mismatch-robust performance is achievable.

Two ablations isolate the key ingredients. The first replaces the calibrated test with a fixed threshold on $|T_l|$, tuned for best spectral efficiency on one scenario and then frozen. The second is always-refine: offset pilots every frame and angle correction every frame, with no test.

Beating the baselines means three things. First, within 5% of the reciprocity-trusting scheme's spectral efficiency when reciprocity holds. Second, recovery of at least half the gap to full training when mismatch is injected. Third, false-alarm rates within the nominal level across all test scenarios, where the frozen threshold drifts.

Why not just…?

Why not just use a tuned threshold on the imbalance? This is the most serious alternative, because a threshold on $|T_l|$ is a test with an uncalibrated level. Our claim is not that the calibrated test detects better at a single operating point. The claim is that its false-alarm rate stays fixed as signal-to-noise ratio, array size, and scenario change, while a frozen threshold's does not. The experiment tunes the threshold on the QuaDRiGa urban microcell and evaluates both methods on the clustered-delay-line and RIS channels. If a signal-to-noise-scaled threshold matches us everywhere, we will say so: the contribution then reduces to the statistic and the multiplicity control.

Why not just refine angles every frame? Always-refine removes the decision entirely. It is a real threat if angle corrections are cheap and accurate. It pays in three ways, however: feedback bits every frame; injected estimation noise on paths that were already correct; and no sequential savings in pilots. The always-refine ablation measures this directly. If it wins at high signal-to-noise ratio, the test's role narrows to low-SNR and overhead-constrained regimes, and we will report that boundary.

Why not just learn the uplink-to-downlink mapping end to end? Amplitude Prediction from Uplink to Downlink CSI against Receiver Distortion in FDD Systems shows that hardware distortion breaks reciprocal angular features. It succeeds only with a separate network that first removes the distortion, trained on a known distortion model. Learning the CSI Recovery in FDD Systems notes that such models are site-specific. A learned mapping also gives no signal when it is wrong. Our test is complementary: it can guard a learned predictor, and we evaluate it on the distortion model from that paper.

Why not just rely on the offline distribution check? The maximum mean discrepancy analysis in Learning the CSI Recovery in FDD Systems establishes similarity over a whole dataset. A single miscalibrated front end, or a moving user, can be wrong while the aggregate looks fine. We show this by injecting per-user mismatch at rates that leave dataset-level discrepancy essentially unchanged, and measuring per-path detection.

Proposed Experiments

We use three channel sources from the file, each targeting a different mismatch cause.

The 3GPP geometric cascaded RIS channel at 27.875/28.125 GHz, from Beamforming Design with Partial Channel Estimation and Feedback for FDD RIS-Assisted Systems, carries controlled angle offsets and calibration errors.

QuaDRiGa urban microcell non-line-of-sight channels with carrier gaps from 120 to 480 MHz, as in Learning the CSI Recovery in FDD Systems, probe naturally frequency-dependent angle differences.

MATLAB clustered-delay-line channels at 5.1/5.3 GHz and 300 km/h, with log-normal amplitude and uniform phase receiver distortion, as in Amplitude Prediction from Uplink to Downlink CSI against Receiver Distortion in FDD Systems, cover hardware distortion and stale angles.

Metrics are: empirical false-alarm rate and false discovery rate under true reciprocity; detection power versus angle error in beamwidths; spectral efficiency and total pilot-plus-feedback overhead; and Bland–Altman bias and limits of agreement for downlink beam gain.

The decisive checks follow from the claims. On null data, the empirical false-alarm rate must match the nominal level across the full signal-to-noise range and on all three channel sources, while the frozen threshold visibly drifts. Detection power must exceed 0.8 for angle errors of a quarter beamwidth at moderate signal-to-noise ratio. With mismatch injected, test-gated correction must recover at least half the spectral-efficiency gap between the always-trust scheme and full training, at under 10% of full training's overhead. With no mismatch, it must cost under 5%.

If leakage in rich scattering inflates false alarms beyond nominal, the test as designed fails there. That outcome would itself be a useful map of where dominant-path reciprocity schemes should not be deployed.